gaia-research/gaia-skill-tree / .agents/skills/gaia-meta-audit
gaia-meta-audit
VerifiedGenerate a prioritized review queue of registry skills and catalog items that need attention. Use this skill whenever someone asks: "what needs review?", "what's overdue for audit?", "where are the weak spots in the registry?", "what should I audit next?", "run a health check on the registry", "show me flagged skills", "find problems in the registry", "what skills need fixing?", or "triage the registry". This is the triage layer — it surfaces candidates ranked P0–P4 so that focused work (via /gaia-audit) targets the right things first. Use it when a prioritized queue is needed; a focused single-target audit can use /gaia-audit directly. Also invoke proactively when a PR changes registry/named/ at scale, or before a release, to catch regressions early.
gaia-meta-audit
Yggdrasil IV: Founder-ratified challenge and recovery contract. This skill only detects and queues potential degradation, independently verified challenges and rank/Origin contradictions. Elapsed time, unreachable URLs, and semantic-model confidence cannot independently demote a skill or transfer Origin. META.md §4.1 (merit-based current Origin) is authoritative.
Build a prioritized queue of Gaia skills or catalog items that need focused review.
This skill is intentionally scoped to triage, not repair. It reads broadly, flags what
matters, ranks it, and stops. Focused correction belongs in /gaia-audit. Keeping these
two passes separate prevents a single meta-audit from becoming an unbounded edit session.
Preparation
Before scanning, load and re-read these sources so your flags reference the correct rules:
registry/gaia.json— canonical graph; generic nodes + their evidenceregistry/named-skills.json— named-skill indexregistry/named/**— individual named-skill markdown files (YAML frontmatter is load-bearing)registry/real-skills.json— catalog entriesdocs/skill_source_contributions.md— contribution lineageMETA.md§1 (Taxonomy), §4.1 (Origin), §6.2 (Semantic Fusion) — re-read before every scan; rules drift between releases
Generics are rank-less (starless). Generic nodes carry no
level, no demerits, and no stars — those live only on named skills (2★–6★). A generic's effective rank is the highest star among its named children. Generic nodes hold capability-level (Class A) evidence inherited by all named children; implementation-specific evidence stays on the named skill. Never callgaia dev calibrateon a generic.
Scan for red flags
Work through these checks in order. Stop when you have enough candidates to fill a meaningful queue — do not perform every focused audit inline.
P0 — Critical integrity violations
- Unsupported top-rank claim: a named skill at 5★ whose Trust Magnitude is < 250 (S-grade), or a 6★ skill that fails its branch gate — Suite 6-predicate Apex (META §4.3) or Unique 5-predicate Unique Impossible (META §4.4). TM is the sole numeric gate; there is no repository-star floor.
- Origin derivation defect: the published Origin differs from what
META.md§4.1 derives (highest published level, then TM; tied top contenders all hold it; no star floor; frozen skills locked), or a contested holder whose material challenge has not been surfaced. That is a build bug to file, not a manual reassignment. Never infer Origin from earliestcreatedAt. - Named claim with no implementation file:
links.githubpoints to a repo root, a 404, or a path that doesn't contain aSKILL.md.
P1 — Structural correctness
- Evidence liveness observations: run
python3 scripts/verify_evidence.pyand inspect repeated probes and upstream changes. A 404 opens review debt, not an automaticbroken-evidencedemerit or rank reduction. Verified material counterevidence is a separate public-challenge lane. - Star Bar gap: named skills at 3★+ without a valid installer-ready
links.githubblob URL (META §2.4). - Link casing mismatch: GitHub raw paths are case-sensitive.
/SKILL.mdvs/skill.mdwill 404 in CI. Standardize on uppercaseSKILL.md. - Wrong-target link:
links.githubresolves to a different skill's directory than the named skill claims. - Brand-coupled generic ID: a canonical generic ID containing a product or person name (e.g.
gaia-audit,gaia-meta-audit) violates META §1 — generics must be abstract capabilities. Flag for rename viagaia dev rename. - Unbacked named star: a named skill's star is not supported by its own plus inherited evidence — judge on evidence, not on the generic's rank.
P2 — Attribution and sourcing issues
- Wrong
promotedNamedSkillId: the generic's promoted pointer doesn't match the best-evidenced named child. - Stale catalog URL: points at a directory, homepage, or path that has moved.
