Agents

Nirecode: The self-driving agent coding framework — secure by default, reliable by design
pronounced "ni-re-code" · for Claude Code and GitHub Copilot

Why Nirecode?
Self-driving by design

- A self-driving development harness — hooks enforce research → tests → code → security → docs as a per-session state machine. Concurrent sessions are tracked independently; any session resumes seamlessly across machines.
- Worktree-based parallel sessions — Run multiple features concurrently without stepping on each other. Writes to the main worktree are blocked by default. See docs/parallel-sessions.md.

Built for output quality

- Two AI providers, one planning loop — Claude drafts and Codex (OpenAI) reviews adversarially, turn by turn, until both agree. The blind spots one model carries, the other catches.
- TDD — tests are written before implementation code. A dedicated subagent handles test authoring in isolation, and the commit hook blocks merging until the test step is satisfied.
- Test cases and security review are scoped to OWASP categories — codified, not optional.

Cross-platform, zero exceptions

- Windows-native, not an afterthought — Claude Code skews Linux/macOS. This framework ships PowerShell-first installers, hooks, and shell conventions so Windows developers get the full workflow without workarounds.
- One installer, every platform — Linux, macOS, and Windows (native and WSL2) all handled by a single branching install; hooks and shell rules detect the platform automatically.

Hook-enforced end-to-end workflow

Most agent frameworks rely on the model to remember process steps. This framework encodes
the dev workflow as a deterministic next-step-driven state machine. After each skill completes,
the model queries bin/workflow/next-step for the next step; a…