Fix ChatGPT subscription generation routing

[gltanaka]

Summary

• route chatgpt/* model smoke tests and normal generation through LiteLLM's Responses API
• bridge Codex login credentials and apply the existing ChatGPT SSE output patch before setup/model tests
• send ChatGPT Responses requests as list-form input built from the final messages, preserving schema-in-prompt instructions
• prevent failed ChatGPT Responses requests from falling through to the unsupported chat-completions endpoint
• retain the existing direct OpenAI GPT-5 Responses behavior and hosted-budget admission checks

Root cause

pdd setup and pdd generate exercised different LiteLLM paths. The setup/model test could validate the Codex subscription, while generation called litellm.completion() for chatgpt/gpt-5.6-sol. That targeted /backend-api/codex/chat/completions, which returned a browser-only Cloudflare challenge. PDD then fell back to Gemini despite the ChatGPT subscription being valid.

The Codex subscription backend works through litellm.responses() and requires list-shaped input.

Tests

• pytest -q tests/test_model_tester.py tests/test_codex_subscription.py — 71 passed
• focused ChatGPT/Codex/Responses regression selection — 43 passed
• existing OpenAI Responses API tests — 4 passed
• python -m py_compile pdd/llm_invoke.py pdd/model_tester.py
• git diff --check

Live verification

Ran an editable install against an authenticated Codex subscription:

PDD_MODEL_DEFAULT=chatgpt/gpt-5.6-sol \
PDD_FORCE=1 \
pdd --local --strength 1.0 generate success_python.prompt \
  --output /home/glt/pdd/codex_responses_smoke.py

The command completed without the Cloudflare response or Gemini fallback:

✓ Step 1 (generate): Cost: $0.000000, Model: chatgpt/gpt-5.6-sol

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[gltanaka]

Changes requested

GitHub does not permit the PR author to submit a formal REQUEST_CHANGES event, but these are blocking findings. The single-call Responses routing is needed and the controlled focused tests support it; the current head is not merge-ready.

P1 — preserve the fix in PDD’s source of truth. The new behavior in pdd/llm_invoke.py:5685-6057 and pdd/model_tester.py:191-272 is not reflected in either checked-in prompt copy. pdd/prompts/llm_invoke_python.prompt:180-183 still says only OpenAI GPT-5 uses responses() and every other model uses completion/batch completion; pdd/prompts/model_tester_python.prompt:17-21 still requires exactly one litellm.completion() call. model_tester also has no correct architecture registration/baseline for pdd/model_tester.py. The repository’s own check confirms this: python -m pdd.ci_drift_heal --dry-run --diff-base origin/main...HEAD --modules llm_invoke model_tester --json exits 1 with llm_invoke = CONFLICT and model_tester = UNBASELINED. A future sync can therefore regenerate the unsupported endpoint and undo this fix. Update the canonical/package prompt copies plus architecture/metadata required by the sync workflow, and make the focused drift check clean.

P1 — the supported batch boundary still uses the broken endpoint. use_responses_api is gated by not use_batch_mode at pdd/llm_invoke.py:5692-5700; a chatgpt/* batch consequently reaches litellm.batch_completion() at pdd/llm_invoke.py:6234-6237. A focused reproduction with PDD_MODEL_DEFAULT=chatgpt/gpt-5.3-codex recorded responses_calls=0, batch_completion_calls=1. That retains the same unsupported chat-completions/Cloudflare failure for a public llm_invoke mode and conflicts with the original first-class-provider contract that every caller can use the subscription. Route batch items through the Responses path, or fail closed with an actionable unsupported-mode error; add a regression proving no ChatGPT request reaches completion or batch completion.

P2 — make the changed subscription tests hermetic. In this dedicated worktree, pytest -q tests/test_model_tester.py tests/test_codex_subscription.py gives 69 passed / 2 failed because test_fallback_reaches_chatgpt_when_anthropic_key_missing and the rewritten structured-output test inherit the project’s legitimate PDD_MODEL_DEFAULT from .env and never reach their mocked ChatGPT row. Pin/delete PDD_MODEL_DEFAULT explicitly in each test as appropriate. With PDD_MODEL_DEFAULT=claude-sonnet-4-6, the same suite passes 71/71.

The contributor’s live pdd generate smoke is useful but only covers an unstructured single request. After the fixes, please also record a real structured subscription invocation and the setup/model-test boundary changed here; batch can remain mocked if a real batch would consume unnecessary allowance.

[gltanaka]

Follow-up review on eeae8fefb — approved after two fix/review cycles.

Resolved the original routing/prompt drift and batch safety issues, then addressed adversarial findings: the ChatGPT setup smoke now rejects empty Responses output, and code-generator text/image inputs retain input_text + input_image structure in the Responses request.

Validation: focused model tester/subscription suite (75 passed); Responses/batch invocation selection (14 passed); syntax and diff checks; ci_drift_heal reports llm_invoke and model_tester IN_SYNC. Current PR-head CodeQL and auto-heal checks are green.

Residual risk: provider calls are mocked; no live ChatGPT subscription smoke was run. I cannot submit a formal GitHub approval because this account is the PR author.

[gltanaka]

Final review/fix loop is complete. The PR head has been rebuilt on current main with the Sol-high-approved product patch (stable patch ID 704cb9725100988db76214d6e6c1efa30c149940).

Validation performed on the equivalent final tree:

• 482 focused product tests passed (test_codex_subscription, test_model_tester, test_llm_invoke)
• 217 protected rollout/verifier tests passed
• exact pre/post rollout loader probes passed (manifest/profile mismatches: 0; coverage: 1.0)
• authenticated ChatGPT subscription smoke passed with chatgpt/gpt-5.6-sol, result PDD_PR2316_SMOKE_OK, finish completed
• adversarial Sol-high review approved after substantive correctness and lifecycle fixes

Prerequisite governance/test-state changes were isolated and merged through #2355, #2356, #2357, #2359, #2360, and #2361. Awaiting this head’s full hosted CI before merge.