Capability Matrix

Supported harnesses: Claude Code · Codex · OpenCode · Pi Mono

Additional (tracked, not blocking): Copilot · Gemini · Kimi

Legend:

• ✅ Universal codegen — adapter emits native output automatically from the manifest.
• ⚠️ Guided per-harness — no native primitive of its own; a documented escape-hatch pattern covers the gap. Emits a WARN (not error) on portable manifests.
• ❌ Unsupported — no viable path on this harness; recorded here, not blocking.
• n/a — not applicable (mechanism differs but functionality is covered).

Reading the "Escape hatch" column: for every ⚠️ cell there's a deliberate reason we don't emit a universal primitive for that capability on that harness yet — the work to express it universally is significant, and a native path already does the job. If you're shipping today, use the escape hatch. If you want a universal primitive, see the Decision tree for authors at the bottom.

Capability table

Capability	Claude Code	Codex	OpenCode	Pi Mono	Escape hatch	Notes
skills	✅	✅	✅	✅	—	Universal codegen
hooks (lifecycle)	✅	✅	✅	✅	—	Universal codegen
commands	✅	✅	✅	✅	—	Universal codegen
mcpServers	✅	✅	✅	⚠️	Pi has no built-in MCP. Ship tools via native tools[] (emitted natively), or bridge an MCP server through a Pi extension via nativeEntry.pimono.	Universal on Claude · Codex · OpenCode. Pi Mono has no MCP; emits WARN when mcpServers is set. See MCP on Pi.
agents[]	✅	✅	✅	⚠️	Pi has no named-agent declaration primitive. Use nativeEntry.pimono to spawn custom agents via a Pi extension.	Pi adapter emits nothing for agents[]; use nativeEntry.pimono for custom agent wiring on Pi.
agents[].model	✅	⚠️	✅	⚠️	Codex/Pi: use the harness's own per-agent model config (or simply omit and accept the harness default).	Claude + OpenCode emit model: frontmatter when set. Codex/Pi have no per-agent file concept; model unset → harness default.
subagentStart	✅	✅	⚠️	✅	OpenCode: intercept subagent tool calls with preToolUse matcher (subagents launch via the subagent tool).	Emits WARN on OpenCode. Pi maps to agent.AgentStart lifecycle event.
subagentStop	✅	✅	⚠️	✅	OpenCode: detect via postToolUse / postToolUseFailure on the subagent tool.	Emits WARN on OpenCode. Pi maps to agent.AgentStop lifecycle event. Pi stop↔subagentStop collision fixed in v0.3.0.
tools[] (first-class)	⚠️	⚠️	✅	✅	Claude/Codex: ship tools via mcpServers — works on all four harnesses (universal).	WARN emitted; OpenCode/Pi emit first-class tools[] natively.
stop / continueWith	⚠️	⚠️	⚠️	⚠️	Each harness already has a stop-class lifecycle hook natively — emit nothing in portable manifests until v0.5.0.	New universal primitive — v0.5.0; all-harness design.
Native-entry passthrough	n/a (JSON)	n/a (JSON)	⚠️	⚠️	OpenCode: drop a .ts file directly into ~/.config/opencode/plugins/<name>/ — Bun runs it as ESM, no codegen needed.	nativeEntry escape hatch — ships in v0.5.0; OpenCode native modules must be .ts (file-drop path).
Inline hook handlers	✅ auto-wrap	✅ auto-wrap	✅	✅	—	Codex/Kimi: auto-wrapped as Node.js command scripts (v0.5.0).

Additional harnesses

Capability	Copilot	Gemini	Kimi	Escape hatch	Notes
skills	✅	✅	✅	—	Universal.
hooks (lifecycle)	⚠️	⚠️	❌	Copilot/Gemini: use HTTP or command handlers (see Hooks).	Kimi supports a subset (see below).
subagentStart / subagentStop	❌	❌	❌	TUI fidelity only — implement per-harness via native plugin config if you really need it.	No universal primitive planned.
tools[]	⚠️	⚠️	❌	Use mcpServers instead (not emitted on these Tier-2 harnesses).	No first-class tool emission; emit WARN.
mcpServers	❌	❌	❌	Wire the MCP server directly into the harness's native config — no agentplugins path needed.	Not on the manifest path; harness-level wiring only.

Kimi supported hooks: preToolUse, userPromptSubmit, sessionStart, notification, permissionRequest. Inline handlers auto-wrapped as Node.js command scripts (v0.5.0).

Decision tree for authors

Does universal codegen cover this capability across all four core harnesses?
  YES → use it; adapter handles the rest
  NO  → does the native escape hatch (Escape hatch column above) cover it today?
    YES → ship the escape hatch now; document the limitation in the manifest
    NO  → can you express it per-harness via nativeEntry + a hand-written module?
      YES → use nativeEntry; emit WARN (not error)
      NO  → is the gap TUI-grade fidelity only?
        YES → acceptable degradation; note in this matrix
        NO  → open a primitive proposal (v0.5.0+ scope)

---

This matrix is the living contract for the project. Update it as capabilities land or gaps are discovered.