[chrarnoldus]

Summary

A custom LLM upstream is private infrastructure. Until now its error bodies reached the user unchanged, which can expose internal URLs, model names, or account details. This PR stops forwarding them:

• HTTP errors: makeErrorReadable now answers every non-2xx custom LLM response with the generic upstream_error used for stealth models. A context overflow is still reported so clients can compact, but with a Kilo-authored message instead of the upstream body.
• Stream errors: rewriteModelResponse replaces error details in stream events with the same generic message: chat completions error, Messages error events, and Responses error / response.failed. This applies whenever shouldRedactErrorResponse is true, which is now custom LLMs plus stealth models.
• Upstream detail is still kept internally: the unrewritten response is still logged, so operators keep the upstream detail.

This is a prep PR for #7291, where custom LLMs can become public.

Verification

• No manual testing; covered by unit tests for HTTP errors, context overflow, and each stream API kind.

Visual Changes

N/A

Reviewer Notes

Stealth models already had HTTP errors redacted. Their stream error events are now redacted too, for consistency.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Executive Summary

Reviewed the custom-LLM/stealth stream-error redaction across the AI-gateway helpers and response rewriters; the redaction correctly targets the real HTTP and stream error shapes, no defects found in the changed code.

Files Reviewed (6 files)

• packages/web-shared/src/lib/ai-gateway/kilo-exclusive-models.ts
• packages/web-shared/src/lib/ai-gateway/llm-proxy-helpers.ts
• packages/web-shared/src/lib/ai-gateway/llm-proxy-helpers.test.ts
• packages/web-shared/src/lib/ai-gateway/models.test.ts
• packages/web-shared/src/lib/ai-gateway/rewriteModelResponse.ts
• packages/web-shared/src/lib/ai-gateway/rewriteModelResponse.test.ts

Detail

• shouldRedactErrorResponse now returns true for provider === 'custom', matching shouldRedactModelNameInMicrodollarUsage; the early providerId === 'custom' branch in makeErrorReadable still routes context overflow through detectContextOverflow and builds a Kilo-authored context_length_exceeded body, and otherwise returns the generic upstream_error. Status is preserved and the upstream detail remains available to internal logging, so no operator signal is lost.
• Stream redaction in rewriteModelResponse_ChatCompletions (frame error), rewriteModelResponse_Messages (type === 'error'), and rewriteModelResponse_Responses (type: 'error' plus response.failed via response.error) matches the shapes the gateway actually receives (ResponseErrorEvent carries flat code/message/param; ResponseFailedEvent carries response.error). The event.response?.error guard is safe because successful responses send error: null.
• Tests cover the changed behavior for HTTP errors, context overflow, and each stream API kind, and the pre-existing providerId: 'custom' tests operate on 2xx bodies or logUnrewrittenResponse, so they are unaffected.

---

_{Reviewed by deepseek-v4.1-flash · Input: 141.3K · Output: 33.5K · Cached: 2.2M}

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

[chrarnoldus]

We don't want this.