REIN (Regulated Execution and Inference Navigation)

REIN is a lightweight and hands-off harness for stricter, more explicit engineering workflow with Codex, Claude Code, and Cursor.

It adds:

• REIN.md for regulated execution and inference guidance
• packaged skills for clarification, inspection, cleanup, triage, verification, and retrospection
• .rein/ as the repo-local artifact root for REIN-generated outputs
• rein index for a local, rebuildable evidence vector index over REIN artifacts

[!NOTE] New in REIN: rein go is the runtime-backed end-to-end flow. It starts the same user-facing interview that rein-interview uses, then carries the task through planning, implementation, cleanup, review, and verification with durable state under .rein/.

[!IMPORTANT] New vector-backed evidence recall: run rein index build to give REIN a local, rebuildable index over protocol docs, skills, codebase maps, completed interviews, specs, and rein go artifacts. When the index exists, rein go automatically attaches matching evidence to plan and stage artifacts with source paths, line numbers, and stale-source status.

New: rein go

If you want one REIN-controlled workflow instead of manually invoking each stage, start here:

rein go "add a CSV export for invoices"

You can then inspect or continue the flow with:

rein go status --slug <slug> --json
rein go resume --slug <slug> --json

If you are using Codex or Claude Code, the same flow can be exposed through wrapper entrypoints like $rein-go and /rein-go.

What rein go gives you:

• the normal user-facing REIN interview, including the clarity score and interview question flow
• durable runtime state for the full task under .rein/
• a runtime-backed path from interview -> plan -> implementation -> cleanup -> review -> verify
• rein go status, rein go resume, and rein go advance for inspecting or continuing the flow
• the option to start from an already-completed interview bundle with --from-interview
• automatic evidence context in plan and stage artifacts when a local rein index is available

New: rein index

REIN can build a local evidence index from protocol docs, skills, codebase maps, completed interviews, specs, and rein go stage artifacts:

rein index build
rein index query "what constraints matter before I edit this?"
rein index status

The index lives under .rein/index/ and is always rebuildable from source artifacts. It uses deterministic local feature-hash vectors, BM25 lexical scoring, and REIN-specific pressure signals like constraints, acceptance criteria, decision boundaries, rejected alternatives, and verification notes. Query results include source paths, line numbers, and stale-source filtering so retrieved context stays evidence, not hidden state.

When the index exists, rein go automatically queries it while generating the plan and later stage instruction artifacts. If the index is missing or stale, the flow still proceeds, but the generated artifact records that evidence status and filters stale sources instead of treating old retrieval results as active context.

Install

npm install -g @jstxn/rein
rein init

Interactive install lets you choose:

• which tool to target (Codex, Claude Code, Cursor, or all)
• the current repository, another repository, or your user-level setup

Non-interactive examples:

rein init --repo                        # default (Codex)
rein init --repo --claude               # Claude Code only
rein init --repo --cursor               # Cursor only
rein init --repo --codex --claude       # both
rein init --repo --codex --claude --cursor  # all
rein init --repo --link                 # advanced: symlink packaged assets instead of copying
rein init --repo /path/to/repo --claude
rein init --user --claude

Core flow commands:

rein go "..."          # start the runtime-backed REIN go flow from a fresh task
rein go --from-interview <slug|path> --json
rein go status --slug <slug> --json
rein go resume --slug <slug> --json
rein go advance --slug <slug> --stage <stage> --status <status> --json
rein index build
rein index query "what prior constraints apply here?" --json
rein index status --json

Direct stage/runtime commands:

rein status              # show what's installed and whether it's outdated
rein update              # re-install REIN surfaces, replacing existing files
rein remove              # uninstall REIN from the repo (interactive)
rein remove --yes        # uninstall without prompts (preserves .rein/ artifacts)
rein interview init --idea "..."             # start a durable interview
rein interview status --slug my-topic        # inspect live interview state
rein interview resume --slug my-topic        # resume a saved interview
rein interview crystallize --slug my-topic --summary '{"intent":"Clarify the implementation scope","desiredOutcome":"A verified implementation plan","transcriptSummary":"The interview captured scope, constraints, and acceptance criteria.","acceptanceCriteria":["Required checks are documented"],"constraints":["Keep the change scoped"],"decisionBoundaries":["Decide whether to plan or implement next"],"executionBridge":["plan"],"inScope":["Crystallize the interview result"],"outOfScope":["Implementing the follow-up task"],"technicalContext":["Artifacts are written under .rein/specs"]}'

The --codex, --claude, and --cursor flags work with status, update, and remove too. Without flags, these commands auto-detect which tools are installed.

rein init --link and rein update --link are advanced options for users who want linked installs. Linked installs update from one source checkout, but they depend on that source path continuing to exist.

What rein init adds

| Surface | Codex (--codex, default) | Claude Code (--claude) | Cursor (--cursor) | |---|---|---|---| | Protocol doc | REIN.md | REIN.md | REIN.md | | Artifact root | .rein/, .rein/codebase/ | .rein/, .rein/codebase/ | .rein/, .rein/codebase/ | | Skills | .codex/skills/ | .claude/commands/ | .cursor/rules/ | | Guidance block | AGENTS.md | CLAUDE.md | AGENTS.md |

All targets install the same 10 skills:

• rein-go
• rein-interview
• rein-inspect
• rein-triage
• rein-plan
• rein-scope
• rein-review
• rein-cleanup
• rein-verify
• rein-retro

By default, rein init copies these surfaces into the target. With --link, REIN symlinks the packaged docs and skills instead, while still writing repo- or user-specific guidance files normally.

How to use it

Typical workflow:

1. Run rein init in the repo you want to use.
2. Start a new Codex, Claude Code, or Cursor session in that repo.
3. Let AGENTS.md / CLAUDE.md route work through REIN.md.
4. Use:
   • rein-go when you want REIN to carry one task through the full runtime-backed flow
   • rein-interview when you want the clarification stage by itself
   • rein-inspect for a durable repo map
   • rein-triage before ambiguous or multi-file changes
   • rein-plan to break complex work into sequenced steps with checkpoints
   • rein-scope when requirements are too large, conflicting, or need negotiation
   • rein-review to self-review your diff before committing
   • rein-cleanup for cleanup/refactor work after behavior is locked
   • rein-verify before declaring completion
   • rein-retro after misses or suspicious shortcuts

Inspiration

• Fear-Regulated Pressure and Verification-First Prompting -- the research paper behind REIN's protocol design
• andrej-karpathy-skills -- concise behavioral guidelines derived from Andrej Karpathy's observations on LLM coding pitfalls

Notes

• REIN is a workflow harness, not a benchmark or proof of superiority.

License

MIT