kg-autoresearch

name: kg-autoresearch description: Autonomous multi-round web research loop (자율 웹 조사/자료조사/문헌 검토). Decomposes a topic into angles, searches, fetches, extracts findings, and queues results for user approval before filing into the wiki. Governed by research-policy.yaml. Use when /kg-query --depth deep says "insufficient wiki coverage" or 새로운 주제를 체계적으로 조사할 때. trigger: /kg-autoresearch

/kg-autoresearch — Supervised Autonomous Research

Autonomy with guardrails. The loop runs unsupervised, but nothing enters the wiki without user approval.

Activate When

• User asks for "자율 웹 조사", "자료조사", "문헌 검토", "literature review"
• /kg-query --depth deep returned "insufficient wiki coverage" for the topic
• User wants systematic external research on a new topic
• User invokes /kg-autoresearch "<topic>"

Do Not Activate When

• Question answerable from existing wiki → /kg-query --depth deep first
• Topic is internal/proprietary (autoresearch uses public web)
• Wiki not bootstrapped (need /kg-init first)
• User wants to ingest a specific file they already have → /kg-ingest <file>

Preconditions

• Wiki bootstrapped (/kg-init has run)
• wiki/.schema/research-policy.yaml — auto-injected at runtime if absent (see below)

Runtime Precondition: research-policy injection

Before running any round, check if wiki/.schema/research-policy.yaml exists. If not:

1. Copy from global: cp ~/.claude/skills/kg/schema/research-policy.yaml wiki/.schema/
2. Create staging folder: mkdir -p wiki/.research-queue/{approved,rejected}
3. Log: [YYYY-MM-DD] autoresearch | injected research-policy.yaml (v1.1 upgrade) This is idempotent — if the file already exists, skip silently.

Flow

Round 0: Decompose

1. Read hot.md for current context
2. Decompose topic into 3-5 research angles
3. Present angles to user for add/remove/modify
4. Lock angles

Rounds 1-N (governed by max_rounds):

For each angle:

1. Search: WebSearch with the angle. Respect source_preferences.
2. Fetch: WebFetch each URL. Track cumulative tokens.
3. Extract: Key claims with quotes, relevance, confidence, contradictions.
4. Queue: Write to wiki/.research-queue/<round>-<angle-slug>-<N>.md with queue-specific frontmatter (NOT ontology types):

   ---
   title: [Source Title]
   queue_status: pending
   date: YYYY-MM-DD
   url: [source URL]
   angle: [research angle]
   round: N
   confidence: medium
   claims: []
   ---
   

5. Gap analysis: Under-covered angles? Emerging sub-angles?

Final: Present & Approve

1. Show summary: N sources queued, K contradictions.
2. For each queued source:
   • approve → re-fetch original URL, run /kg-ingest on fetched content. Move queue file to approved/.
   • reject → move to rejected/
   • defer → leave in queue
3. Update hot.md and log.md.

Safety

• auto_ingest: false by default.
• Token budget enforced per round AND cumulative.
• Source URLs logged for provenance.

When NOT to use

• When the question is answerable from existing wiki (/kg-query --depth deep first)
• When the topic is internal/proprietary (autoresearch uses web)
• When the wiki itself is not bootstrapped (run /kg-init first)

Output Contract

Always return:

Research topic: <topic>

Angles locked:
- <angle 1>
- <angle 2>
- <angle 3-5>

Sources queued: <N>
Contradictions found: <K>

Queue files:
- wiki/.research-queue/<round>-<angle-slug>-<N>.md (×N)

Per-source summary:
- [<title>] <url> | confidence: high|medium|low | <one-line claim>

Approval needed:
For each queued source, user chooses:
- approve → re-fetch, run /kg-ingest, move queue file to approved/
- reject  → move queue file to rejected/
- defer   → leave in queue for next session

Nothing ingested yet: yes

The last line ("Nothing ingested yet: yes") is mandatory. It is the operational guarantee that no wiki content was written without explicit user approval per source.

Examples

Example 1 — fresh topic, internal coverage insufficient

User (after /kg-query --depth deep "graph database for ontology" returned "insufficient wiki coverage"):

/kg-autoresearch "graph database choices for OWL/RDF + Python ergonomics"

Expected behavior:

1. Inject research-policy.yaml if absent.
2. Decompose into 3-5 angles (e.g., RDF/OWL native stores, property graph + RDF*, Python client maturity, license/cost).
3. Lock angles after user confirmation.
4. Run rounds: search → fetch → extract → queue.
5. Present per-source summary + ask approve/reject/defer.
6. Output result block ending with "Nothing ingested yet: yes".

Example 2 — internal-coverage check redirects user

User (without prior /kg-query):

/kg-autoresearch "WRF microphysics scheme comparison"

Expected behavior:

• Stop before research starts.
• Run a quick BM25 lookup or suggest /kg-query "WRF microphysics scheme comparison" --depth deep first.
• Only proceed to autoresearch if internal coverage is genuinely thin.
• Reason: research costs tokens; existing wiki may already answer.

Exceptions and Escalation

• Topic is internal/proprietary → stop and warn that web research is inappropriate. Suggest /kg-query or /kg-elicit instead.
• Internal coverage is sufficient → suggest /kg-query result rather than starting research.
• Sources are low quality → queue with confidence: low and recommend rejection in summary.
• Research-policy.yaml is missing AND global copy unavailable → stop; cannot run without policy.
• Token budget exceeded mid-round → stop the round, queue what was found so far, report partial result.
• Never auto-ingest. Approval gate is non-negotiable.

Quality Gates

Before final answer:

• Every queued source has URL, date, angle, confidence, ≥1 direct quote
• Contradictions across sources are explicitly listed
• No queue file lacks queue_status: pending
• Final summary explicitly states "Nothing ingested yet"
• User approval prompt is unambiguous (approve/reject/defer per source)