Use haiku for databricks

[DOsinga]

Summary

We're seeing compaction being super slow internally. Switching to anthropic models that we use for the main model mostly would be a solution

[tlongwell-block]

/goose

[Copilot]

Pull request overview

This PR switches the Databricks fast model from "gemini-2-5-flash" to "databricks-claude-haiku-4-5" to address slow compaction performance. The change aligns with using Anthropic models that are used for main operations.

Changes:

• Updated default fast model constant to use Claude Haiku 4.5 instead of Gemini 2.5 Flash
• Added environment variable override capability to the compaction test script for easier testing

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
crates/goose/src/providers/databricks.rs	Changed DATABRICKS_DEFAULT_FAST_MODEL from "gemini-2-5-flash" to "databricks-claude-haiku-4-5"
scripts/test_compaction.sh	Added COMPACTION_PROVIDER and COMPACTION_MODEL environment variables to override provider/model settings during testing

[Copilot]

The new fast model "databricks-claude-haiku-4-5" should be added to DATABRICKS_KNOWN_MODELS. This list is used for provider metadata to help users discover available models, and the default fast model should be included.

Suggested change

	pub const DATABRICKS_KNOWN_MODELS: &[&str] = &[
	pub const DATABRICKS_KNOWN_MODELS: &[&str] = &[
	"databricks-claude-haiku-4-5",

[github-actions]

PR #6943 Review: Use haiku for databricks

Summary: This PR changes the Databricks provider's default fast model from gemini-2-5-flash to databricks-claude-haiku-4-5 to address slow compaction performance. The change is minimal and correctly scoped. The test script improvements for environment variable overrides are a nice addition.

Analysis

The fast model is used internally for compaction and session naming via complete_fast() - it's not user-facing. The change is valid because:

1. Model naming is correct: The is_claude_model() function in formats/databricks.rs uses .contains("claude"), so databricks-claude-haiku-4-5 will properly be recognized as a Claude model and receive appropriate cache control handling.

2. Not being in KNOWN_MODELS is fine: I initially considered flagging this, but DATABRICKS_KNOWN_MODELS is only used for UI suggestions during model configuration. The fast model is an internal default and the previous value gemini-2-5-flash wasn't in KNOWN_MODELS either.

3. Test script changes are useful: Adding COMPACTION_PROVIDER and COMPACTION_MODEL environment variable overrides allows easier testing of compaction with different provider/model combinations.

✅ Highlights

• Clean, minimal change that addresses the stated problem
• Test script improvements add useful debugging capability without disrupting existing behavior
• The conditional logic for env var overrides is straightforward and well-documented

🟢 Suggestions

1. Optional: Add to KNOWN_MODELS - Consider adding databricks-claude-haiku-4-5 to DATABRICKS_KNOWN_MODELS for completeness. This isn't required for functionality, but would document that this is a valid model name users could select:

const DATABRICKS_KNOWN_MODELS: &[&str] = &[
    "databricks-claude-sonnet-4",
    "databricks-claude-haiku-4-5",  // Add this
    "databricks-meta-llama-3-3-70b-instruct",
    // ...
];

This PR looks good to merge.

---

Review generated by goose