fix(provider): skip Azure content-filter preamble frames

[kwakayama]

Follow-up to #3815, from sweeping all 16 catalog models against the live gateway.

No version bump — this PR does not cut a release. Stays at 0.1.1241; the release is cut separately.

Problem

Azure OpenAI opens a stream with an informational content-filter report before any content. Captured live from https://api.veryfront.org/ai/gateway/openai/v1/chat/completions on 2026-08-17 — first data chunk for both gpt-5.4 and gpt-5.5:

{"choices":[],"created":0,"id":"","model":"","object":"",
 "prompt_filter_results":[{"prompt_index":0,"content_filter_results":{
   "hate":{"filtered":false,"severity":"safe"},
   "jailbreak":{"detected":false,"filtered":false},
   "self_harm":{"filtered":false,"severity":"safe"},
   "sexual":{"filtered":false,"severity":"safe"},
   "violence":{"filtered":false,"severity":"safe"}}}]}

Empty choices, no usage. The parser required usage on any content-free frame and failed the whole stream:

ProviderRequestError: invalid successful stream (empty choices event had no usage)

gpt-5.2 does not emit it — it is not on the Azure route.

Honest severity: latent, not an outage

No user is affected today, and I want that on the record rather than overstated.

gpt-5.4 and gpt-5.5 never reach this parser: openai-provider.ts:1308-1310 routes reasoning models to the Responses API, and both qualify (isReasoningCapableGpt5 matches gpt-5.<n> for n >= 2, src/provider/shared/openai-reasoning.ts:29-40; supportsDefaultReasoningParams is true for veryfront-cloud, same file :16-19).

Loki corroborates:

• {service_name="veryfront-server"} |= "empty choices event had no usage" over 30d → 0 hits
• every invalid successful stream in veryfront-server over 7d → 45 hits, all reasoning delta was malformed (the Restore Kimi-compatible streams when providers send null deltas #3815 bug). No other variant.

Why fix it anyway

Kimi is Azure-routed and does use this parser. terraform/staging/variables.tf deploys kimi-k26-staging / kimi-k25-staging on Azure AI Foundry, and Moonshot models go through Chat Completions, not Responses. Kimi does not emit a content-filter preamble today — but that is an Azure deployment setting, not a property of the model. Enabling content filtering there would break Kimi instantly, in exactly the way #3815 just repaired.

Secondary: the guard is wrong on the merits. An empty choices array simply carries no content. Usage-only frames use the same shape and were already tolerated; this is the same case. There was no coverage at all — grep -rn prompt_filter_results over extensions/ and src/ returned nothing.

Fix

Skip content-free frames instead of demanding usage. A missing choices key stays rejected — that is a malformed event rather than a content-free one, and that guard keeps its test.

Two files, 41 insertions.

Verification

TDD red → green: the new test fails on main with the exact production error, then passes.

Replaying the verbatim captured streams:

                 BEFORE (main)                          AFTER
gpt-5.4   invalid successful stream (empty…)   parts: {"text-delta":2,"finish":1}
gpt-5.5   invalid successful stream (empty…)   parts: {"text-delta":2,"finish":1}
gpt-5.2   parts: {"text-delta":2,"finish":1}   parts: {"text-delta":2,"finish":1}

Kimi captures still parse fully after this change — text stream 23 reasoning deltas, tool stream through to tool-call + finish.

deno test extensions/ext-llm-openai/ src/utils/version.test.ts — 11 passed (178 steps), 0 failed.

Fixes veryfront/veryfront-issue-inbox#543

[coderabbitai]

Important

Review available on request

• 🔍 Trigger review

Reviews should be triggered manually for repositories with fewer than 10 stars. Select Trigger review above or comment @coderabbitai review to review the latest changes. For a full review, comment @coderabbitai full review.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: cec647d1-7146-4c4c-b143-c857e3fe1183

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[github-actions]

📦 Client bundle boundary

Entrypoint	Modules	Source size	Server leaks
src/index.client.ts	325	1937 KiB	✅ 0

A server module in a client graph aborts hydration in the browser. New leaks fail CI; known leaks are tracked in scripts/lint/client-bundle-baseline.json to burn down.