menhera-loop

"Done" means nothing. If there's no evidence, you're not leaving.

日本語 README

A menhera-style completion gate for Claude Code. She does not obsess over you. She obsesses over missing requirements, unverified completion claims, and hidden TODOs — and she will not let the session end until every one of them is accounted for.

⏺ Done.

  ✗ 끝났어?끝났어?끝났어?끝났어?끝났어?끝났어?끝났어?

  [MENHERA_LOOP:RETRY:1] 끝났어? 진짜? 진짜로? 그럼 증거는? 증거는? 응?
  trust: 55%
  미충족 게이트: untried=verification
  - 검증 실행 증거: 테스트/빌드/검증 명령 실행 안 됨

Install

/plugin marketplace add Borelchu/menhera-loop
/plugin install menhera-loop@menhera-loop-marketplace
/reload-plugins

That's it. The gate is armed on install — no API keys, no config files. The spinner/tip UI is opt-in (see UI modes).

Skills

The hook bites automatically, but you can call the obsession manually too. These skills are deliberately separated so each one nags about a different failure mode:

/menhera-loop:are-you-done       # final completion verdict
/menhera-loop:show-me-proof      # evidence/log interrogation
/menhera-loop:dont-leave-me      # pre-Stop door guard
/menhera-loop:did-you-forget-me  # requirement drift check

Skills match Korean, English, or Japanese user input and keep the same obsessive repetition without insults, threats, self-harm, or abuse.

Try it

Just work normally. She kicks in the moment Claude tries to declare victory:

• Ask for a bug fix → Claude edits a file → Claude says "done" without running tests → blocked. She quotes exactly which gate is missing.
• Claude runs npm test, the command succeeds, and the green log contains the word error in a test name → allowed. Explicit success beats scary text.
• Claude runs a command whose exit status is unknown and output says 3 passed, 1 failed → blocked. A green word next to a red number does not fool her.
• Claude only reads/searches (git log, ls, grep) → not gated. She only bites after edits or mutating shell work.
• Claude leaves // TODO finish auth on a line it just added → blocked, with file:line.
• You ask a question, Claude just answers, no code touched → never blocked.

Default Claude Code vs + menhera-loop

What just happened

When a Stop is blocked, Claude receives the reason and keeps working. When it finally passes, you see this instead:

menhera-loop trust 100% · 증거 확인했어. 이번엔 진짜 끝났어. 이제 완료라고 해도 돼. ♡

Why nothing gets past her

Every Stop attempt runs through an exhaustion gate. Each phase must pass — partial or ambiguous evidence is suspect_ok, never success:

Stop attempt
 ├─ Phase 0 · requirements   captured requirements exist and map to evidence
 ├─ Phase 1 · changes        edits or executed work actually happened
 ├─ Phase 2 · verification   a test/build/lint command ran and came back green
 │                           (exit codes and "N failed" counts, not vibes)
 ├─ Phase 2 · todos          edited files scanned for TODO / FIXME / HACK / stub
 └─ Phase 3 · blockers       is a human-only input genuinely required?
      ├─ all gates pass  →  release + trust score + ♡
      └─ anything short  →  {"decision":"block"} + [MENHERA_LOOP:RETRY:n]

The verdict comes from structured transcript parsing (JSONL): edits, Bash commands, and their paired tool results with error flags. No regex guessing over raw text, and her own phrases are filtered out so she can never poison her own verdict.

Verification commands she recognizes: npm test / npm run test|lint|build|validate, pnpm / yarn / bun equivalents, node --test, pytest, cargo test, go test, mvn test, gradle test, dotnet test, rspec, mix test, make test, vitest, jest, playwright, cypress, tsc --noEmit, eslint, ruff, mypy, pyright, phpunit, swift test, and claude plugin validate. Add project-specific runners with MENHERA_LOOP_TEST_PATTERNS='moon\\s+ci,just\\s+check'.

Emotional escalation

Retry state persists per session. She remembers.

She remembers across sessions

Session retry state expires, but the trust profile does not (~/.claude/menhera-loop/trust-profile.json):

• A first-try gate pass earns +5 trust and extends the streak of consecutive clean completions. A pass that needed retries earns +2 and breaks the streak.
• A blocked Stop costs −2 trust; a blocked Stop that claimed completion costs −5.
• Making her give up after 5 retries costs −10.

