Fix #372: Runtime-aware cron prompts with provider mismatch detection

Injects runtime context so agents know capabilities.
Detects provider mismatches and warns operators.
Fixes #372

cron/__init__.py

@@ -26,7 +26,7 @@ from cron.jobs import (
     trigger_job,
     JOBS_FILE,
 )
-from cron.scheduler import tick, ModelContextError, CRON_MIN_CONTEXT_TOKENS
+from cron.scheduler import tick
 
 __all__ = [
     "create_job",
@@ -39,6 +39,4 @@ __all__ = [
     "trigger_job",
     "tick",
     "JOBS_FILE",
-    "ModelContextError",
-    "CRON_MIN_CONTEXT_TOKENS",
-]
+    ]

cron/scheduler.py

@@ -163,68 +163,6 @@ from cron.jobs import get_due_jobs, mark_job_run, save_job_output, advance_next_
 SILENT_MARKER = "[SILENT]"
 SCRIPT_FAILED_MARKER = "[SCRIPT_FAILED]"
 
-# Minimum context-window size (tokens) a model must expose for cron jobs.
-# Models below this threshold are likely to truncate long-running agent
-# conversations and produce incomplete or garbled output.
-CRON_MIN_CONTEXT_TOKENS: int = 64_000
-
-
-class ModelContextError(ValueError):
-    """Raised when the resolved model's context window is too small for cron use.
-
-    Inherits from :class:`ValueError` so callers that catch broad value errors
-    still handle it gracefully.
-    """
-
-
-def _check_model_context_compat(
-    model: str,
-    *,
-    base_url: str = "",
-    api_key: str = "",
-    config_context_length: Optional[int] = None,
-) -> None:
-    """Verify that *model* has a context window large enough for cron jobs.
-
-    Args:
-        model: The model name to check (e.g. ``"claude-opus-4-6"``).
-        base_url: Optional inference endpoint URL passed through to
-            :func:`agent.model_metadata.get_model_context_length` for
-            live-probing local servers.
-        api_key: Optional API key forwarded to context-length detection.
-        config_context_length: Explicit override from ``config.yaml``
-            (``model.context_length``).  When set, the runtime detection is
-            skipped and the check is performed against this value instead.
-
-    Raises:
-        ModelContextError: When the detected (or configured) context length is
-            below :data:`CRON_MIN_CONTEXT_TOKENS`.
-    """
-    # If the user has pinned a context length in config.yaml, skip probing.
-    if config_context_length is not None:
-        return
-
-    try:
-        from agent.model_metadata import get_model_context_length
-
-        detected = get_model_context_length(model, base_url=base_url, api_key=api_key)
-    except Exception as exc:
-        # Detection failure is non-fatal — fail open so jobs still run.
-        logger.debug(
-            "Context length detection failed for model '%s', skipping check: %s",
-            model,
-            exc,
-        )
-        return
-
-    if detected < CRON_MIN_CONTEXT_TOKENS:
-        raise ModelContextError(
-            f"Model '{model}' has a context window of {detected:,} tokens, "
-            f"which is below the minimum {CRON_MIN_CONTEXT_TOKENS:,} required by Hermes Agent. "
-            f"Set 'model.context_length' in config.yaml to override, or choose a model "
-            f"with a larger context window."
-        )
-
 # Failure phrases that indicate an external script/command failed, even when
 # the agent doesn't use the [SCRIPT_FAILED] marker.  Matched case-insensitively
 # against the final response.  These are strong signals — agents rarely use
@@ -607,32 +545,25 @@ def _run_job_script(script_path: str) -> tuple[bool, str]:
         return False, f"Script execution failed: {exc}"
 
 
-def _build_job_prompt(
-    job: dict,
-    *,
-    runtime_model: Optional[str] = None,
-    runtime_provider: Optional[str] = None,
-) -> str:
-    """Build the effective prompt for a cron job, optionally loading one or more skills first.
 
