auto-research

Files (19)

• SKILL.md
• experimenter.md · agents/
• researcher.md · agents/
• reviewer.md · agents/
• deep-research.md · references/
• evaluation-design.md · references/
• evidence-management.md · references/
• experiment-safety.md · references/
• experimental-research.md · references/
• literature-research.md · references/
• reporting.md · references/
• research-loop.md · references/
• research-protocol.md · references/
• research-journal.mjs · scripts/
• validate-manifest.mjs · scripts/
• research-manifest.json · templates/
• journal.test.ts · tests/
• manifest.test.ts · tests/
• skill.test.ts · tests/

SKILL.md (first lines)

# Auto Research

The current conversation Agent owns the question, evidence, optional delegation and stopping. Use existing tools and history; never start a separate coordinator. Choose the smallest workflow that answers the question.

- **Literature / Research Lite**: search, read, verify sources and synthesize a report. Read only [literature research](references/literature-research.md). No execution_budget, research_id binding, Candidate, evaluator, manifest or separate state file is required. Use supplied evidence when retrieval tools are unavailable. Stay in this mode unless computation or experimental comparison is actually needed.
- **Experimental Research**: before computation or Candidate implementation, read [research protocol](references/research-protocol.md), [experiments and evaluation](references/experimental-research.md) and [safety and recovery](references/experiment-safety.md). Write a versioned evaluation.yaml. This mode requires execution_budget and authorized sandboxed Bash/Notebook; bind research_id, max_seconds and max_executions before the first execution/delegation. Reuse the binding on continuation; a used unbound budget cannot be rebound. Report missing capabilities without host fallback.

Delegation is optional. Built-in roles are auto-research/researcher, auto-research/experimenter and auto-research/reviewer; use only available authorized roles. Retrieval-only delegation uses ordinary runtime limits and does not require research_id. Confined execution is enforced by server-validated sandbox grants and the execution_scope: confined declaration, never by role names or Candidate IDs. Auto Research owns Candidate IDs/directories; candidate_id is optional correlation metadata on the grant path.

Use an append-only research-events.jsonl decision history only for multi-round experiments, cross-Candidate comparison or recovery. manifest.json is an optional, rebuildable projection. A single bounded analysis can retain its protocol, result and actual Execution/Artifact references in the report. Do not initialize empty indexes for literature work. Parent appends decisions with stable event IDs using the Skill journal helper; never edit the projection as a completion ledger. Existing studies require explicit legacy import before journal-based continuation; the strict v1 index validator remains compatible. Existing Session/Execution/Artifact/Provenance records remain authoritative. Never fabricate references, invent a universal quality score or equate process success or Reviewer PASS with scientific success.

From Garhorne0813/pi-science, Apache-2.0.

More from Garhorne0813/pi-science