project-rules-init

Project Rules Init

Create compact canonical rules from source evidence. Source-only means source, effective configuration, tests, and observed commands; it excludes Git, PR, release, owner, deadline, and collaboration policy.

Required resources

Read ../../references/project-context.md, ../../references/candidate-ranking.md, ../../references/language-analyzer.md, ../../references/pattern-card.md, ../../references/coverage-contracts.md, ../../references/conflict-and-lifecycle.md, ../../references/code-chain-discovery.md, ../../references/rule-classification.md, ../../references/rule-content-contract.md, ../../references/confirmation-policy.md, ../../references/adapter-content-contract.md, and ../../references/adapters.json. Derive one to three task_terms from the current user source task and target scope (for example API, module, or feature names). If no concrete Init feature was named, use the source dimensions or module-role terms to be generated. Use one repeated --task-term <term> for each term. Save scan JSON in a system temporary file explicitly outside <project-root> so AI can read it during synthesis; use try/finally and Remove-Item after validation. In PowerShell: $scanResult = Join-Path ([IO.Path]::GetTempPath()) ('project-rules-scan-' + [guid]::NewGuid() + '.json'); try { python ../../scripts/scan_project.py <project-root> --task-term <term-1> --task-term <term-2> > $scanResult; Get-Content -Raw $scanResult; ... } finally { Remove-Item -LiteralPath $scanResult -Force -ErrorAction SilentlyContinue }. Use ../../scripts/inspect_symbol.py for material Python ownership. If a valid manifest v2 exists, stop and use project-rules-update.

Init workflow

1. Scan read-only and build the bounded source-only context payload. Review existing rules and similar documents as candidate evidence, then produce a ranked source-candidate list; preserve sensitive paths as existence-only.
2. Use the language analyzer to verify primary and comparable anchors, then trace representative code chains. Tools provide facts/candidates/conflicts; AI synthesizes evidence; humans decide substantive conflicts, strong constraints, and future direction.
3. Create only anchored pattern cards and lightweight coverage statuses. AI must cite primary and comparable anchors before writing a repeated pattern. Do not import language defaults, documentation, or a single new file as a rule.
4. Ask only when existing rules or similar documents clearly conflict with effective source/config evidence, express a strong constraint, or propose a future direction. Present those conflicts, constraints, and future-direction choices to the human. Weak evidence is pending or omitted; in all modes, leave unresolved claims out of canonical content.
5. Show a compact preview and exact create / manual-only plan. Request one explicit write confirmation; review always precedes write.
6. After approval, atomically write owned canonical files and adapters, manifest v2 last, then within that try run python ../../scripts/validate_outputs.py <project-root> --scan-result $scanResult; the finally cleanup runs after validation.