ai-software-factory

KO | EN

gitlite — search

#java #devops #nodejs #codex #animation #react-native #svg #computer-vision #github #cross-platform

ai-software-factory

(★ 9)

AI Agent pipeline for AI-assisted development. One shared prompt library generates configs for Claude Code, Kiro, Cursor, Codex CLI, and Windsurf across any number of polyrepos and tech stacks.

파일 탐색기

ZIP 다운로드

- factory-check.mjs
- factory-guard.mjs
- factory.mjs
- - 00-introduction.md
  - 01-concepts.md
  - 02-the-pipeline.md
  - 03-the-agent-chain.md
  - 04-path-enforcement.md
  - 05-adapters.md
  - 06-profiles.md
  - 07-cross-repo.md
  - 08-extending.md
  - 09-design-decisions.md
  - README.md
- cross-repo.md
- walkthrough.md
- manifest-backend.yaml
- manifest-frontend.yaml
- workspace.yaml
- bun-hono.md
- go-echo.md
- nextjs-app-router.md
- node-fastify.md
- python-fastapi.md
- python-library.md
- quarkus-reactive.md
- react-rsbuild-microfrontend.md
- react-vite.md
- - backend-builder.md
  - devops-builder.md
  - doc-writer.md
  - frontend-builder.md
  - migration-author.md
  - performance-reviewer.md
  - researcher.md
  - security-reviewer.md
  - spec-writer.md
  - story-writer.md
  - test-verifier.md
  - validator.md
- - feature-factory.md
  - quick-fix.md
  - spike.md
- - feature.ts
  - init.ts
  - install.ts
  - sync.ts
- - claude-code.ts
  - codex.ts
  - index.ts
  - kiro.ts
- - agent-meta.ts
  - contracts.ts
  - detect-stack.ts
  - profile-defaults.ts
  - scope.ts
  - workspace.ts
- cli.ts
- manifest.ts
- render.ts
- adapters.test.ts
- contracts.test.ts
- detect-stack.test.ts
- factory-check.test.ts
- factory-guard.test.ts
- feature.test.ts
- guard.test.ts
- manifest.test.ts
- profile-defaults.test.ts
- render.test.ts
- workspace.test.ts
.gitignore
package.json
pnpm-lock.yaml
README.md
tsconfig.json

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Central factory for a 12-agent software factory pattern. Generates platform-specific files (Claude Code, Kiro, Codex CLI) for many polyrepos from a single source of truth.

Designed for the case: **many repos, many techs, many AI platforms.**

## What it is

- **Prompts library** — 12 agent + 3 skill prompts written once, platform-neutral and stack-neutral.
- **Profile library** — pre-written rule packs per stack (Next.js, Node+Fastify, Go+Echo, Python+FastAPI, etc.).
- **Platform adapters** — code that renders prompts + profile + per-repo manifest into the right files for each AI platform.
- **CLI** — `factory install` reads `.factory.yaml` in any repo and generates everything.

## Architecture: AI and human collaboration

The chain is structured so the human stays in the loop where judgment matters, and steps out where the AI is reliable. Three layers, two roles, three checkpoints.

### Layers

| Layer | Owner | What it does |
|-------|-------|--------------|
| Orchestrator (skill) | AI, driven by human input | Chain logic. Decides which agent to invoke next. Pauses for human checkpoints. Routes failures back to the right builder. Does NOT edit files. |
| Specialist agents (12) | AI, with restricted tools | Each does one job in its own fresh context window. Tool scoping prevents agents from doing each other's work. |
| Reviewer | Human | Approves story, approves brief, reviews diff before merge. Tunes the rules over time. |

