Wiki Research Loop Agent Skill

Wiki Research Loop

Driver that turns a wiki into an auto-grown knowledge base. Layers on top of wiki-builder and wiki-query.

Loop semantics

seed-queue (pending) → next-seed
  → fetch sources via plugins (web | arxiv | github)
  → extract claims
  → dedupe vs index (FTS5; later vector via 3.3.2)
  → compile new page or amend existing
  → upsert page (auto-FTS-index)
  → enqueue follow-up seeds (max-depth gate)
  → mark seed done
  → if budget OR convergence OR kill-switch → halt

Halt conditions (any one trips)

• budget_usd exceeded (loop tracks per-fetcher cost estimate)
• max_pages_per_run written
• max_depth reached on every active branch
• 3 consecutive pages add < 5 % new claims (convergence)
• File ~/.pro-workflow/STOP exists (operator kill-switch)
• wiki.config.md auto_research.enabled: false
• Wiki private: true AND any non-local fetcher selected

Commands

node $SKILL_ROOT/scripts/research-loop.js run <slug> [--max-pages N] [--max-depth N] [--budget-usd 0.50] [--fetchers web,arxiv,github]
node $SKILL_ROOT/scripts/research-loop.js seed <slug> "<query>" [--depth 0] [--parent-id N]
node $SKILL_ROOT/scripts/research-loop.js seeds <slug> [--status pending|active|done|failed]
node $SKILL_ROOT/scripts/research-loop.js cancel <slug>
node $SKILL_ROOT/scripts/research-loop.js status

CLI flags override wiki.config.md for one run only.

Source fetchers

Pluggable. Each lives at scripts/source-fetchers/<name>.js. Interface:

module.exports = {
  name: 'web',
  match: (q) => true,                       // is this fetcher useful?
  estimateCost: (q) => ({ usd: 0, tokens: 0 }),
  fetch: async (q, opts) => [               // returns RawDoc[]
    { url, title, content, fetched_at }
  ]
};

Built-in:

• web.js — Fetches via the user's available WebFetch tool through a stdin/stdout shim. Treats result as plain text/markdown.
• arxiv.js — https://export.arxiv.org/api/query (free, public, no key). Returns abstract + metadata.
• github.js — https://api.github.com/search/repositories + README pull (uses GH_TOKEN if set, otherwise unauthenticated rate limit).

Drop a new file in ~/.pro-workflow/fetchers/<name>.js to add a custom fetcher. Loaded at startup if present.

Budget enforcement

Pre-iteration: sum estimateCost across selected fetchers. If projected cumulative cost would exceed budget_usd, halt.

Post-iteration: track tokens used by the LLM compile step (Anthropic/OpenAI passthrough). Hard-kill on overrun.

Per-fetcher overrides via env: WIKI_LOOP_BUDGET_USD, WIKI_LOOP_MAX_PAGES, WIKI_LOOP_MAX_DEPTH.

Seed queue

SQLite-backed via wiki_seeds table:

| field | meaning |

|-------|---------|

| query | natural-language seed |

| status | pending → active → done\|failed |

| parent_id | seed that produced this one |

| depth | BFS depth from root |

Loop pops by (depth ASC, created_at ASC) so it explores breadth-first.

Convergence detection

After each compiled page, compute Jaccard overlap of claim-text tokens vs the prior 3 pages. If < 5 % novel content for 3 consecutive pages, halt and report converged.

Kill switch

touch ~/.pro-workflow/STOP

Loop checks per-iteration and halts gracefully. Remove file to resume next run.

Privacy guard

If wiki.config.md has private: true, the loop refuses any non-local fetcher and emits a warning. Only raw/ ingestion via manual seeds is allowed.

Reactive trigger (Phase 3.3.4)

scripts/file-watcher.js watches wiki/<slug>/wiki/**/*.md. On user-edited claim, enqueues a verification seed (verify: <claim>) at depth 0. Wired through pro-workflow's file-watcher.js hook.

Cron tick (Phase 3.3.4)

scripts/research-tick.js is launchable from any cron-style runner. Picks the oldest opted-in wiki with pending seeds and runs a single iteration. Hook event: pro-workflow:research-tick.

Output

Each run writes:

<wiki-root>/logs/research-<UTC-timestamp>.md   # human-readable run log
<wiki-root>/derived/run-<UTC-timestamp>.json   # structured stats

Run log lines:

[2026-05-08T10:42Z] seed-3 (depth=1) "memory consolidation in agents"
  fetcher=arxiv hits=3
  fetcher=web hits=2
  compiled wiki/concepts/memory-consolidation.md (claims=7, novel=4)
  enqueued 2 follow-up seeds
  cost so far: $0.04 / $0.50

Integration with wiki-query

Every compiled page goes through wiki-cli.js page so FTS5 stays consistent. The dedupe step calls searchWiki with the candidate claim text to find near-duplicates.

Status (Phase 3.3.1)

Ships: loop driver, seed queue, web/arxiv/github fetchers, budget caps, convergence detector, kill-switch, manual run command.

Defers:

• Vector dedupe (Phase 3.3.2 via sqlite-vec)
• LLM-judged claim novelty (current = Jaccard token overlap)
• Cron + reactive (Phase 3.3.4)