claude/sigit-acp-local-chat-cx6380
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 | /// Author of a turn replayed into a backend when a persisted session reopens. |
| 63 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 64 | pub enum HistoryRole { |
| 65 | User, |
| 66 | Assistant, |
| 67 | } |
| 68 | |
| 69 | /// A prior user/assistant turn restored into a backend's context when a |
| 70 | /// persisted session is reopened (see [`InferenceBackend::restore_history`]). |
| 71 | #[derive(Debug, Clone)] |
| 72 | pub struct HistoryMessage { |
| 73 | pub role: HistoryRole, |
| 74 | pub content: String, |
| 75 | } |
| 76 | |
| 77 | /// Backend errors are plain strings. Callers map them to ACP errors. |
| 78 | pub type BackendError = String; |
| 79 | |
| 80 | /// A sink for streaming assistant text deltas to the UI as they are produced. |
| 81 | /// |
| 82 | /// When a caller passes `Some(sink)`, a streaming-capable backend forwards each |
| 83 | /// text fragment through it as the model emits it; the returned [`TurnResult`] |
| 84 | /// still carries the fully assembled text (and any tool calls). When the sink is |
| 85 | /// `None`, the backend runs in non-streaming mode. Unbounded so the inference |
| 86 | /// task never blocks on a slow consumer. |
| 87 | pub type TokenSink = tokio::sync::mpsc::UnboundedSender<String>; |
| 88 | |
| 89 | // ── The trait ─────────────────────────────────────────────────────────────────── |
| 90 | |
| 91 | /// A swappable inference backend driving siGit Code's agent loop. |
| 92 | #[async_trait] |
| 93 | pub trait InferenceBackend: Send + Sync { |
| 94 | /// Start an assistant turn from a new user message, offering `tools`. |
| 95 | /// |
| 96 | /// If `sink` is `Some`, text is streamed through it as it is generated. A |
| 97 | /// backend may decline to stream a given round (for example, on-device |
| 98 | /// inference cannot stream while it is still deciding whether to call a |
| 99 | /// tool); in that case the text is delivered only via the returned result. |
| 100 | async fn send_message_with_tools( |
| 101 | &self, |
| 102 | text: &str, |
| 103 | tools: &[ToolSpec], |
| 104 | sink: Option<&TokenSink>, |
| 105 | ) -> Result<TurnResult, BackendError>; |
| 106 | |
| 107 | /// Continue the turn by returning tool results. `tools` may be `None` on the |
| 108 | /// final round to force a text answer. `sink` streams that text when set. |
| 109 | async fn send_tool_results( |
| 110 | &self, |
| 111 | results: Vec<ToolResult>, |
| 112 | tools: Option<&[ToolSpec]>, |
| 113 | sink: Option<&TokenSink>, |
| 114 | ) -> Result<TurnResult, BackendError>; |
| 115 | |
| 116 | /// Whether inference runs over the network (a configured provider) rather |
| 117 | /// than on-device. Drives UI labelling so the displayed model can't claim a |
| 118 | /// local model while requests actually go to the cloud. |
| 119 | fn is_remote(&self) -> bool; |
| 120 | |
| 121 | /// Replay prior user/assistant turns into the backend's context after a |
| 122 | /// persisted session is reopened, so the model continues with its history |
| 123 | /// rather than starting blank. Called once on session load, before any new |
| 124 | /// prompt. The default does nothing. |
| 125 | async fn restore_history(&self, _messages: &[HistoryMessage]) {} |
| 126 | } |
| 127 | |
