diff --git a/.env.example b/.env.example
index 1f2cba1da5..908f7ce40e 100644
--- a/.env.example
+++ b/.env.example
@@ -22,9 +22,9 @@ HERMES_BACKEND=openai
 # of OpenRouter.
 #
 # Local server convenience (base URL without /v1):
-# llama.cpp example (see `Hermes-Agent/scripts/launch_llama_cpp_glm47_flash.sh`):
+# llama.cpp example (see `Hermes-Agent/scripts/launch_llama_cpp_hermes_4_36b.sh`):
 # ATROPOS_SERVER_BASE_URL=http://127.0.0.1:8080
-# ATROPOS_SERVER_MODEL=glm-4.7-flash
+# ATROPOS_SERVER_MODEL=hermes-4-36b
 # ATROPOS_SERVER_API_KEY=local
 #
 # Generic OpenAI-compatible (base URL should include /v1):
diff --git a/atropos/envs/agent_env.py b/atropos/envs/agent_env.py
index 789ea88b7c..ecd012b100 100644
--- a/atropos/envs/agent_env.py
+++ b/atropos/envs/agent_env.py
@@ -51,6 +51,13 @@ class AgentEnvConfig(BaseEnvConfig):
     max_containers: int = Field(default=10, description="Nomad mode: maximum containers")
     privileged: bool = Field(default=False, description="Nomad mode: run container privileged")
     acquire_timeout_s: float = Field(default=30.0, description="Slot acquisition timeout (seconds)")
+    purge_job_on_start: bool = Field(
+        default=False,
+        description=(
+            "Nomad mode: stop/purge the sandbox job on startup. This is helpful in local dev and training runs "
+            "to recover from previous crashes that leave the job in a restart backoff state."
+        ),
+    )
     purge_job_on_shutdown: bool = Field(default=True, description="Nomad mode: stop/purge job on shutdown")
 
     # basic agent defaults
@@ -229,6 +236,7 @@ class AgentEnv(BaseEnv, ABC, Generic[AgentEnvConfigT]):
                 max_containers=self.config.max_containers,
                 privileged=self.config.privileged,
                 acquire_timeout=self.config.acquire_timeout_s,
+                purge_job_on_start=bool(self.config.purge_job_on_start),
             )
         )
         await pool.start()
diff --git a/atropos/envs/hermes_compat_test_env.py b/atropos/envs/hermes_compat_test_env.py
index 93b0fe2dd9..be9cc40e0b 100644
--- a/atropos/envs/hermes_compat_test_env.py
+++ b/atropos/envs/hermes_compat_test_env.py
@@ -54,7 +54,7 @@ class HermesCompatTestEnvConfig(AgentEnvConfig):
         default="http://127.0.0.1:8080",
         description="Base URL for an OpenAI-compatible chat server (without /v1).",
     )
-    server_model: str = Field(default="glm-4.7-flash", description="Model name")
+    server_model: str = Field(default="hermes-4-36b", description="Model name")
 
 
 class HermesCompatTestEnv(AgentEnv[HermesCompatTestEnvConfig]):
@@ -79,7 +79,7 @@ class HermesCompatTestEnv(AgentEnv[HermesCompatTestEnvConfig]):
             or os.getenv("LLM_BASE_URL")
             or "http://127.0.0.1:8080"
         )
-        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "glm-4.7-flash"
+        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "hermes-4-36b"
         api_key = os.getenv("ATROPOS_SERVER_API_KEY") or os.getenv("OPENAI_API_KEY") or "local"
 
         env_config = HermesCompatTestEnvConfig(
@@ -97,6 +97,8 @@ class HermesCompatTestEnv(AgentEnv[HermesCompatTestEnvConfig]):
             disabled_toolsets=[],
             # Default to Nomad sandboxing; users can override via --env.* args.
             sandbox_image=os.getenv("ATROPOS_SANDBOX_IMAGE") or "atropos-sandbox:local",
+            # In local dev it's common for a previous crash to leave the job in backoff.
+            purge_job_on_start=True,
             purge_job_on_shutdown=True,
         )
 
diff --git a/atropos/envs/sandbox_terminal_smoke_env.py b/atropos/envs/sandbox_terminal_smoke_env.py
index 9c140a30b6..d850817730 100644
--- a/atropos/envs/sandbox_terminal_smoke_env.py
+++ b/atropos/envs/sandbox_terminal_smoke_env.py
@@ -55,7 +55,7 @@ class SandboxTerminalSmokeEnvConfig(AgentEnvConfig):
         default="http://127.0.0.1:8080",
         description="Base URL for an OpenAI-compatible chat server (without /v1).",
     )
-    server_model: str = Field(default="glm-4.7-flash", description="Model name")
+    server_model: str = Field(default="hermes-4-36b", description="Model name")
 
