[lbarcziova]

BeeAI 0.1.84+ defaults allow_prompt_caching to False for every VertexAIChatModel, which silently stripped the cache_control_injection_points the reasoning agent sets. Every request then re-sent the full context as fresh input: cache read/write dropped to zero and Vertex input cost spiked.

Re-enable it for Vertex Anthropic by passing allow_prompt_caching=True at construction (the adapter honors an explicit value), keyed off the public CHAT_MODEL string rather than a private attribute. Add a unit test that the flag is requested only for Vertex Anthropic, plus an end-to-end test that the transformed request retains its cache injection points.

Assisted-by: Claude Opus 4.8

[qodo-for-packit]

PR Summary by Qodo

Restore prompt caching for Vertex Anthropic models

🐞 Bug fix 🧪 Tests 🕐 20-40 Minutes

---

AI Description

• Explicitly enable caching when constructing Vertex Claude models to preserve reasoning-agent cache
 points.
• Leave Gemini and direct Anthropic caching behavior unchanged.
• Test both the construction argument and cache-point retention in BeeAI’s request transformation.

Diagram

graph TD
  A["CHAT_MODEL"] --> B{"Vertex Claude?"} -->|yes| C["Enable caching"] --> D["BeeAI adapter"] --> F["Request transform"] --> G["Retained cache points"]
  B -->|no| E["BeeAI defaults"] --> D
  H["Reasoning agent"] -->|supplies points| F

Loading

High-Level Assessment

Passing the flag at construction is the direct fix for the adapter’s default; mutating the model afterward was the prior, less reliable approach. Review the separate removal of the private provider-ID normalization for its effect on Vertex Claude adaptive thinking.

Files changed (2) +62 / -36

Bug fix (1) +7 / -9

utils.pyEnable caching when constructing Vertex Claude models +7/-9

Enable caching when constructing Vertex Claude models

• get_chat_model passes allow_prompt_caching=True to BeeAI for Vertex Claude models, while leaving other providers’ construction arguments unchanged. It also removes the post-construction caching assignment and the private LiteLLM provider-ID normalization; Gemini caching remains explicitly disabled.

ymir/agents/utils.py

Tests (1) +55 / -27

test_utils.pyCover caching selection and request transformation +55/-27

Cover caching selection and request transformation

• Parameterized cases check that only Vertex Claude requests the caching constructor argument. A test using the real Vertex adapter verifies that cache injection points survive request transformation when caching is enabled and are stripped when it is disabled.

ymir/agents/tests/unit/test_utils.py

[qodo-for-packit]

Code Review by Qodo

---

Resolved findings

1. Vertex Claude reasoning requests fail ✗ Dismissed

Description

get_chat_model() no longer changes BeeAI's vertexai provider ID to the vertex_ai spelling
needed for adaptive-thinking detection. When reasoning is enabled for a Vertex Claude model, the
deleted correction's documented failure path sends thinking.type=enabled, which Claude Opus 4.8
rejects.

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
Removing the provider-ID correction breaks adaptive-thinking requests for Vertex Claude with reasoning enabled.

## Fix Focus Areas
- ymir/agents/utils.py[77-83]

## Recommended Fix
Keep `allow_prompt_caching=True` at construction, but restore the post-construction `vertexai` to `vertex_ai` provider-ID correction independently of the caching setting. Add coverage for a Vertex Claude model with reasoning enabled.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

---