openclaw - 💡(How to fix) Fix Gateway crashes with 'Agent listener invoked outside active run' during multi-story orchestration loops [1 participants]

Bug Description

The gateway consistently crashes with an unhandled promise rejection (Agent listener invoked outside active run) when an orchestration agent runs a multi-story execution loop involving subagent spawning. The crash occurs at the boundary between completing one story's PR/CI/merge cycle and spawning the next story's subagent.

Environment

• OpenClaw: 2026.4.5 (3e72c03)
• Node.js: v22.22.1
• OS: Windows 10 Pro 10.0.26200 (MINGW64)
• Agent model: anthropic/claude-opus-4-6
• Gateway mode: local, loopback binding

Reproduction

1. Configure an orchestration agent that runs a multi-step workflow (story creation → validation → implementation → code review → PR/CI/merge)
2. Each step is delegated to a subagent (one step per subagent, fresh context each time)
3. Step 8 creates a PR, polls CI every 60s, and merges when CI passes
4. After merge, the orchestration agent attempts to spawn the next story's subagent
5. Gateway crashes consistently at this transition point

The crash has been reproduced 3 times across different runs:

• Run 1: Crashed after R-1 merge, attempting R-2 spawn
• Run 2: Crashed after R-10 merge, attempting R-3 spawn
• Run 3: Crashed after R-3/R-4/R-5 batch merge, attempting R-6 spawn

Stack Trace

[openclaw] Unhandled promise rejection: Error: Agent listener invoked outside active run
    at Agent.processEvents (file:///C:/Users/lanna/AppData/Local/nvm/v22.22.1/node_modules/openclaw/node_modules/@mariozechner/pi-agent-core/src/agent.ts:533:10)
    at file:///C:/Users/lanna/AppData/Local/nvm/v22.22.1/node_modules/openclaw/node_modules/@mariozechner/pi-agent-core/src/agent.ts:380:21
    at Object.onUpdate (file:///C:/Users/lanna/AppData/Local/nvm/v22.22.1/node_modules/openclaw/node_modules/@mariozechner/pi-agent-core/src/agent-loop.ts:539:7)
    at emitUpdate (file:///C:/Users/lanna/AppData/Local/nvm/v22.22.1/node_modules/openclaw/dist/exec-defaults-uj0McX2k.js:1524:8)
    at handleStdout (file:///C:/Users/lanna/AppData/Local/nvm/v22.22.1/node_modules/openclaw/dist/exec-defaults-uj0McX2k.js:1546:4)
    at Object.onSupervisorStdout [as onStdout] (file:///C:/Users/lanna/AppData/Local/nvm/v22.22.1/node_modules/openclaw/dist/exec-defaults-uj0McX2k.js:1610:3)
    at file:///C:/Users/lanna/AppData/Local/nvm/v22.22.1/node_modules/openclaw/dist/exec-defaults-uj0McX2k.js:1007:21
    at Socket.<anonymous> (file:///C:/Users/lanna/AppData/Local/nvm/v22.22.1/node_modules/openclaw/dist/exec-defaults-uj0McX2k.js:568:4)
    at Socket.emit (node:events:519:28)
    at addChunk (node:internal/streams/readable:561:12)

Preceding Gateway Warnings

Before each crash, the subagent announce completion times out repeatedly:

Subagent announce completion direct announce agent call transient failure, retrying 2/4 in 5s: gateway timeout after 120000ms
Subagent announce completion direct announce agent call transient failure, retrying 3/4 in 10s: gateway timeout after 120000ms
Subagent announce completion direct announce agent call transient failure, retrying 4/4 in 20s: gateway timeout after 120000ms

Also observed before the crash:

[diagnostic] lane wait exceeded: lane=session:agent:agent-3:main waitedMs=485113 queueAhead=0

Attempted Mitigations (none resolved the issue)

1. Added 15s then 30s cooldown between PR merge and next subagent spawn — crash still occurs
2. Delegated PR/CI/merge to a dedicated subagent to keep the orchestration agent's connection clean — crash still occurs
3. All runs on gateway --force with fresh startup

Analysis

The crash appears to be a race condition in pi-agent-core where:

1. The subagent completion announce saturates or blocks the gateway WebSocket during CI polling (long-running 60s poll loops over several minutes)
2. When the orchestration agent's turn resumes and it tries to spawn a new subagent, the Agent.processEvents listener fires but the agent run context has already been invalidated
3. The unhandled rejection kills the gateway process

The crash is 100% reproducible in any long-running orchestration loop that involves subagent spawning after a CI polling step.

Expected Behavior

The gateway should gracefully handle subagent announce timeouts without corrupting the listener state, and allow the orchestration agent to continue spawning new subagents after PR/merge operations.