PR #62265: fix(exec): catch onUpdate errors when agent run ends while exec still produces output

• Repository: openclaw/openclaw
• Author: w-sss
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/62265

Description (problem / solution / changelog)

Fixes #62190 Fixes #62256

Problem

When an exec process outlives its parent agent run (e.g. due to tool-call abort, turn timeout, or late stdout delivery after the lane task completes), the onUpdate callback throws "Agent listener invoked outside active run" from pi-agent-core. This becomes an unhandled promise rejection that crashes the gateway.

Multiple users report this on both macOS and Windows with 2026.4.5. The stack trace is always:

Agent.processEvents (pi-agent-core/src/agent.ts:533)
  ← Object.onUpdate (pi-agent-core/src/agent-loop.ts:539)
  ← emitUpdate (bash-tools.exec-runtime)
  ← handleStdout
  ← Object.onSupervisorStdout
  ← Socket.<anonymous>

Root Cause

emitUpdate() guards on session.backgrounded || session.exited, which covers the common cases. However, there is a race where:

1. The agent run completes (lane task done)
2. The exec process is still running or its stdout pipe still has buffered data
3. Late stdout arrives → handleStdout → emitUpdate → opts.onUpdate()
4. pi-agent-core throws because the agent run is no longer active

The session is not yet marked as exited (process still running) and not backgrounded (foreground mode), so the guard passes.

Fix

1. try-catch on onUpdate: Wrap opts.onUpdate() in a try-catch so late-arriving stdout from a finished agent run does not become an unhandled rejection
2. updateSuppressed flag: On first error, set a flag so subsequent stdout/stderr chunks skip onUpdate entirely instead of repeatedly throwing
3. Debug logging: Log a [exec] onUpdate suppressed for session <id> debug message on the first suppression, including the original error message — visible in openclaw logs for troubleshooting without being noisy at info/warn level

The exec process itself continues to run and its output is still captured in the session buffer. Only the streaming updates to the (now-dead) agent run are suppressed.

Limitations and future work

• The catch is broad (catches all errors, not just the expected "Agent listener invoked outside active run"). This is intentional as a defensive measure — any onUpdate failure should not crash the gateway. If more granular handling is needed, error type discrimination can be added later.
• The root cause is a lifecycle mismatch between agent runs and exec processes. A cleaner long-term fix would be for pi-agent-core to expose a cancellation signal or detach hook so exec can proactively stop calling onUpdate when the run ends, rather than discovering it via a thrown error.

Test Plan

• Added test: "does not crash when onUpdate throws (agent run ended while exec still producing output)"
• Verifies exec completes normally, onUpdate was called at least twice, and the thrown error was silently caught
• All existing bash-tools tests pass
• Boundary invariant test updated for upstream setup-registry.runtime.ts allowlist drift

Changed files

• src/agents/bash-tools.exec-runtime.ts (modified, +24/-13)
• src/agents/bash-tools.test.ts (modified, +33/-0)