### The Tier 3 flow (full chain)

```mermaid
sequenceDiagram
    actor Human
    participant Orchestrator
    participant Agents as Specialist Agents

Human->>Orchestrator: /feature-factory <request>
 Note over Orchestrator: triage gate
 Orchestrator->>Agents: invoke researcher (read-only)
 Agents-->>Orchestrator: relevant files, patterns, risks
 Orchestrator->>Agents: invoke story-writer (read-only)
 Agents-->>Orchestrator: user story + acceptance criteria

Note over Human,Orchestrator: ⏸ CHECKPOINT 1
    Orchestrator->>Human: present story
    Human->>Orchestrator: approved

Orchestrator->>Agents: invoke spec-writer (read-only)
 Agents-->>Orchestrator: technical brief (files-that-will-change, API, tests)

Note over Human,Orchestrator: ⏸ CHECKPOINT 2
    Orchestrator->>Human: present brief
    Human->>Orchestrator: approved

Orchestrator->>Agents: invoke migration-author (scoped to migrations)
 Agents-->>Orchestrator: migration files + safety notes (or "Not applicable" if no schema changes)
 Orchestrator->>Agents: invoke backend-builder (scoped edit)
 Agents-->>Orchestrator: files + API contract + test results
 Orchestrator->>Agents: invoke frontend-builder (scoped edit) (passes API contract verbatim)
 Agents-->>Orchestrator: files + test results
 Orchestrator->>Agents: invoke devops-builder (scoped to infra)
 Agents-->>Orchestrator: CI/IaC files (or "Not applicable" if no infra changes)
 Orchestrator->>Agents: invoke test-verifier (test files only)
 Agents-->>Orchestrator: pass/fail per acceptance criterion

opt any AC fails (max 3 iterations)
        Orchestrator->>Agents: re-invoke responsible builder
        Agents-->>Orchestrator: fixed
        Orchestrator->>Agents: re-invoke test-verifier
        Agents-->>Orchestrator: pass/fail
    end

Orchestrator->>Agents: invoke security-reviewer (read-only)
    Agents-->>Orchestrator: security findings (Critical / Important / Minor)
    Orchestrator->>Agents: invoke performance-reviewer (read-only)
    Agents-->>Orchestrator: perf findings (Critical / Important / Minor)
    Orchestrator->>Agents: invoke validator (read-only)
    Agents-->>Orchestrator: project-rule findings (Critical / Important / Minor)

opt Critical findings from any reviewer (max 3 iterations)
        Orchestrator->>Agents: re-invoke responsible builder
        Agents-->>Orchestrator: fixed
        Orchestrator->>Agents: re-invoke the reviewer that flagged it
        Agents-->>Orchestrator: findings
    end

Orchestrator->>Agents: invoke doc-writer (scoped to docs/)
 Agents-->>Orchestrator: CHANGELOG + README + migration guide + suggested PR description

Note over Human,Orchestrator: ⏸ CHECKPOINT 3
    Orchestrator->>Human: final summary + suggested PR title/body
    Note over Human: review diff, open PR
```

### Who decides what

| Decision | Owner | Notes |
|----------|-------|-------|
| Is this worth building? | Human | Trigger the chain. |
| Is the story right? | Human (AI drafts) | Block at CHECKPOINT 1 if not. |
| Is the technical approach sound? | Human (AI drafts) | Block at CHECKPOINT 2 if not. The brief catches architectural mistakes cheaply. |
| Which files to change? | AI (spec-writer) | Bounded by the brief — builders cannot touch files outside this list. |
| What code to write? | AI (builders) | Bounded by the spec + path scoping rules in the per-repo context file. |
| Did the implementation satisfy the story? | AI (test-verifier + validator) | Reported to human; auto-fix loops up to 3 iterations. |
| Is the diff merge-ready? | Human | Block at CHECKPOINT 3 if not. |
| What rules to add when an agent surprises you? | Human | Edit the profile or per-repo context file. This is how the chain improves over time. |

### Why this split

- **Humans are better at:** business judgment, ambiguous trade-offs, catching missing requirements, taste.
- **AI agents are better at:** breadth (reading many files quickly), discipline (checking the same 30 things every time), patience (writing the boring fixtures and edge-case tests).
- **The chain is bad at:** anything not covered by the validator's checklist. That's why the validator's checklist is the most important prompt to keep tuning — every gap the validator misses becomes a new line in its checklist.

The three checkpoints are not bureaucracy. They are where wrong assumptions cost the least to fix. A mistake caught at CHECKPOINT 2 (brief approval) costs a re-prompt. The same mistake caught after the builders run costs hours of rework.

## Repo layout

```
ai-factory/
├── prompts/
│   ├── agents/         ← researcher, story-writer, ... (platform-neutral)
│   └── skills/         ← feature-factory, quick-fix, spike
├── profiles/           ← stack rule packs (nextjs, node-fastify, go-echo, ...)
├── src/
│   ├── cli.ts          ← CLI entrypoint
│   ├── manifest.ts     ← .factory.yaml parsing + validation
│   ├── render.ts       ← prompt/profile composition
│   ├── commands/       ← install, init, sync, feature
│   └── platforms/      ← adapters: one per AI platform
└── examples/           ← sample .factory.yaml manifests
```

> Path examples in this README assume the checkout directory is named `ai-factory`. Adjust paths if you cloned it under a different name (e.g., `ai-software-factory`).

