Add ToolReferenceChunk, ToolFileChunk, BuiltInConnector enum,
ReferenceID union type (int|string), GuardrailConfig with v1/v2
moderation, ConnectorTool for custom connectors, and guardrails
field on chat/agents/conversation requests.
Add AudioTranscriptionRealtime and AudioSpeech to ModelCapabilities.
Move GuardrailConfig from agents/ to chat/ as shared base type.
Remove bundled OpenAPI spec; SDK now tracks upstream Python SDK.
BREAKING: ReferenceChunk.ReferenceIDs changed from []int to
[]ReferenceID. Use IntRef(n) / StringRef(s) constructors.
1. Add README, LICENSE (MIT), .gitignore, Makefile, CHANGELOG
2. Add Version constant and User-Agent header to all requests
3. Rename SetStream to EnableStream (narrower API surface)
4. Fix FinishReason in CompletionStreamChoice to use typed *FinishReason
5. Type conversation entry Content as chat.Content instead of json.RawMessage
6. Graceful unknown type handling — UnknownEntry, UnknownEvent,
UnknownChunk, UnknownMessage, UnknownAgentTool all return data
instead of erroring on unrecognized discriminator values
7. Type agent tools with AgentTool sealed interface + UnmarshalAgentTool
8. Add pagination params to ListConversations and ListLibraries
9. Move openapi.yaml to docs/openapi.yaml
Add doc.go with package-level godoc for all 15 packages, runnable
example functions for chat completion, streaming, embeddings, and
error handling, plus build-tagged integration tests for live API
validation.
Library management with 17 service methods covering full CRUD for
libraries and documents, multipart upload, text content extraction,
processing status, signed URLs, reprocessing, and sharing management.
Agents lifecycle: create, get, list, update, delete, version switching,
version listing, aliases (set/list/delete). Agent types include tools
(function, web_search, code_interpreter, etc.), guardrails, completion
args. Moderation: text + chat endpoints with categories and scores.
Classification: text + chat endpoints with target result scores.
Add retry logic to all HTTP methods:
- doRetry() core loop with configurable max retries
- Exponential backoff with 0.5-1.5x jitter
- Retry-After header support (seconds and HTTP-date)
- Retry on 429 and 5xx; no retry on 4xx client errors
- Context cancellation respected during retry delays
- Multipart uploads also retry via doRetry()
- 9 new tests: 429/500 recovery, exhaustion, no-retry-on-400, backoff math
