Infra health

/infra-health — Harness Telemetry

Reports how the infrastructure is performing, so it can be improved. Reads the event sinks under ~/vault/logs/ — no live cost data (that's /stats). All telemetry is local — nothing leaves this machine.

Run it

python3 ~/.claude/skills/infra-health/scripts/infra_health.py <days>   # default 7

• /infra-health → 7 days · /infra-health today → 1 · /infra-health month → 30

What it shows

• Hooks — per-hook run count, ok/skip/error split, p50/p95 duration. ⚠ flags any hook with errors/blocks. Source: hooks.jsonl.
• Skills — quality + cost — per skill: invocations, auto vs explicit (/slash) split, median output tokens (cost), correction rate (corr%, ⚠ at ≥25% = rework), and in-window errors. Source: skills.jsonl.
• Routing adherence — ok / (ok + missed): are skills firing when a prompt matches their triggers? Lists recent missed (matched triggers, nothing fired) + misfire candidates. Source: routing.jsonl.
• Dead skills — registered (~/.claude/skills/*/SKILL.md) but never invoked. Source: skills.jsonl ∪ workflow.jsonl (all-time), minus the registry.
• Subagents — Task/agent invocation frequency. Source: skills.jsonl (kind=agent).
• Daily jobs — last run, freshness (hours ago), failures. Source: daily-jobs.jsonl.
• Knowledge drift — the semantic-memory eval KPI: latest counters + trend across recent runs (ADR hygiene, KBs past verify_by + max stale days, metric-drift suspects, memory count/index KB). ⚠ flags any counter climbing off its baseline, suspects_recurring > 0 (a flagged item still unactioned a run later — the dead-detector alarm), or a stale date (the weekly --score isn't running). Working = counters flat-low + nothing recurring. Source: knowledge-drift-score.jsonl.
• Sessions — marker count, live (<45m heartbeat), and unpushed count. Source: active-sessions/.
• Recent errors — warm-start FATAL/ERROR tail.

How telemetry is captured (the data model)

Two capture paths: hooks log themselves on the hot path; skill/routing quality is computed offline by skill_analyzer.py in the session-export pipeline (a Skill-tool PostToolUse hook can't measure skills — the tool returns once SKILL.md loads, before the skill does any work). Full schema: ~/code/claude-harness/vault-scripts/TELEMETRY.md.

Sink	Written by	Holds
~/vault/logs/hooks.jsonl	every hook (via hooklib.sh)	hook runs + outcome + duration + exit code
~/vault/logs/skills.jsonl	skill_analyzer.py (session-export, offline)	per-invocation cost + quality: source, output_tokens, tool_calls, errors, correction_next; kind=agent for subagents
~/vault/logs/routing.jsonl	skill_analyzer.py (session-export, offline)	per trigger-matched user turn: verdict (ok/missed/misfire), matched, fired, prompt
~/vault/logs/workflow.jsonl	tool-telemetry-hook.sh (PostToolUse Skill|Task) — the merged hook	real-time skill + subagent invocations: kind (skill/agent), name, project, args, outcome. The single invocation log (absorbed skill-log-hook + the old events.jsonl). Dead-skill history + invocation audit; also read by workflow-gate + vault-audit.
~/vault/logs/daily-jobs.jsonl	run-daily-jobs.sh	daily-job success/failure + timing
~/vault/logs/knowledge-drift-score.jsonl	discrepancy-scan.py --score (weekly, via /wrap-up)	semantic-memory drift KPI: ADR-hygiene + KB-freshness + metric-drift + memory counters per run (eval trend); baseline = 2026-07-02 audit
~/vault/logs/active-sessions/*.json	session-marker-hook.sh	per-session liveness + pushed flag

Using it to improve

• A hook with a high error/block rate or a fat p95 → fix or raise its timeout. (block/deny with exit=2 on a guard hook is working as intended, not a failure — exit_code flags real crashes.)
• A skill with a high corr% (≥25%) → its output keeps getting reworked next turn; tighten it.
• A skill with high median output tokens → cost target to optimize.
• Low routing adherence / recurring missed triggers → a CLAUDE.md routing-rule gap or weak SKILL.md description triggers (matching is heuristic — treat missed/misfire as review candidates).
• A skill with 0 invocations over a month → candidate to retire or fix discoverability.
• A daily job stale beyond its freshness → cron//start-work not firing.
• A persistent unpushed count → sessions ending without /vault-push.
• Suggest concrete changes based on the numbers; this is the feedback loop for the harness.

Note: skill/routing telemetry is reconstructed from session JSONL by skill_analyzer.py, so it accumulates from when that analyzer was wired into session-export. The analyzer is idempotent (dedup-keyed), so Stop/PreCompact re-runs over a growing JSONL never double-count.