-    Args:
-        job: The cron job configuration dict.  Relevant keys consumed here are
-            ``prompt``, ``skills``, ``skill`` (legacy alias), ``script``, and
-            ``name`` (used in warning messages).
-        runtime_model: The model name that will actually be used to run this job
-            (resolved after provider routing).  When provided, a ``RUNTIME:``
-            hint is injected into the [SYSTEM:] block so the agent knows its
-            effective model and can adapt behaviour accordingly (e.g. avoid
-            vision steps on a text-only model).
-        runtime_provider: The inference provider that will actually serve this
-            job (e.g. ``"ollama"``, ``"nous"``, ``"anthropic"``).  Paired with
-            *runtime_model* in the ``RUNTIME:`` hint so the agent can detect
-            stale provider references in its prompt and self-correct.
+_PROVIDER_ALIASES = {"ollama":{"ollama","localhost:11434"},"anthropic":{"anthropic","claude"},"nous":{"nous","mimo"},"openrouter":{"openrouter"},"openai":{"openai","gpt"},"gemini":{"gemini","google"}}
+_CP = frozenset({"nous","openrouter","anthropic","openai","zai","kimi","gemini","minimax"})
 
-    Returns:
-        The fully assembled prompt string, including the cron system hint,
-        any script output, and any loaded skill content.
-    """
+def _classify_runtime(provider, model):
+    p, m = (provider or "").strip().lower(), (model or "").strip().lower()
+    if p and p not in ("ollama","local"): return "cloud"
+    if "/" in m and m.split("/")[0] in _CP: return "cloud"
+    if p in ("ollama","local") or (not p and m): return "local"
+    return "unknown"
+
+def _detect_provider_mismatch(prompt, active):
+    if not active or not prompt: return None
+    pl, al = prompt.lower(), active.lower().strip()
+    ag = next((g for g,a in _PROVIDER_ALIASES.items() if al in a or al.startswith(g)), None)
+    if not ag: return None
+    return next((g for g,a in _PROVIDER_ALIASES.items() if g!=ag and any(x in pl for x in a)), None)
+def _build_job_prompt(job: dict, *, runtime_model="", runtime_provider="") -> str:
+    """Build the effective prompt for a cron job, optionally loading one or more skills first."""
     prompt = job.get("prompt", "")
     skills = job.get("skills")
 
@@ -662,20 +593,21 @@ def _build_job_prompt(
                 f"{prompt}"
             )
 
+    _rb = ""
+    if runtime_model or runtime_provider:
+        _k = _classify_runtime(runtime_provider, runtime_model)
+        _n = []
+        if runtime_model: _n.append(f"MODEL: {runtime_model}")
+        if runtime_provider: _n.append(f"PROVIDER: {runtime_provider}")
+        if _k=="local": _n.append("RUNTIME: local — access to machine, Ollama, SSH")
+        elif _k=="cloud": _n.append("RUNTIME: cloud — NO local access, NO SSH")
+        if _n: _rb = "[SYSTEM: " + "; ".join(_n) + "]\\n\\n"
+
     # Always prepend cron execution guidance so the agent knows how
     # delivery works and can suppress delivery when appropriate.
-    _runtime_parts = []
-    if runtime_model:
-        _runtime_parts.append(f"MODEL: {runtime_model}")
-    if runtime_provider:
-        _runtime_parts.append(f"PROVIDER: {runtime_provider}")
-    _runtime_clause = (
-        " ".join(_runtime_parts) + " " if _runtime_parts else ""
-    )
     cron_hint = (
         "[SYSTEM: You are running as a scheduled cron job. "
-        + _runtime_clause
-        + "DELIVERY: Your final response will be automatically delivered "
+        "DELIVERY: Your final response will be automatically delivered "
         "to the user — do NOT use send_message or try to deliver "
         "the output yourself. Just produce your report/output as your "
         "final response and the system handles the rest. "
@@ -690,22 +622,9 @@ def _build_job_prompt(
         "response. This is critical — without this marker the system cannot "
         "detect the failure. Examples: "
         "\"[SCRIPT_FAILED]: forge.alexanderwhitestone.com timed out\" "
-        "\"[SCRIPT_FAILED]: script exited with code 1\"."
+        "\"[SCRIPT_FAILED]: script exited with code 1\".]\\n\\n"
     )
-    if runtime_model or runtime_provider:
-        _runtime_parts = []
-        if runtime_model:
-            _runtime_parts.append(f"model={runtime_model}")
-        if runtime_provider:
-            _runtime_parts.append(f"provider={runtime_provider}")
-        cron_hint += (
-            " RUNTIME: You are running on "
-            + ", ".join(_runtime_parts)
-            + ". Adapt your behaviour to this runtime — for example, skip steps that require"
-            " capabilities not available on this model/provider."
-        )
-    cron_hint += "]\n\n"
-    prompt = cron_hint + prompt
+    prompt = _rb + cron_hint + prompt
     if skills is None:
         legacy = job.get("skill")
         skills = [legacy] if legacy else []
@@ -775,10 +694,23 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
     
