auto-research

Auto-Research

Turn a broad or long-running request into an Anda Bot loop with a file-backed ledger. The ledger is the source of truth: every iteration reads it before work and writes it after work.

Use this skill to run the work, not to propose a framework. If the user invoked the skill with a real objective, execute until a completion criterion is met, a safety approval is required, or the ledger proves the task is structurally blocked.

Operating Rules

• Ledger first: persist state in files, not conversation memory. Brain memory may provide context, but it does not replace the ledger.
• Ready means execute: once setup is sufficient, start the next work packet, check, retry, or monitor without asking for routine confirmation.
• Autonomy with safety: decide locally unless the next action needs missing credentials, irreversible destructive changes, paid/external side effects, or user-owned policy approval.
• Fresh work packets: each iteration receives only the task spec, recent ledger state, tried directions, and a checkable completion criterion.
• Direction diversity: after a stall, change a structural constraint, source set, decomposition, tool, or validation method; do not tune only wording.
• Separation: workers gather evidence and run checks; the orchestrator judges progress and pivots; a patrol only checks liveness, restarts, or nudges.

Start Or Continue

1. Choose the task directory. Prefer a user-specified directory. Otherwise use auto-research/<slug>/ in the active Anda workspace; for repository work, keep state outside the repo unless the user asked for repo-local artifacts.

2. Create or read:

   state/task_spec.md
   state/progress.json
   state/findings.jsonl
   state/directions.json
   state/checks.md
   logs/events.jsonl
   

3. Normalize task_spec.md to include objective, boundaries, success criteria, allowed side effects, and validation commands or evidence requirements.

4. Ensure progress.json has iteration, status, stale_count, last_seen_ms, and last_finding_count.