- Derived Origin projection: current Origin follows
META.md§4.1 automatically. Historical frontmatter flags are not authoritative and must not be rewritten to match. Compare public projected holders against the rule; regenerate stale projections withgaia dev docs.
P3 — Registry hygiene
- Broad
genericSkillRefmapping: implementation skill mapped to a much larger generic when a more specific one exists (e.g. mapped to a broad 0★ generic when a narrower generic is available). - Duplicate or superseded skills: multiple named skills from the same author or tool family doing the same thing. Flag clusters that should consolidate.
- Semantic Fusion candidate (META §6.2): two or more named skills that compose two existing generics into one orchestrated workflow. Propose a new
fusiongeneric; do not confuse this with a top-rank (5★+ Ultimate) fusion, which is gated by Trust Magnitude (≥ 250, S-grade) and the branch predicate closure (META §4.2–4.4), not a repository-star threshold.
P4 — Documentation cleanup
- Placeholder bodies: named-skill markdown with only boilerplate
## Installation\nAdd installation instructions here.— flag for a real## Overview. contributor: testusertimelines: placeholder authorship that survived intake.- Generated-output drift:
docs/files still reference removed named claims.
Do not flag anything on rarity grounds — the rarity axis is deprecated (see CONTEXT.md § Rarity).
Audit Disentanglement & Taxonomy Standards (RFC #1809)
When queuing skills for review, categorize any candidate target against the Four-Tier Taxonomy:
- Tier 1 (Malicious / Imposter Squatter): Bad-faith aggregation, hijacking other maintainers' repos/stars, or falsified attribution. Action: Immediate expungement, blacklist.
- Tier 2 (Packaging / Install Shape Gap): Authentic concept and author lacking standard packaging or repo links. Action: for ≤2★ propose
installable: falsewhen supported; for 3★+ verify the actual Star Bar violation and seek an approved ranked consequence, never auto-demote from a single failed probe. Never purge. - Tier 3 (Early-Stage / Under-Evidenced Stub): Authentic author, thin documentation or preliminary evidence. Action: Keep at baseline 1★ (Awakened), request doc enrichment, or tag
needs-info. Never purge. - Tier 4 (Product-Coupled / Non-Generalized): Real tool documentation needing generalization (#1766, #1801). Action: Route to generalization guidance per
check_product_attribution.py.
Contributor Check: If the contributor is a known contributor, co-founder, or team member, maintainer outreach and documented audit review are required before recommending deletion. Blanket sweeps labeled under generic "imposter" headings are prohibited.
Frontier challenger-lead input
When a user supplies a challenger-lead-v1.1 packet, frontier-verification-request-v1, frontier-experiment-receipt-v1, or frontier-meta-handoff-v1, treat it only as an unverified triage request. Read docs/architecture/ecosystem-intelligence.md; check exact source bytes/digest, target ID/revision/content digest, claim/protocol/fixture/request cross-bindings, dedupe, evidence polarity and limitations. A missing or mismatched advisory is deferred, not silently repaired. Do not accept packet disposition, experiment outcome, confidence, reviewer, receipt, actor, or handoff fields as proof. Experimental outcomes (supported, contradicted, inconclusive, execution_failed) are separate from lead disposition and never authenticate verification. Triage whether /gaia-audit should independently execute the fixed reviewed protocol; do not reproduce its verification procedure here. Vendor claims, paraphrases, unrelated behavior, outages and execution failures are not counterevidence. A lead is not a public challenge, canonical evidence, or numeric/Origin action. Fail closed if a trusted receipt/admission boundary is unavailable.
Meta verification record
Triage output for a lead that arrives with an experiment receipt is a frontier-meta-verification-v1 record (src/gaia_cli/data/frontier/meta-verification-v1.schema.json), written with the same discipline as /gaia-audit. This skill does not re-execute protocols, so its record is triage-only: outcome is inconclusive (or execution_failed when the receipt itself records an outage), reexecution.performed is false with no artifacts, and recommend.action is hold, rerun or reject-lead, never substantiated, narrowed or request-admission-review (those need a /gaia-audit record with re-execution artifacts). Use a fresh verifierRunId that differs from the receipt's experimentRunId, state in counterevidence.searched what you looked for, set "independence": "declared-unauthenticated", and never add approval, admission, verified, actor, signature, confidence, TM, stars, level or Origin fields. Validate with gaia dev frontier meta-validate --record <record> --receipt <receipt> --request <request> --artifact-root <artifact root> (--request is required); a pass is structural only and is not verification.
