Trace Feedback in Session Turn View#

May 15, 2026

Available in arize-phoenix 15.10.0+

The Session Details view now includes a feedback toolbar on each turn. Click the thumbs-up or thumbs-down icon to create a user_feedback trace annotation on the underlying trace. Click the annotate icon to open the full annotation panel for label and score entry.

Feedback is keyed per-viewer using a client-managed identifier, so clicking the same icon a second time removes it (toggle behavior). The annotation summary in the turn footer updates immediately to reflect the new annotation count.

ATIF v1.7 Trajectory Upload#

May 15, 2026

Available in arize-phoenix-client 2.7.0+

upload_atif_trajectories_as_spans now supports ATIF v1.7, which introduces embedded subagent trajectories, trajectory_id-based linking, and deterministic dispatch steps.

Embedded subagents#

Pass a parent trajectory with subagent_trajectories inline and the full multi-agent span tree is built automatically. No separate upload call is needed:

from phoenix.client import Clientfrom phoenix.client.helpers.atif import upload_atif_trajectories_as_spansclient = Client()parent = {    "schema_version": "v1.7",    "trajectory_id": "parent-traj-abc",    "agent": {"model_name": "claude-opus-4"},    "steps": [        {            "role": "agent",            "content": "Delegating to search subagent.",            "tool_calls": [{"id": "tc1", "name": "search_agent", "arguments": {"query": "latest results"}}],        }    ],    "subagent_trajectories": [        {            "schema_version": "v1.7",            "trajectory_id": "child-traj-xyz",            "agent": {"model_name": "gpt-4o"},            "steps": [                {"role": "agent", "content": "Found 5 results.", "tool_calls": []}            ],        }    ],}upload_atif_trajectories_as_spans(client, [parent], project_name="my-project")

The resulting trace nests the child agent’s spans under the parent’s tool span, preserving the delegation hierarchy in the Phoenix UI.

Trajectory IDs and deterministic span IDs#

Trajectories that include a trajectory_id field use it as the canonical span identity key. Re-uploading the same trajectory produces the same span IDs, making uploads idempotent. Trajectories without a trajectory_id fall back to a stable document hash.

Deterministic dispatch steps#

Steps with llm_call_count: 0 represent orchestration that issued tool calls without an LLM invocation (rule-based routing, hard-coded delegation). Phoenix emits TOOL spans for these steps but does not create a synthetic LLM span, matching the actual execution structure.

Session ID support#

Trajectories can now include an optional session_id in the header. When present, Phoenix groups all spans from that trajectory under the specified session, linking them to related traces in the Sessions view.

trajectory = {    "schema_version": "v1.7",    "trajectory_id": "traj-001",    "session_id": "session-abc123",    "agent": {"model_name": "gpt-4o"},    "steps": [...],}