my-claude-lang

MyCL — my-claude-lang 1.0.46

A semantic-verification layer on top of Claude Code. MyCL imposes a 22-phase development pipeline, ratchets discipline through audit chains, and refuses to fail open. Opt-in per session via /mycl (1.0.5+) — silent until you ask for it. Pure Python 3.8+, no AI agent frameworks, no MCP servers.

Türkçe sürüm: README.tr.md

What MyCL is (and isn't)

MyCL is six Python hooks (UserPromptSubmit, PreToolUse, PostToolUse, Stop, SubagentStop, PreCompact) that, once activated with /mycl:

• Track the active phase in .mycl/state.json
• Seal each phase with an audit entry in .mycl/audit.log
• Enforce per-phase tool allowlists (gate_spec.json) — fail-closed
• Detect spec blocks line-anchored, hash them, and gate downstream tools on developer approval
• Inject phase-specific context (<mycl_active_phase_directive>, <mycl_phase_status>, <mycl_phase_allowlist_escalate>) every turn
• Speak TR + EN bilingually (no labels, blank-line separated)

Before activation (default state for any new session): all six hooks no-op — banner, deny, audit, snapshot — nothing fires. Claude Code behaves exactly as without MyCL installed.

MyCL is not:

• An AI agent framework (we don't reinvent Claude Code)
• LangGraph or any state-machine library (the 22 phases are data in gate_spec.json)
• An MCP server (Plugin Rule C filters out any plugin with .mcp.json)

Quick install

git clone https://github.com/YZ-LLM/my-claude-lang.git
cd my-claude-lang
bash setup.sh

The installer:

1. Verifies Python 3.8+
2. Copies the repo to ~/.claude/mycl/
3. Registers mycl.md skill in ~/.claude/skills/
4. Drops data files into ~/.claude/data/
5. Merges 6 hooks into ~/.claude/settings.json (idempotent — safe to re-run)
6. Smoke-tests py_compile + lib imports

Activation (1.0.5+)

MyCL is opt-in per Claude Code session. After installation:

> /mycl                  # bare activation — MyCL now runs the pipeline
> /mycl todo app yap     # activation + first intent in one prompt

Once /mycl is detected on any prompt, the active session ID is written to .mycl/active_session.txt and MyCL stays on for the rest of that session. Closing Claude Code starts a fresh session ID, so the next time you open it, MyCL is silent again until you type /mycl.

The /mycl token is stripped from the prompt before the model sees the user's actual message; the model receives an <mycl_activation_note> that quotes the stripped message so it knows what the user asked.

Why opt-in? Many sessions don't need the 22-phase pipeline (quick edits, exploration, scripts). MyCL only kicks in when you explicitly ask for it.

The 22 phases

#	Phase	What it does
1	Intent Gathering	Clarify what the developer wants — one question per turn, summary + askq approval
2	Precision Audit	Validate intent across 7 dimensions (auth, failure modes, scope, PII, observability, SLA, idempotency)
3	Engineering Brief	Translate intent to engineering English — silent phase
4	Spec + Approval	Write 📋 Spec block (Objective/MUST/SHOULD/AC/Edge/Technical/OOS) + askq approval
5	Pattern Matching	Learn naming/error/test patterns from existing code (greenfield → skip)
6	UI Build	Frontend with mock data, dev server + browser auto-open (no UI → skip)
7	UI Review	Deferred — free-form intent classify (approve/revise/cancel/sen-de-bak)
8	Database Design	3NF schema, covering indexes, reversible migrations (no DB → skip)
9	TDD Execution	RED → GREEN → REFACTOR per AC; final test suite green
10	Risk Review	4-lens parallel scan (review/simplify/perf/security) + unbounded askq
11	Code Review	Auto-fix correctness/logic/error-handling/dead-code
12	Simplification	Remove premature abstractions, duplicates, unnecessary patterns
13	Performance	Auto-fix N+1, blocking I/O, memory leaks, allocation churn
14	Security (project-wide)	semgrep + npm/pip/cargo audit + secret scanner
15	Unit Tests	Code-coverage tests beyond Phase 9's AC tests
16	Integration Tests	Real cross-module flows (testcontainers, in-memory DB)
17	E2E Tests	Playwright headless (no UI → skip)
18	Load Tests	k6/locust against spec NFRs (no NFR → skip)
19	Impact Review	Downstream effects (importers, contracts, schema, cache)
20	Verification + Mock Cleanup	Spec coverage table + automation barriers + process trace
21	Localized Report	Translate Phase 20 to user's language (skippable)
22	Completeness Audit	Hook-written audit; surfaces uncovered MUSTs

For each phase, see skills/mycl/asama<NN>-*.md.

Architecture