Skip to content

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

gaia-curate-dynamic

Verified

Use for broad, sharded discovery-only Gaia curation; for one source page use gaia-curate. A Sol/Terra-class orchestrator preflights capacity, shards safe volume work to configurable Luna workers, converges risky cases, and stops at human L4 review.

Gaia Curate Dynamic

Read ../gaia-curate/CURATION-CORE.md. Every worker returns one schema-valid discovery-packet-v2; this extension may orchestrate packets but may not relax the core lifecycle or cross its L4 stop. V1 remains readable only for compatibility and is not normative output.

Roles

RoleDefault tierBounded responsibility
orchestratorSol preferred; Terra acceptablepreflight, source shards, budgets, merge, arbitration, L4 synthesis
harvesterLuna Lightfetch/parse/normalize a bounded source shard
mapperLuna Lightchoose among at most three supplied generic options or defer
proposer/refuterLuna High or equivalentisolated adversarial review of risky candidates only
repairLuna High or equivalentone schema-repair attempt against frozen input
assemblerdeterministic firstvalidate, dedupe, merge, persist, render L4 output

Model names are runtime configuration, not part of this skill. Do not silently substitute an unapproved model tier.

Mandatory runtime preflight

Discovery is read-only. Before dispatch, resolve the active harness through explicit overrides such as CLAUDE_BIN, CODEX_BIN, or HERMES_BIN, then PATH. Do not assume a binary location and never edit user configuration.

Persist preflight.json with:

{
  "harness": "claude-code|codex|hermes|other",
  "version": "...",
  "spawnMode": "subagent|process|sequential",
  "requestedConcurrency": 3,
  "observedConcurrency": 1,
  "structuredOutput": true,
  "usageReporting": "native|external|null",
  "budgetEnforcement": "native|orchestrator|null",
  "models": {"orchestrator": "...", "lunaLight": "...", "lunaHigh": "..."}
}

A config value is advisory. Establish capacity with a read-only canary: one Luna Light task, then a two-worker wave. Increase only after both succeed. Effective concurrency is the minimum of requested capacity, observed harness/provider capacity, remaining shards, and remaining budget. On 429/overload, reduce concurrency; do not increase retries. If routing or capacity cannot be established, continue sequentially or stop with the exact missing operator action.

Harness recipes, when configured:

# Claude Code
"$CLAUDE_BIN" -p --model "$LUNA_LIGHT_MODEL" --output-format json < "$PROMPT_FILE"

# Codex; read-only sandbox, explicit working directory
"$CODEX_BIN" exec -m "$LUNA_LIGHT_MODEL" -s read-only -C "$WORKTREE" - < "$PROMPT_FILE"

# Hermes one-shot process; children spawned with delegate_task inherit the parent model
"$HERMES_BIN" chat -Q -q "$BOUNDED_PROMPT" -m "$LUNA_LIGHT_MODEL" \
  --provider "$HERMES_PROVIDER" --max-turns 12 --source tool

Use only the recipe supported by the active harness. Never expose prompts, transcripts, secrets, or auth files in the ledger.

Sharding and convergence

  1. Shard by source or repository affinity, with a bounded source list and output ceiling per harvester. Do not create one process per raw lead.
  2. Validate and exact-dedupe mechanically before semantic or adversarial work.
  3. Send each mapper one candidate plus at most three generic options. Invalid or ambiguous output becomes DEFER; do not infer intent.
  4. Skip adversarial calls for exact duplicates and straightforward existing-generic mappings.
  5. Run isolated proposer/refuter passes only for NEW_GENERIC, fusion proposals, attribution conflicts, and unresolved semantic duplicates. Freeze identical input and hide each response from the other.
  6. Sol/Terra arbitrates disagreements only. Conflicting mappings become DEFER, never a silent merge.
  7. Reconcile against the inventory frozen before dispatch. Use candidate ID, exact named identity and generic ID as separate keys. Verify task/attempt/input/output and actual packet digests; missing or failed candidates remain explicit deferred rows. Require exactly one final outcome per expected candidate. Shared-generic mappings never collapse named rows; conflicting generic definitions defer deterministically, not by completion order.
  8. Present the entire accounted-for inventory, including deferrals, rejections and proven duplicates, at one consolidated L4 checkpoint and stop. Review-ready is not human-approved; local reviewer metadata never grants authority.

Retry and budget policy

FailureAction
429, overload, 5xx, transport timeoutretry at most twice with backoff; reduce concurrency
capacity rejectionsplit/downshift the wave; do not count as candidate failure
invalid structured outputone Luna High repair against the same frozen input
validator or generic-reference failureno blind retry; defer with exact validator output
unavailable or contradictory sourcedefer
budget exhaustedstop dispatch, preserve partials, mark run blocked

Never rerun a successful shard.

Resumable state and cost log

Use generated-output/curate-discovery/<run-id>/. Write state atomically and append one usage.jsonl event per attempt:

{
  "runId": "...", "taskId": "...", "shardId": "...", "attempt": 1,
  "role": "harvester", "tier": "luna-light", "harness": "...",
  "modelResolved": "...", "startedAt": "...", "finishedAt": "...",
  "status": "succeeded", "inputTokens": null, "outputTokens": null,
  "cacheReadTokens": null, "estimatedCostUsd": null, "pricingSource": null,
  "failureClass": null, "inputDigest": "sha256:...", "outputDigest": "sha256:..."
}

Use native usage when exposed, or an approved external cost tool. Store null when unavailable; never invent token or cost values. Resume only failed/incomplete tasks after revalidating the core contract digest and generic snapshot digest.

Discovery mode has no mutation or ship stage. Runtime ledgers and partial worker output stay under generated-output/curate-discovery/<run-id>/, but the deterministic assembler copies each final schema-valid review packet to the Core-owned canonical review path registry-for-review/discovery-packets/ before presenting L4. It must not create intake batches, branches, commits, issues, or PRs. Only the separate explicit post-L4 ratification seam via gaia dev ratify and gaia push in Core may create intake artifacts.

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