docs/compatibility.md
Four agents, one definition.
Agentyx installs the same provider-neutral capabilities into each agent's own documented destinations. This page is generated from the table docs/compatibility.md records, which is checked against each provider's official documentation.
- project-local only
- plan before write
- reversible via lockfile
- never touches $HOME
Where Agentyx installs
Every path is relative to your project root. Agentyx writes nothing else, and never $HOME.
| Capability | Codex | Claude Code | Kimi Code | Qwen Code |
|---|---|---|---|---|
| Project Skills | .agents/skills | .claude/skills | .agents/skills | .qwen/skills |
| Project agents | .codex/agents/<name>.toml | .claude/agents/<name>.md | .agents/agents/<name>.md | .qwen/agents/<name>.md |
| MCP configuration | .codex/config.toml | .mcp.json | .kimi-code/mcp.json | .qwen/settings.json |
| Lifecycle hooks | .codex/hooks.json | .claude/settings.json | user-level config only — Agentyx installs none | .qwen/settings.json shared with MCP |
What doctor can observe
A capability a provider cannot reliably show is unobservable, never dormant. Absence of evidence is never reported as evidence of absence.
| Signal | Codex | Claude Code | Kimi Code | Qwen Code |
|---|---|---|---|---|
| Session lifecycle | yes | yes | unobservable | yes |
| Tool use | yes | yes | unobservable | yes |
| Skill use | unobservable | yes | unobservable | unobservable |
| MCP use | yes | yes | unobservable | yes |
| Context size | unobservable | yes | unobservable | unobservable |
Limits, stated plainly
- Codex codex
- Codex documents no Skill usage signal, so Skill activity is reported as unobservable rather than as zero use. Project hooks need your review and trust through /hooks, so Agentyx reports them as configured, never as active.
- Claude Code claude
- The most observable target: Skill use through the Skill tool or a typed /skill, and a context size on resumed or forked sessions.
- Kimi Code kimi
- Kimi Code supports hooks, but only in ~/.kimi-code/config.toml. Agentyx never writes under $HOME, so this is a policy limit, not a provider limit: Kimi gets the full static analysis and no runtime observation.
- Qwen Code qwenone settings document
- Qwen Code reads MCP servers and hooks from one document, so Agentyx plans and writes it as a single merge under one lock entry. It documents no Skill tool, so a model-selected Skill leaves no signal and Skill use stays unobservable. Project hooks load only for a folder you have trusted.