fix(runtime): harden tool-call pairing and extend run_command timeouts

Prevent invalid tool_calls replay by requiring immediate tool-result adjacency in message building, and raise run_command default and tool-level timeout ceilings to reduce install-command timeout failures.

Co-authored-by: Cursor <cursoragent@cursor.com>

runtime/chat/agent_messages.py

@@ -514,11 +514,30 @@ def build_llm_messages(
                     tool_calls = None
 
             if tool_calls and isinstance(tool_calls, list):
-                # Guard: only include tool_calls if tool results exist later in this trimmed window.
+                # Guard: only include assistant tool_calls when paired tool rows are present.
+                # OpenAI-compatible providers require strict adjacency:
+                # assistant(tool_calls) must be followed immediately by matching tool rows.
                 want_ids = [str(tc.get("id") or "").strip() for tc in tool_calls if isinstance(tc, dict) and str(tc.get("id") or "").strip()]
                 suffix = tool_ids_after[i] if (i >= 0 and i < len(tool_ids_after)) else set()
                 if want_ids and any(tid not in suffix for tid in want_ids):
                     tool_calls = None
+                # Stronger guard than suffix-presence: verify immediate following block.
+                if tool_calls is not None and want_ids:
+                    immediate_ids: set[str] = set()
+                    for j in range(i + 1, len(store_messages)):
+                        nm = store_messages[j]
+                        n_event_type = str(getattr(nm, "event_type", "") or "").strip().lower()
+                        # Ignore reasoning-only rows when checking adjacency.
+                        if n_event_type == "reasoning":
+                            continue
+                        n_role = str(getattr(nm, "role", "") or "")
+                        if n_role != "tool":
+                            break
+                        tcid = _tool_call_id_from_tool_row(getattr(nm, "tool_calls", None))
+                        if tcid:
+                            immediate_ids.add(str(tcid))
+                    if any(tid not in immediate_ids for tid in want_ids):
+                        tool_calls = None
                 if tool_calls is None:
                     out.append(
                         _attach_reasoning_content(

runtime/tools/local_sdk/adapter.py

@@ -66,7 +66,7 @@ class LocalAdapter:
         os.environ[k] = str(value)
         return {"ok": True, "key": k, "value": os.environ.get(k), "deleted": False}
 
-    def run_command(self, *, command: str, cwd: str | None = None, timeout: int = 30) -> dict[str, Any]:
+    def run_command(self, *, command: str, cwd: str | None = None, timeout: int = 300) -> dict[str, Any]:
         cmd = str(command or "").strip()
         if not cmd:
             return {"ok": False, "error_code": "command_required", "error": "command_required"}
@@ -103,7 +103,7 @@ class LocalAdapter:
         except Exception:
             return {"ok": False, "error_code": "disabled", "error": "disabled"}
         try:
-            timeout_s = max(1, min(int(timeout or 30), 600))
+            timeout_s = max(1, min(int(timeout or 300), 600))
             # run_command never follows adapter cd state.
             # It only uses explicit cwd; otherwise defaults to data/workspace.
             workdir = str(resolve_workspace_path(cwd or "data/workspace"))

runtime/tools/public/run_command_tool.py

@@ -12,7 +12,7 @@ def run_command_tool() -> ToolSpec:
         if not command:
             return {"ok": False, "error_code": "command_required", "error": "command_required"}
         cwd = str(args.get("cwd") or "").strip() or None
-        timeout = int(args.get("timeout") or 30)
+        timeout = int(args.get("timeout") or 300)
         return get_local_adapter().run_command(command=command, cwd=cwd, timeout=timeout)
 
     return ToolSpec(
@@ -23,7 +23,7 @@ def run_command_tool() -> ToolSpec:
             "properties": {
                 "command": {"type": "string", "description": "Shell command to execute."},
                 "cwd": {"type": "string", "description": "Optional working directory."},
-                "timeout": {"type": "integer", "description": "Timeout in seconds.", "default": 30},
+                "timeout": {"type": "integer", "description": "Timeout in seconds.", "default": 300},
             },
             "required": ["command"],
             "additionalProperties": False,
@@ -31,7 +31,7 @@ def run_command_tool() -> ToolSpec:
         handler=_handler,
         tags=frozenset({"public", "exec"}),
         risk_level="high",
-        timeout_s=90.0,
+        timeout_s=620.0,
         read_only=False,
     )
 

tests/test_local_public_tools.py

@@ -74,7 +74,7 @@ def test_local_adapter_backend_roundtrip(tmp_path: Path, monkeypatch) -> None:
 
 def test_run_command_tool_handler(monkeypatch) -> None:
     class _Adapter:
-        def run_command(self, *, command: str, cwd: str | None = None, timeout: int = 30):
+        def run_command(self, *, command: str, cwd: str | None = None, timeout: int = 300):
             return {"ok": True, "command": command, "cwd": cwd, "timeout": timeout}
 
     monkeypatch.setattr("oclaw.runtime.tools.public.run_command_tool.get_local_adapter", lambda: _Adapter())