code-factory

---

Why

Most agent skill sets nail idea→ship engineering but have no delivery layer: work gets built the moment it's imagined, nothing lands on a board, and the team can't see what's happening. Code Factory adds the missing half — and welds both halves into one flow with a hard seam:

Nothing is implemented until it's scheduled. Tickets land on the board as suggestions; the Product Manager pulls them into an epic + iteration; only Ready work gets built.

The flow

flowchart LR
    subgraph Architect["🏗 Architect (in-loop)"]
        G["/grill-with-docs"] --> S["/to-spec"] --> T["/to-tickets"]
    end
    subgraph Delivery["📋 PM + Delivery Lead"]
        B[("Board backlog")] --> E["epic + iteration"] --> R["Ready"]
    end
    subgraph Execution["⚙️ Agents (own context)"]
        ENG["engineer<br/>tdd + lazy"] --> REV["reviewer<br/>review-rig"]
    end
    T -->|suggestions| B
    R -->|one ticket, fresh context| ENG
    REV --> DONE(["done — team-visible"])
    W["/wayfinder<br/>big foggy work"] -.-> S
    TR["/triage<br/>raw inbound"] -.-> B
    D["/diagnosing-bugs<br/>something broken"] -.-> ENG

One way in: /navigator knows every skill and routes you — obvious cases route, unclear cases get one clarifying question.

The roster

Every role ships as an agent in both runtimes — agents/<role>.md for Claude Code, codex-agents/cf_<role>.toml for Codex — plus an in-loop skill where the role is conversation-shaped. Skills hold the behaviour; agents are thin role wrappers that compose them.

Role	Sandbox	Wields
Navigator	read-only	the whole map — routes, never executes
Architect	write	grill* · wayfinder · domain-modeling · codebase-design · to-spec · to-tickets · prototype
Product Manager	write	backlog health · prioritization · epics · roadmap · triage
Delivery Lead	write	delivery presets (Scrum / Kanban / Scrumban) · all 10 rituals · flow health · coaching
Engineer	write	implement · tdd · lazy — one Ready ticket per dispatch
Reviewer	read-only	the review rig
Researcher	write (docs)	primary-source research → cited findings file

Skills are shared across roles, not owned 1:1. Every agent returns a structured JSON result, so dispatches can be chained by a composer.

The review rig

Evidence first, judgement second:

diff → deterministic pre-pass ─ fallow static analysis
                              └ codebase-memory blast radius     → evidence pack
     → dimension sub-agents (toggle on/off):
         Standards · Spec · Correctness · Tests · Laziness · Boundaries
     → synthesize (no cross-dimension reranking)
     → graded verdict vs review.schema.json → optional auto-fix

ultra mode widens the same backbone to a per-subsystem × per-dimension grid for whole-mission branches.

The ritual library

Ten rituals, each its own skill, fired on demand or by the preset's cadence:

/planning · /standup · /retro · /refinement · /iteration-review · /roadmap · /estimation · /prioritization-session · /dependency-risk · /capacity

Install

Fastest — any agent, via skills.sh:

npx skills add AaronAbuUsama/code-factory        # pick skills + target agents interactively
npx skills add AaronAbuUsama/code-factory --all  # everything, everywhere, no prompts

Claude Code (plugin, includes the agents + MCP)

claude plugin marketplace add AaronAbuUsama/code-factory
claude plugin install code-factory

Codex — .codex-plugin/ carries the skills; install the seven role agents with:

cp codex-agents/cf_*.toml ~/.codex/agents/

Pi — plugin.yaml.

codebase-memory MCP — the plugin invokes codebase-memory-mcp from your PATH (mcp/mcp-servers.json). Install the binary on each machine; without it, graph-backed features degrade to grep and everything else still works.

First run

/setup