Implements chantier #1 agent-usage-stats from the « maturité v7 » epic (PR #194). Independent branch off main.

What it does

New Go subcommand:

ulk agents stats [--json] [--usage] [--source DIR]

• Composition from framework/agents/registry.json (source of truth) — by category, model, phase, with an internal count-field consistency check.
• Count reconciliation — compares the agent count declared in registry.md, CLAUDE.md, AGENTS.md, framework/agents/CLAUDE.md against the registry and flags drift.
• --usage — best-effort runtime overlay from .ulk-reports/accountability.jsonl, with a loud attribution-coverage caveat. Missing journal degrades gracefully.

Sample output (this repo):

  Total agents : 98  (registry count field: ✓)
  By model      sonnet 71 · opus 24 · haiku 3
  Count reconciliation (source of truth = registry: 98)
    framework/agents/registry.md   98  ✓
    CLAUDE.md                      98  ✓
    AGENTS.md                      98  ✓
    framework/agents/CLAUDE.md     98  ✓

Count reconciliation (the card's first goal)

The drift was real: registry 98 vs CLAUDE.md 95 vs AGENTS.md/agents/CLAUDE.md 94. This PR:

1. Regenerated registry.json/registry.md (main was stale at 97 — first commit).
2. Reconciled all declared counts to 98.

⚠️ Design blocker (important)

accountability.jsonl has an agent field, but attribution is not yet auto-propagated (accountability protocol v1.1 → v1.2 roadmap) — most entries are agent: "unknown". So per-agent invocation stats and reliable "dead agent" detection are NOT derivable today.

Consequence for the card's "identify dead trail (171→~40)" goal: the command cannot reliably flag unused agents yet. It reports attribution coverage instead of claiming disuse, and never presents absence of activity as proof. Unblocking this requires the accountability v1.2 auto-propagation work — a dependency worth surfacing in the epic.

Tests

go build + go vet + go test ./... all green. New tests: loader, Stats, count-field drift, accountability parsing + attribution ratio (internal/registry), declared-count regex against the 4 real doc phrasings + sort helper (cmd).

Code

• framework/cli/internal/registry/agents.go — loader + Stats() + LoadAgentUsage() (parsing/computation)
• framework/cli/cmd/agents.go — command + reconciliation + rendering (thin adapter)

Draft: the registry-composition + reconciliation half is complete and reliable. The usage half is intentionally caveated pending accountability v1.2. Archiving procedure (card AC) is not included — it should follow once usage attribution is trustworthy.

https://claude.ai/code/session_01JrFhrsWDBZVZBRMN6MiXG4

Generated by Claude Code