Rules
How OpenCorvus picks up AGENTS.md, CLAUDE.md, and project rules.
OpenCorvus loads project-specific instructions from a small set of well-known files. These files give the agent durable context — coding style, architectural decisions, must-do/must-not-do lists — that you don’t want to repeat in every prompt.
Sources
| File | Purpose |
|---|---|
AGENTS.md | Generic agent instructions; primary file recognized by OpenCorvus |
CLAUDE.md | Compatibility with Claude Code conventions |
.cursor/rules/*.md | Cursor-style project rules (when configured via instructions) |
instructions glob entries | Anything you list in opencorvus.jsonc |
Configuring extra rule files
{ "instructions": ["CONTRIBUTING.md", "docs/guidelines.md", ".cursor/rules/*.md"],}Paths are resolved relative to the config file. Glob patterns expand at startup.
Session MEMORY.MD and project memory
Beyond static rule files, a Session gains one generated Markdown checkpoint named MEMORY.MD after its first successful conversation compaction. The document is a logical view of the latest completed compaction-summary message. That summary already appears once in compacted replacement history, so OpenCorvus does not inject another copy beside ordinary user Turns. The memory tool can read this checkpoint but cannot mutate it; only a later successful compaction advances it.
Project semantic memory remains a separate cross-Session knowledge store for reusable facts, lessons, profiles, and episodes. Generated memory is not an instruction or authority: required behavior belongs in AGENTS.md or checked-in records, drift-prone facts must be verified, and secrets must never be stored in either memory surface.
When to use which
- Coding standards that always apply →
AGENTS.mdat the repo root. - Per-feature rules or playbooks →
instructionsentry pointing to a file. - Time-sensitive constraints (“do not touch module X until next week”) → put them in
AGENTS.mdand remove when done. Avoid embedding them inopencorvus.jsoncitself.
Skills vs rules
Rules are always-on system-prompt content. Skills are pickable instruction packs the agent invokes by name when the task matches their description. Use rules for project-wide constraints, skills for reusable scoped workflows.