| 128 | // ── Local backend (onde ChatEngine) ────────────────────────────────────────────── |
| 129 | |
| 130 | /// On-device inference. A thin adapter over `onde::ChatEngine`. |
| 131 | pub struct LocalBackend { |
| 132 | engine: Arc<ChatEngine>, |
| 133 | } |
| 134 | |
| 135 | impl LocalBackend { |
| 136 | pub fn new(engine: Arc<ChatEngine>) -> Self { |
| 137 | Self { engine } |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | fn to_onde_tools(tools: &[ToolSpec]) -> Vec<ToolDefinition> { |
| 142 | tools |
| 143 | .iter() |
| 144 | .map(|tool| ToolDefinition { |
| 145 | name: tool.name.clone(), |
| 146 | description: tool.description.clone(), |
| 147 | parameters_schema: tool.parameters_schema.clone(), |
| 148 | }) |
| 149 | .collect() |
| 150 | } |
| 151 | |
| 152 | #[async_trait] |
| 153 | impl InferenceBackend for LocalBackend { |
| 154 | async fn send_message_with_tools( |
| 155 | &self, |
| 156 | text: &str, |
| 157 | tools: &[ToolSpec], |
| 158 | sink: Option<&TokenSink>, |
| 159 | ) -> Result<TurnResult, BackendError> { |
| 160 | // onde's tool-aware path is non-streaming: it has to buffer the whole |
| 161 | // reply to detect tool calls. We can only stream when no tools are on |
| 162 | // offer (a plain answer), which is exactly the tools-disabled case. |
| 163 | if let Some(sink) = sink |
| 164 | && tools.is_empty() |
| 165 | { |
| 166 | let rx = self |
| 167 | .engine |
| 168 | .stream_message(text) |
| 169 | .await |
| 170 | .map_err(|error| error.to_string())?; |
| 171 | return drain_onde_stream(rx, sink).await; |
| 172 | } |
| 173 | |
| 174 | let onde_tools = to_onde_tools(tools); |
| 175 | let result = self |
| 176 | .engine |
| 177 | .send_message_with_tools(text, &onde_tools) |
| 178 | .await |
| 179 | .map_err(|error| error.to_string())?; |
| 180 | Ok(onde_result_to_turn(result)) |
| 181 | } |
| 182 | |
| 183 | async fn send_tool_results( |
| 184 | &self, |
| 185 | results: Vec<ToolResult>, |
| 186 | tools: Option<&[ToolSpec]>, |
| 187 | sink: Option<&TokenSink>, |
| 188 | ) -> Result<TurnResult, BackendError> { |
| 189 | let onde_results: Vec<onde::inference::ToolResult> = results |
| 190 | .into_iter() |
| 191 | .map(|result| onde::inference::ToolResult { |
| 192 | tool_call_id: result.tool_call_id, |
| 193 | content: result.content, |
| 194 | }) |
| 195 | .collect(); |
| 196 | |
| 197 | // The final round passes `tools = None` to force a text answer; that's |
| 198 | // the only round onde can stream, since no further tool calls are parsed. |
| 199 | if let Some(sink) = sink |
| 200 | && tools.is_none() |
| 201 | { |
| 202 | let rx = self |
| 203 | .engine |
| 204 | .stream_tool_results(onde_results, None) |
| 205 | .await |
| 206 | .map_err(|error| error.to_string())?; |
| 207 | return drain_onde_stream(rx, sink).await; |
| 208 | } |
| 209 | |
| 210 | let onde_tools = tools.map(to_onde_tools); |
| 211 | let result = self |
| 212 | .engine |
| 213 | .send_tool_results(onde_results, onde_tools.as_deref()) |
| 214 | .await |
| 215 | .map_err(|error| error.to_string())?; |
| 216 | Ok(onde_result_to_turn(result)) |
| 217 | } |
| 218 | |
| 219 | fn is_remote(&self) -> bool { |
| 220 | false |
| 221 | } |
| 222 | |
| 223 | async fn restore_history(&self, messages: &[HistoryMessage]) { |
| 224 | // Push the saved turns straight into onde's conversation history so the |
| 225 | // next prompt sees them. The system context is re-pushed separately by |
