ACP: Display agent terminal command output in ToolCallCard #355

Problem or Opportunity

The ACP terminal capability lets an agent request that the client (Termul) run a command in a sandboxed subprocess. The Rust backend implements this fully and correctly — all five terminal RPCs (terminal/create, terminal/output, terminal/wait_for_terminal_exit, terminal/kill, terminal/release) are wired in src-tauri/src/acp/manager.rs:898-1022, with TerminalRegistry providing UTF-8-safe buffering, front-truncation, exit polling, and release_all on teardown (src-tauri/src/acp/terminal.rs).

However, the renderer only shows a static label for these tool calls and never displays the command output. In ToolCallCard.tsx:69-82:

if (item.type === 'terminal') {
  // The ACP `terminal` content variant only references a terminal by id; its
  // live output is fetched separately via `terminal/output` (not embedded in
  // the tool call), so we surface the reference rather than inline output.
  const terminalId = (item as { terminalId?: string }).terminalId
  return (
    <div ... >
      {terminalId ? `Terminal ${terminalId}` : 'Terminal'}
    </div>
  )
}

The comment claims output is "fetched separately via terminal/output", but there is no UI code that actually fetches and displays it. The terminal/* RPCs are agent→client requests serviced by the backend; the renderer has no IPC command to retrieve an ACP terminal's output — it only sees tool-call updates that reference a terminalId.

Result: when an agent exercises the terminal capability (e.g. runs npm test, git status, a build script), the command executes correctly on the backend, but the user has no way to see what the command printed. The tool-call card just says "Terminal term-3". This makes the highest-risk client capability (arbitrary command execution) effectively unauditable from the UI, which is a significant trust/usability problem for production use.

Proposed Solution

1. Expose a read-only IPC surface for ACP terminal output. Add a Tauri command (e.g. acp_terminal_output(agent_id, terminal_id)) that returns the current buffered output + running flag + exit status from TerminalRegistry::output (src-tauri/src/acp/terminal.rs). Keep it read-only; the agent remains the sole writer.
2. Render output in ToolCallCard. When a tool-call content item has type === 'terminal', fetch and render the output (collapsible by default, expandable on click). Poll or subscribe while the terminal is running; stop once it exits.
3. Surface exit status. Show exit code / signal (SIGINT/SIGKILL/SIGTERM on Unix) alongside the output, reusing the existing TerminalExitStatus shape.
4. Respect the buffer cap. If output was front-truncated (the truncated flag), show an indicator like "output truncated (first N bytes omitted)".

Success Criteria

• When an agent invokes the terminal capability, the user can see the command's stdout/stderr in the ToolCallCard.
• Output streams/polls while the command runs and settles once it exits.
• Exit code / signal is visible.
• Truncation is indicated when the buffer cap is hit.
• No regression to the agent-side terminal lifecycle (create/output/wait/kill/release still honored).

Additional Context

• Backend (complete, no change needed): src-tauri/src/acp/terminal.rs, src-tauri/src/acp/manager.rs:898-1022.
• Frontend gap: src/renderer/components/chat/ToolCallCard.tsx:69-82.
• The allow_terminal capability defaults to false on AgentConfig (src-tauri/src/acp/config.rs:87-92) and is advertised in client_capabilities (src-tauri/src/acp/client.rs:30-32) — that gating is correct and should stay.
• Related: this is the renderer-side completion of the agent terminal capability (ADR-tagged M6 in client.rs:30-32).