Summary

Render Codex reads of skill instruction files as skill-specific ACP read tool calls.

Before

• When the agent loaded/read a skill, Codex emitted a normal commandExecution item whose parsed commandActions contained a read action for a .../skills/<skill>/SKILL.md path.
• codex-acp mapped every read action the same way, so the UI only saw a generic title like Read file '/path/to/SKILL.md'.
• That made skill loading indistinguishable from ordinary file reads, even though the UI wants to show activity such as Read UI Accessibility skill separately.

After

• createCommandActionEvent() detects read actions whose path points at a skill instruction file (*/skills/<skill>/SKILL.md).
• Those reads are still emitted as standard ACP read tool calls:
  • sessionUpdate: "tool_call"
  • kind: "read"
  • title: "Read <skill> skill"
  • locations: [{ path: <SKILL.md path> }]
  • rawInput: { type: "skill", name: <skill directory>, path }
• Ordinary file reads are unchanged.

Why

The agent-side “loaded/read skill” activity is already present in Codex events as a parsed read command for SKILL.md; codex-acp just treated it like a generic file read. This keeps the ACP protocol shape unchanged while giving Codex-aware UIs a stable way to render skill loading separately from regular file reads (rawInput.type === "skill").

Tests

• npx vitest run src/__tests__/CodexACPAgent/command-action-events.test.ts -u
• npm run typecheck
• npm test