Codex Autoresearch

This is the one skill surface and the only Codex-facing skill. Do not route users to old subskills, slash commands, or separate dashboard/finalizer skills.

Default state machine:

setup -> doctor -> next -> log -> state -> finalize-preview

The job: make one measured improvement loop trustworthy enough that a human can follow it and a future session can resume it.

For Codex

• Use the short command path unless the session is ambiguous or blocked: setup, doctor, next, log, state, then finalize-preview.
• For qualitative or deep-research improvement loops, start with research-start --cwd <project> --slug <slug> --goal "<goal>". It creates the scratchpad, configures quality_gap, validates the command, and records the first baseline as measure unless --no-baseline-log is passed.
• Use advanced diagnostics only when needed. Check node scripts/autoresearch.mjs --help --all from the package root before naming a less common command.
• Use new-segment when the active segment is maxed, stale, phase-changing, or no longer comparable.
• Prefer CLI JSON and durable session state over chat memory: autoresearch.md, autoresearch.jsonl, autoresearch.ideas.md, active last-run/progress snapshots under .git/autoresearch/ in Git repos, fallback autoresearch.last-run.json / autoresearch.progress.json outside Git, and autoresearch.research/<slug>/.
• Keep every packet decision recoverable through METRIC name=value, packet evidence, ASI, continuation data, promotion labels, and the ledger.
• Before another packet, read recommend-next --compact or state --compact; obey blockers. Compact-state field names: docs/concepts.md#state-fields.
• benchmark-lint must prove the primary METRIC contract before product packets are trusted.
• Configure commitPaths or pass --commit-paths for kept results in Git repos.

For the user

• Plain-language prompts work: "/goal @Codex Autoresearch improve this repo."
• Ask only for essentials that materially change setup: goal, benchmark, primary metric, direction, scope, or correctness checks.
• For shippable, product, or final requests, identify product claims before setup. Retrieval, search, ranking, lazy behavior, accessibility, safety, or performance work needs a quality constraint or checks path before promotion.
• Stay on the CLI happy path unless setup is ambiguous, the user asks for the dashboard, packet freshness needs a browser readout, or the canonical action is blocked.
• Report the story: what was tried, what the metric means, the keep/discard/measure/crash/checks decision, the next move, blockers, optional dashboard URL, and verification.

Documentation awareness

Use docs only as needed; do not load everything by default.

• Start/resume or normal operation: docs/start.md, docs/operate.md, and references/loop-operations.md.
• Dashboard, trust, drift, protected paths, unsafe commands, and redaction: docs/trust.md, docs/architecture.md, and references/dashboard-trust.md.
• Deep research, quality gaps, fanout, finalization, or subagent handoffs: docs/finish.md, docs/workflows.md, and references/research-finalize.md.
• Troubleshooting: docs/troubleshooting.md.
• Control-plane failures or cross-surface disagreement: docs/control-plane.md.

Start or resume

1. Identify the owning repo or child package before Git, installs, tests, builds, or autoresearch commands.
2. Check Git status and work around unrelated dirty files.
3. If this repo is the target, use the repo-local plugin. From the wrapper root: node plugins/codex-autoresearch/scripts/autoresearch.mjs .... From the package root: node scripts/autoresearch.mjs ....
4. Read autoresearch.md, autoresearch.jsonl, and autoresearch.ideas.md when present.
5. Use setup-plan for read-only setup guidance when essentials are unclear. Use setup only when essentials are known and files should be created.
6. Run doctor --cwd <project> --check-benchmark --explain before the first trusted packet or any drift-sensitive metric.
7. Use the happy path first: setup -> doctor -> next -> log -> state -> finalize-preview.
8. Before another packet, read recommend-next --compact or state --compact; obey blockers; open detailed diagnostics only when the canonical action is blocked, stale, or unclear.
9. Use state --report for a terminal-first report.text. Governance fields are listed in docs/concepts.md#state-fields.
10. Run serve --cwd <project>, verify liveness, and provide the live dashboard URL only when the user asks, the browser readout matters, or CLI state is not enough.
11. For retrieval/search/ranking/performance work, require quality constraints before promotion.
12. Treat optional task_manifest packet evidence as audit data; quarantine malformed manifests and path escapes without invalidating unrelated metric evidence.
13. Treat benchmark-shaped fixes as diagnostic until proven otherwise. Row-specific detector or citation work is diagnostic repair until holdout, repeat, breadth, or promotion gate proves the broader claim.
14. If session-forensics imports benchmark-overfit or row-specific steering feedback, treat the decision capsule as a trust blocker.
15. Treat runtime freshness as unavailable unless installed runtime version and built-entrypoint fingerprint can be inspected and matched.

Happy-path CLI from plugins/codex-autoresearch:

node scripts/autoresearch.mjs setup --cwd <project> --name "<session>" --metric-name <metric> --direction lower --benchmark-command "<command>"
node scripts/autoresearch.mjs doctor --cwd <project> --check-benchmark --explain
node scripts/autoresearch.mjs next --cwd <project>
node scripts/autoresearch.mjs log --cwd <project> --from-last --status measure --description "Baseline measurement"
node scripts/autoresearch.mjs state --cwd <project> --report
node scripts/autoresearch.mjs finalize-preview --cwd <project>

Active loop contract

After next, log the packet. After log, read the returned continuation object.

