[chrarnoldus]

Summary

On the "Sign in with ChatGPT" route (openai-chatgpt provider on POST /responses), OpenAI rejects some requests that the gateway could easily send in an accepted form. The delegated transform now fixes two of them:

• service_tier: subscription sharing accepts only auto and default, which behave the same as sending no tier. Any other value gets 400 subscription_sharing_unsupported_capability. The transform now always drops the field.
• max_output_tokens: OpenAI rejects values below 16 with 400 integer_below_min_value. The transform now raises them to 16.
• temperature / top_p: the recent reasoning models served on this route reject sampling parameters with 400 unsupported_parameter. The transform now drops both.

The token-refresh change that was originally part of this PR now lives in #8557, so it can be reviewed separately. The remaining error classes are tracked in AISAGE-759.

Verification

• Checked production logs for this route: both rejections occur, with exactly the error codes and parameters this change handles.
• Ran all openai-chatgpt/ suites and upstream-attempt.test.ts against local PostgreSQL; all pass.
• Ran pnpm --filter @kilocode/web-shared typecheck and oxlint on the changed files; both are clean.
• No manual end-to-end test: it needs a live ChatGPT subscription-sharing connection.

Visual Changes

N/A

Reviewer Notes

• Dropping temperature and top_p unconditionally also affects older non-reasoning models (for example gpt-4o) if they are served on this route. For those models it removes a caller preference rather than fixing an error.
• Dropping auto and default assumes they behave the same as sending no tier on this route. Since subscription sharing allows no other tier, sending one gains nothing.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Executive Summary

The incremental commit drops temperature and top_p on the delegated ChatGPT Responses route alongside the existing service_tier and max_output_tokens normalization, and adds a focused test; the ordering leaves no later step re-adding the removed fields.

Assumptions: [INFERENCE] Tests and typecheck were not executed in this read-only review (no code execution permitted), so verification is by reading the changed code and its callers. Dropping sampling parameters also affects any non-reasoning model served by the partner project; the PR description already records this tradeoff. Whether the partner project currently serves non-reasoning models such as gpt-4o could not be confirmed from the repository.

Verification Notes

• provider.transformRequest runs last in apply-provider-specific-logic.ts:409, and no earlier helper re-adds temperature or top_p after the deletes at routing.ts:292-293.
• fixResponsesRequest, removeUnsupportedRequestServiceTier, and the model-setting helpers all run before the provider transform, so the removals are the final write to those fields.
• The new test uses the same transformResponsesRequest helper as the neighboring service_tier and max_output_tokens cases and asserts absence with in checks, so it fails if either field is reintroduced.

Files Reviewed (2 files)

• packages/web-shared/src/lib/ai-gateway/openai-chatgpt/routing.ts
• packages/web-shared/src/lib/ai-gateway/openai-chatgpt/routing.test.ts

Previous Review Summaries (2 snapshots, latest commit df5b912)

Current summary above is authoritative. Previous snapshots are kept for context only.

Previous review (commit df5b912)

Status: No Issues Found | Recommendation: Merge

Executive Summary

The incremental change narrows the PR to two files: service_tier is now dropped unconditionally on the delegated route and the token-expiry work is fully reverted out of the branch.

Assumptions: [INFERENCE] node_modules is not present in the review sandbox, so pnpm --filter @kilocode/web-shared typecheck and the Jest suites could not be re-run; findings are based on reading the changed code and its callers. max_output_tokens: null still passes through untouched, but that matches pre-change behavior on this route.

Verification Notes

• Dropping service_tier remains the last write to it: removeUnsupportedRequestServiceTier and applyKiloExclusiveModelSettings run before provider.transformRequest in apply-provider-specific-logic.ts:371-409, so no later step re-adds a tier.
• The reverted token-refresh code left no dangling references: no remaining imports of rejected-token.ts or expireOpenAiChatGptAccessToken exist in the repository.
• The new clamp only touches numeric values below 16 and leaves absent (undefined) or non-numeric values untouched, consistent with the added tests.

Files Reviewed (2 files)

• packages/web-shared/src/lib/ai-gateway/openai-chatgpt/routing.ts - no issues
• packages/web-shared/src/lib/ai-gateway/openai-chatgpt/routing.test.ts - no issues

Previous review (commit d16453c)

Status: No Issues Found | Recommendation: Merge

Executive Summary

The three delegated-route fixes are small, well-scoped, and consistent with the existing ChatGPT connection handling; the token-expiry compare-and-set under the row lock correctly avoids clobbering a concurrent refresh or reconnect.

Assumptions: [INFERENCE] node_modules is not present in the review sandbox, so pnpm --filter @kilocode/web-shared typecheck and the Jest suites could not be re-run; the review is based on reading the changed code and its callers. max_output_tokens: null and replica lag on readDb are unproven edge cases noted but not flagged, since neither is demonstrably wrong on the changed lines.

Files Reviewed (6 files)

• packages/web-shared/src/lib/ai-gateway/openai-chatgpt/rejected-token.ts - no issues
• packages/web-shared/src/lib/ai-gateway/openai-chatgpt/routing.ts - no issues
• packages/web-shared/src/lib/ai-gateway/openai-chatgpt/routing.test.ts - no issues
• packages/web-shared/src/lib/ai-gateway/openai-chatgpt/store.ts - no issues
• packages/web-shared/src/lib/ai-gateway/providers/upstream-attempt.ts - no issues
• packages/web-shared/src/lib/ai-gateway/providers/upstream-attempt.test.ts - no issues

---

_{Reviewed by deepseek-flash-latest · Input: 36.5K · Output: 7.2K · Cached: 680.6K}

_{Review guidance: REVIEW.md from base branch main}