Each project repo gets a small `.factory.yaml` manifest (~20 lines) declaring layer, stack profile, commands, paths, and target platforms. `factory install` generates the platform files.

## Install

```bash
cd ai-factory
pnpm install            # or `npm install`
pnpm link --global      # or `npm link` — exposes `factory` globally
factory --version       # should print the version
```

That makes `factory` available from any directory. Skip the `link` step if you'd rather invoke via `npx tsx src/cli.ts <command>` from inside the ai-factory checkout.

## Commands at a glance

| Command | What it does |
|---------|--------------|
| `factory init` | Interactive wizard. Creates `.factory.yaml` in the current repo. Detects stack from package.json / go.mod / pyproject.toml. |
| `factory install` | Generates platform files (`.claude/`, `.kiro/`, `AGENTS.md` + `.codex/`, etc.) for one repo, based on its `.factory.yaml` and the chosen profile. |
| `factory sync [--dry-run]` | Reads `factory.workspace.yaml` (or `--workspace <path>`) and runs `install` on every listed repo. Skips repos without a manifest; continues on per-repo failures. |
| `factory feature start <name>` | Scaffolds `<contracts-repo>/features/<name>/` with a `story.md` skeleton + empty `status.yaml`. Pass `--from <path>` to seed `story.md` from a PM-authored markdown file (Claude.ai / ChatGPT / Notion export — anything). The story is the single source of truth, shared across every implementing repo. |
| `factory feature pull <name>` | Copies the feature's `story.md` and any committed contract artifacts from the contracts repo into local `.factory/features/<name>/`. Inputs for the chain in this repo. |
| `factory feature ship <name> --contract <path>` | Marks this repo as having shipped the feature; optionally copies a local API contract back into the contracts repo. Updates `status.yaml`. |
| `factory feature list` | Lists features in the contracts repo with ship counts. |
| `factory feature status <name>` | Shows which repos have shipped a feature, when, and at what commit. |

After [installing](#install) and linking the CLI globally, invoke any of these with just `factory <command>` from any directory. Otherwise run them as `npx tsx /path/to/ai-factory/src/cli.ts <command>`.

## Usage

**Learn the design:** [`docs/book/`](docs/book/README.md) — the architecture book.
Explains the fundamentals and the *why* behind every major decision, so the system is
easy to learn and to extend. Start there if you're new or planning a change.

**Walkthroughs:**
- [`docs/walkthrough.md`](docs/walkthrough.md) — one feature through the full Tier 3 chain in a single repo. What to type at each checkpoint, what the AI returns, common mistakes.
- [`docs/cross-repo.md`](docs/cross-repo.md) — a feature spanning multiple repos via the contract bridge (backend repo → contracts → frontend repo).

In each project repo, create `.factory.yaml` (either run `factory init`, or copy from `examples/` and edit):

```yaml
name: billing-api                          # repo identifier (required)
layer: backend                             # backend | frontend | worker | mobile | fullstack (required)
profile: node-fastify                      # file in profiles/ without .md (required)
# factory-repo: omitted on purpose — the global `factory` binary knows where it is.
# Only set this if you want to pin a specific local checkout for an unusual workflow.
contracts-repo: ../ai-factory-contracts    # cross-repo contract dir (optional, Phase B)

commands:                                  # required — agents read these
  typecheck: pnpm typecheck
  lint: pnpm lint
  test: pnpm test
  acceptance: pnpm test:integration        # optional — separate acceptance/e2e command

paths:                                     # path scoping for agents (all lists optional)
  backend:                                 # Backend Builder may edit
    - src/routes/**
    - src/services/**
  frontend: []                             # Frontend Builder may edit
  migrations:                              # Migration Author may edit
    - prisma/**
  infra:                                   # DevOps Builder may edit
    - .github/workflows/**
  docs:                                    # Doc Writer may edit
    - docs/**
    - CHANGELOG.md
  shared:                                  # readable by either builder
    - packages/shared/**
  tests:                                   # Test Verifier may edit
    - tests/integration/**
  forbidden:                               # no agent may edit
    - .env*
    - "**/secrets.*"

dont-do:                                   # optional — appended to CLAUDE.md
  - Do not call the legacy /v1 endpoints.

platforms:                                 # required — which adapters to run
  - claude-code
  - kiro
  - codex

notes: |                                   # optional — free-form prose appended to CLAUDE.md
  This repo is the authoritative source for billing API contracts.
```