• Only next writes a reusable last-run packet. run remains a raw benchmark probe.
• Use log --from-last instead of retyping parsed metrics.
• keep, ordinary discard, and measure require a finite primary metric.
• Use measure for non-promotional evidence: baselines, no-change probes, environment checks, and diagnostics.
• crash and checks_failed can be logged without inventing sentinel metrics.
• Treat review_required metrics as provisional until ASI acknowledges the review outcome.
• If autoresearch.config.json contains fixedControl, treat the named artifact as control truth. Do not rerun commands matching forbiddenCommandPatterns unless the user explicitly accepts --allow-fixed-control-rerun; prefer reuseCommandHint.
• If run numbers duplicate, segments look stale, or manual log entries were edited, run ledger-doctor --cwd <project> --json before another packet. Use ledger-doctor --repair --yes only after reviewing the JSON health summary; after repair, verify the returned backupPath.
• Read parsed metrics and promotion readiness separately. New keeps default to exploratory unless repeat, holdout, breadth, or explicit promotion metadata make the evidence promotable.
• The loop contract is the authority for whether to spend another packet. sourceCleanliness.blocks.nextPacket=false only says source dirtiness is not the blocker.
• Control-plane contracts are packet brakes too: goal mismatches, missing scoped approvals, stale process residue, unsupported broad claims, and unsafe finalization runways outrank another packet.
• When the metric improves because the benchmark was steered toward known answers, say so.
• If continuation.shouldContinue is true, choose the next hypothesis from ASI, experiment memory, autoresearch.ideas.md, or dashboard lane guidance.
• If continuation.forbidFinalAnswer is true, continue with progress updates instead of returning a final answer.
• Respect packet and wall-clock budgets.
• If correctness checks fail, run checks-inspect before deciding.
• Stop when the user interrupts, the limit or budget is reached, benchmark/checks are blocked, cleanup would be unsafe, a fresh segment is needed, or the goal is genuinely exhausted.

Codex-only Goal completion

• Use completionAudit before a parent agent calls update_goal(status="complete").
• Do not complete a parent Codex Goal while the continuation says the loop is still active.
• Keep Goal state in Codex; Autoresearch only provides codex-goal-brief and completion-audit evidence.

CLI fallback:

node scripts/autoresearch.mjs next --cwd <project> --compact
node scripts/autoresearch.mjs log --cwd <project> --from-last --status keep --description "Describe the kept change"
node scripts/autoresearch.mjs state --cwd <project> --report
node scripts/autoresearch.mjs state --cwd <project> --compact

Dashboard

Use the served dashboard when a live readout is useful:

• Use scripts/autoresearch.mjs serve --cwd <project>.
• Share the served http://127.0.0.1:<port>/ URL by default.
• Static exports are read-only snapshots; serve a fresh dashboard when packet freshness matters.
• Readout only. Use the CLI to do the work.
• The live server accepts only loopback Host headers, sends defensive headers, and keeps the raw ledger endpoint disabled unless --debug-ledger is explicitly used.

Deep research loops

Use a deep-research loop for broad, qualitative, product-study, UX, architecture, or documentation prompts. Study, accept gaps, measure quality_gap, close credible candidates, then start a fresh round when the question is still alive.

1. Start with research-start --cwd <project> --slug <slug> --goal "<goal>". It seeds autoresearch.research/<slug>/, configures the quality_gap benchmark, validates the command, records the first baseline as measure, and prints the resume commands. Use --no-baseline-log only when that first baseline should not be recorded.
2. Keep sources dated and claim-specific in autoresearch.research/<slug>/sources.md.
3. Write the judgment pass in synthesis.md: filter hallucinations, separate evidence from inference.
4. Turn accepted findings into quality-gaps.md.
5. Measure with quality-gap --cwd <project> --research-slug <slug> --list.
6. Preview candidates with gap-candidates; apply only credible high-impact gaps.
7. Log implementation or rejection with ASI.
8. Start a fresh round before claiming there are no more high-impact gaps.

quality_gap=0 only means the accepted checklist for the current round is closed. Read freshRoundSuggested, researchIntegrity, sourceCleanliness, finalization readiness, and plateau reason fields before deciding next steps.

For crashed or timed-out packets with artifact rows, use partial-results --from-last before rerunning expensive work.

Finalize

Use finalization when noisy loop history has useful kept commits.

1. Run finalize-preview --cwd <project> before branch creation.
2. Keep only accepted/current status: "keep" evidence.
3. Compare product claim coverage against accepted evidence.
4. If coverage is missing, report experimental status. Use "Experimental review branch only: product-grade proof is missing."
5. Treat previews and plans as read-only.
6. Review dirty tree, stale plan, overlap, semantic safety, and excluded-file warnings.
7. Session artifacts are excluded by default. Use --include-session-artifacts only when the reviewer explicitly wants them.
8. When state reports current-tree-finalization, run finalize-current-tree --cwd <project> --exclude-session-artifacts. Do not substitute generic finalize-preview as the primary command.
9. Ask before creating branches unless the user already approved finalization.
10. Runway order: preview, approve, create review branches, verify, merge into trunk, verify the merge, cleanup.
11. Do not suggest branch cleanup until merge verification has succeeded.
12. Classify existing review branches before reuse.
13. Report created review branches, files, metric improvement, claim coverage, verification, runway status, and remaining risk.

Subagent handoffs

When Codex uses subagents to work on Autoresearch itself:

• Each lane states scope, evidence source, decision, handoff artifact, and tests.
• No nested subagents.
• Do not run overlapping write lanes. Split by ownership first, then merge through one parent context.
• Reviewers should check the decision-envelope contract, packet freshness, dashboard read-only behavior, finalization artifact policy, and docs/changelog sync.

Verification

Use the narrowest relevant check while iterating. Before claiming plugin work is done, run from plugins/codex-autoresearch:

npm run check

Targeted checks:

npm test
node scripts/autoresearch.mjs --help
node scripts/autoresearch.mjs doctor --cwd . --check-benchmark --explain
node scripts/autoresearch.mjs benchmark-lint --cwd .
node scripts/autoresearch.mjs checks-inspect --cwd . --command "npm test"
git diff --check