Ask the provider why a prompt cache did or did not match - #62
Merged
Merged
Conversation
The OpenAI Responses surface can be asked to compare a request against an earlier response and to say what it found: `comparison_response_id` inside `prompt_cache_options` on the request, `prompt_cache_diagnostics` on the reply. A caller could not use either. The kwarg was not in the parameter allow-list, so it was dropped with an unknown-parameter warning before it reached the wire, and the parser read `output` and `usage` and nothing else, so the verdict was discarded on the way back. Both halves now work, behind a gate narrower than the one on the explicit markers beside them. The markers degrade to today's behaviour where they are unsupported; this field fails the whole request. Measured on 2026-09-12: an Azure v1 Responses deployment answered `prompt_cache_options. comparison_response_id` with HTTP 400, `invalid_request_error`, code `unknown_parameter`, and returned no diagnostics object on any successful reply. So the predicate reads the resolved model family AND a first-party provider, never a deployment label — on a re-hosted surface the model name is whatever an operator typed and is not evidence about the generation underneath. No second response-id field: the Responses `resp_…` id already rides `ResponseMetadata.request_id`, which is the id a later request names as its comparison. The diagnostics object is attached only when the provider returns one, so its absence stays a fact rather than a null. Every other payload, on every other leg and on every model without the capability, is byte-identical to today's.
The predicate's docstring claimed a resolved model family. Nothing resolves one at request time: the generation comes off the model name through the same regex the explicit markers use, and on a re-hosted surface that name is whatever an operator typed on the deployment. What keeps a label out of the decision is the provider set, which admits first-party endpoints alone — and there the name is the model, so no deployment label is ever consulted. The docstring and the test that pins it now say that, because the next reader who widens the set on the strength of a family read would re-earn the measured 400.
This was referenced Sep 12, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The OpenAI Responses surface can be asked to compare a request against an earlier response and to say what it found —
comparison_response_idinsideprompt_cache_optionson the request,prompt_cache_diagnosticson the reply. A caller could use neither: the kwarg was not in the parameter allow-list, so it was dropped with an unknown-parameter warning before it reached the wire, and the parser readoutputandusageand nothing else, so the verdict was discarded on the way back.Both halves work now, behind a gate narrower than the one on the explicit markers beside them, because the failure modes are not alike. The markers degrade to today's behaviour where they are unsupported. This field fails the whole request.
That is measured, not assumed. On 2026-09-12 an Azure v1 Responses deployment answered
prompt_cache_options.comparison_response_idwith HTTP 400,invalid_request_error, codeunknown_parameter, and returned no diagnostics object on any of three successful replies. So the predicate reads the resolved model family and a first-party provider, never a deployment label: on a re-hosted surface the model name is whatever an operator typed, and it is not evidence about the generation underneath.No second response-id field. The Responses
resp_…id already ridesResponseMetadata.request_id, and that is the id a later request names as its comparison; a second field for one fact is two places for it to be wrong. The diagnostics object is attached only when the provider returns one, so its absence stays a fact rather than a null.Every other payload — every other leg, every model without the capability — is byte-identical to today's, and
tests/models/test_openai_prompt_cache_diagnostics.pypins that alongside the rest.