fix: preserve thinking/redacted_thinking blocks in Anthropic message transforms

[bbartels]

Issue for this PR

Closes#14332

Type of change

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

What does this PR do?

Fixes the Anthropic API error: messages.X.content.Y: 'thinking' or 'redacted_thinking' blocks in the latest assistant message cannot be modified.
The Anthropic API requires that thinking and redacted_thinking blocks in assistant messages are sent back byte-for-byte identical in subsequent requests. Three code paths in normalizeMessages / applyCaching in src/provider/transform.ts were inadvertently modifying these blocks:

1. Empty content filtering removed redacted_thinking blocks. The AI SDK represents these as { type: "reasoning", text: "", providerOptions: { anthropic: { redactedData: "..." } } } — the empty text caused them to be filtered out. Fix: only filter reasoning parts that have no providerOptions.
2. Tool-use reordering split assistant messages containing [thinking, tool_call, text] into two separate messages, changing which message the thinking blocks belong to. Fix: skip the reorder when reasoning blocks are present.
3. Cache control hints could be applied to a reasoning block if it was the last content block in the message. Fix: find the last non-reasoning block for cache annotations instead.

How did you verify your code works?

Added tests covering all three fixes to test/provider/transform.test.ts. All 143 tests pass (130 existing + 13 new).
bun test test/provider/transform.test.ts
143 pass, 0 fail

Screenshots / recordings

N/A - not a UI change.

Checklist

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

[github-actions]

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

Potential Duplicate/Related PRs Found:

1. PR fix(provider): skip empty-text filtering for assistant messages in normalizeMessages (#16748) #16750 - fix(provider): skip empty-text filtering for assistant messages in normalizeMessages

   • Directly related to the empty content filtering issue addressed in the current PR

2. PR fix(provider): preserve assistant message content when reasoning blocks present #21370 - fix(provider): preserve assistant message content when reasoning blocks present

   • Addresses similar issue of preserving assistant messages when reasoning blocks are present

3. PR fix(provider): preserve redacted_thinking blocks and fix signature validation #12131 - fix(provider): preserve redacted_thinking blocks and fix signature validation

   • Related to preserving redacted_thinking block integrity

4. PR fix: replace empty text in reasoning messages to preserve thinking block positions #21860 - fix: replace empty text in reasoning messages to preserve thinking block positions

   • Similar fix for preserving thinking block positions with empty text handling

These PRs appear to be addressing overlapping issues with preserving thinking/reasoning blocks in Anthropic message transforms. The current PR (23755) consolidates fixes for three specific code paths (empty content filtering, tool-use reordering, and cache control hints) that were inadvertently modifying these blocks.

[bbartels]

@rekram1-node Any chance you could have a look? ~20 LOC of actual changes, rest is tests

[codeg-dev]

Confirmed this fixes the three transform.ts paths cleanly — verified against messages.X.content.Y: 'thinking' or 'redacted_thinking' blocks in the latest assistant message cannot be modified on claude-opus-4-x (extended thinking, tool-heavy multi-turn).

One gap worth folding in: message-v2.ts strips the thinking signature before transform.ts ever runs, on a path this PR doesn't touch. In toModelMessagesEffect, the reasoning branch is roughly:

if(part.type==="reasoning"){if(differentModel){/* demote to text, drop providerMetadata */}// providerMetadata only kept when !differentModel}

with differentModel = ${model.providerID}/${model.id} !== ${msg.info.providerID}/${msg.info.modelID}. For Anthropic thinking blocks the providerMetadata carries the cryptographic signature, so dropping it on differentModel re-triggers the exact same cannot be modified 400 on the next turn — even with this PR's transform.ts fix applied.

It reliably reproduces whenever provider/model identity shifts between turns: fallback-model switches, variant changes, and proxy setups where the stored providerID differs from the live one (we hit it through a proxy front where requests resolve under a different provider id than what's persisted in the session). Stripping text/tool metadata there is correct; stripping thinking metadata is not.

Minimal fix that pairs with this PR:

if(part.type==="reasoning"){constisAnthropicReasoning=model.providerID==="anthropic"||model.providerID.includes("anthropic")||model.api?.npm==="@ai-sdk/anthropic"// keep the signature for Anthropic even across differentModel...(differentModel&&!isAnthropicReasoning ? {} : {providerMetadata: part.metadata})}

Happy to open a small follow-up PR (rebased on this) if useful.

---

Disclosure: investigation + draft assisted by an AI agent; reproduction and the message-v2.ts gap were validated against a live multi-provider (CLIProxyAPI) deployment.

[codeg-dev]

Heads-up on the empty-reasoning filter in this PR — the providerOptions != null check is too loose and still lets the 400 through in one case:

if(part.type==="reasoning"){returnpart.text!==""||part.providerOptions!=null}

With Anthropic interleaved thinking across several tool-call rounds, a turn that is cut off (e.g. output-length / max-tokens) can leave trailing reasoning parts that have empty text and providerOptions: {} (an empty object — still != null). Those get kept and re-sent as malformed thinking blocks, and Anthropic rejects the next turn with the very error this PR targets:

messages.N.content.M: `thinking` or `redacted_thinking` blocks in the latest assistant message cannot be modified.

We hit this in production after applying an equivalent of this filter: the error recurred on a long Opus extended-thinking + multi-tool session. Tightening the predicate to require an actual signature/redactedData fixed it:

if(part.type==="reasoning"){if(part.text!=="")returntrueconstpo=part.providerOptions??{}// keep ONLY genuine redacted_thinking / signature-only blocksreturnObject.values(po).some((v)=>v!=null&&(v.signature!=null||v.redactedData!=null))}

Empty reasoning with {} / no signature is a truncated remnant and should be dropped, not replayed. Might be worth folding into this PR. (Related: #30046 fixes the upstream message-v2.ts strip that drops the signature on differentModel.)