fix: filter empty text content blocks to prevent LLM API 400 errors

[RomneyDa]

Summary

Empty or whitespace-only text content blocks in array message content cause 400 errors across multiple providers (Anthropic, Bedrock, Venice, and other OpenAI-compatible proxies). This has been reported in 12 open issues.

Root causes fixed:

• chatMessageIsEmpty() only checked string content — array content like [{type:"text", text:""}] was not detected as empty, so these messages passed through filtering
• Empty text parts in array content were never stripped — addSpaceToAnyEmptyMessages() only handled the string case, leaving empty text parts in arrays to reach provider APIs
• Bedrock _convertMessageContentToBlocks didn't filter empty text — blank text blocks were passed directly to the Bedrock API

Changes:

• core/llm/messages.ts: Fix chatMessageIsEmpty to detect empty array content. Replace addSpaceToAnyEmptyMessages with stripEmptyContentParts that removes empty text parts from arrays
• core/llm/openaiTypeConverters.ts: Extract toAssistantContent helper that filters empty text parts from assistant array content. Filter empty text parts in multi-media user messages
• core/llm/llms/Bedrock.ts: Skip empty/whitespace text in _convertMessageContentToBlocks
• packages/openai-adapters/src/apis/Bedrock.ts: Filter empty text parts in user message conversion
• core/llm/countTokens.ts: Use new stripEmptyContentParts

What this does NOT change:

• Existing " " fallbacks for string content (preserves compatibility with LM Studio, Ollama, etc.)
• toResponsesInput behavior (preserves Responses API ID chains)
• Bedrock error on empty assistant turns (keeps error surfacing for debugging)

Issues Fixed

Fixes #9232, fixes #9765, fixes #9767, fixes #10148, fixes #10293, fixes #10504, fixes #10804, fixes #11045, fixes #11264, fixes #11446, fixes #11497, fixes #11728

Test plan

• Existing openaiTypeConverters.test.ts tests pass (24/24)
• Existing countTokens.test.ts tests pass
• Prettier and lint checks pass
• Manual testing with Bedrock provider
• Manual testing with Anthropic direct provider
• Manual testing with OpenAI-compatible proxy (Venice, OpenWebUI)
• Verify tool call flows still work (assistant messages with empty content + tool calls)

[continue]

Docs Review: No documentation updates needed.

This PR fixes internal message handling to prevent empty/whitespace text content blocks from being sent to LLM APIs. These are implementation-level changes that:

• Fix 400 errors across multiple providers (Anthropic, Bedrock, Venice, etc.)
• Don't introduce new user-facing configuration options
• Don't change how developers interact with Continue

The fixes improve reliability behind the scenes—users will simply stop seeing the errors without needing to know or configure anything differently.

[cubic-dev-ai]

2 issues found across 5 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="core/llm/openaiTypeConverters.ts">

<violation number="1" location="core/llm/openaiTypeConverters.ts:150">
P1: Whitespace-only assistant string content is still sent instead of being treated as empty.</violation>

<violation number="2" location="core/llm/openaiTypeConverters.ts:187">
P1: Whitespace-only user string content is not normalized and can still be sent to providers.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[RomneyDa]

this is a somewhat risky one, pending further evaluation

[RomneyDa]

Closing in favor of #11877 which takes a more targeted approach — fixing empty text content blocks at each provider boundary (+15/-6 lines) instead of modifying the shared upstream message filtering layer. See the comment on #11877 for rationale.