Branchline — release rehearsal for humans and coding agents
Branchline fills the gap between code review and incident response. It turns a real local Git diff or public HTTPS repository into a release-impact rehearsal, so engineering teams can see the trade-offs before a deployment reaches users.
How it works
- Map evidence, not intuition. Branchline reads the selected Git boundary and identifies changed contracts, source consumers, tests, configuration, dependencies, and migrations. Every signal is linked to committed repository evidence; it never executes target code.
- Rehearse decision paths deterministically. A release captain can test full rollout, canary, compatibility adapter, and rollback paths. The state changes are transparent scenario rules—not production telemetry, forecasts, or opaque AI scores.
- Challenge the path. Branchline seals selected, redacted evidence into an immutable Release Council packet. Contract, test/observability, rollout, and security specialists return reports that are schema-validated against that exact hash. Disagreements remain visible.
- Keep the human accountable. A human records approved, blocked, accepted-risk, or request-evidence decisions with rationale and follow-ups. The durable decision ledger is included in Markdown and JSON exports.
- Make it enforceable and portable. A flat policy file can require compatibility safeguards, test confidence, Council coverage, and a human decision in local workflows or GitHub Actions. Portable skills, native Codex/Claude Code plugins, and a read-only MCP server bring the same evidence into agent harnesses.
Why Codex and GPT-5.6
Codex accelerated the full product build: architecture, fixture design, Git analysis, deterministic simulation, persistence, UI, automated verification, portable agent integrations, and the demo workflow. GPT-5.6 is used through the Branchline Codex skills and read-only MCP surface to interpret narrow, redacted evidence and return reviewable mitigation or specialist perspectives. Neither model changes deterministic risk metrics, writes to a repository, deploys code, or makes the release decision.
Judge-ready path
No account is needed. Clone the repository, run npm install, npm run fixture:reset, and npm run dev, then connect the included two-commit fixture. The repository includes clear setup instructions, sample data, distribution checks, policy/CI examples, and an under-three-minute product demo source.
Built With
- codex
- github-actions
- gpt-5.6
- model-context-protocol
- next.js
- node.js
- react
- remotion
- simple-git
- sqlite
- typescript
- vitest
