autoresearch

name	autoresearch
description	Autonomous iterative experimentation loop for measurable Robot SF tasks; use when the user wants an improvement loop with baseline, experiments, and keep/discard decisions.
category	research-iteration
kind	atomic
phase	analysis
requires_write	true
requires_slurm	false
requires_benchmark_artifacts	false
delegates_to	[]
output_schema	skill_run_summary.v1

Autoresearch

When to use

Use this skill when the task has measurable goals and needs an iterative experiment loop (hypothesis, baseline, test, keep/discard decisions), not a single-pass edit.

Read First

• AGENTS.md
• docs/dev_guide.md
• .specify/memory/constitution.md
• .github/copilot-instructions.md
• docs/ai/awesome_copilot_adaptation.md

Workflow

1. Define the experiment contract

   • Restate the goal, metric command, extraction method, scope, constraints, and stop condition.
   • If any input is missing, ask for it or switch to a clarifying skill.

2. Establish a baseline

   • Run the metric command on the current branch before changing code.
   • Record baseline and metadata in output/ai/autoresearch/<slug>/results.tsv (or approved location).

3. Iterate on one hypothesis at a time

   • Make the smallest useful edit.
   • Commit the experiment snapshot before running it.
   • Run the metric command and any repo-required validation gate.
   • Keep only if metric improves and constraints are satisfied.
   • Revert the change if it regresses, stalls, or becomes too complex for the gain.

4. Use repo-native validation

   • Prefer scripts/dev/pr_ready_check.sh for broad validation unless the user gave a tighter command.
   • Prefer scripts/dev/ wrappers and config-first commands from the dev guide.
   • Treat fallback or degraded benchmark outcomes as caveats, not wins.

5. Report the result

   • Summarize the baseline, best result, discarded attempts, and remaining risks.
   • Call out any assumption that stayed unresolved.

Proof and Guardrails

• Do not add dependencies or change public contracts unless explicitly allowed.
• Keep benchmark wording conservative.
• Use one concept per experiment.
• Stop when the user budget is reached or the metric stops improving.
• If benchmark evidence is invalid or non-reproducible, stop and report that status as incomplete.

Output

Summarize:

• experiment contract,
• baseline and best result,
• discarded attempts and reasons,
• unresolved assumptions and next risk.

Guardrails

• Stay within the skill scope declared in .agents/skills/skills.yaml.
• Prefer repository scripts and canonical docs before ad-hoc commands.
• Record blockers and validation gaps instead of overstating completion.