Best possible solution:

Close this partial branch and let maintainers review or land #95542 as the canonical fix for the upstream_error fallback cluster.

Do we have a high-confidence way to reproduce the issue?

Yes at source level: current main joins embedded error payload text, filters it through an auth/auth_permanent/billing-only helper, and treats a null result classification as a completed candidate rather than fallback-worthy failure.

Is this the best way to solve the issue?

No. The matcher addition is useful but too narrow; the best fix must also classify the embedded provider payload boundary, which #95542 does with focused tests and proof.

Security review:

Security review cleared: The diff only adds one internal matcher string and does not touch dependencies, workflows, secrets, package metadata, or code execution surfaces.

AGENTS.md: found and applied where relevant.

What I checked:

• PR diff scope: This PR changes only ERROR_PATTERNS.serverError by adding the upstream request failed matcher string. (src/agents/embedded-agent-helpers/failover-matches.ts:123, 920ff5a45349)
• Current embedded payload filter still drops server errors: Current main routes embedded error payload text through a helper that only returns auth, auth_permanent, or billing; server_error/timeout classifications return null. (src/agents/embedded-agent-runner/result-fallback-classifier.ts:106, c6386178973d)
• Null result classification stops fallback advancement: When a candidate returns ok and the result classifier returns null, the fallback runner builds success for that candidate instead of advancing to the next fallback model. (src/agents/model-fallback.ts:418, c6386178973d)
• Canonical replacement is viable: The replacement PR is open, non-draft, mergeable, clean, and carries proof: sufficient plus status: 👀 ready for maintainer look. (658ff234cb28)
• Canonical replacement covers the missing boundary: The replacement PR maps structured upstream_error/server_error payloads and updates the embedded provider error payload classifier to return server_error with focused regression coverage. (src/agents/embedded-agent-runner/result-fallback-classifier.ts:106, 658ff234cb28)
• Related cluster search: Live GitHub search shows the canonical bug report and overlapping same-day candidate PRs; fix #95519: [Bug]: Fallback should trigger on provider upstream_error / LLM request failed #95542 is the proof-positive clean candidate in the cluster.

Likely related people:

• vincentkoc: Local blame points the central classifier and fallback-loop regions to Vincent Koc in this checkout, and recent live history shows adjacent model fallback work with this person as co-author. (role: current line-history contributor; confidence: medium; commits: 9fd9aa5fcdd2; files: src/agents/embedded-agent-runner/result-fallback-classifier.ts, src/agents/model-fallback.ts)
• steipete: Live commit history shows recent result-fallback classifier, model-fallback, and model-failover documentation work defining this fallback boundary. (role: recent fallback area contributor; confidence: high; commits: 0314819f918a, 6719528316c0, 4b0f16d496e5; files: src/agents/embedded-agent-runner/result-fallback-classifier.ts, src/agents/model-fallback.ts, docs/concepts/model-failover.md)
• fuller-stack-dev: Live history shows recent work passing provider status/code/type descriptors through the same failover classification surface. (role: structured provider-signal contributor; confidence: medium; commits: 89975eea24a5, 0314819f918a; files: src/agents/embedded-agent-helpers/errors.ts, src/agents/embedded-agent-helpers/errors-provider-structured-signals.test.ts)
• takhoffman: The earlier embedded provider business-denial fallback classifier change lists this person in the approval/co-author metadata, making them relevant to this classifier boundary. (role: adjacent reviewer; confidence: medium; commits: 18f94fc83a72; files: src/agents/embedded-agent-runner/result-fallback-classifier.ts)

Codex review notes: model internal, reasoning high; reviewed against c6386178973d.