Long tool calls hit tool_timeout without actionable recovery (prod)

[wubingjie1st]

Summary

FinClaw agent runs on the sales demo stop with internal_stop_reason=tool_timeout
(~17× in ~14 days). Long-running tool work does not consistently background or return
a non-fatal, actionable next step.

Environment

• Host: ChatKit demo (finclaw.finogeeks.club) / FinClaw claw runtime
• Evidence: audit/agent-runs.jsonl
• Window: ~14 days ending 2026-07-15
• Sample timeout run metadata (no user content): tool_call_count=18,
  tool_runtime_ms≈10347, agent_tool_bundle=workspace

Actual

• stop_reason=error, internal_stop_reason=tool_timeout
• Count: 17 runs
• Samples:
  • run_id=448a80d1ed7570933146128c91cd38cf tenant a60018fc… ended 2026-07-01T04:00:00
  • run_id=527c85bb34547d64c1e9a03952a9ca34 tenant a60018fc… ended 2026-07-01T04:40:41
  • run_id=cf24586e50f6b069aa5d57cf4faf3ec3 tenant a60018fc… ended 2026-07-02T01:55:41

Expected

Long-running tool work either:

1. runs in background with a job / status path, or
2. surfaces a non-fatal timeout with a clear user-visible next step

Scoped tests in finclaw for timeout budget / recovery.

Out of scope

• Raising demo host wall-clock without a product contract
• ChatKit-only deploy changes