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

Review follow-up for the outside-diff Russian collision-preservation finding: that review was anchored to the old 9cdb84c3 head. The current stacked head includes 4e8e7d5 from #652, which already updates the Russian baseUrl row, adds the fixed-provider-endpoints section, and documents the same-named custom-provider preservation case. No additional code or documentation change is needed for that item.

Status after the prerequisite landed: #652 is merged (48f2e8362), so the discovery contract no longer blocks this PR.

I reran the failing Windows job and looked at what it actually reported. The rerun came back cancelled at 12m05s, not failed:

job 90733427766   conclusion: cancelled
started 23:11:50Z   completed 23:23:55Z   (12m05s)
Test: cancelled     later steps: skipped

That is the test job hitting timeout-minutes: 12, which gh pr checks renders as fail. It is not this preset failing. I filed #717 with the run-by-run evidence — the last green Windows run took 11 minutes against the 12-minute ceiling, and every dev run after it was cancelled too, so this is repository-wide rather than specific to #653.

Your code checks out locally on my side:

bun test tests/baseten*.test.ts tests/provider-registry-parity.test.ts tests/model-discovery-contract.test.ts
  38 pass / 0 fail
bun x tsc --noEmit   clean

So this is waiting on CI capacity, not on a change from you. Two things worth doing when you get a chance:

1. Rebase onto current dev. feat(providers): add bounded model discovery contract #652 has landed, and rebasing lets the discovery contract you build on be the merged version rather than the PR-branch version. Worth re-checking whether any modelDiscovery declaration here should now use a lower per-provider limit from the merged contract instead of a local constant.
2. Nothing else — once CI: windows-latest test job cancelled at the 12-minute ceiling, reported as a failing check #717 is sorted and CI can finish, this should go green as-is.

One note for the remaining batches in #572: the first failure on this PR was a genuine Test failure before the rerun, so please do not read this comment as "Windows failures are always timeouts." The distinction is in gh api repos/lidge-jun/opencodex/commits/<sha>/check-runs — conclusion: failure means look at the logs, conclusion: cancelled means the job ran out of time.

Review follow-up is complete for the current head (7e0f7793):

• Baseten-scoped fixes are on this PR, including exact Kimi reasoning allowlist assertions, a non-tautological custom-destination test, zh-CN wording, and documented 1 MiB / 256-row discovery limits across maintained locales.
• Shared contract hardening was intentionally split into fix(providers): harden discovery contract cleanup #743 now that feat(providers): add bounded model discovery contract #652 is merged. It covers .. path-segment rejection plus deterministic registry/cache cleanup in tests.
• I did not apply the out-of-diff auth-cors.ts suggestion because its reported failure mode does not exist: file-backed admin tokens are loaded into management auth state and accepted by requireManagementAuth, while their required ocx_admin_... format is already blocked from provider forwarding by the proxy-admission secret prefix guard.

Local verification for this PR: typecheck, privacy scan, focused tests, docs build, and the full suite all pass (6036 pass / 2 skip / 0 fail).

CI note: the only remaining red check is the Windows matrix job. Its single failure is unrelated to this PR's diff: storage-mutation-race.test.ts completed the test body, then Windows failed to remove its temporary directory with EBUSY: resource busy or locked during teardown. The same job passed 6033 tests, and Ubuntu/macOS plus both npm-global Windows checks are green.

I attempted both the single-job and failed-jobs rerun APIs, but GitHub requires upstream repository admin rights for this run. A maintainer will need to re-run the failed Windows job: https://github.com/lidge-jun/opencodex/actions/runs/30522540376/job/90805964297

NEEDS-CHANGES — evidence only. The code is the strongest of the current preset submissions.

I read the diff rather than the thread first, and there is nothing to argue with technically. https://inference.baseten.co/v1, liveModels: true with sane 1 MiB / 256-row discovery bounds, preserveCustomDestination: true, and no hardcoded selectable catalog — only capability metadata for specific live slugs at registry.ts:412-447, which is the right place to draw that line. tests/baseten-provider.test.ts binds the derived login/init/dashboard contracts, authenticated discovery, the real chat endpoint and Bearer header, reasoning allowlists and their omission for unsupported models, slash-ID routing in both forms, and non-canonical base URL preservation.

I also read the 24 existing reviews so I would not add a 25th copy of something already settled. Those rounds covered the Russian collision documentation, MD028 formatting, numeral agreement, containsAny/containsAll semantics, capability priority before discovery truncation, Kimi reasoning allowlists, the previously tautological custom-destination test, the 1 MiB/256-row documentation across locales, the zh-CN schema/protocol wording, and the discovery-cache and .. handling that moved to #743. As far as I can tell those are resolved or withdrawn.

What is left is a gate that applies to every preset regardless of code quality. MAINTAINERS.md:33-42 treats a canonical preset as a credential-destination change and requires, before merge:

• documented OpenAI-compatible endpoints, including authenticated GET /v1/models since this entry declares liveModels — the diff links the Model APIs overview, API key, reasoning and vision pages, but not a clear primary-source citation of the exact chat endpoint and /v1/models contract
• terms of service and the operating legal entity — missing
• resale or routing authorization — not applicable; Baseten's own material describes them operating deployment and inference APIs, so I am reading them as a first-party host, not a reseller
• a named maintenance owner — missing, and this one matters more than it looks. When an endpoint changes shape a year from now, someone has to notice
• a citable verification date — missing

The PR body requests security review, which is right, but none of the 24 reviews is an approval or an explicit maintainer security sign-off, so that requirement is still open.

What happens next: add the ToS and operating entity, a dated primary-source citation for the chat endpoint and authenticated /v1/models, and name a maintenance owner. Then request explicit security review. This merges cleanly against current dev and dev CI is fully green now.

Thanks for the patience through 24 rounds — the code got better for it.

Addressed the evidence-only changes requested and restacked the single Baseten commit onto current dev (6d54e5f3).

• The PR description now cites Baseten primary sources for the exact Chat Completions endpoint and Bearer-authenticated GET /v1/models, records the 2026-07-31 verification date, identifies Baseten Labs, Inc. and its Terms, explains why resale authorization is not applicable, names @olddonkey as maintenance owner, and discloses no affiliation or compensation.
• The obsolete feat(providers): add bounded model discovery contract #652 stack/dependency wording is removed; [Provider] Promote a verified batch of OpenAI-compatible inference providers #572 is linked without a closing keyword.
• Final local verification: focused provider/contract tests 186/0, full suite 6,494 pass / 4 skip / 0 fail, typecheck and privacy scan clean, docs build 146 pages.
• The current head is ea129f5f; GitHub now reports the PR mergeable.

@Ingwannu, this is an explicit request for the maintainer security review/sign-off required for this credential-destination change. @lidge-jun, the evidence gate from your requested-changes review is now addressed; please re-review when CI completes.

CI follow-up for head ea129f5f: the Windows matrix job is red because Bun 1.3.14 crashed with an internal assertion failure, not because a Baseten assertion failed. The log shows the suite reached storage-mutation-race.test.ts, passed the first two cases, then Bun emitted panic ... Internal assertion failure / This indicates a bug in Bun, not your code after 661s. Ubuntu and macOS main jobs, all npm-global jobs, and the local 6,494-test suite are green.

Failed job: https://github.com/lidge-jun/opencodex/actions/runs/30676226719/job/91303936452

I am not changing the provider diff for this runner crash. A maintainer can rerun the job if a green Windows check is required.