| 226 | // the caller, so only user/assistant turns flow through here. |
| 227 | for message in messages { |
| 228 | let chat_message = match message.role { |
| 229 | HistoryRole::User => onde::inference::ChatMessage::user(message.content.clone()), |
| 230 | HistoryRole::Assistant => { |
| 231 | onde::inference::ChatMessage::assistant(message.content.clone()) |
| 232 | } |
| 233 | }; |
| 234 | self.engine.push_history(chat_message).await; |
| 235 | } |
| 236 | } |
| 237 | } |
| 238 | |
| 239 | /// Drain an onde streaming receiver, forwarding each token to `sink` and |
| 240 | /// assembling the full text. onde reports stream failures as a final chunk whose |
| 241 | /// `finish_reason` is `"error: …"`; surface those as a backend error. |
| 242 | async fn drain_onde_stream( |
| 243 | mut rx: tokio::sync::mpsc::Receiver<onde::inference::StreamChunk>, |
| 244 | sink: &TokenSink, |
| 245 | ) -> Result<TurnResult, BackendError> { |
| 246 | let mut text = String::new(); |
| 247 | while let Some(chunk) = rx.recv().await { |
| 248 | if !chunk.delta.is_empty() { |
| 249 | text.push_str(&chunk.delta); |
| 250 | // The receiver is the UI; if it's gone the turn is being cancelled, |
| 251 | // so stop assembling rather than spinning the model to completion. |
| 252 | if sink.send(chunk.delta).is_err() { |
| 253 | break; |
| 254 | } |
| 255 | } |
| 256 | if chunk.done { |
| 257 | if let Some(reason) = chunk.finish_reason |
| 258 | && let Some(message) = reason.strip_prefix("error: ") |
| 259 | { |
| 260 | return Err(message.to_string()); |
| 261 | } |
| 262 | break; |
| 263 | } |
| 264 | } |
| 265 | Ok(TurnResult { |
| 266 | text, |
| 267 | tool_calls: Vec::new(), |
| 268 | }) |
| 269 | } |
| 270 | |
| 271 | /// Convert an `onde` tool-aware result into the neutral [`TurnResult`]. |
| 272 | fn onde_result_to_turn(result: onde::inference::ToolAwareResult) -> TurnResult { |
| 273 | TurnResult { |
| 274 | text: result.text, |
| 275 | tool_calls: result |
| 276 | .tool_calls |
| 277 | .into_iter() |
| 278 | .map(|call| ToolCall { |
| 279 | id: call.id, |
| 280 | name: call.function_name, |
| 281 | arguments: call.arguments, |
| 282 | }) |
| 283 | .collect(), |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | // ── OpenAI-compatible backend ───────────────────────────────────────────────────── |
| 288 | |
| 289 | /// Inference against any OpenAI-compatible Chat Completions endpoint. |
| 290 | /// |
| 291 | /// Conversation state is held client-side and replayed on every request, so the |
| 292 | /// endpoint can be stateless. Standard OpenAI function-calling is used end to |
| 293 | /// end (`tools`, `choices[].message.tool_calls`, `role: "tool"` follow-ups). |
| 294 | pub struct OpenAiBackend { |
| 295 | base_url: String, |
| 296 | api_key: String, |
| 297 | model: String, |
| 298 | http: reqwest::Client, |
| 299 | /// The full message list sent on each request (system + turns + tool results). |
| 300 | history: Mutex<Vec<serde_json::Value>>, |
| 301 | } |
| 302 | |
| 303 | impl OpenAiBackend { |
| 304 | /// Build a backend for `{base_url, api_key, model}`, seeding the optional |
| 305 | /// system prompt. `base_url` should include the API root (e.g. ending in |
| 306 | /// `/v1`); the chat path is appended. |
| 307 | pub fn new( |
| 308 | base_url: impl Into<String>, |
