fix(providers): route OpenAI responses-only models to /v1/responses (#5842)

[diegosouzapw]

Closes #5842

Root cause

The native openai provider is hard-wired to https://api.openai.com/v1/chat/completions with no per-model endpoint routing, yet the curated catalog ships gpt-5.5-pro and gpt-5.4-pro — models OpenAI only serves via /v1/responses. Every request (and "Test all models", which goes through the internal /v1/chat/completions route → handleChatCore) 404s with "only supported in v1/responses" / "not a chat model".

Fix (wires existing plumbing — no new mechanisms)

• Registry (open-sse/config/providers/registry/openai/index.ts): tag gpt-5.5-pro / gpt-5.4-pro with targetFormat: "openai-responses" — the per-model translation mechanism already used by gh/codex/opencode; chatCore picks it up via resolveChatCoreTargetFormat and translates Chat↔Responses both ways.
• Heuristic (open-sse/config/providerModels.ts::getModelTargetFormat): dynamically-synced OpenAI ids ending in -pro (e.g. o1-pro, gpt-5.2-pro from the reporter's list — not in our catalog) resolve to openai-responses. Scoped to the openai alias; mirrors the gh executor's /codex/i routing (9router#102). Keeping the heuristic inside getModelTargetFormat keeps the URL and the body translation in lockstep from a single source of truth.
• Executor (open-sse/executors/default.ts::buildUrl): new case "openai" swaps /chat/completions → /responses when the model's targetFormat is openai-responses, honoring a custom base URL (proxy/gateway) for both endpoints.

Out of scope (noted in the issue): gpt-3.5-turbo-instruct is a legacy /v1/completions model — not in our catalog and OmniRoute has no legacy completions upstream.

Validation (TDD, Hard Rule #18)

• tests/unit/openai-responses-only-models-5842.test.ts written first → 5 failures on the unfixed code (registry untagged, heuristic absent, buildUrl pinned to chat/completions) → 8/8 pass after the fix.
• Neighboring suites: all *responses* + *target-format* unit files → 252/252 pass (includes custom-model-target-format, copilot-gemini-claude-route-no-responses, chatcore-target-format).
• check:provider-consistency → OK (171 REGISTRY entries).
• typecheck:core → only the pre-existing open-sse/executors/base.ts errors, byte-identical on the clean release tip (untouched by this PR).
• ESLint on changed files → clean.

Plan: _tasks/fixes-v3.8.43/5842-openai-responses-only-models.plan.md

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[janeza2]

does this fix cover to codex oauth provider?

[diegosouzapw]

CI: Vitest / dast-smoke (full build) / semgrep pass; the Fast Quality Gates job fails only at its TIA unit step, and every failing test there belongs to the 8 pre-existing base-red files tracked in #5798 — verified: provider-translate-path-golden fails identically (2 pass / 1 fail) on this branch and on the pristine tip, and the opencode-executor failures reproduce on the tip (this PR's -pro heuristic is scoped to the openai alias; opencode resolves via its own alias). This PR's own guards: openai-responses-only-models-5842 8/8, neighboring responses/target-format suites 252/252, executor base suites 71/71. Merging with admin, same rationale as #5896/#5902.