Stream OpenAI and Gemini tokens into :messages mode

[twist900]

Why

LangEx.stream(..., modes: [:messages]) is supposed to yield {:message_delta, ...} for every built-in provider. The graph layer already injects :on_token via ChatModel, but only Anthropic read it. OpenAI and Gemini always did a blocking JSON POST, so GPT/Gemini graphs emitted node_start / node_end / :done and never a token.

This is a cherry-pick of 5bfac18 onto main. #43 landed the same feature, but it targeted feat/resumable-member-interrupts after that branch had already been squash-merged as #41. GitHub shows #43 as merged; origin/main still had blocking JSON POSTs for GPT/Gemini.

What

• LangEx.LLM.OpenAI and LangEx.LLM.Gemini honor :on_token (1-arity) or :stream: true and SSE-stream the body. The public return stays {:ok, %Message.AI{}, usage}.
• Without those opts the request is unchanged: one JSON completion, no stream key, so invoke/3 and existing batch tests keep working.
• Tool-call / function-call payloads are assembled into the final Message.AI and are not emitted as content deltas.
• Gemini thought parts (thought: true) are excluded from content.
• ChatModel, Graph.Stream, and Pregel are untouched.

How

OpenAI (POST {base_url}/chat/completions, :base_url still works for OpenRouter):

• Body: stream: true, stream_options: %{include_usage: true} so the last chunk carries usage.
• SSE: data: {json} lines, terminated by data: [DONE].
• Content: choices[0].delta.content → on_token. Empty / null first-chunk content is ignored.
• Tool calls: delta.tool_calls[] assembled by index (id, function.name, string arguments fragments).
• Usage: last chunk has choices: [] and prompt_tokens / completion_tokens. Intermediate usage: null is ignored.

Gemini (POST .../models/{model}:streamGenerateContent?alt=sse, same x-goog-api-key):

• Batch URL stays :generateContent. Streaming is the URL, not a stream field in the JSON body.
• Content: candidates[0].content.parts[].text → on_token.
• Function calls: parts[].functionCall with required name and args as a JSON object; same id / same name across chunks is merged. id is kept when present.
• Usage: last chunk usageMetadata.promptTokenCount / candidatesTokenCount.

Parsers live in LangEx.LLM.OpenAI.SSE and LangEx.LLM.Gemini.SSE. Transport copies Anthropic: Req :into callback + process-dictionary accumulator.

Test plan

Automated (already green locally: mix compile --warnings-as-errors, mix format --check-formatted, and mix test, 632 tests / 0 failures):

• OpenAI SSE parser: content deltas assemble "Hello"; empty/null content is not a token; tool-call fragments assemble one ToolCall and never fire on_token; [DONE] ignored; usage from the choices: [] chunk
• OpenAI adapter: :on_token / :stream: true send stream: true + stream_options.include_usage and set :into; without those opts the body has no stream key; :base_url still used
• Gemini SSE parser: text + functionCall object merge (same name / same id) + parallel calls in one chunk + usageMetadata; thought parts stay out of content
• Gemini adapter: streaming hits streamGenerateContent?alt=sse with x-goog-api-key; batch still hits generateContent
• ChatModel.node(model: "gpt-4o") under LangEx.stream(..., modes: [:messages]) yields two {:message_delta, ...} events before {:done, {:ok, _}}

Manual / live (needs API keys; not in CI):

• Graph with ChatModel.node(model: "gpt-4o") (or OpenRouter via :base_url) streamed with modes: [:messages] — confirm content chunks appear before :done, and invoke/3 on the same graph still returns a full Message.AI
• Same for model: "gemini-2.0-flash" — confirm streamGenerateContent traffic, thought text not in the AI message, usage non-zero
• A tool-calling turn on each provider — confirm argument fragments never show up as :messages deltas, and the final Message.AI.tool_calls is complete

Made with Cursor