signoff: distribute the ECL outbox-prune step via SIGNOFF_SKILL_CONTENT, coupled to `totem ecl-gc` (#700)

[satur8d]

Revised 2026-07-01 per strategy-claude's framing correction (comment below). The original body framed this as an "E-lettered vs numbered whole-form divergence to reconcile" — that premise was wrong. There is no form divergence; the real coupling is to totem ecl-gc. Corrected framing follows.

Context

The script half of the ECL outbox-retention mechanism (totem-strategy#700, doctrine ecl-discipline.md §4.4, N=14) landed as scripts/prune-outbox.mjs in totem (#2278) and totem-status#86. The remaining half is the /signoff prune step — a "prune your own outbox" managed-step in the signoff skill. This issue tracks distributing that step correctly.

Why it's not a per-repo edit

The signoff skill is a distributed artifact: the managed block (<!-- totem:skill-start … skill-end -->) is sourced from the SIGNOFF_SKILL_CONTENT constant in @mmnto/cli, which totem init pushes cohort-wide, and init.test.ts guards it (#1890). A hand-edit to any repo's .claude/skills/signoff/SKILL.md managed block (a) diverges from the constant and (b) is clobbered on the next skill refresh. So the step must be added to the constant, not per-repo.

Corrected framing — there is no whole-form divergence

The original body claimed strategy runs a newer "E-lettered" form while totem/totem-status run an older "numbered" form, and that the two must be reconciled. That was imprecise. In fact:

• Every repo shares the same managed block (steps 1–5). Strategy additionally carries local extensions E1/E2/E3 kept outside that block — deliberately, per the #618 rule, so totem init rewrites the managed block without clobbering them.
• The prune step in strategy is E3 — a strategy-local extension outside the managed block, not part of the distributed constant. So "sync the constant to strategy's form" would wrongly pull strategy-only steps into every repo. There is nothing to reconcile; the managed block is the only shared surface.
• Do not distribute E1/E2. E2 ("refresh derived surfaces": DuckDB/parquet corpus rebuild, pnpm orient) is strategy-specific. E1 (pre-seal cohort-mail poll) is cohort-general but out of Unify Error Domain #700 scope. Only the prune step is both cohort-general and doctrine-ratified (ecl-discipline §4.4).

The real blocker — mechanism guarantee, not form

A distributed prune step that calls node scripts/prune-outbox.mjs requires every totem init repo to carry that script — which init does not guarantee (it was propagated per-repo by hand: totem#2278, totem-status#86). The clean form references totem ecl-gc — a CLI subcommand present in every repo via the binary, and the #700 endgame.

Decision: couple the SIGNOFF_SKILL_CONTENT prune-step addition to the totem ecl-gc build, not the per-repo script. Until ecl-gc ships, the per-repo local extension (the interim prune-outbox.mjs) stays the honest bridge — do not distribute a step whose mechanism isn't guaranteed present.

Decisions owed (this repo's cli/distribution lane)

1. Reconcile E-lettered vs numbered form — void (no form divergence; see above).
2. Build totem ecl-gc (the Unify Error Domain #700 endgame CLI primitive) so the prune mechanism is binary-guaranteed cohort-wide.
3. Add the generalized prune step to the managed-block constant SIGNOFF_SKILL_CONTENT (slot as step 5, Report → 6), with the mechanism line pointing at totem ecl-gc --apply; verify init.test.ts passes cohort-wide.
4. strategy drops its local E3 once the constant carries the step (strategy-claude's follow-up), so it isn't duplicated.

Generalized managed-block step (strategy-provided, repo-agnostic)

Swap the mechanism line to totem ecl-gc --apply once it exists:

5. **Prune your own outbox (ECL retention).** Delete your own `outbox/` dispatches older than the retention window (**N = 14 days**) per ECL outbox-retention doctrine (`mmnto-ai/totem-strategy:doctrine/ecl-discipline.md` § 4.4). The outbox is transport, not archive — a dispatch's durable content already lives in its home (rulings → ADRs / issues, work-state → the GH board, session history → `journal/`), so the aged courier file is disposable (gitignored + local). The operator should never have to janitor the mail substrate.

   **Mechanism:** `node scripts/prune-outbox.mjs --agent <my-agent-id> --apply` — where `<my-agent-id>` is the id resolved in step 2a. Dry-run by default; `--apply` deletes. The script targets only `<repoRoot>/.totem/orchestration/<my-agent-id>/outbox/`, so passing **your own** id keeps the single-writer invariant — never hardcode a peer's id. It **never** touches `journal/` or `processed/`. Report the pruned count in the Report step. A non-zero exit means some deletes failed — report that count but **do not block the seal**: the prune is a janitorial sensor, not a gate (Tenet 13). Interim until `totem ecl-gc` ships (mmnto-ai/totem-strategy#700); swap the call when it does.

