tt-a1i/harnessed

記録された最新のソース活動: 2026年3月29日 08:22

SkillsMP カタログ更新: 2026年11月12日 12:52

収集済み skills: 7

GitHub スター: 2
GitHub フォーク: 0

AI アシスタントでインストール

このプロンプトをコピーして、使用中の AI アシスタントに送ってください。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/tt-a1i/harnessed.

リポジトリの README

README の元ファイル：README.md

README の元ファイルを読む

README 全文を表示

# Harnessed [中文版](README.zh-CN.md) Independent quality verification for AI coding agents. Code isn't done until an isolated evaluator independently reviews it. ## Why AI coding agents evaluate their own work — and get it wrong. Research shows self-evaluation disagrees with independent evaluation 31% of the time, and AI-generated code initially passes only 42% of comprehensive tests. The fix is simple: don't let the same agent that wrote the code judge the code. ## What It Does Harnessed adds an independent QA loop to your AI coding workflow: ``` Standalone: Task → Contract → Code → QA → Fix Loop → Gate → Done Complementary: Task → SP Planning → Code → QA → Fix Loop → Gate → Done Micro: Task → Code → Gate → Done ``` - **Contract Writing** — Testable acceptance criteria before coding begins. After drafting, every user requirement is mapped to a criterion and gaps are filled automatically. - **Independent QA** — Isolated evaluator path grades code against the contract. High-risk tasks use corroborating reviewers, security-sensitive tasks add a dedicated security reviewer, and reviewer disagreement triggers tie-break escalation. - **Verification Gate** — Structured evidence (file:line citations) required before "done". If code changed since QA last ran, the gate re-runs QA before accepting. ## Installation ### Claude Code (Plugin Marketplace — recommended) ```bash # Register the Harnessed marketplace (one-time) claude plugin marketplace add tt-a1i/harnessed # Install the plugin claude plugin install harnessed ``` Updates are automatic — Claude Code syncs plugins on startup. ### Manual Installation (alternative) ```bash # Clone the repository git clone https://github.com/tt-a1i/harnessed.git # Option A: Install as a project plugin (project-local) cp -r harnessed/ your-project/harnessed/ # Option B: Install globally for all projects cp -r harnessed/ ~/.claude/plugins/harnessed/ ``` > If you downloaded a zip instead of cloning, run `chmod +x harnessed/hooks/session-start`. ### After Installation Verify: start a new Claude Code session — you should see Harnessed respond with "I'll draft acceptance criteria before coding" on your first coding task. Harnessed automatically adds `.harnessed/` to your project's `.gitignore` on first use. ### With Superpowers Harnessed detects Superpowers automatically. When both are installed: - Superpowers handles planning, TDD, and process discipline - Harnessed handles independent QA and verification - No duplication, no conflict When both are installed, Harnessed QA runs after Superpowers' process completes. If Superpowers is detected but no specs exist for the current task, Harnessed falls back to Standalone Mode. ## How It Works ### Standalone Mode (no Superpowers) 1. You give a coding task 2. **Contract Writing** generates testable acceptance criteria → `.harnessed/contract.md` 3. Agent writes code 4. **Independent QA** dispatches isolated reviewers - Primary evaluator sees: diff + contract + project context - Evaluators do NOT see: generator's reasoning or self-assessment - High-risk tasks use corroborating review, disagreement-triggered tie-break review, and explicit calibration status - Security-sensitive tasks add heuristic security issue flagging plus Semgrep/CodeQL/bandit outputs when available - QA reports now record **Confidence** and **Uncertainty** so weak evidence is downgraded instead of overstated - Overall grade: **SHIP** / **SHIP_WITH_HUMAN_REVIEW** / **ITERATE** / **BLOCKED** 5. If ITERATE: agent fixes issues, QA re-runs (max 3 rounds) 6. If SHIP or SHIP_WITH_HUMAN_REVIEW: **Verification Gate** collects file:line evidence for every criterion 7. Task complete with full audit trail ### Complementary Mode (with Superpowers) Same as above, but contract-writing is skipped (Superpowers' specs are used instead). ### Three-Tier Verification - **Tier 1 (Code Review)** — Always available. Evaluator reads the diff and checks criteria. - **Tier 1.5 (HTTP Smoke Tests)** — Auto-detected when a dev server is running but no test suite is present. Evaluator uses curl/HTTP requests to verify behavior against the live server. - **Tier 2 (Execution Verification)** — Auto-detected when test frameworks are present. Evaluator runs tests and interacts with the application. ### Project-Level Learning The independent QA evaluator categorizes each failure it finds, and those categories are recorded in `.harnessed/failure-patterns.md` across tasks. During contract writing, patterns with Count ≥ 2 automatically inform acceptance criteria so the same class of bug doesn't recur. One-off patterns decay after 90 days. ## Example QA Report After running independent QA, Harnessed produces a structured report at `.harnessed/qa-report.md`: ```markdown # QA Report ## Overview - **Verification Tier:** 2 - **Overall Grade:** ITERATE - **Criteria Passed:** 7/9 - **Critical Issues:** 0 ## Per-Criterion Evaluation ### Criterion: "Selected theme persists across page reloads" - **Grade:** FAIL - **Evidence:** src/hooks/useTheme.ts:24 — localStorage.setItem is called but getItem on line 8 reads from wrong key ("theme" vs "color-theme") - **Finding:** Theme preference is written to localStorage under key "color-theme" but