| 309 | api_key: impl Into<String>, |
| 310 | model: impl Into<String>, |
| 311 | system_prompt: Option<String>, |
| 312 | ) -> Self { |
| 313 | let mut history = Vec::new(); |
| 314 | if let Some(prompt) = system_prompt { |
| 315 | history.push(serde_json::json!({ "role": "system", "content": prompt })); |
| 316 | } |
| 317 | Self { |
| 318 | base_url: base_url.into(), |
| 319 | api_key: api_key.into(), |
| 320 | model: model.into(), |
| 321 | http: reqwest::Client::new(), |
| 322 | history: Mutex::new(history), |
| 323 | } |
| 324 | } |
| 325 | |
| 326 | fn tools_json(tools: &[ToolSpec]) -> Vec<serde_json::Value> { |
| 327 | tools |
| 328 | .iter() |
| 329 | .map(|tool| { |
| 330 | // parameters_schema is a JSON string; parse it, defaulting to an |
| 331 | // empty object schema if malformed. |
| 332 | let parameters: serde_json::Value = serde_json::from_str(&tool.parameters_schema) |
| 333 | .unwrap_or_else(|_| serde_json::json!({ "type": "object", "properties": {} })); |
| 334 | serde_json::json!({ |
| 335 | "type": "function", |
| 336 | "function": { |
| 337 | "name": tool.name, |
| 338 | "description": tool.description, |
| 339 | "parameters": parameters, |
| 340 | } |
| 341 | }) |
| 342 | }) |
| 343 | .collect() |
| 344 | } |
| 345 | |
| 346 | /// POST the current history (plus `tools`) and apply the assistant reply to |
| 347 | /// history, returning the neutral turn result. Streams via SSE when `sink` |
| 348 | /// is set; otherwise reads a single JSON response. |
| 349 | async fn complete( |
| 350 | &self, |
| 351 | tools: Option<&[ToolSpec]>, |
| 352 | sink: Option<&TokenSink>, |
| 353 | ) -> Result<TurnResult, BackendError> { |
| 354 | let url = format!("{}/chat/completions", self.base_url.trim_end_matches('/')); |
| 355 | let streaming = sink.is_some(); |
| 356 | |
| 357 | let mut body = serde_json::json!({ |
| 358 | "model": self.model, |
| 359 | "messages": *self.history.lock().await, |
| 360 | "stream": streaming, |
| 361 | }); |
| 362 | if let Some(tools) = tools |
| 363 | && !tools.is_empty() |
| 364 | { |
| 365 | body["tools"] = serde_json::Value::Array(Self::tools_json(tools)); |
| 366 | } |
| 367 | |
| 368 | let response = self |
| 369 | .http |
| 370 | .post(&url) |
| 371 | .bearer_auth(&self.api_key) |
| 372 | .json(&body) |
| 373 | .send() |
| 374 | .await |
| 375 | .map_err(|error| format!("request to {url} failed: {error}"))?; |
| 376 | |
| 377 | if !response.status().is_success() { |
| 378 | let status = response.status(); |
| 379 | let detail = response.text().await.unwrap_or_default(); |
| 380 | return Err(format!("endpoint returned {status}: {detail}")); |
| 381 | } |
| 382 | |
| 383 | if let Some(sink) = sink { |
| 384 | self.consume_stream(response, sink).await |
| 385 | } else { |
| 386 | self.consume_json(response).await |
| 387 | } |
| 388 | } |
| 389 | |
| 390 | /// Parse a single non-streaming chat-completion response. |
| 391 | async fn consume_json(&self, response: reqwest::Response) -> Result<TurnResult, BackendError> { |
| 392 | let parsed: ChatCompletion = response |
| 393 | .json() |
| 394 | .await |
| 395 | .map_err(|error| format!("response parse error: {error}"))?; |
| 396 | |
| 397 | let message = parsed |
| 398 | .choices |
| 399 | .into_iter() |
| 400 | .next() |
| 401 | .map(|choice| choice.message) |
| 402 | .ok_or_else(|| "endpoint returned no choices".to_string())?; |
