Repository navigation
fix: restore native count request contract and safe diagnostics - #50
Merged
Merged
Conversation
…sable AFC Three scoped corrections to the Gemini Developer API preflight counter and request config. 1. _count_request_payload keeps its documented generateContentRequest envelope and now carries the REST-REQUIRED nested `model` field as `models/<selected model>`. The endpoint path and the nested field are both built from the same plan.model, so they cannot name different models. Eligibility still admits only bare `gemini-` ids, so no double-prefix is possible. 2. request_budget() still returns None when native counting is unavailable, but the reason is now visible through a best-effort provider:request_budget_unavailable event, registered on the existing observability.events contribution. Payload carries provider, method, a fixed reason code, the selected model when it is a plain bounded string, and an integer http_status when one exists. Reason codes: unsupported_model, unsupported_route, request_plan_unavailable, request_projection_unavailable, invalid_output_limit, http_error, invalid_response. _count_request_tokens returns (count, reason, http_status) and never emits; request_budget is the single reporting site, so one failed call produces exactly one report. No exception text, response body, request body, URL, credential or environment value enters the payload or the matching log line; exc_info dumps removed from both count paths. A failing subscriber cannot convert a successful or None count into a product failure (ordinary Exception caught, BaseException and cancellation propagate). With no hooks channel the same fixed reason is logged as a warning. 3. AutomaticFunctionCallingConfig(disable=True) is now set for every GenerateContentConfig, not only tool-carrying requests. It is applied before extra_request_params, so a supported explicit override still wins, and it remains an SDK-local field excluded from the count projection. Tests added for both items are WRITTEN BUT UNEXECUTED: this source-only lane is barred from running pytest. See the lane DONE.json for the exact DTU commands and the specific assumptions that need a real run to confirm. Generated with Amplifier Co-Authored-By: Amplifier <240397093+microsoft-amplifier@users.noreply.github.com>
Publish a model identifier in unavailable-count diagnostics only when it exactly matches a documented limit-table key. Generated with Amplifier Co-Authored-By: Amplifier <240397093+microsoft-amplifier@users.noreply.github.com>
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
2 participants
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.
What changed
generateContentRequest.modelfield for nativecountTokensrequests asmodels/<effective model>, derived from the same request plan as the endpoint while preserving the projected generation request.provider:request_budget_unavailablediagnostic with fixed reason codes and an optional HTTP status. It never publishes request/response content, URLs, exception text, credentials, or environment values; model telemetry is restricted to documented exact model IDs.extra_request_paramsoverrides are applied.Why
Native counting needs the complete documented request envelope. When counting is unavailable, callers need a safe explanation without changing the established
None-or-budget decision contract or leaking request data.Validation
pytest -q -p no:cacheprovider -m 'not live' tests, in an isolated environment with Python 3.12.3, Core 1.6.1, google-genai 2.23.0, and httpx 0.28.1.INVALID_ARGUMENTfor the old body, and HTTP 200 with 236 native tokens after adding only the required nested model.The same reviewed provider and Loop #59 were validated together with the existing Context implementation:
Each first SDK prompt count was 443,422. The nonstreaming next-request difference was 302 tokens (0.0681%), within the predeclared 1% tolerance; the other three comparisons were exact. The effective policy budget was 999,200 and its compaction target 499,600.
The final live run made 10 count requests and four generations, all HTTP 200. It verified normalized count/generation payload parity, durable next-request reduction, retained system/first-user/overlay content and final five tool pairs, and no unavailable-count fallback or AFC warning. The 64 synthetic tool results were individually below the unchanged 131,072-byte ingress limit.
Scope:
gemini-3.8-flash, Developer API, SDK 2.23.0, synthetic history, declared tools with modeNONE. This is not an endurance test or live post-compaction tool-execution test. Earlier harness failures were corrected without product changes or weakening assertions; these figures describe the final successful run, not all attempts.Independent read-only review found no ship blockers. The initial commits' statements that tests were unexecuted describe their creation time; the results above were executed afterward against the final candidate.
Compatibility
request_budget()retains its existing public result behavior. Explicit supported AFC overrides remain possible; overriding AFC can now produce the existing override warning on tool-less requests too. No dependency changes or breaking changes are intended.