 
 class SandboxTerminalSmokeEnv(AgentEnv[SandboxTerminalSmokeEnvConfig]):
@@ -80,7 +80,7 @@ class SandboxTerminalSmokeEnv(AgentEnv[SandboxTerminalSmokeEnvConfig]):
             or os.getenv("LLM_BASE_URL")
             or "http://127.0.0.1:8080"
         )
-        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "glm-4.7-flash"
+        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "hermes-4-36b"
         api_key = os.getenv("ATROPOS_SERVER_API_KEY") or os.getenv("OPENAI_API_KEY") or "local"
 
         env_config = SandboxTerminalSmokeEnvConfig(
@@ -98,6 +98,7 @@ class SandboxTerminalSmokeEnv(AgentEnv[SandboxTerminalSmokeEnvConfig]):
             disabled_toolsets=[],
             # Default to Nomad sandboxing; users can override via --env.* args.
             sandbox_image=os.getenv("ATROPOS_SANDBOX_IMAGE") or "atropos-sandbox:local",
+            purge_job_on_start=True,
             purge_job_on_shutdown=True,
         )
 
diff --git a/atropos/envs/swe_smith_oracle_env.py b/atropos/envs/swe_smith_oracle_env.py
index 79f384495c..fa08a37633 100644
--- a/atropos/envs/swe_smith_oracle_env.py
+++ b/atropos/envs/swe_smith_oracle_env.py
@@ -74,7 +74,7 @@ class SweSmithOracleEnv(AgentEnv[SweSmithOracleEnvConfig]):
             or os.getenv("LLM_BASE_URL")
             or "http://127.0.0.1:8080"
         )
-        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "glm-4.7-flash"
+        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "hermes-4-36b"
         api_key = os.getenv("ATROPOS_SERVER_API_KEY") or os.getenv("OPENAI_API_KEY") or "local"
 
         env_config = SweSmithOracleEnvConfig(
diff --git a/atropos/envs/test_env.py b/atropos/envs/test_env.py
index 7c7f08e011..e078943b30 100644
--- a/atropos/envs/test_env.py
+++ b/atropos/envs/test_env.py
@@ -65,7 +65,7 @@ class SimpleTestEnvConfig(AgentEnvConfig):
         description="Base URL for an OpenAI-compatible server (without /v1)",
     )
     server_model: str = Field(
-        default="glm-4.7-flash",
+        default="hermes-4-36b",
         description="Model name",
     )
 
@@ -104,7 +104,7 @@ class SimpleTestEnv(AgentEnv[SimpleTestEnvConfig]):
             or os.getenv("LLM_BASE_URL")
             or "http://127.0.0.1:8080"
         )
-        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "glm-4.7-flash"
+        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "hermes-4-36b"
         api_key = os.getenv("ATROPOS_SERVER_API_KEY") or os.getenv("OPENAI_API_KEY") or "local"
 
         env_config = SimpleTestEnvConfig(
diff --git a/atropos/envs/toolserver_smoke_env.py b/atropos/envs/toolserver_smoke_env.py
index 4b39af468d..234e0d7452 100644
--- a/atropos/envs/toolserver_smoke_env.py
+++ b/atropos/envs/toolserver_smoke_env.py
@@ -33,7 +33,7 @@ class ToolServerSmokeEnvConfig(AgentEnvConfig):
         default="http://127.0.0.1:8080",
         description="Base URL for an OpenAI-compatible chat server (without /v1).",
     )
-    server_model: str = Field(default="glm-4.7-flash", description="Model name")
+    server_model: str = Field(default="hermes-4-36b", description="Model name")
 
 
 class ToolServerSmokeEnv(AgentEnv[ToolServerSmokeEnvConfig]):
@@ -58,7 +58,7 @@ class ToolServerSmokeEnv(AgentEnv[ToolServerSmokeEnvConfig]):
             or os.getenv("LLM_BASE_URL")
             or "http://127.0.0.1:8080"
         )
-        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "glm-4.7-flash"
+        model = os.getenv("ATROPOS_SERVER_MODEL") or os.getenv("LLM_MODEL") or "hermes-4-36b"
         api_key = os.getenv("ATROPOS_SERVER_API_KEY") or os.getenv("OPENAI_API_KEY") or "local"
 
