fix(ai): advance Anthropic cache marker across tool results

[sethkarten]

Summary

• include tool-result messages when placing Anthropic-style rolling cache markers
• allow tool message text content to carry cache_control
• add a regression test for sessions with tool-call-only assistant turns

Linear: ENG-5823

Why

Prime Agent's Python kernel loop often emits assistant messages with content: null after the first turn, while each kernel result is sent as a tool message. The cache-marker scan skipped tool messages, so it walked back to an earlier assistant text block and left the rolling breakpoint frozen there. Every later tool call and result was then resent as uncached input.

Test plan

• npm test -w packages/ai
• npx biome check packages/ai/src/providers/openai-completions.ts packages/ai/test/openai-completions-cache-control-format.test.ts
• npx tsgo --noEmit

---

Note

Low Risk
Narrow change to where cache markers are attached on Anthropic-format completions; improves caching behavior without touching auth or data handling.

Overview
Fixes Anthropic-compatible rolling prompt cache placement so the ephemeral cache_control marker can land on the latest tool message, not only user or assistant turns.

In openai-completions.ts, the backward scan in addCacheControlToLastConversationMessage and the helpers addCacheControlToMessage / addCacheControlToTextContent now treat role: "tool" like other conversational messages. That matters when assistant turns are tool-call-only (e.g. empty text) and each kernel step ends with a tool result—the marker previously stuck on an earlier assistant block, so later tool traffic was treated as uncached input.

Adds a regression test that asserts the last outbound message is a tool payload with cache_control on its text part, plus a short changelog note.

^{Reviewed by Cursor Bugbot for commit eb64556. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix Anthropic cache marker to advance across tool result messages

Previously the rolling cache_control marker could only land on user or assistant messages, so it never advanced past a tool result. Expands addCacheControlToLastConversationMessage, addCacheControlToMessage, and addCacheControlToTextContent in openai-completions.ts to also consider role: 'tool' messages when placing the ephemeral cache marker. Adds a test confirming the marker lands on the final tool message in a conversation ending with a tool result.

Changes since #1927 opened

• Updated model identifier in test case [eb64556]

^{Macroscope summarized be91ea9.}