English · 日本語

The recommended path is the plugin marketplace, from inside any Claude Code session:

/plugin marketplace add nguyenhx2/agent-harness-bootstrap
/plugin install harness-bootstrap@agent-harness-bootstrap
/plugin install spec-builder@agent-harness-bootstrap

Updates then arrive with /plugin update. For an offline, pinned install, download the bundle from the latest release and unzip it into your skills directory:

unzip agent-harness-bootstrap.zip -d ~/.claude/skills/

Claude Code picks the skills up on the next session. Each skill directory carries a VERSION file, so ~/.claude/skills/harness-bootstrap/VERSION tells you what you have.

The blocking hooks declare onFailure: "block", so a guard that crashes or times out refuses the call instead of letting it through. That key needs Claude Code 2.1.295 or later.

It depends on what you already have.

Each skill stands alone. If you start with /harness-bootstrap and there are no specs, it offers to hand over to spec-builder.

flowchart TD
    A["1. Read the code"]
    R[/"Inventory Report"/]
    B["2. Ask what code<br/>cannot decide"]
    C{"3. Plan OK?"}
    D["4. Scaffold"]
    K[/"ADDED / KEPT"/]
    X[/"CONFLICT queue"/]
    E(["5. Summary and next step"])

    A --> R --> B --> C
    C -->|"change it"| B
    C -->|"yes"| D
    D --> K --> E
    D --> X -->|"you reconcile"| E

    classDef det fill:#2D6A4F,stroke:#081C15,stroke-width:1px,color:#D8F3DC
    classDef mod fill:#5A189A,stroke:#240046,stroke-width:1px,color:#E0AAFF
    classDef hum fill:#1D3557,stroke:#0B1B2B,stroke-width:1px,color:#A8DADC
    classDef art fill:#495057,stroke:#212529,stroke-width:1px,color:#F8F9FA
    class A,B mod
    class C,E hum
    class D det
    class R,K,X art

1. It reads your code first. With any source in the repo the analysis is mandatory. The Inventory Report lists the stack, the modules, the conventions, the risky operations and the gaps, and you see it before anything is written.
2. It asks what code cannot decide. The first question is the repo's purpose (code, research, strategy, docs, in any combination); later questions are asked only where that purpose needs them. The closed choices come as one pre-answered questionnaire of at most two calls, in your language. Anything the analysis already answered is not asked again.
3. It shows the plan. Files created, kept and modified, and the roster with each seat's model and effort. You confirm or change it.
4. It scaffolds. The scaffolder never overwrites a file. Each one is ADDED, KEPT (identical) or CONFLICT (exists and differs). A CONFLICT is the reconciliation queue, not an error. It also writes a short CLAUDE.md into each module, and ends by checking that every link between seats, hooks, rules and commands resolves.
5. You get a summary and a next step, usually /task-resume.

The harness is not finished until the loop has run once:

/new-task          # asks first, then writes docs/tasks/active/TASK-NNN-<slug>.md and its board row
/task-resume       # reads the task's checkpoint and picks up from the next step
/review-changes    # the review gate, before you open a PR

• /new-task opens a task only for work you agreed to, and only after you say yes.
• Task state then changes through python .claude/scripts/task.py, which keeps the file, its folder and the board row in step. A hand edit of a board row is refused.
• An empty board is an answer, not an error. In a brownfield repo the analysis seeds the board from its gap list, so there is real work on day one.

python eval/guardrail_eval.py

From this repository, the eval scaffolds a real harness (strict, long-session and database hooks on) and fires known-bad and known-good payloads at it, per hook flavour. No model is consulted, and CI runs the same command. benchmark/benchmark.py replays the must-block payloads against a harnessed repo and the same repo without .claude/.

In your own repo, python .claude/scripts/harness-check.py names the file and line of any link that points at nothing.

rm -rf .claude/

That removes every agent, rule, hook and command. The scaffold also wrote AGENTS.md, CLAUDE.md, the docs/ tree and the per-module CLAUDE.md files; they are plain files, and git status lists them. There is no daemon, no service and nothing outside the repository to clean up.

• Tailored Build - why the roster is smaller than the catalogue
• Harness View - see what you just built, and score it
• Troubleshooting - when a hook does not fire