SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillswiki-room

wiki-room

Agents & models

View repository

Updated 10-15 22:33

About this skill

Skill: wiki-room

Tier C of wiki-core v2 (concept wiki-core-relations): when the main session context is already long/rotted, DO NOT load more details into the main session — open ONE room (subagent), load fully, then return a compressed conclusion. The main session keeps the map (Tier A); details live in a use-and-discard room.

WHAT

Purpose and context

  • Purpose: answer questions needing deep details from many wiki pages WITHOUT bloating the main session context — details are read in a use-and-discard room (subagent); the main session receives only a compressed conclusion with sources.
  • Trigger (when to use):
    • The main session is already long (many rounds of summarizing) and deep details from many wiki pages are needed.
    • The question requires reading >3 full wiki pages + their relation neighbors.
    • User says "mở room", "context rot", "nạp thêm chi tiết wiki", "đào sâu wiki", /wiki-room.
  • Non-goals:
    • DO NOT use when only 1-2 specific pages are needed — reading directly is cheaper than a room (speed criterion).
    • Not long-term memory — retained knowledge must be distilled into the wiki via /ingest or page updates.
    • Do not modify the wiki; the room only reads and synthesizes.

Mental model

câu hỏi → brief Tier A (list id + trang stale + episode mồi) → budget B → MỘT room (subagent, depth 1) đọc Tier B + hàng xóm 1 bước → kết luận nén ≤500 từ có nguồn → phiên chính hành động.

Input and output contract

Field Required? Meaning
In question yes what needs to be explored deeply in the wiki
In budget B no (default 12) cap on the number of Tier B pages, including 1-step relation neighbors
In list of page ids self-assembled at W01 from the question + wiki/index.md + wiki/stale.json + episode retrieval
Out compressed conclusion yes ≤500 words, every claim accompanied by its source page; stale pages flagged; budget reached → write "budget reached, not yet read: "; missing data → "missing: <what>"

"Done" = the main session receives the compressed conclusion and acts on it, without pasting full page text.

Rules and capabilities

Three hard rules (copy VERBATIM into the room prompt):

  • RULE-01 (MUST): Depth cap = 1 (G3): the room MUST NOT open a child room/subagent for any reason. Missing data → write "missing: <what>" in the conclusion, and let the main session decide.
  • RULE-02 (MUST): Hard budget: read up to B pages. Hit the ceiling → STOP reading, synthesize from what is available (circuit breaker — return a truncated result with the note "budget reached, not yet read: ").
  • RULE-03 (MUST): Sourced compressed conclusion: every claim accompanied by its source page (concepts/x.md); a source page that is stale (per the brief) must be annotated with the flag "(stale — re-verify before trusting)".

Anti-patterns (auxiliary rules for the main session):

  • RULE-04 (MUST NOT): Open multiple rooms in parallel for one question (loses control of the total budget) — one question, one room.

  • RULE-05 (MUST NOT): Use a room as long-term memory — a room is use-once-discard; knowledge that persists must be distilled into the wiki via /ingest or a page update, not left in the room transcript.

  • RULE-06 (SHOULD NOT): Call a room when the main session is still short — reading directly is always faster.

  • Capabilities: read the wiki (index, stale list, frontmatter relations:); episodic semantic retrieval; spawn exactly one read-only subagent. No wiki writes.

Failure boundaries

  • Question needs only 1-2 pages → do not open a room, read directly (valid refusal).
  • Room hits budget B → partial: truncated conclusion + list of unread pages.
  • Room lacks data → write "missing: <what>", the main session decides (do not open a child room).
  • Source page stale → still use it but flag for re-verification.

HOW

Main workflow

Step Type Inputs Action Outputs/exit Failure/next
W01 judgment question, wiki/index.md, wiki/stale.json, mem-rank.py Draft the Tier A brief (main session, cheap) list of page ids + stale flags + seed episodes ≤2 pages → read directly, stop
W02 deterministic list of ids, frontmatter relations: Fix budget B BEFORE opening (default 12, including 1-step neighbors) B + list within the ceiling —
W03 effect brief, B, 3 hard rules Open EXACTLY ONE room (Explore/general-purpose subagent) room running —
W04 judgment room conclusion Receive the compressed conclusion (≤500 words + sources), next action main session decision budget hit/missing → B01

Details of each step (source of truth for W01–W04):

  1. Draft the Tier A brief (main session, cheap): extract the list of page ids to dig into — from the question + wiki/index.md
    • wiki/stale.json (related stale pages must be explicitly stated in the brief as "possibly outdated"). Seed further with episodic SEMANTIC retrieval: python3 harness/scripts/mem-rank.py retrieve "<câu hỏi>" --kind-filter episode — catch "what the previous session did" without needing [[wikilink]]; bring the surfaced files/episodes into the brief for the room.
  2. Set the budget BEFORE opening: at most B Tier B pages (default 12) + 1-step relationship neighbors (read relations: frontmatter of the pages in the list — also counts toward B).
  3. Open EXACTLY ONE room: spawn 1 Explore/general-purpose subagent with a prompt containing: the question, the id list, budget B, and the 3 hard rules in the Rules section (copy them verbatim into the prompt).
  4. Receive a compressed conclusion (≤500 words + source citations), the main session acts on that — DO NOT paste the full text of the pages the room read into the main session.

Branches

ID Kind Guard Behavior Skip / failure Rejoin
B01 recovery room hits budget B or reports "missing: <what>" main session decides: use a truncated conclusion, or read directly the few missing pages do not open a child room, do not open a second room in parallel W04

Validation and stopping

The room stops when it has read the full list or hits B (deterministic circuit breaker by page count). Check by review: conclusion ≤500 words, every claim has a source page, stale pages are flagged. One question = at most one room.

Examples

  • Positive: long session, ask "how does wiki-core v2 handle stale across tiers?" → brief of 6 pages + 3 neighbors (B=12) → one room → 400-word conclusion citing concepts/wiki-core-relations.md, one page tagged "(stale — re-verify before trusting)".
  • Boundary/failure: the question needs 20 pages, B=12 → the room stops at page 12, returns a conclusion with "hit budget, not yet read: <8 pages>"; do not open a child room.
  • Boundary: only need concepts/solid-what-how.md → do not open a room, read directly.

Origin

  • Source: concept wiki-core-relations §2.4 Tier C + council 020726 guardrail G3/G6
  • Date: 2026-07-02