repo-context-ledger: Instructions file for Claude Code

Claude Code instructions

@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 claude; 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 Git-tracked Context Packs, stable specs, and handoffs as the cross-Agent source of truth. Follow the repository context ledger workflow without asking the user to run lifecycle commands. Never message or steer another user-owned task unless the user explicitly requested cross-task coordination.