telemetry-inspect

telemetry-inspect

One-shot health check for OrchestKit's telemetry pipeline. Reports writer activity, file sizes, schema lock coverage, orphan files, and growth warnings. Use when verifying the pipeline is flowing correctly or debugging a missing writer.

When to use

• Before or after a risky hook refactor, to prove telemetry still writes as expected
• Weekly health check on a long-running project
• When analytics output looks suspicious — inspect the underlying data first
• When adding a new telemetry file and wanting to confirm it's picked up
• Auditing which files are schema-locked vs. drift-vulnerable

What it checks

1. Writer activity — for each registered telemetry file, recent write count (from mtime scan) and last-write delta
2. File health — size (warn at 256 KB, critical at 1 MB), line count, mtime
3. Schema lock status — which files have validators in lib/telemetry-schemas.ts
4. Orphan detection — files on disk under .claude/{telemetry,logs,state,feedback}/ that aren't in the registry (possible stale writer or new file needing schema), plus inventory rows whose writer hook no longer exists in src/hooks/src/ (dead writer → orphan)
5. Growth trend — bytes per hour since session start (fire alert if > 100 KB/hr)
6. Coordination layer (M168) — live counts from sessions.db (running sessions, held locks, pending worktree links, skill invocations) plus write throughput from coordination-metrics.jsonl
7. Runtime fired-census — which hooks actually fired, from the per-invocation records run-hook.mjs writes to ~/.claude/analytics/hook-timing.jsonl. Checks 1–4 answer "is the file being written?"; this answers "is the writer running at all?" — the upstream question. A hook can be wired and reachable (the closure gate proves that statically) and still never fire: #2886 shipped exactly that with 12 green tests.

Usage

telemetry-inspect
telemetry-inspect --session sess-abc123
telemetry-inspect --json

Default mode: terminal-friendly ASCII report. --json emits a structured result suitable for piping into another tool or uploading.

Output shape (ASCII mode)

Telemetry Health — 2026-07-09 11:50
────────────────────────────────────

Schema-locked files (7)
  .claude/telemetry/pre-compact-decisions.jsonl  ◆ 3 lines  1.1 KB  ✓ healthy
  .claude/telemetry/image-responses.jsonl        ◆ 0 lines  —       ✗ no writes
  .claude/logs/decisions.jsonl                   ◆ 0 lines  —       ✗ no writes
  .claude/logs/subagent-spawns.jsonl             ◆ 6 lines  3 KB    ✓ healthy
  .claude/state/edit-history.jsonl               ◆ 94 lines 412 KB  ⚠ rotate
  .claude/state/ork-metrics-*.json               ◆ (N/A)    2.1 KB  ✓ healthy
  .claude/logs/skill-channels.jsonl              ◆ 12 lines 4 KB    ✓ healthy

Unlocked telemetry files (14)
  .claude/feedback/changelog-decisions.json      ○ 4 KB    ✗ no schema
  .claude/feedback/learned-patterns.json         ○ 8 KB    ✗ no schema
  (...14 more...)

Orphan files (1)
  .claude/feedback/skill-usage.json  — delisted, writer unwired since #959

Runtime fired-census (7d window)
  ALIVE  59   IDLE 3   NEVER 83   UNOBSERVABLE 54
  tree 8.75.0 · installed 8.73.0/8.74.0  <-- SKEW: NEVER is not a verdict
  never-fired writers (the ones that matter here):
    posttool/dirty-file-tracker   [PostToolUse]   0 fires, ever

Summary
  Pipeline health:  GREEN  (21/21 expected writers active)
  Schema coverage:  7/21 (33%)
  Largest file:     edit-history.jsonl (412 KB)
  Hotspot:          edit-history.jsonl  +40 KB/hr

Implementation plan (for an agent/LLM running this skill)

