Rigor distributes Skills, hooks, CLI and templates. It does not need to be an MCP service or replace GitHub, Exa, Hugging Face or PaperMeld. Those tools are used only when their evidence is relevant; available equivalents can resolve the same uncertainty.
| Layer | Owns | Avoid |
|---|---|---|
| Rigor plugin | Reusable research/implementation/validation/cleanup/recovery workflows and deterministic tools | Per-project paths, fixed datasets/metrics, global user settings |
| Global AGENTS.md | Language, autonomy, truthful reporting, short Rigor routing, user boundaries | Duplicating the plugin manual or forcing every skill every turn |
| Project AGENTS.md | Actual commands/runtime, data/model contracts, acceptance, resource and path ownership | Repeating global preferences, universal ML settings copied from another repository |
| DESIGN/HANDOFF/MEMORY | Current design, continuation point, verified durable facts respectively | Using AGENTS as a growing experiment diary |
Distributable templates: templates/global.md and templates/deep-learning.md. Copy only after substituting actual project commands and constraints. No global/project file is overwritten by installing the plugin. Existing applicable instructions keep their precedence. If old AGENTS mandates every research/memory gate, remove that duplication deliberately when adopting the new global template; a plugin cannot override higher-priority instructions.
Project setup creates .codex/rigor.json schema 2 with protected_paths, sensitive_paths, memory and library. Paths in protected/sensitive arrays are project-relative directory or file scopes (not globs). Configure library { "root": "...", "code_root": "..." } and memory only when used. Empty objects are valid. Sensitive structured edits need a tracked scope; defaults protect .git/.codex from cleanup but do not block normal tasks. No universal ML acceptance profile.
For old Rigor: preserve the old policy, review each useful value and create schema 2 deliberately. Unsupported keys produce errors; setup never overwrites. Keep old runtime cache/state until its tasks have been reviewed; do not treat it as task garbage.
Native role templates under templates/agents are optional explicit examples, not installed automatically. Current runtime presets may already cover every role. Adding a role requires merging a named [agents.<name>] entry and its config_file with the host's actual schema, without overwriting existing roles. Role files request model/effort; they do not establish a sandbox. Prefer the per-call new exec route when read-only runtime enforcement is necessary, and still treat external connector scope separately.
Source directory is user-named Rigor; distribution identifier is rigor. Package with scripts/package.py OUTPUT.tar.gz; its root is lowercase rigor and excludes Git/private state/build caches. Validate unpacked manifest with the official plugin validator. Register/install with the host's current official plugin CLI/UI, then explicitly trust hooks and start a fresh session. Do not assume editing source updates installed cache. For developer iteration use the official plugin-creator cachebuster/reinstall workflow.
Hooks provide short route/continuity/assignment guidance and narrow known-structured-edit guards. Ordinary shell/MCP internals, disabled hooks or unavailable event payloads are not enforced. Unknown structured response success is not accepted as proof. There is no universal shell security guarantee. See docs/validation.md for what was actually tested on the host.