VSCode and Cursor bind Ctrl+G to "Find Next" at the editor level, so
the keystroke never reaches the embedded terminal — Ctrl+G to open
\$EDITOR was effectively dead inside those IDEs.
Alt+G is unbound in both editors and reaches the TUI cleanly as
`\x1bg` → `key.meta && ch === 'g'` after parse-keypress. Accept it
alongside the existing isAction(key, ch, 'g') check, and document the
fallback in README + the hotkeys panel.
The Ctrl+G handler was toggling the alt-screen by hand
(`\x1b[?1049l` ... `\x1b[?1049h`) without releasing stdin or kitty
keyboard mode, so the launched editor would lose keystrokes (Ink kept
swallowing them) and editors that don't speak CSI-u (e.g. nano) would
print "Unknown sequence" for every Ctrl-key.
Switch to `withInkSuspended` from @hermes/ink, the same helper
`/setup` already uses. It pauses Ink, removes stdin listeners, drops
raw mode, disables kitty/modifyOtherKeys + mouse + focus reporting,
runs the editor, then restores everything with a full repaint.
Previously _copy_reasoning_content_for_api only padded reasoning_content
when the assistant message had tool_calls. DeepSeek V4 thinking mode
requires the field on every assistant turn, including plain text replies
without tool_calls.
- Remove the 'source_msg.get("tool_calls") and' guard
- Update test: plain assistant turns now get padded for DeepSeek/Kimi
Fixes#15213
- webhooks.md: adds a Video Tutorial section under the intro with a
responsive YouTube iframe (WNYe5mD4fY8).
- configuration.md: adds a Video Tutorial subsection under Auxiliary
Models with a responsive YouTube iframe (NoF-YajElIM).
Both use a 16:9 aspect-ratio wrapper so the embeds scale cleanly on
mobile. Verified with `npm run build` — MDX parses clean, no new
warnings or broken links introduced.
Adds a 'Video Guide' section pointing at the walkthrough of a Hermes agent
abliterating Gemma with OBLITERATUS, so the agent can surface it when the
user wants a visual overview before running the workflow.
- add a written-cell bitmap so selection can distinguish rendered spaces from blank padding
- preserve code indentation without markdown-specific rendering hacks
- clamp selection highlight to real row content so blank drag margins do not render or copy
- keep successful copy actions quiet while preserving usage and failure feedback
- accept forwarded Cmd+C for selection copy in SSH sessions even when Hermes runs on Linux
- keep local Linux Alt+C from acting as copy and update TUI hotkey hints for remote shells
- add reusable overlay key and help-text helpers for picker-style overlays
- make model, session, skills, and pager hints consistently support Esc/q close behavior
- expand short model aliases like sonnet/opus via static catalogs during startup runtime resolution
- keep startup alias resolution network-free and add regression tests in models and tui gateway suites
- run the requested ui-tui lint+format pass and include resulting formatting updates
- guard text-measure cache eviction key in hermes-ink so ui-tui type-check stays green
The Codex Responses API rejects input_text inside assistant messages —
only output_text and refusal are valid content types for assistant role.
_chat_content_to_responses_parts() previously hardcoded all text content
to input_text regardless of the message role. When an assistant message
had list-format content (multimodal or structured), this produced invalid
input_text parts that the API rejected with:
Invalid value: 'input_text'. Supported values are: 'output_text' and 'refusal'.
Fix: add a role parameter to _chat_content_to_responses_parts() that
selects output_text for assistant messages and input_text for user
messages. Thread this through _chat_messages_to_responses_input() and
_preflight_codex_input_items().
Fixes#15687
When a user sends /stop during a streaming API call, the outer poll loop
detects _interrupt_requested and closes the HTTP connection. However, the
inner _call() thread catches the connection error and enters its retry
loop — opening a FRESH connection without checking the interrupt flag.
On slow providers like ollama-cloud, each retry attempt blocks for the
full stream-read timeout (120s+). With 3 retry attempts this caused
510+ second delays between /stop and actual response — the agent appeared
completely unresponsive despite the stop being acknowledged.