| 403 | |
| 404 | let text = message.content.clone().unwrap_or_default(); |
| 405 | let tool_calls: Vec<ToolCall> = message |
| 406 | .tool_calls |
| 407 | .iter() |
| 408 | .flatten() |
| 409 | .map(|call| ToolCall { |
| 410 | id: call.id.clone(), |
| 411 | name: call.function.name.clone(), |
| 412 | arguments: call.function.arguments.clone(), |
| 413 | }) |
| 414 | .collect(); |
| 415 | |
| 416 | // Record the assistant turn so later tool results have context. |
| 417 | self.history.lock().await.push(message.into_history_value()); |
| 418 | |
| 419 | Ok(TurnResult { text, tool_calls }) |
| 420 | } |
| 421 | |
| 422 | /// Consume an OpenAI Server-Sent Events stream, forwarding content deltas to |
| 423 | /// `sink` and reassembling any tool calls (which arrive fragmented across |
| 424 | /// chunks, keyed by `index`). |
| 425 | async fn consume_stream( |
| 426 | &self, |
| 427 | response: reqwest::Response, |
| 428 | sink: &TokenSink, |
| 429 | ) -> Result<TurnResult, BackendError> { |
| 430 | use futures::StreamExt; |
| 431 | |
| 432 | let mut stream = response.bytes_stream(); |
| 433 | // Newlines are ASCII, so splitting raw bytes on `\n` never bisects a |
| 434 | // multibyte UTF-8 sequence; we only lossily decode whole lines. |
| 435 | let mut buffer: Vec<u8> = Vec::new(); |
| 436 | let mut text = String::new(); |
| 437 | let mut tool_accum: Vec<StreamingToolCall> = Vec::new(); |
| 438 | let mut done = false; |
| 439 | |
| 440 | while let Some(item) = stream.next().await { |
| 441 | let bytes = item.map_err(|error| format!("stream read error: {error}"))?; |
| 442 | buffer.extend_from_slice(&bytes); |
| 443 | |
| 444 | while let Some(pos) = buffer.iter().position(|&b| b == b'\n') { |
| 445 | let line: Vec<u8> = buffer.drain(..=pos).collect(); |
| 446 | let line = String::from_utf8_lossy(&line); |
| 447 | let line = line.trim(); |
| 448 | |
| 449 | let Some(data) = line.strip_prefix("data:") else { |
| 450 | continue; |
| 451 | }; |
| 452 | let data = data.trim(); |
| 453 | if data == "[DONE]" { |
| 454 | done = true; |
| 455 | break; |
| 456 | } |
| 457 | if data.is_empty() { |
| 458 | continue; |
| 459 | } |
| 460 | |
| 461 | let chunk: StreamCompletion = match serde_json::from_str(data) { |
| 462 | Ok(chunk) => chunk, |
| 463 | // Skip keep-alive comments and anything we can't parse rather |
| 464 | // than aborting a turn over one malformed frame. |
| 465 | Err(_) => continue, |
| 466 | }; |
| 467 | |
| 468 | let Some(choice) = chunk.choices.into_iter().next() else { |
| 469 | continue; |
| 470 | }; |
| 471 | if let Some(content) = choice.delta.content |
| 472 | && !content.is_empty() |
| 473 | { |
| 474 | text.push_str(&content); |
| 475 | if sink.send(content).is_err() { |
| 476 | // Consumer dropped (turn cancelled) — stop reading. |
| 477 | done = true; |
| 478 | break; |
| 479 | } |
| 480 | } |
| 481 | for delta in choice.delta.tool_calls.into_iter().flatten() { |
| 482 | let index = delta.index.unwrap_or(0) as usize; |
| 483 | if tool_accum.len() <= index { |
| 484 | tool_accum.resize_with(index + 1, StreamingToolCall::default); |
| 485 | } |
| 486 | let slot = &mut tool_accum[index]; |
| 487 | if let Some(id) = delta.id { |
| 488 | slot.id = id; |
| 489 | } |
| 490 | if let Some(function) = delta.function { |
| 491 | if let Some(name) = function.name { |
| 492 | slot.name = name; |
