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.

[chaodu-agent]

🔴 Needs Rebase + Rework — Thinking-content feature is valuable, but the branch is stale and duplicates code already merged to main.

Baseline Check (Step 0)

Field	Value
State	OPEN
Mergeable	CONFLICTING
Created	2026-04-08 (23 days ago)
Last commit	2026-04-08
Author	@brettchien
Labels	closing-soon
Base branch	fix/dedupe-tool-call-display (PR #138)

Main has: AcpEvent::Thinking (unit variant, no payload), ToolEntry/ToolState/sanitize_title in adapter.rs, compose_display with 4-param signature (tool_lines, text, streaming, tool_display). Thinking events trigger 🤔 emoji only — content is discarded.

Net-new: AcpEvent::Thinking(String) payload, thought_buf accumulation, blockquote rendering ("> 🤔 Thinking"), needs_thinking_separator() heuristic.

四問框架

1. 解決什麼問題

When an agent enters extended thinking mode, the only visible signal is a 🤔 emoji. The actual thinking content streamed via agent_thought_chunk events is parsed and discarded.

2. 怎麼解決

Changes AcpEvent::Thinking to carry the delta string, accumulates thinking deltas into thought_buf, renders as a blockquote above the tool list in Discord messages.

3. 考慮過什麼

Block-boundary heuristic (needs_thinking_separator) handles the case where claude-agent-acp emits both thinking and thinking_delta as the same shape with no boundary marker.

4. 最佳方案嗎

The feature is great UX, but the implementation needs rebasing — most of the tool-tracking infra it adds already landed on main.

🟢 INFO

• Surfacing thinking content as a blockquote is a great UX improvement.
• needs_thinking_separator() is a thoughtful heuristic for detecting block boundaries.
• Graceful fallback: Thinking("") still triggers the 🤔 reaction.
• Good defensive coding: ToolDone without prior ToolStart is handled.

🟡 NIT

1. Hardcoded blockquote header — "> 🤔 _Thinking_\n" is hardcoded. Main's config already has emoji_thinking — consider using it for consistency.
2. No truncation on thought_buf — Extended thinking sessions could produce very long blockquotes that hit Discord's 2000-char message limit.
3. Visibility mismatch — ToolEntry and ToolState are declared pub in discord.rs but were pub(crate) / private in main's adapter.rs.

🔴 SUGGESTED CHANGES

1. Merge conflict — needs rebase. Main already has ToolEntry, ToolState, sanitize_title, and id-based tool dedup in src/adapter.rs (merged from PR #138 or similar). This PR re-introduces all of that in src/discord.rs. After rebase, the only net-new code should be: (1) AcpEvent::Thinking(String) in protocol.rs, (2) thought_buf accumulation + needs_thinking_separator(), (3) blockquote rendering in compose_display.

2. compose_display signature mismatch — Main's compose_display takes (tool_lines, text, streaming, tool_display). The PR's version takes (tool_lines, thought, text) — dropping streaming and tool_display which control compact/full/none tool rendering and the "typing…" indicator. These features would regress. The thought parameter needs to be added to main's existing signature.

3. protocol.rs duplication — The Thinking → Thinking(String) change is the core of this PR and is correct, but the diff also re-introduces ToolStart { title } → ToolStart { id, title } which main already has. After rebase this should be a one-line change.

[chaodu-agent]

超渡法師 Review — PR #139

1. What problem does it solve?

When an ACP backend agent enters extended thinking mode, the only visible signal in Discord is a 🤔 emoji reaction. The actual thinking content streamed via agent_thought_chunk events is parsed but immediately discarded. Users have no visibility into what the agent is reasoning about.

2. How does it solve it?

• protocol.rs — AcpEvent::Thinking becomes AcpEvent::Thinking(String), extracting update.content.text from agent_thought_chunk events.
• discord.rs — Adds thought_buf to accumulate thinking deltas. needs_thinking_separator() heuristic detects block boundaries using sentence-terminating punctuation (including CJK: 。！？). compose_display renders thinking as > 🤔 _Thinking_ blockquote above tool list and answer text.

3. What was considered?

• Upstream protocol fix acknowledged as the cleaner long-term solution.
• Backwards compatible — when no thinking events are emitted, blockquote is skipped.
• Stacked on fix: dedupe tool call display by toolCallId and sanitize titles #138 (both touch AcpEvent and compose_display).

4. Is this the best approach?

The feature design is solid — blockquote rendering of thinking content is the right UX pattern for Discord. The heuristic is a pragmatic workaround for a protocol gap. However, the implementation needs a rebase.

---

Traffic Light

🟢 INFO — Excellent UX improvement

• Thinking content as blockquotes gives users real-time visibility into agent reasoning.
• needs_thinking_separator() is well-reasoned with good CJK punctuation support.
• Graceful degradation: Thinking("") still triggers 🤔 reaction; backends without thinking events are unaffected.
• PR description is exceptionally thorough — explains the protocol gap, heuristic limitations, and upstream fix path.

🔴 SUGGESTED CHANGES

1. Rebase required — PR is in CONFLICTING merge state. Main's codebase has been significantly refactored since the branch was created. Streaming logic (stream_prompt, compose_display, ToolEntry, ToolState) moved from discord.rs to adapter.rs. The PR branch doesn't have adapter.rs at all. After rebase, the diff should be ~30-40 lines net-new.

2. compose_display signature mismatch — Main's compose_display takes (tool_lines, text, streaming, tool_display) with compact/full/none tool rendering modes. The PR's version takes (tool_lines, thought, text), dropping streaming and tool_display. Merging as-is would regress the tool display feature. Add thought as a 5th parameter to main's existing signature.

3. protocol.rs diff is inflated — The diff re-introduces tool_id extraction and ToolStart { id, title } / ToolDone { id, title, status } changes already on main (from fix: dedupe tool call display by toolCallId and sanitize titles #138). After rebase, the protocol.rs change should be ~12 lines.

🟡 NIT

1. No thought_buf truncation — Extended thinking can produce thousands of tokens. The blockquote will eventually exceed Discord's 2000-char limit. Consider truncating to show only the last N characters during streaming.
2. Hardcoded emoji — "> 🤔 _Thinking_" is hardcoded, but main's config already has emoji_thinking. Use the configured emoji for consistency.

---

Verdict

🔴 Changes requested — Great feature, needs rebase. The thinking blockquote is the right UX pattern. After rebasing onto current main, port the thinking changes to adapter.rs (where compose_display now lives). The net-new diff should be ~30-40 lines.