[etwk]

Summary

• Add native Gemini generateContent, countTokens, and transactional SSE streamGenerateContent?alt=sse, including standard Vertex AI project/location paths.
• Preserve native payloads, tool calls, thought signatures, authentication, and trailing usage. Buffer through clean EOF and candidate completion; abort failed committed transports without replay.
• Add independent PROXY_GEMINI_BUFFER_MS configuration (default 600000 ms), Google-shaped errors, API-key archive redaction, regression coverage, and setup/operations documentation.
• Report explicit Gemini thinkingLevel values in access and retry logs, accepting camelCase and snake_case fields. Do not infer levels from token budgets or model defaults.

Reliability fixes

• Decode array-wrapped non-SSE HTTP errors without losing authentication retry classification or error context.
• Accept ProtoJSON candidate-count spellings and numeric representations.
• Make gateway-forbidden Gemini 408/429/5xx failures terminal to SDKs that ignore X-Should-Retry.
• Accept SSE control frames and empty data messages without mistaking them for model output or completion.
• Reject contradictory prompt blocks accompanied by candidates instead of hiding incomplete generation.
• Preserve URL-derived model names in archived attempts, including local retries.
• Preserve HTTP error-body read failures before validation heuristics; incomplete or malformed Gemini error JSON cannot establish a permanent validation error. Retain authoritative retry hints and HTTP cooldowns.
• Preserve attempt deadline causes during HTTP/2 response-body reads, preventing proxy timeouts from becoming terminal client_gone failures across Gemini, Responses, and Messages.
• Honor native google.rpc.RetryInfo.retryDelay for HTTP and in-band SSE errors when no valid positive HTTP Retry-After exists. Round fractional seconds up and prevent long delays from overflowing local retry/deadline arithmetic.

Codex CLI review

• Initial feature: codex review --uncommitted twice; three initial P2 findings fixed before the original PR was published.
• Full-PR follow-up: codex review --base origin/main; both reproduced P2 findings fixed (error-body read failures and native RetryInfo cooldowns).
• The next codex review --uncommitted reproduced an HTTP/2 cancellation-cause regression; fixed at the transport-body boundary with a cross-protocol regression and 3 compiled-proxy HTTP/2 checks.
• Final codex review --uncommitted: no actionable regressions. The reviewer also passed full tests, race tests, 32-bit tests, vet, and diff checks.

Verification

• go test -race -p 1 ./... -count=1
• GOARCH=386 go test ./... (final Codex review)
• go vet ./...
• make release-test (9 tests)
• Compiled-proxy follow-up HTTP checks: 31 passed (16 thinking/stream/archive checks, 11 error/cooldown checks, 1 long-cooldown check with the timed live switch disabled, and 3 HTTP/2 deadline checks).
• Go sources formatted with gofmt.