CodeZero is a GitHub-native engineering agent platform for turning product intent into production-ready pull requests. Open an Issue, mention the agent, or trigger a repository workflow; CodeZero drafts the PRD, understands the codebase, plans the smallest safe change, writes the code in an isolated sandbox, verifies the result, reviews the diff, and opens a draft PR with local validation steps.
The idea is simple: humans write the intent, AI handles the code path. Not as a fragile one-shot prompt, but as a traceable engineering system with durable workflows, multi-agent orchestration, repository intelligence, sandbox execution, quality gates, memory governance, and human approval where it matters.
- Issue to PRD to PR: convert GitHub Issues into structured PRDs, implementation plans, verified diffs, and draft PRs.
- Zero-code operator flow: product or engineering leads describe what should change; agents handle the coding loop.
- Repository intelligence: build a Repo Navigation Graph and evidence-backed ContextPack before editing.
- Isolated execution: each Issue runs in its own sandbox, branch, artifact set, and quality gate trail.
- Human control: PRD approval, policy gates, review subagents, and memory update proposals keep the system inspectable.
- Local verification built in: generated PRs include checkout, install, test, and run instructions.
- OpenAI-compatible providers: designed for OpenAI, DeepSeek, Qwen, or any compatible model gateway.
- Operator console: Run Console, Settings Console, Memory Inbox, Trace Replay API, and Golden Issue Eval CLI.
flowchart TD
GH["GitHub Issue / Comment / Label"] --> TP["Repository Trigger Policy"]
TP --> API["Fastify Webhook API"]
API --> WF["Durable Workflow Orchestrator"]
WF --> PRD["PRD Agent"]
PRD --> GATE{"Human approval required?"}
GATE -->|Yes| UI["Web Review Board"]
UI --> WF
GATE -->|No| SB["Per-Issue Sandbox"]
WF --> SB
SB --> IDX["Codebase Intelligence"]
IDX --> GRAPH["Repo Navigation Graph"]
GRAPH --> MEM["Approved Memory Retrieval"]
MEM --> CP["Evidence-backed ContextPack"]
CP --> PLAN["Minimal Change Planner"]
PLAN --> IMPL["Implementation Agent"]
IMPL --> QA["Quality Gates"]
QA --> REV["Review Subagent"]
REV --> PRW["PR Writer"]
PRW --> PR["Draft PR + Local Verification"]
PR --> LEARN["Memory / Project Map Proposal"]
- Trigger: GitHub webhook,
@agentmention, label, or manual import creates a task. - Understand: the PRD agent extracts goals, risks, acceptance criteria, and complexity.
- Plan: repository indexing, navigation graph, approved memory, and ContextPack narrow the work area.
- Implement: the implementation agent makes the smallest safe code change in an isolated sandbox.
- Verify: build, lint, test, typecheck, screenshot hooks, policy checks, and review subagents run before PR creation.
- Publish: CodeZero pushes a branch and opens a draft PR with evidence, risks, and local verification commands.
apps/
api/ Fastify API, GitHub webhooks, settings and task routes
web/ Next.js Run Console, Settings Console and Memory Inbox
worker/ Workflow worker and repository task execution
packages/
agent-runtime/ model provider and structured agent primitives
codebase-intelligence/ indexing, hybrid search, ContextPack and repo graph
config/ YAML config loading and validation
github/ GitHub Issue, branch and PR integration
memory/ approved memory and memory proposal store
orchestrator/ task state machine and workflow decisions
persistence/ file/Postgres task persistence
sandbox/ Docker/worktree sandbox abstraction
skills/ platform skill loader and built-in skills
tool-gateway/ audited tool execution boundary
verification/ test, screenshot and local verification helpers
workflows/ issue-to-PR workflow composition
pnpm install
cp .env.example .env
cp config/agents.example.yaml config/agents.yaml
cp config/repositories.example.yaml config/repositories.yaml
cp config/sandbox.example.yaml config/sandbox.yaml
cp config/policies.example.yaml config/policies.yaml
cp config/tools.example.yaml config/tools.yamlEdit .env with an OpenAI-compatible model provider and GitHub token:
OPENAI_BASE_URL=https://api.openai.com/v1
OPENAI_API_KEY=...
OPENAI_MODEL=...
GITHUB_TOKEN=...
GITHUB_WEBHOOK_SECRET=...
AGENT_TRIGGER_MENTION=@agent-prdStart local dependencies and services:
docker compose -f infra/docker/docker-compose.yml up -d
pnpm dev:api
pnpm dev:worker
pnpm dev:webOpen the web console at http://localhost:3000.
pnpm check
pnpm eval:goldenpnpm check runs lint, typecheck, tests, and build. pnpm eval:golden scores the sample golden issues in evals/golden-issues against candidate artifacts and writes the report to artifacts/eval-report.md.
Runtime configuration lives in YAML files under config/:
agents.yaml: model providers, agent roles, and routing.repositories.yaml: repository trigger policy, queue limits, and permissions.sandbox.yaml: execution mode, workspace paths, and sandbox settings.policies.yaml: approval rules, blocked paths, and guardrail policy.tools.yaml: tool gateway permissions and timeout defaults.
The Web Settings Console can edit and validate these files during local operation.
- Documentation Index
- System Architecture
- Workflow Blueprint
- Operations Guide
- Product Requirements
- Repo Navigation Graph
- Codebase Intelligence
- Memory Architecture
- Prompt and Skill Design
Historical planning notes are kept in docs/archive.
The MVP runs locally and includes GitHub Issue ingestion, repository trigger policy, repository queue and concurrency limits, PRD generation, human PRD approval, Repo Navigation Graph MVP, ContextPack generation, Tool Gateway JSON action fallback, Trace Replay API, Run Console, Settings Console, Memory Inbox, Golden Issue Eval CLI/CI, Repository Onboarding, sandbox execution, quality gates, Review subagent, and draft PR creation.
Next hardening areas include approval recovery, stricter tool input schemas, security scanning, richer eval assertions, and deeper graph adapters for larger repositories.