Repository navigation
fix(fallback): check the first hard-lock combo destination - #7588
Conversation
Check the first resolved combo target against the locked OpenAI source after the existing pre-admission preview. Keep the hard-lock rewrite classification unchanged. Cover sampled-provider admission followed by a forbidden model pick, and describe taken events as guard admission rather than upstream contact or billing.
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info
📝 Walkthrough
Merge Risk: ⚪ Minimal · up to
|
|
✅ Deterministic PR hygiene checks passed. |
Ingwannu
left a comment
There was a problem hiding this comment.
Reviewed the complete diff at ce11e67e8a52291e6a9342ce7ded3b0df787f013. The actual first combo destination is now checked against the original locked OpenAI source after concrete alias/redirect resolution, independently of the random representative selected during pre-admission. The positive and refusal controls distinguish same-provider model fallback from provider fallback and assert the refused destination receives no send. Documentation now correctly describes taken as guard admission rather than upstream contact or billing.
Independent capped Linux/Bun 1.4.3 with fresh private homes: bun test tests/codex-integration/main-account-hard-lock-fallback-off.test.ts — 7 pass / 0 fail / 31 assertions. Current-head hosted CI 39898754531, including the gates/Typecheck step, succeeded. The current review-thread snapshot has no unresolved findings or standing maintainer objection.
The full local suite/typecheck were not repeated on the shared capped host; the completed hosted Ubuntu run supplies broader validation. Approval is bound to this source head and diff, and does not certify a later dev union, native platforms or live provider/client behavior. All external sends in the focused behavior fixtures were stubbed; no real account, runtime configuration, service or OS trust state was changed.
Summary
When main-account hard lock selects a random combo, its admission preview and actual first pick can differ. A previously admitted provider hop could therefore allow a same-provider model hop even when model fallback is disabled.
Check the actual first destination, after aliases and redirects resolve, against the locked source before child dispatch. Keep the existing send budget and account-limit checks, and clarify that a
takenevent records guard admission rather than a physical or billed send.Verification
ce11e67e8a52291e6a9342ce7ded3b0df787f013. The checkout log recordsrefs/pull/7588/mergeatfad50002a9c063f29962c4eaab8458579634edc8, with base parentc51b0b4c875469548e70219a1be276a2395490d2and source parentce11e67e8a52291e6a9342ce7ded3b0df787f013; its Git tree matches the source tree. All four Ubuntu test shards, gates, structure, storage policy, API usage, Docker smoke and the final CI aggregate succeeded. The macOS/Windows jobs were skipped under PR policy.Checklist
Summary by CodeRabbit