Summary

• preserve text.format through transformer defaults and fast-session tuning
• resolve prompt_cache_retention from the request, provider options, or model config
• add focused regression coverage for the request-contract fields

Stack

• base: #312

note: greptile review for oc-chatgpt-multi-auth. cite files like lib/foo.ts:123. confirm regression tests + windows concurrency/token redaction coverage.

Greptile Summary

this pr fixes two quiet but impactful request-contract bugs: text.format was being silently dropped on transformer re-assignment (now preserved via object spread in both the normal and fast-session paths), and prompt_cache_retention had no resolution cascade (now resolved body → providerOptions.openai → modelConfig, consistent with how textVerbosity and include work). it also wires the codex-max short alias into the model map and tightens the PromptCacheRetention type from bare string to a branded open union for better autocomplete.

• resolvePromptCacheRetention follows the same three-tier priority pattern already established by resolveTextVerbosity and resolveInclude — no concurrency risk, pure function.
• text.format preservation is correct in both the normal path (body.text = { ...body.text, verbosity: ... }) and the fast-session path (body.text = { ...body.text, verbosity: "low" }).
• no windows filesystem or token safety risks introduced — changes are in-memory request transforms only.
• two vitest coverage gaps remain: the new codex-max alias is not asserted in model-map.test.ts, and the fast-session + text.format combination has no dedicated test.

Confidence Score: 5/5

• safe to merge — logic is correct, no runtime risk, two coverage gaps are non-blocking follow-ups.
• the core fixes (format preservation and cache-retention cascade) are correct, follow established patterns in the codebase, and are well-tested for the primary paths. the two p2 items (missing alias test, missing fast-session+format test) are coverage improvements, not bugs. no concurrency, windows filesystem, or token safety risks introduced.
• test/model-map.test.ts and test/request-transformer.test.ts could use the two additional coverage cases called out in comments, but neither blocks merge.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[incoming RequestBody] --> B{body.prompt_cache_retention set?}
    B -- yes --> E[use body value]
    B -- no --> C{providerOptions.openai\n.promptCacheRetention set?}
    C -- yes --> E
    C -- no --> D{modelConfig\n.promptCacheRetention set?}
    D -- yes --> E
    D -- no --> F[omit field — no-op]
    E --> G[assign body.prompt_cache_retention]

    A --> H{body.text.format set?}
    H -- yes --> I[spread: body.text = ...\nbody.text + verbosity override]
    H -- no --> I
    I --> J{fastSession?}
    J -- yes --> K[spread: body.text = ...\nbody.text + verbosity = low\nformat preserved]
    J -- no --> L[body.text.format preserved]
    K --> L

This is a comment left during a code review.
Path: test/model-map.test.ts
Line: 14-19

Comment:
**missing vitest coverage for new `codex-max` alias**

`addReasoningAliases("codex-max", "gpt-5.1-codex-max")` is added in model-map.ts but the `MODEL_MAP` snapshot test never asserts that `codex-max`, `codex-max-high`, or `codex-max-xhigh` resolve to `"gpt-5.1-codex-max"`. if the alias registration order or canonical target ever regresses, nothing catches it.

```suggestion
	it("keeps codex families canonical", () => {
		expect(MODEL_MAP["gpt-5-codex"]).toBe("gpt-5-codex");
		expect(MODEL_MAP["gpt-5.3-codex-spark-high"]).toBe("gpt-5-codex");
		expect(MODEL_MAP["gpt-5.1-codex-max-xhigh"]).toBe("gpt-5.1-codex-max");
		expect(MODEL_MAP["codex-max"]).toBe("gpt-5.1-codex-max");
		expect(MODEL_MAP["codex-max-high"]).toBe("gpt-5.1-codex-max");
		expect(MODEL_MAP["codex-mini-latest"]).toBe("gpt-5.1-codex-mini");
	});
```

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: test/request-transformer.test.ts
Line: 684-706

Comment:
**missing fast-session + `text.format` coverage**

the existing `preserves text.format` test exercises the normal path. the fast-session branch (`shouldApplyFastSessionTuning = true`) also does `body.text = { ...body.text, verbosity: "low" }`, but no test passes `text.format` through that branch to confirm format survives there too. without it, a future change to the fast-session block could silently clobber `text.format`.

add a sibling test that calls `transformRequestBody` with the `isFastSession` flag set to `true` and a non-null `text.format`, then asserts `result.text?.format` equals the input value and `result.text?.verbosity === "low"`.

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (5): Last reviewed commit: "Merge main into feat/openai-parity-pr5" | Re-trigger Greptile}