EvalPort adapter for Phoenix datasets/experiments now exists (external package, no code changes needed here)

ATIF converter emits dangling parent refs for subagents spawned without a tool call

[agents] traces for llm tools are incomplete

[harbor] capturing instructions as datasets that are versioned in phoenix

[harbor] trace propagation from inside the sandbox / container back to the task

[mcp] MVP benchmark: pytest eval harness mounting Phoenix MCP on pydantic-ai

[mcp] `phoenix-trace-rca` — root-cause analysis on failing production traces (port/extend `debug_trace`)

Adopt uv workspaces for shared internal Python modules and a single publish flow

[agents] re-implement api key error handling in chat route

[annotations] add the ability to add a span note inline