Fix: add an _interrupt_requested check at the top of the streaming retry
loop so the agent exits immediately instead of retrying.
Also fix log truncation: all session key logging in gateway/run.py used
[:20] or [:30] slices, which truncated 'agent:main:telegram:dm:5690190437'
(33 chars) to 'agent:main:telegram:' — losing the identifying chat type
and user ID. Replace with full keys to make logs debuggable.
Reported by user Sidharth Pulipaka via Telegram on ollama-cloud provider.
The post-graceful-drain is-active poll used a fixed 10s timeout, but
systemd's hermes-gateway.service has RestartSec=30 — so systemd won't
respawn the unit for 30s after exit-75, and our poll gives up during
the cooldown. Result: every 'hermes update' printed
⚠ hermes-gateway drained but didn't relaunch — forcing restart
followed by a redundant 'systemctl restart' that kicked the newly-
respawning gateway again (and re-started WhatsApp / Discord a second
time in the process).
Fix: read RestartUSec from the unit via 'systemctl show' and set the
poll budget to max(10s, RestartSec + 10s slack). Units without
RestartSec set (or value=infinity) fall back to the original 10s.
Observed timeline from journalctl before fix:
08:56:22.262 old PID exits 75
08:56:32.707 systemd logs Stopped -> Started (10.4s gap, > 10s budget)
After fix the poll covers 40s — comfortably inside RestartSec + slack.
Validation:
- RestartUSec parser tested against '30s', '100ms', '1min 30s',
'infinity', '', 'garbage', '500us', '2min' — all correct.
- Against the live hermes-gateway.service: parses to 30.0s.
- tests/hermes_cli/test_update_gateway_restart.py: 41/41 pass.
Makes hermes -z usable by sweeper without mutating user config.
- Top-level -m/--model and --provider flags that apply to -z/--oneshot
(mirrors hermes chat's plumbing).
- HERMES_INFERENCE_MODEL env var as the parallel to HERMES_INFERENCE_PROVIDER
for CI / scripted invocations.
- resolve_runtime_provider() gets the requested provider; when --model is
given without --provider, detect_provider_for_model() auto-selects the
provider that serves it (same semantic as /model in an interactive session).
- --provider without --model errors out with exit 2 — carrying a config
model across to a different provider is usually wrong, and silently
picking the provider's catalog default hides the mismatch.
Config defaults still used when both flags are omitted (existing behavior).
Validation (all live against OpenRouter):
-z 'x' ....................... uses config default (opus-4.7)
-z 'x' --model haiku-4.5 ..... haiku-4.5 via auto-detected openrouter
-z 'x' --model ... --provider pair as given
HERMES_INFERENCE_MODEL=... -z haiku-4.5 via env var
-z 'x' --provider anthropic .. exits 2 with error to stderr
* feat: add `hermes -z <prompt>` one-shot mode
Top-level flag that runs a single prompt and prints ONLY the final
response text to stdout. No banner, no spinner, no tool previews, no
session_id line — stdout is machine-readable, stderr is silent.
Tools, memory, rules, and AGENTS.md in the CWD are loaded as normal.
Approvals are auto-bypassed (sets HERMES_YOLO_MODE=1 for the call).
Bypasses cli.py entirely — goes straight to AIAgent.chat().
* feat(oneshot): handle interactive-callback gaps explicitly
Document (and where needed, patch) the interactive surfaces that have
no user to answer in oneshot mode:
- clarify — inject a callback that tells the agent to pick the
best default and continue (previously returned a
generic 'not available in this execution context'
error that wastes a tool call)
- sudo password — terminal_tool already gates on HERMES_INTERACTIVE
(we don't set it); sudo fails gracefully
- shell hooks — HERMES_ACCEPT_HOOKS=1 auto-approves; also falls
back to deny on non-tty stdin
- dangerous cmd — HERMES_YOLO_MODE=1 short-circuits before input()
- secret capture— tool returns gracefully when no callback wired
Live-tested: agent asked clarify(['red','blue']) and got 'red' back,
replied with only 'red'.
