claude/zen-feynman-0u78dk
claude/code-feature-parity-q003hm
claude/elegant-carson-l1menh
claude/sigit-acp-local-chat-cx6380
claude/sigit-cloud-agent-expansion-reox0a
claude/tool-permission-system
claude/zen-feynman-0u78dk
development
feature/agent-tools-multiedit-glob-todos-remember
feature/background-commands
feature/commit-coauthor-attribution
feature/headless-mode
feature/init-command
feature/load-local-model-explicitly
feature/session-persistence-compaction
feature/sigit-code-cloud
feature/subagent-tool
feature/tool-permission-system
feature/tui-repo-tabs
feature/tui-tabs
main
release/v1.3.1
| 1 | //! Inference backend abstraction. |
| 2 | //! |
| 3 | //! The agent loop only needs to send a turn (optionally with tools) and return |
| 4 | //! tool results. This module defines that seam as the `InferenceBackend` trait |
| 5 | //! plus a few neutral types, with two implementations: |
| 6 | //! |
| 7 | //! - `LocalBackend` runs on-device through the `onde` crate (`ChatEngine`). |
| 8 | //! - `OpenAiBackend` talks to any OpenAI-compatible HTTP endpoint, configured by |
| 9 | //! `base_url`, `api_key`, and `model`. |
| 10 | //! |
| 11 | //! The trait exposes neither `onde` nor OpenAI types, so the loop does not depend |
| 12 | //! on a specific backend. |
| 13 | //! |
| 14 | //! The whole backend seam is wired up only through the interactive client, which |
| 15 | //! is `#[cfg(unix)]` (see `run_interactive` in `main.rs` and `mod tui` in |
| 16 | //! `chat.rs`). On non-Unix targets the binary runs ACP-only and drives `onde` |
| 17 | //! directly, so every item here is legitimately unused there. Suppress the |
| 18 | //! dead-code lint on those targets only — Unix builds still get full coverage. |
| 19 | #![cfg_attr(not(unix), allow(dead_code))] |
| 20 | |
| 21 | use std::sync::Arc; |
| 22 | |
| 23 | use async_trait::async_trait; |
| 24 | use onde::inference::{ChatEngine, ToolDefinition}; |
| 25 | use serde::Deserialize; |
| 26 | use tokio::sync::Mutex; |
| 27 | |
| 28 | // ── Neutral types ─────────────────────────────────────────────────────────────── |
| 29 | |
| 30 | /// A tool the model may call, in a provider-neutral form. `parameters_schema` is |
| 31 | /// a JSON Schema encoded as a string (matching how siGit already declares tools). |
| 32 | #[derive(Debug, Clone)] |
| 33 | pub struct ToolSpec { |
| 34 | pub name: String, |
| 35 | pub description: String, |
| 36 | pub parameters_schema: String, |
| 37 | } |
| 38 | |
| 39 | /// A tool call requested by the model. |
| 40 | #[derive(Debug, Clone)] |
| 41 | pub struct ToolCall { |
| 42 | pub id: String, |
| 43 | pub name: String, |
| 44 | /// Arguments as a JSON-encoded string. |
| 45 | pub arguments: String, |
| 46 | } |
| 47 | |
| 48 | /// The output of executing one tool call, fed back to the model. |
| 49 | #[derive(Debug, Clone)] |
| 50 | pub struct ToolResult { |
| 51 | pub tool_call_id: String, |
| 52 | pub content: String, |
| 53 | } |
| 54 | |
| 55 | /// The result of one assistant turn: free text and/or tool calls. |
| 56 | #[derive(Debug, Clone, Default)] |
| 57 | pub struct TurnResult { |
| 58 | pub text: String, |
| 59 | pub tool_calls: Vec<ToolCall>, |
| 60 | } |
| 61 | |
| 62 | /// Backend errors are plain strings. Callers map them to ACP errors. |
| 63 | pub type BackendError = String; |