Then run from your project repo:

```bash
factory install        # if you ran `pnpm link --global`
# or:
npx tsx /path/to/ai-factory/src/cli.ts install
```

This reads your manifest, loads the matching profile, and writes platform-specific files (e.g., `.claude/agents/*.md` + `CLAUDE.md` for Claude Code).

## Keeping things in sync (updating)

Three layers can drift over time: the central `ai-factory` checkout, the global `factory` binary, and the per-repo generated files. Update flow:

### Refresh the central factory

```bash
cd /path/to/ai-factory
git pull                # pulls new prompts, profiles, adapter changes
pnpm install            # only if dependencies changed
```

The `pnpm link --global` from initial setup still points at this directory, so the global `factory` command picks up new code automatically — no re-link needed unless you blew away `node_modules`.

### Refresh one project

In any repo that already has `.factory.yaml`:

```bash
cd /path/to/your-project
factory install
git status                                    # see what changed
git add CLAUDE.md .claude .factory.yaml       # whatever the diff shows
git commit -m "chore: update factory artifacts"
```

That's it. Re-running install regenerates everything from the current state of the central factory. The new agents / prompts / profile rules land in place.

### Refresh many projects at once

If you maintain a workspace file listing your repos:

```bash
factory sync               # re-installs every listed repo
factory sync --dry-run     # preview what would happen
```

Useful when a central prompt or profile change needs to propagate across 5+ repos.

### What gets preserved vs overwritten

| Overwritten on every install | Preserved |
|------------------------------|-----------|
| `CLAUDE.md` | `.factory.yaml` (your manifest — never overwritten) |
| `.claude/agents/*.md`, `.claude/skills/*/SKILL.md` | `.gitignore` (your changes stay) |
| `.claude/hooks/factory-guard.mjs` + `.claude/hooks/factory-scope.json` (if `forbidden:` or any path allow-list is set) | `.claude/settings.json` — **merged, not overwritten**: only the factory's path-guard `PreToolUse` hook is added/refreshed; your other settings and hooks are kept |
| `.kiro/steering/*`, `.kiro/FACTORY.md` (if Kiro platform) | Anything else in the repo (`src/`, `tests/`, etc.) |
| `AGENTS.md`, `.codex/agents/*`, `.codex/orchestrator/*.sh`, `.codex/FACTORY.md` (if Codex platform) | `.codex/runs/**` (run history — never touched) |

**Hard rule:** never hand-edit generated files. Edit the manifest, the profile, or the central prompts — then re-run `factory install`. Otherwise your edits are lost next sync.

### When you need to change something — where to edit

| You want to change | Edit |
|--------------------|------|
| The repo's commands, paths, or repo-specific don't-do rules | `.factory.yaml` in that project repo |
| Architecture rules / conventions for a whole stack | `ai-factory/profiles/<name>.md`, then sync all repos using that profile |
| An agent's behavior (e.g., make validator stricter) | `ai-factory/prompts/agents/<name>.md` |
| The orchestration chain | `ai-factory/prompts/skills/feature-factory.md` |
| Platform-specific output shape | `ai-factory/src/platforms/<name>.ts` |

After editing anything in the central `ai-factory/`: `git push` → in each project: `factory install` (or one-shot `factory sync`).

