Skip to content
cccskills
FreePublished on GitHub

gaia-curate-chain

Fixed-topology, discovery-only extension for Gaia curation. Atomically checkpoints each candidate transition, retries only the failed field, and stops at human L4 review.

The English summary is being prepared. Showing the original description.

Run score Not tested

See how to install

Files included (1)

  • SKILL.md4.2 KB

SKILL.md (original)

Review the instructions the agent will receive before you install.

Gaia Curate Chain

Read ../gaia-curate/CURATION-CORE.md; it owns the lifecycle and packet. Use this extension when recoverability matters more than throughput.

Checkpoint contract

Persist a run ledger under generated-output/curate-discovery/<run-id>/run.json containing the core contract digest, generic snapshot digest, source cursor, active candidate ID, completed candidate IDs, deferred rows, and next instruction. Write to a temporary file, validate it, then atomically rename it; never leave a partially written checkpoint.

Process exactly one candidate at a time. Validate its discovery-packet-v2 after every transition and persist each final review packet to Core's registry-for-review/discovery-packets/ path. V1 remains readable only for compatibility. Advance only when the current transition validates and its input digest matches the checkpoint. On resume, revalidate contract and generic snapshot digests; preserve raw source records but move stale mappings to DEFER.

Keep current-state blockers and sourceRevisions separate from historical events. Each blocker records checkId, status (open or resolved), summary, and resolutionEvidence when resolved. On resume, read gaia_cli.curationCheckpoint.compactCheckpoint(ledger, freshFailingChecks) first; only checks actually rerun during this resume, with fresh failure evidence, belong in freshFailingChecks. Resolved blockers with evidence stay resolved; unproven resolutions and open blockers remain active. Never promote old failure history into a fresh check. Persist new failures as open blockers before the next checkpoint. The CLI's .gaia/curation/runs/<run-id>/state.json stores the same fields under checkpoint; gaia curate --resume <run-id> [--status] reports that compact state and pauses on active blockers. Neither ledger format grants L4 approval.

Bounded repair

Route only the failing candidate and field back to its worker. Allow at most two retries for transient fetch failure and one Luna High/equivalent repair for invalid structured output. A deterministic schema, hash, or generic-reference failure is not blindly retried.

If retries are exhausted or output remains invalid/ambiguous, emit DEFER with the stable core reason code and exact resume instruction: candidate ID, failed field, validator code, source cursor, and required operator action. Never synthesize or infer a decision.

The chain ends at L4 human review. Shared reconciliation accounts for every expected candidate exactly once (review-ready, deferred, rejected or proven duplicate), including failures. Review-ready is not human-approved. It does not collect evidence, calculate trust, mutate the registry, regenerate docs, create intake batches, commit, push, or open a PR.

Explicit downstream handoff

After an L4 decision, write a handoff record; do not silently end the caller:

  • New external discoveries: after independently verified L4 human approval bound to the exact reconciled inventory and manifest, ratify the packet via gaia dev ratify with a trusted authorization envelope. Local metadata is not authority; the descriptive flags remain (--assessment, --reviewed-by, --approval-ref, --reason, --acknowledge-human-review), which attaches the schema-defined l4Resolution, then run gaia push --from-file <packet.json>. Do not manually translate packets to skills.yml or fabricate resolutions; the adapter owns that deterministic translation and refuses unresolved identity or non-blob provenance.
  • Rows that already came from /gaia-draft-curate: preserve batchId and issue/PR links, return them as needs-evidence to /ev-pipeline, then hand verified rows to a maintainer on review/meta/* for the CLI-only meta-shift in CONTRIBUTING.md §1C. Do not submit the same intake twice.

Record the route, artifact path, remaining gate, and exact next command. This compatibility handoff is data only; the chain remains discovery-only.

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

Measure CI churn cost for a pull request: how many commits and how many CI-compute seconds were burned on avoidable "fix the CI" push rounds. Use when you want to understand the efficiency cost of a PR's iteration cycle, get actionable pre-push check suggestions, or include a churn summary in a pipeline report. Triggers: "ci churn", "how much CI did we burn", "avoidable commits", "ci efficiency", "wasted CI time", "churn report", "/ci-churn". Called automatically by /fp-drift at pipeline close.

The English summary is being prepared. Showing the original description.

gaia-research/gaia-skill-tree23Updated Oct 24, 2026

Produce the rendered evidence a human-gated frontend PR owes its reviewer: capture each changed surface (plus the before state for anything modified, plus mobile width), then generate one self-contained local HTML page the founder can open and review without checking anything out. Use when a PR is gated by the "Frontend changes are HUMAN-GATED" section of CLAUDE.md, or when the user says "/design-gate", "screenshot evidence for this PR", "prep this for founder review", "build the design evidence page", "what does this look like before and after". The page comes from a template script, not hand-written markup. Output is session-local and never committed. Do NOT use for non-gated changes: typo/grammar/factual fixes in an unchanged block, dead link targets, alt/ARIA fixes with no visual change, cache-bust bumps, or refactors with no rendered diff.

The English summary is being prepared. Showing the original description.

gaia-research/gaia-skill-tree23Updated Oct 24, 2026

One-fix-at-a-time, approval-gated design fix recovery. Given a checklist of fixes (handover doc, CHECKLIST.md, or inline list), applies each individually, presents a diff, and waits for an explicit APPROVE / REJECT / NO CHANGES decision before committing. Trigger phrases: /design-iteration, "go through the checklist one by one", "fix these one at a time", "bucket recovery", "design fix pass".

The English summary is being prepared. Showing the original description.

gaia-research/gaia-skill-tree23Updated Oct 24, 2026

Playbook for correcting a named skill's star level to match its computed Trust Magnitude grade, with an independent post-calibration verification pass and a CLI-logged timeline event (never hand-written). Use when: recalibrating a skill's level, fixing a stale or drifted star rating, running a batch of `gaia dev calibrate` corrections, or confirming a past calibration actually landed. Triggers: "calibrate <skill>", "recalibrate the trust magnitude backlog", "promote/demote <skill> to N stars", "why is <skill> still N stars", "verify a calibration actually landed", "Hall of Heroes ranking is wrong".

The English summary is being prepared. Showing the original description.

gaia-research/gaia-skill-tree23Updated Oct 24, 2026

Close out a completed EPIC / sprint end-to-end: author a fact-verified meta-post announcing what shipped, archive the EPIC's handovers, snapshot orchestrator memory, gate the integration→main merge on CI-green + explicit human go, drive the version bump (merge-commit conventions that auto-sync classifies), and verify the release reached the registry/PyPI. Use when the user says "close out EPIC <n>", "wrap up the sprint", "ship the meta shift", "/epic-close", or asks to finalize an integration branch that is ready to merge to main. Generalizes the EPIC 1002 (Yggdrasil II) close-out. This skill orchestrates — it authors founder docs directly (superadmin) and delegates code/regen to workers; it NEVER merges to main without the human gate.

The English summary is being prepared. Showing the original description.

gaia-research/gaia-skill-tree23Updated Oct 24, 2026

Run this skill for Phase 3 of the evidence verification pipeline — the adversarial audit step. Use when you need to check the evidence data lake for bad data before ingestion: dead links, wrong URL formats (tree/ vs blob/), subjective wording ("elite", "high-quality"), stale migration notes, benchmark catalog misuse/vendor-claim leakage, or skills whose star evidence conflicts with classified evidence level. Triggers on phrases like: "audit the data lake", "adversarial check", "ev-adversarial-audit", "check for noise in evidence", "flag bad evidence", "run the audit phase", "quality check the by-type files", or any reference to Phase 3 of the pipeline.

The English summary is being prepared. Showing the original description.

gaia-research/gaia-skill-tree23Updated Oct 24, 2026

See all skills by gaia-research

Report a problem with this skill