wiki

name	wiki
description	Project knowledge base — living, cross-referenced, always current. Initialize, ingest new knowledge, query existing, lint for staleness. Powered by Karpathy's LLM Wiki pattern.
argument-hint	<init|ingest|query|lint> [topic or question]

/wiki Skill

Purpose

Maintains a living knowledge base for the project. Unlike memory (append-only historical notes) or docs (static decision records), the wiki is actively maintained — pages are updated when knowledge changes, cross-references are kept current, stale information is cleaned up.

The wiki answers: "What is the current truth about X in this project?"

Location

.atl/wiki/
├── index.md                    ← Auto-maintained table of contents
├── {topic-1}.md                ← Knowledge pages (kebab-case)
├── {topic-2}.md
└── ...

Four Modes

init Mode

Usage: /wiki init (no argument)

Scaffolds .atl/wiki/ in the current project. Safe to re-run — it's a no-op if the wiki is already initialized.

Process:

1. Check for .atl/wiki/ directory.

   • Does not exist → create it.
   • Exists with index.md containing the initialized template → idempotent no-op, print wiki: already initialized (N pages) and exit.
   • Exists but empty or missing index.md → treat as not-yet-initialized and continue.

2. Create .atl/wiki/index.md with this template:

   # Project Wiki
   
   > Current truth about this project — maintained automatically from `<!-- learning -->` markers via `/save-learnings`, or manually via `/wiki ingest`.
   
   ## Pages
   
   (none yet — pages appear here as knowledge accumulates)
   
   ## Health
   
   - Last ingest: never
   - Last lint: never
   
   ## How this works
   
   - **Content flows in** from inline learning markers captured at session end (`atl learning-capture`), plus any manual `/save-learnings` or `/wiki ingest` invocation.
   - **Pages are topic-based** (one concept per page) and reflect current truth — old facts get replaced, not appended.
   - **Run `/wiki query <question>`** to search wiki content.
   - **Run `/wiki lint`** periodically to catch stale / orphan / contradicting pages.
   

3. Report one-line status: wiki initialized at .atl/wiki/ (ready for ingest)

When init fires automatically:

• First session of a project where the user has atl setup-hooks installed but the project has no .atl/wiki/ yet — the SessionStart hook invokes /wiki init as part of the wiki-bootstrap step.
• Manual: user runs /wiki init directly.

What init does NOT do:

• It does not ingest existing knowledge sources. Use /wiki ingest for that after init.
• It does not create topic pages. Pages are created by ingest or save-learnings when content exists.

---

ingest Mode

Usage: /wiki ingest (no argument needed)

Scans all project knowledge sources and updates wiki pages:

Sources scanned:

1. <!-- learning --> markers in the current session transcript — primary source, see learning-capture rule
2. .atl/journal/*.md — chronological per-agent learning record (the post-Q4 single layer; supersedes the retired agent-memory/)
3. .atl/docs/*.md — finalized decisions from brainstorms
4. .atl/brain-storms/*.md (completed only) — decision context
5. Recent conversation context — what was just discussed/built

Process:

1. Read all sources
2. For each piece of knowledge, determine the topic (e.g., "caching", "authentication", "order-management")
3. If wiki page exists for that topic → update (merge new info, resolve contradictions, keep latest truth)
4. If no wiki page exists → create new page
5. Update cross-references (backlinks between related pages)
6. Update index.md with new/changed pages

Page format:

# {Topic Title}

> Last updated: {date}
> Sources: [journal](../journal/2026-01-15_api-agent.md), [brainstorm](../brain-storms/auth-design.md)

## Summary
{2-3 sentence overview of this topic in the project}

## Current State
{What is true RIGHT NOW — not history, not plans, just current reality}

## Key Decisions
{Important decisions made about this topic, with brief reasoning}

## Patterns & Rules
{Established patterns, conventions, rules for this topic}

## Known Issues
{Current problems or limitations}

## Related
- [{related-topic-1}]({related-topic-1}.md)
- [{related-topic-2}]({related-topic-2}.md)

Important: Wiki pages reflect current truth. If old memory says "we use pattern X" but later memory says "pattern X caused problems, switched to Y" — the wiki page says "we use Y" (not both).