| 64 | |
| 65 | // ── The trait ─────────────────────────────────────────────────────────────────── |
| 66 | |
| 67 | /// A swappable inference backend driving siGit Code's agent loop. |
| 68 | #[async_trait] |
| 69 | pub trait InferenceBackend: Send + Sync { |
| 70 | /// Start an assistant turn from a new user message, offering `tools`. |
| 71 | async fn send_message_with_tools( |
| 72 | &self, |
| 73 | text: &str, |
| 74 | tools: &[ToolSpec], |
| 75 | ) -> Result<TurnResult, BackendError>; |
| 76 | |
| 77 | /// Continue the turn by returning tool results. `tools` may be `None` on the |
| 78 | /// final round to force a text answer. |
| 79 | async fn send_tool_results( |
| 80 | &self, |
| 81 | results: Vec<ToolResult>, |
| 82 | tools: Option<&[ToolSpec]>, |
| 83 | ) -> Result<TurnResult, BackendError>; |
| 84 | |
| 85 | /// Whether inference runs over the network (a configured provider) rather |
| 86 | /// than on-device. Drives UI labelling so the displayed model can't claim a |
| 87 | /// local model while requests actually go to the cloud. |
| 88 | fn is_remote(&self) -> bool; |
| 89 | } |
| 90 | |
| 91 | // ── Local backend (onde ChatEngine) ────────────────────────────────────────────── |
| 92 | |
| 93 | /// On-device inference. A thin adapter over `onde::ChatEngine`. |
| 94 | pub struct LocalBackend { |
| 95 | engine: Arc<ChatEngine>, |
| 96 | } |
| 97 | |
| 98 | impl LocalBackend { |
| 99 | pub fn new(engine: Arc<ChatEngine>) -> Self { |
| 100 | Self { engine } |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | fn to_onde_tools(tools: &[ToolSpec]) -> Vec<ToolDefinition> { |
| 105 | tools |
| 106 | .iter() |
| 107 | .map(|tool| ToolDefinition { |
| 108 | name: tool.name.clone(), |
| 109 | description: tool.description.clone(), |
| 110 | parameters_schema: tool.parameters_schema.clone(), |
| 111 | }) |
| 112 | .collect() |
| 113 | } |
| 114 | |
| 115 | #[async_trait] |
| 116 | impl InferenceBackend for LocalBackend { |
| 117 | async fn send_message_with_tools( |
| 118 | &self, |
| 119 | text: &str, |
| 120 | tools: &[ToolSpec], |
| 121 | ) -> Result<TurnResult, BackendError> { |
| 122 | let onde_tools = to_onde_tools(tools); |
| 123 | let result = self |
| 124 | .engine |
| 125 | .send_message_with_tools(text, &onde_tools) |
| 126 | .await |
| 127 | .map_err(|error| error.to_string())?; |
| 128 | Ok(onde_result_to_turn(result)) |
| 129 | } |
| 130 | |
| 131 | async fn send_tool_results( |
| 132 | &self, |
| 133 | results: Vec<ToolResult>, |
| 134 | tools: Option<&[ToolSpec]>, |
| 135 | ) -> Result<TurnResult, BackendError> { |
| 136 | let onde_results: Vec<onde::inference::ToolResult> = results |
| 137 | .into_iter() |
| 138 | .map(|result| onde::inference::ToolResult { |
| 139 | tool_call_id: result.tool_call_id, |
| 140 | content: result.content, |
| 141 | }) |
| 142 | .collect(); |
| 143 | let onde_tools = tools.map(to_onde_tools); |
| 144 | let result = self |
| 145 | .engine |
| 146 | .send_tool_results(onde_results, onde_tools.as_deref()) |
| 147 | .await |
| 148 | .map_err(|error| error.to_string())?; |
| 149 | Ok(onde_result_to_turn(result)) |
| 150 | } |
| 151 | |
| 152 | fn is_remote(&self) -> bool { |
| 153 | false |
| 154 | } |
| 155 | } |
| 156 | |
| 157 | /// Convert an `onde` tool-aware result into the neutral [`TurnResult`]. |
