feat: surface streamed thinking content as a Discord blockquote

[marvin-69-jpg]

Summary

When a Claude (or any other ACP backend) agent enters extended thinking mode, the only visible signal in Discord today is the 🤔 emoji on the bot's message. The actual thinking content streamed by the agent over the ACP protocol (agent_thought_chunk events) is parsed and discarded:

"agent_thought_chunk" => {
    Some(AcpEvent::Thinking)   // ← unit variant, no payload
}

AcpEvent::Thinking => {
    reactions.set_thinking().await;   // ← only the emoji
}

This PR surfaces that content as a blockquote at the top of the streaming Discord message, similar to how the native Claude Code CLI shows thinking blocks above tool calls and the final answer:

> 🤔 Thinking
> Let me start by reading the README and checking my notes.
>
> There's a local commit ahead of origin/main, I should reset to
> origin/main first before branching.

🔧 `git fetch && git checkout main && git pull --ff-only`...

(final answer streams in here)

Discord users — especially in multi-user team channels where the bot is doing non-trivial work — get a much clearer picture of what the agent is reasoning about, before the answer arrives.

Implementation

• AcpEvent::Thinking carries the delta string instead of being a unit variant.
• classify_notification reads update.content.text (the same shape agent_message_chunk already uses).
• stream_prompt accumulates thinking deltas into a thought_buf alongside the existing text_buf, and re-renders the message on each event.
• compose_display gains a thought parameter and emits a blockquote-prefixed section above the tool list when thought is non-empty.

The block-boundary heuristic

claude-agent-acp emits both thinking (the start of a new block) and thinking_delta (continuation) as the same agent_thought_chunk shape with no marker for the boundary:

// claude-agent-acp/dist/acp-agent.js, around line 1463
case "thinking":
case "thinking_delta":
    update = {
        sessionUpdate: "agent_thought_chunk",
        content: { type: "text", text: chunk.thinking },
    };

When the model produces several thinking blocks in a row separated only by tool calls, the deltas concatenate into a wall of text like ".There's a local commit" with no whitespace between sentences. We work around this with a small heuristic in needs_thinking_separator:

fn needs_thinking_separator(prev: &str, new: &str) -> bool {
    let last = prev.chars().last()?;
    let first = new.chars().next()?;
    let sentence_end = matches!(last, '.' | '!' | '?' | '。' | '！' | '？');
    let starts_word = first.is_alphabetic();
    sentence_end && starts_word
}

If the previous chunk ends in sentence-terminating punctuation and the new chunk starts with a letter (no leading whitespace), insert a \n\n. This is imperfect but covers the common case without requiring upstream protocol changes — token-level deltas within a single block almost always include leading whitespace, so the heuristic doesn't over-correct.

A cleaner long-term fix would be for claude-agent-acp to thread the content_block_start event through to agent_thought_chunk (or add a block_start: true flag in the update payload). Happy to follow up upstream if maintainers prefer that route.

Files changed

• src/acp/protocol.rs — AcpEvent::Thinking(String); classify_notification extracts text from agent_thought_chunk.
• src/discord.rs — new thought_buf in stream_prompt; new needs_thinking_separator helper; compose_display takes thought and renders blockquote.

+75 / -8.

Compatibility

• Backwards compatible with any ACP backend that doesn't emit thinking events at all (the Thinking variant is just never produced; compose_display skips the blockquote when thought_buf is empty).
• Preset-agnostic: this is a render-side change inside discord.rs, not gated on agent.preset.
• No new dependencies.

Testing

Running in production for the past day against a multi-user Discord channel mixing English and Traditional Chinese conversations. The heuristic correctly inserts paragraph breaks between distinct thinking blocks and leaves token-level deltas inside a single block alone.

Note about base branch

This PR is layered on top of #138 (fix: dedupe tool call display by toolCallId and sanitize titles) — both PRs touch the AcpEvent enum and the compose_display signature, and #138 is the more critical bug fix. I've based this branch on fix/dedupe-tool-call-display so they apply cleanly together; if you'd rather merge them in the opposite order I can rebase. The diff stat above is for the thinking change alone, not cumulative.