Automation that turns GitHub issues into merged PRs with human oversight, powered by Claude Code.
Most of the time, you already know what needs to change — the bug, the feature, the refactor. What you don't have is the hour to sit at a terminal and shepherd an AI through it. Code Factory moves the entire interaction to GitHub: describe the work in an issue, review an AI-generated plan on a draft PR, leave comments, and merge when you're satisfied. GitHub becomes the interface — no terminal session required.
Issues in, merged PRs out, with you in the driver's seat the entire time.
# Continuous polling (runs until stopped — best in a tmux/screen session)
tmux new -s factory
python3 code_factory.py --repo owner/repo
# Single pass (process one item and exit)
python3 code_factory.py --once --repo owner/repo
# Auto-detect repo from current directory
cd /path/to/your-repo
python3 /path/to/code_factory/code_factory.py --onceA single Python script checks a GitHub repo for actionable work, then orchestrates Claude Code to handle it. Every change goes through a plan-first workflow — Claude proposes a plan as a draft PR, waits for human review, and only implements after approval.
- Claim — picks an unassigned issue and self-assigns
- Plan — creates a draft PR with an implementation plan
- Review — waits for human feedback on the plan (see Reviewing Plans)
- Implement — writes the code using TDD after plan approval
- Verify — runs tests and CI checks
- Code Review — waits for human code review (see Reviewing Code)
- Merge — after human code review approval
When the bot creates a draft PR with a plan, leave a comment on the PR to provide feedback. Do not use GitHub's "Approve" review button — the bot reads PR comments, not review approvals.
| What you want | What to comment |
|---|---|
| Approve the plan | LGTM, looks good, ship it, approved |
| Approve with tweaks | Looks good, but also add tests for X |
| Request major changes | This approach won't work because X. Instead, try Y. |
| Ask a question | What about edge case X? |
The bot will classify your comment and either start implementing, revise the plan, or reply with clarification.
After implementation, the bot marks the PR ready for review and adds the bot:review-requested label. At this stage, use GitHub's normal code review workflow — leave review comments, request changes, or approve. The bot will address review feedback or merge after approval.
Existing work is always finished before new work is started:
| Priority | What | Action |
|---|---|---|
| 1 | PRs with code review feedback | Address reviewer comments |
| 2 | PRs with plan feedback | Incorporate feedback or proceed |
| 3 | Accepted plans | Implement the approved plan |
| 4 | Unclaimed issues | Claim and propose a plan |
- Claude Code CLI installed
ghCLI authenticated with repo permissions- Python 3
# Continuous loop — polls every 5 minutes, dispatches work automatically
python3 code_factory.py --repo owner/repo
# Single pass — find one item, process it, exit
python3 code_factory.py --once --repo owner/repo
# Auto-detect repo from current directory
python3 code_factory.py --oncecode_factory.py # Orchestrator: poll, route, manage phases, invoke Claude Code
prompts/ # Prompt templates for LLM-dependent phases
phase1_claim_and_plan.md
phase2_process_feedback.md
phase4_implement.md
phase6_process_review.md
phase6_apply_fixes.md
skills/
git-contribute/
SKILL.md # Claude Code skill for interactive invocation
TROUBLESHOOTING.md # Diagnostics and manual fix recipes
docs/ # Design specs and implementation plans
tests/
test_code_factory.py # Unit tests
The workflow uses these GitHub labels (created automatically on first run):
bot:plan-proposed— draft PR with a plan awaiting human reviewbot:plan-accepted— plan approved, ready for implementationbot:in-progress— PR currently being processedbot:review-requested— implementation complete, awaiting code review