| 493 | } |
| 494 | if let Some(arguments) = function.arguments { |
| 495 | slot.arguments.push_str(&arguments); |
| 496 | } |
| 497 | } |
| 498 | } |
| 499 | } |
| 500 | |
| 501 | if done { |
| 502 | break; |
| 503 | } |
| 504 | } |
| 505 | |
| 506 | let tool_calls: Vec<ToolCall> = tool_accum |
| 507 | .iter() |
| 508 | .filter(|call| !call.name.is_empty()) |
| 509 | .enumerate() |
| 510 | .map(|(index, call)| ToolCall { |
| 511 | id: if call.id.is_empty() { |
| 512 | format!("call_{index}") |
| 513 | } else { |
| 514 | call.id.clone() |
| 515 | }, |
| 516 | name: call.name.clone(), |
| 517 | arguments: call.arguments.clone(), |
| 518 | }) |
| 519 | .collect(); |
| 520 | |
| 521 | // Record the assistant turn so later tool results have context. |
| 522 | self.history |
| 523 | .lock() |
| 524 | .await |
| 525 | .push(streamed_assistant_history(&text, &tool_calls)); |
| 526 | |
| 527 | Ok(TurnResult { text, tool_calls }) |
| 528 | } |
| 529 | } |
| 530 | |
| 531 | /// One tool call being reassembled from streamed deltas. |
| 532 | #[derive(Default)] |
| 533 | struct StreamingToolCall { |
| 534 | id: String, |
| 535 | name: String, |
| 536 | arguments: String, |
| 537 | } |
| 538 | |
| 539 | /// Rebuild the assistant message for replay in history after a streamed turn, |
| 540 | /// preserving any tool calls so the follow-up request is well-formed. Mirrors |
| 541 | /// [`ResponseMessage::into_history_value`] for the non-streaming path. |
| 542 | fn streamed_assistant_history(text: &str, tool_calls: &[ToolCall]) -> serde_json::Value { |
| 543 | let mut message = serde_json::json!({ "role": "assistant" }); |
| 544 | message["content"] = if text.is_empty() { |
| 545 | serde_json::Value::Null |
| 546 | } else { |
| 547 | serde_json::Value::String(text.to_string()) |
| 548 | }; |
| 549 | if !tool_calls.is_empty() { |
| 550 | message["tool_calls"] = serde_json::json!( |
| 551 | tool_calls |
| 552 | .iter() |
| 553 | .map(|call| serde_json::json!({ |
| 554 | "id": call.id, |
| 555 | "type": "function", |
| 556 | "function": { |
| 557 | "name": call.name, |
| 558 | "arguments": call.arguments, |
| 559 | } |
| 560 | })) |
| 561 | .collect::<Vec<_>>() |
| 562 | ); |
| 563 | } |
| 564 | message |
| 565 | } |
| 566 | |
| 567 | #[async_trait] |
| 568 | impl InferenceBackend for OpenAiBackend { |
| 569 | async fn send_message_with_tools( |
| 570 | &self, |
| 571 | text: &str, |
| 572 | tools: &[ToolSpec], |
| 573 | sink: Option<&TokenSink>, |
| 574 | ) -> Result<TurnResult, BackendError> { |
| 575 | self.history |
| 576 | .lock() |
| 577 | .await |
| 578 | .push(serde_json::json!({ "role": "user", "content": text })); |
| 579 | self.complete(Some(tools), sink).await |
| 580 | } |
| 581 | |
| 582 | async fn send_tool_results( |
| 583 | &self, |
| 584 | results: Vec<ToolResult>, |
| 585 | tools: Option<&[ToolSpec]>, |
| 586 | sink: Option<&TokenSink>, |
| 587 | ) -> Result<TurnResult, BackendError> { |
| 588 | { |
| 589 | let mut history = self.history.lock().await; |
| 590 | for result in results { |
| 591 | history.push(serde_json::json!({ |
| 592 | "role": "tool", |
| 593 | "tool_call_id": result.tool_call_id, |
| 594 | "content": result.content, |
| 595 | })); |
| 596 | } |
| 597 | } |
| 598 | self.complete(tools, sink).await |
| 599 | } |
| 600 | |
| 601 | fn is_remote(&self) -> bool { |
| 602 | true |
| 603 | } |
