Approve All

Approve-all — a one-confirmation human bulk-approval, prepared not forged

This is the fast path to draining the review backlog: promote every approvable claim to canonical in one go. The catch it respects — the AI must not sign for you. So it does everything up to the write (find the batch, show it, confirm it) and then hands you the exact one-line command to run yourself. You fire it; the log records by: human because a human really did.

Why the AI hands off instead of committing

canonical · by human is supposed to attest that a human approved (ADR-0004, BUREAU.md). The decision log is unsigned, so by: is a claim, not proof — the whole gate rests on the AI never writing by: human. An in-session "yes" cannot be seen at the log layer: if the AI's process ran the approval, every by: human entry would become indistinguishable from a fabricated one. Running the command yourself (via the ! prefix) keeps the attestation true at one keystroke. approve --all is also a documented weakening (ADR-0005) — it promotes the backlog sight-unseen — which is all the more reason it is the human's to fire.

Steps

1. Locate the workspace. Read bureau.json for the workspace dir (default canon); honor --workspace <name>. If no workspace exists, tell the user to run bureau:init first and stop.

2. Build and split the queue. Run node "${CLAUDE_PLUGIN_ROOT}/press/bin/gazette.mjs" review --json --dir <workspace> — invoke it with the ${CLAUDE_PLUGIN_ROOT} variable literally; never substitute a hand-picked ~/.claude/plugins/cache/…/<version>/… path (a stale cache silently runs the wrong engine). Parse {items, counts, total} and split by kind:

   • approvable — approve and reapprove: exactly what approve --all promotes.
   • not covered — confirm-dependencies, resolve-conflict, repair-edge: approve --all leaves these untouched (a confirm can't be an approve, a conflict needs a decision, a broken edge needs an edit). They stay in the queue.

   If review --json exits non-zero or returns output that is not valid JSON, report the raw error and stop — never fabricate the queue or hand off a run command against an unknown backlog.

3. Stop if there is nothing to approve. If the approvable set is empty, report "nothing to bulk-approve" plus any not-covered items, and stop. Never confirm approving an empty batch.

4. Show the batch, then confirm in-session. List the approvable titles (what would become canonical) and the not-covered items. Then take the confirmation with AskUserQuestion: "Approve all N page(s) as canonical — by you (by: human)?" with options Approve all and Cancel. On Cancel, stop — nothing is approved.

5. Hand off the one line (never run it yourself). On Approve all, resolve the absolute path to gazette.mjs (expand ${CLAUDE_PLUGIN_ROOT}) and print the ready-to-run command, --dir pre-filled, for the human to fire with the ! prefix:

   ! <abs path>/press/bin/gazette.mjs approve --all --dir <workspace> --by human
   

   Explain briefly: running it via ! executes it as you, so by: human is true; it is non-interactive, so it proceeds without a y/n and content-binds each page (capturing every digest, refusing any page it cannot bind). Do not run gazette approve yourself, and do not pass --by human from your own process.

6. Offer the reviewed alternative. Note that if approving sight-unseen is too blunt, a per-page reviewed batch keeps human authorship AND records exactly the bytes vetted: gazette review --json > decisions.json → prune it → approve --from decisions.json --by human; and that bureau:codex-review can pre-screen the queue first.

Rules

1. The AI never approves and never signs --by human. It prepares and hands off; the human runs the one line. This is the non-negotiable core of the gate.
2. Confirm before handing off. Always show the batch and take an explicit in-session confirmation first — an empty or unconfirmed batch is never handed a run command.
3. Only approve/reapprove are in scope. Name the not-covered items so the user knows the queue is not fully cleared; route them to bureau:review.
4. Name the weakening. approve --all promotes sight-unseen (ADR-0005); say so, and offer the approve --from per-page alternative.
5. Read-only until the handoff. The skill only reads the queue; the sole state change happens when the human runs the printed line.

Output format

Report, in order: (1) the workspace; (2) would approve — the count and titles that become canonical; (3) not covered — the confirm/resolve/repair items left in the queue; (4) your one line — the exact ! … approve --all --dir <workspace> --by human command, plus the approve --from reviewed-batch alternative.

Example

Scope note

This skill ONLY prepares a human bulk-approval and hands off the command. It does not run any approval, capture sessions, compile the logbook, resolve conflicts, or render the board. It is invoked by bureau:approve-all, and is the bulk companion to bureau:review (per-page human gate) and bureau:codex-review (Codex pre-screens the queue as your representative).