         env_config = ToolServerSmokeEnvConfig(
@@ -76,6 +76,7 @@ class ToolServerSmokeEnv(AgentEnv[ToolServerSmokeEnvConfig]):
             # Self-contained ToolServer for local smoke.
             tool_server_url="inprocess",
             sandbox_image=os.getenv("ATROPOS_SANDBOX_IMAGE") or "atropos-sandbox:local",
+            purge_job_on_start=True,
             purge_job_on_shutdown=True,
         )
 
diff --git a/atropos/slots/pool.py b/atropos/slots/pool.py
index d6ace16b7d..1ee741cafa 100644
--- a/atropos/slots/pool.py
+++ b/atropos/slots/pool.py
@@ -50,6 +50,9 @@ class SlotPoolConfig:
     health_check_interval: float = 30.0  # Seconds between health checks
     scale_cooldown: float = 60.0  # Seconds between scale operations
 
+    # Job lifecycle
+    purge_job_on_start: bool = False  # Purge any pre-existing job before starting (local dev/training friendly)
+
 
 class SlotPool:
     """
@@ -144,7 +147,11 @@ class SlotPool:
             if not await self.nomad.is_healthy():
                 raise RuntimeError(f"Nomad is not reachable at {self.config.nomad_address}")
 
-            # Check if job exists
+            if self.config.purge_job_on_start:
+                logger.info(f"Purging any existing Nomad job: {self.config.job_id}")
+                await self.nomad.stop_job(self.config.job_id, purge=True)
+
+            # Check if job exists (after optional purge)
             job = await self.nomad.get_job(self.config.job_id)
 
             if job is None:
@@ -397,11 +404,30 @@ class SlotPool:
                 for task_name, st in task_states.items():
                     events = (st or {}).get("Events") or []
                     if isinstance(events, list) and events:
-                        last = events[-1]
-                        desc = last.get("DisplayMessage") or last.get("Message") or last.get("Type") or ""
-                        if desc:
-                            parts.append(f"{task_name}: {desc}")
+                        # Include a few recent events; the latest can be a generic restart message
+                        # while the true root cause is slightly earlier (e.g. image pull failure).
+                        recent = events[-3:]
+                        msgs: List[str] = []
+                        for ev in recent:
+                            desc = ev.get("DisplayMessage") or ev.get("Message") or ev.get("Type") or ""
+                            if desc:
+                                msgs.append(desc)
+                        if msgs:
+                            parts.append(f"{task_name}: " + " | ".join(msgs))
             return "; ".join(parts)
+
+        def _alloc_events_lower(detail: Dict[str, Any]) -> str:
+            task_states = detail.get("TaskStates") or {}
+            texts: List[str] = []
+            if isinstance(task_states, dict):
+                for _task_name, st in task_states.items():
+                    events = (st or {}).get("Events") or []
+                    if isinstance(events, list):
+                        for ev in events[-10:]:
+                            desc = ev.get("DisplayMessage") or ev.get("Message") or ev.get("Type") or ""
+                            if desc:
+                                texts.append(desc)
+            return " ".join(texts).lower()
         
         while time.time() - start < timeout:
             allocs = await self.nomad.get_job_allocations(self.config.job_id)
@@ -417,13 +443,23 @@ class SlotPool:
                     detail = await self.nomad.get_allocation(alloc.id)
                     if isinstance(detail, dict):
                         summary = _summarize_alloc_detail(detail)
-                        lowered = summary.lower()
+                        lowered = _alloc_events_lower(detail) or summary.lower()
                         if "failed to pull" in lowered or "pull access denied" in lowered:
                             raise RuntimeError(
                                 "Nomad allocation failed to start due to a Docker image pull error. "
                                 f"Allocation {alloc.id[:8]}: {summary}\n"
                                 "If you're using a local image tag (e.g. `atropos-sandbox:local`) on macOS, "
-                                "make sure the image is loaded into Docker (build with `docker buildx build --load ...`)."
+                                "make sure the image is loaded into Docker, e.g.:\n"
+                                "  docker buildx build --load -t atropos-sandbox:local -f Hermes-Agent/atropos/Dockerfile Hermes-Agent/atropos"
+                            )
+                        if "exceeded allowed attempts" in lowered:
+                            raise RuntimeError(
+                                "Nomad allocation is crash-looping and has entered restart backoff. "
+                                f"Allocation {alloc.id[:8]}: {summary}\n"
+                                "Inspect logs with:\n"
+                                f"  nomad alloc logs -stderr -task sandbox-server {alloc.id}\n"
+                                "Common causes include: missing local Docker image tag, container entrypoint error, "
+                                "or sandbox-server startup failure."
                             )
             
