lawrencewzen

Hero Coding — Git skill for Claude Code

Git community

极简自主编码 Agent：丢需求进 inbox，出 git commit · Minimal autonomous coding agent harness.

How to install Hero Coding

This entry records only its repository, not the path inside it, so there is no exact command to give. Open lawrencewzen/hero-coding and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Hero Coding does

极简自主编码 Agent：丢需求进 inbox，出 git commit · Minimal autonomous coding agent harness.

Alternatives in Git

Hermes IDE — One console for every coding agent 333 ★
Harness Drift — Continuous-drift sweep — arch warnings, dead code, and Sonar drift outside the PR lifecycle 324 ★
Diri — Native macOS orchestrator for coding agents — run Claude Code, Codex, Cursor, Gemini and shells in parallel ac 272 ★

README

hero-coding

[English](./README.md) · [中文](./README.zh.md)

A minimal harness for autonomous coding agents. Drop a user story in `inbox/`, get git commits out.

Philosophy

**The harness is the thinking layer. The agent is a replaceable executor.**

Reasoning lives in the loop, not in the weights. A non-reasoning model running inside a harness with dispatcher, verifier, reviewer, and guardrails behaves like a reasoning system as a whole — focused on the task, caught when it strays, retried with concrete feedback.

Architecture

inbox/us-001.md
       │
       ▼
   Dispatcher  ── watches inbox/, creates an isolated git worktree per story
       │
       ▼
   ┌── Round Loop ─────────────────────────────────┐
   │                                               │
   │   Worker   ── coding agent, native ReAct loop │
   │     │        atomic git commits per change    │
   │     ▼                                         │
   │   Verifier ── runs the story's `verify:`      │
   │     │        commands (tests / lint / type)   │
   │     ▼                                         │
   │   Reviewer ── reads story + verifier output   │
   │              + git diff, returns PASS / FAIL  │
   │                                               │
   └─────────────────┬─────────────────────────────┘
                     │
              ┌──────┴──────┐
            PASS           FAIL
              │             │
              ▼             ▼
          done/        feedback appended → next round

This is a deliberately reduced **Execute-Verify** harness — the simplified form of the PEV (Plan-Execute-Verify) pattern with the Plan stage left outside the harness boundary. No LLM Planner is run inside hero-coding, by design — see `docs/architecture-research-2026.md` §E.4 for the rationale.

**Where Plan happens.** Plan typically lives in an upstream conversation between the user and an interactive coding agent (Claude Code, Codex, Cursor, etc.). The user discusses the change,