fix: drop the redundant stream_options patch that OpenAI 2.13.0 rejects

[thomasluizon]

Ticket: #364 in the ticket repository, https://github.com/thomasluizon/orbit-tickets/issues/364

main is red and this unblocks it. Every open pull request against this repository inherits the failure, including #494.

What broke

#479, the Dependabot nuget-minor-patch group, bumped OpenAI 2.12.0 to 2.13.0. Nine AiIntentServiceStreamingTests went red.

It was invisible on main because API Tests only ran on pull_request, so a red main surfaced as a single failing SonarCloud Analysis, whose "Test with Coverage" job is what actually runs the suite.

Root cause, traced rather than inferred

The streamed round asked for usage through the experimental SCME0001 JsonPatch API:

options.Patch.Set("$.stream_options.include_usage"u8, true);

In 2.13.0, ChatCompletionOptions gained a PropagateSet override for that path which dereferences a null internal StreamOptions, so it throws NullReferenceException. Measured directly against the installed 2.13.0 assembly:

patch	result
$.foo (top-level bool)	OK
$.a.b (arbitrary nested)	OK
$.stream_options.include_usage	NullReferenceException

So this is specific to stream_options, not a general JsonPatch break. ChatCompletionOptions.StreamOptions is internal in 2.13.0 and there is no public typed replacement, so Patch was the only documented route.

AiIntentService swallowed the NRE into Result.Failure(AiUnavailable), which is why every assertion only ever reported IsSuccess false with no cause. I surfaced it by temporarily rethrowing.

Why this deletes the call instead of pinning the package

The patch was never needed. The SDK already emits "stream_options":{"include_usage":true} for a streamed round on its own. Verified by dumping the outgoing request body with the call removed:

{"temperature":0.1,"messages":[...],"model":"gpt-test","max_completion_tokens":8192,
 "stream":true,"stream_options":{"include_usage":true}}

That holds on both 2.12.0 and 2.13.0.

So the options were: pin to 2.12.0 (2.13.0 is the latest on nuget.org, so there is no forward fix to take instead), or delete code that does nothing. Deleting is strictly better: we stay on the current SDK, we take no Dependabot ignore, and an experimental-API suppression leaves the codebase.

The new test is the guard for that reasoning

We no longer request usage ourselves, so we now depend on SDK behaviour. SendWithToolsAsync_Streaming_RequestsUsageInStreamOptions asserts the outgoing request carries stream_options.include_usage. If a future SDK stops emitting it, per-user token accounting would silently record nothing, and ORB-100's per-request cost bound would quietly stop working. This test fails instead.

API Tests now runs on push to main

arch-map.yml, guards.yml and codeql.yml already do. test.yml did not, which is the reason a red main was only visible as one confusingly-named failing check.

Verification

Full suite, locally:

project	result
Orbit.Domain.Tests	547 passed, 0 failed
Orbit.Application.Tests	3208 passed, 0 failed
Orbit.Infrastructure.Tests	2169 passed, 0 failed
Orbit.Analyzers.Tests	32 passed, 0 failed

Production

The deployed API on 2.13.0 would fail every streamed Astra turn, because the NRE happens before the request is sent and the caller sees AiUnavailable. Confirm live chat after this deploys. Flagged on the ticket as the urgent half.

Nobody merged past a red gate

Worth stating, because it is the natural suspicion. #479 merged at 23:24:26Z, and the only Unit Tests run on that head started at 23:27:56Z, three and a half minutes later, failing at 23:30:03Z. The failure did not exist at merge time.

The missing push gate is the whole explanation: once the package was on main, nothing re-ran the suite there, and the only signal was a job called SonarCloud Analysis whose name gives no hint that it is what runs the tests.

[pullfrog]

Important

The new push trigger reaches pull request only jobs that use an unset github.base_ref, so API Tests will fail on every main push.

Reviewed changes I reviewed the OpenAI streaming regression fix, its request-shape coverage, and the expanded API test workflow trigger.

• Streaming patch removal: Removes the redundant experimental JsonPatch call that OpenAI 2.13.0 rejects while retaining the SDK's emitted usage option.
• Usage regression coverage: Adds an exact outbound JSON assertion for stream_options.include_usage so usage accounting cannot silently lose its SDK prerequisite.
• Main branch test coverage: Adds push coverage for main, but currently exposes pull request only assumptions in two workflow jobs.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes I reviewed the workflow guards added since the prior Pullfrog review and checked them against the complete PR diff and the API Tests job dependency graph.

• Push safe workflow: Scoped migration-guard and openapi-breaking to pull request events, preserving their github.base_ref comparisons while leaving build and unit tests active on pushes to main.

  ^{｜ View workflow run ｜ Using GPT Sol ｜ 𝕏}

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud