Customized domain spec: VC to developer knowledge brain

[chapter37haptics]

Summary

• Spec for customizing gbrain from VC/executive use case (people, companies, deals) to developer use case (goals, decisions, processes, concepts)
• 4 entity types with hard MECE boundaries, mapped 1:1 to the /goal development workflow
• 12-file change manifest (3 skill rewrites, 3 skill patches, 4 code patches, 2 unchanged) with execution order
• Page templates for all 4 types plus structured debug-trail timeline entries
• Architecture compliance audited against all 4 docs in docs/architecture/
• 5 Codex cross-model reviews incorporated (retrieval gap, architecture, infra impact, scope narrowing, MECE)

Test plan

• Read the spec at docs/specs/customized-domain.md (Context section first)
• Verify the 12-file change manifest covers the full closed-loop (detect, write, store, retrieve, present, enrich)
• Confirm MECE boundaries table has hard rules for all 6 entity pairs
• Implementation is a follow-up — this PR is spec only (1 new file, 0 code changes)

🤖 Generated with Claude Code

[chapter37haptics]

Codex Review (gpt-5.3-codex) — PR #4

Gate: PASS | Tokens: 68,137

Findings (3 fixed, 2 noted)

1. Tier inconsistency — types.ts and markdown.ts placement (FIXED)

types.ts and markdown.ts were listed under Tier 2 in the manifest table but marked **HIGH (Tier 1)** in the Priority column. An implementer following the spec top-to-bottom would see them in Tier 2 and defer them, but they are compile-time prerequisites — inferType() returns goal/decision/process, which must exist in the PageType union or TypeScript rejects the build.

Fix applied: Moved both to the Tier 1 "Code/config files to PATCH" section (items garrytan#9 and garrytan#10). Removed from Tier 2 table.

---

2. Execution order — code-layer must precede skill-layer (FIXED)

The original execution order put quality.md first (delegation root reasoning), with code changes at steps 7-9. This is correct for the delegation chain but risky for incremental testing: if you implement signal-detector/SKILL.md (step 3) and test it before types.ts (step 8), the agent tries to write a page with type: goal and hits a runtime/compile error because the type doesn't exist yet.

Fix applied: Reordered to code-first:

1. types.ts (compile-time prerequisite)
2. markdown.ts (type inference)
3. link-extraction.ts (auto-link)
4. quality.md (delegation root)
5. brain-ops/SKILL.md → ... → brain-first.md

---

3. Missing PageType audit step (FIXED)

The spec names 10 specific files but doesn't require a grep pass to find all PageType consumers — switches, whitelists, and filters in diagnostics, tests, and secondary paths could silently reject developer types.

Fix applied: Added verification step 10: grep -rn 'PageType\|ALL_PAGE_TYPES' src/ --include='*.ts' to catch hidden consumers not named in the manifest.

---

4. CE weakness — scope boundary implicit (NOTED)

The 4-way MECE is strong on mutual exclusivity but collective exhaustiveness only holds under the assumption of "goal-driven developer memory." External references, long-lived project containers, and documentation artifacts are implicitly forced into goals/ or concepts/. Acceptable for v1 scope. If scope expands later, an explicit boundary statement or a 5th type would be needed.

---

5. CLAUDE.md surface area (NOTED)

For a spec-only PR, replacing CLAUDE.md (1312 deletions) and adding CLAUDE-original-read-only.md (1445 additions) is review noise. The spec itself is the deliverable — consider splitting the CLAUDE.md changes into a separate commit or PR to keep this PR focused.

---

Recommendation: Address finding #1 (tier inconsistency) before implementation because an implementer following the spec literally could defer compile-time prerequisites, causing type errors that block all downstream changes. Findings #4 and #5 are cosmetic for v1 scope.

🤖 Reviewed by Codex (gpt-5.3-codex) via /gstack-codex review