fix(kimi): spawn's own prepared instance no longer counts against the concurrency cap (#714)

roboco/runtime/orchestrator.py

@@ -3120,7 +3120,9 @@ class AgentOrchestrator:
         # existing-running check above stays first, so a live agent is never
         # replaced by this bail. Fail-open: a tracker read error never blocks.
         route = await self._resolve_agent_route(agent_id, task_id)
-        skip_reason = await self._spawn_gate_skip_reason(route.provider_type.value)
+        skip_reason = await self._spawn_gate_skip_reason(
+            route.provider_type.value, exclude_agent_id=agent_id
+        )
         if skip_reason:
             self._mark_task_handled(task_id)
             logger.info(
@@ -3145,7 +3147,9 @@ class AgentOrchestrator:
         # parked case is already handled above; this guards the window between
         # the pre-check and the launch. Fail-open: a tracker read error never
         # blocks spawning.
-        skip_reason = await self._spawn_gate_skip_reason(config.provider_type)
+        skip_reason = await self._spawn_gate_skip_reason(
+            config.provider_type, exclude_agent_id=agent_id
+        )
         if skip_reason:
             return self._bail_prepared_instance(
                 instance, agent_id, task_id, config.provider_type, skip_reason
@@ -10186,39 +10190,52 @@ Start by:
             return settings.kimi_max_concurrent
         return None
 
-    def _live_provider_instance_count(self, provider_type: str) -> int:
+    def _live_provider_instance_count(
+        self, provider_type: str, exclude_agent_id: str | None = None
+    ) -> int:
         """Count ``_instances`` entries for *provider_type* still holding a slot.
 
         Mirrors ``_existing_running_instance``'s liveness definition: any
         state other than OFFLINE/WAITING_LONG occupies (or is about to
-        occupy) a container.
+        occupy) a container. *exclude_agent_id* drops the agent being
+        spawned from the count — ``_prepare_agent_spawn`` registers its
+        STARTING instance before the post-prepare gate re-check, and with
+        cap=1 a self-counting check cancels every spawn forever.
         """
         return sum(
             1
-            for inst in self._instances.values()
+            for agent_id, inst in self._instances.items()
-            if inst.config is not None
+            if agent_id != exclude_agent_id
+            and inst.config is not None
             and inst.config.provider_type == provider_type
             and inst.state not in (AgentState.OFFLINE, AgentState.WAITING_LONG)
         )
 
-    def _provider_spawn_at_capacity(self, provider_type: str | None) -> bool:
+    def _provider_spawn_at_capacity(
+        self, provider_type: str | None, exclude_agent_id: str | None = None
+    ) -> bool:
         """True when *provider_type* is at its concurrency cap, if any."""
         cap = self._provider_concurrency_cap(provider_type)
         if cap is None or provider_type is None:
             return False
-        return self._live_provider_instance_count(provider_type) >= cap
+        live = self._live_provider_instance_count(provider_type, exclude_agent_id)
+        return live >= cap
 
-    async def _spawn_gate_skip_reason(self, provider_type: str | None) -> str | None:
+    async def _spawn_gate_skip_reason(
+        self, provider_type: str | None, exclude_agent_id: str | None = None
+    ) -> str | None:
         """Why ``spawn_agent`` should bail before launch, or None to proceed.
 
         Checked in order: provider-parked (rate-limited/overloaded), then the
         provider's concurrency cap (currently kimi-only, see
         ``_provider_concurrency_cap``). The returned string is both the skip
         reason and the log event name — same shape both callers already used.
+        *exclude_agent_id* is the agent this spawn is FOR: its own registered
+        instance must never count against its own capacity check.
         """
         if await self._provider_spawn_parked(provider_type):
             return "Spawn skipped: provider rate-limited (parked)"
-        if self._provider_spawn_at_capacity(provider_type):
+        if self._provider_spawn_at_capacity(provider_type, exclude_agent_id):
             return "Spawn skipped: kimi concurrency cap reached"
         return None
 

tests/unit/runtime/test_kimi_concurrency_cap.py

@@ -68,12 +68,21 @@ def _wire(monitor: dict[str, Any], provider_value: str) -> Any:
             auth_token=None,
         )
 
-    async def _prepare(*_a: Any, **_k: Any) -> Any:
+    async def _prepare(agent_id: str, *_a: Any, **_k: Any) -> Any:
+        # Registers the STARTING instance under the requested agent_id like the
+        # real _prepare_agent_spawn does — the post-prepare gate re-check runs
+        # against _instances WITH this entry present (the self-count regression
+        # only reproduces with it registered).
         monitor["prepare_calls"] += 1
         cfg = SimpleNamespace(provider_type=provider_value, model="model")
         inst = AgentInstance(
-            agent_id="spawning-agent", state=AgentState.STARTING, config=None
+            agent_id=agent_id,
+            state=AgentState.STARTING,
+            config=AgentConfig(
+                agent_id=agent_id, blueprint_path=Path(), provider_type=provider_value
+            ),
         )
+        monitor["orch"]._instances[agent_id] = inst
         return cfg, inst, None
 
     return _readiness_gate, _git_context, _route, _prepare
@@ -85,6 +94,7 @@ def _wire_orch(
     monitor: dict[str, Any],
     provider_value: str,
 ) -> None:
+    monitor["orch"] = orch
     _rg, _gc, _route, _prepare = _wire(monitor, provider_value)
     monkeypatch.setattr(orch, "_readiness_gate", _rg)
     monkeypatch.setattr(orch, "_resolve_spawn_git_context", _gc)
@@ -147,6 +157,32 @@ def test_live_provider_instance_count_and_capacity() -> None:
 # ---------------------------------------------------------------------------
 
 
+@pytest.mark.asyncio
+async def test_sole_kimi_spawn_passes_its_own_post_prepare_cap_check(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """The 2026-07-29 fleet wedge: with cap=1 and NO other kimi instance, the
+    post-prepare gate re-check counted the spawn's own just-registered
+    STARTING instance (1 >= 1) and cancelled every kimi spawn forever. The
+    agent being spawned must be excluded from its own capacity check."""
+    orch = _make_orchestrator()
+    monitor = {"route_calls": 0, "prepare_calls": 0}
+    _wire_orch(orch, monkeypatch, monitor, "kimi")
+
+    launched: list[bool] = []
+
+    async def _launch(*_a: Any, **_k: Any) -> AgentInstance:
+        launched.append(True)
+        return AgentInstance(agent_id="main-pm", state=AgentState.ACTIVE, config=None)
+
+    monkeypatch.setattr(orch, "_launch_spawn", _launch)
+
+    await orch.spawn_agent(agent_id="main-pm", task_id="task-1")
+
+    assert monitor["prepare_calls"] == 1
+    assert launched == [True]
+
+
 @pytest.mark.asyncio
 async def test_second_concurrent_kimi_spawn_skipped_while_first_live(
     monkeypatch: pytest.MonkeyPatch,