🤖 Fix Anthropic thinking error with context-aware filtering

[ammar-agent]

Problem

When users interrupt a thinking stream and resume with Anthropic's extended thinking enabled, the API returns an error:

Expected `thinking` but found `text`

This happens because:

1. User interrupts during thinking phase → assistant message has only reasoning, no text
2. On resume, filterEmptyAssistantMessages() strips reasoning-only messages
3. Anthropic API receives thinking request without prior reasoning context → error

Solution

Implement context-aware filtering that preserves reasoning-only messages when needed:

• Preserve reasoning-only partial messages for Anthropic with thinking enabled
• Filter them in all other cases (existing behavior)

This gives Anthropic's thinking API the context it needs while keeping other flows clean.

Implementation

filterEmptyAssistantMessages() now accepts optional context:

• provider - AI provider name
• thinkingLevel - Thinking setting

Filtering logic:

• Has text/tools → Keep (unchanged)
• Anthropic + thinking enabled + partial → Keep (NEW)
• Otherwise → Filter (unchanged)

Testing

New integration test with 5 tricky message scenarios:

• Reasoning-only messages
• Empty text content
• Reasoning + empty text
• Multiple reasoning blocks
• Whitespace-only text

Tests both Anthropic and OpenAI to ensure the fix doesn't break other providers.

Uses direct HistoryService manipulation (not timing-dependent interrupts) for reliability.

Results:

• ✅ All 297 unit tests pass
• ✅ Typecheck passes
• 🔄 Integration tests run on both providers

Related

This builds on #170 which changed the sentinel from [INTERRUPTED] to [CONTINUE] for better model understanding.

Generated with cmux

[chatgpt-codex-connector]

💡 Codex Review

cmux/src/utils/messages/modelMessageTransform.ts

Lines 485 to 502 in 8dcff40

	export function transformModelMessages(messages: ModelMessage[], provider: string): ModelMessage[] {
	// Pass 0: Coalesce consecutive parts to reduce JSON overhead from streaming (applies to all providers)
	const coalesced = coalesceConsecutiveParts(messages);
	// Pass 1: Split mixed content messages (applies to all providers)
	const split = splitMixedContentMessages(coalesced);
	// Pass 2: Provider-specific reasoning handling
	let reasoningHandled: ModelMessage[];
	if (provider === "openai") {
	// OpenAI: Strip all reasoning parts (Anthropic's text-based reasoning is incompatible with OpenAI's format)
	reasoningHandled = stripReasoningParts(split);
	// Then filter out any messages that became empty after stripping
	reasoningHandled = filterReasoningOnlyMessages(reasoningHandled);
	} else if (provider === "anthropic") {
	// Anthropic: Filter out reasoning-only messages (API rejects messages with only reasoning)
	reasoningHandled = filterReasoningOnlyMessages(split);
	} else {

Preserve reasoning-only messages through all transforms

The new filterEmptyAssistantMessages keeps reasoning‑only partials when Anthropic thinking is enabled, but those preserved entries are still unconditionally removed a few steps later in transformModelMessages. The Anthropic branch at the provider transform (filterReasoningOnlyMessages) drops every assistant message whose content consists solely of reasoning regardless of thinkingLevel. As a result, resuming a thinking stream still sends no reasoning context to Anthropic, so the API can continue to return Expected thinkingbut foundtext``. The transform needs the same context-aware logic (e.g., skip filtering partial reasoning messages when thinking is on) or the earlier change is effectively a no-op.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.