1. List known files — read lib/telemetry-schemas.ts's SCHEMA_LOCKED inventory for the 7 locked paths. Extend with the unlocked paths listed in the skill-local references/telemetry-inventory.md.
2. Cross-check inventory writers against the registry closure — never trust the inventory blindly:
   • Run node src/hooks/scripts/validate-registry.mjs (from the OrchestKit repo root; skip gracefully if not in the OrchestKit repo) and confirm it passes.
   • For every writer named in the inventory (e.g. posttool/metrics-bridge), verify the source file exists: test -f src/hooks/src/<writer>.ts.
   • Any inventory writer that is NOT a live src/hooks/src/ file is ORPHANED — report 🔴, never "expected-but-empty". Its file(s) on disk are orphans regardless of freshness, and the inventory row is stale (flag it for removal).
3. For each file:
   • Use Glob to resolve .claude/state/ork-metrics-*.json pattern → may be multiple
   • Use Read with limit: 10 to see shape and Bash wc -l for line count
   • Use Bash stat for mtime + size
4. Classify health:
   • size > 1 MB → critical
   • size > 256 KB → warn
   • mtime > 7 days → "no recent writes"
   • line count 0 → "no writes"
5. Orphan scan — Bash find .claude/{telemetry,logs,state,feedback} -type f cross-check against registered paths. Any on-disk files not in inventory → orphan. Merge in dead-writer orphans from step 2.
6. Runtime fired-census — steps 1–5 ask "is the file being written?". This asks the upstream question: "does the writer hook run at all?"
   • Run node src/hooks/scripts/fired-census.mjs --json (OrchestKit repo only; skip gracefully elsewhere — it exits 0 and prints an absent-census notice if ~/.claude/analytics/hook-timing.jsonl doesn't exist yet).
   • Report the class counts (totals), and the versions.skewed line verbatim when true — the census is written by the INSTALLED plugin across ALL projects, while the closure comes from this tree, so NEVER on a hook newer than the installed build is expected, not a finding.
   • Cross-reference: for every inventory writer from step 2 that is a live file, look up its census row. A writer that is ALIVE but whose file has 0 lines is a real defect (it runs and produces nothing). A writer that is NEVER is the #2886 class — surface it, but as advisory.
   • Never call a hook dead from NEVER alone. UNOBSERVABLE rows (dispatcher children) cannot appear by construction, and rare events (WorktreeRemove, PreCompact) look identical to a broken guard from here. Read the guard before concluding.
7. Render report — ASCII table by default, JSON if --json argument passed.

Core logic is deterministic + read-only. Do NOT write to any telemetry file — this skill is an observer.

Coordination layer (M168 #1915)

The SQLite coordination layer lives outside .claude/, at ~/.local/state/orchestkit/:

Source	What it tells you
sessions.db	live session / lock / worktree state (SQLite)
events.jsonl	coordination event stream (goal_converged, chain_stale, …)
coordination-metrics.jsonl	sessions.db write throughput counters (#1915)

Live counts — the DB file is a standard SQLite database; read it with sqlite3 (read-only SELECTs only):

DB="$HOME/.local/state/orchestkit/sessions.db"
[ -f "$DB" ] || echo "coordination layer idle (no multi-session activity yet)"
sqlite3 "$DB" "SELECT COUNT(*) FROM sessions WHERE status='running'"                  # live sessions
sqlite3 "$DB" "SELECT COUNT(*) FROM skill_invocation"                                 # skill invocations

Write throughput — coordination-metrics.jsonl is append-only {ts, metric, count} lines emitted async by lib/metrics-emitter.ts on every sessions.db write. Event rate ≈ recent sessions_db_write lines:

M="$HOME/.local/state/orchestkit/coordination-metrics.jsonl"
[ -f "$M" ] && tail -200 "$M" | grep -c '"sessions_db_write"'

Degrade gracefully: if sqlite3 is absent or the DB / metrics file doesn't exist, report "coordination layer idle" — never error. Like the rest of this skill, these are read-only observations.

Upstream OTel metric notes

When inspecting Claude Code's own OTel metrics (downstream of this skill — claude_code.* in your collector):

• CC 2.1.129+: claude_code.pull_request.count now also counts PRs/MRs filed via MCP tools (e.g., GitHub MCP create_pull_request), not just shell commands run through the Bash tool. Dashboards built before 2.1.129 will see a step-function increase at the cutover, annotate, don't alert. See ../monitoring-observability/references/metrics-collection.md for the join pattern that distinguishes MCP- from shell-filed PRs.
• CC 2.1.161+: OTEL_RESOURCE_ATTRIBUTES values are now attached as labels on all metric datapoints, enabling dimensional slicing (team, repo, environment). Existing dashboards keep working; new dashboards should use label selectors to segment usage.
• CC 2.1.145+: claude_code.tool OTEL spans carry agent_id + parent_agent_id, and background subagent spans nest under the dispatching Agent tool span. Build the trace tree by querying on parent_agent_id — enables per-skill fan-out timing and cost attribution for multi-agent skills (brainstorm, explore, implement); no schema change needed.
• CC 2.1.174+: /usage exposes CC-native per-component attribution — cache misses, long context, subagents, and per-skill/agent/plugin/MCP cost breakdowns over 24h/7d (surfaced first in the VSCode Account & usage dialog). Treat it as a cross-check source in the health report: if ork telemetry shows a skill/agent active but CC attribution shows zero usage for it (or vice versa), flag the divergence as a possible missing writer or stale install rather than trusting either side alone.
• CC 2.1.202+: telemetry from workflow-spawned agents (a /workflows run) carries workflow.run_id + workflow.name attributes, so a whole workflow run is reconstructable from OTel — group all agent events on workflow.run_id to rebuild one run's fan-out, and slice by workflow.name. See ../analytics/references/otel-fields.md ("From 2.1.202") for the field table; filter select(.["workflow.run_id"] != null) first, since ordinary session events lack these attributes.

Related

• lib/telemetry-schemas.ts — source of truth for schema-locked paths
• analytics — aggregates data across sessions (different use case)
• M121 "Observability Consolidation" milestone