| 604 | |
| 605 | async fn restore_history(&self, messages: &[HistoryMessage]) { |
| 606 | // Splice the saved turns in after the seeded system prompt and before |
| 607 | // any new request, matching the OpenAI chat history shape. |
| 608 | let mut history = self.history.lock().await; |
| 609 | for message in messages { |
| 610 | let role = match message.role { |
| 611 | HistoryRole::User => "user", |
| 612 | HistoryRole::Assistant => "assistant", |
| 613 | }; |
| 614 | history.push(serde_json::json!({ "role": role, "content": message.content })); |
| 615 | } |
| 616 | } |
| 617 | } |
| 618 | |
| 619 | // ── OpenAI response shapes ──────────────────────────────────────────────────────── |
| 620 | |
| 621 | #[derive(Debug, Deserialize)] |
| 622 | struct ChatCompletion { |
| 623 | #[serde(default)] |
| 624 | choices: Vec<CompletionChoice>, |
| 625 | } |
| 626 | |
| 627 | #[derive(Debug, Deserialize)] |
| 628 | struct CompletionChoice { |
| 629 | message: ResponseMessage, |
| 630 | } |
| 631 | |
| 632 | #[derive(Debug, Deserialize)] |
| 633 | struct ResponseMessage { |
| 634 | #[serde(default)] |
| 635 | content: Option<String>, |
| 636 | #[serde(default)] |
| 637 | tool_calls: Option<Vec<ResponseToolCall>>, |
| 638 | } |
| 639 | |
| 640 | impl ResponseMessage { |
| 641 | /// Reconstruct the assistant message for replay in history, preserving any |
| 642 | /// tool calls so the follow-up request is well-formed. |
| 643 | fn into_history_value(self) -> serde_json::Value { |
| 644 | let mut message = serde_json::json!({ "role": "assistant" }); |
| 645 | message["content"] = match self.content { |
| 646 | Some(text) => serde_json::Value::String(text), |
| 647 | None => serde_json::Value::Null, |
| 648 | }; |
| 649 | if let Some(tool_calls) = self.tool_calls { |
| 650 | message["tool_calls"] = serde_json::json!( |
| 651 | tool_calls |
| 652 | .into_iter() |
| 653 | .map(|call| serde_json::json!({ |
| 654 | "id": call.id, |
| 655 | "type": "function", |
| 656 | "function": { |
| 657 | "name": call.function.name, |
| 658 | "arguments": call.function.arguments, |
| 659 | } |
| 660 | })) |
| 661 | .collect::<Vec<_>>() |
| 662 | ); |
| 663 | } |
| 664 | message |
| 665 | } |
| 666 | } |
| 667 | |
| 668 | #[derive(Debug, Deserialize)] |
| 669 | struct ResponseToolCall { |
| 670 | id: String, |
| 671 | function: ResponseFunction, |
| 672 | } |
| 673 | |
| 674 | #[derive(Debug, Deserialize)] |
| 675 | struct ResponseFunction { |
| 676 | name: String, |
| 677 | #[serde(default)] |
| 678 | arguments: String, |
| 679 | } |
| 680 | |
| 681 | // ── OpenAI streaming (SSE) chunk shapes ───────────────────────────────────────── |
| 682 | |
| 683 | #[derive(Debug, Deserialize)] |
| 684 | struct StreamCompletion { |
| 685 | #[serde(default)] |
| 686 | choices: Vec<StreamChoice>, |
| 687 | } |
| 688 | |
| 689 | #[derive(Debug, Deserialize)] |
| 690 | struct StreamChoice { |
| 691 | #[serde(default)] |
| 692 | delta: StreamDelta, |
| 693 | } |
| 694 | |
| 695 | #[derive(Debug, Default, Deserialize)] |
| 696 | struct StreamDelta { |
| 697 | #[serde(default)] |
| 698 | content: Option<String>, |
| 699 | #[serde(default)] |
| 700 | tool_calls: Option<Vec<StreamToolCallDelta>>, |
| 701 | } |
| 702 | |
| 703 | #[derive(Debug, Deserialize)] |
| 704 | struct StreamToolCallDelta { |
| 705 | #[serde(default)] |
| 706 | index: Option<u32>, |
