Autoresearch

Use the parent for research decisions and orchestration workers for experiment execution. Measure consistently. Retain only verified improvements. Stop on explicit resource or convergence limits.

Orchestration Contract

Start a new autoresearch session in Plan mode. Also require Plan mode before materially changing an approved session's objective, primary metric or direction, benchmark or correctness commands, write scope, or hard resource, cost, regression, or convergence limits. If Plan mode is required but inactive, ask the user to switch and stop. An unchanged approved contract may resume and receive new hypothesis batches outside Plan mode.

Always invoke $orchestration for implementation. Never fall back to direct parent implementation or another handoff mechanism. Follow its host selection, plan manifest, team sizing, validation ownership, reconciliation, failure, and completion contracts.

The parent owns the session contract, evidence synthesis, hypothesis selection and ordering, and stop decisions. It may delegate read-only repository investigation when useful, but research workers return evidence rather than hypotheses or plans. Keep the parent's execution work to orchestration, integrity checks, compact result review, and selection of the next batch.

Implementation workers execute parent-supplied ordered hypothesis batches. For a sequential search, default to one worker. Pack multiple related hypotheses into one brief up to orchestration's sizing limit. Never assign one worker to each idea by default. Split only when hypotheses are independent, dependency waves require it, or one brief would exceed the size limit.

A worker may make local adjustments needed to execute an assigned hypothesis. It must report a materially different research direction instead of executing it.

Plan and Session Contract

Resolve the objective, primary metric and direction, benchmark and correctness commands, allowed/off-limits paths, run/runtime/command/cost/regression limits, convergence window, and reporting cadence before approval. Infer safe facts from the request and repository. Ask only when a missing choice changes the experiment.

Defaults: 20 runs, two hours wall time, 10 minutes per benchmark, five minutes per correctness check, no new paid API spend, and convergence after five consecutive valid runs without a new retained best. Explicit --max-runs and --max-runtime values are hard limits.

The Plan-mode response must include the resolved contract, an evidence-backed ordered initial hypothesis batch with its completion or early-stop criteria, and orchestration's required plan section and manifest. Adding, removing, or reordering hypotheses inside the approved contract is follow-on planning, not a material contract change.

Delegated Execution

Read references/worker-loop.md before constructing an implementation brief. Include its applicable instructions with the approved contract, ordered batch, exact paths and commands, current session and best-result state or first-batch status, and batch stopping criteria. Orchestration owns the remaining prompt and result fields.

The first implementation worker creates the isolation and session artifacts and records the unchanged baseline. Each worker leaves detailed measurements and logs in those artifacts and returns only the compact batch receipt required by the worker reference. The parent reconciles that receipt with the session module's JSON status. It reads raw benchmark output only when a decision or integrity check requires it. It then selects another batch or stops. Further batches under the unchanged contract remain follow-on work within the approved outcome.

Progress and Completion

Use orchestration's host-native progress surface. Send parent-authored updates only from settled evidence at the baseline, completed batch, material best change, blocker, or final stop. Do not relay per-run narration. Render the session module's exact bar, counts, metrics, budgets, and convergence facts, and name the next parent-selected batch without recording it as settled work.

Finish with ### 🏁 Autoresearch complete — <stop reason>, baseline/best/delta/confidence, status counts, kept-file tree, exact checks, worktree/branch, and remaining cleanup or integration. Keep METRIC lines, JSONL, commands, and diagnostics undecorated. A resource limit is not convergence.