fix(llm): retry statusless mid-stream provider errors (quota/billing/transient)

[devin-ai-integration]

Summary

Provider errors that surface mid-stream as a bare openai.APIError with no HTTP status code (the SSE data.error path in openai/_streaming.py — the request already returned 200 OK, then the error appears inside the stream) were getting zero retries and failing on the first attempt. Quota/billing exhaustion is the common case, but so is any transient mid-stream blip.

The SDK retry policy only retried errors with a known status code:

policy=retry_policies.any(
    retry_policies.provider_suggested(),
    retry_policies.network_error(),
    retry_policies.http_status((429, 500, 502, 503, 504)),  # status_code is None -> skipped
)

The pre-SDK engine instead retried on code is None or litellm._should_retry(code) — i.e. it did retry statusless errors. Restore that by adding a statusless retry policy:

def _retry_statusless_provider_errors(context):
    n = context.normalized
    if n.is_abort:            # never retry a user cancellation
        return False
    return n.status_code is None

added to DEFAULT_MODEL_RETRY.policy. Retry count (5) and backoff (2→90s, ~62s total) are unchanged. The runner still refuses to replay a stream once content has been emitted (emitted_retry_unsafe_event), so this only retries statusless failures that occur before any content streams — the common shape for quota/rate errors — and a user abort is never retried.

If the provider is genuinely exhausted, the error still propagates and the scan fails after retries — intentionally not a graceful/resumable stop (non-interactive SaaS runs can't be resumed interactively). The pre-existing persistent-RateLimitError handling in run_strix_scan is left untouched.

Testing

• tests/test_model_retry.py (new): drives the composed DEFAULT_MODEL_RETRY policy — statusless → retry, statusless abort → no retry, 400 → no retry, 429/5xx → retry.
• Smoke-tested the real SDK decision path (agents.run_internal.model_retry._evaluate_retry) with DEFAULT_MODEL_RETRY: statusless-before-content → retry=True (delay 2s), attempt 2 → 4s, attempt > max(5) → False, statusless-after-content → False (SDK replay guard), 400 → False. Before/after: with the old policy the statusless quota error returns retry=False; with this PR it returns retry=True.
• ruff, mypy, and the suite pass locally. (Two pre-existing failures in tests/test_runner_root_prompt.py are unrelated — they fail identically on main, from a test fixture missing settings.runtime.)

Link to Devin session: https://app.devin.ai/sessions/dad023e379e942f287bcf6822463b7a4
Requested by: @0xallam

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

This PR restores retries for statusless provider errors raised before streamed content is emitted. The main changes are:

• Adds an abort-aware retry policy for errors without an HTTP status.
• Composes the policy into the default model retry settings.
• Adds tests for statusless, aborted, client, rate-limit, and server errors.

Confidence Score: 5/5

This looks safe to merge.

• No blocking issues found in the updated retry policy or its tests.

Important Files Changed

Filename	Overview
strix/config/models.py	Adds statusless provider errors to the default retry policy while excluding user aborts.
tests/test_model_retry.py	Adds focused coverage for the updated model retry decisions.

_{Reviews (4): Last reviewed commit: "fix(llm): retry statusless mid-stream pr..." | Re-trigger Greptile}

[devin-ai-integration]

@greptile

[devin-ai-integration]

@greptile

[0xallam]

@greptile