Skip to content

gaia-research/gaia-skill-tree / .agents/skills/gaia-review-meta-close

gaia-review-meta-close

Verified

Close out a review/meta branch that lands an evidence-ingest / curation batch: create the named nodes, ingest verified evidence, appraise Trust Magnitude, calibrate stars, wire suites, correct upstream naming, stage ONLY the intended artifacts (dropping CRLF churn and blocking leaks), validate, and open the PR. Use when a curation/ingest batch is verified and ready to land, or when the user says "close the review/meta branch", "land the ingest batch", "wrap up the intake PR", "/gaia-review-meta-close". Wraps the deterministic hygiene in scripts/review_meta_close.py and gates the judgment calls (calibration, Origin, suite promotion, naming, merge) behind explicit human checkpoints.

/gaia-review-meta-close

Turns the messy, rules-everywhere review/meta close-out into a bounded pipeline: the script owns the mechanical parts, the human owns the judgment parts. It exists because the ingest → appraise → calibrate → suite → name → stage → PR sequence has ~10 precise rules (Origin gate, suite-manifest survival, CRLF churn, badge/Class-S discipline, UTF-8 validate) that are easy to get wrong one at a time.

Backed by scripts/review_meta_close.py (status / stage / validate / check).

When to run

After /ev-pipeline (or /gaia-curate-chain) has verified the evidence and an L4 reviewer approved the batch. This skill lands it; it does not discover or re-verify evidence.

Upstream handoff: /gaia-ingest-batch hands off to this skill once evidence is ingested and appraised — it passes the branch, the affected contributor handles, the per-skill TM/grade, and the intake issues to Resolves. If you arrive here directly (nodes + evidence already in place), run from §3 onward; if the nodes/evidence do not yet exist, do §1–§3 first.

Operating principle

  • Programmatic by default. Every registry mutation goes through a gaia dev verb (Programmatic-First Policy). Never hand-edit frontmatter or timelines.
  • Human gates catch what the pipes missed. Stop at each GATE below for an explicit decision. These are the four places prior phases most often ship a wrong star, a stripped suite, a capability-vs-upstream name mismatch, or an accidental merge. A gate is a genuine stop — present the finding, wait.
  • Prefix every gaia dev call with GAIA_OPERATOR_OVERRIDE=1 PYTHONPATH=src and pass --no-build on every mutation; run exactly one build at the end.

Pipeline

0. Branch

Work on review/meta/<slug> cut from origin/main. Confirm you are on it (git rev-parse --abbrev-ref HEAD) before any edit — a commit landing on the wrong branch is a real failure mode. Commit + push after each logical unit.

1. Create nodes (dependency order)

Named-skill evidence cannot be written until the node exists, and a named node needs its generic ref to exist first. Order:

  1. Generic bucket: gaia dev add "<Name>" --id <slug> --type basic --description "..." --no-build (skip if the generic already exists — gaia dev list --generic | grep <slug>).
  2. Named skill: gaia dev add "<Name>" --named --contributor <handle> --generic-ref <slug> --description "..." --title "<Title>" --extra-fields '{"links":{"github":"<blob-url>"}}' --no-build.

Contributor handles must be lowercase (Panniantong → panniantong); the display name/title can keep original casing. Blob URL must be blob/<branch>/<path>, never tree/.

2. Ingest evidence (one row at a time)

For each verified row in the data lake, gaia dev evidence <contributor/skill> "<url>" --type <type> <typed-flags> --notes "..." --source-started-at YYYY-MM-DD --no-build. Use the numeric flags the type supports (--stars, --views, --citations, --reviewers, --commits, --contributors, --skill-count-in-repo, --percentile). A benchmark-result needs its full reproducibility payload (attestor, dataset/input hashes) — a mirrored third-party benchmark that can't supply them is EXCLUDED from TM; record it in the lake, not as a registry row with faked hashes.

3. Appraise (use the RIGHT tool)

TM is a derived value read after build — never write it. Read it with the canonical inspector, which loads the named .md evidence:

GAIA_OPERATOR_OVERRIDE=1 PYTHONPATH=src python3 scripts/inspectTrustMagnitude.py --skill <contributor/skill>

Or run scripts/trust_appraise.py --skill <contributor/skill> (which reads the named .md frontmatter, loads generic and named maps, and computes live Trust Magnitude and Grade). For a frozen skill the output labels FROZEN - locked TM ...: the lock is its standing and the recomputation beside it is information only (never restamp a frozen skill).

4. GATE — calibration (TM band, branch rules, and Origin)

Per META.md: TM bands map S≥250→5★, A≥100→4★, B≥50→3★, C≥20→2★, ungraded(<20)→1★. Every 4★ promotion requires Overall Trust Grade A (TM ≥ 100) and live verified blob evidence pointing to a concrete implementation file (The Star Bar, META.md §2.4).

However, Origin requirements fork by branch:

  • Unique branch (standalone skill, no suiteComponents): 4★ Unique and 5★ Unique Ultimate require only that the skill is its own bucket's Origin, which is automatic: the highest-merit implementation (published level, then TM; ties all hold it; no star floor; META.md §4.1). There is no fusion-prerequisite Origin requirement. Before proposing 4★ Unique, check the bucket: grep -rl "genericSkillRef: <slug>$" registry/named/. If a higher-merit skill holds the bucket, the A-grade skill caps at 3★; gaia dev calibrate reports this. Origin cannot be set by hand.
  • Suite branch (suite capstone carrying suiteComponents): 4★ Extra does NOT require bucket-level Origin on its generic bucket. The capstone represents a multi-skill suite rather than a single bucket monopoly. Suite origin requirements apply at 5★ Ultimate (where the proposer must hold Origin on ≥1 of the suiteComponents per META.md §4.2). Existing 4★ suite capstone precedents include leonxlnx/taste-skill (4★ Extra, origin: false on generic design-generation, TM 172.43 Grade A) and gsd-build/get-shit-done (4★ Extra).