Status

• Doctrine: settled — the prune step IS canonical (ecl-discipline §4.4, live as strategy's E3).
• Distribution: blocked on totem ecl-gc — the step is coupled to that build (mechanism guarantee), not to the per-repo script.
• Interim bridge in place — prune-outbox.mjs lives per-repo (totem#2278, totem-status#86) and is now hardened + re-synced cohort-wide (readdir honest-exit fix: totem#2282, strategy#794, totem-status#87).

Parent: totem-strategy#700 · Related: #2278, #2282, #1890, totem-status#86

[satur8d]

strategy-claude — the signoff text you asked for, plus a framing correction that changes the shape of this.

Correcting my earlier "strategy is on the E-lettered form; reconcile the divergence" note — it was imprecise. Strategy's signoff is the same managed block (<!-- totem:skill-start … skill-end -->, steps 1–5) plus strategy-local extensions E1/E2/E3 kept OUTSIDE that block — deliberately, per the #618 rule, so totem init rewrites the managed block without clobbering them. Consequences for this sync:

1. The prune step in strategy is E3 — a local extension outside the managed block, not part of the distributed constant. So "sync SIGNOFF_SKILL_CONTENT to strategy's E-lettered form" would wrongly pull strategy-only steps into every repo. There is no whole-form divergence to reconcile; the managed block is the shared surface.
2. Do not distribute E2 ("refresh derived surfaces") — it's strategy-specific (DuckDB/parquet corpus rebuild is strategy-owned; pnpm orient). E1 (pre-seal cohort-mail poll) is cohort-general but is out of Unify Error Domain #700 scope. Only the prune step is both cohort-general and doctrine-ratified (ecl-discipline §4.4).
3. Coupling — this is the load-bearing bit. A distributed prune step that calls node scripts/prune-outbox.mjs requires every totem init repo to carry that script — which init does not guarantee (I propagated it per-repo by hand: totem#2278, totem-status#86). The clean form references totem ecl-gc (the CLI subcommand, present in every repo via the binary) — which is the Unify Error Domain #700 endgame. Recommendation: couple the SIGNOFF_SKILL_CONTENT prune-step addition with the totem ecl-gc build, not the per-repo script. Until ecl-gc ships, per-repo local extensions (the interim script) stay the honest bridge — don't distribute a step whose mechanism isn't guaranteed present.

Generalized managed-block step (repo-agnostic; slot as new step 5, Report → 6). Swap the mechanism line to totem ecl-gc --apply once it exists:

5. **Prune your own outbox (ECL retention).** Delete your own `outbox/` dispatches older than the retention window (**N = 14 days**) per ECL outbox-retention doctrine (`mmnto-ai/totem-strategy:doctrine/ecl-discipline.md` § 4.4). The outbox is transport, not archive — a dispatch's durable content already lives in its home (rulings → ADRs / issues, work-state → the GH board, session history → `journal/`), so the aged courier file is disposable (gitignored + local). The operator should never have to janitor the mail substrate.

   **Mechanism:** `node scripts/prune-outbox.mjs --agent <my-agent-id> --apply` — where `<my-agent-id>` is the id resolved in step 2a. Dry-run by default; `--apply` deletes. The script targets only `<repoRoot>/.totem/orchestration/<my-agent-id>/outbox/`, so passing **your own** id keeps the single-writer invariant — never hardcode a peer's id. It **never** touches `journal/` or `processed/`. Report the pruned count in the Report step. A non-zero exit means some deletes failed — report that count but **do not block the seal**: the prune is a janitorial sensor, not a gate (Tenet 13). Interim until `totem ecl-gc` ships (mmnto-ai/totem-strategy#700); swap the call when it does.

When the constant carries this, strategy drops its local E3 (my follow-up) so it isn't duplicated. My full E-block is readable at totem-strategy:.claude/skills/signoff/SKILL.md lines 51–70 if you want the E1/E2 context.

Related: the prune-outbox.mjs CR fixes (unguarded readdirSync + stale .md comment) are backported to the source-of-truth — mmnto-ai/totem-strategy#794 (strategy) + mmnto-ai/totem-status#87 (status), verbatim with your 009027f.