Agent Harness Bootstrap

English · 日本語

Two Claude Code skills and a viewer:

• spec-builder writes a spec that both you and the AI can read.
• harness-bootstrap fits an agent team to your repo.
• harness-view shows the result, scores it, and switches any part of it off.

This wiki is the reference. The README explains why any of it exists, and the landing page is the short version. Start there if you have read neither.

flowchart TD
    IN[/"Idea, notes, legacy docs"/]
    CODE[/"Your codebase"/]
    SB["spec-builder"]
    SPEC[/"docs/specs/ contract"/]
    HB["harness-bootstrap"]
    H[/".claude/ harness"/]
    HV["harness-view"]

    IN --> SB --> SPEC --> HB
    CODE --> HB
    HB --> H -->|"scan, score"| HV

    classDef det fill:#2D6A4F,stroke:#081C15,stroke-width:1px,color:#D8F3DC
    classDef mod fill:#5A189A,stroke:#240046,stroke-width:1px,color:#E0AAFF
    classDef art fill:#495057,stroke:#212529,stroke-width:1px,color:#F8F9FA
    class SB,HB mod
    class HV det
    class IN,SPEC,H,CODE art

Purple runs a model, green is a deterministic program, grey is a file on disk.

Start here

If you want to	Read
Install it and run it once	Getting Started
Understand why the build is small on purpose	Tailored Build
Look up a seat, a rule, a hook, a command, a flag	the reference pages below
See what your harness actually looks like	Harness View
Work out why something is not firing	Troubleshooting
Ask the obvious question	FAQ

Reference

These five pages are generated from the repository by scripts/build_wiki.py, which is why their counts can be trusted:

• Agent Reference - the seats, their model and effort, and when each is installed
• Rule Reference - which rules load in every session and which are path-scoped
• Hook Reference - what blocks what, and the exit code that does it
• Command Reference - every slash command the harness installs
• Flag Reference - what each questionnaire answer turns on

An edit made in the wiki UI is overwritten by the next build. Change the asset instead.

The one-paragraph version

spec-builder turns whatever you brought - an idea, a meeting transcript, legacy documents, a bare repo - into one contract with stable requirement IDs. harness-bootstrap reads your code, then builds the .claude/ harness that fits it: a roster derived from the contract and the modules that exist, rules scoped to paths that exist, and hooks that block rather than advise. The guardrails are shell scripts and exit codes, so the safety floor cannot depend on the model. harness-view reads the result and scores it, also with no model in the loop.

Proof you can run yourself

python eval/guardrail_eval.py                        # the safety floor, per hook flavour
python benchmark/benchmark.py                        # harnessed vs a bare repo
python harness-bootstrap/scripts/port.py --self-test # the Cursor and Codex adapter
python scripts/check_numbers.py                      # every published figure vs the scripts

No number on this wiki is remembered. If a page and a script disagree, the script is right and the page is a bug.