harness（harness 文档）

Harness

Build the verification infrastructure that makes agent work trustworthy.

Principles

• Environment > instruction — the harness matters more than the prompt
• Mechanical enforcement > prose — hooks, CI, health checks, and scripts beat wishes
• Separate builder from judge — harness builds the rig, verify uses it
• Smallest useful harness first — add layers in order, stop when the repo becomes reliably verifiable
• Progressive disclosure — keep the core workflow here, load patterns on demand

Handoffs

• Need to review a diff, branch, or PR on real surfaces → use verify
• Need to update AGENTS.md, README.md, specs, or repo docs → use docs

The 7-Layer Stack

1. Boot — single command starts the app
2. Smoke — a fast proof the app is alive
3. Interact — agent can exercise the real surface
4. E2e — key user flows work end to end
5. Enforce — hooks, CI gates, lint rules, or mechanical checks
6. Observe — logs, health endpoints, traces, machine-readable signals
7. Isolate — worktrees or containers do not collide

Workflow

1. Audit

Grade the repo across these dimensions:

• bootable
• testable
• observable
• verifiable

For each, report:

• status: pass / partial / fail
• evidence: file or command
• gap: what is missing

Use references/grading.md. Lowest dimension sets the overall grade.

2. Setup

Build missing layers in this order:

Boot → Smoke → Interact → E2e → Enforce → Observe → Isolate

Each step should be independently useful. Stop once the repo is reliably verifiable; do not build a cathedral because you got excited.

3. Improve

Tighten weak or flaky layers:

• remove mock-only confidence theater
• replace one-off checks with reusable scripts or hooks
• add logs and health signals agents can query
• make parallel work safe when agent collisions are real

4. Hand Off

When the repo reaches C+ and can be judged honestly, hand off to verify. If harness changes created doc drift, hand off to docs.

Anti-Patterns

• Mock-only tests — pass by construction, verify nothing
• Self-evaluation — builder grading its own work
• Docs-only fixes disguised as harness work
• Routine PR review here — that's verify
• Perfect harness upfront — iterate from real failure modes

Output

After harness work, report:

• grade before and after
• dimensions with evidence
• files changed
• remaining gaps ranked by impact
• verify readiness
• recommended next handoff: verify, docs, or human review

References

• references/grading.md — harness quality grading scale with mechanical criteria
• references/setup-patterns.md — boot, smoke, e2e, observability, and isolation patterns
• references/industry-examples.md — external patterns and justification for harness investment