forth-ai-harness v0.0.0
GitHub ↗

concrete idea shipped product

The lifecycle

Four stages, owned alternately by the PO (you) and Eng (the plugin). Pick a stage to spotlight its commands.

Get started — your first feature

    The engine inside every command

    No command is one-shot. Each runs a loop until a machine-checkable gate is green — and, with --codex, until a second AI agrees too.

    1

    Generate

    A subagent drafts the artifact or patch in a fresh context window.

    2

    Codex pass

    An independent AI critiques — write phases isolate it in its own git worktree.

    3

    Check

    Automated gates run: lint, typecheck, tests, a11y, Playwright, ship-gates.

    4

    Converged?

    Green + Codex approves → mirror to GitHub, swap labels, suggest the next command. Else loop.

    exit = Claude self-assessment ✓ AND Codex verdict = approved | n/a AND gate green
    A conjunction — the generator can never approve its own work alone. Flat 3× → PLATEAU.md; Claude ✓ vs Codex ✗ 3× → DISAGREEMENT.md.

    Commands

    PO Eng Offline / ops

    cap iteration cap · Codex tier · <tag> artifact mirrored to GitHub · label issue transition

    Skills

    Reusable behavior modules. Universal ones load on every command; phase-specific ones power one stage.

    Agents

    Specialized subagents commands fan out to — each runs in a fresh context window so the main thread stays under budget.

    What makes it FoWork-specific

    The generic lifecycle is tailored to the work-ssot monorepo — its branch model, deploy signal, and demo flow.