What happened
The repository describes “Software Factory for Claude Code (v2),” a workflow that starts with GitHub issues labeled factory. Running /factory run drains the open queue, while /factory run <n> runs one issue. Intake assesses whether an issue is worth building and at what scope, can request more information or reject it, and rates risk and surfaces. A script then routes the job to reviewers and models using config.json.
Accepted work moves through a spec writer, builder, checks gate, reviewers, and approver. The builder writes code, commits and pushes a branch, then opens a draft pull request intended to close the issue. The gate runs the project’s factory/checks.sh; if it fails, reviewers do not run. Reviewers can request changes, sending the work through another round. On approval, the approver synchronizes with main, runs the gate again, and merges with gh pr merge. If the approver escalates, the pull request is left for a person and the issue receives a factory:needs-human label.
The included roles are intake, spec-writer, builder, code reviewer, security reviewer, UX reviewer, UI reviewer, and approver. Reviewer selection is conditional: UX and UI reviewers are added for UI changes, while the security reviewer runs when intake flags security. One Claude Code session orchestrates jobs, with concurrency configurable and a default of three. Each job gets its own git worktree and branch. Job state lives under factory/jobs/, allowing /factory run to resume work after a session ends.
The installation instructions require Claude Code 2.1.257 or later, Python 3, and an authenticated gh account able to push to the repository. The installer copies the skill, agent definitions, scripts, and dashboard, then asks the user to fill in project-specific test, lint, typecheck, and build commands in checks.sh, tailor reviewers, create GitHub labels, and verify that checks pass on main.
The README describes examples from a test run on a small private calculator repository, using concurrency 2 and Haiku for every agent. A low-risk function change passed with only code review; a UI change also used UX and UI reviewers. Another task was narrowed by intake, failed a gate because the builder edited a protected test file, and went to a human after three rounds. A crypto-miner issue was rejected at intake. The project is GPLv3; it says closed-source or commercial use requires a commercial license. It advises starting at concurrency 1 and reviewing every PR until the workflow is trusted.
What the archive adds
New Runtime covered Claude Code Game Studios on 2026-11-06, a distinct Claude Code setup organized as a game-development studio. This repository is a new event, applying role-based orchestration to issue-to-PR software work rather than game production.
Trend signal
This supports AI-native organizations move toward review and orchestration: the factory routes work, applies checks, delegates review, and reserves merge or escalation for accountable humans. The repository makes that operating model concrete in a single-session workflow, while its examples remain a small test run rather than evidence of broad organizational results.
Why now
The repository is explicitly designed around GitHub issues and Claude Code, with install steps that adapt the factory to a project’s own checks. That makes the design relevant to teams exploring coding agents as an operational workflow, not merely as an interactive coding assistant. Its recommendation to start with one concurrent job signals that trust and review capacity should be established before increasing throughput.
Why it matters
The useful design choice is not simply assigning different prompts to agents. The workflow puts boundaries around what they can do: only labeled issues enter; repository-specific checks gate review; jobs have isolated worktrees and branches; and a human merges approved changes. Failed checks halt downstream review, while unresolved work can be escalated rather than forced through an automated success path.
Teams can adapt this structure without copying every role or model choice. The most important prerequisite is a reliable project check suite: the factory’s gate is only as meaningful as the tests, lint, typecheck, and build commands configured for it. The README’s example of a protected-test-file violation illustrates why the gate and human review need to remain part of the loop. The repository offers a practical starting point, not proof that issue-to-PR automation is safe for every codebase.
Limits and uncertainty
The reported examples come from a small private calculator repository and a deliberately cheap test configuration with all agents on Haiku; they do not establish performance on larger or production systems. The repository does not report aggregate success rates, review time saved, defect rates, or costs across a representative workload. Its dashboard’s cost figure is described as API-equivalent agent pricing read from Claude Code transcripts; on a subscription plan, it indicates spend allocation rather than the actual bill. Teams must also validate their own checks, permissions, routing, and failure handling before relying on the workflow.
Skill to build now
This week, make a small issue-to-PR pilot on a low-risk repository. Configure checks.sh with the project’s real tests, lint, typecheck, and build commands; confirm the checks pass on main; then run with concurrency 1 and review every pull request. Keep merge permission with a human, and record which failures trigger escalation before adding more issue types or parallel jobs.
Reader takeaway
ZazenCodes offers a concrete Claude Code workflow for automating the path from GitHub issue to reviewed pull request. Its strongest lesson is to automate bounded steps behind project-specific checks, isolated job state, and human approval—not to treat an agent’s confidence as evidence that a change is ready to merge.