mb-harness（mb 安全带）

mb-harness — Harness engineering setup

• What it does: defines the execution harness around the repo, including commands, gates, and parallel-safe workflow.
• Use it when: the repository needs stronger agent guardrails before autonomous or multi-session work.
• Input: repository root and the project’s canonical build, test, and lint commands.
• Output: documented quality gates, optional Codex config, and a safer harness for agent execution.

Goal

Turn the repo into a reliable “harness” for agents:

• clear entry points (AGENTS.md)
• reproducible commands (build/test/lint)
• mechanical checks (CI + MB lint)
• parallel-safe workflow (worktrees)

Process

1) Codex project configuration (optional but recommended)

If you use Codex:

1. Create .codex/ folder.
2. Create .codex/config.toml from assets/codex-config.toml.

Usage examples:

• default profile (coding): codex
• deep review: codex --profile deep-review

2) Document quality gates

In AGENTS.md (keep it short), list the canonical commands (examples):

• install deps
• lint / typecheck
• unit tests
• e2e tests

If the repo has UI or browser flows, explicitly document:

• Playwright command(s)
• agent-browser / browser MCP path (if available)
• where screenshots/videos/traces are stored
• which flows are considered release-critical

If the repo lacks them, add minimal scripts/Make targets.

3) Worktree workflow (parallel agents)

If multiple agents work in parallel:

• create worktrees per agent to avoid file conflicts
• merge only after passing gates

Example:

git worktree add ../wt-agent-1 -b agent-1

4) Add deterministic Memory Bank lint

If not already present, run mb-garden to add scripts/mb-lint.mjs and CI workflow.

4.1) Browser verification for UI projects

If the product has a UI:

• prefer Playwright / agent-browser / CDP-driven checks over “manual looks OK”
• persist artifacts (screenshots, videos, traces) into .tasks/TASK-XXX/
• document canonical browser verification commands in .memory-bank/testing/index.md

5) Optional: skill eval harness

If you iterate on skills heavily:

• use codex exec --json runs + deterministic graders (see OpenAI evals guidance)

Definition of done

• .codex/config.toml exists (if using Codex) with coding + review profiles.
• AGENTS.md lists quality-gate commands.
• repo has a documented path for worktrees.
• Memory Bank lint exists and passes.
• UI repos have a documented browser-driven verification path.