audit-activation

audit-activation

Reports whether OrchestKit's specialized sub-agents are actually being activated, from real spawn telemetry — not vibes. Scope is agent activation, not skill quality.

It answers: "Do my specialized agents get spawned, or does the model default to generic Explore/general-purpose?" Read-only — it never edits agents.

Why this exists: a 2026-06 audit found only ~14% of agent spawns hit the 36-agent catalog vs ~74% generic, with 17/36 agents dormant — and that agents fire ~1:1 with how often a high-traffic skill names them via subagent_type=. Description rewrites ("use proactively") were A/B-tested and gave Δ0, so this skill measures wiring + usage, not description prose. See docs/feat--activation-audit/.

Quick Reference

CRITICAL: Task Management is MANDATORY (CC 2.1.16)

TaskCreate(subject="Audit activation: agent spawn analysis",
  description="Computing generic-vs-specialist split + dormancy from spawn telemetry",
  activeForm="Auditing agent activation")
TaskCreate(subject="Read spawn telemetry", activeForm="Reading subagent-spawns.jsonl")
TaskCreate(subject="Compute split + dormancy", activeForm="Computing split and dormant agents")
TaskCreate(subject="Classify & render", activeForm="Classifying agents and rendering report")
TaskUpdate(taskId="3", addBlockedBy=["2"])
TaskUpdate(taskId="4", addBlockedBy=["3"])

Workflow

1. Run the script FIRST — every audit starts by running (or, when execution is impossible, explicitly referencing) the deterministic collector:

   bash "${CLAUDE_PLUGIN_ROOT}/skills/audit-activation/scripts/run-activation-audit.sh"   # add --json for machine output
   

   It reads the spawn telemetry, joins it against src/agents/, and prints the split, per-agent fire counts, never-fired list, and top-5 concentration. Never eyeball the JSONL by hand when the script exists.

2. Inventory — Glob src/agents/*.md (exclude README/INDEX/CONTRIBUTING) for the catalog.

3. Read telemetry — the FRESH stream is .claude/logs/subagent-spawns.jsonl (writers: pretool/task/spawn-intent-logger + subagent-start/subagent-validator). The legacy ~/.claude/analytics/agent-usage.jsonl is DEAD (orphaned in a refactor) — never read it.

4. Compute — all checks from Read("${CLAUDE_PLUGIN_ROOT}/skills/audit-activation/rules/activation-checks.md"); the Report Contract below lists the mandatory ones.

5. Classify — bucket every agent using the Four Buckets below (full procedure: Read("${CLAUDE_PLUGIN_ROOT}/skills/audit-activation/rules/activation-status.md")).

6. Render — output per Read("${CLAUDE_PLUGIN_ROOT}/skills/audit-activation/references/output-format.md"), satisfying the Report Contract.

Report Contract (every audit MUST include all six)

1. Data-source line (first line of the report) — verbatim form: Data: scripts/run-activation-audit.sh over .claude/logs/subagent-spawns.jsonl (window: <start> → <end>). The literal path .claude/logs/subagent-spawns.jsonl MUST appear in this line — "telemetry from the spawn log" or any paraphrase is a contract violation. Also state, once, that the legacy ~/.claude/analytics/agent-usage.jsonl was NOT read (dead stream). A report that presents numbers without citing the script and the literal file path is invalid.
2. Spawn split — generic (Explore/general-purpose/Plan) vs ork-catalog vs other-plugin, as counts AND percentages (e.g. "412 generic (74%) / 78 ork-catalog (14%) / 67 other (12%)"). Never percentages alone.
3. Concentration — the top-5 agents' share of all catalog spawns as a percentage (e.g. "top-5 = 81% of catalog spawns"), plus the top-5 list with fire counts.
4. Never-fired list — COMPLETE, by name. Enumerate EVERY catalog agent absent from the telemetry, one per line with its reference count and bucket. Never truncate to "17 dormant, e.g. these 4" — all names, every time.
5. Window caveat — verbatim, in every report: "Never fired" means absent from the available telemetry window — a strong signal, but NOT proof of zero lifetime use; the window may have gaps. State the window dates next to it.
6. Four-bucket classification table (below) + fix recommendations — wiring changes only (see Hard Rules).

