fix(review): stop replay after uncertain DNS failure

[seonghobae]

Issue #1106 replay boundary

A review-tagged orchestrator/free passthrough completion treated a temporary DNS exception, including one nested inside a transport exception, as proof that no provider send began. A candidate transport may already have sent an attempt before surfacing that exception. The request then advanced to another provider without an idempotency contract.

This change keeps the existing ordinary virtual-route DNS behavior. On the review-free route, DNS no longer authorizes cross-provider replay. A statusless retryable transport failure now returns nonretryable typed 502 provider_outcome_unknown. Direct local-slot admission failure and explicit provider model refusal keep their existing distinct behavior.

Exact evidence

• Parent: PR Review gateway: admit tool requests by evidence and stop unsafe free replay #1227 b63cb4823eaac40106c215b9cd2a31e5ec495401.

• RED: python -m pytest -q tests/test_passthrough_provider_failover.py::test_free_review_dns_failure_does_not_authorize_replay → 2 failed. Both direct and wrapped synthetic DNS cases returned the fallback model after two mocked provider calls.

• GREEN at e5481c58abfa0de0fa6ffe487b63382facc1bc9a: the same two cases plus explicit model refusal and local-slot controls → 5 passed. Both DNS cases stop after one mocked provider call with typed nonretryable 502.

• Bounded local run with the existing project .venv (Python 3.14, OpenAI SDK 2.54.0): /Users/seonghobae/orca/workspaces/contextual-orchestrator/fix-review-gateway-own-free-pool-admission-and-r/.venv/bin/python -m pytest -q tests/test_passthrough_provider_failover.py tests/test_provider_error_taxonomy.py tests/test_rate_limit_aware_admission.py → 184 passed, one inherited pytest config warning, process exit 0. git diff --check passed. A preliminary system-Python run used SDK 3.0.0 and failed the pinned-version assertion; it is not the project-interpreter result.

• Local composition probe: cherry-picked this exact patch onto fix(review): reject empty text batch before submission #1243@89b7beafbde9a6952af1861e719e99daf97fa56d without conflict. Probe HEAD 230d2f6d75d9a4d80c87f7c75c3fab9d28011527 has tree 2b22bc2bfc13e7e1f384a0c17d00054c9f2d765e, matching git merge-tree. With the same project interpreter, /Users/seonghobae/orca/workspaces/contextual-orchestrator/fix-review-gateway-own-free-pool-admission-and-r/.venv/bin/python -m pytest -q tests/test_review_gateway_admission_contract_1106.py tests/test_passthrough_provider_failover.py tests/test_provider_error_taxonomy.py tests/test_cost_router_boundaries.py tests/test_batch_request_lineage.py from that probe → 262 passed, one inherited pytest config warning, exit 0. This is local composed-code evidence, not a PR-head hosted gate.

Mocked transport calls are boundary evidence, not wire delivery evidence. Keep Draft/HOLD. No merge, deploy, real provider request, company content, or broad load.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}