harness-engineering-skills

English | 中文

Harness Engineering Skills

Stometa's public curated Claude Code skillset — a small, opinionated set of skills we use ourselves, published periodically.

Why this repo

This is the public companion to Stometa's private stometa-skillset. We dogfood a larger internal skillset day-to-day; selected skills are extracted, polished, and published here in batches. The goal is to share the workflows that actually hold up under real engineering work — not a pile of prototypes.

The first batch ships two skills: review-loop (already proven in daily use) and harness (multi-agent orchestration for larger tasks). Both are installable as a single Claude Code plugin.

Workflow at a glance

harness is a cybernetics-inspired orchestrator: planning and execution live in separate sessions so context cannot leak; every checkpoint runs against a fresh sub-agent to reset eigenbehavior; an engine script owns state and enforces hard gates so the LLM cannot self-certify; and a cross-model peer (a different vendor's CLI) reviews before PR so you never merge on a single model's opinion. Persistent retro feeds learnings back into future tasks — that's the closing loop of the cybernetic system.

flowchart TB
    H((Human))
    H -->|"harness plan task-id"| HOST

    subgraph HOST["Orchestrator host — pick one (symmetric)"]
        direction LR
        CC["Claude Code CLI"]
        CX["Codex CLI"]
    end

    HOST --> ENG[["harness-engine.sh<br/>single source of truth<br/>state · phase machine · hard gates"]]

    ENG --> S1
    subgraph S1["Session 1 — Planning (recommended host: Claude Code)"]
        direction TB
        PL["Orchestrator = Planner<br/>brainstorm + draft spec.md"]
        SE["harness-spec-evaluator<br/>fresh sub-agent (Claude)"]
        OK1["spec.md approved"]
        PL --> SE
        SE -->|revise| PL
        SE -->|approve| OK1
    end

    S1 -. session ends — planning context discarded .-> S2

    subgraph S2["Session 2 — Execution (recommended host: Codex)"]
        direction TB
        CPL{{"For each Checkpoint NN"}}
        GEN["harness-generator<br/>fresh sub-agent per CP<br/>TDD + verification preloaded"]
        EVL{"harness-evaluator<br/>fresh sub-agent per CP<br/>Tier 1 deterministic + Tier 2 LLM"}
        MORE{"more checkpoints?"}
        E2E["E2E Evaluator<br/>cross-checkpoint data-flow audit"]
        RL[["review-loop<br/>cross-model quality gate"]]
        FV["full-verify<br/>tests · coverage ≥ threshold · lint · types"]
        PR["Open PR"]
        RT["harness-retro<br/>fresh sub-agent"]

        CPL --> GEN --> EVL
        EVL -->|FAIL / REVIEW| GEN
        EVL -->|PASS| MORE
        MORE -->|yes| CPL
        MORE -->|no| E2E
        E2E -->|FAIL| GEN
        E2E -->|PASS| RL
        RL --> FV --> PR --> RT
    end

    subgraph RLSUB["review-loop · cross-LLM peer review"]
        direction LR
        PEER["Peer reviewer CLI<br/>codex OR gemini<br/>(allowlisted)"]
        HEV["Host LLM evaluates<br/>ACCEPT / REJECT / INSIST"]
        FRESH["Fresh peer session<br/>final approval pass"]
        DONE["pass-review-loop"]

        PEER -->|FINDING fN| HEV
        HEV -->|fix + commit| PEER
        PEER -.CONSENSUS.-> FRESH --> DONE
    end
    RL -. invokes .-> RLSUB

    RT --> RD[(".harness/retro/<br/>cross-task learnings<br/>git-tracked, persistent")]
    RD -. informs future tasks .-> H

    classDef antiDrift stroke:#d97706,stroke-width:2px;
    class GEN,EVL,SE,RT antiDrift;
    classDef gate stroke:#059669,stroke-width:2px;
    class ENG,RL,FV gate;

Legend — orange-bordered nodes are the fresh-sub-agent drift firewalls; green-bordered nodes are the engine-enforced gates that the LLM cannot bypass.

Hosts and roles

The model running each role is decoupled from the model hosting the session — that's why the same pipeline works whether you start in Claude Code or Codex.