The AIAgent.flush_memories pre-compression save, the gateway
_flush_memories_for_session, and everything feeding them are
obsolete now that the background memory/skill review handles
persistent memory extraction.
Problems with flush_memories:
- Pre-dates the background review loop. It was the only memory-save
path when introduced; the background review now fires every 10 user
turns on CLI and gateway alike, which is far more frequent than
compression or session reset ever triggered flush.
- Blocking and synchronous. Pre-compression flush ran on the live agent
before compression, blocking the user-visible response.
- Cache-breaking. Flush built a temporary conversation prefix
(system prompt + memory-only tool list) that diverged from the live
conversation's cached prefix, invalidating prompt caching. The
gateway variant spawned a fresh AIAgent with its own clean prompt
for each finalized session — still cache-breaking, just in a
different process.
- Redundant. Background review runs in the live conversation's
session context, gets the same content, writes to the same memory
store, and doesn't break the cache. Everything flush_memories
claimed to preserve is already covered.
What this removes:
- AIAgent.flush_memories() method (~248 LOC in run_agent.py)
- Pre-compression flush call in _compress_context
- flush_memories call sites in cli.py (/new + exit)
- GatewayRunner._flush_memories_for_session + _async_flush_memories
(and the 3 call sites: session expiry watcher, /new, /resume)
- 'flush_memories' entry from DEFAULT_CONFIG auxiliary tasks,
hermes tools UI task list, auxiliary_client docstrings
- _memory_flush_min_turns config + init
- #15631's headroom-deduction math in
_check_compression_model_feasibility (headroom was only needed
because flush dragged the full main-agent system prompt along;
the compression summariser sends a single user-role prompt so
new_threshold = aux_context is safe again)
- The dedicated test files and assertions that exercised
flush-specific paths
What this renames (with read-time backcompat on sessions.json):
- SessionEntry.memory_flushed -> SessionEntry.expiry_finalized.
The session-expiry watcher still uses the flag to avoid re-running
finalize/eviction on the same expired session; the new name
reflects what it now actually gates. from_dict() reads
'expiry_finalized' first, falls back to the legacy 'memory_flushed'
key so existing sessions.json files upgrade seamlessly.
Supersedes #15631 and #15638.
Tested: 383 targeted tests pass across run_agent/, agent/, cli/,
and gateway/ session-boundary suites. No behavior regressions —
background memory review continues to handle persistent memory
extraction on both CLI and gateway.
_check_compression_model_feasibility calls get_model_context_length
without provider=, so Codex OAuth users get 1,050,000 (from models.dev
for 'openai') instead of the actual 272,000 limit. This happens because
_infer_provider_from_url maps chatgpt.com → 'openai' (not 'openai-codex'),
skipping the Codex-specific resolution branch entirely.
Result: compression threshold set at 85% of 1.05M = 892K — conversations
never trigger compression, the context grows unbounded, and when gateway
hygiene eventually forces compression, the Codex endpoint drops the
oversized streaming request ('peer closed connection without sending
complete message body').
Fix: forward self.provider to get_model_context_length so provider-
specific resolution branches (Codex OAuth 272K, Copilot live /models,
Nous suffix-match) fire correctly.
Reported by user on GPT 5.5 via Codex OAuth Pro (paste.rs/vsra3).
Follow-up to PR #15658. The feature PR introduced page-scoped slots
(<page>:top / <page>:bottom inside every built-in page) but only
touched the Shell slots catalogue. Adds proper narrative coverage so
plugin authors find the feature.