### TL;DR

```bash
(cd /path/to/ai-factory && git pull)        # 1. refresh central
cd /path/to/your-project                    # 2. enter project
factory install                             # 3. regenerate
```

## Status

### Phase A — foundation (shipped)

- ✅ Manifest parsing + validation (`.factory.yaml`)
- ✅ Render engine (template substitution + context-file composition)
- ✅ Platform-neutral agent prompts (12 agents, 3 skills)
- ✅ Stack profiles: Next.js App Router, Node+Fastify, Go+Echo, Python+FastAPI, Bun+Hono, Quarkus Reactive (Java), React + Vite, React + rsbuild + Module Federation (micro-frontend), Python library
- ✅ **Claude Code adapter** — generates `CLAUDE.md` + `.claude/agents/*` + `.claude/skills/*/SKILL.md`, plus a `PreToolUse` path-guard hook (`.claude/hooks/factory-guard.mjs` + `.claude/hooks/factory-scope.json` + merged `.claude/settings.json`) that **enforces** the manifest's `forbidden:` list and per-agent allow-lists at the tool level — out-of-scope edits are blocked, not just discouraged by prose
- ✅ `factory install` command

### Phase B — multi-platform + multi-repo (shipped)

- ✅ **Kiro adapter** — generates `.kiro/steering/*` + `.kiro/FACTORY.md`
- ✅ **Codex CLI adapter** — generates `AGENTS.md` + `.codex/agents/*` + executable bash orchestrators in `.codex/orchestrator/*.sh` + `.codex/FACTORY.md`
- ✅ `factory init` — interactive manifest wizard with stack auto-detection
- ✅ `factory sync` — workspace-wide refresh driven by `factory.workspace.yaml`
- ✅ `factory feature start / pull / ship / list / status` — cross-repo contract bridge (MVP)
- ✅ Documentation: [`docs/walkthrough.md`](docs/walkthrough.md) (single repo) + [`docs/cross-repo.md`](docs/cross-repo.md) (polyrepo)

### Build-on-demand (not blockers)

These are deferred until you actually need them. Each is straightforward to add when the use case shows up.

**Adapters**

- ⏳ Cursor / Windsurf — not implemented. Both are rules-file (context-injection) tools with no generatable enforcement hook, so an adapter would be prompt-only (like the Kiro IDE flow). Add one by implementing `PlatformAdapter` in `src/platforms/` and registering it in `src/platforms/index.ts` — see Chapter 8 of the book.

**More stack profiles** — add by writing a markdown file under `profiles/` matching the shape of the existing profiles (architecture rules, don't-do, default commands, default paths). Likely candidates when you hit them:

- Rust + Axum / Actix
- SvelteKit (fullstack)
- Nuxt 3 (fullstack)
- Django (Python)
- React Native / Flutter (mobile)
- Ruby on Rails
- Spring Boot (Java, blocking)
- .NET / ASP.NET Core

**Chain ↔ contract-bridge integration** — currently the user invokes `factory feature pull / ship` manually around the chain. A future iteration can have the skill orchestrator auto-pull on start and auto-ship on completion.

**Contract-format validation** — ensure the backend repo's emitted contract format (OpenAPI, proto, Zod, etc.) matches what the frontend repo's spec-writer expects.

**Status locking** — protect against two developers running `feature ship` on the same repo simultaneously (rare in practice).

~~Global `factory` binary~~ — **shipped.** `bin/factory.mjs` is a tsx-spawn shim; `pnpm link --global` (or `npm link`) installs it globally so `factory <command>` works from any directory.

## How prompts work

Each prompt in `prompts/agents/` and `prompts/skills/` is platform-neutral. References to the project context document use the template variable `{{CONTEXT_FILE}}`, which the adapter substitutes at install time (e.g., `CLAUDE.md`, `AGENTS.md`, `.kiro/steering/project.md`).

Stack-specific content (commands, paths, conventions) does NOT live in the prompts. It comes from:

- **Manifest** (per-repo) — commands, paths, layer, repo-specific don't-do rules.
- **Profile** (shared) — architecture rules, conventions, don't-do, default paths/commands.

The render engine in `src/render.ts` composes manifest + profile into the platform's context file. The adapter writes that file plus the agent/skill files in the platform's format.

## How profiles work

A profile is a markdown file under `profiles/`. It contains:
- Architecture rules
- Don't-do list
- Conventions
- Default paths (documentation only — see note below)
- Default commands (documentation only — see note below)

The profile body is **inlined verbatim** into CLAUDE.md (or the platform's context file) under the `## Profile rules` section, including its own markdown headings. When you write a profile, structure it as a self-contained section because its `## Architecture rules` heading ends up nested inside CLAUDE.md's `## Profile rules`.

The "Default paths" and "Default commands" YAML blocks in the profile are **reference documentation only** — they are not parsed. The real values come from the manifest's `paths:` and `commands:` blocks. Treat them as the suggested starting point a manifest author should copy.

