GET /state is a lightweight, read-only endpoint that returns the current CRMPipelineState object without mutating any data. It exposes the episode UUID, the number of steps executed so far, and the active task identifier. No request body is needed. Use it to monitor agent progress, enforce step thresholds before submitting, or debug episode isolation across concurrent runs.
Request
Content-Type header or request body is required.
Response
Before any /reset call
If /reset has never been called since server startup, episode_id will be null and step_count will be 0:
Response Fields
string | null
The UUID that uniquely identifies the current active episode. This is the
same value embedded in
last_action_feedback after a /reset call, and
the path parameter you supply to /grader/{episode_id}.Returns null if no episode has been initialized yet (i.e., /reset has
not been called). Once /reset is called, this value is stable for the
lifetime of the episode and changes only when /reset is called again.integer
The total number of
POST /step calls executed in the current episode.
This counter increments on every /step call regardless of whether
the action succeeded or failed — even -0.05 penalty actions increment it.The minimum step threshold before SUBMIT_PIPELINE is accepted is 3.
Check this field if you need to programmatically verify that the submission
guard has been cleared.string
The task identifier active for the current episode. Possible values are
"t1", "t2", and "t3", corresponding to the Easy, Medium, and Hard
difficulty tiers respectively. This value is set from the TASK_ID
environment variable on the server at the time /reset is called.Use Cases
Checking the minimum-step guard
TheSUBMIT_PIPELINE action is blocked with reward = -0.15 if step_count < 3. Query /state before submitting to avoid the penalty:
Recovering the episode UUID mid-run
If you lost the UUID from the reset response’slast_action_feedback, retrieve it from /state at any time during the episode:
Verifying task identity
When running automated evaluation harnesses across multiple tasks, confirm the active task before sending task-specific actions:Debugging episode isolation
Each call to/reset creates a new episode_id. If episode_id is unexpectedly null or step_count has reset to 0, your agent called /reset without your knowledge (or the server restarted). The state endpoint makes this immediately visible:
GET /state never mutates server state. It is safe to call at any frequency
— for example, after every /step call in a polling loop — without affecting
rewards or episode progression.