HarnessFlow

English | 中文

A harness that drives AI coding agents from idea to shipped work — Matt-aligned main chain under hf-* names, plus progress recovery, auto mode, review discipline, and demo acceptance.

Main-chain skill content is adapted from mattpocock/skills (MIT; copying authorized). HarnessFlow keeps progress.md, interactive/auto, and cross-stage hf-review.

Install

python scripts/install.py --target cursor --dest /path/to/project
python scripts/install.py --target opencode --dest /path/to/project
./install.sh --target both --dest /path/to/project
./install.ps1 -Target both -Dest C:\path\to\project

• Cursor: copies skills into .cursor/skills/ and writes an always-on .cursor/rules/harness-flow.mdc (paths rewritten). Unrelated project skills are preserved. Use --mode symlink to follow this checkout.
• OpenCode: installs under .opencode/skills/ (generated, gitignored); top-level skills/ remains the source of truth.
• Claude Code: install as a plugin from this repo’s marketplace, or vendor skills/ into the project.

In this repo itself (OpenCode): python scripts/install.py --target opencode --dest .

Usage

1. One-time project setup

Ask the agent to run hf-grill-with-docs and create the product layer from skills/hf-workflow/references/product-layer-templates.md (CONTEXT.md, product/…, docs/adr/, features/). Specs and tickets live under features//.

2. Start work (talk naturally)

The agent should load hf-workflow first. Example prompts:

3. Follow the main chain

hf-workflow
  → hf-grill-with-docs
  → hf-to-product-architecture — hf-review (product architecture) —
  → hf-to-spec            — hf-review (spec) —
  → hf-to-architecture    — hf-review (architecture) —
  → hf-to-tickets
  → hf-implement          — hf-review (incl. code gate) —
  → hf-ship

What you should see on disk as you go:

Exploration path: 模式: 探索 → conclusion.md (never ship; no promoting exploration code).

4. Recover from disk (don’t rely on chat)

Read product/progress.md and each features//progress.md to see stage and next step. Artifact layout is defined by the stage skills and skills/hf-workflow/references/product-layer-templates.md.

Rules of thumb

• Spec / product architecture / feature architecture / code should get independent hf-review (code gate is Standards + Spec inside that skill).
• Underspecified choices: propose a default → product/assumptions.md → continue.
• Say auto only when you want passing reviews to advance without waiting; degraded same-session review is a hard stop in auto.

5. Execution modes

• interactive (default): wait for your confirmation after reviews and demo acceptance.
• auto: you must say so explicitly. Passing review advances with auto-approved. Floors: implement/review in subagents, no degraded self-approve, assumptions ledgered; present demo evidence at the next human interaction.

Skills (core)

Meta: hf-tdd, hf-grilling, hf-domain-modeling, hf-codebase-design.

License

MIT. Main-chain skill prose adapted from mattpocock/skills (MIT).