read back under key "theme", so it never persists. - **Action Required:** Align the localStorage key in useTheme.ts lines 8 and 24. ### Criterion: "Dark mode applies to all UI elements" - **Grade:** PARTIAL - **Evidence:** src/components/Header.tsx:15 — className does not include theme variable - **Finding:** Header component does not consume the theme context. Background stays white in dark mode. - **Action Required:** Import and apply theme class in Header.tsx. ``` ## Skills | Skill | Invocation | Purpose | |-------|-----------|---------| | using-harnessed | Auto-loaded at session start | Routes to other skills, manages modes | | contract-writing | `harnessed:contract-writing` | Generates acceptance criteria | | independent-qa | `harnessed:independent-qa` | Dispatches isolated QA evaluator | | verification-gate | `harnessed:verification-gate` | Collects completion evidence | ## Artifacts All Harnessed artifacts are written to `.harnessed/` in your project root: ``` .harnessed/ ├── .lock # Advisory session lock (PID + timestamp) ├── contract.md # Acceptance criteria (written in both modes) ├── failure-patterns.md # Recurring failure categories (persistent across tasks) ├── qa-report.md # Primary QA evaluation report ├── qa-report-secondary.md # Corroborating reviewer report (high-risk tasks) ├── qa-report-security.md # Security reviewer report (security-sensitive tasks) ├── qa-report-tiebreak.md # Tie-break reviewer report (disagreement resolution) ├── qa-state.md # Iteration count and dispatch timestamp ├── verification-summary.md # Completion evidence └── archive/ # Archived artifacts from previous tasks ``` **Note:** Concurrent Claude Code sessions in the same project directory are not fully supported. Harnessed uses an advisory lock (`.harnessed/.lock`) to detect and warn about concurrent sessions, but this is best-effort only — artifacts from one session may still overwrite another's. ## Cost Harnessed trades speed for confidence. Costs are not only token costs. - **Token cost** — A typical standard task uses 1-2 QA rounds. High-risk tasks may use corroborating review, security review, and tie-break review. - **Latency / cycle time** — Independent QA, reruns, and high-risk corroboration add time before completion. The point is to catch issues before release, not to optimize for the fastest possible green check. - **Developer friction** — Review comments, explicit evidence gathering, and pending-human-review carry-through add workflow friction. That friction is deliberate when the task is risky, but it is still a real cost. ## When Not to Use the Full Pipeline Use the lighter path when: - the task is truly micro and inert - the user explicitly asks to skip QA - you are doing throwaway exploration that will be discarded Do **not** use the lighter path for auth, security, destructive data changes, public endpoints, or release-blocking work. ## Trust Model Harnessed is designed to **lower self-evaluation bias and increase issue-finding rate**. It does **not** claim perfect objectivity. - Independent evaluators are more trustworthy than self-assessment, but they still have evaluator bias - Confidence and uncertainty are first-class outputs - High-risk tasks use corroborating review and tie-break escalation - Security review is **security issue flagging / heuristic review**, not a complete security audit ## Injection Mitigation Harnessed uses **prompt-level mitigation** for untrusted artifacts such as contracts and diffs. It treats them as data, not instructions. This helps, but it is **not a complete safety boundary**. System-level isolation, tool allowlists, and explicit trusted/untrusted context partitioning remain part of the architecture, especially for high-risk review paths. ## Customization - **Skip QA for a task:** Tell Claude "skip QA" or "treat this as a micro task" - **Skip the contract:** Tell Claude "skip the contract, just do QA" - **Disable entirely:** Tell Claude "don't run Harnessed" — it will comply but note that verification was skipped - Harnessed always respects explicit user instructions. The automated gates prevent the *agent* from skipping verification on its own, not the user from choosing to skip it. ## Design Principles 1. **Verification > Process** — A simple plan with verified results beats an elaborate plan with unverified results 2. **Independence > Thoroughness** — A quick independent check catches more than a thorough self-review 3. **Evidence > Assertion** — "It works because file:42 shows X" beats "It works because I wrote it correctly" 4. **Skepticism > Trust** — Default stance is "prove it works" not "assume it works" ## Troubleshooting **Harnessed not loading (no contract generated on first task):** - Run the hook manually: `bash harnessed/hooks/session-start` — it should output JSON containing the meta-skill - If you get an error, check that `python3` is installed (used for JSON escaping) - If you downloaded a zip, ensure the hook is executable: `chmod +x harnessed/hooks/session-start` **QA evaluator fails to produce a report:** - This usually means the evaluator prompt was too large. Try a smaller diff or fewer criteria. - After one retry, Harnessed marks the task as BLOCKED and escalates to you. **Artifacts appearing in git status:** - Add `.harnessed/` to your project's `.gitignore` **Concurrent sessions:** - Running multiple Claude Code sessions on the same project is not fully supported — Harnessed warns via advisory lock but cannot guarantee artifact isolation ## License MIT

リポジトリの README を読み込んでいます…