* Handle PostToolUse for Bash: check last-command-result.json for hangs. * Only fires a warning -- does not trip the breaker or block execution.

* Handle PostToolUseFailure: track consecutive failures, trip on threshold.

* Check if a session-scoped allow is active for a category. * Session allows expire after 6 hours (covers long sessions but not overnight). * @param

* Increment a counter in the hook timing log. * @param {string} hook - Hook name (e.g., 'circuit-breaker', 'quality-gate') * @param {string} metric

* Increment trust counters in harness.json for contextual appropriateness. * Tracks sessions completed and campaigns completed this session. * Non-c

* True when an absolute, normalized path points into the Claude Code native * auto-memory directory: <home>/.claude/projects/<slug>/memory/ where <sl

* Returns true if the given branch is owned by a persistent campaign * (campaign frontmatter has worktree_status: active). * Scans .planning/campaig

* Load permission events from both sources, normalized to * { ts, tool, target, decision }. * * audit.jsonl entries (event === 'permission-request'

* Append a compact permission event to the dedicated audit JSONL. * Observer-only: failures are swallowed, never affects the hook outcome. * One app

* Log session cost data to .planning/telemetry/session-costs.jsonl. * * Two-layer approach: * 1. Real tokens: Read Claude Code's session JSONL fo