autoresearch

name: autoresearchdescription: "Autonomous iteration loop: modify, verify, keep/discard against any metric"version: 2.3.0

Autoresearch — Autonomous Goal-directed Iteration

Safety Invariants (all subcommands)

• Never push, publish, or deploy without explicit user approval.
• Bounded by default. Override with Iterations: unlimited.
• All results logged to autoresearch/{subcommand}-{YYMMDD}-{HHMM}/ directory.
• Chain handoff via handoff.json. Evals reads *-results.tsv.

Dispatch (bare $autoresearch)

Parse the invocation in this order:

Print a banner on every invocation: [autoresearch] mode: classic | orchestrator | wizard.

Subcommands

Subcommand routing

To run a subcommand yourself (an orchestrator hop or a --chain target), Read its protocol file and follow it: <name>.md (relative to this skill's directory), or autoresearch.md for the classic loop. The commands are user-typed only, so the model loads them from here instead of invoking them.

Universal Flags

Getting Unstuck: Fork-and-Flush

Metric loops run fork-and-flush rounds on a fixed schedule: at 33% and 66% of the budget, plus early when 5 iterations pass with no keep. A round forks K fresh-context branches from the best commit so far, runs each for a small budget, and keeps the winner. A fresh context means the files and results log only, with none of the leader's reasoning. Half the branches (by default) are random and pick their own ideas; the other half are guided, each given one distinct experiment by a stateless Advisor that reads the results log. Protocol: references/fork-and-flush.md (arXiv:2610.07447, arXiv:2610.07625).

Orchestrator

Activated when a plain-language goal is given without Metric:/Verify:. Classifies the goal into a Goal archetype — see references/orchestrator-routing.md for the archetype table and router decision table.

Resolve every scripts/... path below relative to this installed skill directory, never relative to the caller's working directory.

For performance, latency, throughput, runtime, CPU, memory, allocations, binary size, power, energy, bandwidth, or compute-cost goals, read references/behavior-preserving-optimization.md. The working metric ranks candidates; independent behavior evidence decides whether a candidate survives.

Two modes based on archetype:

• Orchestration loop — predicate-bearing archetypes (ship-ready, optimize-metric, fix-broken, harden, build-feature, explore). Goal has a mechanical Success predicate; the loop runs until that predicate is met.
• Single-pass dispatch — subjective/terminal archetypes (document, what-to-build, decide-design). Routes once to the fitting subcommand (learn / improve / reason), lets it self-terminate, then reports. No loop, no Plateau, no ship gate.

Orchestration Loop Steps

Backed by scripts/orchestrate.sh (deterministic seam — all routing logic lives there). Subcommands exposed: classify, next-hop, units, plateau, screen-cmd, verdict, validate-state, screen-state-predicate.

1. Classify — scripts/orchestrate.sh classify "<goal>" → archetype label + mode.
2. Derive predicate — reuse plan logic to produce a concrete Success predicate: exact shell command + expected output. For optimize-metric, run the full plan/wizard derivation internally.
3. Confirm — ONE request_user_input showing: archetype, mode, concrete predicate (command + expected output), terminal choice (stop-at-verified vs proceed-to-ship). Misclassifications are caught here, not mid-run.
4. Round-0 dry-run — prove the predicate command runs and returns a value; safety-screen every derived command via screen-cmd; print projected cycle budget. Stop here if --dry-run.
5. Loop until predicate satisfied: a. Assess state via cheap signals (last handoff.json, regression verdict, error count) + affected-test verify. b. scripts/orchestrate.sh next-hop orchestrator-state.json → next subcommand. c. Run subcommand via Subcommand routing (its own bounded inner loop). d. Record per-hop outcome ∈ {progressed, no-op, failed, blocked}. e. Fold hop's handoff.json into orchestrator-state.json. f. scripts/orchestrate.sh units → recompute Units remaining and append the number or unknown to the state's units_remaining array.
6. Stop conditions (checked after each hop):
   • Predicate met → ship gate (only if ship is in the pipeline) else CONVERGED.
   • Cycles > ceiling (default 50, override --max-cycles N) → stop + report CEILING. Checked first: a reached ceiling is never extended by fork rounds.
   • scripts/orchestrate.sh plateau orchestrator-state.json → exit 0 (flat): unless --no-fork or fork_rounds_since_progress is already 2, run one fork-and-flush round from the incumbent (references/fork-and-flush.md), add 1 to fork_rounds_since_progress, and continue; otherwise stop + report PLATEAU. Exit 1: continue. Exit 3: stop + report BLOCKED.
   • Hop outcome blocked/failed with no alternative route → checkpoint + stop + report BLOCKED.

Orchestrator State

orchestrator-state.json — orchestrator-owned, additive. Tracks: goal, archetype, predicate, terminal-choice, units_remaining history, cycle count, per-hop pipeline log with outcomes, current incumbent, fork_rounds_since_progress (reset to 0 whenever units improve). Each hop's handoff.json is unchanged (single-hop bridge); the orchestrator reads it and folds it in. Two clearly-owned state objects, no overlap.

Orchestrator Safety Invariants

• Never auto-approve ship/deploy/push. The orchestrator never passes --auto to ship; deploy always requires explicit user approval.
• Data-migration behind anchored DB-URL allowlist. Reuses regression's allowlist — host must be localhost/127.0.0.1/container hostname, or database name carries _test/_ci suffix. Bare substring match does not qualify. Anything else refused.
• screen-cmd on every derived command — run before the loop starts AND on every command read from a persisted state file on resume. Persisted commands are never trusted; resume re-screens the pinned predicate via screen-state-predicate and refuses on refuse.
• No un-screened commands mid-loop. The autonomous loop cannot introduce new shell commands that bypass screen-cmd.
• Predicate pinned, not re-derived. Round-0 writes the derived Success predicate verbatim into orchestrator-state.json; every cycle and every resume reuses that exact string so "done" is reproducible across runs.
• Validate the ledger before routing. validate-state gates orchestrator-state.json (required fields + coarse types); a malformed ledger is not trusted to route from.
• Independent verify before convergence. High-impact changes accepted on the working signal set pending_verify; next-hop routes to a verify hop (held-out / adversarial check) before DONE or ship. The verify hop never auto-approves ship.
• Unknown-units cycles excluded from Plateau counter. A cycle where units returns unknown (e.g. runner crash) is not counted as zero-progress; repeated unknown routes to BLOCKED.