Classification: the Four Buckets (rules/activation-status.md)

Bucket by reference counts, never by description quality. For each dormant agent, count its real spawn references: grep -rc "subagent_type=ork:<name>" src/skills/ (plus agent:/team-map mentions in src/agents/). Show the evidence: cite that grep command in the report and put each agent's ref-count (with an example source file, e.g. src/skills/cover/SKILL.md) in the classification table — telemetry proves firing, only the grep proves wiring. Always render all four buckets — including ACTIVE, and including empty buckets as "(0)" — even when the question asks only about dormant agents, citing rules/activation-status.md as the procedure.

DEAD requires zero references anywhere — an agent with even one reference is never DEAD; it is mis-triggered or niche. Only DEAD agents are prune candidates.

Hard Rules

• Never recommend rewriting agent descriptions. A/B-tested: description rewrites gave Δ0 routing benefit. Dormancy is a skill-wiring problem — the fix is always adding a subagent_type=ork:<name> spawn from a high-traffic skill (e.g. implement, cover, review-pr), never prose changes.
• Classify by reference count, not description quality. The grep above is the evidence; "the description is vague" is not.
• State the telemetry window AND the never-fired caveat in EVERY answer — including classification-only or prune-decision answers, not just full reports. If the words "never fired" or "dormant" appear anywhere in your output, the caveat sentence from Report Contract #5 must appear too.
• Name the telemetry file by its literal path in every answer. The string .claude/logs/subagent-spawns.jsonl must appear verbatim in your data-source line (it is the fresh stream you read); never substitute a paraphrase like "the spawn telemetry", and never cite the dead agent-usage.jsonl as a source.
• Answer with the report itself, never with only a description of the process. Even when you cannot execute the script in the current context, render the full contract-format report (state which numbers are from the latest available run vs illustrative).
• Read-only: this skill never edits agents.

Example report shape (illustrative numbers — imitate the FORM exactly)

Data: scripts/run-activation-audit.sh over .claude/logs/subagent-spawns.jsonl (window: 2026-05-23 → 2026-07-14)
Not read: ~/.claude/analytics/agent-usage.jsonl (dead stream — orphaned writer)

## Spawn split
Total 560 spawns: 412 generic (74%) / 78 ork-catalog (14%) / 70 other-plugin (12%)

## Concentration
Top-5 = 63/78 catalog spawns (81%): web-research-analyst 26, code-quality-reviewer 21,
test-generator 8, debug-investigator 5, backend-system-architect 3

## Never-fired agents (6 of 20 — ALL listed, refs via grep -rc "subagent_type=ork:<name>" src/skills/)
- emulate-engineer      (3 refs)  -> MIS-TRIGGERED
- expect-agent          (3 refs)  -> MIS-TRIGGERED
- design-system-architect (2 refs) -> NICHE (rare domain, wired)
- multimodal-specialist (1 ref)   -> NICHE (rare domain, wired)
- infrastructure-architect (2 refs) -> NICHE (rare domain, wired)
- event-driven-architect (1 ref)  -> MIS-TRIGGERED

Caveat: "Never fired" means absent from the available telemetry window — a strong
signal, but NOT proof of zero lifetime use; the window may have gaps.

## Buckets (rules/activation-status.md)
ACTIVE (14) | MIS-TRIGGERED (3) | NICHE (3) | DEAD (0 — none: every agent has >=1 reference)

## Fixes (wiring only — never description rewrites)
- emulate-engineer: add subagent_type=ork:emulate-engineer spawn from cover (src/skills/cover/SKILL.md)
- expect-agent: add subagent_type=ork:expect-agent spawn from review-pr

Chain

After this audit, run the deeper experiment if you suspect descriptions: docs/feat--activation-audit/agent-routing-experiment.mjs (isolated A/B selection harness).

Related Skills

• telemetry-inspect — validates the telemetry data-plane this skill reads from
• analytics — raw usage queries across projects
• doctor — broader plugin health (manifests, hooks, memory budget)

Eval note: this skill is direct-only (disable-model-invocation: true), so quality evals run in TIER-1 unit mode (--force-skill) — routed evals are impossible by construction.