Add Mistral platform + Mistral dialect + array-content handling (provider 5)

The 8th registry row: id "mistral", MISTRAL_API_KEY > auth.json "mistral" scope, https://api.mistral.ai/v1 with KIGI_MISTRAL_BASE_URL override, enrichment-backed metadata with the tool-calling listing restriction (embed/moderation/OCR noise). Mistral is NOT a pure-pattern provider — an adversarial review found two doc-confirmed blockers that no test exercises (no e2e covers a chat POST), so the gate-green registry row alone would have shipped it DOA. A research workflow pinned the exact wire shapes against the mistralai/client-python SDK source (adversarially verified), then both were fixed: 1. stream_options 422: Mistral's strict Pydantic validator rejects the stream_options.include_usage field kigi injects on every streaming request (the SDK's request model has no such field). New ChatCompat::Mistral dialect strips it (plus the kigi-private message fields, like Passthrough). Streaming usage falls back to token estimation. 2. Reasoning content arrays: Mistral reasoning models return content as Union[str, List[ContentChunk]] on both streaming and non-streaming, which the flat Option<String> path could not decode -> aborted turn. A UNIVERSAL lenient deserializer (#[serde(from = "Raw..")] on ChatResponseMessage + ChatChunkDelta) accepts string-or-array, routing {type:text} chunks to the answer and the nested text of {type:thinking} chunks to reasoning_content, tolerant of the OPEN chunk union (unknown types ignored, never fatal). String content stays byte-identical for every other provider (kimi/deepseek/groq/BYOK). Review refuted all seven attack lines (no regression, no crash, exhaustive) and flagged one coverage gap, now closed: a stream-consumer integration test drives a full thinking -> transition -> answer chunk sequence and proves it yields the same reasoning-sibling + assistant-answer result as the reasoning_content string path. Also folds a verified quirk matrix for all 23 remaining API providers into providers-plan.md, tiered by real difficulty (self-enriching OpenRouter/ Vercel; bare-array Together listing; Messages-dialect MiniMax reusing the Anthropic machinery; non-Bearer Azure/Bedrock; router wildcards; the OAuth block).
2026-07-21 10:36:33 -04:00
parent 49e6414c29
commit 9953a26b8d
8 changed files with 491 additions and 14 deletions
@@ -47,6 +47,22 @@ pub(crate) fn adapt_chat_completions_body_for(
        kigi_sampling_types::ChatCompat::Passthrough => {
            strip_kigi_private_message_fields(body);
        }
+        kigi_sampling_types::ChatCompat::Mistral => {
+            strip_kigi_private_message_fields(body);
+            strip_stream_options(body);
+        }
+    }
+}
+
+/// Mistral's strict Pydantic validator 422-rejects `stream_options`
+/// (`extra_forbidden` on `stream_options.include_usage`; its request model
+/// has no such field). kigi injects `stream_options.include_usage` on every
+/// streaming request for the other providers, so strip the whole object for
+/// Mistral. Streaming usage falls back to token estimation (as for any
+/// provider that omits streaming usage).
+fn strip_stream_options(body: &mut Value) {
+    if let Some(obj) = body.as_object_mut() {
+        obj.remove("stream_options");
    }
 }

@@ -393,6 +409,38 @@ mod tests {
        assert_eq!(body["messages"][0].get("model_id"), None);
    }

+    #[test]
+    fn mistral_dialect_strips_stream_options_and_private_fields() {
+        use kigi_sampling_types::ChatCompat;
+        // Mistral 422s on stream_options (extra_forbidden) and doesn't know
+        // kigi's private message fields; OpenAI-style reasoning_effort stays.
+        let mut body = json!({
+            "model": "mistral-medium-latest",
+            "reasoning_effort": "high",
+            "stream": true,
+            "stream_options": { "include_usage": true },
+            "messages": [
+                { "role": "assistant", "content": "hi",
+                  "reasoning_content": "internal", "model_id": "kigi/x" }
+            ]
+        });
+        adapt_chat_completions_body_for(ChatCompat::Mistral, &mut body);
+        assert_eq!(
+            body.get("stream_options"),
+            None,
+            "stream_options must be stripped"
+        );
+        assert_eq!(body["stream"], json!(true), "stream flag stays");
+        assert_eq!(
+            body["reasoning_effort"],
+            json!("high"),
+            "OpenAI-style effort passes through (Mistral accepts it natively)"
+        );
+        assert_eq!(body["messages"][0].get("reasoning_content"), None);
+        assert_eq!(body["messages"][0].get("model_id"), None);
+        assert_eq!(body["messages"][0]["content"], json!("hi"));
+    }
+
    #[test]
    fn passthrough_dialect_leaves_openai_body_verbatim() {
        use kigi_sampling_types::ChatCompat;
@@ -517,6 +517,85 @@ mod tests {
        }
    }

+    /// Deserialize a full chunk from a JSON `delta` so it flows through the
+    /// `#[serde(from = "RawChatChunkDelta")]` content-split path (Mistral
+    /// sends array `content`).
+    fn chunk_from_delta(delta: serde_json::Value) -> ChatCompletionChunk {
+        serde_json::from_value(serde_json::json!({
+            "id": "c",
+            "object": "chat.completion.chunk",
+            "created": 0,
+            "model": "mistral-medium-latest",
+            "choices": [{ "index": 0, "delta": delta, "finish_reason": null }],
+        }))
+        .expect("mistral chunk deserializes")
+    }
+
+    /// End-to-end: a Mistral reasoning stream (array `content` thinking
+    /// deltas → a transition delta carrying both a thinking and a text chunk
+    /// → plain-string answer deltas) is split by deserialization and consumed
+    /// into the SAME reasoning-sibling + assistant-answer result the
+    /// `reasoning_content` string path produces. Closes the array-path
+    /// integration gap.
+    #[tokio::test]
+    async fn mistral_reasoning_array_stream_splits_into_reasoning_and_answer() {
+        let chunks: Vec<Result<ChatCompletionChunk, SamplingError>> = vec![
+            Ok(chunk_from_delta(serde_json::json!({ "content": [
+                { "type": "thinking",
+                  "thinking": [{ "type": "text", "text": "Let me think. " }] }
+            ]}))),
+            Ok(chunk_from_delta(serde_json::json!({ "content": [
+                { "type": "thinking",
+                  "thinking": [{ "type": "text", "text": "It's 22." }] }
+            ]}))),
+            // Transition: one array with a closing thinking chunk AND the
+            // first answer text chunk.
+            Ok(chunk_from_delta(serde_json::json!({ "content": [
+                { "type": "thinking", "thinking": [{ "type": "text", "text": " Done." }] },
+                { "type": "text", "text": "Answer: " }
+            ]}))),
+            // Answer phase: plain-string deltas (no longer arrays).
+            Ok(chunk_from_delta(serde_json::json!({ "content": "22." }))),
+            Ok(final_chunk(FinishReason::Stop)),
+        ];
+        let raw = stream::iter(chunks).boxed();
+        let events = collect(stream_chat_completions(
+            raw,
+            None,
+            rid(),
+            Duration::from_secs(60),
+        ))
+        .await;
+
+        // Channel tokens: thinking rode the Reasoning channel, answer the Text
+        // channel — never crossed.
+        let mut reasoning = String::new();
+        let mut answer = String::new();
+        for e in &events {
+            if let SamplingEvent::ChannelToken { channel, text, .. } = e {
+                match channel {
+                    SamplingChannel::Reasoning => reasoning.push_str(text),
+                    SamplingChannel::Text => answer.push_str(text),
+                }
+            }
+        }
+        assert_eq!(reasoning, "Let me think. It's 22. Done.");
+        assert_eq!(answer, "Answer: 22.");
+
+        // The accumulated final response carries the same split.
+        match events.last().unwrap() {
+            SamplingEvent::Completed { response, .. } => {
+                let r = response
+                    .reasoning_items()
+                    .next()
+                    .expect("array thinking became a reasoning sibling");
+                let rs::SummaryPart::SummaryText(t) = &r.summary[0];
+                assert_eq!(t.text, "Let me think. It's 22. Done.");
+            }
+            other => panic!("expected Completed, got {other:?}"),
+        }
+    }
+
    #[tokio::test]
    async fn tool_call_stream_emits_deltas_and_assembles_final_call() {
        // First chunk has id + name + part of arguments.