feat: relay overlong prompt errors to the harness

[mikasenghaas]

Summary

• relay an overlong prompt to the harness like any other provider failure
• return the provider's original error message with a deterministic 400
• stop synthesizing a context_length body and stopping the trace
• remove the now-unused OverlongPromptError type and model_error phrase-sniffing
• raise a plain ProviderError(status_code=400) at the renderer pre-flight and Responses context_length_exceeded sites

First half of the compaction stack; the harness-side recovery lands on top in the follow-up PR.

Breaking

• Interception no longer stops a trace on an overlong prompt. A harness that does not recover (e.g. by compacting) fails the rollout like any other provider error.
• "context_length" is gone as a stop condition, so Trace.is_truncated no longer reports it.

Verification

• uv run pytest -q tests/v1 — passed; live E2E tests skipped without PRIME_API_KEY
• uv run ruff check / uv run ruff format --check

🤖 Generated with Claude Code

Note

Replace OverlongPromptError with ProviderError(status_code=400)

• Removes OverlongPromptError class and its heuristic phrase-based detection from errors.py; model_error now always returns ProviderError with an explicit status code.
• Renderer preflight overflow in TrainClient.get_response now raises ProviderError(status_code=400) instead of wrapping in OverlongPromptError.
• response_from_wire maps context_length_exceeded and invalid_prompt to status 400, rate limits to 429, and everything else to 502 via model_error.
• Removes dedicated OverlongPromptError branches in both streaming and non-streaming InterceptionServer.request handlers; these errors now flow through the generic RolloutError path.
• Trace.is_truncated no longer treats context_length as a truncation condition.
• Behavioral Change: overlong prompts no longer produce a special context_length trace stop or custom 400 body; they surface as generic ProviderError failures with the provider's status. Callers checking for OverlongPromptError specifically or relying on Trace.is_truncated for context-length stops need to update.

^{Macroscope summarized fe0dad7.}

---

Note

Medium Risk
Breaking behavior for rollouts and any code checking OverlongPromptError or is_truncated for context-length stops; core model-call error handling changes but is scoped to overlong-prompt semantics.

Overview
Overlong prompts are no longer treated as a special “clean truncation” path inside interception. OverlongPromptError is removed from verifiers.v1.errors, along with model_error’s phrase-based context-length detection.

Renderer pre-flight overflow in TrainClient and Responses context_length_exceeded now surface as ProviderError with HTTP 400 (using the renderer’s OverlongPromptError only at the catch site in train). The interception server drops dedicated handlers that stopped the trace with "context_length" and a synthetic rollout stopped: context_length body; those failures follow the normal RolloutError / ProviderError HTTP relay with the provider message.

Trace.is_truncated no longer treats context_length as a stop condition. Harnesses that relied on automatic trace stop or that exception type must handle 400 provider failures explicitly (e.g. compaction in a follow-up).

^{Reviewed by Cursor Bugbot for commit fe0dad7. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This changes production handling of overlong prompts across the renderer, interception server, harness-visible HTTP responses, and trace status: failures that were previously recorded as clean context-length truncations become relayed provider errors. Because this alters existing rollout and reporting semantics across multiple components, human review is appropriate.

^{You can add or adjust custom eligibility rules. Learn more.}