STOP. Present a table (skill | branch | TM | grade | bucket occupancy / suite status | proposed star) and get explicit operator approval. Then apply approved calibrations: gaia dev calibrate <contributor/skill> <N★> --no-build (star glyph, not bare int). A skill landing at 1★ is not Named — that is a legitimate outcome for ungraded Stage-1 skills; the redaction validator handles it, status: named is fine.

5. GATE — suites (use gaia dev fuse, not update-named)

If the batch includes a suite capstone: gaia dev update-named --suite-components alone does NOT survive gaia dev build — gaia dev docs strips suiteComponents from frontmatter unless a suite manifest at registry/suites/<contributor>/<slug>.json backs it. Establish suites with:

gaia dev fuse <generic-id> --name "<Name>" --description "..." \
  --prereqs <comp-generic-ids> --named-capstone <contributor/capstone> \
  --suite-components <contributor/comp-1>,<contributor/comp-2>,... --no-build

Verify after build that grep -c suiteComponents registry/named/<contributor>/<capstone>.md is ≥1. STOP for the capstone rank decision (recipe-driven TM can overstate a young repo — a founder ruling may hold it below its TM band).

6. GATE — upstream naming

The intake slug is often the capability name, not the real upstream skill (e.g. ux-audit for ui-ux-pro-max, format-output for i-have-adhd). Compare each named id against its links.github folder/file. STOP and confirm which should be corrected — intentional command-name divergences (documented in the intake issue) are left alone. To correct: gaia dev rename <contributor/old> <contributor/new> (named-skill rename is supported), then gaia dev update-named <contributor/new> --name "<Upstream Name>" --title "<Upstream Name>". The genericSkillRef (capability bucket) does not change — only the named id + display.

7. Build once, then hygiene + stage

GAIA_OPERATOR_OVERRIDE=1 PYTHONPATH=src python3 -m gaia_cli.main dev build   # single build
python scripts/review_meta_close.py status -v          # classify real / EOL-noise / leaks
python scripts/review_meta_close.py stage --contributor <h1,h2,...> --apply   # stage intended only
python scripts/review_meta_close.py check               # must be CLEAN

The script stages ONLY: registry/{named,nodes,suites}, tracked registry/*.md, docs/graph/* (Class S — Guard E), docs/badges/registry.json, and docs/badges/_assets/<h>/ + docs/og/<h>/ for the handles you pass. It renormalizes to LF (killing CRLF churn) and never stages a leak. Badges are only 2★+ (1★ carry none). Everything else the build touched (version-bump churn on unrelated profiles, other contributors' badges) is intentionally left out — never git add -A. If stage lists real changes OUTSIDE the allowlist, review them: they are either version-churn (discard) or need another --contributor.

8. Validate + commit + push

python scripts/review_meta_close.py validate    # gaia dev validate, UTF-8-safe, must PASS
git commit -m "..."   # conventional; note calibrations, renames, suite in body
git push

Merge conflicts land only in generated docs/graph/* — take --ours, complete the merge, re-run the single build, re-stage docs/graph/*. Never hand-merge them.

9. PR

Open (or edit) the PR with --body-file (real newlines):

  • Resolves #<n> for each fully-closed intake issue. A suite issue with sub-skills held for separate promotion is listed but NOT auto-resolved — say so explicitly.
  • Summary table: skill | TM | grade | calibrated stars.
  • Note the pre-ingest fixes (e.g. a tree/→blob/ correction) and any renames.
  • Include an Entrypoints section (usually "registry-data + CLI PR; no new user-facing pages" for a pure ingest).
  • Verify: gh pr view <n> --json body,mergeable.

10. GATE — merge

Do not merge. Report the PR URL, CI status, and the calibration table. The final merge to main is a human decision (this skill never merges).

What the script owns vs what you own

Mechanical (script)Judgment (you, at a GATE)
classify real vs EOL-noise vs leakswhich star each skill lands at (§4)
LF-renormalized allowlist stagingOrigin bucket resolution (§4)
refuse/flag leaks, never git add -Asuite capstone rank (§5)
UTF-8-safe gaia dev validatewhich names to correct vs leave (§6)
staged-EOL-noise + conflict preflightwhich issues to Resolves (§9); merge (§10)

Known tooling frictions (worked around here; report persistent ones)

  • gaia dev build strips suiteComponents without a suite manifest → use gaia dev fuse (§5).
  • gaia dev validate crashes on Windows (cp1252 on ✓) → the script runs it UTF-8-safe.
  • trust_appraise.py --skill resolves named frontmatter and computes live TM/Grade (fixed in Issue #1742).
  • Windows checkouts introduce CRLF churn on every touched file → the script's LF-renormalized allowlist drops it.
receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementationStars: 322kUpdated Claude CodeCodexVerifiedInstall
requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirementsStars: 322kUpdated Claude CodeCodexVerifiedInstall
using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallbackStars: 322kUpdated Claude CodeCodexVerifiedInstall
verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions alwaysStars: 322kUpdated Claude CodeCodexVerifiedInstall