arize-phoenix-v20.4.0
Breaking Changes📦 arize-phoenixView on GitHub →
⚠ 1 breaking✨ 10 features🐛 12 fixes🔧 11 symbols
Summary
This release introduces new features like in-process Phoenix MCP toolset for agents, retrieval relevance evaluator, and AI Query for trace filter DSL. It also includes several bug fixes and dependency updates.
⚠️ Breaking Changes
- The GET /v1/model_providers endpoint no longer returns custom providers or a `next_cursor`. Built-in entries now expose `provider` instead of `kind` + `provider_key`. Custom providers have been moved to the unreleased GET /v1/custom_model_providers endpoint. No published client is affected by this change as the endpoint was unreleased.
✨ New Features
- Added in-process Phoenix MCP toolset for agents.
- Applied the 2026 brand to the Phoenix documentation.
- Added recording of Harbor jobs in Phoenix client.
- Added retrieval relevance evaluator for evals.
- Added support for gemini-3.7-flash in the playground.
- Enabled phoenix-gql mutations by default with approval in manual mode for pxi.
- Replaced client-action tools with browser-action meta-tools for pxi.
- Added GET /model_providers REST endpoint to the server.
- Added PATCH /projects/{project_identifier}/retention endpoint to the server.
- Added AI Query for the trace filter DSL.
🐛 Bug Fixes
- Forbade extra fields on UI context models for agents.
- Polished password reset and user management for auth.
- Preserved prompt cache usage for chat completions.
- Updated built-in model token prices.
- Updated arize-phoenix-client to version 3.2.0.
- Updated arize-phoenix-evals to version 3.5.0.
- Updated arize-phoenix-evals to version 3.5.1.
- Prevented echoing submitted tool JSON in evaluator prompt validation errors.
- Handled missing annotation labels with pandas 3.
- Parenthesized an operator key in a postgres JSON read for mcp/sql.
- Parenthesized BETWEEN and IN keys in a postgres JSON read for mcp/sql.
- Upgraded to anthropic SDK v1.