fix(provider): do not place Bedrock cachePoint after reasoning blocks

[giggling-ginger]

Issue for this PR

Closes #36517

Type of change

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

What does this PR do?

On Amazon Bedrock (Anthropic Claude with extended thinking), opencode puts a prompt-cache breakpoint on the last couple of messages. The AI SDK emits message-level bedrock.cachePoint after every content part. When an assistant message ends with a reasoning block, that produces:

ValidationException: Cache point cannot be inserted after reasoning block.

The turn fails, and every retry rebuilds the same invalid request, so the session is stuck until that message ages out of history.

This PR:

1. For Bedrock assistant messages that end in reasoning, anchors cachePoint on the last non-reasoning content part (or skips it for reasoning-only messages) instead of message-level options.
2. Updates the LLM cache policy so it never marks a pure-reasoning assistant message, and prefers text / last non-reasoning parts.
3. Adds regression tests for both the AI SDK transform path and the Bedrock Converse / cache-policy path.

Bug path

flowchart LR
  A[applyCaching<br/>message-level cachePoint] --> B[AI SDK convert]
  B --> C[push cache after<br/>all content parts]
  C --> D{last part is<br/>reasoning?}
  D -->|yes| E[wire: text → reasoning → cachePoint]
  E --> F[Bedrock 400<br/>ValidationException]
  F --> G[message stays in history]
  G --> H[retry rebuilds same request]
  H --> F
  D -->|no| I[OK — cache after text/tool]

Loading

Fix decision

flowchart TD
  S[Bedrock assistant message<br/>needs a cache breakpoint] --> Q1{last content part<br/>is reasoning?}
  Q1 -->|no| M[keep message-level<br/>cachePoint — old behavior]
  Q1 -->|yes| Q2{any non-reasoning<br/>part exists?}
  Q2 -->|yes| A[anchor cachePoint on<br/>last non-reasoning part]
  Q2 -->|no| K[skip cachePoint<br/>reasoning-only message]
  A --> OK[wire: text → cachePoint → reasoning ✅]
  K --> OK2[no illegal cache after reasoning ✅]
  M --> OK3[wire ends with text/tool ✅]

Loading

Before / after wire layout

flowchart TB
  subgraph before [Before — illegal]
    direction TB
    B1[text] --> B2[reasoning] --> B3[cachePoint ❌]
  end
  subgraph after [After — legal]
    direction TB
    A1[text] --> A2[cachePoint ✅] --> A3[reasoning]
  end

Loading

How did you verify your code works?

• cd packages/opencode && bun test test/provider/transform.test.ts (includes three new cases for bug(bedrock): cachePoint is placed after a reasoning block → ValidationException "Cache point cannot be inserted after reasoning block" #36517)
• cd packages/llm && bun test test/provider/bedrock-converse.test.ts test/cache-policy.test.ts
• cd packages/opencode && bun typecheck
• cd packages/llm && bun typecheck

New regression coverage (6 tests, all pass):

Package	Test
opencode transform	anchors cachePoint on last non-reasoning part
opencode transform	skips cachePoint for reasoning-only assistant
opencode transform	keeps message-level cache when not trailing reasoning
llm cache-policy	latest-assistant anchors on text before reasoning
llm cache-policy	skips markers for reasoning-only assistant
llm bedrock-converse	does not place cachePoint after trailing reasoning

Screenshots / recordings

Flow diagrams: Mermaid charts above (bug path, fix decision, before/after wire).

Test results (local bun test, 6 new cases, all green):

packages/opencode  transform.test.ts          3 pass
packages/llm       cache-policy + converse    3 pass

Checklist

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

[github-actions]

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