[FD-3805] feat(prompt_agent): grade on max-steps terminal instead of retrying

[williamloo-scale]

What

When a solver's A2A terminal reports failure with error_code == "max_turns_reached", PromptAgentTaskStep.execute now returns the context (logging a warning) instead of raising RuntimeError. The task then flows on to grading (cua_evaluate) with the partial trajectory, which is already recorded in prompt_responses before the failure branch.

Adds MAX_TURNS_ERROR_CODE = "max_turns_reached" to a2a_agent/protocol.py as the cross-process contract string.

Why

A max-steps/max-turns outcome is an expected budget exhaustion, not a retryable crash. Previously it raised, and the Temporal step-attempt retry (cap 3 in the scaleapi worker) re-ran the whole prompt — each attempt burning another full step budget (e.g. 3 × 500 = 1500 steps) and never grading. Now it grades once, no retry.

Other failure modes (no error_code, or a different one) still raise and retry exactly as before.

Companion change (separate repo)

The harness that produces the terminal must emit error_code: "max_turns_reached" on its max-turns branch for this to trigger. That lives in the agent harness/gateway package (not this repo). This PR is forward-compatible: with no harness emitting the code yet, behavior is unchanged. The harness PR and this should land together.

Tests

tst/unit/task_step/test_prompt_agent.py:

• test_max_turns_terminal_grades_without_raising — failed terminal + max_turns_reached → execute returns context, records the prompt_response with the code; no raise.
• test_other_failed_terminal_still_raises — failed terminal without the code still raises RuntimeError.

Full test_prompt_agent.py suite passes (15 tests).

🤖 Generated with Claude Code

Confidence Score: 5/5

The PR appears safe to merge; both previous findings are addressed and no new blocking issue remains.

Summary

Budget-exhausted agent runs return their saved context so downstream steps can grade partial work instead of retrying the prompt.

• Max-turns failures now reach grading without retrying.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Agent terminal received] --> B[Save response and available trajectory]
  B --> C{Failed terminal?}
  C -->|No| F[Continue to grading]
  C -->|Yes| D{max_turns_reached?}
  D -->|Yes| F
  D -->|No| E[Raise failure]
  F --> G[Verifiers allow budget exhaustion]

Loading

_{Reviews (2) · Last reviewed commit: "fix(verifiers): grade max-turns runs ins..." · Reviewed by Greptile}