Summary

After upgrading a local OpenClaw source install to 2026.5.7, ACP/Codex sessions can report completed successfully even though the task did not actually run. The only assistant output is the initial progress sentence, and no requested investigation/results are produced.

Environment

• OpenClaw: 2026.5.7
• OS: macOS 26.3.1 arm64
• Runtime: local gateway on loopback
• ACP agent: codex
• Model/provider in parent agent: openai-codex/gpt-5.5

What happened

Two separate ACP/Codex runs were started with sessions_spawn(runtime="acp", agentId="codex") for code investigation tasks.

Both returned an internal completion event with successful status, but the actual child session contained only a short initial progress message such as:

Vou localizar o fluxo de inbound do WhatsApp até a detecção/autorização de comandos...

or:

Vou mapear o fluxo real primeiro...

No tool work or final report was produced, despite the task explicitly asking for a final report and minimal/no progress updates.

Expected behavior

The ACP run should either:

1. execute the requested task and return the final result, or
2. fail/timeout with a clear error if the ACP/Codex adapter stopped before doing the work.

It should not mark the task as successfully completed after only the initial progress text.

Actual behavior

• status: completed successfully
• child session transcript contains only the initial progress sentence
• no substantive task execution happened
• parent session context did not grow abnormally, so this does not look like a context explosion case

Why this matters

This makes ACP task completion unreliable: callers receive a success event but the requested work was not done.

Notes

This was observed after moving the same local environment to 2026.5.7. It may be related to ACP runtime completion semantics or the Codex ACP adapter returning/closing after the first assistant message.