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).
This commit is contained in:
2026-07-21 10:36:33 -04:00
parent 49e6414c29
commit 9953a26b8d
8 changed files with 491 additions and 14 deletions
@@ -1206,6 +1206,90 @@ mod tests {
);
}
/// Mistral-cycle e2e: embed pollution restricted away, and the Mistral
/// dialect (strips stream_options, handles reasoning arrays) is mapped.
#[tokio::test(flavor = "multi_thread")]
#[serial_test::serial]
async fn mistral_listing_restricts_and_maps_mistral_dialect() {
let platform_server = wiremock::MockServer::start().await;
wiremock::Mock::given(wiremock::matchers::method("GET"))
.and(wiremock::matchers::path("/models"))
.and(wiremock::matchers::header("Authorization", "Bearer msk-1"))
.respond_with(wiremock::ResponseTemplate::new(200).set_body_json(
serde_json::json!({ "data": [
{ "id": "devstral-latest", "object": "model" },
{ "id": "mistral-embed", "object": "model" }
]}),
))
.expect(1)
.mount(&platform_server)
.await;
let modelsdev_server = wiremock::MockServer::start().await;
wiremock::Mock::given(wiremock::matchers::method("GET"))
.and(wiremock::matchers::path("/api.json"))
.respond_with(wiremock::ResponseTemplate::new(200).set_body_json(
serde_json::json!({ "mistral": { "models": {
"devstral-latest": {
"limit": {"context": 262144, "output": 65536},
"tool_call": true
},
"mistral-embed": { "limit": {"context": 8000} }
}}}),
))
.expect(1)
.mount(&modelsdev_server)
.await;
let cache_dir = tempfile::tempdir().unwrap();
let _base = kigi_test_support::EnvGuard::set(
kigi_models::MISTRAL_BASE_URL_ENV,
platform_server.uri(),
);
let _mdev = kigi_test_support::EnvGuard::set(
crate::agent::enrichment_fetch::MODELS_DEV_URL_ENV,
format!("{}/api.json", modelsdev_server.uri()),
);
let _mdev_cache = kigi_test_support::EnvGuard::set(
crate::agent::enrichment_fetch::MODELS_DEV_CACHE_DIR_ENV,
cache_dir.path(),
);
let endpoints = crate::agent::config::EndpointsConfig::default();
let keys = crate::agent::models::PlatformApiKeys::test_single(
kigi_models::PlatformId::Mistral,
"msk-1",
);
let result = tokio::task::spawn_blocking(move || {
fetch_platform_models_blocking(&endpoints, None, &keys)
})
.await
.unwrap()
.expect("fetch must succeed");
assert_eq!(
result
.models
.iter()
.map(|m| m.id.as_deref().unwrap_or_default())
.collect::<Vec<_>>(),
vec!["mistral/devstral-latest"],
"embed (enrichment-known, not tool-calling) must be dropped"
);
let entry = &result.models[0];
assert_eq!(entry.context_window.get(), 262_144);
assert_eq!(entry.max_completion_tokens, Some(65_536));
let model_entry = crate::agent::config::ModelEntry::from_config_entry(entry);
let creds = crate::agent::config::ResolvedCredentials {
api_key: Some("msk-1".into()),
base_url: entry.base_url.clone(),
auth_type: kigi_chat_state::AuthType::ApiKey,
auth_scheme: Default::default(),
};
let cfg = crate::agent::config::sampling_config_for_model(&model_entry, creds, None);
assert_eq!(
cfg.chat_compat,
kigi_sampling_types::ChatCompat::Mistral,
"mistral entries must map to the Mistral dialect (stream_options strip)"
);
}
#[test]
fn get_env_keys_parses_strings_and_rejects_non_strings() {
use crate::agent::config::EnvKeys;