Skip to main content
A published case is a reviewed task, a frozen starting world and a Hue-owned outcome check, built from one production trace. Your agent then runs against a fresh copy of that world per case with hue eval, and Hue grades the sealed outcome. This page covers the review path from trace to case and which traces qualify. For what a world is, see simulations.

​
Traces that can become cases

Record with the Hue SDK transport and full content, and name the MCP server on tool spans:
  • Use captureContent: true in TypeScript or capture_content=True in Python. A trace exported without content cannot be reviewed into a case (content_not_recorded).
  • Pass the MCP server identity on tool spans (mcp: on hue.tool), so the trace names the app each call reached.
  • The trace must have reached an app Hue simulates. A trace that reached none is refused with no_simulated_app (HTTP 422).
Case creation also refuses a trace that is still running (running; retry after it ends), evidence over 8 MiB, and a trace not recorded by Hue’s SDK (not_hue_sdk) unless the project accepts generic sources, whose traces then need a full review (generic_task_unreviewed). Direct OTLP and metadata-only capture record traces you can inspect, but not traces you can turn into cases.

​
In the Hue UI

  1. Open Traces and open the trace.
  2. Choose Create case and answer the one question: “Did the agent complete the task correctly?”
  3. Choose Build case. Hue compiles a draft: the task, the starting world the trace shows, and the criteria that decide the outcome.
  4. Review the draft at /case-conversions/<id>/quick: the task, the criteria and the facts the world rests on.
  5. Choose Publish case. Publishing saves the eval-set version when Hue may; otherwise choose Save eval-set version.

​
Over the Hue MCP

With a Read and write connection, a coding agent can run the same review. Writes need the user’s go-ahead; on an organization connection every call takes the case’s project as project_id.
  1. get_trace for the trace’s current revision.
  2. add_case_conversion with trace_id, expected_trace_revision and an idempotency_key.
  3. Poll get_case_conversion until the build finishes, then call it with include_content: true to read the task, the starting world and the criteria you are about to accept. Every write below takes the revision of your latest read as expected_revision; read the draft again after each write before the next one.
  4. If the draft asks “Was this run correct?”, answer with update_case_conversion and run_was_correct.
  5. Review the criteria and accept them: update_case_conversion with reviewed_criteria (accepted_criteria_digest), reviewed_task and authored_closed_world.
  6. Publish only when get_case_conversion reports ready: true, which means trust checks 1 and 2 pass and check 3’s empty run fails: publish_case_conversion with name and the current expected_revision.
The case_from_trace MCP prompt runs this sequence. To save the eval-set version afterwards, use Save eval-set version in Hue, the freeze_eval_set_version tool, or hue eval --save-version. The MCP tool reference lists each tool’s arguments.

​
Tiers

A case publishes at a tier that says how much of its world is verified. Run results report per tier, and judges are advisory: they never decide a case. A trace that still needs a person (T5) is never published.

​
Next step

Make your agent eval-ready once, then run it against the case: Make your agent eval-ready.