openclaw - 💡(How to fix) Fix [Bug]: claude-cli harness deregisters on session abort, cascades to subagent failure and announce loop on 2026.4.24 [1 comments, 2 participants]

Summary

On OpenClaw v2026.4.24 (macOS), a subagent session using claude-cli/claude-opus-4-6 was aborted mid-run when the session went quiet during large file reads. After the abort, the claude-cli harness deregistered, and every subsequent dispatch failed instantly with the same error — including the runtime's own announce/completion mechanism. The session's auto-announce retried 3 times and gave up with no user-visible notification.

This is related to (but distinct from) #71957 "Gateway startup race." That issue covered harness registration at startup. This issue covers harness deregistration at runtime when a CLI session is aborted.

Steps to reproduce

1. Spawn a subagent with runtime=subagent, model=claude-cli/claude-opus-4-6, mode=run
2. Give it a task that involves reading large files (~80-128KB) with no streaming output for 180+ seconds
3. The gateway's no-output timeout aborts the CLI session
4. Check gateway logs — harness immediately reports not registered
5. Any subsequent dispatches (including the session's own auto-announce) fail instantly

Gateway log timeline (exact, 2026-04-26 ~08:35 PDT)

08:35:03 [agent/cli-backend] claude live session start: provider=claude-cli model=claude-opus-4-6 activeSessions=2
08:36:07 [agent/cli-backend] claude live session turn: durationMs=133734 rawLines=1689
08:44:36 [agent/cli-backend] claude live session close: provider=claude-cli model=claude-opus-4-6 reason=abort
08:44:37 [ws] ⇄ res ✗ errorCode=UNAVAILABLE errorMessage=Error: Requested agent harness "claude-cli" is not registered and PI fallback is disabled.
08:45:07 [ws] ⇄ res ✗ errorCode=UNAVAILABLE (announce retry 1)
08:45:08 [ws] ⇄ res ✗ errorCode=UNAVAILABLE (announce retry 2, cached=true)
08:46:07 [agent/cli-backend] claude live session close: reason=idle  (a *separate* session, re-registers harness)
08:46:14 [ws] ⇄ res ✗ errorCode=UNAVAILABLE (announce retry 3 give-up, cached=true)
08:46:14 [warn] Subagent announce give up (retry-limit) run=d86a5df6 retries=3 endedAgo=97s

Key observation: the harness re-registered automatically when a separate CLI session closed cleanly at 08:46:07 — but the announce retries had already given up at 08:46:14.

Impact

1. Subagent work is silently lost — the session failed partway through, no output written, no notify to user
2. Auto-announce retry loop also fails (harness still down during the 3-retry window)
3. User gets no feedback — the spawning session already ended (Discord/Telegram flow), so completion events cannot reach them

Related

• #71957 covered harness registration at startup (fixed on main, ships next release)
• This is the "harness deregistration on CLI abort" issue that #71957's body explicitly calls out as a separate issue: "The harness deregistration on CLI abort is a separate issue that amplifies this one"

Suggested fixes

• Don't deregister the harness on session abort — the harness itself (the plugin + backend) is still alive; only the CLI session died
• If deregistration is intentional: retry announce after harness re-registration, not just on a fixed timer
• Increase default no-output timeout for subagent sessions (vs. interactive sessions) — large file reads are legitimate long-quiet operations

Environment