fix: preserve Anthropic redacted_thinking blocks in message sanitization

[threepointone]

Problem

_sanitizeMessageForPersistence in AIChatAgent was incorrectly stripping Anthropic's redacted_thinking blocks. These blocks are stored as reasoning parts with empty text but carry critical data in providerMetadata.anthropic.redactedData. When stripped, the Anthropic API rejects subsequent requests because the conversation history is incomplete.

Fixes #978

Root Cause

The sanitizer filtered out all empty reasoning parts unconditionally. This was designed for OpenAI's ephemeral reasoning placeholders but also caught Anthropic's redacted_thinking blocks, which must be preserved for round-tripping.

Solution

Refactored _sanitizeMessageForPersistence from a filter-then-strip approach to a strip-then-filter approach:

1. Strip OpenAI ephemeral metadata (itemId, reasoningEncryptedContent) from all parts first
2. Filter reasoning parts that are truly empty — no text AND no remaining providerMetadata

This is provider-agnostic: any provider that stores meaningful data in providerMetadata on reasoning parts will be preserved automatically. OpenAI placeholders (whose metadata was stripped in step 1) are still correctly removed. Future providers with similar patterns will work without code changes.

How providers use providerMetadata on reasoning parts

Provider	Metadata	Empty text?	Must preserve?
Anthropic	{ anthropic: { redactedData: "..." } }	Yes	Yes — required for API round-trip
Anthropic	{ anthropic: { signature: "..." } }	No (has thinking text)	Yes — but was never affected
OpenAI	{ openai: { itemId: "...", reasoningEncryptedContent: "..." } }	Can be	No — ephemeral, stripped

Changes

• packages/ai-chat/src/index.ts — Reordered sanitization steps + updated JSDoc
• packages/ai-chat/src/tests/sanitize-messages.test.ts — Added 2 new tests:
  • Anthropic redacted_thinking blocks (empty text + providerMetadata.anthropic) are preserved
  • OpenAI empty reasoning placeholders are still removed after strip-then-filter

Testing

All 219 tests pass across 24 test files (npm run test).

[changeset-bot]

🦋 Changeset detected

Latest commit: 14a93f6

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@cloudflare/ai-chat	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/cloudflare/agents@980

npm i https://pkg.pr.new/cloudflare/agents/@cloudflare/ai-chat@980

npm i https://pkg.pr.new/cloudflare/agents/@cloudflare/codemode@980

npm i https://pkg.pr.new/cloudflare/agents/hono-agents@980

commit: 14a93f6