| 158 | fn onde_result_to_turn(result: onde::inference::ToolAwareResult) -> TurnResult { |
| 159 | TurnResult { |
| 160 | text: result.text, |
| 161 | tool_calls: result |
| 162 | .tool_calls |
| 163 | .into_iter() |
| 164 | .map(|call| ToolCall { |
| 165 | id: call.id, |
| 166 | name: call.function_name, |
| 167 | arguments: call.arguments, |
| 168 | }) |
| 169 | .collect(), |
| 170 | } |
| 171 | } |
| 172 | |
| 173 | // ── OpenAI-compatible backend ───────────────────────────────────────────────────── |
| 174 | |
| 175 | /// Inference against any OpenAI-compatible Chat Completions endpoint. |
| 176 | /// |
| 177 | /// Conversation state is held client-side and replayed on every request, so the |
| 178 | /// endpoint can be stateless. Standard OpenAI function-calling is used end to |
| 179 | /// end (`tools`, `choices[].message.tool_calls`, `role: "tool"` follow-ups). |
| 180 | pub struct OpenAiBackend { |
| 181 | base_url: String, |
| 182 | api_key: String, |
| 183 | model: String, |
| 184 | http: reqwest::Client, |
| 185 | /// The full message list sent on each request (system + turns + tool results). |
| 186 | history: Mutex<Vec<serde_json::Value>>, |
| 187 | } |
| 188 | |
| 189 | impl OpenAiBackend { |
| 190 | /// Build a backend for `{base_url, api_key, model}`, seeding the optional |
| 191 | /// system prompt. `base_url` should include the API root (e.g. ending in |
| 192 | /// `/v1`); the chat path is appended. |
| 193 | pub fn new( |
| 194 | base_url: impl Into<String>, |
| 195 | api_key: impl Into<String>, |
| 196 | model: impl Into<String>, |
| 197 | system_prompt: Option<String>, |
| 198 | ) -> Self { |
| 199 | let mut history = Vec::new(); |
| 200 | if let Some(prompt) = system_prompt { |
| 201 | history.push(serde_json::json!({ "role": "system", "content": prompt })); |
| 202 | } |
| 203 | Self { |
| 204 | base_url: base_url.into(), |
| 205 | api_key: api_key.into(), |
| 206 | model: model.into(), |
| 207 | http: reqwest::Client::new(), |
| 208 | history: Mutex::new(history), |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | fn tools_json(tools: &[ToolSpec]) -> Vec<serde_json::Value> { |
| 213 | tools |
| 214 | .iter() |
| 215 | .map(|tool| { |
| 216 | // parameters_schema is a JSON string; parse it, defaulting to an |
| 217 | // empty object schema if malformed. |
| 218 | let parameters: serde_json::Value = serde_json::from_str(&tool.parameters_schema) |
| 219 | .unwrap_or_else(|_| serde_json::json!({ "type": "object", "properties": {} })); |
| 220 | serde_json::json!({ |
| 221 | "type": "function", |
| 222 | "function": { |
| 223 | "name": tool.name, |
| 224 | "description": tool.description, |
| 225 | "parameters": parameters, |
| 226 | } |
| 227 | }) |
| 228 | }) |
| 229 | .collect() |
| 230 | } |
| 231 | |
| 232 | /// POST the current history (plus `tools`) and apply the assistant reply to |
| 233 | /// history, returning the neutral turn result. |
| 234 | async fn complete(&self, tools: Option<&[ToolSpec]>) -> Result<TurnResult, BackendError> { |
| 235 | let url = format!("{}/chat/completions", self.base_url.trim_end_matches('/')); |
| 236 | |
| 237 | let mut body = serde_json::json!({ |
| 238 | "model": self.model, |
| 239 | "messages": *self.history.lock().await, |
| 240 | "stream": false, |
| 241 | }); |
