fix(provider): skip empty-text filtering for assistant messages in normalizeMessages (#16748)

[altendky]

Issue for this PR

Closes #16748
Related: #13286, #16246, #15074, #10970, #14716, #6176, #9364, #8010

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

normalizeMessages() in transform.ts:62-66 filters empty text parts from all message roles, including assistant. When an empty text part between two reasoning blocks with cryptographic signatures is removed, the positional context encoded in the signatures changes, and the Anthropic API rejects the replayed message with:

`thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified.

Fix: Skip the empty-text/reasoning filter for assistant messages. Assistant messages must be replayed verbatim because Anthropic thinking block signatures are positionally sensitive. The filter is still applied to user and tool messages where Anthropic rejects empty content.

Changes:

• transform.ts: Add if (msg.role === "assistant") return msg before the filter. Also simplified the remaining filter to only check text parts (since reasoning parts only appear in assistant messages, which are now skipped).
• transform.test.ts: Added a failing test reproducing the exact scenario (reasoning with signatures around an empty text part). Updated 5 existing tests whose expectations assumed filtering applied to assistant messages.

How did you verify your code works?

• All 106 transform tests pass, 0 failures
• The new test constructs [reasoning(sig1), text(""), reasoning(sig2), text("..."), tool-call], runs through ProviderTransform.message() with an Anthropic model, and asserts all 5 parts are preserved
• Before the fix: Expected length: 5, Received length: 4
• After the fix: passes

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my search results, I found related PRs addressing thinking block and empty content filtering issues, but these appear to be related fixes rather than duplicates of the current test reproduction PR (#16750):

Related PRs:

• fix(provider): preserve redacted_thinking blocks and fix signature validation #12131 - fix(provider): preserve redacted_thinking blocks and fix signature validation - Addresses thinking block signature validation
• fix: expand Anthropic detection and strip whitespace-only text blocks #12634 - fix: expand Anthropic detection and strip whitespace-only text blocks - Related to whitespace/empty content handling
• fix: filter empty content blocks for Bedrock provider #14586 - fix: filter empty content blocks for Bedrock provider - Empty content filtering for another provider
• fix: strip reasoning parts when switching to non-interleaved models #11572 - fix: strip reasoning parts when switching to non-interleaved models - Related to reasoning/thinking block handling
• fix: strip incompatible thinking blocks when switching to Claude #8958 - fix: strip incompatible thinking blocks when switching to Claude - Thinking block compatibility

The current PR #16750 is a test reproduction case that demonstrates the issue described in #16748, whereas the related PRs are attempted fixes or related provider handling. No exact duplicate reproduction test was found.

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[altendky]

History of the Modified Code

Timeline

Date	Commit	What happened
2025-11-28	13f89fdb8 (PR #4811)	First empty-message filter added in message-v2.ts. A community contributor (DS / @Tarquinen) added .filter((msg) => msg.parts.length > 0) before convertToModelMessages() to prevent crashing when empty messages hit the AI SDK. Only filtered at the UIMessage level — no part-level filtering.
2026-01-05	c285304a (no PR, direct commit)	Anthropic empty-content filter added in transform.ts by Aiden Cline. Anthropic's API rejects messages with empty content, so this filters out empty text and reasoning parts from array content, and removes messages that become empty. Applied to all roles (user, assistant, tool) indiscriminately. Tests all used role: "assistant" messages.
2026-01-20	021e42c0b (no PR, direct commit)	differentModel guard added in message-v2.ts by Aiden Cline. When switching between providers, stale providerMetadata (e.g. Anthropic signatures sent to OpenAI) caused 400 errors. Fix: only include providerMetadata when the model matches. This introduced the conditional ...(differentModel ? {} : { providerMetadata: part.metadata }).
2026-02-13	0d90a22f9 (PR #13439)	Opus 4.6 adaptive thinking enabled on vertex/bedrock/anthropic. This is when Opus 4.6 started producing interleaved reasoning blocks with cryptographic signatures, creating the conditions for the bug.

How the bug emerged

The empty-content filter from Jan 5 (c285304a) was written before Opus 4.6 adaptive thinking existed (Feb 13). At the time, reasoning parts with empty text were genuinely useless — there were no cryptographic signatures to preserve. The filter treated all roles identically.

Once Opus 4.6 arrived, it began producing patterns like:

reasoning(sig1) → text("") → reasoning(sig2) → text("answer")

The empty text("") is structurally significant — the signatures encode positional context. There are two independent stripping points:

1. transform.ts: The normalizeMessages filter removes text("") from assistant messages, shifting the reasoning block positions and invalidating signatures.

2. message-v2.ts: Even if transform.ts is fixed, the AI SDK's internal convertToLanguageModelPrompt strips text parts where text === "" && providerOptions == null. The differentModel guard from Jan 20 was correct for its purpose (don't send foreign metadata), but when part.metadata is undefined for an empty text part from the same model, the resulting providerMetadata: undefined causes the AI SDK to strip the part.

PRs in flight

• PR fix(provider): skip empty-text filtering for assistant messages in normalizeMessages (#16748) #16750 (this branch's upstream PR, opened 2026-03-09): The fix under review. Skips empty-text filtering for assistant messages with reasoning blocks in transform.ts, and adds the providerMetadata: {} fallback in message-v2.ts.

• PR fix: preserve thinking block signatures and fix compaction headroom asymmetry #14393 (opened 2026-02-20, still open): An earlier, broader attempt at fixing the signature stripping problem that proposed removing the differentModel guard entirely (always pass metadata through). That PR also bundled a compaction headroom fix. It was never merged — presumably because always passing foreign provider metadata was considered too risky. The current branch's approach is more surgical: keep the guard but patch the specific empty-text-with-no-metadata case.

Summary

The code being modified was never designed for the Opus 4.6 world. The empty-content filter (c285304a, Jan 5) predates adaptive thinking by 5+ weeks and assumed empty parts are always safe to strip. The differentModel metadata guard (021e42c0b, Jan 20) was solving a real cross-provider problem but created a gap where same-model empty text parts lose their metadata lifeline. Both pieces of code were correct when written — the bug is an emergent interaction with Opus 4.6's signature-sensitive thinking blocks.