Skip to content

gaia-research/gaia-skill-tree / .agents/skills/gaia-steward-full

gaia-steward-full

Verified

Run Gaia Steward end to end: observe every routine's detectors, triage debt by authority, land Class A repairs on the integration branch, drive a bounded batch of Class B dispatches through independent verifiers, re-observe, and leave a quiet founder desk or a precise list of decisions. Use for "/gaia-steward-full", "run the steward", "full maintenance pass", "what does the repo need", "resume the steward run", or a scheduled maintenance wake. Thin entrypoint: each phase is an atomic gaia-steward-* skill.

gaia-steward-full

The root of Gaia Steward Full. It owns no logic of its own: the deterministic runner (gaia steward full) decides every state transition, and the atomic skills below each do one job. This file maps a mode to a command and an action to a skill, and states the rules that hold across all of them.

Modes

InvocationCommandSkill
/gaia-steward-fullgaia steward full run --jsonthis loop
--observegaia steward full run --observe --jsongaia-steward-observe
--triagegaia steward full run --triage --jsongaia-steward-triage
--routine Ngaia steward full run --triage --routine N --jsongaia-steward-triage
--debt IDgaia steward full run --triage --debt ID --jsongaia-steward-triage
--verifygaia steward full run --verify --jsongaia-steward-verify
--retrogaia steward full run --retro --jsongaia-steward-retro
--statusgaia steward full status --jsongaia-steward-desk
--resume RUNgaia steward full run --resume RUN --jsonthis loop
--afkadd --afk to any runthis loop

The runner itself never prompts in any mode; --afk is recorded on the run for audit and tells the driver to queue founder decisions instead of asking. An awaiting-action run keeps the run lease until its TTL, so a second invocation meanwhile exits 4 (observing-existing).

Scheduled runs add --trigger scheduled. Class A landing additionally needs --integration-branch dev/<name> --push (or GAIA_STEWARD_INTEGRATION_BRANCH); the runner refuses main, master and the remote default branch.

--routine takes 1 to 19. Routine 18 is GitHub issue hygiene (issue-hygiene); routine 19 is routine pull-request maintenance (pr-maintenance, skill gaia-steward-pr-maintenance). Issue-write authority is founder-ratified policy (ruling #10, A1-A5) but not rolled out: issueHygiene.writesEnabled is false, so no closure actions are emitted, and there is no live closure dry run, pilot or burn-down. Closures need ratification AND write opt-in AND an attributable Steward owner and worker on every receipt; changing the kill switch alone grants no authority.

Step 0: refresh the censuses

Before the first run of any mode that observes or triages, refresh the GitHub issue census that routine 18 reads:

python scripts/steward_issue_census.py --repo gaia-research/gaia-skill-tree --out .gaia/steward/inputs/github-issues.json

Exit 0 is complete. Exit 3 is incomplete: the file is written with its completeness flags false, and routine 18 reports not evaluated. Any other exit: delete the file, so the sensor reports its input missing. The unattended driver python scripts/steward_drive.py runs this step itself; --no-census skips it. A census older than the policy's maxCensusAgeSeconds is never used.

Routine 19 reads its own census the same way: python scripts/steward_pr_census.py (read-only GETs; exit 3 means incomplete, so the sensor reports coverage unknown).

The driver loop

The runner makes zero model calls. Whatever needs a model or a forge comes back as a pending action. Drive it like this, and nothing else:

  1. Run the mode's command. Read status, nextAction and actions.

  2. For each pending action, hand it to exactly one owner:

    Action typeOwner
    agent.builda fresh builder agent following gaia-steward-lane, given only params.promptPath
    agent.verifya different, fresh agent following gaia-steward-verify
    forge.open_prgaia-steward-land (draft PR to the integration branch)
    forge.close_issuegaia-steward-issue-hygiene (one proven closure, performed by scripts/steward_issue_close.py)

| forge.merge_pr, forge.close_pr | gaia-steward-pr-maintenance (shared preparation, then explicit execute --operation-id; unattended driver refuses) | | lead.integrate | the orchestrator, merging into the integration branch only |

  1. Record each outcome: gaia steward full record RUN STEP --input outcome.json.
  2. Resume: gaia steward full run --resume RUN --json. Repeat until the run is completed or blocked.
  3. Render the desk (gaia-steward-desk) and stop.

The runner tracks every step in a durable ledger (refs/heads/steward/ledger, or a local file when no remote answers). Resuming never repeats a recorded step, and a second concurrent invocation loses the lease and refuses cleanly. Only the Lead writes the shared ledger; --ledger git is unchanged and is never replaced by local state to get a green run.

