Autoresearch — Bounded Multi-Round KB Expansion

You are executing an autonomous research loop against Jay's Agentic-KB. The loop combines the BACKFILL and INGEST workflows already defined in the KB's CLAUDE.md into a bounded agent cycle that terminates on saturation.

This skill does NOT introduce new infrastructure. It orchestrates existing primitives: WebSearch, WebFetch, raw/ stub creation, kb ingest-file, wiki/index.md reads, and wiki/log.md appends. Do not build a parallel toolchain. Do not modify the KB schema.

Config

Collect these before starting. Ask only for what's missing.

topic              (required)   The subject to research.
max_rounds         (default 3)  Hard cap on loop iterations.
pages_per_round    (default 5)  Max raw/ stubs created per round.
source_allowlist   (default []) Domains the user trusts. Empty = any.
source_denylist    (default []) Domains to skip.
mode               (default wiki)  'wiki' = full INGEST into wiki/.
                                   'raw-only' = stage in raw/, don't INGEST yet.
dry_run            (default no) If yes, list planned fetches but don't write.

If the user says "autoresearch X" with no other config, use defaults and tell them the config you chose in one line before starting.

Loop Contract

For round ∈ 1..max_rounds:

1. Gap detection. Read wiki/index.md and any existing pages whose title or tags intersect topic. Produce an explicit gap list: concepts referenced but not yet documented, frameworks mentioned with no framework page, [UNVERIFIED] markers, stale last_checked dates, and [[wiki-links]] that resolve to nothing. If round > 1, diff against the previous gap list.
2. Query formulation. For each gap, generate one precise search query. Do not broaden to the parent topic — that was last round. Drop any query that matches a query already run this session.
3. Fetch. Run up to pages_per_round WebSearch + WebFetch calls in parallel. Skip domains in source_denylist; require source_allowlist if non-empty. If a fetch fails, log it and continue — never retry via bash.
4. Raw stubs. For each accepted source, write to raw/framework-docs/{slug}.md (or the most fitting raw/ subdir) with the full source frontmatter: source_url, author, date_published, date_ingested, tags. One source = one raw file. Never modify existing raw/ files — create-only.
5. INGEST. Unless mode=raw-only, run the KB's INGEST workflow on each new raw file: create wiki/summaries/{slug}.md, update or create the relevant concept/pattern/framework pages, cross-link bidirectionally, update wiki/index.md, append to wiki/log.md, and add an entry to wiki/recently-added.md under today's heading.
6. Saturation check. Compute new_concepts = (concepts after round) − (concepts before round). If new_concepts == 0, HALT. Otherwise continue.
7. Round log. Append one block to wiki/log.md:

   ## autoresearch {YYYY-MM-DD HH:MM} — round N/{max_rounds}
   topic: {topic}
   queries: [...]
   raw_created: [...]
   wiki_pages_created: [...]
   wiki_pages_updated: [...]
   new_concepts: N
   contradictions: [...]

Halt Conditions (any triggers exit)

• round == max_rounds
• new_concepts == 0 in the most recent round (saturation)
• Every query in step 2 is a duplicate of one already run
• User interrupts
• Rate limit / repeated fetch failure (≥3 consecutive failures → halt, don't silently degrade)

On halt, emit a final report (see Output Contract).

Safety Rules

1. Never modify raw/ — create-only. Existing files are immutable per the KB rules.
2. Respect allowlist/denylist before every fetch.
3. Confidence honesty — new pages default to confidence: medium unless ≥2 independent sources corroborate the core claim. Downgrade to low when a single low-authority source is the only evidence.
4. Reviewed flag stays false — every page autoresearch creates or updates must have reviewed: false and an empty reviewed_date: "". Never flip to true — that is Jay's signature. Automation does not sign.
5. Counter-arguments are mandatory — every concept and synthesis page this skill creates must include a Counter-arguments & Gaps section. If the sources all agree, the section names what they do NOT show and what would change the verdict. One-sided compilation is a failure mode, not an output.
6. Contradictions — if a new source contradicts an existing wiki claim, add an inline > [!contradiction] source vs. {existing-page} block AND record it in the round log. Never silently overwrite.
7. No orphan pages — every new wiki page gets ≥1 inbound link before the round closes. If you cannot find a parent, link from the relevant MoC in wiki/mocs/ or from wiki/index.md.
8. No hallucinated sources — every citation resolves to a real fetched URL in raw/. If a claim lacks a source, mark [UNVERIFIED].
9. Hot cache discipline — do not touch wiki/hot.md. Hot-cache promotion is a separate, human-approved decision.

Output Contract — Final Report

When the loop halts, produce this exact shape (markdown, save to outputs/autoresearch-{topic-slug}-{YYYY-MM-DD}.md):

# Autoresearch: {topic}

Run: {YYYY-MM-DD HH:MM}  |  Rounds: {N}/{max_rounds}  |  Halt: {reason}

## Summary
{2-3 sentences: what was known before, what was added, what the wiki now
says that it didn't.}

## Raw sources added ({count})
- [[raw/.../slug.md]] — {title} ({domain}) — confidence: {high|medium|low}

## Wiki pages created ({count})
- [[wiki/.../slug.md]] — {type} — 1-line purpose

## Wiki pages updated ({count})
- [[wiki/.../slug.md]] — {what changed in ≤10 words}

## Contradictions surfaced ({count})
- {existing claim} vs {new source} → decision: {keep|flag|re-verify}

## Remaining gaps (not fetched — user decision needed)
- {gap} — reason not fetched (rate limit | denylist | ambiguous | low-confidence)

## Next-step suggestions
1. {concrete action with owner — Jay or a follow-up autoresearch run}

Do not embed raw source bodies in the report — link to them.

Anti-Patterns (never do these)

• Running more than max_rounds "just to be thorough" — the cap exists so the loop terminates deterministically.
• Ingesting a source that matches an existing raw file slug — collisions mean stop and rename, not overwrite.
• Touching files outside raw/, wiki/, and outputs/.
• Calling kb compile or /graphify inside the loop — those are post-run steps the user invokes explicitly.
• Emitting the final report before writing it to outputs/ — save first, then print the path.

Invocation Example

User: /autoresearch supervisor-worker multi-agent patterns
       max_rounds=2 pages_per_round=4 source_allowlist=[arxiv.org,anthropic.com]

Agent: Starting autoresearch.
       topic=supervisor-worker multi-agent patterns
       max_rounds=2 pages_per_round=4
       allowlist=[arxiv.org, anthropic.com]
       mode=wiki dry_run=no

       — Round 1 —
       gap list: [3 items]  queries: [...]  raw added: 4  wiki updated: 7
       new_concepts: 5 — continue

       — Round 2 —
       gap list: [2 items]  queries: [...]  raw added: 2  wiki updated: 3
       new_concepts: 0 — HALT (saturation)

       Report → outputs/autoresearch-supervisor-worker-2026-04-17.md

Activation Confirmation

When this skill is invoked, respond with:

Autoresearch active. I'll run a bounded loop against the KB on {topic} using max_rounds={N} and pages_per_round={M}. I will stop on saturation or the cap, whichever comes first. Save the final report to outputs/. Proceed?

Then wait for confirmation before the first WebSearch call.