| 242 | if let Some(tools) = tools |
| 243 | && !tools.is_empty() |
| 244 | { |
| 245 | body["tools"] = serde_json::Value::Array(Self::tools_json(tools)); |
| 246 | } |
| 247 | |
| 248 | let response = self |
| 249 | .http |
| 250 | .post(&url) |
| 251 | .bearer_auth(&self.api_key) |
| 252 | .json(&body) |
| 253 | .send() |
| 254 | .await |
| 255 | .map_err(|error| format!("request to {url} failed: {error}"))?; |
| 256 | |
| 257 | if !response.status().is_success() { |
| 258 | let status = response.status(); |
| 259 | let detail = response.text().await.unwrap_or_default(); |
| 260 | return Err(format!("endpoint returned {status}: {detail}")); |
| 261 | } |
| 262 | |
| 263 | let parsed: ChatCompletion = response |
| 264 | .json() |
| 265 | .await |
| 266 | .map_err(|error| format!("response parse error: {error}"))?; |
| 267 | |
| 268 | let message = parsed |
| 269 | .choices |
| 270 | .into_iter() |
| 271 | .next() |
| 272 | .map(|choice| choice.message) |
| 273 | .ok_or_else(|| "endpoint returned no choices".to_string())?; |
| 274 | |
| 275 | let text = message.content.clone().unwrap_or_default(); |
| 276 | let tool_calls: Vec<ToolCall> = message |
| 277 | .tool_calls |
| 278 | .iter() |
| 279 | .flatten() |
| 280 | .map(|call| ToolCall { |
| 281 | id: call.id.clone(), |
| 282 | name: call.function.name.clone(), |
| 283 | arguments: call.function.arguments.clone(), |
| 284 | }) |
| 285 | .collect(); |
| 286 | |
| 287 | // Record the assistant turn so later tool results have context. |
| 288 | self.history.lock().await.push(message.into_history_value()); |
| 289 | |
| 290 | Ok(TurnResult { text, tool_calls }) |
| 291 | } |
| 292 | } |
| 293 | |
| 294 | #[async_trait] |
| 295 | impl InferenceBackend for OpenAiBackend { |
| 296 | async fn send_message_with_tools( |
| 297 | &self, |
| 298 | text: &str, |
| 299 | tools: &[ToolSpec], |
| 300 | ) -> Result<TurnResult, BackendError> { |
| 301 | self.history |
| 302 | .lock() |
| 303 | .await |
| 304 | .push(serde_json::json!({ "role": "user", "content": text })); |
| 305 | self.complete(Some(tools)).await |
| 306 | } |
| 307 | |
| 308 | async fn send_tool_results( |
| 309 | &self, |
| 310 | results: Vec<ToolResult>, |
| 311 | tools: Option<&[ToolSpec]>, |
| 312 | ) -> Result<TurnResult, BackendError> { |
| 313 | { |
| 314 | let mut history = self.history.lock().await; |
| 315 | for result in results { |
| 316 | history.push(serde_json::json!({ |
| 317 | "role": "tool", |
| 318 | "tool_call_id": result.tool_call_id, |
| 319 | "content": result.content, |
| 320 | })); |
| 321 | } |
| 322 | } |
| 323 | self.complete(tools).await |
| 324 | } |
| 325 | |
| 326 | fn is_remote(&self) -> bool { |
| 327 | true |
| 328 | } |
| 329 | } |
| 330 | |
| 331 | // ── OpenAI response shapes ──────────────────────────────────────────────────────── |
| 332 | |
| 333 | #[derive(Debug, Deserialize)] |
| 334 | struct ChatCompletion { |
| 335 | #[serde(default)] |
| 336 | choices: Vec<CompletionChoice>, |
| 337 | } |
| 338 | |
| 339 | #[derive(Debug, Deserialize)] |
| 340 | struct CompletionChoice { |
| 341 | message: ResponseMessage, |
| 342 | } |
| 343 | |
| 344 | #[derive(Debug, Deserialize)] |
| 345 | struct ResponseMessage { |
| 346 | #[serde(default)] |
| 347 | content: Option<String>, |
| 348 | #[serde(default)] |
| 349 | tool_calls: Option<Vec<ResponseToolCall>>, |
| 350 | } |
| 351 | |
| 352 | impl ResponseMessage { |
| 353 | /// Reconstruct the assistant message for replay in history, preserving any |
| 354 | /// tool calls so the follow-up request is well-formed. |
| 355 | fn into_history_value(self) -> serde_json::Value { |
| 356 | let mut message = serde_json::json!({ "role": "assistant" }); |
| 357 | message["content"] = match self.content { |
| 358 | Some(text) => serde_json::Value::String(text), |
| 359 | None => serde_json::Value::Null, |
| 360 | }; |
| 361 | if let Some(tool_calls) = self.tool_calls { |
| 362 | message["tool_calls"] = serde_json::json!( |
| 363 | tool_calls |
| 364 | .into_iter() |
| 365 | .map(|call| serde_json::json!({ |
| 366 | "id": call.id, |
| 367 | "type": "function", |
| 368 | "function": { |
| 369 | "name": call.function.name, |
| 370 | "arguments": call.function.arguments, |
| 371 | } |
| 372 | })) |
| 373 | .collect::<Vec<_>>() |
| 374 | ); |
| 375 | } |
| 376 | message |
| 377 | } |
| 378 | } |
| 379 | |
| 380 | #[derive(Debug, Deserialize)] |
| 381 | struct ResponseToolCall { |
| 382 | id: String, |
| 383 | function: ResponseFunction, |
| 384 | } |
| 385 | |
| 386 | #[derive(Debug, Deserialize)] |
| 387 | struct ResponseFunction { |
| 388 | name: String, |
| 389 | #[serde(default)] |
| 390 | arguments: String, |
| 391 | } |
| 392 | |
| 393 | #[cfg(test)] |
| 394 | mod tests { |
| 395 | use super::*; |
| 396 | |
| 397 | #[test] |
| 398 | fn tools_json_wraps_function_schema() { |
| 399 | let tools = vec![ToolSpec { |
| 400 | name: "read_file".to_string(), |
| 401 | description: "Read a file".to_string(), |
| 402 | parameters_schema: r#"{"type":"object","properties":{"path":{"type":"string"}}}"# |
| 403 | .to_string(), |
| 404 | }]; |
| 405 | let json = OpenAiBackend::tools_json(&tools); |
| 406 | assert_eq!(json[0]["type"], "function"); |
| 407 | assert_eq!(json[0]["function"]["name"], "read_file"); |
| 408 | assert_eq!( |
| 409 | json[0]["function"]["parameters"]["properties"]["path"]["type"], |
| 410 | "string" |
| 411 | ); |
| 412 | } |
| 413 | |
| 414 | #[test] |
| 415 | fn malformed_schema_falls_back_to_empty_object() { |
| 416 | let tools = vec![ToolSpec { |
| 417 | name: "x".to_string(), |
| 418 | description: String::new(), |
| 419 | parameters_schema: "not json".to_string(), |
| 420 | }]; |
| 421 | let json = OpenAiBackend::tools_json(&tools); |
| 422 | assert_eq!(json[0]["function"]["parameters"]["type"], "object"); |
| 423 | } |
| 424 | |
| 425 | #[test] |
| 426 | fn assistant_message_with_tool_calls_round_trips() { |
| 427 | let message = ResponseMessage { |
| 428 | content: None, |
| 429 | tool_calls: Some(vec![ResponseToolCall { |
| 430 | id: "call_1".to_string(), |
| 431 | function: ResponseFunction { |
| 432 | name: "read_file".to_string(), |
| 433 | arguments: r#"{"path":"a.rs"}"#.to_string(), |
| 434 | }, |
| 435 | }]), |
| 436 | }; |
| 437 | let value = message.into_history_value(); |
| 438 | assert_eq!(value["role"], "assistant"); |
| 439 | assert!(value["content"].is_null()); |
| 440 | assert_eq!(value["tool_calls"][0]["id"], "call_1"); |
| 441 | assert_eq!(value["tool_calls"][0]["type"], "function"); |
| 442 | assert_eq!(value["tool_calls"][0]["function"]["name"], "read_file"); |
| 443 | } |
| 444 | } |