Knowledge Extraction & Documentation Sync

Extracts knowledge from Amp threads and synchronizes project documentation.

Pipeline

REQUEST → THREADS → TOPICS → CODE → DOCS

Phase 1: Discover Threads

Start from user request:

# "Document last 2 weeks"
find_thread after:2w

# "Summarize auth work"
find_thread "authentication"

# "What touched the SDK?"
find_thread file:packages/sdk

# Combined
find_thread "refactor" after:1w file:packages/api

Phase 2: Extract Topics

Spawn parallel Task agents (2-3 threads each):

Task prompt:
"Read threads [T-xxx, T-yyy] using read_thread.
Goal: 'Extract topics, decisions, changes'

Return JSON:
{
  'topics': [{
    'name': 'topic name',
    'threads': ['T-xxx'],
    'summary': '1-2 sentences',
    'decisions': ['...'],
    'patterns': ['...'],
    'changes': ['...']
  }]
}"

Collect outputs → Oracle synthesizes:

Oracle: "Cluster these extractions. Deduplicate.
Latest thread wins conflicts. Output unified topic list."

See reference/extraction-prompts.md for full templates.

Phase 3: Verify Against Code

For each topic, verify claims:

Topic: "Added retry logic to API client"

1. finder "retry logic API client"
   → finds src/api/client.ts

2. gkg__search_codebase_definitions "retry"
   → RetryPolicy class at L45

3. gkg__get_references "RetryPolicy"
   → 12 usages across 4 files

→ Confirmed: topic matches code

Phase 4: Map to Docs

Discover existing documentation:

# Structure
gkg__repo_map on docs/, .claude/skills/, */AGENTS.md

# Find existing mentions
Grep "topic keyword" docs/
Grep "RetryPolicy" AGENTS.md

Read target files before updating:

Read docs/ARCHITECTURE.md
→ Note structure, sections, voice
→ Identify insertion point

See reference/doc-mapping.md for target file conventions.

Phase 5: Reconcile

Oracle compares three sources:

Oracle prompt:
"Compare:
1. TOPICS: [extracted]
2. CODE: [verified state]
3. DOCS: [current content]

Output:
- GAPS: topics not in docs
- STALE: docs ≠ code
- UPDATES: [{file, section, change, rationale}]"

Phase 6: Apply Updates

Text updates:

Read target → edit_file with precise changes
Preserve structure and voice

Architecture diagrams:

mermaid with citations:
{
  "code": "flowchart LR\n  A[Client] --> B[RetryPolicy]",
  "citations": {
    "Client": "file:///src/api/client.ts#L10",
    "RetryPolicy": "file:///src/api/retry.ts#L45"
  }
}

Parallel updates: Multiple unrelated files → spawn Task per file.

Concrete Example

User: "Document the auth refactor from last week"

1. find_thread "auth" after:7d
   → [T-abc, T-def, T-ghi]

2. Task agents extract (parallel):
   Agent A: T-abc → {topics: [{name: "JWT migration"}]}
   Agent B: T-def, T-ghi → {topics: [{name: "session cleanup"}]}

3. Oracle clusters:
   → "Auth Refactor" with sub-topics

4. Verify:
   mcp__gkg__search_codebase_definitions "JWTService"
   → confirmed at packages/auth/jwt.ts

5. Map docs:
   Grep "authentication" AGENTS.md
   → Section exists at line 45

6. Oracle reconciles:
   → Gap: JWT migration not documented
   → Update: AGENTS.md auth section

7. Apply:
   edit_file AGENTS.md add JWT details
   mermaid auth flow diagram

Tool Quick Reference

Quality Checklist

- [ ] Topics verified against code (GKG)
- [ ] Existing docs read before updating
- [ ] Changes surgical, not wholesale
- [ ] Mermaid diagrams have citations
- [ ] Terminology matches existing docs

Troubleshooting

Thread not found: Try topic keywords, widen date range

Too many threads: Add file: filter, narrow dates

Topic ≠ code: Code is truth; note as "planned" or "historical"

Doc structure unclear: Read first, match existing patterns