Domain Extract

Analyze existing project sources to propose domain knowledge rules for approval.

Dispatch

Based on $ARGUMENTS after "domain":

• domain extract → Step 1-5 (full extraction)
• domain list → Step 6 (list only)
• domain sync-vault → Step 7 (vault sync only)

Step 1: Gather sources

Read these project sources (skip any that don't exist):

1. CLAUDE.md — sections below <!-- forge:custom --> (user-written content)
2. Auto-memory — resolve the project's memory directory:
   • Compute project path hash: the memory lives in ~/.claude/projects/-<path-with-dashes>/memory/
   • Read MEMORY.md index, then read each linked memory file
   • Filter: keep only entries with business/domain content (not purely technical)
3. CLAUDE_ERRORS.md — read all entries. Flag errors where the root cause is "Claude didn't know X about the business"
4. Agent memory — read .claude/agent-memory/implementer/*.md and .claude/agent-memory/architect/*.md
   • Filter: keep domain-relevant learnings (business logic, API behavior, deployment procedures)
5. Existing rules — read all .claude/rules/*.md. Identify files that mix domain + technical content
6. Git log — run git log --oneline -50 to identify domain-related commits (keywords: fix, deploy, API, auth, config)

Step 2: Classify findings

For each piece of domain knowledge found, classify:

• Category: deploy, api-integration, business-logic, data-model, auth-flow, external-service, workflow
• Source: which file(s) it came from
• Confidence: high (explicit in CLAUDE.md/rules), medium (in memory/errors), low (inferred from git)
• Overlap: does it duplicate existing .claude/rules/domain/*.md content?

Discard:

• Pure technical patterns (import order, test fixtures, lint rules)
• Already covered by existing domain rules
• One-time fixes that won't recur

Step 3: Generate proposals

Group findings by category. For each group, propose a domain rule file:

═══ DOMAIN EXTRACT ═══
Project: {{project_name}}
Sources analyzed: {{count}}
Existing domain rules: {{count in .claude/rules/domain/}}

── PROPOSED DOMAIN RULES ──

{{number}}. {{filename}}.md (from: {{sources}})
   globs: {{suggested glob patterns}}
   Content preview:
     "{{3-5 line summary of what the rule will contain}}"
   → Create? (y/n/edit)

Also propose a Role section if CLAUDE.md doesn't have ## Role yet:

── ROLE SECTION ──
Detected role from project context:
  "{{proposed role description}}"
  → Add to CLAUDE.md ## Role? (y/n/edit)

And a Domain section if CLAUDE.md doesn't have ## Domain yet:

── DOMAIN SECTION ──
Detected domain context:
  "{{proposed domain summary}}"
  → Add to CLAUDE.md ## Domain? (y/n/edit)

Step 4: Wait for user approval

STOP and WAIT for the user to respond to each proposal (y/n/edit).

If the user says "edit", ask what to change and regenerate that specific proposal. If the user says "all", approve all proposals.

Step 5: Create approved files

For each approved domain rule:

1. Create .claude/rules/domain/ directory if not exists
2. Write the file with frontmatter:
   yamlCopy

   ---
   globs: "{{patterns}}"
   description: "{{one-line description}}"
   domain: {{category}}
   last_verified: {{today YYYY-MM-DD}}
   ---
   

3. Content: imperative mood, concise facts, <40 lines per file
4. All content in English (Claude-consumed content rule)

For approved Role/Domain sections:

1. Read current CLAUDE.md
2. If ## Role marker exists, replace content. If not, add section before ## Arquitectura
3. If ## Domain marker exists, replace content. If not, add section after ## Role

Show summary:

═══ DOMAIN EXTRACT — DONE ═══
Created: {{N}} domain rules in .claude/rules/domain/
Updated: CLAUDE.md (Role: ✓/✗, Domain: ✓/✗)
Total domain knowledge: {{N}} files

Step 6: Domain list (for /forge domain list)

List all files in .claude/rules/domain/ with metadata:

═══ DOMAIN RULES ═══
File              Domain    Globs                          Last Verified   Status
─────────────────────────────────────────────────────────────────────────────────
ratios.md         trading   **/*ratio*,**/*price*          2026-03-15      ⚠ stale (>90d)
deploy-vps.md     infra     **/*deploy*,docker-compose*    2026-03-25      ✓ current
jira-api.md       jira      **/*jira*,**/*issue*           2026-03-20      ✓ current

Status rules:

• ✓ current — last_verified within 90 days
• ⚠ stale — last_verified older than 90 days
• ✗ no date — missing last_verified field

If no .claude/rules/domain/ exists: "No domain rules found. Run /forge domain extract to generate from existing sources."

Step 7: Vault sync (for /forge domain sync-vault)

For each file in .claude/rules/domain/ that has domain_source: vault://... in frontmatter:

1. Resolve path: replace vault:// with ~/vault/
2. Read the vault note at that path
3. Compare key facts in the vault note vs the domain rule
4. If differences found:
   Copy

   ══ VAULT SYNC ══
   File: ratios.md
   Source: vault://decisions/ratio-engine.md
   
   Changes detected:
     + New: "ONs use factor 0.001 for sub-peso bonds"
     ~ Changed: exit ratio now uses mid-price, not ask
   
   → Update domain rule? (y/n/edit)
   

5. If approved: update the domain rule content, set last_verified to today
6. If vault note not found: warn and skip

If no files have domain_source: "No vault-linked domain rules found. Add domain_source: vault://path to a domain rule frontmatter to enable sync."

Constraints

• NEVER auto-create files without user approval — always show proposals first
• Domain rules go in .claude/rules/domain/, never in .claude/rules/ root
• All generated content must be in English (Claude-consumed)
• Keep each domain rule file under 40 lines
• Frontmatter fields: globs (required), description (required), domain (required), last_verified (required), domain_source (optional)
• If a project has no meaningful domain knowledge in its sources, say so honestly — don't generate empty placeholder files