Changes
- extending-the-dashboard.md:
- Frontmatter description + intro bullet now mention page-scoped slots
- New TOC entry "Augmenting built-in pages (page-scoped slots)"
- New dedicated subsection after "Replacing built-in pages"
explaining the heavy-vs-light tradeoff, listing the pages that
expose slots, and showing a worked manifest + IIFE example with
tab.hidden: true
- Cross-link from the tab.override section pointing readers to the
lighter augmentation option
- web-dashboard.md:
- Bullet mentioning "page-scoped slots (inject widgets into
built-in pages without overriding them)"
Validation
- TOC anchor "#augmenting-built-in-pages-page-scoped-slots" matches
the generated heading slug
- Code fences balanced (64, even)
- Pre-existing docusaurus build errors (skills.json, api-server.md
link) reproduce on bare main -- not introduced here
* fix(terminal): three-layer defense against watch_patterns notification spam
Background processes that stack notify_on_complete=True with watch_patterns
can flood the user with duplicate, delayed notifications — matches deliver
asynchronously via the completion queue and continue arriving minutes after
the process has exited. The docstring warning against this (PR #12113) has
proven insufficient; agents still misuse the combination.
Three layered defenses, each sufficient on its own:
1. Mutual exclusion (terminal_tool.py): When both flags are set on a
background process, drop watch_patterns with a warning. notify_on_complete
wins because 'let me know when it's done' is the more useful signal and
fires exactly once. Extracted as _resolve_notification_flag_conflict() so
the rule is testable in isolation.
2. Suppress-after-exit (process_registry.py): _check_watch_patterns() now
bails the moment session.exited is True. Post-exit chunks (buffered reads
draining after the process is gone) no longer produce notifications. This
is the fix flagged as future work in session 20260418_020302_79881c.
3. Global circuit breaker (process_registry.py): Per-session rate limits don't
catch the sibling-flood case — N concurrent processes can each stay under
8/10s and still collectively spam. New WATCH_GLOBAL_MAX_PER_WINDOW=15 cap
trips a 30-second cooldown across ALL sessions, emits a single
watch_overflow_tripped event, silently counts dropped events, and emits a
watch_overflow_released summary when the cooldown ends.
Also updates the tool schema + docstring to document the new behavior.
Tests: 8 new tests covering all three fixes (suppress-after-exit x2,
mutual-exclusion resolver x4, global breaker trip/cooldown/release x2).
All 60 tests across test_watch_patterns.py, test_notify_on_complete.py,
test_terminal_tool.py pass.
Real-world trigger: self-inflicted in session 20260425_051924 — three
concurrent hermes-sweeper review subprocesses each set watch_patterns=
['failed validation', 'errored'] AND notify_on_complete=True, then iterated
over multiple items, producing enough matches per process to defeat the
per-session cap while staying under the global cap that didn't yet exist.
* fix(terminal): aggressive 1-per-15s watch_patterns rate limit + strike-3 promotion
Per Teknium's direction, the watch_patterns rate limit is now much more
aggressive and self-healing.
## New rule — per session
- HARD cap: 1 watch-match notification per 15 seconds per process.
- Any match arriving inside the cooldown window is dropped and counts as
ONE strike for that window (many drops in the same window still = 1 strike).
- After 3 consecutive strike windows, watch_patterns is permanently disabled
for the session and the session is auto-promoted to notify_on_complete
semantics — exactly one notification when the process actually exits.
- A cooldown window that expires with zero drops resets the consecutive
strike counter — healthy cadence is forgiven.
## Schema + docstring rewritten
The tool schema description now gives the model explicit guidance:
- notify_on_complete is 'the right choice for almost every long-running task'
- watch_patterns is for RARE one-shot signals on LONG-LIVED processes
- Do NOT use watch_patterns with loops/batch jobs — error patterns fire every
iteration and will hit the strike limit fast
- Mutual exclusion is stated on both parameter descriptions
- 1/15s cooldown and 3-strike promotion are stated in the watch_patterns
description so the model sees the contract every turn
## Removed
- WATCH_MAX_PER_WINDOW (8/10s) and WATCH_OVERLOAD_KILL_SECONDS (45) — the
new 1/15s limit subsumes both; keeping them would double-count.
- _watch_window_hits / _watch_window_start / _watch_overload_since fields
on ProcessSession. Replaced by _watch_last_emit_at / _watch_cooldown_until
/ _watch_strike_candidate / _watch_consecutive_strikes.
