v1.1.4

[1.1.4] — 2026-06-19

Provider-translation correctness release. Tool/function calling, sampling parameters, completion-token limits, finish-reason normalization, multimodal input, and streaming fidelity are now correctly translated across native and OpenAI-compatible providers. Bundles an 11-bump dependency sweep and an internal shared-helper refactor (~1,800 lines of duplication removed) with no public API breaks. Fixes every issue labelled release-1.1.4: #139, #140, #141, #142, #143, #144, and #145.

Fixed

• Tool / function calling dropped on native non-OpenAI providers (issue #139, PR #214): Anthropic, Gemini, Bedrock (Claude), and Cohere now forward tools / tool_choice in each provider's native shape, parse tool_use / tool_calls responses (JSON-object arguments → JSON string), assemble streaming tool-call deltas with index remapping, and handle multi-turn tool round-trips. OpenAI-compatible providers preserve streamed tool_calls deltas via a shared decoder.
• Sampling / output params silently dropped (issue #140): a shared internal/openaicompat request builder forwards the full OpenAI-shaped body (top_p, stop, seed, penalties, response_format, n, …); native providers translate to upstream field names and warn-and-drop only what an API cannot express.
• max_completion_tokens ignored by non-OpenAI providers (issue #141, PR #213): the limit is normalized at the gateway seam so every provider honors it, and OpenAI / Azure o-series reasoning models receive max_completion_tokens instead of the rejected max_tokens.
• finish_reason not normalized to OpenAI values (issue #142): Anthropic / Bedrock / Cohere native stop reasons now map to stop | length | tool_calls | content_filter, so clients can detect truncation and tool use.
• Anthropic multimodal images dropped and tool-role messages malformed (issue #143): image content maps to Anthropic content blocks (vision no longer degrades to text-only) and tool results emit tool_result blocks inside a user turn instead of a bare {"role":"tool"} object the API rejects.
• Gemini system prompt prepended to the user message (issue #144): routed through the dedicated top-level systemInstruction field (Gemini 1.5+) instead of being smuggled into the first user turn, and no longer lost when a system message has no following user turn.
• Streaming and per-provider translation fidelity (issue #145): Anthropic streaming usage is parsed (non-zero token / cost metrics); DeepSeek reasoning_content and prompt-cache tokens are surfaced; Cohere embeddings input_type is configurable (defaults to search_document).

Changed

• Dependency sweep (11 bumps): aws-sdk-go-v2 (config / credentials / bedrockruntime), OpenTelemetry (otel + otlptracegrpc), go-chi/chi v5, modernc.org/sqlite, lib/pq, codecov-action, and upload-artifact, plus SA1019 deprecation fixes surfaced by the bumps.
• Internal shared-helper refactor: openaicompat.StreamSSE / PostChat / PostStream collapse ~23 OpenAI-compatible providers' duplicated request and stream plumbing; new core.NormalizeToolChoice / NormalizeFinishReason / Ptr helpers and a shared anthropicwire tool mapper; interface{} → any modernization. ~1,800 net lines removed with no behavior change beyond unifying SSE decode-error handling.

Contributors

Thanks to everyone who shipped this release:

• Provider-translation correctness + dependency sweep (#205) — @MitulShah1
• Tool/function calling across providers (#214) and max_completion_tokens fallback (#213) — @Rachit-Gandhi

---

Full changelog: https://github.com/ferro-labs/ai-gateway/blob/v1.1.4/CHANGELOG.md