External agents: slash command autocomplete is offered mid-prompt, but only a leading command runs

Reproduction steps

1. Open a Claude Code external agent thread in the Agent Panel.
2. Type some text, then / and part of a skill name, e.g. let's test this: /pr.
3. The completion menu appears; accept /pr and send.

Current vs. Expected behavior

Current: the completion menu offers commands anywhere in the prompt, but a command that isn't at the start of the message is sent to the agent as plain text, and the skill never runs. The same /pr sent at the start of a message loads the skill as expected.

Expected, either:

• only offer slash command completion at the start of the prompt (matching what the agent executes), or
• route mid-prompt commands so they actually run, as @ mentions work anywhere.

Today the UI suggests mid-prompt commands work when they don't, so users find out only from the agent's reply.

Related

• #53161: a commenter notes "The slash command only work at the start of the prompt and not anywhere alse in the prompt"

Zed version and system specs

Zed: 1.25.3 Agent: Claude Code via ACP OS: macOS (Darwin 24.6.0)