fix(agents): survive long overnight runs

- compact at memory-safe token budget (~20k) not the 131k window; gpt-oss
  hits oMLX's prefill memory guard ~47k tokens on a 24GB machine
- repair invalid_tool_calls (malformed JSON args) before echoing history —
  LangChain serializes the raw malformed string which oMLX 500s on
- retry invokes (4x) with an oMLX watchdog restart between attempts
- ALWAYS finalize: salvage picks from findings even after invoke failures

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016DvdRZkyCy9NJpHYDdPqmS

agents/graph.py

@@ -89,6 +89,55 @@ def _short(args: dict) -> str:
         return str(args)[:60]
 
 
+def _sanitize_tool_calls(messages: list) -> None:
+    """Re-encode historical tool-call arguments to valid JSON, in place.
+
+    gpt-oss sometimes emits a tool call (especially run_analysis with a big multi-line code
+    string) whose raw `arguments` aren't valid JSON. It executes fine once, but when that
+    assistant turn is echoed back in the next request, oMLX's FunctionCall validator rejects
+    it with a 500 ('arguments must be valid JSON') and the whole agent dies. We rewrite the
+    stored raw arguments from the cleanly-parsed args dict so every echo is valid JSON.
+    """
+    for m in messages:
+        ak = getattr(m, "additional_kwargs", None)
+        if not isinstance(ak, dict):
+            continue
+        raws = ak.get("tool_calls")
+        if not raws:
+            continue
+        parsed = getattr(m, "tool_calls", None) or []
+        for i, raw in enumerate(raws):
+            fn = raw.get("function")
+            if not isinstance(fn, dict):
+                continue
+            args = parsed[i]["args"] if i < len(parsed) and isinstance(parsed[i].get("args"), dict) else {}
+            fn["arguments"] = json.dumps(args)
+
+
+def _repair_invalid_tool_calls(ai) -> list[tuple[str, str]]:
+    """Convert `invalid_tool_calls` (malformed-JSON args) into valid empty-args calls, in place.
+
+    LangChain serializes invalid_tool_calls with their RAW malformed argument string — which
+    oMLX's validator 500s on at the next request (this bypasses additional_kwargs, so the
+    sanitizer above can't catch it). We convert each into a valid call with args={} and return
+    (id, name) pairs so the loop can attach an error ToolMessage telling the model to retry
+    with valid JSON.
+    """
+    import uuid
+
+    repaired: list[tuple[str, str]] = []
+    invalid = getattr(ai, "invalid_tool_calls", None) or []
+    if not invalid:
+        return repaired
+    for itc in invalid:
+        name = ((itc.get("name") or "unknown").split("<|")[0].strip()) or "unknown"
+        tc_id = itc.get("id") or f"call_{uuid.uuid4().hex[:8]}"
+        ai.tool_calls.append({"name": name, "args": {}, "id": tc_id, "type": "tool_call"})
+        repaired.append((tc_id, name))
+    ai.invalid_tool_calls = []
+    return repaired
+
+
 def _prompt_tokens(ai, messages: list) -> int:
     """Best available count of tokens the last prompt consumed (for compaction timing)."""
     um = getattr(ai, "usage_metadata", None) or {}
@@ -197,7 +246,10 @@ def run_agent(
     tools_by_name = {t.name: t for t in tools}
     tool_llm = build_llm(tools=tools)
     plain_llm = build_llm()  # no tools — used for summarization
-    ctx_len = get_settings().llm_context_length
+    # Compaction ceiling is the MEMORY-safe KV budget, not the model's 131k window: gpt-oss
+    # hits oMLX's prefill memory guard around ~47k tokens on a 24GB Mac, so we compact well
+    # below that to keep long (2hr) runs alive.
+    compact_budget = get_settings().compaction_token_budget
 
