Problem
The pr-flow and board-ops skills embed multi-line bash blocks that agents (and humans) reconstruct from scratch on every use. During the #2556 review loop the same blocks were rebuilt repeatedly:
- the review-wait loop (pr-flow 7b, ~30 lines) — rebuilt 5×
- the requestReviews GraphQL block (pr-flow 7a, hardcoded bot id) — 5×
- the paginated round-fetch + jq (pr-flow 8) — every round
- the card status move (pr-flow 1/6, board-ops) — 2×, and the most fragile: a half-pasted block silently no-ops
Each reconstruction is re-reviewed/re-approved from scratch, and any transcription slip (wrong jq filter, missed --paginate) fails in ways the skill prose can only warn about.
Proposal
Add a small maintainer-workflow category to root scripts/ (joining the existing gate-lease / skill-eval precedent), with npm aliases and stable stdout contracts:
| Script |
npm alias |
Prints |
scripts/pr-review-request.mjs |
pr:review-request -- --pr N |
confirmation; owns the bot id + GraphQL mutation |
scripts/pr-review-wait.mjs |
pr:review-wait -- --pr N --expected K |
exactly one of ROUND=posted|ended-without-review|timed-out; non-zero exit on API/parse failure (never retries blind) |
scripts/pr-review-fetch.mjs |
pr:review-fetch -- --pr N [--review ID] |
the round's comments (id, path:line, body), paginated |
scripts/board-card-status.mjs |
board:status -- --issue N --status "In Review" [--board 28] |
card: <Status> only on verified success; all ids resolved by name at run time |
- Node
.mjs with sibling *.test.mjs (pure logic tested; gh invocations behind a thin injected runner), matching the existing scripts/ conventions — flat layout, scripts/lib/ for shared gh/paging helpers.
SKILL.mds updated to invoke, not transcribe: each skill keeps the decision tables and the why, and names the command instead of embedding the bash.- Shell out to
gh for auth, like the existing sweeps.
Why root scripts/ and not skill-embedded scripts
scripts/*.mjs is already inside the test (node --test), lint, and format scopes; skill folders are outside all of them.- These encode repo facts (Copilot bot id, owner/repo, board invariants) — repo tooling like
sdk-watch.mjs, not skill prose.
- Node ≥22.19 is a checked repo requirement; Python is not.
Surfaced while running the review loop on #2556.
Problem
The
pr-flowandboard-opsskills embed multi-line bash blocks that agents (and humans) reconstruct from scratch on every use. During the #2556 review loop the same blocks were rebuilt repeatedly:Each reconstruction is re-reviewed/re-approved from scratch, and any transcription slip (wrong jq filter, missed
--paginate) fails in ways the skill prose can only warn about.Proposal
Add a small maintainer-workflow category to root
scripts/(joining the existing gate-lease / skill-eval precedent), with npm aliases and stable stdout contracts:scripts/pr-review-request.mjspr:review-request -- --pr Nscripts/pr-review-wait.mjspr:review-wait -- --pr N --expected KROUND=posted|ended-without-review|timed-out; non-zero exit on API/parse failure (never retries blind)scripts/pr-review-fetch.mjspr:review-fetch -- --pr N [--review ID]scripts/board-card-status.mjsboard:status -- --issue N --status "In Review" [--board 28]card: <Status>only on verified success; all ids resolved by name at run time.mjswith sibling*.test.mjs(pure logic tested;ghinvocations behind a thin injected runner), matching the existingscripts/conventions — flat layout,scripts/lib/for shared gh/paging helpers.SKILL.mds updated to invoke, not transcribe: each skill keeps the decision tables and the why, and names the command instead of embedding the bash.ghfor auth, like the existing sweeps.Why root scripts/ and not skill-embedded scripts
scripts/*.mjsis already inside the test (node --test), lint, and format scopes; skill folders are outside all of them.sdk-watch.mjs, not skill prose.Surfaced while running the review loop on #2556.