feat: multi-provider LLM SDK v0.1.0 (openai, gemini, deepseek, zai, kimi, anthropic) - #1
Merged
Merged
Conversation
Holistic adversarial review of b2aff9b: 3 sequential agents (contract diff, bug hunt, robustness/coverage/docs) + personal verification of every finding. Streaming/SSE: - parser goroutine no longer leaks on abort/cancel/timeout (pumpSSE now signals done; parser selects on send) - SSE line cap actually works: accumulate ReadSlice fragments to the 1 MiB limit instead of dying at the 64 KiB bufio buffer - partial-output failure is returned with its partial result and never retried (was silently retried, duplicating user-visible output) - keepalive activity flag resets after data events Retry semantics: - buffered 429 whose Retry-After outlives the context surfaces as *RateLimitError (was a bare ctx error hiding the 429) - streamRejected classifier narrowed: requires an explicit rejection phrase, so context-length 400s no longer permanently downgrade to buffered - cancellation during backoff no longer mislabeled "retry exhausted" Provider fidelity: - anthropic ListModels paginates with after_id (page_id never advanced) - o-series/gpt-5 receive max_completion_tokens instead of rejected max_tokens - gemini trailing chunk without usageMetadata no longer wipes accumulated usage - nested OpenAI error envelope parsed into APIError.Message/Code - RateLimitError.Unwrap exposes Status/Retryable via errors.As - learn-once fallbacks moved to Provider: shared by every Chat() client (each client used to re-pay the provider's rejection) Registry: - WithProvider overrides validated at wiring time like custom providers - unauthenticated hint names the real env var (DEEPSEEK_API_KEY) Tests: 18 new (RED-first), incl. CallStream failure orchestration suite that was at zero coverage. Coverage 81.0% -> 83.4%. Lint clean (errcheck, staticcheck S1016 conversions). Docs: README Errors contract extended, PLAN typos fixed.
Second adversarial review cycle (2 sequential hunts + personal verification of every finding, RED-first fixes): Wire fidelity: - Anthropic extended-thinking tool loops: ThinkingSignature added to Message/ChatResult; thinking blocks captured buffered + signature_delta streaming, replayed as first block with signature on request side - zero-value/unknown message roles rejected loudly at the SDK boundary (was: dropped silently on Anthropic/Gemini, sent as user on OpenAI) - Gemini: usage-only stream chunks no longer dropped; functionResponse name recovered from the assistant tool_call (loud error if unresolvable) - canonical finish reasons: unmapped provider stop reasons map to "" (openai/anthropic/gemini), never leak raw values - OpenAI finish-reason default aligned to canonical mapping Streaming/retry: - premature close (200 SSE headers, no completion signal) is a retryable error, never a silent empty success; after deltas -> partial + error - stream path learns reasoning_effort:none like the buffered path - learn-retry and effort-learn budget-guarded; (nil, nil) return hole closed - SetRequestTimeout swapped via atomic pointer (data race fix) Tests/docs: - 45+ new tests (RED-first), coverage 83.4% -> 97.7% - README rewritten: full contract, streaming guarantees, thinking signatures, thread safety, error classification - PLAN.md implementation-status section added
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.
Summary
Canonical Go SDK for six LLM providers, extracted and hardened from odek's battle-tested internal client. Two full adversarial review cycles (contract, concurrency, wire-fidelity — every finding personally verified and fixed RED-first).
Highlights
Quality gates
Test plan