[andrewpark-scaleai]

Why

Some step failures can't be fixed by a retry, and the agent-env worker retries them anyway. A step that finds its input missing from the run's context fails the same way on every attempt, because each attempt starts from that same context. Datadog shows the scale of it over the last 3 days of run_task_step failures:

Check	Steps failed	Failed all 3 attempts
load_artifact: no artifact recorded by step X	1,868	1,868
deploy_env: env already deployed	387	386
rubrics_verifier: PromptResponse not found in context	160	160

Each of those spends two more attempts and their backoff before the run fails. The worker can't tell these apart from a transient error, because every one is a plain RuntimeError.

What

• agent_env.task_step.StepContextError (a RuntimeError) marks a step the run's context can't satisfy. It covers two cases. Something the step needs from an earlier step is missing, or something it would create is already there.
• 47 raises across 23 step files now use it. They cover every not found in context.deployed_envs / deployed_agents / deployed_sandboxes check, every is already deployed check, the PromptResponse ... not found in context checks, and No artifact recorded by step. A codemod made the change, and rerunning it changes nothing.
• test_step_context_error.py fails if a step raises one of these as a plain RuntimeError again.

Callers that catch RuntimeError behave the same. The worker change that reads the new type is scaleapi/scaleapi#174619.

Validation

The new guard test fails on the step files before this change and passes after it. Three existing tests now expect StepContextError. In the full unit suite, 14,520 tests pass and 2 fail. test_mcp_cli_builder.py::test_codegen_loads_sibling_env_file fails on main too, because its subprocess can't import httpx here. test_a_repl_left_idle_is_closed_... passed when rerun.

Confidence Score: 5/5

The PR appears safe to merge; no blocking issues remain.

Summary

Adds StepContextError, a RuntimeError subclass, so callers can identify steps that cannot succeed with the run's current context.

• Task steps mark context failures with a distinct error type.

The previous finding is addressed. No new actionable issues were found in the changes since the last review.

_{Reviews (2) · Last reviewed commit: "test: the StepContextError guard reads r..." · Reviewed by Greptile}