Lucas RayanandClaude Sonnet 5.5
b536ed13e7
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
45 lines
2.9 KiB
Markdown
45 lines
2.9 KiB
Markdown
# OpenWolf Operating Protocol
|
|
|
|
You are working in an OpenWolf-managed project. These rules apply every turn.
|
|
|
|
OpenWolf's hooks handle the bookkeeping: they maintain `.wolf/anatomy.md` and `.wolf/memory.md` after writes, track reads, and surface anatomy hints when you read files. Do not update those two files manually unless your agent has no OpenWolf hooks installed (Gemini CLI, Cursor).
|
|
|
|
## STATUS.md: read first, keep fresh
|
|
|
|
`.wolf/STATUS.md` is the handoff document. Read it FIRST when resuming a session; it replaces re-reading memory, plans, and code to reconstruct context.
|
|
|
|
Keep it fresh: when the user signals a quest is done ("done", "ship it", "next phase", "/clear", "wrap up"), move finished items to the done section, write the next quest (objective, files, decisions), and bump the date. Do this before responding "done" on any multi-file task and before suggesting `/clear`. A stale STATUS.md wastes the next session.
|
|
|
|
## File navigation
|
|
|
|
1. Before reading an unfamiliar file, grep `.wolf/anatomy.md` for its path to get a one-line description and token estimate. Do NOT read anatomy.md whole; it is an index, not a document.
|
|
2. If the description answers your question, skip the full read. For large files, prefer Read with offset/limit over whole-file reads.
|
|
3. If a file is not in anatomy.md, search with Grep/Glob. Regenerate the index with `openwolf scan`.
|
|
|
|
## Code generation
|
|
|
|
1. Before generating code, check `.wolf/cerebrum.md`: respect `## Do-Not-Repeat` (past mistakes), `## Key Learnings`, and `## User Preferences`.
|
|
2. Update cerebrum.md whenever you learn something: a user correction or preference, a project convention not obvious from code, an API surprise, a gotcha that would trip a fresh session, a significant decision and its why. The bar is LOW; a redundant entry costs nothing, a missing one repeats the discovery next session.
|
|
|
|
## Bug logging
|
|
|
|
Before fixing any bug: grep `.wolf/buglog.json` for the error message or filename; the fix may already be known.
|
|
|
|
After fixing any bug, failed test, failed build, or user-reported problem: append an entry with `id`, `timestamp`, `error_message`, `file`, `root_cause`, `fix`, `tags`, `occurrences`, `last_seen`. Also log when you edit a file more than twice to get it right. The threshold is LOW.
|
|
|
|
## Token discipline
|
|
|
|
- Never re-read a file already read this session unless it changed since.
|
|
- Prefer anatomy descriptions and targeted Grep over full file reads.
|
|
- If appending to a file, do not read the entire file first.
|
|
|
|
## Session end
|
|
|
|
Before wrapping up: update `.wolf/STATUS.md`, write a one-line session summary to `.wolf/memory.md` (`| HH:MM | description | file(s) | outcome | ~tokens |`), and record any learnings or bugs from the session in cerebrum.md / buglog.json.
|
|
|
|
## On-demand skills
|
|
|
|
- `/designqc`: screenshot-based design review of the running app (uses `openwolf designqc`).
|
|
- `/reframe`: UI framework selection, migration, and anti-generic design audits.
|
|
- `/security-audit`: security review of the project.
|