             if healthy_count >= min_count:
diff --git a/cli.py b/cli.py
index c1d0b9f880..e2e026a337 100755
--- a/cli.py
+++ b/cli.py
@@ -472,14 +472,33 @@ class HermesCLI:
         self.console = Console()
         self.compact = compact if compact is not None else CLI_CONFIG["display"].get("compact", False)
         self.verbose = verbose if verbose is not None else CLI_CONFIG["agent"].get("verbose", False)
+
+        self.backend = (backend or os.getenv("HERMES_BACKEND") or "openai").strip().lower()
+        if self.backend not in {"openai", "atropos"}:
+            self.console.print(
+                f"[bold yellow]Warning:[/] unknown backend '{self.backend}', falling back to 'openai'"
+            )
+            self.backend = "openai"
         
         # Configuration - priority: CLI args > env vars > config file
-        self.model = (
-            model
-            or os.getenv("LLM_MODEL")
-            or os.getenv("ATROPOS_SERVER_MODEL")
-            or CLI_CONFIG["model"]["default"]
-        )
+        #
+        # Note: For the Atropos backend we intentionally prefer `ATROPOS_SERVER_MODEL`
+        # over `LLM_MODEL`, because `LLM_MODEL` is commonly an OpenRouter-style ID
+        # (e.g. "anthropic/claude-sonnet-4") and will not exist on local servers.
+        if model:
+            self.model = model
+        elif self.backend == "atropos":
+            self.model = (
+                os.getenv("ATROPOS_SERVER_MODEL")
+                or os.getenv("LLM_MODEL")
+                or CLI_CONFIG["model"]["default"]
+            )
+        else:
+            self.model = (
+                os.getenv("LLM_MODEL")
+                or os.getenv("ATROPOS_SERVER_MODEL")
+                or CLI_CONFIG["model"]["default"]
+            )
 
         env_openai_base_url = os.getenv("OPENAI_BASE_URL")
         if env_openai_base_url:
@@ -507,13 +526,6 @@ class HermesCLI:
             or os.getenv("OPENROUTER_API_KEY")
         )
         self.max_turns = max_turns if max_turns != 20 else CLI_CONFIG["agent"].get("max_turns", 20)
-
-        self.backend = (backend or os.getenv("HERMES_BACKEND") or "openai").strip().lower()
-        if self.backend not in {"openai", "atropos"}:
-            self.console.print(
-                f"[bold yellow]Warning:[/] unknown backend '{self.backend}', falling back to 'openai'"
-            )
-            self.backend = "openai"
         
