Sdk V3 Step

Two-level monthly budget readout.

plan_turns_remaining counts down the plan's included allowance; monthly_turns_remaining counts down to the hard stop (allowance plus any customer-authorized overage). Once the plan allowance is exhausted the run keeps going and overage_turns_used tracks the extra units billed at month end. Design: docs/plans/SDK_OVERAGE_ALLOWANCE.md.

Harness introspection for one step, present only when the request set debug=true. Rides the response envelope — deliberately NOT the Turn model, which stays the lean protocol surface.

Shapes mirror the playground inspector (services/evals/runner.py):

• examples_recalled: the few-shot SQL examples selected for this turn ({question, note?}, ranked order).
• prompt_focus: column-narrowing record — {narrowing, columns_total, columns_kept, kept_columns: {table: [column, ...]}}; {narrowing: false} when the layer carries no embeddings or focusing failed.
• context: what fed the prompt — {table_names, metric_names, marts, tags, agent_id}.
• docs: docs turns only — the sub agent's full search trace (queries, ranked hits with scores/content, used flags, answer).
• compaction: present only on steps that compacted the transcript — {messages_before, messages_after, tokens_before, tokens_after, summarized}. Counts only; the summary text lives in the transcript.

What the pre-send figure check found on this step.

Present on every step that produced an answer, whether or not verify_answers was on: the check is what says a figure is not in the data, and a caller running with it off still wants to know that.