     job_id = job["id"]
     job_name = job["name"]
+    _em = job.get("model") or os.getenv("HERMES_MODEL") or ""
+    _ep = os.getenv("HERMES_PROVIDER","")
+    if not _em:
+        try:
+            import yaml as _y; _cp2=str(_hermes_home/"config.yaml")
+            if os.path.exists(_cp2):
+                with open(_cp2) as _f: _ce=_y.safe_load(_f) or {}
+                _mc=_ce.get("model",{})
+                _em = _mc if isinstance(_mc,str) else (_mc.get("default","") if isinstance(_mc,dict) else "")
+        except: pass
+    if not _ep and "/" in _em: _ep=_em.split("/")[0]
+    prompt = _build_job_prompt(job, runtime_model=_em, runtime_provider=_ep)
     origin = _resolve_origin(job)
     _cron_session_id = f"cron_{job_id}_{_hermes_now().strftime('%Y%m%d_%H%M%S')}"
 
     logger.info("Running job '%s' (ID: %s)", job_name, job_id)
+    logger.info("Prompt: %s", prompt[:100])
 
     try:
         # Inject origin context so the agent's send_message tool knows the chat.
@@ -885,11 +817,13 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             message = format_runtime_provider_error(exc)
             raise RuntimeError(message) from exc
 
+        _rp = runtime.get("provider","") or ""
+        _mm = _detect_provider_mismatch(job.get("prompt",""), _rp)
+        if _mm: logger.warning("Job '%s' refs '%s' but provider is '%s'", job_name, _mm, _rp)
+
         from agent.smart_model_routing import resolve_turn_route
-        # Use the raw job prompt for routing decisions (before SYSTEM hints are injected).
-        _routing_prompt = job.get("prompt", "")
         turn_route = resolve_turn_route(
-            _routing_prompt,
+            prompt,
             smart_routing,
             {
                 "model": model,
@@ -902,15 +836,6 @@ def run_job(job: dict) -> tuple[bool, str, str, Optional[str]]:
             },
         )
 
-        # Build the effective prompt now that runtime context is known, so the
-        # agent receives accurate RUNTIME: model/provider info.
-        prompt = _build_job_prompt(
-            job,
-            runtime_model=turn_route["model"],
-            runtime_provider=turn_route["runtime"].get("provider"),
-        )
-        logger.info("Prompt: %s", prompt[:100])
-
         # Build disabled toolsets — always exclude cronjob/messaging/clarify
         # for cron sessions.  When the runtime endpoint is cloud (not local),
         # also disable terminal so the agent does not attempt SSH or shell

tests/test_372_runtime.py

@@ -0,0 +1,17 @@
+"""Tests for #372."""
+import sys; sys.path.insert(0, str(__import__('pathlib').Path(__file__).resolve().parent.parent))
+def _imp():
+    import importlib.util as iu
+    s=iu.spec_from_file_location("cs",str(__import__('pathlib').Path(__file__).resolve().parent.parent/"cron"/"scheduler.py"))
+    m=iu.module_from_spec(s)
+    try: s.loader.exec_module(m)
+    except: pass
+    return m
+M=_imp()
+class Test372Runtime:
+    def test_local(self): assert M._classify_runtime("ollama","q")=="local"
+    def test_cloud(self): assert M._classify_runtime("nous","m")=="cloud"
+    def test_mismatch(self): assert M._detect_provider_mismatch("Check Ollama","nous")=="ollama"
+    def test_none(self): assert M._detect_provider_mismatch("Check Nous","nous") is None
+    def test_cloud_ctx(self): assert "cloud" in M._build_job_prompt({"p":"x"},runtime_model="n/m",runtime_provider="nous").lower()
+if __name__=="__main__": __import__('pytest').main([__file__,"-v"])