v0.29.0

One question per destructive call, the server's own when it asks, and tool results in the person's language.

New features

• One question, not two (#77). metadata["agent-kit"].serverConfirmTools lists an extension's tools whose server asks the person itself before acting, through MCP elicitation (moodle-mcp before deleting or resetting a course). In guided mode the kit lets them through without its own question, so the person answers once, the server's. A kept "no, never" still holds; plan mode still stops them; autonomous still denies them unless extensionGate.autonomous allows it.
• Tool results in the person's language (#79). A tool label can give the line shown under the call, ToolLabel.result(result), or result per language in a manifest's labels. The kit's own extensions (sources, knowledge, memory, awareness) no longer show the model's English line when the call succeeded, since their label already says what happened; a failure still shows. The agent's formatResult comes first, and the model's result is unchanged. quietResults() and withToolResults() are exported.
• @falkenslab/agent-kit-catalog 0.6.0 accepts serverConfirmTools and a label's result.

Fixes

• A tool that always asks asks on every call (#78). A confirmTools or destructiveHint tool offered "yes, this session", after which it ran without asking for the rest of the session. It's now offered only yes, no and "no, never".

Captain Whiskers

Captain Whiskers 0.6.1, the example agent, now built on agent-kit 0.29.0, as a Windows app: download the installer. It's unsigned, so SmartScreen warns about it; it asks for a Claude key on first start and keeps everything in ~/.captain-whiskers. The same captain as with 0.28.0, on the new kit.