---

query Mode

Usage: /wiki query how does caching work in this project?

1. Read index.md to find relevant pages
2. Read those pages
3. Synthesize an answer from wiki content
4. Cite which wiki pages the answer came from

This is useful when:

• New developer (or new Claude session) needs to understand a topic
• You forgot how something was decided/implemented
• You want a quick refresher before making changes

---

lint Mode

Usage: /wiki lint

Health check for the entire wiki:

Checks performed:

1. Stale pages — last updated > 30 days ago, source files have changed since → flag for review
2. Contradictions — two pages say conflicting things → flag for resolution
3. Orphan pages — page exists but no other page references it → suggest connections
4. Missing pages — topic referenced in another page but no page exists → suggest creation
5. Duplicate topics — two pages cover the same topic → suggest merge
6. Index sync — index.md matches actual files in wiki/ → fix if out of sync

Output:

Wiki Health Report:
──────────────────────────
📊 Total pages: 12
✅ Healthy: 9
⚠️  Stale (>30 days): 2 (caching-patterns.md, email-setup.md)
❌ Contradiction found: auth.md says "15min token" but jwt-config.md says "30min token"
🔗 Orphan: database-indexes.md (no incoming links)
📝 Missing: "rate-limiting" referenced in api-endpoints.md but no page exists
──────────────────────────

Fixing automatically...
✅ index.md synced
✅ Created rate-limiting.md (stub)
⚠️  Review needed: auth.md vs jwt-config.md contradiction
⚠️  Review needed: 2 stale pages

Auto-fixable issues are fixed silently. Contradictions and stale content are reported for human review.

---

Integration with Other Systems

Learning markers → /save-learnings → Wiki

The normal flow when atl setup-hooks is installed:

1. Claude drops <!-- learning topic=... --> markers inline during the conversation (per learning-capture rule)
2. SessionStart of the next session → atl session-start runs atl learning-capture --previous-transcripts, which scans the previous session's transcripts for markers
3. If markers found → the next session's first turn invokes /save-learnings --from-markers --transcripts ...
4. /save-learnings updates journal/, agent children/, skill learnings/, and wiki pages (replace/update)

Example propagation:

<!-- learning topic: redis-cache; body: TTL should be 30 min, not 15 -->
  → journal/{date}_{agent}.md: append historical note with date
  → wiki/redis-cache.md: UPDATE "TTL is 30 minutes" (replace old "15 minutes")
  → (if domain-specific) agents/{agent}/children/redis-cache.md: append or create

Without markers (or without hooks), manual /wiki ingest and manual /save-learnings still work.

Agent Startup → Wiki

Agents read relevant wiki pages at session start (per knowledge-system rule). Agent doesn't read ALL pages — only those related to its domain:

• API Agent → pages about api patterns, database, auth, caching
• Flutter Agent → pages about ui patterns, state management, navigation
• Selection is based on page topics matching agent's responsibility area

/brainstorm done → Wiki

When a brainstorm completes, its decisions are ingested into the wiki automatically.

---

Wiki Page Lifecycle

New knowledge discovered (conversation, brainstorm, learning)
    ↓
Determine topic
    ↓
Page exists?
├── Yes → Update page (merge, resolve contradictions, update date)
└── No → Create new page (from template)
    ↓
Update cross-references (backlinks)
    ↓
Update index.md

Important Rules

1. Wiki = current truth. Not history, not plans. What is true RIGHT NOW.
2. Update, don't append. If a fact changes, the old version is replaced, not kept alongside.
3. Cross-reference always. Every page should link to related pages. Orphans are flagged by lint.
4. Auto-maintained. Humans rarely edit wiki directly. It's maintained by /save-learnings, /brainstorm done, and /wiki ingest.
5. Agent-readable. Pages are structured for both human and AI consumption — clear sections, no ambiguity.
6. Topic-based, not date-based. Unlike journal (date-based) or memory (date-based), wiki is organized by topic. One page per concept.
7. Lint regularly. Run /wiki lint periodically (monthly or when something feels off).

Accumulated Learnings

(Auto-rebuilt by /save-learnings from learnings/*.md frontmatter. Do not edit by hand. Currently empty — populates as the skill is used and edge-case learnings accumulate.)