Cloud agents are ephemeral and steward/ledger does not keep their receipts or proof: run Class B with --evidence git and follow cloud-custody.md. Persistence fails closed (BLOCKED: provenance unavailable); never re-dispatch.

Only Routine 19 critical operation identity, immutable proof, global slot/accounting and outcomes use its separately provisioned canonical coordination ref; ordinary run state never moves there. Local run expiry/action supersession cannot refund or cancel that shared sender. Every success-shaped PR report (including already-*) binds shared head/evidence; unknown stays recoverable. Missing coordination/current authority/clock config refuses execution. Full-run completion does not settle an outstanding critical operation.

Agents, models and effort

  • Builders: params.allowedModels (Haiku or Sonnet). Verifier: Sonnet only, and never an agent listed in params.forbiddenAgentIds. Routine 19 dispatches are different: the active environment's roster profile (python scripts/steward_pr.py roster …) chooses agents by capability; no model name is canon, and the lane's maxInFlight and maxDispatchesPerRun still cap them.
  • Every agent invocation names an explicit thinking effort. default and inherit are refused by record.
  • Parallel agent.build actions from one run may be handed to builder subagents in one batch (founder decision steward-class-b-throughput-2026-11-06). Each builder gets only its own params.promptPath; any other fan-out needs the orchestrator's approval.
  • No other models are authorized.

Hard limits (enforced in code, repeated here)

  • Never merge into main, push to main, force-push, tag, release or publish. The one exception is the ratified, narrowly scoped route forge.merge_pr (routine 19): a registered producer's pull request, one exact head SHA, an independent verification, GitHub's own branch protection, a merge commit, never a squash. It is never a general permission and is revoked by the kill switch.
  • Class B: maxInFlight 8 (the policy value; the code accepts 1 to 8), maxDispatchesPerRun 8 (1 up to maxInFlight; one run emits up to the lesser of it and the free lane slots, as parallel agent.build actions over distinct debts), maxAttempts 2, cooldown 3600 s. Founder decision steward-class-b-throughput-2026-11-06.
  • Closures (forge.close_issue) need GH_TOKEN or GITHUB_TOKEN. Without one they are reported under the driver's unperformed list with the reason, and nothing is written. A closure the executor refuses (exit 2) is recorded as a refused outcome with its reason code (B8 H-3a): the step settles, the pair is emitted again against a later census, and a pair refused three times becomes one founder question. It is never re-sent within the same pass. Need more throughput? File a founder-reviewable amendment (founder/steward/decisions/); never loop the CLI to get around the cap.
  • Unknown coverage blocks Class A for the run. A not-evaluated sensor is never healthy.
  • Class C goes to the desk. Founder availability is not permission, and --afk only queues decisions while A/B work continues.
  • The builder never verifies its own work; the machine can reject or escalate but never accept.

Reused skills

Steward finds and bounds the work; existing skills do the specialised parts.

Debt or situationSkill
any Class B dispatchgaia-steward-lane (the builder's contract)
trust_calibration_drift, trust_stamp_stalegaia-tm-inspect first: verify the effective-evidence context before any calibration
timeline_causality_driftgaia-trace-timeline to reconstruct the order; the fix stays human-gated
issue_closure_ready, issue_suite_review, issue_founder_decision (routine 18)gaia-steward-issue-hygiene (A closures), gaia-steward-lane (B suite review), gaia-steward-desk (C questions)
generated_projection_driftgaia-docs-sync
ci_failure, CI churnci-churn
routine_pr_* (routine 19)gaia-steward-pr-maintenance; Class C items go to gaia-steward-desk
escalated work too large for one dispatchgaia-triage, then gaia-issue-resolver or feature-pipeline on its own PR

Stop conditions

Stop and report when the run is completed or blocked, a lease is lost, a forge closure comes back blocked (the driver halts; no further closures this run), or an action asks for anything outside the table above.

Where the rest lives

  • Architecture contract: founder/handovers/2026-10-20-steward-full-architecture.md; issue hygiene: founder/handovers/2026-10-21-steward-issue-hygiene-architecture.md
  • Coverage catalog and checker: founder/steward/routine-coverage.json, scripts/check_steward_coverage.py; routine 18 guide: founder/steward/routines/issue-hygiene.md
  • Policy: founder/steward/POLICY.yaml; decisions: founder/steward/decisions/
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: 472kUpdated 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