## Kept
- Global circuit breaker across all sessions (15/10s → 30s cooldown) as a
secondary safety net for concurrent siblings. Still valuable when 20
short-lived processes each fire once — none individually violates the
per-session limit.
- Suppress-after-exit guard.
- Mutual exclusion resolver at the tool entry point.
## Tests
- 6 new tests in TestPerSessionRateLimit covering: first match delivers,
second in cooldown suppressed, multi-drop = single strike, 3 strikes
disables + promotes, clean window resets counter, suppressed count
carried to next emit.
- Global circuit breaker tests rewritten to use fresh sessions instead of
hacking removed per-window fields.
- 50/50 watch_patterns + notify_on_complete tests pass.
- 60/60 including test_terminal_tool.py pass.
* feat(dashboard): page-scoped plugin slots for built-in pages
Dashboard plugins can now inject components into specific built-in
pages (Sessions, Analytics, Logs, Cron, Skills, Config, Env, Docs,
Chat) without overriding the whole route.
Previously, plugins could only:
1. Add new tabs (tab.path)
2. Replace whole built-in pages (tab.override)
3. Inject into global shell slots (header-*, footer-*, pre-main, ...)
None of those let a plugin add a banner, card, or widget to an
existing page. The new <page>:top / <page>:bottom slots close that
gap, reusing the existing registerSlot() API.
Changes
- web/src/plugins/slots.ts: 18 new KNOWN_SLOT_NAMES entries
(sessions:top, sessions:bottom, analytics:top, ..., chat:bottom),
grouped under "Shell-wide" vs "Page-scoped" in the docblock
- web/src/pages/*: each built-in page now renders
<PluginSlot name="<page>:top" />
as the first child of its outer wrapper and
<PluginSlot name="<page>:bottom" />
as the last child -- zero visual cost when no plugin registers
- plugins/example-dashboard: registers a demo banner into
sessions:top via registerSlot(), with matching slots entry in
the manifest -- so freshly-setup users can see what page-scoped
slots look like without writing any plugin code
- website/docs: new "Page-scoped slots" table in the plugin
authoring guide, with a worked example
- tests/hermes_cli/test_web_server.py: round-trip test for
colon-bearing slot names (sessions:top, analytics:bottom, ...)
Validation
- npm run build: clean (tsc -b + vite build, 2761 modules)
- scripts/run_tests.sh tests/hermes_cli/test_web_server.py::TestDashboardPluginManifestExtensions: 5/5 pass
* fix(terminal): three-layer defense against watch_patterns notification spam
Background processes that stack notify_on_complete=True with watch_patterns
can flood the user with duplicate, delayed notifications — matches deliver
asynchronously via the completion queue and continue arriving minutes after
the process has exited. The docstring warning against this (PR #12113) has
proven insufficient; agents still misuse the combination.
Three layered defenses, each sufficient on its own:
1. Mutual exclusion (terminal_tool.py): When both flags are set on a
background process, drop watch_patterns with a warning. notify_on_complete
wins because 'let me know when it's done' is the more useful signal and
fires exactly once. Extracted as _resolve_notification_flag_conflict() so
the rule is testable in isolation.
2. Suppress-after-exit (process_registry.py): _check_watch_patterns() now
bails the moment session.exited is True. Post-exit chunks (buffered reads
draining after the process is gone) no longer produce notifications. This
is the fix flagged as future work in session 20260418_020302_79881c.
3. Global circuit breaker (process_registry.py): Per-session rate limits don't
catch the sibling-flood case — N concurrent processes can each stay under
8/10s and still collectively spam. New WATCH_GLOBAL_MAX_PER_WINDOW=15 cap
trips a 30-second cooldown across ALL sessions, emits a single
watch_overflow_tripped event, silently counts dropped events, and emits a
watch_overflow_released summary when the cooldown ends.
Also updates the tool schema + docstring to document the new behavior.
Tests: 8 new tests covering all three fixes (suppress-after-exit x2,
mutual-exclusion resolver x4, global breaker trip/cooldown/release x2).
All 60 tests across test_watch_patterns.py, test_notify_on_complete.py,
test_terminal_tool.py pass.
