fix(hooks): AgentInvocation logs every background Agent spawn as subagent_stop (duration 0) — async-spawn detection misses the current tool_response, and start records are gone before AgentWatchdog can count them #2387

Version

LifeOS 7.40.4 (main @ 5e2f2e8), Claude Code 2.1.30x (observed 2026-10-23; Agent tool schema unchanged in 2.1.310).

What is broken

AgentInvocation.hook.ts runs on PreToolUse and PostToolUse for Agent. Since v1.4.0 it knows PostToolUse fires at spawn time for a background agent and detects that case by matching tool_response against /Spawned successfully|running and will receive instructions via mailbox/i: on a match it logs subagent_spawned_async, otherwise subagent_stop with a computed duration.

In the current harness the match never succeeds for a background spawn. Every dispatch produces a subagent_start and a subagent_stop with duration_seconds: 0 in the same second — 7 of 7 dispatches across two sessions on 2026-10-23 — so subagent-events.jsonl records every background agent as a zero-second run, and anything built on those events (durations, the Pulse agents view) sees agents that start and stop at once.

The transcript's tool_result for the same spawn does contain both phrases ("Spawned successfully. (This tool result is internal metadata …)\nagent_id: …\nname: …\nThe agent is now running and will receive instructions via mailbox."), so the PostToolUse tool_response payload the hook receives must differ from the transcript text. I have not captured the raw payload — the hook does not log it.

A second thing I noticed while tracing it: the start record is deleted at PostToolUse in both branches (:204-211 runs before the async check, and the v1.4.0 header says the spawn-ack removes it by design, liveness coming from subagents/agent-*.jsonl mtimes). That means AgentWatchdog.ts, which counts active agents from the starts file, sees zero during a background run even after #2362 fixes the file name — the record is gone within the spawn second. Confirmed: agent-starts.json was {} while a 13-minute agent was working.

Where

• LifeOS/install/hooks/AgentInvocation.hook.ts:217-220 (string coercion + regex), :204-211 (duration + delete starts[key]), :222-230 (the event).
• LifeOS/install/LIFEOS/TOOLS/AgentWatchdog.ts:41 (active agents = starts file; see fix(watchdog): read agent-starts.json #2362).

Repro

1. Dispatch any background Agent (subagent_type: general-purpose, a brief that runs a few minutes).
2. tail -2 MEMORY/OBSERVABILITY/subagent-events.jsonl → subagent_start and subagent_stop with "duration_seconds":0, same timestamp.
3. cat MEMORY/OBSERVABILITY/agent-starts.json while the agent is still working → {}.

Suggested fix

Stop inferring the spawn from response text: with the current Agent tool every main-conversation dispatch is a background spawn, so PostToolUse:Agent is always the spawn-ack — log subagent_spawned_async there and take the real stop from the SubagentStop hook event (correlated by agent_id/name, since that payload lacks subagent_type/description), or from the transcript going quiet. If the response text stays a signal, log the raw tool_response shape once under a debug flag so the regex can be fixed against what the harness actually sends. For the watchdog, read liveness from the same subagents/agent-*.jsonl mtimes the header names instead of the starts file, so #2362 and this fix together give it something to watch.

🤖 Generated with Claude Code