Token Audit — Duplication → Redirect → Measured Savings

Scope: $ARGUMENTS

Describe which workflow copies to audit. Include:

• The canonical source (council canonical: presets/agy/workflows/cc-council.md)
• The redirect targets to verify (the 5 council copies on codex/claude/cursor/opencode/gemini)
• The baseline in bytes (council baseline: 19327, about 19KB)

Rules

C1 — Council duplication

Every non-canonical council copy MUST be a minimal redirect stub: a GENERATED header plus a pointer to the canonical copy, under 400 bytes each. A copy carrying the full Steps 0-5 body is a C1 violation. Stubs are produced ONLY by regenerating with inject-ccep-bootstrap.ts + render-agent-commands.ts — never by hand-editing the 5 copies.

C2 — Fallback sync

COUNCIL_FALLBACK_PROFILE MUST be generated from src/core/ccep/workflows/council.yml, never hand-synced. The generated mirror lives at src/core/ccep/council-fallback.generated.ts and is covered by bun run check:drift: any drift between the YAML roster and the fallback artifact fails the gate. A hardcoded fallback literal in profiles.ts is a C2 violation.

S1 — STOP/scope is report-only

Count STOP gates and scope mentions, report the count (for example: 2 STOP gates across the 5 council phases), and never block on it. S1 is report-only by design: no refactor, no gate failure may follow from the count alone.

Official method: bytes/4 + bytes/%

Tokens are estimated as bytes/4. Savings are reported in bytes and in bytes/% against the baseline:

measured = sum of the 5 stub sizes in bytes
saved    = 19327 - measured
pct      = saved / 19327 * 100
tokens   ~= measured / 4

There is no cutoff threshold: always report the measured numbers, even when the reduction is small.

Gates

• bun run check:council-redirect — every council copy is a stub under 400 bytes pointing at the canonical file. Fails naming each offending copy; passes on a clean tree so work can proceed. Prints the measured total with bytes/% saved.
• bun run check:drift — regenerates the fallback artifact from council.yml plus all derived presets, then fails on any uncommitted diff.

Regeneration (never manual)

bun run scripts/inject-ccep-bootstrap.ts
bun run scripts/render-agent-commands.ts

Both passes are idempotent: a second run changes zero bytes. If check:council-redirect fails after a regeneration, the stub format regressed — fix the generator, not the copies.

Completion criteria

• 5/5 council copies are stubs (<400 bytes each, no full Steps body, GENERATED banner + canonical pointer).
• Fallback artifact matches council.yml and drift is clean.
• Savings reported in bytes and bytes/% vs the 19327 baseline, with the S1 STOP/scope count attached.