Describe ACP tool calls before Claude Code requests them

• Status: Todo
• Kind: Feature
• Authors: jp
• Date: 2026-09-25
• Label: domain=llm
• Label: llm-provider=anthropic
• Label: package=jp_cli
• Label: package=jp_llm
• Label: package=jp_mcp
• Label: type=follow-up

With the Claude Code provider, a tool call's formatter (and any question it asks) only starts once Claude Code sends the call's permission request. Claude Code sends parallel calls to JP's MCP server one after another, so when the model asks for two ticket_create calls with long titles, the second call's shorter_title inquiry only starts after the user has answered the first call's approval prompt. The user waits several seconds between the two prompts.

For providers JP calls directly, this already overlaps: a call's arguments are complete once it finishes streaming, so its formatter runs while an earlier call's approval prompt is open (PR #1175).

Why ACP calls can't overlap today

JP builds an ACP call's ToolCallRequest from the permission request's raw_input (jp_llm acp/protocol.rs, the permission handler). The model's streamed tool_use blocks only produce a pending row with the tool name; their input is ignored.

The streamed input does arrive early. In a manual run of PR #1175, the second call's complete input reached JP at 12:36:48.596, and Claude Code sent the call itself at 12:37:13.782, 9 ms after the first call's result went back.

It can't be used as-is:

• The stream and the permission request travel separately, and stream events can arrive after the permission request.
• Calls made inside a subagent (parent_tool_use_id set) are dropped from the stream.
• Claude Code may change the input before it asks (e.g. through a PreToolUse hook). Not verified.

Proposal

Describe the call speculatively, and only keep the result if it matches:

1. The ACP protocol parses a finished tool_use block's input and emits it as a hint, keyed by the tool-call id.
2. For a tool with format = "unattended", the coordinator starts the formatter from the hint, including routing its questions.
3. When the permission request arrives, compare its raw_input with the hint. Equal: reuse the description and answers. Different: discard them and describe the call again.
4. A call Claude Code never sends (cancelled or restarted turn) discards its description; any open question is recorded as withdrawn.
5. A remembered N for the tool still skips the formatter.

The comparison guarantees the approval prompt never shows a description built from other arguments. The worst case is one wasted assistant request.

Not this

Marking tools read-only so Claude Code runs them in parallel: ticket_create writes a file, so the mark would be false, and it isn't confirmed that Claude Code uses it that way.

Open

• Whether PreToolUse hooks can rewrite tool input in practice, and how often the hint and raw_input differ.
• Whether the MCP server should keep the speculative description (it owns formatter runs) or the Host should.