Runs end with internal_stop_reason=malformed_tool_call (prod cluster) #5

Summary

FinClaw-hosted agent runs on the sales demo terminate with
internal_stop_reason=malformed_tool_call (~17× in ~14 days). Users see a failed
turn when the model emits an unparseable / invalid tool call.

Hypothesis (unverified): repair / recovery path is insufficient — fix agent should
prove or disprove against current main.

Environment

• Host: ChatKit demo (finclaw.finogeeks.club) running FinClaw claw runtime
• Evidence source: Admin tenant execution logs → audit/agent-runs.jsonl (not fleet.log)
• Window: ~14 days ending 2026-07-15

Actual

• stop_reason=error, internal_stop_reason=malformed_tool_call
• Count: 17 runs in window
• Samples (redacted tenant prefix):
  • run_id=239eba9e579cb4fb1c416d00f7f82a40 tenant a60018fc… ended 2026-07-01T04:44:44
  • run_id=a782690eb8d5263a2e130dd2d097d783 tenant a60018fc… ended 2026-07-06T09:54:26
  • run_id=33a4deddd858284a8716659ff53d00d5 tenant d3821216… ended 2026-07-14T07:29:31

Expected

Malformed tool calls are repaired when possible, or the run fails with a clear,
recoverable user-visible path. Regression coverage prevents silent reintro in finclaw
agent-loop / claw.

Out of scope

• ChatKit Admin / compose-only changes
• Ops config (API keys, MCP auth, Feishu)