Skip to content

gaia-research/gaia-skill-tree / .agents/skills/gaia-draft-curate

gaia-draft-curate

Verified

Triage and classify pending Gaia skill intake batches in registry-for-review/. Use for gaia push intake proposals, draft skill batches, and deciding whether proposals should proceed to canonical curation. This skill is read-only and never promotes skills itself.

gaia-draft-curate

Review and classify pending skill proposals staged by gaia push. This is the intake decision gate between registry-for-review/ and canonical curation.

Ownership

This skill owns only intake batches and their linked intake issues/PRs. It does not:

  • mutate registry/, registry/gaia.json, or batch JSON files;
  • perform broad issue triage (use /gaia-triage);
  • deeply audit an existing canonical curation PR (use /gaia-curation-review);
  • promote accepted proposals (hand off to /gaia-curate-chain or /gaia-curate).

Pre-curation context

When invoked directly rather than through /gaia-full-pipeline, run /ev-scout before reviewing the batch. Provide its broad, explicitly unverified scout packet to the decision-maker. Do not turn scouting into evidence grading or registry mutation. Reuse an existing packet for the same intake when available.

Workflow

1. Establish a clean review state

git status --short --branch
python3 scripts/validate_intake.py

If the working tree contains unrelated changes, report them and do not alter them. If intake validation fails, stop and report the malformed batch.

2. Inventory immutable batches

Inspect registry-for-review/skill-batches/*.json, excluding .gitkeep. Record for each batch:

  • batchId, generation time, source, and proposed items;
  • proposal ID, name, type, description, source repository, supplied evidence, and similarity hints;
  • knownSkills as informational context only.

Never edit or rewrite an intake batch. These are immutable submission records.

3. Link batches to intake tracking

Use GitHub CLI when authenticated:

gh issue list --state open --search 'label:intake'
gh pr list --state open --label intake --json number,title,headRefName,url

Match batchId in titles, bodies, or comments. Report:

  • batches without a linked issue/PR;
  • linked issues/PRs without a local batch;
  • duplicate batches or proposals.

If GitHub authentication is unavailable, state that limitation rather than guessing.

4. Prepare the decision table

Use only facts present in the batch and verified canonical lookups. Do not invent evidence grades or impose repository-star/recency thresholds here; deep evidence research belongs to the downstream curation workflow.

#Proposed IDNameTypeExisting matchSupplied evidenceOpen questionsDecision
1skill-idSkill namebasic/fusiongeneric or noneURLs/artifactsmissing factspending

Check for obvious canonical duplicates with read-only commands such as:

gaia dev list --generic --named

For each proposal, collect a rationale and, when applicable, the target canonical ID for a rename or duplicate decision.

5. Collect explicit decisions

Work one batch at a time. Valid decisions are:

  • accept — proceed to canonical curation;
  • rename <new-id> — accept under a corrected ID;
  • duplicate <existing-id> — already represented;
  • needs-evidence — hold and state exactly what is missing;
  • reject — decline with a reason.

Every decision must include a short rationale. Do not silently discard proposals.

6. Produce a handoff packet

Summarize:

  • source batchId and linked issue/PR;
  • accepted/renamed/held/dropped proposals;
  • decision rationales;
  • canonical mappings;
  • evidence and URL questions still requiring research;
  • proposed branch scope (review/gaia-push/* for intake, review/meta/* for canonical changes).

The packet may be saved as /tmp/gaia-draft-curate-<batchId>.md. Write a repository Markdown artifact only when explicitly requested. Never modify the source batch.

7. Hand off only with confirmation

Ask whether to continue. By default, hand accepted items to /gaia-curate-chain for discovery-packet validation and L4 review only. The chain does not mutate the registry or sync docs. Use /gaia-curate only when the user explicitly chooses the lower-overhead path for a low-risk batch.

Issue-form intakes carry no batch, so that gate cannot read them: the workflow stops with "Intake issue has no canonical Batch ID row" before any recovery path runs, and recovery=true cannot help because the body has no metadata table or collapsible skill blocks. When L4 ratifies a topology that differs from the submitted proposal, reconcile it through the CLI rather than hand-writing a record: author the ratified topology as a skills: YAML, run gaia push --from-file <yaml> --no-issue so the CLI writes the batch, render the canonical body with build_intake_issue_body, and use it to replace the issue body while keeping the original submission verbatim and the ruling's provenance (upstream commit, source and packet SHA-256) as decision history. Commit the batch on a review/meta/* branch, since the workflow resolves the batch from that branch, and dry-run intake_recovery.parseIssueBody against the new body first — otherwise the gate fails publicly on the issue.

A completed topology approval is already the L4 decision — do not re-litigate it. Success leaves three durable signals: intake:topology-approved and intake:evidence-review on the issue; a bot comment carrying <!-- gaia-intake-approval {"topologyApprovedBy": …} --> with the batchId, batchSha256 and issueBodySha256; and an immutable workflow artifact (30-day retention) that the evidence stage re-verifies. Treat them as authorization and go straight to the evidence stage. Two rules follow: never edit an approved issue body — the tooling re-hashes it and fails closed with EXIT_STALE, so even a cosmetic "Decision: (pending)" tweak costs a fresh topology approval (add a comment instead) — and keep the branch named in the Batch Branch row alive, since the workflow resolves that row before main.

After L4 topology approval, a maintainer applies intake:topology-approved on the intake issue. The Intake Approval workflow validates the immutable batch, posts its machine-readable handoff, and applies intake:evidence-review; it does not seed or promote. An agent then prepares the Stage-1 seed and optional Phase-0 discovery plan, and marks it intake:evidence-ready for human review. Only a maintainer applying intake:evidence-approved to an evidence-ready intake opens the one draft review/meta/intake-<issue> promotion PR. Keep all verified evidence, CLI-only ingestion, provenance, and generated artifacts on that single PR.

8. Intake & Named Implementation Rules (Yggdrasil II)

  • Star ranks are non-authoritative at intake: Star levels (2★–6★) are never pre-declared in intake proposals. All named entries enter at the 2★ floor by default; higher ranks are derived canonically downstream via Trust Magnitude appraisal (trust-appraise).
  • Named entry slugs match upstream SKILL.md: A named implementation takes the format <contributor>/<skill_name>, where <skill_name> is the slugified name: declared in the upstream SKILL.md frontmatter (e.g. Bhanunamikaze/seo), NOT an arbitrary generic ID.

The handoff must preserve the originating batch and issue/PR links. This skill itself does not run gaia dev add, gaia dev evidence, or any other mutating command.

Optional Semantic Advisory Sidecar

To assist batch classification without altering intake state, an operator may optionally generate an auxiliary read-only advisory report (docs/agents/jev.md):

# Offline dry-run by default ($0 spend); add --live to query TypeSafe API
python scripts/jev_advisory.py --mode mapping --input registry-for-review/skill-batches/<batch>.json

The runner provides non-authoritative candidate alignment hints and falls back to worker-luna when uncertain, offline, or budget-limited. All batch decisions remain human-confirmed.

Final report

Batches reviewed: N
Tracking links: N found, N missing
Decisions: N accepted, N renamed, N held, N duplicated, N rejected
Promotion triggered: no (or explicit downstream handoff and branch)
Validation: commands run and results
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall