Hare/yoi
1
1
Fork 0
Code Actions Packages Releases Activity

refactor: fold provider crate into engine

Browse Source
This commit is contained in:
Hare
2026-07-15 00:58:29 +09:00
parent 1505c5e8a3
commit 839fc7b40c
33 changed files with 263 additions and 279 deletions
Download Patch File Download Diff File Expand all files Collapse all files
crates/llm-engine/Cargo.toml
+7 -1
View File
@@ -5,6 +5,10 @@ version = "0.2.1"
edition.workspace = true
license.workspace = true
[features]
default = []
codex = ["dep:base64", "dep:chrono"]
[dependencies]
serde = { workspace = true, features = ["derive"] }
serde_json = { workspace = true }
@@ -12,11 +16,13 @@ thiserror = { workspace = true }
tracing = { workspace = true }
async-trait = { workspace = true }
futures = { workspace = true }
tokio = { workspace = true, features = ["macros", "rt-multi-thread", "time"] }
tokio = { workspace = true, features = ["fs", "macros", "rt-multi-thread", "sync", "time"] }
tokio-util = "0.7"
reqwest = { version = "0.13", default-features = false, features = ["stream", "json", "native-tls", "http2"] }
eventsource-stream = "0.2"
zstd = "0.13"
base64 = { version = "0.22.1", optional = true }
chrono = { version = "0.4", default-features = false, features = ["serde", "clock"], optional = true }
llm-engine-macros = { workspace = true }
[dev-dependencies]
crates/llm-engine/src/lib.rs
+1
View File
@@ -45,6 +45,7 @@ pub(crate) mod callback;
pub mod event;
pub mod interceptor;
pub mod llm_client;
pub mod providers;
pub mod prune;
pub mod state;
pub mod timeline;
crates/llm-engine/src/llm_client/auth.rs
-8
View File
@@ -40,12 +40,4 @@ pub enum AuthRequirement {
pub trait AuthProvider: Send + Sync + std::fmt::Debug {
/// 1 リクエスト分の認証ヘッダを返す。refresh が必要なら内部で行う。
async fn headers(&self) -> Result<Vec<(HeaderName, HeaderValue)>, ClientError>;
/// Conversation header / request compression が必要な backend profile かどうか。
///
/// transport は呼び出し側の具象型を知らないため、この hook だけで
/// 追加の wire behavior を切り替える。
fn is_codex_backend(&self) -> bool {
false
}
}
crates/llm-engine/src/llm_client/retry.rs
+4 -4
View File
@@ -1,15 +1,15 @@
//! LLM response stream を開く前の transient error 向けリトライポリシー。
//!
//! Engine が `LlmClient::stream` の open error に対して `is_retryable` を見て
//! retry / backoff / TUI event / cancellation をまとめて管理する。
//! `LlmClient::stream` の open error に対して `is_retryable` を見て
//! retry / backoff / cancellation をまとめて管理する。
//! SSE 読み出し開始後の失敗は対象外。
use std::time::Duration;
/// 指数バックオフ + ジッター + 累積タイムアウトを表すポリシー。
///
/// `Default` は llm-engine 全体の固定値を返す。manifest 経由の上書きが
/// 必要になったら拡張する（現状は不要 → `tickets/llm-engine-transient-retry.md`）。
/// `Default` は llm-engine 全体の固定値を返す。呼び出し側からの上書きが
/// 必要になったら拡張する。
#[derive(Debug, Clone)]
pub struct RetryPolicy {
/// 指数の基準値。`base * 2^attempt` を `cap` で頭打ちにした上限から
crates/llm-engine/src/llm_client/scheme/anthropic/capability.rs
+2 -2
View File
@@ -1,7 +1,7 @@
//! Anthropic scheme の wire-level 既定 capability。
//!
//! モデル ID 固有のテーブル(`claude-*` など)は高レベル構築層
//! (`provider::capability`)の責務。ここでは未知モデルでも「この wire で
//! モデル ID 固有のテーブル(`claude-*` など)は client construction layer
//! の責務。ここでは未知モデルでも「この wire で
//! 安全に送れる最小共通項」を返すだけに留める。
use crate::llm_client::capability::{
crates/llm-engine/src/llm_client/scheme/gemini/capability.rs
+1 -1
View File
@@ -1,7 +1,7 @@
//! Gemini scheme の wire-level 既定 capability。
//!
//! モデル ID 固有のテーブル(`gemini-*` バージョン別の reasoning 有無)は
//! 高レベル構築層(`provider::capability`)の責務。ここでは wire の
//! client construction layer の責務。ここでは wire の
//! 保守的 default のみ。
use crate::llm_client::capability::{
crates/llm-engine/src/llm_client/scheme/openai_chat/capability.rs
+2 -2
View File
@@ -1,7 +1,7 @@
//! OpenAI Chat Completions scheme の wire-level 既定 capability。
//!
//! モデル ID 固有のテーブル(`gpt-5` 系など)は高レベル構築層
//! (`provider::capability`)の責務。ここでは wire の保守的 default のみ。
//! モデル ID 固有のテーブル(`gpt-5` 系など)は client construction layer
//! の責務。ここでは wire の保守的 default のみ。
use crate::llm_client::capability::{
CacheStrategy, ModelCapability, StructuredOutput, ToolCallingSupport,
crates/llm-engine/src/llm_client/scheme/openai_responses/scheme_impl.rs
+3 -3
View File
@@ -65,7 +65,7 @@ impl Scheme for OpenAIResponsesScheme {
if !self.send_max_output_tokens && config.max_tokens.is_some() {
warnings.push(ConfigWarning::unsupported(
"max_tokens",
"OpenAI Responses (ChatGPT backend)",
"OpenAI Responses compatible backend",
));
}
// Same for `temperature` / `top_p` on compatible backends that
@@ -74,13 +74,13 @@ impl Scheme for OpenAIResponsesScheme {
if config.temperature.is_some() {
warnings.push(ConfigWarning::unsupported(
"temperature",
"OpenAI Responses (ChatGPT backend)",
"OpenAI Responses compatible backend",
));
}
if config.top_p.is_some() {
warnings.push(ConfigWarning::unsupported(
"top_p",
"OpenAI Responses (ChatGPT backend)",
"OpenAI Responses compatible backend",
));
}
}
crates/llm-engine/src/llm_client/transport.rs
+115 -43
View File
@@ -1,8 +1,8 @@
//! `HttpTransport<S: Scheme>`: すべての LLM wire scheme を共通の 1 本の
//! HTTP クライアントで扱う。
//!
//! 旧 `providers/{anthropic,openai,gemini,ollama}.rs` を置き換える。
//! scheme 固有の差分は [`Scheme`] trait 実装に委譲する。
//! scheme 固有の差分は [`Scheme`] trait 実装に委譲し、backend 固有の
//! HTTP policy は [`TransportPolicy`] で明示的に差し込む。
use std::pin::Pin;
use std::sync::Arc;
@@ -63,6 +63,71 @@ impl ResolvedAuth {
}
}
/// Request body encoding policy used by [`HttpTransport`].
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum RequestBodyEncoding {
/// Send the request body as plain JSON.
Json,
/// Send the request body as zstd-compressed JSON with `Content-Encoding: zstd`.
ZstdJson,
}
impl Default for RequestBodyEncoding {
fn default() -> Self {
Self::Json
}
}
/// Conversation header policy used by [`HttpTransport`].
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum ConversationHeaderPolicy {
/// Do not derive any transport headers from [`Request::cache_key`].
None,
/// Send OpenAI-compatible conversation headers from [`Request::cache_key`].
OpenAiCompatible {
/// Send the legacy `session_id` header in addition to `session-id`.
include_legacy_session_id: bool,
/// Send `thread-id` with the same value as `session-id`.
include_thread_id: bool,
/// Send `x-client-request-id` with the same value as `session-id`.
include_client_request_id: bool,
},
}
impl Default for ConversationHeaderPolicy {
fn default() -> Self {
Self::None
}
}
/// Backend-specific HTTP transport policy.
#[derive(Debug, Clone, PartialEq, Eq, Default)]
pub struct TransportPolicy {
/// How to serialize and encode the HTTP request body.
pub request_body_encoding: RequestBodyEncoding,
/// Optional backend-specific conversation headers derived from request metadata.
pub conversation_headers: ConversationHeaderPolicy,
}
impl TransportPolicy {
/// Plain JSON requests with no derived conversation headers.
pub fn standard() -> Self {
Self::default()
}
/// OpenAI-compatible backend profile that uses zstd JSON bodies and
/// conversation headers derived from [`Request::cache_key`].
pub fn openai_compatible_zstd() -> Self {
Self {
request_body_encoding: RequestBodyEncoding::ZstdJson,
conversation_headers: ConversationHeaderPolicy::OpenAiCompatible {
include_legacy_session_id: true,
include_thread_id: true,
include_client_request_id: true,
},
}
}
}
fn header_value_for_diagnostics(headers: &HeaderMap, name: &str) -> Option<String> {
headers
.get(name)
@@ -120,6 +185,7 @@ pub struct HttpTransport<S: Scheme> {
base_url: String,
auth: ResolvedAuth,
capability: ModelCapability,
policy: TransportPolicy,
}
impl<S: Scheme> HttpTransport<S> {
@@ -141,9 +207,16 @@ impl<S: Scheme> HttpTransport<S> {
base_url,
auth,
capability,
policy: TransportPolicy::default(),
}
}
/// Set a backend-specific HTTP transport policy.
pub fn with_transport_policy(mut self, policy: TransportPolicy) -> Self {
self.policy = policy;
self
}
/// カスタム HTTP クライアントを差し込む（テスト等）。
pub fn with_http_client(mut self, client: reqwest::Client) -> Self {
self.http_client = client;
@@ -206,13 +279,6 @@ impl<S: Scheme> HttpTransport<S> {
Ok(headers)
}
fn is_codex_backend(&self) -> bool {
match &self.auth {
ResolvedAuth::Custom(provider) => provider.is_codex_backend(),
_ => false,
}
}
fn apply_stream_headers(
&self,
headers: &mut HeaderMap,
@@ -220,18 +286,26 @@ impl<S: Scheme> HttpTransport<S> {
) -> Result<(), ClientError> {
headers.insert(ACCEPT, HeaderValue::from_static("text/event-stream"));
if self.is_codex_backend()
if let ConversationHeaderPolicy::OpenAiCompatible {
include_legacy_session_id,
include_thread_id,
include_client_request_id,
} = self.policy.conversation_headers
&& let Some(cache_key) = request.cache_key.as_deref()
{
let value = HeaderValue::from_str(cache_key).map_err(|e| {
ClientError::Config(format!("invalid Codex conversation header: {e}"))
ClientError::Config(format!("invalid conversation header value: {e}"))
})?;
// Send both current hyphenated conversation headers and the
// legacy underscore form for compatibility with existing backends.
headers.insert(HeaderName::from_static("session-id"), value.clone());
headers.insert(HeaderName::from_static("thread-id"), value.clone());
headers.insert(HeaderName::from_static("session_id"), value.clone());
headers.insert(HeaderName::from_static("x-client-request-id"), value);
if include_thread_id {
headers.insert(HeaderName::from_static("thread-id"), value.clone());
}
if include_legacy_session_id {
headers.insert(HeaderName::from_static("session_id"), value.clone());
}
if include_client_request_id {
headers.insert(HeaderName::from_static("x-client-request-id"), value);
}
}
Ok(())
@@ -242,19 +316,22 @@ impl<S: Scheme> HttpTransport<S> {
body: &serde_json::Value,
headers: &mut HeaderMap,
) -> Result<RequestBody, ClientError> {
if !self.is_codex_backend() {
return Ok(RequestBody::Json(body.clone()));
match self.policy.request_body_encoding {
RequestBodyEncoding::Json => Ok(RequestBody::Json(body.clone())),
RequestBodyEncoding::ZstdJson => {
let raw = serde_json::to_vec(body)?;
let raw_json_bytes = raw.len();
let compressed =
zstd::stream::encode_all(std::io::Cursor::new(raw), 3).map_err(|e| {
ClientError::Config(format!("failed to zstd-compress request: {e}"))
})?;
headers.insert(CONTENT_ENCODING, HeaderValue::from_static("zstd"));
Ok(RequestBody::CompressedJson {
bytes: compressed,
raw_json_bytes,
})
}
}
let raw = serde_json::to_vec(body)?;
let raw_json_bytes = raw.len();
let compressed = zstd::stream::encode_all(std::io::Cursor::new(raw), 3)
.map_err(|e| ClientError::Config(format!("failed to zstd-compress request: {e}")))?;
headers.insert(CONTENT_ENCODING, HeaderValue::from_static("zstd"));
Ok(RequestBody::CompressedJson {
bytes: compressed,
raw_json_bytes,
})
}
}
@@ -383,6 +460,7 @@ impl<S: Scheme + Clone> Clone for HttpTransport<S> {
base_url: self.base_url.clone(),
auth: self.auth.clone(),
capability: self.capability.clone(),
policy: self.policy.clone(),
}
}
}
@@ -446,7 +524,8 @@ impl<S: Scheme + Clone + 'static> LlmClient for HttpTransport<S> {
"path": path,
"auth_kind": auth_kind(&self.auth),
"required_auth": format!("{:?}", self.scheme.required_auth()),
"codex_backend": self.is_codex_backend(),
"request_body_encoding": format!("{:?}", self.policy.request_body_encoding),
"conversation_headers": format!("{:?}", self.policy.conversation_headers),
"cache_key_present": request.cache_key.is_some(),
"stream_open_timeout_ms": DEFAULT_STREAM_OPEN_TIMEOUT.as_millis() as u64,
}),
@@ -677,9 +756,7 @@ mod tests {
use serde_json::json;
#[derive(Debug)]
struct TestAuthProvider {
codex: bool,
}
struct TestAuthProvider;
#[async_trait]
impl AuthProvider for TestAuthProvider {
@@ -695,10 +772,6 @@ mod tests {
),
])
}
fn is_codex_backend(&self) -> bool {
self.codex
}
}
#[derive(Clone)]
@@ -827,10 +900,9 @@ mod tests {
}
#[tokio::test]
async fn codex_backend_adds_conversation_headers_and_zstd_body() {
let transport = transport(ResolvedAuth::Custom(Arc::new(TestAuthProvider {
codex: true,
})));
async fn transport_policy_adds_conversation_headers_and_zstd_body() {
let transport = transport(ResolvedAuth::Custom(Arc::new(TestAuthProvider)))
.with_transport_policy(TransportPolicy::openai_compatible_zstd());
let request = Request::new().user("hello").cache_key("segment-123");
let mut headers = transport.build_headers().await.unwrap();
transport
@@ -869,7 +941,7 @@ mod tests {
raw_json_bytes,
} = encoded
else {
panic!("Codex backend request body must be zstd-compressed");
panic!("transport policy should zstd-compress request body");
};
assert!(raw_json_bytes > 0);
let decoded = zstd::stream::decode_all(std::io::Cursor::new(compressed)).unwrap();
@@ -878,7 +950,7 @@ mod tests {
}
#[tokio::test]
async fn non_codex_request_does_not_get_codex_only_headers_or_compression() {
async fn standard_policy_does_not_get_conversation_headers_or_compression() {
let transport = transport(ResolvedAuth::ApiKey("api-key".to_string()));
let request = Request::new().user("hello").cache_key("segment-123");
let mut headers = transport.build_headers().await.unwrap();
@@ -900,7 +972,7 @@ mod tests {
assert!(headers.get(CONTENT_ENCODING).is_none());
let RequestBody::Json(decoded) = encoded else {
panic!("non-Codex request body must remain normal JSON");
panic!("standard transport policy should keep request body as JSON");
};
assert_eq!(decoded["prompt_cache_key"], "segment-123");
}
crates/provider/src/codex_oauth/auth_json.rs → crates/llm-engine/src/providers/codex/auth_json.rs
+3 -3
View File
@@ -1,7 +1,7 @@
//! `~/.codex/auth.json` の読み書き。
//!
//! Codex CLI と schema を共有するが、yoi は知らないフィールドを
//! 失わないようファイル全体を `serde_json::Value` で保持し、必要箇所
//! Codex CLI と schema を共有するが、知らないフィールドを失わないよう
//! ファイル全体を `serde_json::Value` で保持し、必要箇所
//! のみアクセスする。書込は `mode 0o600` を再設定（Codex CLI 同様）、
//! ファイルロックは取らない（manager 側で guarded reload）。
@@ -106,7 +106,7 @@ pub async fn load(path: &Path) -> Result<AuthSnapshot, CodexAuthError> {
/// 既存ファイルを再読込し、`tokens.{id_token,access_token,refresh_token}` と
/// `last_refresh` を更新して書き戻す。Codex CLI の `persist_tokens` 相当。
///
/// 並行する Codex CLI / 別 yoi プロセスが先に refresh していた場合の
/// 並行する Codex CLI / 別プロセスが先に refresh していた場合の
/// fields を保護するため、書込前に再 load して merge する。
pub async fn persist_refreshed(
path: &Path,
crates/provider/src/codex_oauth/error.rs → crates/llm-engine/src/providers/codex/error.rs
+1 -1
View File
@@ -6,7 +6,7 @@
use std::path::PathBuf;
use llm_engine::llm_client::ClientError;
use crate::llm_client::ClientError;
use thiserror::Error;
#[derive(Debug, Error)]
crates/provider/src/codex_oauth/jwt.rs → crates/llm-engine/src/providers/codex/jwt.rs
View File
crates/provider/src/codex_oauth/mod.rs → crates/llm-engine/src/providers/codex/mod.rs
+5 -9
View File
@@ -4,11 +4,11 @@
//!
//! 設計:
//!
//! - llm-engine は [`AuthProvider`] trait しか知らず、実体である
//! [`CodexAuthProvider`] はこのクレートに置く（feedback_llm_engine_scope）
//! - HTTP transport は [`AuthProvider`] trait だけを見て、実体である
//! [`CodexAuthProvider`] はこの optional module に置く
//! - access_token JWT の `exp` を読み、`now` 以下で proactive refresh
//! （Codex CLI と同じバッファなし）
//! - 並行する Codex CLI / 別 yoi の refresh と取り違えないよう、
//! - 並行する Codex CLI / 別プロセスの refresh と取り違えないよう、
//! refresh 直前に再 load して account_id 一致を確認（guarded reload）
//! - ファイルロックは取らず、書込前に再 load + diff merge で吸収
//! - Codex の Keyring storage は対象外。auth.json 不在ならエラーで案内
@@ -21,9 +21,9 @@ mod refresh;
use std::path::PathBuf;
use std::sync::Arc;
use crate::llm_client::{ClientError, auth::AuthProvider};
use async_trait::async_trait;
use chrono::{Duration, Utc};
use llm_engine::llm_client::{ClientError, auth::AuthProvider};
use reqwest::header::{HeaderName, HeaderValue};
use tokio::sync::Mutex;
@@ -152,7 +152,7 @@ impl CodexAuthProvider {
.map_err(|e| CodexAuthError::InvalidHeader(format!("ChatGPT-Account-Id: {e}")))?;
out.push((HeaderName::from_static("chatgpt-account-id"), acc_val));
// Cloudflare WAF は ChatGPT backend アクセス元を `originator` /
// Cloudflare WAF は互換 backend アクセス元を `originator` /
// `User-Agent` で識別する。Codex CLI が送る固定値を流用しないと
// HTML challenge (403) を返されて SSE に到達できない。
out.push((
@@ -188,10 +188,6 @@ impl AuthProvider for CodexAuthProvider {
.map_err(CodexAuthError::to_client_error)?;
Self::build_headers(&snap).map_err(CodexAuthError::to_client_error)
}
fn is_codex_backend(&self) -> bool {
true
}
}
/// `access_token` の JWT `exp` を見て、期限切れなら true。
crates/provider/src/codex_oauth/refresh.rs → crates/llm-engine/src/providers/codex/refresh.rs
+1 -1
View File
@@ -80,7 +80,7 @@ async fn response_with_timeout(
.await
.map_err(|_| {
CodexAuthError::RefreshTransient(format!(
"codex_oauth_refresh timed out after {}s",
"oauth_token_refresh timed out after {}s",
timeout.as_secs()
))
})?
crates/llm-engine/src/providers/mod.rs
+4
View File
@@ -0,0 +1,4 @@
//! Optional built-in provider/backend helpers.
#[cfg(feature = "codex")]
pub mod codex;
crates/llm-engine/src/timeline/tool_call_collector.rs
+2 -2
View File
@@ -130,13 +130,13 @@ mod tests {
let mut timeline = Timeline::new();
timeline.on_tool_use_block(collector.clone());
timeline.dispatch(&Event::tool_use_start(0, "tool_empty", "ListWorkers"));
timeline.dispatch(&Event::tool_use_start(0, "tool_empty", "ListItems"));
timeline.dispatch(&Event::tool_use_stop(0));
let calls = collector.take_collected();
assert_eq!(calls.len(), 1);
assert_eq!(calls[0].id, "tool_empty");
assert_eq!(calls[0].name, "ListWorkers");
assert_eq!(calls[0].name, "ListItems");
assert!(calls[0].input.is_object());
assert_eq!(
calls[0].input,
crates/manifest/Cargo.toml
+1
View File
@@ -18,4 +18,5 @@ toml = { workspace = true }
tracing = { workspace = true }
[dev-dependencies]
serial_test = "3.4.0"
tempfile = { workspace = true }
crates/manifest/src/config.rs
+1 -1
View File
@@ -723,7 +723,7 @@ fn resolve_auth_file(auth: &mut Option<AuthRef>, base: &Path) {
/// モデル宣言に含まれる `auth.file` が絶対パスであることを検証する。
/// ref / scheme / model_id 等の論理的な有効性（ref があるか、inline が
/// 揃っているか）の検証はカタログを知る `crates/provider` 側で行う。
/// 揃っているか）の検証は [`crate::model_catalog`] で行う。
fn validate_model_paths(model: &ModelManifest, field: &'static str) -> Result<(), ResolveError> {
if let Some(AuthRef::ApiKey { file: Some(p), .. }) = &model.auth {
ensure_absolute(field, p)?;
crates/manifest/src/lib.rs
+1
View File
@@ -1,6 +1,7 @@
mod config;
pub mod defaults;
mod model;
pub mod model_catalog;
pub mod paths;
pub mod plugin;
mod profile;
crates/manifest/src/model.rs
+4 -4
View File
@@ -5,7 +5,7 @@
//! 直書き）の両方を受け入れるため、すべてのフィールドを `Option` として
//! 持つ 1 つの型 [`ModelManifest`] に統合している。実解決（ref をプロバイダ
//! カタログ / モデルカタログから引いて `scheme` や `model_id` を埋める）
//! は `crates/provider` の責務で、本モジュールはデータ表現のみを提供する。
//! は [`crate::model_catalog`] の責務で、本モジュールはデータ表現のみを提供する。
//!
//! 同じ型を partial（カスケード層）と完成形（最終マニフェスト）の両方で
//! 使うことで、merge と最終変換の重複を避ける。
@@ -24,7 +24,7 @@ pub use llm_engine::llm_client::capability::{ModelCapability, ReasoningControl,
/// - ref + 一部 override: ref で基底を引き、`auth` 等だけ書き換え
/// - 完全 inline: `ref` を省略して `scheme` / `model_id` / `auth` を直書き
///
/// どの形が有効かの判定は `provider::resolve_model_manifest` が担う。
/// どの形が有効かの判定は [`crate::model_catalog::resolve_model_manifest`] が担う。
/// 本クレートは「どこから取るか」を表現するだけで、未設定かどうかを
/// 理由にした hard error は出さない。
#[derive(Debug, Clone, Default, Serialize, Deserialize, PartialEq)]
@@ -98,7 +98,7 @@ pub enum SchemeKind {
/// 認証の参照。
///
/// 実際のトークン値の解決（local secret store / file 読取、OAuth refresh 等）は
/// `crates/provider` で行う。ここはあくまで「どこから取るか」の宣言。
/// worker-side factory で行う。ここはあくまで「どこから取るか」の宣言。
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq, Default)]
#[serde(tag = "kind", rename_all = "snake_case")]
pub enum AuthRef {
@@ -113,7 +113,7 @@ pub enum AuthRef {
#[serde(default)]
file: Option<PathBuf>,
},
/// ChatGPT OAuth（`~/.codex/auth.json`）。実装は `llm-auth-codex-oauth` チケット
/// ChatGPT OAuth（`~/.codex/auth.json`）。
#[serde(rename = "codex_oauth")]
CodexOAuth,
/// Typed local secret-store reference. The profile resolver preserves this
crates/provider/src/catalog.rs → crates/manifest/src/model_catalog.rs
+5 -5
View File
@@ -4,19 +4,19 @@
//! - builtin モデル: `resources/models/builtin.toml`
//! - user override: `<config_dir>/{providers,models}.toml`
//!
//! `<config_dir>` の解決は [`manifest::paths::config_dir`] を参照。
//! `<config_dir>` の解決は [`crate::paths::config_dir`] を参照。
//! どちらの override も「あれば builtin を置換、無ければ builtin」と
//! いう一方向の差し替え（マージしない）。providers / models は独立に
//! 読み、片方だけ user override も可。
//!
//! [`resolve_model_manifest`] が `manifest::ModelManifest`（ref / inline
//! [`resolve_model_manifest`] が [`crate::ModelManifest`]（ref / inline
//! 両形）を最終的な [`ModelConfig`] に解決する単一の入口で、wire 層
//! に渡す前のバリデーションもここで行う。
use std::path::{Path, PathBuf};
use crate::{AuthRef, ModelManifest, SchemeKind};
use llm_engine::llm_client::capability::ModelCapability;
use manifest::{AuthRef, ModelManifest, SchemeKind};
use serde::{Deserialize, Serialize};
const BUILTIN_PROVIDERS: &str = include_str!("../../../resources/providers/builtin.toml");
@@ -178,7 +178,7 @@ fn auth_hint_to_ref(hint: &AuthHint) -> AuthRef {
/// ユーザーが書いた設定が silent に無視されて builtin に戻る挙動は
/// 気付きにくいため）。
pub fn load_providers() -> Result<Vec<ProviderEntry>, CatalogError> {
if let Some(path) = manifest::paths::user_catalog_override("providers.toml")
if let Some(path) = crate::paths::user_catalog_override("providers.toml")
&& path.is_file()
{
return load_providers_from(&path);
@@ -211,7 +211,7 @@ pub fn load_providers_from(path: &Path) -> Result<Vec<ProviderEntry>, CatalogErr
/// builtin + user override を解決してモデルカタログを返す。
pub fn load_models() -> Result<Vec<ModelEntry>, CatalogError> {
if let Some(path) = manifest::paths::user_catalog_override("models.toml")
if let Some(path) = crate::paths::user_catalog_override("models.toml")
&& path.is_file()
{
return load_models_from(&path);
crates/provider/Cargo.toml
-25
View File
@@ -1,25 +0,0 @@
[package]
name = "provider"
version = "0.1.0"
edition.workspace = true
license.workspace = true
[dependencies]
async-trait = { workspace = true }
base64 = "0.22.1"
chrono = { version = "0.4", default-features = false, features = ["serde", "clock"] }
llm-engine = { workspace = true }
manifest = { workspace = true }
secrets = { workspace = true }
reqwest = { version = "0.13", features = ["json", "native-tls"] }
serde = { workspace = true, features = ["derive"] }
serde_json = { workspace = true }
thiserror = { workspace = true }
tokio = { workspace = true, features = ["sync", "fs", "rt", "time"] }
toml = { workspace = true }
tracing = { workspace = true }
[dev-dependencies]
serial_test = "3.4.0"
tempfile = { workspace = true }
wiremock = "0.6.5"
crates/provider/README.md
-33
View File
@@ -1,33 +0,0 @@
# provider
## Role
`provider` resolves model/provider configuration and constructs provider-specific LLM clients for `llm-engine`.
## Boundaries
Owns:
- builtin and user provider/model catalog resolution
- model reference expansion into concrete model config
- auth reference resolution through supported mechanisms
- provider/scheme capability and context-window metadata
- provider-specific client construction
Does not own:
- Engine turn lifecycle (`llm-engine`)
- secret storage internals (`secrets`)
- Worker lifecycle (`worker`)
- product CLI parsing (`yoi`)
## Design notes
Provider API facts drift. Keep wire-format, auth, catalog, and capability differences here so Engine semantics remain stable.
Codex OAuth is a separate integration from normal provider secret refs because its local file shape and lifecycle differ.
## See also
- [`../../docs/design/provider-model-boundary.md`](../../docs/design/provider-model-boundary.md)
- [`../../docs/design/profiles-manifests-prompts.md`](../../docs/design/profiles-manifests-prompts.md)
crates/tui/Cargo.toml
-1
View File
@@ -23,7 +23,6 @@ manifest = { workspace = true }
secrets = { workspace = true }
session-store = { workspace = true }
fs4 = { workspace = true }
provider = { workspace = true }
ticket = { workspace = true }
serde = { workspace = true, features = ["derive"] }
minijinja = "2.19.0"
crates/tui/src/setup_model.rs
+1 -1
View File
@@ -3,7 +3,7 @@ use std::io::{self, Write};
use std::path::{Path, PathBuf};
use std::process::ExitCode;
use provider::catalog::{self, AuthHint, ModelEntry, ProviderEntry};
use manifest::model_catalog::{self as catalog, AuthHint, ModelEntry, ProviderEntry};
use toml::Value;
use toml::map::Map;
crates/worker/Cargo.toml
+3 -2
View File
@@ -11,12 +11,12 @@ default = []
[dependencies]
async-trait = { workspace = true }
clap = { version = "4.6.0", features = ["derive"] }
llm-engine = { workspace = true }
llm-engine = { workspace = true, features = ["codex"] }
session-store = { workspace = true }
secrets = { workspace = true }
manifest = { workspace = true }
mcp = { workspace = true }
protocol = { workspace = true }
provider = { workspace = true }
client = { workspace = true }
serde = { workspace = true, features = ["derive"] }
serde_json = { workspace = true }
@@ -46,6 +46,7 @@ futures-util = { version = "0.3", features = ["sink"] }
[dev-dependencies]
dotenv = "0.15.0"
futures = { workspace = true }
serial_test = "3.4.0"
tempfile = { workspace = true }
wat = "1.241.2"
yoi-plugin-pdk = { workspace = true }
crates/worker/src/controller.rs
+2 -2
View File
@@ -1449,11 +1449,11 @@ where
let manifest = worker.manifest();
// `build_client` がここに到達する前に同じマニフェストで成功している
// ため、カタログ解決も必ず通る。念のため失敗時は "unknown" に落とす。
let resolved = provider::catalog::resolve_model_manifest(&manifest.model).ok();
let resolved = manifest::model_catalog::resolve_model_manifest(&manifest.model).ok();
let context_window = resolved
.as_ref()
.map(|cfg| cfg.context_window)
.unwrap_or(provider::catalog::DEFAULT_CONTEXT_WINDOW);
.unwrap_or(manifest::model_catalog::DEFAULT_CONTEXT_WINDOW);
let (provider_name, model_id) = match resolved {
Some(cfg) => {
let name = match cfg.scheme {
crates/worker/src/lib.rs
+2 -1
View File
@@ -8,6 +8,7 @@ pub mod fs_view;
pub mod hook;
pub(crate) mod in_flight;
pub mod ipc;
pub mod model_client;
pub mod prompt;
pub mod runtime;
pub mod segment_log_sink;
@@ -30,11 +31,11 @@ pub use manifest::{
AuthRef, ModelManifest, SchemeKind, Scope, WorkerManifest, WorkerManifestConfig,
WorkerMetaConfig,
};
pub use model_client::{ProviderError, build_client};
pub use prompt::catalog::{CatalogError, PromptCatalog, WorkerPrompt};
pub use prompt::loader::PromptLoader;
pub use prompt::system::{SystemPromptContext, SystemPromptError, SystemPromptTemplate};
pub use protocol::{ErrorCode, Event, Method, TurnResult, WorkerStatus};
pub use provider::{ProviderError, build_client};
pub use runtime::dir::RuntimeDir;
pub use segment_log_sink::SegmentLogSink;
pub use shared_state::WorkerSharedState;
crates/provider/src/lib.rs → crates/worker/src/model_client.rs
+75 -23
View File
@@ -1,5 +1,4 @@
//! Worker マニフェストの [`ModelManifest`] を [`Box<dyn LlmClient>`]
//! に落とすファクトリ。
//! [`ModelManifest`] を [`Box<dyn LlmClient>`] に落とす worker-side factory。
//!
//! 段階:
//! 1. `ModelManifest` を [`catalog::resolve_model_manifest`] で
@@ -11,11 +10,7 @@
//! `catalog::resolve_model_manifest` が [`ModelConfig`] に詰め込む）
//!
//! llm-engine は低レベル基盤に留める方針なので、高レベル側で必要に
//! なる認証ストア解決（Codex OAuth の `~/.codex/auth.json` 読取等）は
//! このクレートに追加する。
pub mod catalog;
pub mod codex_oauth;
//! なる認証ストア解決と secret store 解決は worker 側で行う。
use std::sync::Arc;
@@ -26,13 +21,14 @@ use llm_engine::llm_client::{
Scheme, anthropic::AnthropicScheme, gemini::GeminiScheme, openai_chat::OpenAIScheme,
openai_responses::OpenAIResponsesScheme,
},
transport::{HttpTransport, ResolvedAuth},
transport::{HttpTransport, ResolvedAuth, TransportPolicy},
};
use llm_engine::providers::codex::CodexAuthProvider;
use manifest::{AuthRef, ModelManifest, SchemeKind};
use manifest::{AuthRef, ModelManifest, SchemeKind, model_catalog as catalog};
use secrets::{SecretStore, SecretValue};
pub use catalog::{ModelConfig, ResolveError as CatalogResolveError};
pub use manifest::model_catalog::{ModelConfig, ResolveError as CatalogResolveError};
/// プロバイダ構築時のエラー。
#[derive(Debug, thiserror::Error)]
@@ -111,7 +107,7 @@ fn resolve_auth_with_resolver(
Err(ProviderError::ApiKeyMissing { scheme })
}
AuthRef::CodexOAuth => {
let provider = codex_oauth::CodexAuthProvider::from_default_home()
let provider = CodexAuthProvider::from_default_home()
.map_err(|e| ProviderError::Config(e.to_string()))?;
Ok(ResolvedAuth::Custom(Arc::new(provider)))
}
@@ -144,6 +140,7 @@ fn build_transport<S: Scheme>(
scheme: S,
config: &ModelConfig,
resolved: ResolvedAuth,
policy: TransportPolicy,
) -> Result<Box<dyn LlmClient>, ProviderError> {
if !resolved.matches(scheme.required_auth()) {
return Err(ProviderError::AuthMismatch {
@@ -160,21 +157,39 @@ fn build_transport<S: Scheme>(
.clone()
.unwrap_or_else(|| scheme.default_capability());
let base_url = effective_base_url(&scheme, config);
Ok(Box::new(HttpTransport::new(
scheme,
config.model_id.clone(),
base_url,
resolved,
capability,
)))
Ok(Box::new(
HttpTransport::new(
scheme,
config.model_id.clone(),
base_url,
resolved,
capability,
)
.with_transport_policy(policy),
))
}
fn build_from_config(config: &ModelConfig) -> Result<Box<dyn LlmClient>, ProviderError> {
let resolved = resolve_auth(config.scheme, &config.auth)?;
match config.scheme {
SchemeKind::Anthropic => build_transport(AnthropicScheme::new(), config, resolved),
SchemeKind::OpenaiChat => build_transport(OpenAIScheme::new(), config, resolved),
SchemeKind::Gemini => build_transport(GeminiScheme::new(), config, resolved),
SchemeKind::Anthropic => build_transport(
AnthropicScheme::new(),
config,
resolved,
TransportPolicy::standard(),
),
SchemeKind::OpenaiChat => build_transport(
OpenAIScheme::new(),
config,
resolved,
TransportPolicy::standard(),
),
SchemeKind::Gemini => build_transport(
GeminiScheme::new(),
config,
resolved,
TransportPolicy::standard(),
),
SchemeKind::OpenaiResponses => {
// ChatGPT backend (codex-oauth) は `max_output_tokens` /
// `temperature` / `top_p` を 400 で弾くため、その経路では
@@ -183,7 +198,12 @@ fn build_from_config(config: &ModelConfig) -> Result<Box<dyn LlmClient>, Provide
let scheme = OpenAIResponsesScheme::new()
.with_send_max_output_tokens(send_to_official)
.with_send_sampling_params(send_to_official);
build_transport(scheme, config, resolved)
let policy = if matches!(config.auth, AuthRef::CodexOAuth) {
TransportPolicy::openai_compatible_zstd()
} else {
TransportPolicy::standard()
};
build_transport(scheme, config, resolved, policy)
}
}
}
@@ -208,10 +228,34 @@ mod tests {
use super::*;
use serial_test::serial;
use std::io::Write;
use std::path::PathBuf;
use std::path::{Path, PathBuf};
struct TestSecrets(std::collections::BTreeMap<String, String>);
struct ConfigDirGuard {
prev: Option<String>,
}
impl ConfigDirGuard {
fn new(path: &Path) -> Self {
let prev = std::env::var("YOI_CONFIG_DIR").ok();
// SAFETY: tests using this guard are marked `#[serial]`.
unsafe { std::env::set_var("YOI_CONFIG_DIR", path) };
Self { prev }
}
}
impl Drop for ConfigDirGuard {
fn drop(&mut self) {
unsafe {
match &self.prev {
Some(v) => std::env::set_var("YOI_CONFIG_DIR", v),
None => std::env::remove_var("YOI_CONFIG_DIR"),
}
}
}
}
impl SecretResolver for TestSecrets {
fn get_secret(&self, id: &str) -> Result<SecretValue, secrets::Error> {
self.0
@@ -311,7 +355,11 @@ mod tests {
}
#[test]
#[serial]
fn ref_manifest_builds_client() {
let dir = tempfile::tempdir().unwrap();
let _guard = ConfigDirGuard::new(dir.path());
// Ollama は AuthRef::None で構築できる end-to-end path。
let manifest = ModelManifest {
ref_: Some("ollama-local/llama3.1".into()),
@@ -326,7 +374,11 @@ mod tests {
}
#[test]
#[serial]
fn inline_manifest_builds_client() {
let dir = tempfile::tempdir().unwrap();
let _guard = ConfigDirGuard::new(dir.path());
// Form C: 完全直書き。Ollama 相当を AuthRef::None で構築。
let manifest = ModelManifest {
scheme: Some(SchemeKind::Anthropic),
crates/worker/src/worker.rs
+5 -5
View File
@@ -3048,7 +3048,7 @@ impl<C: LlmClient, St: Store> Worker<C, St> {
fn build_compactor_client(&self) -> Result<Box<dyn LlmClient>, WorkerError> {
if let Some(ref compaction) = self.manifest.compaction {
if let Some(ref model_config) = compaction.model {
let client = provider::build_client(model_config)?;
let client = crate::model_client::build_client(model_config)?;
return Ok(client);
}
}
@@ -3065,7 +3065,7 @@ impl<C: LlmClient, St: Store> Worker<C, St> {
memory_cfg: &manifest::MemoryConfig,
) -> Result<Box<dyn LlmClient>, WorkerError> {
if let Some(ref m) = memory_cfg.extract_model {
let client = provider::build_client(m)?;
let client = crate::model_client::build_client(m)?;
return Ok(client);
}
let worker = self.engine.as_ref().expect("worker taken during run");
@@ -3502,7 +3502,7 @@ impl<C: LlmClient, St: Store> Worker<C, St> {
memory_cfg: &manifest::MemoryConfig,
) -> Result<Box<dyn LlmClient>, WorkerError> {
if let Some(ref m) = memory_cfg.consolidation_model {
let client = provider::build_client(m)?;
let client = crate::model_client::build_client(m)?;
return Ok(client);
}
let worker = self.engine.as_ref().expect("worker taken during run");
@@ -5070,7 +5070,7 @@ pub enum WorkerError {
ManifestResolve(#[source] ResolveError),
#[error(transparent)]
Provider(#[from] provider::ProviderError),
Provider(#[from] crate::model_client::ProviderError),
#[error("compaction thrash: context still exceeds threshold immediately after compact")]
CompactThrash,
@@ -5299,7 +5299,7 @@ fn prepare_worker_common_from_scope(
let delegation_scope =
DelegationScope::from_config(&manifest.delegation_scope).map_err(WorkerError::Scope)?;
let client = provider::build_client(&manifest.model)?;
let client = crate::model_client::build_client(&manifest.model)?;
let prompts = PromptCatalog::load(loader, manifest.worker.prompt_pack.as_deref())?;
let memory_layout = manifest.memory.as_ref().and_then(|mem| {
filesystem_authority