code-factory

Agent skills for dynamic software delivery. Each skill works standalone in Claude Code or Codex; together they cover the path from rough idea to merged PR.

Workflow

Rough idea → writing-specs → execute-dynamic-workflow → PR with green CI → babysit-pr
                  ↑                                            |
                  |            User approves spec/plan         |
                  +------------------ feedback ----------------+

1. writing-specs — Collaborate with the user on a design spec with EARS requirements, then run an adversarial spec review (Codex, clean context) before writing the artifact. This is the human checkpoint.
2. execute-dynamic-workflow — Deliver the spec autonomously: an expensive planner writes a granular plan directory, tasks are elaborated just-in-time and implemented in parallel git worktrees by difficulty-matched Codex/Claude implementers, adversarial reviewers (cross-model on high-risk tasks) attack each diff, and the full check suite runs at phase boundaries behind a deterministic hygiene gate. One approval gate after planning; everything else is flags, not questions. Ships as a workflow shape plus tested example scripts (full-delivery.js for spec-to-PR, mini-feature.js for small scoped changes) that custom workflows copy and edit.
3. babysit-pr — Monitor the PR (or a whole stack — Graphite and GitHub gh stack both supported) in a long-running loop: snapshot status with a deterministic script, fix CI bottom-up, answer review comments (🤖-prefixed), push once per iteration, with stale-CI and give-up guards so the loop can't spin.

Standalone utilities:

• codex-mcp — The verified recipe for driving the Codex MCP: tool parameters, model names, sandbox/approval flags (including the .git writable_roots trick that makes commits work under workspace-write), effort ladder, relay template for workflow scripts, and failure signatures. Other skills reference it instead of hardcoding Codex details.
• audit-tests — Audit an existing test suite for low-value, brittle, duplicative, or AI-generated tests and produce a precise cleanup plan. Its test-quality doctrine is also enforced at review time by the adversarial reviewer.

Skills

Skill	Description
writing-specs	Design specs with EARS requirements, clarification markers, brownfield gap analysis, and adversarial spec review
execute-dynamic-workflow	Spec → PR via dynamic workflows: worktree-parallel implementation, cross-model adversarial review, phase-boundary verification
babysit-pr	Long-running PR/stack monitor: fix CI, respond to reviews, push verified fixes — Graphite, gh-stack, or standalone
codex-mcp	Verified known-good configuration for the Codex MCP (models, sandbox, approvals, worktree cwd rules)
audit-tests	Test-suite audit: keep/delete/merge/upgrade grading with a behavior-value rubric

Agents

Subagent contracts used by execute-dynamic-workflow (Claude Code):

Agent	Description
workflow-planner	Spec → compact plan directory (plan.md TOC, shared preface, per-phase sketches)
dynamic-elaborator	Phase sketch → verified just-in-time task list, one small file per task
adversarial-reviewer	Clean-context diff attack: requirement fraud, hostile inputs, test quality, plan-vocabulary leaks

Design principles

• Spec-minimal frontmatter (name + description only) so every skill loads in Claude Code, Codex, and anything else that reads the Agent Skills spec.
• Deterministic scripts for what must not vary (scan-plan-vocab.sh, check-pr.sh): parsing and gating live in bash; judgment lives in prose.
• Hard rules stated first, with the enforcement stacked: implementer prompts, reviewer attack lists, and a script gate all independently catch plan-vocabulary leaks and dishonest tests.
• Worktree isolation over file locks: parallel implementers each get their own checkout; integration merges and cleans up in the same step.
• Flags over questions: after the last human gate, autonomous runs surface judgment calls at the end instead of stopping.

Installation

Claude Code

claude plugin marketplace add xmtplabs/code-factory
claude plugin install code-factory@code-factory

Or from a local clone:

git clone https://github.com/xmtplabs/code-factory.git
claude plugin marketplace add /path/to/code-factory
claude plugin install code-factory@code-factory

Codex

Codex discovers skills in ~/.agents/skills:

git clone https://github.com/xmtplabs/code-factory.git
mkdir -p ~/.agents/skills
for s in /path/to/code-factory/skills/*/; do
  ln -s "$s" ~/.agents/skills/"$(basename "$s")"
done

(Note: the execute-dynamic-workflow example scripts orchestrate via Claude Code's Workflow tool; on Codex the skill's shape and the codex-mcp/babysit-pr/writing-specs/audit-tests skills work directly.)

OpenCode

git clone https://github.com/xmtplabs/code-factory.git

Add to your opencode.json:

{
  "plugins": [
    {
      "name": "code-factory",
      "path": "/path/to/code-factory"
    }
  ]
}