Best possible solution:

Keep the current main body-sniffing normalization as the canonical fix and close this replacement branch as redundant hygiene rather than merging another provider-transport change.

Do we have a high-confidence way to reproduce the issue?

Yes for shipped behavior: the linked bug report gives a concrete v2026.6.1 probe path and logs showing HTTP 200 with an empty content-type failing as invalid_provider_content_type. Current main should not reproduce that failure because the guard now normalizes missing-content-type SSE bodies before SDK parsing.

Is this the best way to solve the issue?

No for keeping this PR open: the current main guard is the better canonical fix because the source PR reported that an Accept header alone did not stop the live endpoint from returning no content type. The PR's scoped Accept header is compatible hygiene, but not required to solve the remaining observable problem.

Security review:

Security review cleared: The PR changes scoped provider request handling tests, one request-option helper, and JSDoc only; it does not add dependencies, workflows, credentials handling, package resolution, or new code-execution paths.

AGENTS.md: found and applied where relevant.

What I checked:

• Current main implementation: Current main normalizes successful streaming SDK responses with missing content type by sniffing a cloned body and synthesizing text/event-stream for SSE bytes or JSON for JSON bytes before the OpenAI SDK consumes the stream. (src/agents/provider-transport-fetch.ts:304, e0018382eb00)
• Current main regression coverage: Current main already covers the observed missing-content-type SSE path, including a ChatGPT Codex /responses URL, and asserts the stream parses successfully. (src/agents/provider-transport-fetch.test.ts:245, e0018382eb00)
• PR delta compared with current main: Against current main, the PR does not change provider-transport-fetch.ts; its unique runtime change is the scoped ChatGPT Responses Accept: text/event-stream request option plus tests/JSDoc. (src/agents/openai-transport-stream.ts:1846, ecc88e19f9cc)
• Git history provenance: git blame ties the current main missing-content-type normalization guard to commit 506c2ee, which is the best local proof commit for the implemented behavior. (src/agents/provider-transport-fetch.ts:304, 506c2ee18186)
• Fix timestamp: The proof commit is on current main with commit timestamp 2026-06-04T22:34:53-04:00. (506c2ee18186)
• Release provenance: No tag contains the proof commit, git merge-base --is-ancestor succeeds for current main and fails for v2026.6.1, so this is implemented on current main but not proven shipped in the latest release. (506c2ee18186)

Likely related people:

• Peter Steinberger: Git blame and shortlog show the current main OpenAI transport and guarded fetch implementation, including the missing-content-type normalization guard, comes from the same proof commit by this author in the local history. (role: current implementation owner; confidence: high; commits: 506c2ee18186; files: src/agents/provider-transport-fetch.ts, src/agents/provider-transport-fetch.test.ts, src/agents/openai-transport-stream.ts)
• baanish: The related source PR and closed bug report supplied the live ChatGPT OAuth reproduction, after-fix terminal proof, and original replacement-branch commits that this PR preserved. (role: reporter and source contributor; confidence: medium; commits: 68def1774edf, dd3d67d41414, 50d84e1b3852; files: src/agents/openai-transport-stream.ts, src/agents/provider-transport-fetch.test.ts, src/commands/migrate/skill-selection-prompt.ts)

Codex review notes: model gpt-5.5, reasoning high; reviewed against e0018382eb00; fix evidence: commit 506c2ee18186, main fix timestamp 2026-06-04T22:34:53-04:00.