Real-world trigger: self-inflicted in session 20260425_051924 — three
concurrent hermes-sweeper review subprocesses each set watch_patterns=
['failed validation', 'errored'] AND notify_on_complete=True, then iterated
over multiple items, producing enough matches per process to defeat the
per-session cap while staying under the global cap that didn't yet exist.
* fix(terminal): aggressive 1-per-15s watch_patterns rate limit + strike-3 promotion
Per Teknium's direction, the watch_patterns rate limit is now much more
aggressive and self-healing.
## New rule — per session
- HARD cap: 1 watch-match notification per 15 seconds per process.
- Any match arriving inside the cooldown window is dropped and counts as
ONE strike for that window (many drops in the same window still = 1 strike).
- After 3 consecutive strike windows, watch_patterns is permanently disabled
for the session and the session is auto-promoted to notify_on_complete
semantics — exactly one notification when the process actually exits.
- A cooldown window that expires with zero drops resets the consecutive
strike counter — healthy cadence is forgiven.
## Schema + docstring rewritten
The tool schema description now gives the model explicit guidance:
- notify_on_complete is 'the right choice for almost every long-running task'
- watch_patterns is for RARE one-shot signals on LONG-LIVED processes
- Do NOT use watch_patterns with loops/batch jobs — error patterns fire every
iteration and will hit the strike limit fast
- Mutual exclusion is stated on both parameter descriptions
- 1/15s cooldown and 3-strike promotion are stated in the watch_patterns
description so the model sees the contract every turn
## Removed
- WATCH_MAX_PER_WINDOW (8/10s) and WATCH_OVERLOAD_KILL_SECONDS (45) — the
new 1/15s limit subsumes both; keeping them would double-count.
- _watch_window_hits / _watch_window_start / _watch_overload_since fields
on ProcessSession. Replaced by _watch_last_emit_at / _watch_cooldown_until
/ _watch_strike_candidate / _watch_consecutive_strikes.
## Kept
- Global circuit breaker across all sessions (15/10s → 30s cooldown) as a
secondary safety net for concurrent siblings. Still valuable when 20
short-lived processes each fire once — none individually violates the
per-session limit.
- Suppress-after-exit guard.
- Mutual exclusion resolver at the tool entry point.
## Tests
- 6 new tests in TestPerSessionRateLimit covering: first match delivers,
second in cooldown suppressed, multi-drop = single strike, 3 strikes
disables + promotes, clean window resets counter, suppressed count
carried to next emit.
- Global circuit breaker tests rewritten to use fresh sessions instead of
hacking removed per-window fields.
- 50/50 watch_patterns + notify_on_complete tests pass.
- 60/60 including test_terminal_tool.py pass.
The auto-restart path in `hermes update` verifies systemd unit health with
`time.sleep(3)` + a single `systemctl is-active` call. The unit's
Stopped -> Started transition after a graceful SIGUSR1 exit (or a hard
restart) is not always complete inside that 3s window, so the verify
races and reports 'drained but didn't relaunch' even though systemd is
about to bring the unit back up a fraction of a second later. Users
then see a spurious warning, a redundant fallback `systemctl restart`
fires, and adapters (Discord, WhatsApp) get restarted twice.
Replace the three sleep+oneshot sites with a small `_wait_for_service_active()`
closure that polls `is-active` every 0.5s for up to 10s. Behaviour
is unchanged when the unit is healthy or truly dead — only the race
window around a clean restart is now handled correctly.
Tests: tests/hermes_cli/test_update_gateway_restart.py (41/41).
`hermes tools` → "reconfigure existing" listed Spotify twice because
the Apr 24 refactor that moved Spotify into plugins/spotify/ (PR #15174)
left the entry in CONFIGURABLE_TOOLSETS. _get_effective_configurable_toolsets()
unconditionally appended get_plugin_toolsets() on top, so the same
'spotify' key showed up from both sources.
Dedupe by key — built-in CONFIGURABLE_TOOLSETS entry wins (it has the
nicer label and description). Also guards against future bundled plugins
that share a toolset key with a built-in.
Generalize the temperature-specific 400 retry that shipped in PR #15621 so
the same reactive strategy covers any provider that rejects an arbitrary
request parameter — — not just temperature.
- agent/auxiliary_client.py:
* New _is_unsupported_parameter_error(exc, param): matches the same six
phrasings the old temperature detector did plus 'unrecognized parameter'
and 'invalid parameter', against any named param.
* _is_unsupported_temperature_error is now a thin back-compat wrapper so
existing imports and tests keep working.
* The max_tokens → max_completion_tokens retry branch in call_llm and
async_call_llm now (a) gates on 'max_tokens is not None' so we do not
pop a key that was never set and silently substitute a None value on
the retry, and (b) also matches the generic helper in addition to the
legacy 'max_tokens' / 'unsupported_parameter' substring checks — picking
up phrasings like 'Unknown parameter: max_tokens' that previously slipped
through.
- tests/agent/test_unsupported_parameter_retry.py: 18 new tests covering
the generic detector across params, the back-compat wrapper, and the two
hardenings to the max_tokens retry branch (None gate + generic phrasing).
Credit: retry-generalization pattern from @nicholasrae's PR #15416. That PR
also proposed the reactive temperature retry which landed independently via
PR #15621 + #15623 (co-authored with @BlueBirdBack). This commit salvages
the remaining hardening ideas onto current main.
When the auxiliary compression model's context is smaller than the main
model's compression threshold, _check_compression_model_feasibility
auto-lowers the session threshold. Previously it set:
new_threshold = aux_context
This let the raw message list grow to exactly aux_context tokens. But
compression and flush_memories actually send system_prompt + tool_schemas
+ messages to the aux model. With 50+ tools that overhead is 25-30K
tokens, so the full request overflowed aux with HTTP 400.
Subtract a headroom estimate from aux_context before setting the new
threshold: the actual tool-schema token count (from
estimate_request_tokens_rough) plus a 12K allowance for the system
prompt (not yet built at __init__ time) and flush-instruction overhead.
Clamp to MINIMUM_CONTEXT_LENGTH so the session still starts even with
an unusually heavy tool schema.
This fixes the 'flush_memories overflow on busy toolsets' path that
Teknium flagged — where main and aux can be nominally the same model
but still 400 because the threshold left no room for the request
overhead. Same fix also protects the normal compression summarisation
request on the same binding aux.
Tests: two new regression tests cover the headroom reservation and the
MINIMUM_CONTEXT_LENGTH floor. Two existing tests updated for the new
(lower) threshold values now that empty-tools still produces a 12K
static headroom deduction.
Universal reactive fix for 'HTTP 400: Unsupported parameter: temperature'
across all providers/models — not just Codex Responses.
The same backend can accept temperature for some models and reject it for
others (e.g. gpt-5.4 accepts but gpt-5.5 rejects on the same OpenAI
endpoint; similar patterns on Copilot, OpenRouter reasoning routes, and
Anthropic Opus 4.7+ via OAI-compat). An allow/deny-list by model name does
not scale.
call_llm / async_call_llm now detect the concrete 'unsupported parameter:
temperature' 400 and transparently retry once without temperature. Kimi's
server-managed omission and Opus 4.7+'s proactive strip stay in place —
this is the safety net for everything else.
Changes:
- agent/auxiliary_client.py: add _is_unsupported_temperature_error helper;
wire into both sync and async call_llm paths before the existing
max_tokens/payment/auth retry ladder
- tests/agent/test_unsupported_temperature_retry.py: 19 tests covering
detector phrasings, sync + async retry, no-retry-without-temperature,
and non-temperature 400s not triggering the retry
Builds on PR #15620 (codex_responses fallback) which stripped temperature
up front for that one api_mode. This PR closes the gap for every other
provider/model combo via reactive retry.
Credit: retry approach and detector originate from @BlueBirdBack's PR #15578.
Co-authored-by: BlueBirdBack <BlueBirdBack@users.noreply.github.com>
