feat(cockpit): PostToolUse tool-mirror hook — zero-token live worker activity (3c spike, closes #71)

[robercano-ghbot]

Summary

Spike (Cockpit 3c, part 3 of 3 — design in #68, unblocked by #70) adding a harness-side PostToolUse tool-mirror hook that records worker Bash/Edit/Write tool calls to .claude/state/worker-tools.jsonl at zero agent tokens. Ships disabled by default — enabling is a one-line wiring change pending the go/no-go decision below.

Closes #71.

What's here

• .claude/scripts/log-worker-tool.sh — PostToolUse hook. Mirrors only Bash/Edit/Write into JSONL {ts, tool, summary(one-line), path} (path = cwd for Bash → worktree-path attribution; file_path for Edit/Write). JSON built via node (safe escaping, single-line collapse, 200-char Bash cap). Strict no-op on any failure (always exit 0). Rotation capped to ${WORKER_TOOLS_MAX_LINES:-2000} via atomic temp+rename, mirroring log-event.sh. CLAUDE_WORKER_TOOLS_FILE override for tests. Not wired into .claude/settings.json — a commented, copy-paste wiring snippet lives in the script header.
• .claude/scripts/log-worker-tool.test.sh — 17 checks (discovered by the test gate): per-tool capture, non-target drop, empty/malformed stdin no-op, multi-line collapse, JSON escaping, rotation cap, unwritable-dir exit-0.

Gates (self-adapter .claude/self/gates.json)

build / lint / test — all pass (test includes checks.sh + smoke-fanout.sh).

Review

Consensus (all) met: correctness (opus) APPROVE, tests (sonnet) APPROVE. Non-blocking nits noted for follow-up: Edit/Write summary field not asserted (trivial derivative of well-tested path); no explicit 200-char truncation test.

Go/no-go evaluation

• Signal vs noise: synthetic 15-event implementer burst → 8 records mirrored (2 Edit, 1 Write, 5 Bash), 7 Read/Grep/Glob dropped (~53% of raw calls become records; every record actionable). No noise source found.
• Log growth: ~105 bytes/record (844 bytes for the 8-record burst). At the 2000-line cap, steady-state ceiling ≈ 300–500 KB. Growth bounded regardless of session length.
• Zero worker interference: exit-0-always confirmed under injected failures — empty/truncated stdin, unwritable dir, node absent from PATH — no error output, no corruption (0-byte file worst case).
• 0 agent tokens: structural — hook runs in the Claude Code PostToolUse mechanism, outside the agent sandbox/context (per the repo's verified 2026-07-07 note that Pre/PostToolUse hooks run outside the sandbox). The agent never sees the hook's stdin/stdout.
• Recommendation: GO, staying disabled by default until a short live trial confirms cwd/path from real PostToolUse payloads resolve to the worktree paths cockpit expects (this spike exercised synthetic hook-event JSON only). Cockpit read/attribution path deferred to the go decision, per plan.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SxWgee2T3w45nnPw5yZ4iu