Summary

OpenClaw diagnostics can retain stale native tool-call activity for a session after stuck-session recovery, session reset, or session replacement. Future work on the same sessionKey can then be classified as blocked_tool_call even when the original tool process is no longer running and the affected session transcript has already been reset/archived.

This is narrower than the existing stuck-session umbrella issues: the suspected defect is not just that a session stalls, but that the diagnostic activity tracker can keep an orphaned activeWorkKind=tool_call / activeTool=bash entry long enough to poison later recovery decisions for the same session key.

Environment

• OpenClaw: 2026.5.20 (e510042)
• OS: Linux 7.0.10-arch1-1 x86_64
• Node: v24.11.1
• Gateway: systemd user service, loopback gateway
• Runtime observed: embedded Codex app-server / native tool execution

Observed behavior

A local gateway repeatedly emitted stalled-session diagnostics with this shape, redacted:

stalled session: sessionId=<redacted> sessionKey=<redacted>
  state=processing
  reason=blocked_tool_call
  classification=blocked_tool_call
  activeWorkKind=tool_call
  activeTool=bash
  activeToolAgeMs=<very large>
  lastProgress=<old tool start>
  recovery=checking

The active tool ages were on the order of many hours. The two stale tool ids found in local state resolved only to reset/trajectory archives, not live active sessions. One stale tool was a long foreground bash command that had started a dev server; another was an rg command. At investigation time, there was no matching live process for the dev-server case, and the current gateway stability view showed no active work after recovery.

Recovery later emitted an abort/drain outcome for an embedded run, for example:

stuck session recovery outcome: status=aborted action=abort_embedded_run
  activeWorkKind=embedded_run
  aborted=true drained=true forceCleared=false released=0

The concerning part is that stale native tool activity and embedded-run/reply-run recovery appear to be tracked by separate state paths. If the native tool never emits the matching completion event, or if recovery/reset replaces the session before the completion event is reconciled, the diagnostic tracker can continue to report tool_call as active and drive blocked_tool_call classification for future turns.

Expected behavior

Stuck-session recovery, session reset, and session replacement should clear or reconcile diagnostic active-work state for the affected sessionId/sessionKey, including native tool calls.

After a recovery aborts/drains an embedded run or a session is reset/replaced:

1. stale activeTools entries for the old session/run should not continue to classify later turns as blocked_tool_call;
2. blocked_tool_call should mean there is still an owned active native tool, not just an orphaned diagnostic record;
3. if the original tool cannot be cancelled or observed, the diagnostic state should be explicitly evicted/quarantined with a structured recovery event.

Actual behavior

A stale activeTool=bash record can remain associated with a session key for many hours, repeatedly producing session.stalled with classification=blocked_tool_call. Recovery can abort/drain the embedded run but does not clearly guarantee that stale native tool activity for the session key is cleared.

Source-level suspect

Relevant current source paths:

• src/logging/diagnostic-run-activity.ts
  • recordToolStarted adds native tools to activeTools.
  • recordToolEnded removes them.
  • recordRunCompleted clears active tools/model calls/embedded runs.
  • markDiagnosticEmbeddedRunEnded can clear run activity, but callers can opt out.
• src/auto-reply/reply/reply-run-registry.ts
  • markReplyRunDiagnosticWorkEnded calls markDiagnosticEmbeddedRunEnded(..., clearRunActivity: false).
• src/logging/diagnostic-session-attention.ts
  • stale tool_call activity is classified as blocked_tool_call.
• src/logging/diagnostic.ts
  • isBlockedToolCallRecoveryEligible allows recovery once the blocked tool call crosses the abort threshold.
• src/logging/diagnostic-stuck-session-recovery.runtime.ts
  • recovery can abort active embedded work, but the cleanup contract for orphaned native tool activity is not obvious from the observed outcome.

The missing primitive may be something like a targeted diagnostic cleanup/reconciliation path, for example clearDiagnosticSessionActivity({ sessionId, sessionKey, reason }), called when recovery aborts/drains a run, when a session is reset/replaced, and when a diagnostic tool-call owner is no longer present.

Suggested fix shape

1. Add a diagnostic activity cleanup primitive that removes or quarantines all active tool/model/embedded-run state for a given sessionId and/or sessionKey.
2. Call it from stuck-session recovery when reason=blocked_tool_call recovery aborts/drains a run or determines the owner is gone.
3. Call it from session reset/replacement paths after the old session is archived or superseded.
4. Add regression coverage for:
   • native tool_call starts and never emits completion;
   • session is reset/replaced or embedded recovery aborts/drains;
   • later work on the same sessionKey is not classified as blocked_tool_call from the old tool;
   • genuine active native tools still classify as blocked until completion/abort.
5. Emit a structured diagnostic event when stale tool activity is evicted, so operators can distinguish a real running tool from recovered stale state.

Related issues

This is a narrow follow-up to prior stuck-session/recovery work rather than a duplicate:

• [Bug]: Stuck processing sessions are detected but never aborted — gateway requires external restart to recover #71127 - canonical stuck-session recovery issue, now closed after base recovery landed.
• Stuck session ghost blocks event loop, causes CLI timeout and sustained high CPU (11.5hr+) #77115 - ghost session / stale activity family; comments discuss stale model_call and tool_call recovery gaps.
• [Bug]: Stuck Session Recovery 机制双重失效 + Session 预处理每次耗时过长 #76038 - stuck-session recovery skip/no-op cases and active-work recovery policy.
• A single stalled agent session blocks the entire Gateway event loop (isolation failure) #84903 - stalled session isolation failure and stale diagnostic activity affecting gateway health.
• High CPU, extreme control-plane RPC latency, and unstable polling after upgrade from 2026.4.24 to 2026.4.29/2026.5.2 #76562 - broad gateway pressure issue; includes reports of blocked_tool_call / hung tool calls contributing to gateway instability.
• Codex app-server emits notification:turn/started then goes silent; embedded run wedges for the full stuck-session recovery window #85251 and Diagnostic recovery aborts active embedded runs after terminal-looking app-server progress #85532 - related embedded-run abort/recovery behavior, but focused more on app-server/embedded progress semantics than orphaned native tool-call state.

Impact

A single orphaned native tool diagnostic record can keep a session lane looking blocked long after the original command is gone. The user-visible effect is delayed or lost replies, repeated stalled-session logs, and confusing recovery outcomes that make the gateway appear to still be blocked by bash when there is no corresponding live tool process.