Planr Agent Instructions

Planr is being shaped into a local-first planning and execution coordination tool for Codex, Claude Code, Cursor, and generic MCP clients.

Use the product plan package in .planr/plans/product/planr/ as the product source of truth for new implementation work:

PRODUCT_SPEC.md for scope and non-goals.
TECH_ARCHITECTURE.md for ownership boundaries.
API_AND_DATA_MODEL.md for project, plan, map, item, log, review, and API contracts.
TASKS.md for executable coding-agent tasks.

Frozen contracts live in docs/contracts/.

Core rules:

Build Planr as a self-owned product with original implementation, docs, and naming.
Use the public product flow: idea -> product plan -> build plan -> map -> pick -> log -> review/evidence -> recovery/package -> close.
Keep map graph state authoritative for item status, links, picks, reviews, approvals, and completion.
Keep product and build plans first-class for rich scope, ownership, verification, and narrative context.
Require log-backed closure: files changed, commands run, tests/review results, and blocked or unverified items.
Support Codex, Claude Code, and Cursor as peers through MCP and CLI workflows.
Never modify, replace, or re-upload an already published release. Fix release defects only by bumping and publishing a new version.

Proofloop contract

Project profile: .proofloop/project-profile.json.
Sol is the sole writer and final decision owner.
Luna is disabled unless a future run contract explicitly enables a read-only role; the current run uses luna_mode: null.
BUILD may change production code and product-plan contracts, but not tracked tests or run a full suite.
HARDEN requires a separate run contract and admits at most one permanent invariant into one existing test file.
Validation commands and evidence paths come only from the active run contract.