Repository navigation
Description
Problem
A permission ask shows the tool call's kind and title and nothing else, so a request whose title is generic gives the user no way to know what they are approving.
Observed on macOS, 2026-09-14. The whole row reads:
BUDDY BOT · ASKS
other: Question from MCP server
[ Allow ] [ Always allow ] [ Deny ] [ Cancel ]
other is ToolKind::Other through name_of, and it is given top billing despite being the least informative thing available. "Question from MCP server" is the tool call's title. The question itself is never forwarded. The user is asked to Allow or Deny with the subject withheld.
This is the shape that trains people to click Allow, which makes it worse than a cosmetic gap: the consent gate is the one place the product asks the user to make a security decision, and it is the one place the copy says least.
Cause
permission_ask at src-tauri/src/acp_wire.rs:712-735 builds PermissionAsk from exactly two fields of the tool call:
title: request.tool_call.fields.title.clone().unwrap_or_else(|| "(untitled)".to_string()),
kind: request.tool_call.fields.kind.as_ref().map(name_of),PermissionAsk (acp_wire.rs:88) therefore carries only request, title, kind, options. But agent-client-protocol 2.1's ToolCallUpdateFields also carries, all of which are dropped on the floor:
content: Option<Vec<ToolCallContent>>— where a question's text would beraw_input: Option<serde_json::Value>— the arguments the tool was called withlocations: Option<Vec<ToolCallLocation>>— what it would touchname: Option<String>— the programmatic tool name, which alone would say more thanother
Then src/chat.js:208 renders what little survived:
body.textContent = ask.kind ? `${ask.kind}: ${ask.title}` : ask.title;Expected
- Show what is actually being asked. Prefer the tool call's
content; fall back toraw_inputwhen there is no content. - Do not lead with
kind.otheris noise; the toolnameis not. - When a tool would touch a file or a path, say which —
locationscarries it. - An ask with genuinely nothing to show should say that plainly rather than printing
(untitled)or a bare kind, so an empty ask is distinguishable from one whose detail was dropped. - A long or structured payload must not break the 420pt surface (ADR-0012), and
raw_inputis arbitrary JSON — it needs a bound and a readable shape, not a dump.
Worth deciding in the issue
raw_input is untrusted and arbitrary. It reaches the DOM through textContent today, which is safe; keep it that way, and note this interacts with #677 if a Markdown renderer lands on the same surface.
How to verify
Attach an MCP server that asks a question, or exercise the fake at src-tauri/src/harness.rs:1787, which already posts a synthetic session/request_permission and is the cheapest way to drive the row.
Related
- ADR-0018 (the Harness authenticates itself; ai-buddy brokers the decision but holds nothing)
- ADR-0012 (the 420pt chat surface)
- feat(chat): Render Markdown in a reply instead of its source #677 (Markdown on the same surface)
- Spike: how an attached Harness authenticates, and what ai-buddy must never hold #371 (untrusted content reaching the user)
Reporter still is in the originating chat; GitHub's attachment upload is a browser affordance no agent can reach, so drag the PNG onto this issue if a picture is wanted in-thread.
— Claude Code agent, on @omesser's behalf.
Metadata
Metadata
Assignees
Labels
Projects
- StatusShow more project fieldsDone
Activity