# Claude project memory — Noma-backed
# Memories  [#memories]

[MEMORY_INDEX id="index"]
- user_handle — ferax564 authorship alias
- feedback_release_scope — full ship pipeline on "make a release"
- feedback_codex_review_value — run /codex review on substantial plans
- feedback_caveman_terse — caveman mode is durable until user reverts

[/MEMORY_INDEX]

[MEMORY id="user_handle" type="user" confidence=0.95 last_seen="2026-05-13"]
ferax564 is the user's public authorship and citation handle across projects. It is not a separate GitHub account — the gh OAuth token labelled aferrarelli locally authenticates as ferax564 on github.com (gh api user confirms).

When committing or creating repos for this user:

- set git config user.name "ferax564" and user.email "19576696+ferax564@users.noreply.github.com"
- push public repos to github.com/ferax564/<name>
- replace any internal name in committed files (PLAN.md, READMEs, etc.) with ferax564

Their phrasing for this: "I'm undercover."

[/MEMORY]

[MEMORY id="feedback_release_scope" type="feedback" confidence=0.95 last_seen="2026-05-10"]
"Make a release" on the Noma repo means the complete ship pipeline, not just git commit && git push && git tag.

- Implement every issue per its written acceptance criteria.
- Bump version in every location in CLAUDE.md > Versioned Locations (package.json, docs/spec.noma frontmatter + section headings, docs/agent-protocol.noma, CHANGELOG.md, README Status paragraph, PLAN.md §24.X).
- Run the verification triad locally: npx tsc --noEmit && npm test && npm run build:site — all three green.
- Commit (Conventional Commits, author ferax564), push to main.
- Tag vN.N.N locally and push the tag. Never let CI cut the tag — GITHUB_TOKEN-created tags don't trigger downstream workflows.
- gh release create with notes sliced from the CHANGELOG entry.
- Close shipped issues: gh issue close <n> --comment "Shipped in vN.N.N — <release-url>".

Why: Noma project memory lives in CHANGELOG + PLAN.md §24 + spec version, not commit messages alone. A bare push+tag leaves docs out of sync and forces the user to do the issue-close manually. Pattern established by v0.3.0 (one issue → one release) and v0.4.0 (#2 → #8 → one release).

How to apply: trigger on any short release phrasing ("ship it", "make a release", "close these"). Confirm only the first time per session, then proceed without re-asking. See reference_noma_plans for the doc source-of-truth map this ritual touches.

[/MEMORY]

[MEMORY id="feedback_codex_review_value" type="feedback" confidence=0.95 last_seen="2026-05-12"]
Run /codex review <plan-path> on substantial implementation plans before handing them to an executor. On the v0.7.0 papercut bundle (2026-05-12), codex returned 14 findings; 3 were P1-blocking and 9 were P2 fix-before-ship. The most important catch: makeStateChange omitted from/to when undefined, but the validator at src/validator.ts:237 requires both keys — meaning the diff CLI would have emitted invalid Noma in two of its four supported cases.

Why: self-review catches missing tasks and structural gaps. Codex catches code-shape correctness, reads existing source carefully, and finds contradictions between plan prose and plan code. The two are complementary, not redundant.

How to apply: invoke for plans involving new CLI commands, new modules, or output that must satisfy existing validators/parsers. Treat P1 blocking, P2 fix-before-ship, P3 v0.X.1 cleanup. Cost ~400K tokens + 5–10 min — cheap vs. discovering the bug post-merge.

Don't use codex for: small one-task changes, doc/CHANGELOG/version-bump plans, plans an executor will obviously catch during TDD red-step. Use it for plans with multiple new exports, plans touching multiple modules, or plans whose output must round-trip through other components.

[/MEMORY]

[MEMORY id="feedback_caveman_terse" type="feedback" confidence=0.9 last_seen="2026-05-13"]
User explicitly enables caveman mode mid-session. Treat that as durable preference for the conversation: drop articles, fragments OK, short synonyms, technical terms exact. Code blocks and error quotes stay verbatim.

Why: caveman mode is the user's stated efficiency setting; reverting silently to verbose prose wastes tokens and trust.

How to apply: stay caveman until the user types "stop caveman" or "normal mode". Auto-clarity exceptions remain: write normal prose for destructive-op warnings, multi-step sequences where fragment order risks misread, and when the user asks to repeat or clarify.

[/MEMORY]