         # Parse and validate toolsets
         self.enabled_toolsets = toolsets
diff --git a/scripts/launch_llama_cpp_hermes_4_36b.sh b/scripts/launch_llama_cpp_hermes_4_36b.sh
new file mode 100755
index 0000000000..8e6f14be4e
--- /dev/null
+++ b/scripts/launch_llama_cpp_hermes_4_36b.sh
@@ -0,0 +1,63 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Launch a local llama.cpp OpenAI-compatible server running Hermes 4.3 36B (GGUF).
+#
+# Requires:
+# - `llama-server` installed (e.g. `brew install llama.cpp`)
+#
+# Note: Port choice can conflict with other local dev servers. If 8080 is already
+# in use, override via `LLAMA_CPP_PORT=...`.
+#
+# Usage:
+#   Hermes-Agent/scripts/launch_llama_cpp_hermes_4_36b.sh
+#
+# Override defaults:
+#   LLAMA_CPP_HOST=127.0.0.1 LLAMA_CPP_PORT=8082 \
+#   LLAMA_CPP_HF_REPO=NousResearch/Hermes-4.3-36B-GGUF \
+#   LLAMA_CPP_HF_FILE=hermes-4_3_36b-Q4_K_M.gguf \
+#   LLAMA_CPP_ALIAS=hermes-4-36b \
+#   Hermes-Agent/scripts/launch_llama_cpp_hermes_4_36b.sh
+
+HOST="${LLAMA_CPP_HOST:-127.0.0.1}"
+PORT="${LLAMA_CPP_PORT:-8080}"
+HF_REPO="${LLAMA_CPP_HF_REPO:-NousResearch/Hermes-4.3-36B-GGUF}"
+HF_FILE="${LLAMA_CPP_HF_FILE:-hermes-4_3_36b-Q4_K_M.gguf}"
+ALIAS="${LLAMA_CPP_ALIAS:-hermes-4-36b}"
+
+if ! command -v llama-server >/dev/null 2>&1; then
+  echo "Error: llama-server not found in PATH."
+  echo "Install via Homebrew: brew install llama.cpp"
+  exit 1
+fi
+
+echo "Launching llama.cpp server..."
+echo "  host:  $HOST"
+echo "  port:  $PORT"
+echo "  repo:  $HF_REPO"
+echo "  file:  $HF_FILE"
+echo "  alias: $ALIAS"
+echo
+echo "Suggested env vars for Hermes/Atropos integration:"
+echo "  export ATROPOS_SERVER_BASE_URL=http://${HOST}:${PORT}"
+echo "  export ATROPOS_SERVER_MODEL=${ALIAS}"
+echo "  export ATROPOS_SERVER_API_KEY=local"
+echo
+
+if command -v lsof >/dev/null 2>&1; then
+  if lsof -nP -iTCP:"$PORT" -sTCP:LISTEN >/dev/null 2>&1; then
+    echo "Error: port $PORT is already in use."
+    echo "Pick a different port, e.g.:"
+    echo "  LLAMA_CPP_PORT=8082 Hermes-Agent/scripts/launch_llama_cpp_hermes_4_36b.sh"
+    exit 1
+  fi
+fi
+
+exec llama-server \
+  --host "$HOST" \
+  --port "$PORT" \
+  --hf-repo "$HF_REPO" \
+  --hf-file "$HF_FILE" \
+  --alias "$ALIAS" \
+  -c 32768 \
+  -n -1
diff --git a/tests/test_data/checkpoint_test_dataset.jsonl b/tests/test_data/checkpoint_test_dataset.jsonl
new file mode 100644
index 0000000000..150e1c0afc
--- /dev/null
+++ b/tests/test_data/checkpoint_test_dataset.jsonl
@@ -0,0 +1,15 @@
+{"prompt": "Test prompt 0: What is 2+2? Just answer briefly.", "test_id": 0}
+{"prompt": "Test prompt 1: What is 2+2? Just answer briefly.", "test_id": 1}
+{"prompt": "Test prompt 2: What is 2+2? Just answer briefly.", "test_id": 2}
+{"prompt": "Test prompt 3: What is 2+2? Just answer briefly.", "test_id": 3}
+{"prompt": "Test prompt 4: What is 2+2? Just answer briefly.", "test_id": 4}
+{"prompt": "Test prompt 5: What is 2+2? Just answer briefly.", "test_id": 5}
+{"prompt": "Test prompt 6: What is 2+2? Just answer briefly.", "test_id": 6}
+{"prompt": "Test prompt 7: What is 2+2? Just answer briefly.", "test_id": 7}
+{"prompt": "Test prompt 8: What is 2+2? Just answer briefly.", "test_id": 8}
+{"prompt": "Test prompt 9: What is 2+2? Just answer briefly.", "test_id": 9}
+{"prompt": "Test prompt 10: What is 2+2? Just answer briefly.", "test_id": 10}
+{"prompt": "Test prompt 11: What is 2+2? Just answer briefly.", "test_id": 11}
+{"prompt": "Test prompt 12: What is 2+2? Just answer briefly.", "test_id": 12}
+{"prompt": "Test prompt 13: What is 2+2? Just answer briefly.", "test_id": 13}
+{"prompt": "Test prompt 14: What is 2+2? Just answer briefly.", "test_id": 14}
diff --git a/tests/test_data/checkpoint_test_resume_partial.jsonl b/tests/test_data/checkpoint_test_resume_partial.jsonl
new file mode 100644
index 0000000000..e001d382b5
--- /dev/null
+++ b/tests/test_data/checkpoint_test_resume_partial.jsonl
@@ -0,0 +1,5 @@
+{"prompt": "Test prompt 0: What is 2+2? Just answer briefly.", "test_id": 0}
+{"prompt": "Test prompt 1: What is 2+2? Just answer briefly.", "test_id": 1}
+{"prompt": "Test prompt 2: What is 2+2? Just answer briefly.", "test_id": 2}
+{"prompt": "Test prompt 3: What is 2+2? Just answer briefly.", "test_id": 3}
+{"prompt": "Test prompt 4: What is 2+2? Just answer briefly.", "test_id": 4}