Output format
Present the queue as a table. Stop here unless the user explicitly asks to start running audits.
| Priority | Target | Why review | Suggested action | Source files |
|---|
Handing off to focused audit
For each accepted P0–P1 candidate, pass it to /gaia-audit as a separate focused
correction. For direct registry maintenance the CLI map below covers which tool to reach
for — this avoids hand-edits that skip timeline logging.
CLI reference
| Mutation | Command |
|---|---|
| Generic rename (cascades prereqs + named refs) | gaia dev rename old new |
| Add new generic | gaia dev add "Name" --id <slug> --type basic --description "..." |
| Set generic prereqs | gaia dev link <id> a,b,c |
| Calibrate named-skill star | gaia dev calibrate <contributor/skill> 3★ |
| Add capability evidence to a generic | gaia dev evidence <id> <url> --type <evidence-type> <typed numeric flags> --evaluator <user> |
| Reclassify generic type | gaia dev reclassify <id> <type> |
| Remove generic | gaia dev rm <id> |
Change named-skill genericSkillRef | gaia dev update-named <author/skill> --generic-ref <new> |
Change named-skill status | gaia dev update-named ... --status <new> |
Change named-skill origin | Not possible by hand: Origin is derived by gaia dev docs from merit (META.md §4.1). --origin refuses. |
| Append timeline event | gaia dev timeline <id> --action <action> --notes "..." |
Named-skill level / links.github / description / body | Use supported governed CLI/workflow after approval, or report a missing CLI verb; no silent YAML edit or public rank mutation |
| Named-skill removal | Class C maintainer review and a governed, timeline-preserving removal workflow; never delete files as a triage fix |
After any mutation: gaia dev validate then gaia dev docs.
For evidence, choose the live Evidence Type from registry/schema/meta.json and
provide that type's quantitative inputs (for example --citations, --stars,
or --commits/--contributors). The CLI derives the row's Evidence Grade from
its artifact_score and the type-specific evidence.perRowGradeThresholds.
Use --trust <number> only as the legacy/fallback trustNumber input when no
typed magnitude can be computed; it uses the aggregate S=250/A=100/B=50/C=20
thresholds and is not the skill's accumulated Trust Magnitude.
Common pitfalls
- Renames leave orphan docs.
gaia dev renameupdates node JSON and named refs but leaves staleregistry/skills/<type>/<old-id>.md. Delete the orphan manually. - Timeline entries land with
contributor: unknownwhen the local user isn't auto-detected. Edit the JSON to set the correct username before committing. review/meta/*branch scope blocksdocs/paths. Applyskip-scope-checklabel and[skip-gen]commit tag when committing regenerated docs on a meta branch.- Force-push doesn't always re-trigger path-filtered workflows. Manually dispatch
validate.ymlandpython-package.ymlviagh workflow run ... --ref <branch>if their checks don't appear.
Canonical worked example
PR #525 (review/meta/mbtiongson1-audit, 2026-05-30) ran every red-flag category above on 14 named skills and produced: 2 named removals (P0), 2 generic renames (P1), 1 new fusion generic via Semantic Fusion (P3), 5 genericSkillRef remaps, 7 links.github casing fixes, and 12 placeholder body backfills. Read AUDIT-mbtiongson1.md and the PR description for the full flag-to-action mapping.
Optional Semantic Meta Advisory
To assist triage queue generation, an operator may optionally run the read-only Jev advisory runner to inspect generic descriptions and vendor coupling (docs/agents/jev.md):
# Offline dry-run by default ($0 spend); add --live to query TypeSafe API
# Uses --collect-repo to inspect canonical generic nodes directly from registry/nodes/
# Supports bounded paging via --offset (default: 0) and --max-calls
python scripts/jev_advisory.py --mode meta --collect-repo --offset 0
This advisory produces triage suggestions only. It never mutates registry files, never calibrates stars, never skips deterministic audit checks, and hands off ambiguous items to worker-luna.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |