[flora131]

Summary

models.dev removed moonshotai/Kimi-K2.6 from Baseten (Kimi-K3 replaces it), so two packages/ai tests no longer typecheck on a fresh build: on main and on every branch. This updates them, following pi's practice of updating tests when a model leaves the catalog (pi e86b0547, 99fe1489), so drift stays visible. pi still has the K2.6 test.

Changes

• image-tool-result.test.ts: the live Baseten provider block (runs only with BASETEN_API_KEY) uses Kimi-K3, which also accepts image input.

• baseten-models.test.ts: the toggle-only reasoning test used K2.6 as its example. No current Baseten model is toggle-only, but the path is still live in the generator and runtime, so the test builds a toggle-only model on Kimi-K3 and keeps its runtime assertions (levels off/high, chat_template_args.enable_thinking: true, no reasoning_effort).

• baseten-qwen-provider.test.ts (coding-agent): models.dev removed ten more Baseten models after the first commit (DeepSeek-V4-Pro, Kimi-K2.5/K2.6/K2.7-Code, Nemotron-120B-A12B, both Inkling models, GLM-4.7, GLM-5, GLM-5.1). The exact-list assertion now matches the current 14-model catalog.

Notes

• The first CI run also failed codemode-worker-admission "holds shared capacity until actual worker exit after timeout" on Windows. That test is load-sensitive and unrelated to this change (it passes in isolation). It is tracked separately.
• The two assertions that only restated the fixture were dropped, so the generator's toggle-only mapping has no live model to check until Baseten lists one again.
• The Baseten default (zai-org/GLM-5.3) is still in the catalog; no product change is needed.

Assistant-model: Claude Opus 5.5
Assistant-workflow: inline
Assistant-duration: 30m converged, estimated 15m
Assistant-verification: bun typecheck failed before and passed after: packages/ai typecheck on origin/main 35bee21 with a freshly generated catalog (2 TS2345 -> 0)
Assistant-verification: vitest packages/ai passed: baseten-models (5 tests)
Assistant-verification: npm run check passed: biome, typecheck, shrinkwrap
Assistant-verification: npm run test --workspace=@bastani/atomic passed except 1 load-sensitive codemode timing test, which passed alone twice (codemode-worker-admission, 8/8)
User-preference: when models.dev removes a model, update the tests that name it rather than decoupling tests from live IDs, so drift stays visible (matches pi)
Co-authored-by: Norin Lavaee nlavaee@umich.edu

Confidence Score: 5/5

Safe to merge based on the reviewed changes: they update test model references and catalog expectations without changing product behavior.

Summary

Updates Baseten tests after models.dev removed older models.

• Baseten AI tests use Kimi-K3 while keeping toggle-only reasoning covered.
• The coding-agent test expects only Baseten model IDs still listed.

No actionable issues or changed-line rule violations were found. No merge-blocking issues remain.

_{Reviews (2) · Last reviewed commit: "test(coding-agent): match the Baseten ca..." · Reviewed by Greptile}

Dismissed because a newer commit was pushed; Greptile will re-review the current head.