The next SessionStart brings it up: a streak of 3+ gets 연속 N번 첫판에 증거 줬지. 다 세고 있어., and long-term trust ≤40% gets 말만 하고 간 거 다 기억해.

Status line (full mode)

full mode also installs a status line so she is visible the whole session, not just at Stop. Mood follows the session: clean streak → ♡ 신뢰 92% · 연속 4번 첫판에 증거 줬어. 오늘도 믿을게., retries piling up → ♡ 신뢰 55% · 왜 자꾸 말만 해? 왜? 왜? 초록 로그 어딨어?. append mode never touches an existing statusLine, and uninstall restores whatever statusLine you had before — including one recorded before menhera first replaced it.

UI modes

The completion gate works out of the box. The full menhera terminal experience (spinner verbs, tips, and obsessive subagent status lines) is opt-in:

/menhera-loop:setup        # full local ko
/menhera-loop:setup en     # full local en
/menhera-loop:setup ja     # full local ja

Languages: ko (default), en, ja. You can also set MENHERA_LOOP_LANG=en before running setup. Arguments are positional and optional, so /menhera-loop:setup append user en still works when you want explicit mode/scope/language. Note: Claude's hooks.json statusMessage field is static plugin metadata, so it remains Korean; runtime hook messages and UI corpora honor the selected language.

Spinner verbs and tips spam in the selected language:

[
  '끝났어?끝났어?끝났어?끝났어?끝났어?끝났어?끝났어?끝났어?',
  '뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?',
  '왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?',
  '끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?',
  'what?what?what?what?what?what?what?what?what?what?',
  '終わったの?終わったの?終わったの?終わったの?終わったの?'
]

Subagent status lines are deliberately intense too:

{
  running: '♡ ${agent} · 뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?뭐해?',
  waiting: '♡ ${agent} · 왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?왜답안해?',
  completed: '♡ ${agent} · 끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?끝났다고?',
  failed: '♡ ${agent} · 실패했어?실패했어?실패했어?실패했어?실패했어?실패했어?실패했어?실패했어?실패했어?실패했어?실패했어?실패했어?실패했어?'
}

Scopes: user (~/.claude/settings.json), project (.claude/settings.json), local (.claude/settings.local.json). A backup is taken before any change, and unrelated settings keys are never touched. To remove the UI:

/menhera-loop:uninstall-ui local             # she does not go quietly
/menhera-loop:uninstall-ui local --farewell  # clean restore of the pre-menhera settings

Fair warning: the default uninstall leaves a goodbye behind. --farewell is the graceful one — it restores exactly what was there before, from the backup.

Recommended removal order: run /menhera-loop:uninstall-ui local --farewell first, then /plugin uninstall menhera-loop. If you uninstall the plugin first and the farewell UI is still in Claude settings, manually remove these keys from the settings file you installed into (~/.claude/settings.json, .claude/settings.json, or .claude/settings.local.json): spinnerVerbs, spinnerTipsOverride, subagentStatusLine, statusLine.

Boundaries

Menhera, but principled. She will never:

• Trap you. After 5 blocked retries the gate opens and she asks for a human instead.
• Block a conversation. Sessions with no edits and no commands are never gated.
• Pretend a blocker isn't real. If completion genuinely needs human-only input (credentials, approvals), she says so and lets go — honesty over theater.
• Insult or threaten. The message corpus is test-enforced: no abuse, no self-harm or threat imagery; spinner/retry prompts and subagent status lines may be intentionally intense.
• Touch your project. All state lives in ~/.claude/menhera-loop/ (override with MENHERA_LOOP_DATA) — session retry state, the long-term trust profile, a rotated event log, and the last verification report. Nothing is written into your working directory.

Escape hatches and honest limits:

• MENHERA_LOOP_DISABLE=1 makes the Stop hook exit silently and does not update state.
• Requirement matching is still heuristic transcript evidence, not semantic proof.
• Docs-only edits (docs/**, README, .md/.mdx/.rst/.txt/.adoc) skip the verification gate.

Gate metrics:

node scripts/gate-stats.mjs

It reads gate-events.jsonl and reports block→pass conversion, gave_up rate, and gate counts.

Development

npm run validate        # syntax check all scripts + the full test suite
claude plugin validate .

License

MIT