openclaw - 💡(How to fix) Fix Aborted subagent with empty content silently marked `done`, never auto-announces — parent black-holes [1 comments, 2 participants]

Summary

When a subagent's underlying inference call gets aborted mid-response (Codex AbortError, OpenAI Responses API timeout, network drop, etc.), the subagent emits an empty assistant message (content: []) with stopReason: "aborted" and errorMessage: "This operation was aborted". The gateway then marks the run as status: done with abortedLastRun: false, and the auto-announce queue silently drops the empty result. The parent agent waits indefinitely for an announcement that never comes.

Reproduction

1. Spawn a long-running subagent with mode: "run" and runtime: "subagent" that does heavy reasoning over many tool calls (in our case Codex GPT-5.5 doing 16 minutes of upstream-source archaeology)
2. Wait for the underlying inference call to hit its provider-internal timeout (Codex's per-response timeout, distinct from runTimeoutSeconds)
3. Observe the subagent's transcript: final entry is an openclaw:prompt-error custom event followed by an empty content: [] assistant message
4. The gateway flags status: done, but no completion event is delivered to the parent

Forensic timeline (real example, 2026-04-25)

20:28:01 AST  Spawn subagent (Codex GPT-5.5, mode=run)
20:44:21 AST  Last successful tool call (exec)
20:44:22.570  Codex emits openclaw:prompt-error
              error: "This operation was aborted | This operation was aborted"
20:44:22.614  Final assistant message:
              {
                content: [],            // <-- empty
                stopReason: "aborted",
                errorMessage: "This operation was aborted",
                usage: { totalTokens: 0 }
              }
20:44:24.876  Gateway: status=done, abortedLastRun=false
              Auto-announce: silently swallowed (no content to relay)
              Parent: waits forever, no completion event arrives

Why it's a bug

Two distinct gateway invariants are violated:

1. abortedLastRun: false is wrong when the assistant message has stopReason: "aborted" and errorMessage: "This operation was aborted". The assistant message itself reports the abort; the gateway should propagate that into abortedLastRun: true (or a richer status like aborted).
2. Silent announce drop: the auto-announce queue should not produce a no-op for an aborted-with-empty-content terminal state. Either deliver a failure announcement ("subagent X aborted at step Y, no result available") or surface as a parent-visible error.

Either fix would prevent the parent from black-holing.

Suggested fix sketch

In the subagent runtime path (something like subagent-announce.ts / pi-embedded-runner/run/attempt.ts), when the final assistant message has both stopReason: "aborted" and zero output tokens:

• Set runRecord.status = "aborted" (not "done")
• Set abortedLastRun = true
• Either:
  • (a) Auto-retry once with a fresh inference call, or
  • (b) Synthesize a completion event with status=failed and the error message, so the parent can act on it

Workaround

Parent must poll sessions_list with kind=subagent and check for stale "done" entries with empty output, instead of trusting that auto-announce will fire. We're adding a heartbeat-time watchdog on our side, but this defeats the purpose of the announce mechanism.

Environment

• OpenClaw 2026.4.24
• Subagent: Codex openai-codex/gpt-5.5 (free via Pro subscription)
• Mode: run, runtime: subagent
• runTimeoutSeconds: 18000 global default (so OpenClaw timeout was NOT the trigger — the abort came from inside the inference layer)