| 707 | #[serde(default)] |
| 708 | id: Option<String>, |
| 709 | #[serde(default)] |
| 710 | function: Option<StreamFunctionDelta>, |
| 711 | } |
| 712 | |
| 713 | #[derive(Debug, Deserialize)] |
| 714 | struct StreamFunctionDelta { |
| 715 | #[serde(default)] |
| 716 | name: Option<String>, |
| 717 | #[serde(default)] |
| 718 | arguments: Option<String>, |
| 719 | } |
| 720 | |
| 721 | #[cfg(test)] |
| 722 | mod tests { |
| 723 | use super::*; |
| 724 | |
| 725 | #[test] |
| 726 | fn tools_json_wraps_function_schema() { |
| 727 | let tools = vec![ToolSpec { |
| 728 | name: "read_file".to_string(), |
| 729 | description: "Read a file".to_string(), |
| 730 | parameters_schema: r#"{"type":"object","properties":{"path":{"type":"string"}}}"# |
| 731 | .to_string(), |
| 732 | }]; |
| 733 | let json = OpenAiBackend::tools_json(&tools); |
| 734 | assert_eq!(json[0]["type"], "function"); |
| 735 | assert_eq!(json[0]["function"]["name"], "read_file"); |
| 736 | assert_eq!( |
| 737 | json[0]["function"]["parameters"]["properties"]["path"]["type"], |
| 738 | "string" |
| 739 | ); |
| 740 | } |
| 741 | |
| 742 | #[test] |
| 743 | fn malformed_schema_falls_back_to_empty_object() { |
| 744 | let tools = vec![ToolSpec { |
| 745 | name: "x".to_string(), |
| 746 | description: String::new(), |
| 747 | parameters_schema: "not json".to_string(), |
| 748 | }]; |
| 749 | let json = OpenAiBackend::tools_json(&tools); |
| 750 | assert_eq!(json[0]["function"]["parameters"]["type"], "object"); |
| 751 | } |
| 752 | |
| 753 | #[test] |
| 754 | fn streamed_assistant_history_omits_empty_tool_calls() { |
| 755 | let value = streamed_assistant_history("hello", &[]); |
| 756 | assert_eq!(value["role"], "assistant"); |
| 757 | assert_eq!(value["content"], "hello"); |
| 758 | assert!(value.get("tool_calls").is_none()); |
| 759 | } |
| 760 | |
| 761 | #[test] |
| 762 | fn streamed_assistant_history_preserves_tool_calls() { |
| 763 | let calls = vec![ToolCall { |
| 764 | id: "call_0".to_string(), |
| 765 | name: "read_file".to_string(), |
| 766 | arguments: r#"{"path":"a.rs"}"#.to_string(), |
| 767 | }]; |
| 768 | let value = streamed_assistant_history("", &calls); |
| 769 | assert!(value["content"].is_null()); |
| 770 | assert_eq!(value["tool_calls"][0]["id"], "call_0"); |
| 771 | assert_eq!(value["tool_calls"][0]["type"], "function"); |
| 772 | assert_eq!(value["tool_calls"][0]["function"]["name"], "read_file"); |
| 773 | assert_eq!( |
| 774 | value["tool_calls"][0]["function"]["arguments"], |
| 775 | r#"{"path":"a.rs"}"# |
| 776 | ); |
| 777 | } |
| 778 | |
| 779 | #[test] |
| 780 | fn assistant_message_with_tool_calls_round_trips() { |
| 781 | let message = ResponseMessage { |
| 782 | content: None, |
| 783 | tool_calls: Some(vec![ResponseToolCall { |
| 784 | id: "call_1".to_string(), |
| 785 | function: ResponseFunction { |
| 786 | name: "read_file".to_string(), |
| 787 | arguments: r#"{"path":"a.rs"}"#.to_string(), |
| 788 | }, |
| 789 | }]), |
| 790 | }; |
| 791 | let value = message.into_history_value(); |
| 792 | assert_eq!(value["role"], "assistant"); |
| 793 | assert!(value["content"].is_null()); |
| 794 | assert_eq!(value["tool_calls"][0]["id"], "call_1"); |
| 795 | assert_eq!(value["tool_calls"][0]["type"], "function"); |
| 796 | assert_eq!(value["tool_calls"][0]["function"]["name"], "read_file"); |
| 797 | } |
| 798 | } |