repo-context-ledger / rules

---
description: Route Cursor through the repository's verified, cross-Agent context ledger.
alwaysApply: true
---

Read and follow the repository root `AGENTS.md`. Use already loaded task context first. For a missing or changed background question, run `context --query "<task>"` before broad documentation exploration; do not repeat routing for a routine follow-up whose context is sufficient. Read only the Context Bundle's Required reads initially. Never recursively read `docs/ai`, `docs/specs`, or `docs/changes`. Do not open completed Change bodies unless the plan selects one, a required Pack cites one for a named reason, or the user asks for historical reasoning. Required reads are a starting route, not a maximum code-reading limit. Expand context only after stating the unresolved question, and always expand whenever callers, implementations, configuration, persistence, permissions, concurrency, retries, tests, or external boundaries can affect the behavior. A same-window 'continue' on a known active task keeps its session and epoch without another resume. For a real window/tool handover, paused task, or missing task identity, resolve only the missing state and use an owned unique Resume Capsule. Continue the same Ledger session with `resume --query "<keywords>" --tool cursor`; keep the returned continuation epoch and pass `--epoch <n>` to later writes. Never read, resume, pause, checkpoint, finish, or invalidate another principal's private session unless an explicit unexpired grant authorizes that exact access. Read-only grants never permit writes. A foreign overlap without a grant permits only Git-tracked Pack/spec/Change guidance. Treat Cursor Memory as a private cache, never as the repository source of truth. Run ledger lifecycle commands autonomously; never delegate them to the user. Never message or steer another user-owned task unless the user explicitly requested cross-task coordination.