fix: extend empty-text gate to Bedrock and Vertex signature namespaces

[omer-koren]

Small follow-up to your PR that extends the hasSignedReasoning gate to cover AWS Bedrock and GCP Vertex AI. Feel free to squash into your PR, take the ideas, or close — whichever is easiest for you.

The bug

The gate currently only matches direct-Anthropic signatures:

const hasSignedReasoning = msg.parts.some(
  (p) => p.type === "reasoning" && (p.metadata as any)?.anthropic?.signature != null,
)

When Claude is hosted on AWS Bedrock (amazon-bedrock/anthropic.*) or GCP Vertex AI, the reasoning part's metadata stores the signature under a provider-scoped namespace:

• metadata.bedrock.signature — AWS Bedrock
• metadata.vertex.signature — GCP Vertex AI

The AI SDK's Anthropic adapter keeps the namespace on the host that answered the call, so the gate never fires for Bedrock/Vertex users. The empty-text substitution is skipped, and on compaction/replay the message still trips Anthropic's:

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

Repro

On Bedrock Claude Opus 4.7 (amazon-bedrock/anthropic.claude-opus-4-7), inspecting a reasoning part from the SQLite session DB:

{
  "type": "reasoning",
  "text": "...",
  "metadata": { "bedrock": { "signature": "Es4EClkIDRABGAIq..." } }
}

With only the anthropic namespace checked, hasSignedReasoning returns false → empty text parts between signed reasoning blocks remain "" → Anthropic rejects on the next turn or manual compact.

The fix

Extend the check to any of the three provider namespaces:

const hasSignedReasoning = msg.parts.some(
  (p) =>
    p.type === "reasoning" &&
    ((p.metadata as any)?.anthropic?.signature != null ||
      (p.metadata as any)?.bedrock?.signature != null ||
      (p.metadata as any)?.vertex?.signature != null),
)

The "non-Anthropic providers' reasoning doesn't position-validate" rationale from your existing comment still holds — OpenAI Responses, Gemini thinking, Copilot reasoningOpaque don't write under these three keys, so they continue to have empty text filtered normally.

Tests

Two new tests alongside your existing one, covering the Bedrock and Vertex paths. Same structure as leaves empty text alone when reasoning has no Anthropic signature — just flipping the metadata namespace.

• 29 tests pass locally (bun test test/session/message-v2.test.ts)
• LSP diagnostics clean on the changed file
• Production-verified against a real Bedrock Opus 4.7 session that was reliably reproducing the cannot be modified error on manual compact — the error no longer triggers after this change

Verification

The aws-sdk/client-bedrock-runtime adapter and google-cloud/vertexai both preserve the provider namespace on streamed response metadata rather than normalizing to anthropic — so the extension is necessary, not redundant.

---

Thanks for the original fix — saved us from a persistent production issue. Happy to adjust/squash/close as you prefer.