Concepts
Capability Matrix
Which agents enforce reflexes natively, which decisions each honors, and how agentreflex degrades elsewhere.
The same reflex is meant to run everywhere. But agents differ in what they expose, so coverage falls into two tiers.
Tiers
- Native — the agent exposes a pre-tool hook, so a reflex can
deny(and sometimesask) before a tool runs. - Advisory — no hook. agentreflex compiles your reflexes into
AGENTS.md, so the model is steered but nothing is hard-blocked.
Support today
| Agent | Tier | pass | deny | ask |
|---|---|---|---|---|
| Claude Code | Native | ✅ | ✅ | ✅ |
| Cursor | Native | ✅ | ✅ | ✅ |
| Copilot CLI | Native | ✅ | ✅ | ✅ |
| Gemini CLI | Native | ✅ | ✅ | — |
| Windsurf | Native | ✅ | ✅ | — |
| OpenCode | Native | ✅ | ✅ | — |
| Codex | Advisory | ✅ | guidance | guidance |
Where ask is unsupported (—), the call is allowed through rather than blocked. modify is reserved in the protocol but no agent honors it yet.
Degradation
A reflex you write once is portable. On native agents deny enforces deterministically; on advisory targets it becomes strongly-worded guidance in AGENTS.md. You always know which tier you're in:
arx doctor