harness-kit

Harness Kit

English · 简体中文

Harness Engineering toolkit for Claude Code — a Claude Code Plugin (15 skills + 8 framework agents + project templates) that brings disciplined AI-driven development to fullstack and backend projects. Claude-native (the framework agents ship as plugin agents — no per-project copy).

Goal: humans only do "describe the requirement" and "step in when AI can't"; everything else — 7-agent pipeline, verify gates, structured documents — runs automatically.

What's inside

This is a Claude Code Plugin packaging that gives any project fifteen AI skills:

Pipeline skills (six task shapes; the AI picks the right one from your natural-language description)

• /harness-kit:harness — full 7-stage pipeline (RA → SA → GR → Dev → CR → QA → Delivery). Use for real feature / bug / refactor work.
• /harness-kit:harness-plan — design-only mode: runs RA + SA + GR, stops with a verdict before any code is written. Use to vet a design.
• /harness-kit:harness-explore — research / feasibility mode: light RA + a findings.md with citations. No design, no code. Use for "can we even do X?"
• /harness-kit:harness-goal — open-ended Dev + QA loop bounded by a measurable success criterion and a budget. Use for "keep improving until coverage > 80%" type tasks.
• /harness-kit:harness-batch — runs T-01…T-NN sequentially through pm-orchestrator, each task in its own sub-agent context. Stops on strong signals (verify_all FAIL, pm-orchestrator FAIL, intervention STOP, safety hook block). Use for /harness-plan decompositions, accumulated backlogs, post-checkup sweeps, or imported task lists — instead of invoking /harness N times.
• /harness-kit:harness-stream — like batch, but the task pool stays alive: it re-reads BATCH_PLAN.md every iteration, so tasks you add mid-run (in chat or by appending to the pool / an ADD intervention) get planned and executed without re-invoking. Best-effort completion (a failed task is marked and skipped, the stream keeps going) with the same hard-safety stops as batch. A task that needs human input (clarification, a human-reserved decision, or authorization for a safety-critical action) is deferred (a distinct needs-human status — set aside, its dependents blocked) and surfaced together at stream end in STREAM_REPORT.md's ## Needs your input section, so the stream never sits stopped waiting on you. Use when you want to fire requirements as they occur to you and only watch results. Complex multi-part requirements are auto-decomposed at ingest into dependency-staged sub-task rows (simple ones stay one row; rows you author — ADD / hand-written — run as-written). Ambient mode: just invoke with no pool-id — a default pool (docs/batches/default/) is auto-created and a UserPromptSubmit hook (gated by .harness/ambient.flag) makes every chat message a heartbeat that folds requirements into the pool and drains it; no /loop, no re-invocation, no keyword. It is session-scoped — a SessionStart hook auto-clears the flag each new session, so re-invoke /harness-stream to resume.

Setup skills

• /harness-kit:harness-init — bootstrap Harness skeleton in a new project (asks 6 questions, generates .harness/ + .claude/ + AI-GUIDE.md + stub CLAUDE.md / copilot-instructions.md in ~30s)
• /harness-kit:harness-adopt — non-invasively add Harness to an existing project (detects stack, extracts conventions, prompts before applying)
• /harness-kit:harness-upgrade — bring an already-initialized but stale project up to the current plugin layout (relocate scripts to .harness/scripts/, content-refresh depth-sensitive scripts for correct root derivation, re-install the pre-commit hook, rewire settings, regenerate verify_all while preserving your B.* checks — dry-run preview, idempotent, proven with a green verify_all)
• /harness-kit:harness-language — set, switch (English ↔ Chinese), or refresh a project's output-language policy by surgically rewriting only the three policy surfaces (.harness/rules/00-core.md section + CLAUDE.md line + .github/copilot-instructions.md line) to the target language's current canonical text. Self-bootstraps the text from the plugin templates (so an old project can pull a refreshed policy), non-destructive, idempotent, dry-run preview, .bak per file.