fix(retry): treat provider usage-limit errors as terminal, not transient

[samirhvbr]

Problem

A subscription usage-limit rejection — e.g. ChatGPT's 5-hour / weekly cap,
returned as HTTP 429 with usage_limit_reached — is currently treated as an
ordinary transient throttle and retried by both retry layers:

• the SDK ModelSettings retry (DEFAULT_MODEL_RETRY, up to 5×), and
• Strix's turn replay (_is_transient_model_error → _run_cycle, up to 5×).

Neither can succeed: the window won't reset for (often) hours. So a run that hits
the cap spends minutes thrashing exponential backoff — flooding the log with
openai.agents: Error streaming response — and burns more of the same exhausted
quota before it finally stops.

Fix

Classify usage-limit errors as terminal so the run fails fast and lands on
the resumable stop path that already exists (run_strix_scan catches the
RateLimitError, logs the strix --resume <run> hint, stops the root):

• codex.is_usage_limit_error(exc) — text-marker predicate, mirroring the
  existing is_content_guardrail_error.
• Excluded from _is_transient_model_error (Strix replay layer).
• Excluded from the SDK retry policy via a thin wrapper _default_model_retry_policy.

Ordinary 429 throttles are unchanged (still retried) — covered by the existing
test_rate_limit_is_retried. New regression tests assert usage-limit is terminal
and that a plain 429 is not misclassified.

ruff format, ruff check, and the suite pass locally.

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR adds text-based detection of provider usage-limit errors and excludes matching failures from both model retry layers so exhausted subscription windows fail quickly.

• Adds usage-limit markers and a shared classifier.
• Composes the classifier into the SDK retry policy.
• Excludes matching failures from turn replay.
• Adds regression coverage for usage-limit and ordinary 429 behavior.

Confidence Score: 4/5

The PR needs adjustment before merging because usage-limit failures from LiteLLM-routed providers can now fail the scan instead of producing the promised resumable stop.

Retry suppression is provider-agnostic, while terminal rate-limit recovery is restricted to OpenAI's exception class, leaving supported non-OpenAI routes on the generic failure path.

Files Needing Attention: strix/core/execution.py, strix/config/models.py, and strix/core/runner.py

Important Files Changed

Filename	Overview
strix/config/codex.py	Adds the shared text-marker predicate used to identify usage-limit failures.
strix/config/models.py	Prevents SDK retries for matching errors, but applies the classification across all provider routes.
strix/core/execution.py	Suppresses turn replay globally, exposing non-OpenAI provider exceptions to a runner that cannot convert them into resumable stops.
tests/test_execution_transient_retry.py	Covers OpenAI usage-limit and ordinary-rate-limit classification but not a LiteLLM-routed exception.
tests/test_model_retry.py	Verifies the composed retry policy declines matching text while retaining ordinary retry behavior.

Prompt To Fix All With AI

### Issue 1
strix/core/execution.py:135-136
**Non-OpenAI limits fail**

When a LiteLLM-routed provider raises a marker-bearing usage-limit error, this classification suppresses both retry layers, but `run_strix_scan` only handles `openai.RateLimitError` as resumable. The provider exception therefore reaches the generic failure path, marking the root failed and re-raising instead of showing the resume hint.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (1): Last reviewed commit: "fix(retry): treat provider usage-limit e..." | Re-trigger Greptile}

[greptile-apps]

Non-OpenAI limits fail

When a LiteLLM-routed provider raises a marker-bearing usage-limit error, this classification suppresses both retry layers, but run_strix_scan only handles openai.RateLimitError as resumable. The provider exception therefore reaches the generic failure path, marking the root failed and re-raising instead of showing the resume hint.

Knowledge Base Used:

• Scan execution lifecycle
• Configuration and provider settings

Prompt To Fix With AI

This is a comment left during a code review.
Path: strix/core/execution.py
Line: 135-136

Comment:
**Non-OpenAI limits fail**

When a LiteLLM-routed provider raises a marker-bearing usage-limit error, this classification suppresses both retry layers, but `run_strix_scan` only handles `openai.RateLimitError` as resumable. The provider exception therefore reaches the generic failure path, marking the root failed and re-raising instead of showing the resume hint.

**Knowledge Base Used:**
- [Scan execution lifecycle](https://app.greptile.com/strix-org-3/-/custom-context/knowledge-base/usestrix/strix/-/docs/scan-execution.md)
- [Configuration and provider settings](https://app.greptile.com/strix-org-3/-/custom-context/knowledge-base/usestrix/strix/-/docs/configuration-and-provider-settings.md)

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[Copilot]

Pull request overview

This PR updates Strix’s retry behavior to treat provider “usage-limit reached” (plan/subscription window exhausted) errors as terminal rather than transient, preventing repeated retries that cannot succeed until the provider’s window resets.

Changes:

• Add codex.is_usage_limit_error(exc) predicate based on provider error text markers.
• Exclude usage-limit errors from Strix’s execution replay retry (execution._is_transient_model_error).
• Exclude usage-limit errors from the SDK-level model retry policy by composing the existing retry policy with an additional “not usage limit” guard.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
strix/config/codex.py	Adds is_usage_limit_error and marker definitions for detecting subscription usage-limit rejections.
strix/core/execution.py	Treats usage-limit errors as non-transient to stop Strix turn replay retries.
strix/config/models.py	Wraps the default SDK retry policy with an additional predicate to prevent retrying usage-limit errors.
tests/test_model_retry.py	Adds regression test ensuring usage-limit errors are not retried by the SDK retry policy.
tests/test_execution_transient_retry.py	Adds regression tests ensuring usage-limit errors are terminal in Strix replay logic and ordinary 429s remain transient.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.