Summary

Running OpenClaw 2026.5.26 (npm, Windows 10). During an agent task involving a long uninterrupted sequence of Write tool calls (~21 files written across 3 new workspace directories under ~/.openclaw/workspace-<name>/), the diagnostic watcher classified the session as active_work_without_progress and force-aborted the embedded run at the default 360s abort threshold (max(5min, stuckSessionWarnMs × 3)), even though tool calls were firing continuously and committing to disk the entire time.

Logs

00:57:00 [agent/cli-backend] cli exec: provider=claude-cli model=opus promptChars=111 trigger=user useResume=true session=present resumeSession=a683dcc44eeb reuse=reusable historyPrompt=present
00:57:00 [agent/cli-backend] claude live session reuse: provider=claude-cli model=claude-opus-4-7
00:59:07 [diagnostic] stalled session: sessionId=unknown sessionKey=agent:main:main state=processing age=128s queueDepth=1 reason=active_work_without_progress classification=stalled_agent_run activeWorkKind=embedded_run lastProgress=embedded_run:started lastProgressAge=128s recovery=none
00:59:37 [diagnostic] stalled session: ... age=158s ... lastProgress=embedded_run:started lastProgressAge=158s recovery=none
00:00:07 [diagnostic] stalled session: ... age=188s ... recovery=none
... (repeats every 30s, ages 218/248/278/308/338s, all lastProgress=embedded_run:started) ...
01:03:07 [diagnostic] stalled session: ... age=368s ... lastProgress=embedded_run:started lastProgressAge=368s recovery=checking
01:03:07 [agent/cli-backend] claude live session close: provider=claude-cli model=claude-opus-4-7 reason=abort
01:03:07 [agent/cli-backend] claude live session turn failed: provider=claude-cli model=claude-opus-4-7 durationMs=367432 error=AbortError
01:03:07 [diagnostic] stuck session recovery: sessionId=37f0eb00-d091-4693-84df-40f242152bd7 sessionKey=agent:main:main age=368s action=abort_embedded_run aborted=true drained=true released=0
01:03:07 [diagnostic] stuck session recovery outcome: status=aborted action=abort_embedded_run sessionId=37f0eb00-d091-4693-84df-40f242152bd7 sessionKey=agent:main:main activeSessionId=37f0eb00-d091-4693-84df-40f242152bd7 activeWorkKind=embedded_run lane=session:agent:main:main aborted=true drained=true forceCleared=false released=0

Note lastProgress stays pinned at embedded_run:started for the entire 368-second window — the diagnostic watcher never sees any progress event, even though Write tool calls were executing throughout.

Expected

Active tool-call execution should count as forward progress. The embedded run is not stalled — it is doing exactly what it was asked to do — but the watcher's progress signal only updates on coarser events (e.g. embedded_run:started/turn boundaries), missing all the work in between.

Actual

lastProgress remains at embedded_run:started for the full duration. At 360s the recovery path fires abortAndDrainEmbeddedPiRun and the live Claude CLI session is terminated mid-flow with error=AbortError. The user-facing turn appears to end abruptly; the next user message is required to resume.

Reproduction shape

Any agent turn that performs a long uninterrupted sequence of tool calls (file writes, multi-step refactors, large restoration tasks) without intervening model turns will hit this. In this case: writing IDENTITY/USER/SOUL/TOOLS files across three new sibling-agent workspace directories.

Workaround

Raise diagnostics.stuckSessionAbortMs in openclaw.json. Setting it to 1800000 (30 min) eliminates the kill in normal long-tool-call runs. But the underlying heuristic gap remains: tool-call activity does not register as progress to the stall detector, so the workaround just moves the cliff.

Related

• [Bug]: Agent stall detector hard-coded 120s threshold kills legitimate long model calls on local vLLM #85826 — same heuristic, different surface (long model calls on local vLLM)
• Diagnostic recovery aborts active embedded runs after terminal-looking app-server progress #85532 — recovery aborting active embedded runs after terminal-looking app-server progress
• Codex app-server emits notification:turn/started then goes silent; embedded run wedges for the full stuck-session recovery window #85251 — Codex app-server emits notification:turn/started then goes silent

Suggested directions (for maintainer judgment)

1. Emit a progress event when a tool call starts and/or completes, so lastProgress advances during tool-heavy turns.
2. Distinguish embedded-run idle (no tool calls in flight) from embedded-run busy (tool call active) in the classifier, and only treat the idle case as active_work_without_progress.
3. Document diagnostics.stuckSessionAbortMs more prominently in onboarding/upgrade notes, since the default 360s is short enough to bite users on routine long sessions on first encounter.

Environment

• OpenClaw: 2026.5.26 (npm)
• OS: Windows 10 (10.0.19045, x64)
• Node: 24.14.1
• Provider: claude-cli (Claude Opus 4.7)