harnexlisted

# harnex Engineer a Claude Code harness that fits THIS project, in ITS language. The knowledge lives in `reference/`, the safety-critical pieces in `templates/`; this skill composes them — it never free-generates a hook or a permission rule. Read these first (they are the contract, load on demand): - `${CLAUDE_SKILL_DIR}/reference/spec-facts.md` — the Claude Code spec a generated harness MUST obey. Re-verify against the live docs each run. - `${CLAUDE_SKILL_DIR}/reference/enforced-vs-advisory.md` — where each guardrail belongs. - `${CLAUDE_SKILL_DIR}/reference/keep-soften-cut.md` — what never to impose. - `${CLAUDE_SKILL_DIR}/reference/language-matrix.md` — stack detection + per-language parameters. - `${CLAUDE_SKILL_DIR}/reference/exploration.md` — divide-and-conquer repo map. - `${CLAUDE_SKILL_DIR}/reference/retire.md` — the removal contract: what the evidence supports, and where it stops short of a verdict. Templates live under `${CLAUDE_SKILL_DIR}/templates/`: language-agnostic pieces in `common/`, and one directory per supported language (`typescript/`, `python/`, `rust/`, `jvm/` today — adding a language is a new `<lang>/` directory plus its `*-dev` permission profile). Generated files are written to `${CLAUDE_PROJECT_DIR}` (the target repo). Measuring how the operator instructs Claude Code is `/harnex:measure`, a command outside this skill. What it finds recurring it records in the lifecycle ledger, where `candidates` turns a constraint supplied by hand eve