To add a profile for a new stack:
1. Create `profiles/<your-stack>.md`.
2. Follow the structure of `profiles/nextjs-app-router.md` as a template.
3. Reference it in a manifest with `profile: <your-stack>`.

## How adapters work

Each adapter implements `PlatformAdapter` in `src/platforms/index.ts`:

```ts
export interface PlatformAdapter {
 name: Platform;
 contextFileName: string;
 generate(args: {
 targetRoot: string;
 manifest: Manifest;
 agents: PromptFile[];
 skills: PromptFile[];
 profileBody: string;
 }): Promise<PlatformWriteResult>;
}
```

`src/platforms/claude-code.ts` is the reference implementation; `kiro.ts` and `codex.ts` are the other two. To add a platform, implement `PlatformAdapter` and register it in `src/platforms/index.ts`.

### Enforced path scoping (Claude Code)

On Claude Code, path scoping is **enforced**, not just advised:

- The `forbidden:` list is blocked session-wide by a `PreToolUse` hook
  (`.claude/hooks/factory-guard.mjs` + a merged `.claude/settings.json`).
- Each editing agent (`backend`, `frontend`, `tests`, `migrations`, `infra`,
  `docs`) gets a per-agent `PreToolUse` hook in its frontmatter that blocks
  edits outside its allow-list. Lists are **opt-in**: an agent with no list in
  the manifest is unenforced (prompt-only); an empty list means "edit nothing".

> **Opt-in / upgrading existing repos:** the guard is generated from the keys
> present in *your* `.factory.yaml`, which is never overwritten by `install`. A
> manifest written before these keys existed gains enforcement only for the keys
> it already has. To enforce `migrations`/`infra`/`docs` in an existing repo, add
> those keys to `.factory.yaml` (copy the profile's "Default paths" as a starting
> point) and re-run `factory install`.

Limitations: enforcement covers `Write`/`Edit`/`MultiEdit`/`NotebookEdit` only —
a builder's `Bash` access can still write files, so the guard is a guardrail, not
a sandbox. Kiro and Codex have no hook mechanism, so they remain prompt-only.

## Cross-repo coordination

For features that touch multiple repos (e.g., backend repo emits a contract, frontend repo consumes it), the factory uses a separate **contracts repo** as the bridge — the place where the user story lives once and where the API contract is exchanged between repos.

**Workflow:**

```bash
# 1. In the backend repo (or wherever the story originates):
factory feature start invoice-reminders
# → scaffolds <contracts-repo>/features/invoice-reminders/story.md + status.yaml
# Edit story.md, then commit + push the contracts repo.

# 2. In the backend repo, pull the (now-committed) story:
factory feature pull invoice-reminders
# → copies story.md into .factory/features/invoice-reminders/
# Then run the chain (Tier 3) referencing that story as input.

# 3. When the backend chain produces an API contract artifact:
factory feature ship invoice-reminders \
  --contract docs/api/invoice-reminders.openapi.yaml \
  --commit $(git rev-parse HEAD)
# → copies the contract into the contracts repo
# → appends this repo to status.yaml's shipped list
# Commit + push the contracts repo.

# 4. In the frontend repo:
factory feature pull invoice-reminders
# → pulls story.md AND api.openapi.yaml into .factory/features/<name>/
# Run the chain with the story + contract as inputs.

# 5. When the frontend ships:
factory feature ship invoice-reminders --commit $(git rev-parse HEAD)
# → marks this repo shipped in status.yaml
```

**On-disk in the contracts repo:**

```
<contracts-repo>/
└── features/
 └── invoice-reminders/
 ├── story.md # authored once, shared
 ├── api.openapi.yaml # backend writes; frontend reads
 └── status.yaml # append-only ship log
```

See [`docs/cross-repo.md`](docs/cross-repo.md) for the full worked example with two repos and the orchestrator skill flow.

**Current scope:** The chain *consumes* a pulled feature bundle — the orchestrator reads `story.md` and any published `api.*` contract from `.factory/features/<name>/`, and the spec-writer/frontend-builder build on that contract. What's still manual is the **pull/ship invocation itself**: you run `factory feature pull` / `ship` by hand around the chain; the orchestrator doesn't auto-pull on start or auto-ship on completion. Auto-pull/ship is a follow-up (see "Build-on-demand" under Status).

## License

MIT License

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

ai-software-factory

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README