fix: show reasoning on the reasoning stream so thoughts render

[addyCooks]

Closes #853.

Description

The thought dropdown rendered but expanded to nothing. Three changes:

• source/ai-sdk-client/chat/chat-handler.ts streamed tokens are batched behind a 150ms timer, but the reasoning/text routing flag was driven by the *-start / *-end markers around a batch rather than by the deltas that filled it. The delta type is now the source of truth: reasoning-delta / text-delta flush the pending batch before flipping the flag, and the four start/end cases collapse into a single flush-only case. Routing is now independent of provider ordering.
• source/acp/acp-conversation.ts / acp-agent.ts leading whitespace-only reasoning no longer emits an agent_thought_chunk, is no longer stored on the message, and is skipped on session-history replay. The callback accumulates only what it emits, so replaySessionHistory re-sending the stored value renders identically to the live stream.
• plugins/vscode/media/chat-panel.js a thought section is only opened once there is non-whitespace text, and endCurrentTextBlock() moved inside that guard so an empty chunk can't split the answer.

Upstream references for the orderings this has to survive: @ai-sdk/openai emits reasoning-end on part.done only when store is set, otherwise it defers to output_item.done; @ai-sdk/openai-compatible emits reasoning-start without ever closing an active text part. Some providers emit reasoning deltas with no start marker at all.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update

Changeset

• Added a changeset (pnpm changeset) describing this change for the changelog

Testing

Automated Tests

• New features include passing tests in .spec.ts/tsx files
• All existing tests pass (pnpm test:all completes successfully)
• Tests cover both success and error scenarios

Manual Testing

• Tested with Ollama
• Tested with OpenRouter
• Tested with OpenAI-compatible API driven end-to-end against a local stub OpenAI-compatible SSE endpoint (real --acp server → real chat-panel.js), not a hosted provider; see the review reply for exactly what was covered
• Tested MCP integration (if applicable)

Checklist

• Code follows project style guidelines
• Self-review completed
• Documentation updated (if needed)
• No breaking changes (or clearly documented)
• Appropriate logging added using structured logging

[will-lamerton]

Hey @addyCooks - nice diagnosis - the batch-flush ordering is the real root cause, and the chat-handler tests genuinely reproduce it (the deltas are enqueued synchronously, so flushTimer is still pending when the flip arrives, which is exactly the failing condition). flushPending() also cleans up three copies of clear-then-flush. One thing I'd like changed before merge, plus a few smaller items.

Requested change

Routing is still driven by start/end markers rather than by the delta type.

isReasoning is only ever assigned in the *-start / *-end cases, so any reasoning-delta arriving while isReasoning === false still lands on onToken. That happens whenever a provider emits reasoning deltas without a preceding reasoning-start, or interleaves reasoning deltas after text-start because reasoning-end was deferred - which is precisely the ordering unreliability the PR description cites as the root cause. The fix hardens the flush but leaves the flag's source of truth unchanged, so the bug can still recur on a provider whose ordering diverges from the two cases covered by the tests.

Deriving the flag from the delta itself makes it ordering-independent and reduces the start/end handlers to pure flush points:

case 'reasoning-delta':
	if (!isReasoning) {
		flushPending();
		isReasoning = true;
	}
	accumulatedReasoning += chunk.text;
	tokenBuffer += chunk.text;
	...
case 'text-delta':
	if (isReasoning) {
		flushPending();
		isReasoning = false;
	}
	...

This also neutralises isReasoning = false on reasoning-end, which is currently a small latent hazard if a provider emits further reasoning deltas after closing the item. Worth a test with a bare reasoning-delta (no reasoning-start) to lock it in.

Smaller items

Live and replay disagree on leading whitespace. acp-conversation.ts:151 appends the token to streamedReasoning before the early return, so the raw value including the leading \n\n is what gets stored at line 244. replaySessionHistory then re-sends message.reasoning verbatim once it passes the new .trim().length > 0 guard. The same session renders as "foo" live and "\n\nfoo" after loadSession. Either trimStart() on the replay's text: field, or store the trimmed value.

Whitespace-only reasoning is still persisted. reasoning: streamedReasoning || undefined (acp-conversation.ts:244) keeps "\n\n" on the message even though nothing was emitted for it. streamedReasoning.trim() ? streamedReasoning : undefined stops it at the source and leaves the acp-agent guard as backwards-compat for sessions already on disk.

The user-visible piece has no coverage. plugins is excluded from Biome (biome.json:18) and AVA only globs plugins/**/*.spec.ts, so nothing lints or exercises plugins/vscode/media/chat-panel.js. The guard itself reads correctly - thoughtText && (currentThoughtBox || thoughtText.trim()) opens on first non-whitespace and keeps appending afterwards, and moving endCurrentTextBlock() inside the guard is an improvement over the old unconditional call. But every "Manual Testing" box is unchecked, and the dropdown rendering is the reported symptom, so could you do one manual pass in the extension and tick the relevant box?

Test shape nit. acp-conversation.spec.ts invokes capturedCallbacks.onReasoningToken(...) after runAcpConversation has already resolved. It passes because the closure over streamedReasoning outlives the turn, but it is asserting against a finished turn, so it would not catch a regression that moved the per-turn let streamedReasoning = '' reset out of the loop.

[addyCooks]

Thanks @will-lamerton all four addressed in c2e8e11. Rebased onto current main.

Routing by delta type. reasoning-delta / text-delta now own the flag (flush, then flip); the four start/end cases collapse into one flush-only case, so isReasoning = false on reasoning-end is gone. Three new tests, each verified to fail on the marker-driven version including the bare-delta case you asked for and the post-reasoning-end hazard.

Whitespace. Fixed at the source instead of on replay: the callback accumulates only what it emits (streamedReasoning ? token : token.trimStart()), so the stored value is exactly the thought chunks and replaySessionHistory can keep re-sending it verbatim. A test asserts assistant.reasoning === thoughtTexts(updates).join(''). Plus streamedReasoning.trim() ? … : undefined as you wrote it; the acp-agent guard stays, commented as back-compat.

chat-panel.js coverage. Your premise was right when you wrote it, but main has since picked up chat-panel-thoughts.spec.ts + chat-panel-harness.ts, which run the real panel in a VM sandbox — and under source/, AVA does glob it. Four tests added there, all failing without the guard, including one that locks in the endCurrentTextBlock() move.

Test shape. Tokens now stream from inside chat(), so assertions run against a live turn. Added resets streamed reasoning between turns, the regression the old shape couldn't catch.

Manual pass worth being precise: not a live extension session. I drove the real --acp server against a stub OpenAI-compatible SSE endpoint and replayed the captured session/updates through the real chat-panel.js. The deferred-reasoning-end ordering reproduces #853's screenshot exactly against main:

BEFORE  thought: "<md>\n\n</md>"   ← "Thought for 0s", renders to nothing
        answer:  "The user wants the capital of France. That is Paris.The capital of France is Paris."

AFTER   thought: "The user wants the capital of France. That is Paris."
        answer:  "The capital of France is Paris."

So the thoughts were never lost they were printing as the assistant's reply, and the empty bubble was the leading \n\n opening a section on its own.

I ticked Tested with OpenAI-compatible API on that basis with an inline caveat. Happy to untick if you'd rather that box mean a live session only that pass is still outstanding.

CI is green, full suite included.

[will-lamerton]

Nice work @addyCooks :)

[addyCooks]

Thanks @will-lamerton! : )