     date_str = datetime.now().astimezone().strftime("%A %Y-%m-%d")
     messages: list = [
@@ -231,15 +283,43 @@ def run_agent(
             ended_reason = "timeout"
             break
         steps_used += 1
-        try:
-            ai = tool_llm.invoke(messages)
-        except Exception as e:
-            reasoning_trace.append(f"[error] llm.invoke: {e}")
+        _sanitize_tool_calls(messages)  # keep echoed tool-call args valid JSON (oMLX 500 guard)
+        ai = None
+        for attempt in range(4):
+            try:
+                ai = tool_llm.invoke(messages)
+                break
+            except Exception as e:
+                err = str(e)[:100]
+                reasoning_trace.append(f"[retry {attempt + 1}/4] invoke failed: {err}")
+                logger.warning("{}: invoke failed (attempt {}/4): {}", persona.id, attempt + 1, err)
+                time.sleep(4)
+                if attempt >= 1:  # watchdog: restart oMLX if it's wedged, then keep going
+                    from llm import server as llm_server
+
+                    if not llm_server.server_up():
+                        logger.warning("{}: oMLX down — watchdog restarting it", persona.id)
+                        llm_server.ensure_ready(warm=False)
+        if ai is None:
+            reasoning_trace.append("[error] llm.invoke failed after 4 retries")
             ended_reason = "error"
-            logger.warning("{}: llm.invoke failed: {}", persona.id, e)
+            logger.warning("{}: giving up research after 4 failed invokes (will still finalize)", persona.id)
             break
         messages.append(ai)
 
+        # malformed-JSON tool calls: make them serializable + tell the model to retry
+        repaired = _repair_invalid_tool_calls(ai)
+        repaired_ids = {tc_id for tc_id, _ in repaired}
+        for tc_id, name in repaired:
+            messages.append(
+                ToolMessage(
+                    content=f"ERROR: your {name} call had malformed JSON arguments. "
+                    "Call it again — arguments must be a valid JSON object.",
+                    tool_call_id=tc_id,
+                )
+            )
+            reasoning_trace.append(f"[repair] malformed {name} args -> asked model to retry")
+
         tool_calls = getattr(ai, "tool_calls", None)
         if not tool_calls:
             # Force depth: if the model tries to finalize before the TIME floor (or step floor),
@@ -259,6 +339,8 @@ def run_agent(
             break
 
         for tc in tool_calls:
+            if tc["id"] in repaired_ids:  # already answered with a retry-error ToolMessage
+                continue
             # gpt-oss can leak harmony channel tokens into the tool name (e.g. "get_bars<|channel|>")
             name = tc["name"].split("<|")[0].strip()
             args = tc.get("args", {}) or {}
@@ -270,7 +352,7 @@ def run_agent(
             messages.append(ToolMessage(content=str(result), tool_call_id=tc["id"]))
             reasoning_trace.append(f"{name}({_short(args)}) -> {str(result)[:160]}")
 
-        if compaction.should_compact(_prompt_tokens(ai, messages), ctx_len):
+        if compaction.should_compact(_prompt_tokens(ai, messages), compact_budget):
             messages, running_summary = compaction.compact(plain_llm, messages, running_summary)
 
         if checkpointer:  # incremental save after each round
@@ -279,14 +361,17 @@ def run_agent(
     if steps_used >= step_budget and ended_reason == "natural":
         ended_reason = "budget"
 
-    # --- finalize (always attempted, even on timeout, to salvage picks) ---
+    # --- finalize (ALWAYS attempted — even after invoke failures, salvage the findings) ---
     picks: list[Pick] = []
     errors: list[str] = []
-    if ended_reason != "error":
+    try:
         picks, errors = _finalize(persona, running_summary, reasoning_trace)
+    except Exception as e:
+        errors = [f"finalize crashed: {e}"]
 
     if ended_reason == "error":
-        status = AgentStatus.ERROR
+        # salvaged picks still count — mark partial so execution can use them
+        status = AgentStatus.PARTIAL if picks else AgentStatus.ERROR
     elif ended_reason == "timeout":
         status = AgentStatus.TIMEOUT
     elif picks: