claude/linked-issue-acp-tui-u4vvpa
claude/github-issue-40-0au5ua
claude/linked-issue-acp-tui-u4vvpa
claude/minor-release-cut-vkugow
claude/sigit-acp-local-chat-cx6380
claude/sigit-cloud-agent-expansion-reox0a
claude/tool-permission-system
claude/zen-feynman-0u78dk
development
feature/acp-cancel-request
feature/acp-session-list-cursor
feature/acp-v1-conformance
feature/agent-tools-multiedit-glob-todos-remember
feature/cargo-workspace
feature/hooks
feature/load-local-model-explicitly
feature/sigit-agent-profile
feature/sigit-code-cloud
feature/xcode-mcp-server
fix/acp-replay-loaded-session-history
fix/acp-replay-reasoning-and-attachments
fix/dsml-doubled-bar-markers
main
release/v1.5.10
release/v1.5.6
release/v1.6.1
| 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 seam is consumed by both surfaces: the interactive client (`#[cfg(unix)]`, |
| 15 | //! see `run_interactive` in `main.rs` and `mod tui` in `chat.rs`) and the ACP |
| 16 | //! server's prompt loop. Some items are still reached only through the |
| 17 | //! Unix-only interactive paths, so the dead-code lint stays suppressed on |
| 18 | //! non-Unix targets only — Unix builds keep 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, ChatMessage, ChatRole, EngineStatus, 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 Code already declares |
| 32 | /// tools). |
| 33 | #[derive(Debug, Clone)] |
| 34 | pub struct ToolSpec { |
| 35 | pub name: String, |
| 36 | pub description: String, |
| 37 | pub parameters_schema: String, |
| 38 | } |
| 39 | |
| 40 | /// A tool call requested by the model. |
| 41 | #[derive(Debug, Clone)] |
| 42 | pub struct ToolCall { |
| 43 | pub id: String, |
| 44 | pub name: String, |
| 45 | /// Arguments as a JSON-encoded string. |
| 46 | pub arguments: String, |
| 47 | } |
| 48 | |
| 49 | /// The output of executing one tool call, fed back to the model. |
| 50 | #[derive(Debug, Clone)] |
| 51 | pub struct ToolResult { |
| 52 | pub tool_call_id: String, |
| 53 | pub content: String, |
| 54 | } |
| 55 | |
| 56 | /// The result of one assistant turn: free text and/or tool calls. |
| 57 | #[derive(Debug, Clone, Default)] |
| 58 | pub struct TurnResult { |
| 59 | pub text: String, |
| 60 | pub tool_calls: Vec<ToolCall>, |
| 61 | } |
| 62 | |
| 63 | /// Backend errors are plain strings. Callers map them to ACP errors. |
| 64 | pub type BackendError = String; |
| 65 | |
| 66 | /// Rough context budget for a conversation, in estimated tokens (see |
| 67 | /// [`estimate_tokens`]). When a snapshot exceeds this, the agent loops compact |
| 68 | /// history before the next tool round. |
| 69 | pub const DEFAULT_CONTEXT_TOKEN_BUDGET: usize = 24_000; |
| 70 | |
| 71 | /// How many trailing messages survive a compaction verbatim (the rest are |
| 72 | /// folded into the summary). |
| 73 | pub const COMPACT_KEEP_LAST: usize = 6; |
| 74 | |
| 75 | /// The summarization request sent to the model when compacting history. |
| 76 | const SUMMARIZE_PROMPT: &str = "Summarize this coding session so far: decisions made, \ |
| 77 | files touched, current state, open items. Be concise and factual."; |
| 78 | |
| 79 | /// Render a history snapshot as a plain-text transcript, with tool calls and |
| 80 | /// tool results spelled out as prose rather than left in their wire shapes. |
| 81 | /// |
| 82 | /// Compaction summarizes a conversation that is, by definition, thick with |
| 83 | /// `tool_calls` and `role: "tool"` messages — but the summarization round asks |
| 84 | /// for a plain answer and so sends no `tools` array. Forwarding the raw shapes |
| 85 | /// in that request produces tool blocks with no schema to validate against, |
| 86 | /// which strict endpoints reject outright: Anthropic answers 400, so every |
| 87 | /// compaction of a session that had ever run a tool failed, permanently, no |
| 88 | /// matter how small the history was. Flattening to text keeps everything the |
| 89 | /// summary actually needs and drops the shapes that only make sense alongside |
| 90 | /// a tool schema. It also sidesteps orphaned `tool_call_id`s and role- |
| 91 | /// alternation rules, neither of which a transcript can violate. |
| 92 | fn transcript_for_summary(history: &[serde_json::Value]) -> String { |
| 93 | let mut lines: Vec<String> = Vec::new(); |
| 94 | for message in history { |
| 95 | let role = message["role"].as_str().unwrap_or("user"); |
| 96 | // The system prompt is carried over verbatim, so it needn't be summarized. |
| 97 | if role == "system" { |
| 98 | continue; |
| 99 | } |
| 100 | |
| 101 | let mut parts: Vec<String> = Vec::new(); |
| 102 | if let Some(text) = message["content"].as_str() |
| 103 | && !text.trim().is_empty() |
| 104 | { |
| 105 | parts.push(text.to_string()); |
| 106 | } |
| 107 | for call in message["tool_calls"].as_array().into_iter().flatten() { |
| 108 | parts.push(format!( |
| 109 | "called {}({})", |
| 110 | call["function"]["name"].as_str().unwrap_or("tool"), |
| 111 | call["function"]["arguments"].as_str().unwrap_or_default(), |
| 112 | )); |
| 113 | } |
| 114 | if parts.is_empty() { |
| 115 | continue; |
| 116 | } |
| 117 | |
| 118 | let label = if role == "tool" { "tool result" } else { role }; |
| 119 | lines.push(format!("{label}: {}", parts.join("\n"))); |
| 120 | } |
| 121 | lines.join("\n\n") |
| 122 | } |
| 123 | |
| 124 | /// Crude token estimate for a history snapshot: serialized characters / 4. |
| 125 | /// Deliberately model-agnostic — it only needs to be in the right ballpark to |
| 126 | /// decide when compaction is worth an extra inference round. |
| 127 | pub fn estimate_tokens(history: &[serde_json::Value]) -> usize { |
| 128 | let chars: usize = history |
| 129 | .iter() |
| 130 | .map(|message| message.to_string().chars().count()) |
| 131 | .sum(); |
| 132 | chars / 4 |
| 133 | } |
| 134 | |
| 135 | /// A sink for streaming assistant text deltas to the UI as they are produced. |
| 136 | /// |
| 137 | /// When a caller passes `Some(sink)`, a streaming-capable backend forwards each |
| 138 | /// text fragment through it as the model emits it; the returned [`TurnResult`] |
| 139 | /// still carries the fully assembled text (and any tool calls). When the sink is |
| 140 | /// `None`, the backend runs in non-streaming mode. Unbounded so the inference |
| 141 | /// task never blocks on a slow consumer. |
| 142 | pub type TokenSink = tokio::sync::mpsc::UnboundedSender<String>; |
| 143 | |
| 144 | // ── The trait ─────────────────────────────────────────────────────────────────── |
| 145 | |
| 146 | /// A swappable inference backend driving siGit Code's agent loop. |
| 147 | #[async_trait] |
| 148 | pub trait InferenceBackend: Send + Sync { |
| 149 | /// Start an assistant turn from a new user message, offering `tools`. |
| 150 | /// |
| 151 | /// If `sink` is `Some`, text is streamed through it as it is generated. A |
| 152 | /// backend may decline to stream a given round (for example, on-device |
| 153 | /// inference cannot stream while it is still deciding whether to call a |
| 154 | /// tool); in that case the text is delivered only via the returned result. |
| 155 | async fn send_message_with_tools( |
| 156 | &self, |
| 157 | text: &str, |
| 158 | tools: &[ToolSpec], |
| 159 | sink: Option<&TokenSink>, |
| 160 | ) -> Result<TurnResult, BackendError>; |
| 161 | |
| 162 | /// Continue the turn by returning tool results. `allow_tool_calls` controls |
| 163 | /// whether `tools` are offered again; the complete catalog remains available |
| 164 | /// so a disabled round can recognize and suppress tool-shaped model output. |
| 165 | /// `sink` streams assistant text when set. |
| 166 | async fn send_tool_results( |
| 167 | &self, |
| 168 | results: Vec<ToolResult>, |
| 169 | tools: &[ToolSpec], |
| 170 | allow_tool_calls: bool, |
| 171 | sink: Option<&TokenSink>, |
| 172 | ) -> Result<TurnResult, BackendError>; |
| 173 | |
| 174 | /// Record tool results in the conversation history *without* asking the |
| 175 | /// model to continue the turn. Used when a turn is abandoned mid-round |
| 176 | /// (the user cancelled at the permission gate): by then the assistant |
| 177 | /// message carrying the tool calls is already in history, and leaving them |
| 178 | /// unanswered makes strict OpenAI-compatible endpoints reject every later |
| 179 | /// request in the session. |
| 180 | async fn record_cancelled_tool_results(&self, results: Vec<ToolResult>); |
| 181 | |
| 182 | /// Whether inference runs over the network (a configured provider) rather |
| 183 | /// than on-device. Drives UI labelling so the displayed model can't claim a |
| 184 | /// local model while requests actually go to the cloud. |
| 185 | fn is_remote(&self) -> bool; |
| 186 | |
| 187 | /// A serializable snapshot of the conversation history, one JSON object per |
| 188 | /// message (`{"role": ..., "content": ...}` at minimum). The snapshot is |
| 189 | /// what the session store persists; it includes any seeded system message |
| 190 | /// so [`InferenceBackend::restore_history`] can replace state wholesale. |
| 191 | async fn history_snapshot(&self) -> Vec<serde_json::Value>; |
| 192 | |
| 193 | /// Replace the conversation history with a previously saved snapshot. |
| 194 | /// Backends that cannot represent every entry (e.g. on-device history has |
| 195 | /// no tool-call structure) flatten what they can and drop the rest. |
| 196 | async fn restore_history(&self, history: Vec<serde_json::Value>); |
| 197 | |
| 198 | /// Shrink the conversation history: summarize everything so far with one |
| 199 | /// extra (non-streaming) inference round, then rebuild history as |
| 200 | /// `[system message, summary, last keep_last non-system messages]`. On |
| 201 | /// error the original history is left in place. |
| 202 | async fn compact_history(&self, keep_last: usize) -> Result<(), BackendError>; |
| 203 | } |
| 204 | |
| 205 | // ── Local backend (onde ChatEngine) ────────────────────────────────────────────── |
| 206 | |
| 207 | /// On-device inference. A thin adapter over `onde::ChatEngine`. |
| 208 | pub struct LocalBackend { |
| 209 | engine: Arc<ChatEngine>, |
| 210 | /// History restored before an on-device model is loaded. Onde ignores |
| 211 | /// `push_history` while unloaded, so keep the snapshot here until the |
| 212 | /// engine can accept it instead of silently dropping a loaded session. |
| 213 | pending_history: Mutex<Option<Vec<serde_json::Value>>>, |
| 214 | } |
| 215 | |
| 216 | impl LocalBackend { |
| 217 | pub fn new(engine: Arc<ChatEngine>) -> Self { |
| 218 | Self { |
| 219 | engine, |
| 220 | pending_history: Mutex::new(None), |
| 221 | } |
| 222 | } |
| 223 | |
| 224 | async fn apply_pending_history(&self) -> Result<(), BackendError> { |
| 225 | if self.pending_history.lock().await.is_none() { |
| 226 | return Ok(()); |
| 227 | } |
| 228 | let status = self.engine.info().await.status; |
| 229 | if status != EngineStatus::Ready { |
| 230 | return Err(format!( |
| 231 | "on-device model is not ready to restore session history (status: {status})" |
| 232 | )); |
| 233 | } |
| 234 | let Some(history) = self.pending_history.lock().await.take() else { |
| 235 | return Ok(()); |
| 236 | }; |
| 237 | |
| 238 | self.engine.clear_history().await; |
| 239 | for entry in history { |
| 240 | let role = entry["role"].as_str().unwrap_or(""); |
| 241 | let content = entry["content"].as_str().unwrap_or("").to_string(); |
| 242 | // Tool-call-only assistant entries and empty tool results carry no |
| 243 | // text a plain chat history can replay; drop them. |
| 244 | if content.is_empty() && role != "user" && role != "system" { |
| 245 | continue; |
| 246 | } |
| 247 | let message = match role { |
| 248 | "system" => ChatMessage::system(content), |
| 249 | "user" => ChatMessage::user(content), |
| 250 | "assistant" => ChatMessage::assistant(content), |
| 251 | // Tool results flatten to plain text (MVP; acceptable loss). |
| 252 | "tool" => ChatMessage::user(format!("[tool result]\n{content}")), |
| 253 | _ => continue, |
| 254 | }; |
| 255 | self.engine.push_history(message).await; |
| 256 | } |
| 257 | Ok(()) |
| 258 | } |
| 259 | } |
| 260 | |
| 261 | fn to_onde_tools(tools: &[ToolSpec]) -> Vec<ToolDefinition> { |
| 262 | tools |
| 263 | .iter() |
| 264 | .map(|tool| ToolDefinition { |
| 265 | name: tool.name.clone(), |
| 266 | description: tool.description.clone(), |
| 267 | parameters_schema: tool.parameters_schema.clone(), |
| 268 | }) |
| 269 | .collect() |
| 270 | } |
| 271 | |
| 272 | #[async_trait] |
| 273 | impl InferenceBackend for LocalBackend { |
| 274 | async fn send_message_with_tools( |
| 275 | &self, |
| 276 | text: &str, |
| 277 | tools: &[ToolSpec], |
| 278 | sink: Option<&TokenSink>, |
| 279 | ) -> Result<TurnResult, BackendError> { |
| 280 | self.apply_pending_history().await?; |
| 281 | // onde's tool-aware path is non-streaming: it has to buffer the whole |
| 282 | // reply to detect tool calls. We can only stream when no tools are on |
| 283 | // offer (a plain answer), which is exactly the tools-disabled case. |
| 284 | if let Some(sink) = sink |
| 285 | && tools.is_empty() |
| 286 | { |
| 287 | let rx = self |
| 288 | .engine |
| 289 | .stream_message(text) |
| 290 | .await |
| 291 | .map_err(|error| error.to_string())?; |
| 292 | return drain_onde_stream(rx, sink).await; |
| 293 | } |
| 294 | |
| 295 | let onde_tools = to_onde_tools(tools); |
| 296 | let result = self |
| 297 | .engine |
| 298 | .send_message_with_tools(text, &onde_tools) |
| 299 | .await |
| 300 | .map_err(|error| error.to_string())?; |
| 301 | Ok(onde_result_to_turn(result)) |
| 302 | } |
| 303 | |
| 304 | async fn send_tool_results( |
| 305 | &self, |
| 306 | results: Vec<ToolResult>, |
| 307 | tools: &[ToolSpec], |
| 308 | allow_tool_calls: bool, |
| 309 | sink: Option<&TokenSink>, |
| 310 | ) -> Result<TurnResult, BackendError> { |
| 311 | self.apply_pending_history().await?; |
| 312 | let onde_results: Vec<onde::inference::ToolResult> = results |
| 313 | .into_iter() |
| 314 | .map(|result| onde::inference::ToolResult { |
| 315 | tool_call_id: result.tool_call_id, |
| 316 | content: result.content, |
| 317 | }) |
| 318 | .collect(); |
| 319 | |
| 320 | // A forced-text round is the only round onde can stream, since no |
| 321 | // further tool calls are parsed. |
| 322 | if let Some(sink) = sink |
| 323 | && !allow_tool_calls |
| 324 | { |
| 325 | let rx = self |
| 326 | .engine |
| 327 | .stream_tool_results(onde_results, None) |
| 328 | .await |
| 329 | .map_err(|error| error.to_string())?; |
| 330 | return drain_onde_stream(rx, sink).await; |
| 331 | } |
| 332 | |
| 333 | let onde_tools = allow_tool_calls.then(|| to_onde_tools(tools)); |
| 334 | let result = self |
| 335 | .engine |
| 336 | .send_tool_results(onde_results, onde_tools.as_deref()) |
| 337 | .await |
| 338 | .map_err(|error| error.to_string())?; |
| 339 | Ok(onde_result_to_turn(result)) |
| 340 | } |
| 341 | |
| 342 | async fn record_cancelled_tool_results(&self, _results: Vec<ToolResult>) { |
| 343 | // onde's public API cannot append tool-result history entries without |
| 344 | // running another inference round, so the dangling tool call stays in |
| 345 | // its history. The chat template replays it as-is, which local models |
| 346 | // tolerate — worst case the model re-issues the call next turn. |
| 347 | } |
| 348 | |
| 349 | fn is_remote(&self) -> bool { |
| 350 | false |
| 351 | } |
| 352 | |
| 353 | async fn history_snapshot(&self) -> Vec<serde_json::Value> { |
| 354 | if let Some(history) = self.pending_history.lock().await.as_ref() { |
| 355 | return history.clone(); |
| 356 | } |
| 357 | // onde's `history()` already flattens tool entries: assistant tool |
| 358 | // calls become plain assistant text and tool results are omitted, so |
| 359 | // the snapshot is lossy for tool-heavy turns (acceptable in this MVP). |
| 360 | self.engine |
| 361 | .history() |
| 362 | .await |
| 363 | .iter() |
| 364 | .map(|message| { |
| 365 | serde_json::json!({ |
| 366 | "role": message.role.to_string(), |
| 367 | "content": message.content, |
| 368 | }) |
| 369 | }) |
| 370 | .collect() |
| 371 | } |
| 372 | |
| 373 | async fn restore_history(&self, history: Vec<serde_json::Value>) { |
| 374 | *self.pending_history.lock().await = Some(history); |
| 375 | if self.engine.info().await.status == EngineStatus::Ready { |
| 376 | // Ready was just observed, so failure here can only mean a status |
| 377 | // transition; keep the pending copy for the next inference call. |
| 378 | let _ = self.apply_pending_history().await; |
| 379 | } |
| 380 | } |
| 381 | |
| 382 | async fn compact_history(&self, keep_last: usize) -> Result<(), BackendError> { |
| 383 | self.apply_pending_history().await?; |
| 384 | let snapshot = self.engine.history().await; |
| 385 | // One plain (tool-free) inference round produces the summary. On error |
| 386 | // history is untouched — send_message only mutates it on success, and |
| 387 | // whatever it appended is wiped by the clear below anyway. |
| 388 | let result = self |
| 389 | .engine |
| 390 | .send_message(SUMMARIZE_PROMPT) |
| 391 | .await |
| 392 | .map_err(|error| error.to_string())?; |
| 393 | // Local models may reason in <think> blocks; keep only the visible part. |
| 394 | let (_think, summary) = crate::chat::strip_think_blocks(&result.text); |
| 395 | |
| 396 | self.engine.clear_history().await; |
| 397 | // Leading system messages carry the session context; keep them all. |
| 398 | for message in snapshot |
| 399 | .iter() |
| 400 | .take_while(|message| message.role == ChatRole::System) |
| 401 | { |
| 402 | self.engine.push_history(message.clone()).await; |
| 403 | } |
| 404 | self.engine |
| 405 | .push_history(ChatMessage::user(format!( |
| 406 | "[Conversation summary]\n{summary}" |
| 407 | ))) |
| 408 | .await; |
| 409 | let non_system: Vec<&ChatMessage> = snapshot |
| 410 | .iter() |
| 411 | .filter(|message| message.role != ChatRole::System) |
| 412 | .collect(); |
| 413 | let tail_start = non_system.len().saturating_sub(keep_last); |
| 414 | for message in &non_system[tail_start..] { |
| 415 | self.engine.push_history((*message).clone()).await; |
| 416 | } |
| 417 | Ok(()) |
| 418 | } |
| 419 | } |
| 420 | |
| 421 | /// Drain an onde streaming receiver, forwarding each token to `sink` and |
| 422 | /// assembling the full text. onde reports stream failures as a final chunk whose |
| 423 | /// `finish_reason` is `"error: …"`; surface those as a backend error. |
| 424 | async fn drain_onde_stream( |
| 425 | mut rx: tokio::sync::mpsc::Receiver<onde::inference::StreamChunk>, |
| 426 | sink: &TokenSink, |
| 427 | ) -> Result<TurnResult, BackendError> { |
| 428 | let mut text = String::new(); |
| 429 | while let Some(chunk) = rx.recv().await { |
| 430 | if !chunk.delta.is_empty() { |
| 431 | text.push_str(&chunk.delta); |
| 432 | // The receiver is the UI; if it's gone the turn is being cancelled, |
| 433 | // so stop assembling rather than spinning the model to completion. |
| 434 | if sink.send(chunk.delta).is_err() { |
| 435 | break; |
| 436 | } |
| 437 | } |
| 438 | if chunk.done { |
| 439 | if let Some(reason) = chunk.finish_reason |
| 440 | && let Some(message) = reason.strip_prefix("error: ") |
| 441 | { |
| 442 | return Err(message.to_string()); |
| 443 | } |
| 444 | break; |
| 445 | } |
| 446 | } |
| 447 | Ok(TurnResult { |
| 448 | text, |
| 449 | tool_calls: Vec::new(), |
| 450 | }) |
| 451 | } |
| 452 | |
| 453 | /// Convert an `onde` tool-aware result into the neutral [`TurnResult`]. |
| 454 | fn onde_result_to_turn(result: onde::inference::ToolAwareResult) -> TurnResult { |
| 455 | TurnResult { |
| 456 | text: result.text, |
| 457 | tool_calls: result |
| 458 | .tool_calls |
| 459 | .into_iter() |
| 460 | .map(|call| ToolCall { |
| 461 | id: call.id, |
| 462 | name: call.function_name, |
| 463 | arguments: call.arguments, |
| 464 | }) |
| 465 | .collect(), |
| 466 | } |
| 467 | } |
| 468 | |
| 469 | // ── OpenAI-compatible backend ───────────────────────────────────────────────────── |
| 470 | |
| 471 | /// Inference against any OpenAI-compatible Chat Completions endpoint. |
| 472 | /// |
| 473 | /// Conversation state is held client-side and replayed on every request, so the |
| 474 | /// endpoint can be stateless. Standard OpenAI function-calling is used end to |
| 475 | /// end (`tools`, `choices[].message.tool_calls`, `role: "tool"` follow-ups). |
| 476 | pub struct OpenAiBackend { |
| 477 | base_url: String, |
| 478 | api_key: String, |
| 479 | model: String, |
| 480 | http: reqwest::Client, |
| 481 | /// The full message list sent on each request (system + turns + tool results). |
| 482 | history: Mutex<Vec<serde_json::Value>>, |
| 483 | } |
| 484 | |
| 485 | impl OpenAiBackend { |
| 486 | /// Build a backend for `{base_url, api_key, model}`, seeding the optional |
| 487 | /// system prompt. `base_url` should include the API root (e.g. ending in |
| 488 | /// `/v1`); the chat path is appended. |
| 489 | pub fn new( |
| 490 | base_url: impl Into<String>, |
| 491 | api_key: impl Into<String>, |
| 492 | model: impl Into<String>, |
| 493 | system_prompt: Option<String>, |
| 494 | ) -> Self { |
| 495 | let mut history = Vec::new(); |
| 496 | if let Some(prompt) = system_prompt { |
| 497 | history.push(serde_json::json!({ "role": "system", "content": prompt })); |
| 498 | } |
| 499 | Self { |
| 500 | base_url: base_url.into(), |
| 501 | api_key: api_key.into(), |
| 502 | model: model.into(), |
| 503 | http: reqwest::Client::new(), |
| 504 | history: Mutex::new(history), |
| 505 | } |
| 506 | } |
| 507 | |
| 508 | fn tools_json(tools: &[ToolSpec]) -> Vec<serde_json::Value> { |
| 509 | tools |
| 510 | .iter() |
| 511 | .map(|tool| { |
| 512 | // parameters_schema is a JSON string; parse it, defaulting to an |
| 513 | // empty object schema if malformed. |
| 514 | let parameters: serde_json::Value = serde_json::from_str(&tool.parameters_schema) |
| 515 | .unwrap_or_else(|_| serde_json::json!({ "type": "object", "properties": {} })); |
| 516 | serde_json::json!({ |
| 517 | "type": "function", |
| 518 | "function": { |
| 519 | "name": tool.name, |
| 520 | "description": tool.description, |
| 521 | "parameters": parameters, |
| 522 | } |
| 523 | }) |
| 524 | }) |
| 525 | .collect() |
| 526 | } |
| 527 | |
| 528 | /// POST the current history and apply the assistant reply to history. |
| 529 | /// `tools` is always the known catalog, while `allow_tool_calls` determines |
| 530 | /// whether it is advertised to the model and whether returned calls may run. |
| 531 | /// Streams via SSE when `sink` is set; otherwise reads a single JSON response. |
| 532 | /// |
| 533 | /// A reply whose only attempt at a tool call was a block that couldn't be |
| 534 | /// parsed gets one retry: the model is told the call didn't run and asked |
| 535 | /// to make it again through the structured interface. Without that the |
| 536 | /// turn ends on whatever prose preceded the broken block, and the call is |
| 537 | /// silently lost. |
| 538 | async fn complete( |
| 539 | &self, |
| 540 | tools: &[ToolSpec], |
| 541 | allow_tool_calls: bool, |
| 542 | sink: Option<&TokenSink>, |
| 543 | ) -> Result<TurnResult, BackendError> { |
| 544 | let (result, malformed) = self.request(tools, allow_tool_calls, sink).await?; |
| 545 | if malformed == 0 || !allow_tool_calls || !result.tool_calls.is_empty() { |
| 546 | return Ok(result); |
| 547 | } |
| 548 | |
| 549 | log::warn!( |
| 550 | "dropped {malformed} unparseable inline tool call(s); asking the model to retry" |
| 551 | ); |
| 552 | self.history.lock().await.push(serde_json::json!({ |
| 553 | "role": "user", |
| 554 | "content": MALFORMED_TOOL_CALL_RETRY, |
| 555 | })); |
| 556 | if let Some(sink) = sink |
| 557 | && !result.text.trim().is_empty() |
| 558 | { |
| 559 | // The retry's text follows what already streamed; keep it from |
| 560 | // running on into the previous sentence. |
| 561 | let _ = sink.send("\n\n".to_string()); |
| 562 | } |
| 563 | let (retry, retry_malformed) = self.request(tools, allow_tool_calls, sink).await?; |
| 564 | if retry_malformed > 0 && retry.tool_calls.is_empty() { |
| 565 | log::warn!( |
| 566 | "the retry also wrote {retry_malformed} unparseable inline tool call(s); \ |
| 567 | ending the turn without them" |
| 568 | ); |
| 569 | } |
| 570 | Ok(TurnResult { |
| 571 | text: join_reply_text(&result.text, &retry.text), |
| 572 | tool_calls: retry.tool_calls, |
| 573 | }) |
| 574 | } |
| 575 | |
| 576 | /// One chat-completion round trip. Returns the turn plus how many inline |
| 577 | /// tool-call blocks were dropped because they couldn't be parsed. |
| 578 | async fn request( |
| 579 | &self, |
| 580 | tools: &[ToolSpec], |
| 581 | allow_tool_calls: bool, |
| 582 | sink: Option<&TokenSink>, |
| 583 | ) -> Result<(TurnResult, usize), BackendError> { |
| 584 | let url = format!("{}/chat/completions", self.base_url.trim_end_matches('/')); |
| 585 | let streaming = sink.is_some(); |
| 586 | |
| 587 | let mut body = serde_json::json!({ |
| 588 | "model": self.model, |
| 589 | "messages": *self.history.lock().await, |
| 590 | "stream": streaming, |
| 591 | }); |
| 592 | if allow_tool_calls && !tools.is_empty() { |
| 593 | body["tools"] = serde_json::Value::Array(Self::tools_json(tools)); |
| 594 | // OpenAI specifies `auto` as the default when tools are present, |
| 595 | // but not every OpenAI-compatible gateway implements that default. |
| 596 | // Sending it explicitly is important for the cloud path: otherwise |
| 597 | // a model can describe the action it intends to take as text and |
| 598 | // end the turn without returning a tool call at all. |
| 599 | body["tool_choice"] = serde_json::Value::String("auto".to_string()); |
| 600 | } |
| 601 | |
| 602 | let response = self |
| 603 | .http |
| 604 | .post(&url) |
| 605 | .bearer_auth(&self.api_key) |
| 606 | .json(&body) |
| 607 | .send() |
| 608 | .await |
| 609 | .map_err(|error| format!("request to {url} failed: {error}"))?; |
| 610 | |
| 611 | if !response.status().is_success() { |
| 612 | let status = response.status(); |
| 613 | let body = response.text().await.unwrap_or_default(); |
| 614 | return Err(describe_api_error(status, &body)); |
| 615 | } |
| 616 | |
| 617 | if let Some(sink) = sink { |
| 618 | self.consume_stream(response, sink, tools, allow_tool_calls) |
| 619 | .await |
| 620 | } else { |
| 621 | self.consume_json(response, tools, allow_tool_calls).await |
| 622 | } |
| 623 | } |
| 624 | |
| 625 | /// Parse a single non-streaming chat-completion response. |
| 626 | async fn consume_json( |
| 627 | &self, |
| 628 | response: reqwest::Response, |
| 629 | tools: &[ToolSpec], |
| 630 | allow_tool_calls: bool, |
| 631 | ) -> Result<(TurnResult, usize), BackendError> { |
| 632 | let parsed: ChatCompletion = response |
| 633 | .json() |
| 634 | .await |
| 635 | .map_err(|error| format!("response parse error: {error}"))?; |
| 636 | |
| 637 | let message = parsed |
| 638 | .choices |
| 639 | .into_iter() |
| 640 | .next() |
| 641 | .map(|choice| choice.message) |
| 642 | .ok_or_else(|| "endpoint returned no choices".to_string())?; |
| 643 | |
| 644 | let mut text = message.content.clone().unwrap_or_default(); |
| 645 | let mut tool_calls: Vec<ToolCall> = message |
| 646 | .tool_calls |
| 647 | .iter() |
| 648 | .flatten() |
| 649 | .map(|call| ToolCall { |
| 650 | id: call.id.clone(), |
| 651 | name: call.function.name.clone(), |
| 652 | arguments: call.function.arguments.clone(), |
| 653 | }) |
| 654 | .collect(); |
| 655 | |
| 656 | let extracted = crate::inline_tool_calls::extract(&text, tools); |
| 657 | let malformed = extracted.malformed; |
| 658 | if malformed > 0 { |
| 659 | log::warn!("dropped {malformed} unparseable inline tool call(s) from the reply"); |
| 660 | } |
| 661 | |
| 662 | if !allow_tool_calls { |
| 663 | let recovered = extracted.calls; |
| 664 | text = extracted.text; |
| 665 | let suppressed = tool_calls.len() + recovered.len(); |
| 666 | if suppressed > 0 { |
| 667 | let names = tool_calls |
| 668 | .iter() |
| 669 | .map(|call| call.name.as_str()) |
| 670 | .chain(recovered.iter().map(|call| call.name.as_str())) |
| 671 | .collect::<Vec<_>>() |
| 672 | .join(", "); |
| 673 | log::warn!( |
| 674 | "suppressed {suppressed} tool call(s) from a forced-text response: {names}" |
| 675 | ); |
| 676 | } |
| 677 | self.history |
| 678 | .lock() |
| 679 | .await |
| 680 | .push(streamed_assistant_history(&text, &[])); |
| 681 | return Ok(( |
| 682 | TurnResult { |
| 683 | text, |
| 684 | tool_calls: Vec::new(), |
| 685 | }, |
| 686 | malformed, |
| 687 | )); |
| 688 | } |
| 689 | |
| 690 | // Some models write a tool call out as literal `<tool_call>` text |
| 691 | // instead of using the structured field (see `inline_tool_calls`). |
| 692 | // Recover it, or the turn ends with the tag rendered as prose and |
| 693 | // whatever the model meant to do is dropped. |
| 694 | if tool_calls.is_empty() { |
| 695 | let cleaned = extracted.text; |
| 696 | let recovered = extracted.calls; |
| 697 | if !recovered.is_empty() { |
| 698 | log::warn!( |
| 699 | "recovered {} tool call(s) the model emitted as text instead of a structured call", |
| 700 | recovered.len() |
| 701 | ); |
| 702 | let tool_calls: Vec<ToolCall> = recovered |
| 703 | .into_iter() |
| 704 | .enumerate() |
| 705 | .map(|(index, call)| ToolCall { |
| 706 | id: format!("call_recovered_{index}"), |
| 707 | name: call.name, |
| 708 | arguments: call.arguments, |
| 709 | }) |
| 710 | .collect(); |
| 711 | // Record the recovered shape, not the raw tag: the tool |
| 712 | // results that follow have to answer an assistant message |
| 713 | // that actually carries these calls, or the next request is |
| 714 | // rejected for orphaned tool results. |
| 715 | self.history |
| 716 | .lock() |
| 717 | .await |
| 718 | .push(streamed_assistant_history(&cleaned, &tool_calls)); |
| 719 | return Ok(( |
| 720 | TurnResult { |
| 721 | text: cleaned, |
| 722 | tool_calls, |
| 723 | }, |
| 724 | malformed, |
| 725 | )); |
| 726 | } |
| 727 | if malformed > 0 { |
| 728 | // Keep the broken block out of history as well as the reply; |
| 729 | // see `inline_tool_calls` for why it poisons later turns. |
| 730 | self.push_reply_without_calls(&cleaned).await; |
| 731 | return Ok(( |
| 732 | TurnResult { |
| 733 | text: cleaned, |
| 734 | tool_calls, |
| 735 | }, |
| 736 | malformed, |
| 737 | )); |
| 738 | } |
| 739 | } else if malformed > 0 || !extracted.calls.is_empty() { |
| 740 | // Structured calls arrived with inline blocks beside them. As on |
| 741 | // the streaming path, inline calls that parse run alongside the |
| 742 | // structured ones, broken blocks are dropped, and neither kind of |
| 743 | // block is left in the reply or the history. |
| 744 | let offset = tool_calls.len(); |
| 745 | tool_calls.extend( |
| 746 | extracted |
| 747 | .calls |
| 748 | .into_iter() |
| 749 | .enumerate() |
| 750 | .map(|(index, call)| ToolCall { |
| 751 | id: format!("call_recovered_{}", offset + index), |
| 752 | name: call.name, |
| 753 | arguments: call.arguments, |
| 754 | }), |
| 755 | ); |
| 756 | self.history |
| 757 | .lock() |
| 758 | .await |
| 759 | .push(streamed_assistant_history(&extracted.text, &tool_calls)); |
| 760 | return Ok(( |
| 761 | TurnResult { |
| 762 | text: extracted.text, |
| 763 | tool_calls, |
| 764 | }, |
| 765 | malformed, |
| 766 | )); |
| 767 | } |
| 768 | |
| 769 | // Record the assistant turn so later tool results have context. |
| 770 | self.history.lock().await.push(message.into_history_value()); |
| 771 | |
| 772 | Ok((TurnResult { text, tool_calls }, 0)) |
| 773 | } |
| 774 | |
| 775 | /// Record a text-only assistant reply. An empty one is skipped: strict |
| 776 | /// endpoints reject an assistant message with neither content nor tool |
| 777 | /// calls, and there is nothing in it worth replaying. |
| 778 | async fn push_reply_without_calls(&self, text: &str) { |
| 779 | if !text.is_empty() { |
| 780 | self.history |
| 781 | .lock() |
| 782 | .await |
| 783 | .push(streamed_assistant_history(text, &[])); |
| 784 | } |
| 785 | } |
| 786 | |
| 787 | /// Consume an OpenAI Server-Sent Events stream, forwarding content deltas to |
| 788 | /// `sink` and reassembling any tool calls (which arrive fragmented across |
| 789 | /// chunks, keyed by `index`). |
| 790 | async fn consume_stream( |
| 791 | &self, |
| 792 | response: reqwest::Response, |
| 793 | sink: &TokenSink, |
| 794 | tools: &[ToolSpec], |
| 795 | allow_tool_calls: bool, |
| 796 | ) -> Result<(TurnResult, usize), BackendError> { |
| 797 | use futures::StreamExt; |
| 798 | |
| 799 | let mut stream = response.bytes_stream(); |
| 800 | // Newlines are ASCII, so splitting raw bytes on `\n` never bisects a |
| 801 | // multibyte UTF-8 sequence; we only lossily decode whole lines. |
| 802 | let mut buffer: Vec<u8> = Vec::new(); |
| 803 | let mut text = String::new(); |
| 804 | let mut tool_accum: Vec<StreamingToolCall> = Vec::new(); |
| 805 | let mut done = false; |
| 806 | // Recovers a tool call the model wrote as literal `<tool_call>` text |
| 807 | // instead of a structured delta. Scanning here (rather than after the |
| 808 | // stream) keeps the tag off the UI: content goes straight to `sink` as |
| 809 | // it arrives, so by the time a whole turn is assembled the tag has |
| 810 | // already been rendered. See `inline_tool_calls`. |
| 811 | let mut scanner = crate::inline_tool_calls::StreamScanner::new(tools); |
| 812 | let mut recovered: Vec<ToolCall> = Vec::new(); |
| 813 | let mut malformed = 0usize; |
| 814 | |
| 815 | while let Some(item) = stream.next().await { |
| 816 | let bytes = item.map_err(|error| format!("stream read error: {error}"))?; |
| 817 | buffer.extend_from_slice(&bytes); |
| 818 | |
| 819 | while let Some(pos) = buffer.iter().position(|&b| b == b'\n') { |
| 820 | let line: Vec<u8> = buffer.drain(..=pos).collect(); |
| 821 | let line = String::from_utf8_lossy(&line); |
| 822 | let line = line.trim(); |
| 823 | |
| 824 | let Some(data) = line.strip_prefix("data:") else { |
| 825 | continue; |
| 826 | }; |
| 827 | let data = data.trim(); |
| 828 | if data == "[DONE]" { |
| 829 | done = true; |
| 830 | break; |
| 831 | } |
| 832 | if data.is_empty() { |
| 833 | continue; |
| 834 | } |
| 835 | |
| 836 | // An endpoint that fails after the status line is already sent |
| 837 | // reports it in the stream instead, as an ordinary `data:` |
| 838 | // frame holding an error envelope. siGit Code Cloud does this |
| 839 | // when the upstream fails mid-turn. It has no `choices`, so |
| 840 | // without this it parses as an empty chunk and is skipped, and |
| 841 | // the turn ends looking like the model simply said nothing. |
| 842 | if let Some(message) = api_error_message(data) { |
| 843 | return Err(message); |
| 844 | } |
| 845 | |
| 846 | let chunk: StreamCompletion = match serde_json::from_str(data) { |
| 847 | Ok(chunk) => chunk, |
| 848 | // Skip keep-alive comments and anything we can't parse rather |
| 849 | // than aborting a turn over one malformed frame. |
| 850 | Err(_) => continue, |
| 851 | }; |
| 852 | |
| 853 | let Some(choice) = chunk.choices.into_iter().next() else { |
| 854 | continue; |
| 855 | }; |
| 856 | if let Some(content) = choice.delta.content |
| 857 | && !content.is_empty() |
| 858 | { |
| 859 | let mut cancelled = false; |
| 860 | for event in scanner.push(&content) { |
| 861 | match event { |
| 862 | crate::inline_tool_calls::ScanEvent::Text(chunk) => { |
| 863 | text.push_str(&chunk); |
| 864 | if sink.send(chunk).is_err() { |
| 865 | // Consumer dropped (turn cancelled). |
| 866 | cancelled = true; |
| 867 | break; |
| 868 | } |
| 869 | } |
| 870 | crate::inline_tool_calls::ScanEvent::ToolCall(call) => { |
| 871 | if allow_tool_calls { |
| 872 | log::warn!( |
| 873 | "recovered tool call '{}' the model emitted as text instead of a structured call", |
| 874 | call.name |
| 875 | ); |
| 876 | } |
| 877 | recovered.push(ToolCall { |
| 878 | id: format!("call_recovered_{}", recovered.len()), |
| 879 | name: call.name, |
| 880 | arguments: call.arguments, |
| 881 | }); |
| 882 | } |
| 883 | crate::inline_tool_calls::ScanEvent::Malformed(block) => { |
| 884 | log_malformed_block(&block); |
| 885 | malformed += 1; |
| 886 | } |
| 887 | } |
| 888 | } |
| 889 | if cancelled { |
| 890 | done = true; |
| 891 | break; |
| 892 | } |
| 893 | } |
| 894 | for delta in choice.delta.tool_calls.into_iter().flatten() { |
| 895 | let index = delta.index.unwrap_or(0) as usize; |
| 896 | if tool_accum.len() <= index { |
| 897 | tool_accum.resize_with(index + 1, StreamingToolCall::default); |
| 898 | } |
| 899 | let slot = &mut tool_accum[index]; |
| 900 | if let Some(id) = delta.id { |
| 901 | slot.id = id; |
| 902 | } |
| 903 | if let Some(function) = delta.function { |
| 904 | if let Some(name) = function.name { |
| 905 | slot.name = name; |
| 906 | } |
| 907 | if let Some(arguments) = function.arguments { |
| 908 | slot.arguments.push_str(&arguments); |
| 909 | } |
| 910 | } |
| 911 | } |
| 912 | } |
| 913 | |
| 914 | if done { |
| 915 | break; |
| 916 | } |
| 917 | } |
| 918 | |
| 919 | // Flush what was held back. A partial marker is just text; a |
| 920 | // tool-call block that never closed is dropped like any other |
| 921 | // unparseable one. |
| 922 | match scanner.finish() { |
| 923 | Some(crate::inline_tool_calls::ScanEvent::Text(leftover)) => { |
| 924 | text.push_str(&leftover); |
| 925 | let _ = sink.send(leftover); |
| 926 | } |
| 927 | Some(crate::inline_tool_calls::ScanEvent::Malformed(block)) => { |
| 928 | log_malformed_block(&block); |
| 929 | malformed += 1; |
| 930 | } |
| 931 | Some(crate::inline_tool_calls::ScanEvent::ToolCall(_)) | None => {} |
| 932 | } |
| 933 | |
| 934 | let mut tool_calls: Vec<ToolCall> = tool_accum |
| 935 | .iter() |
| 936 | .filter(|call| !call.name.is_empty()) |
| 937 | .enumerate() |
| 938 | .map(|(index, call)| ToolCall { |
| 939 | id: if call.id.is_empty() { |
| 940 | format!("call_{index}") |
| 941 | } else { |
| 942 | call.id.clone() |
| 943 | }, |
| 944 | name: call.name.clone(), |
| 945 | arguments: call.arguments.clone(), |
| 946 | }) |
| 947 | .collect(); |
| 948 | tool_calls.extend(recovered); |
| 949 | |
| 950 | if !allow_tool_calls && !tool_calls.is_empty() { |
| 951 | let names = tool_calls |
| 952 | .iter() |
| 953 | .map(|call| call.name.as_str()) |
| 954 | .collect::<Vec<_>>() |
| 955 | .join(", "); |
| 956 | log::warn!( |
| 957 | "suppressed {} tool call(s) from a forced-text response: {names}", |
| 958 | tool_calls.len() |
| 959 | ); |
| 960 | tool_calls.clear(); |
| 961 | } |
| 962 | |
| 963 | // Record the assistant turn so later tool results have context. |
| 964 | if malformed > 0 && tool_calls.is_empty() { |
| 965 | self.push_reply_without_calls(&text).await; |
| 966 | } else { |
| 967 | self.history |
| 968 | .lock() |
| 969 | .await |
| 970 | .push(streamed_assistant_history(&text, &tool_calls)); |
| 971 | } |
| 972 | |
| 973 | Ok((TurnResult { text, tool_calls }, malformed)) |
| 974 | } |
| 975 | } |
| 976 | |
| 977 | /// Sent as a user turn after a reply whose tool call couldn't be parsed. |
| 978 | const MALFORMED_TOOL_CALL_RETRY: &str = "[siGit Code] Your last reply tried to call a tool by \ |
| 979 | writing the call out as text, and it could not be parsed, so nothing ran and the user did \ |
| 980 | not see it. Do not write tool calls or tool results as text. Make the call again using the \ |
| 981 | tool-calling interface, or answer in plain prose if no tool is needed."; |
| 982 | |
| 983 | fn log_malformed_block(block: &str) { |
| 984 | log::warn!( |
| 985 | "dropped an unparseable inline tool call ({} chars): {}", |
| 986 | block.len(), |
| 987 | block.chars().take(200).collect::<String>() |
| 988 | ); |
| 989 | } |
| 990 | |
| 991 | /// Join the visible text of a reply and its retry the way the stream showed |
| 992 | /// them: as separate paragraphs. |
| 993 | fn join_reply_text(first: &str, retry: &str) -> String { |
| 994 | match (first.trim().is_empty(), retry.trim().is_empty()) { |
| 995 | (true, _) => retry.to_string(), |
| 996 | (false, true) => first.to_string(), |
| 997 | (false, false) => format!("{first}\n\n{retry}"), |
| 998 | } |
| 999 | } |
| 1000 | |
| 1001 | /// One tool call being reassembled from streamed deltas. |
| 1002 | #[derive(Default)] |
| 1003 | struct StreamingToolCall { |
| 1004 | id: String, |
| 1005 | name: String, |
| 1006 | arguments: String, |
| 1007 | } |
| 1008 | |
| 1009 | /// Rebuild the assistant message for replay in history after a streamed turn, |
| 1010 | /// preserving any tool calls so the follow-up request is well-formed. Mirrors |
| 1011 | /// [`ResponseMessage::into_history_value`] for the non-streaming path. |
| 1012 | fn streamed_assistant_history(text: &str, tool_calls: &[ToolCall]) -> serde_json::Value { |
| 1013 | let mut message = serde_json::json!({ "role": "assistant" }); |
| 1014 | message["content"] = if text.is_empty() { |
| 1015 | serde_json::Value::Null |
| 1016 | } else { |
| 1017 | serde_json::Value::String(text.to_string()) |
| 1018 | }; |
| 1019 | if !tool_calls.is_empty() { |
| 1020 | message["tool_calls"] = serde_json::json!( |
| 1021 | tool_calls |
| 1022 | .iter() |
| 1023 | .map(|call| serde_json::json!({ |
| 1024 | "id": call.id, |
| 1025 | "type": "function", |
| 1026 | "function": { |
| 1027 | "name": call.name, |
| 1028 | "arguments": call.arguments, |
| 1029 | } |
| 1030 | })) |
| 1031 | .collect::<Vec<_>>() |
| 1032 | ); |
| 1033 | } |
| 1034 | message |
| 1035 | } |
| 1036 | |
| 1037 | #[async_trait] |
| 1038 | impl InferenceBackend for OpenAiBackend { |
| 1039 | async fn send_message_with_tools( |
| 1040 | &self, |
| 1041 | text: &str, |
| 1042 | tools: &[ToolSpec], |
| 1043 | sink: Option<&TokenSink>, |
| 1044 | ) -> Result<TurnResult, BackendError> { |
| 1045 | self.history |
| 1046 | .lock() |
| 1047 | .await |
| 1048 | .push(serde_json::json!({ "role": "user", "content": text })); |
| 1049 | self.complete(tools, true, sink).await |
| 1050 | } |
| 1051 | |
| 1052 | async fn send_tool_results( |
| 1053 | &self, |
| 1054 | results: Vec<ToolResult>, |
| 1055 | tools: &[ToolSpec], |
| 1056 | allow_tool_calls: bool, |
| 1057 | sink: Option<&TokenSink>, |
| 1058 | ) -> Result<TurnResult, BackendError> { |
| 1059 | { |
| 1060 | let mut history = self.history.lock().await; |
| 1061 | for result in results { |
| 1062 | history.push(serde_json::json!({ |
| 1063 | "role": "tool", |
| 1064 | "tool_call_id": result.tool_call_id, |
| 1065 | "content": result.content, |
| 1066 | })); |
| 1067 | } |
| 1068 | } |
| 1069 | self.complete(tools, allow_tool_calls, sink).await |
| 1070 | } |
| 1071 | |
| 1072 | async fn record_cancelled_tool_results(&self, results: Vec<ToolResult>) { |
| 1073 | let mut history = self.history.lock().await; |
| 1074 | for result in results { |
| 1075 | // Cancellation races the in-flight continuation request. That |
| 1076 | // request appends its tool results before awaiting HTTP, so cleanup |
| 1077 | // must be safe whether cancellation won just before or just after |
| 1078 | // the append. |
| 1079 | let matching_call = history.iter().rposition(|message| { |
| 1080 | message["role"] == "assistant" |
| 1081 | && message["tool_calls"].as_array().is_some_and(|calls| { |
| 1082 | calls.iter().any(|call| call["id"] == result.tool_call_id) |
| 1083 | }) |
| 1084 | }); |
| 1085 | let already_recorded = matching_call.is_some_and(|call_index| { |
| 1086 | history[call_index + 1..].iter().any(|message| { |
| 1087 | message["role"] == "tool" && message["tool_call_id"] == result.tool_call_id |
| 1088 | }) |
| 1089 | }); |
| 1090 | if already_recorded { |
| 1091 | continue; |
| 1092 | } |
| 1093 | history.push(serde_json::json!({ |
| 1094 | "role": "tool", |
| 1095 | "tool_call_id": result.tool_call_id, |
| 1096 | "content": result.content, |
| 1097 | })); |
| 1098 | } |
| 1099 | } |
| 1100 | |
| 1101 | fn is_remote(&self) -> bool { |
| 1102 | true |
| 1103 | } |
| 1104 | |
| 1105 | async fn history_snapshot(&self) -> Vec<serde_json::Value> { |
| 1106 | self.history.lock().await.clone() |
| 1107 | } |
| 1108 | |
| 1109 | async fn restore_history(&self, history: Vec<serde_json::Value>) { |
| 1110 | // The snapshot includes the seeded system message, so a wholesale |
| 1111 | // replacement restores exactly what was saved. |
| 1112 | *self.history.lock().await = history; |
| 1113 | } |
| 1114 | |
| 1115 | async fn compact_history(&self, keep_last: usize) -> Result<(), BackendError> { |
| 1116 | let snapshot: Vec<serde_json::Value> = self.history.lock().await.clone(); |
| 1117 | |
| 1118 | let system = snapshot |
| 1119 | .first() |
| 1120 | .filter(|message| message["role"] == "system") |
| 1121 | .cloned(); |
| 1122 | |
| 1123 | // Ask the endpoint for a summary of the conversation so far, through the |
| 1124 | // ordinary completion machinery (non-streaming). The request carries the |
| 1125 | // conversation as a flattened transcript in a single user message rather |
| 1126 | // than the live history: this round offers no tools, and a tool-shaped |
| 1127 | // history sent without a tool schema is rejected upstream (see |
| 1128 | // `transcript_for_summary`). |
| 1129 | let mut request = Vec::new(); |
| 1130 | if let Some(system) = system.clone() { |
| 1131 | request.push(system); |
| 1132 | } |
| 1133 | request.push(serde_json::json!({ |
| 1134 | "role": "user", |
| 1135 | "content": format!( |
| 1136 | "{}\n\n{SUMMARIZE_PROMPT}", |
| 1137 | transcript_for_summary(&snapshot), |
| 1138 | ), |
| 1139 | })); |
| 1140 | *self.history.lock().await = request; |
| 1141 | |
| 1142 | let summary = match self.complete(&[], false, None).await { |
| 1143 | Ok(result) => result.text, |
| 1144 | Err(error) => { |
| 1145 | // Roll back the summarization request; the turn never happened. |
| 1146 | *self.history.lock().await = snapshot; |
| 1147 | return Err(error); |
| 1148 | } |
| 1149 | }; |
| 1150 | |
| 1151 | let non_system: Vec<serde_json::Value> = snapshot |
| 1152 | .iter() |
| 1153 | .filter(|message| message["role"] != "system") |
| 1154 | .cloned() |
| 1155 | .collect(); |
| 1156 | let tail_start = non_system.len().saturating_sub(keep_last); |
| 1157 | let mut tail = non_system[tail_start..].to_vec(); |
| 1158 | // Drop leading tool results whose assistant tool-call message was |
| 1159 | // summarized away — strict endpoints reject orphaned `role: "tool"` |
| 1160 | // entries on the very next request. |
| 1161 | while tail |
| 1162 | .first() |
| 1163 | .is_some_and(|message| message["role"] == "tool") |
| 1164 | { |
| 1165 | tail.remove(0); |
| 1166 | } |
| 1167 | |
| 1168 | let mut rebuilt = Vec::new(); |
| 1169 | if let Some(system) = system { |
| 1170 | rebuilt.push(system); |
| 1171 | } |
| 1172 | rebuilt.push(serde_json::json!({ |
| 1173 | "role": "user", |
| 1174 | "content": format!("[Conversation summary]\n{summary}"), |
| 1175 | })); |
| 1176 | rebuilt.extend(tail); |
| 1177 | *self.history.lock().await = rebuilt; |
| 1178 | Ok(()) |
| 1179 | } |
| 1180 | } |
| 1181 | |
| 1182 | // ── OpenAI error shape ──────────────────────────────────────────────────────── |
| 1183 | |
| 1184 | /// The OpenAI error envelope: `{"error": {"message": ..., "type": ...}}`. |
| 1185 | /// |
| 1186 | /// Every endpoint siGit talks to speaks it, including siGit Code Cloud, whose |
| 1187 | /// messages are written for the person reading them ("Monthly siGit Code Cloud |
| 1188 | /// allowance reached..."). Worth unwrapping rather than pasting the raw body |
| 1189 | /// into the editor's error banner. |
| 1190 | #[derive(Debug, Deserialize)] |
| 1191 | struct ApiErrorEnvelope { |
| 1192 | error: ApiErrorBody, |
| 1193 | } |
| 1194 | |
| 1195 | #[derive(Debug, Deserialize)] |
| 1196 | struct ApiErrorBody { |
| 1197 | #[serde(default)] |
| 1198 | message: Option<String>, |
| 1199 | } |
| 1200 | |
| 1201 | /// How much of an unparseable error body to keep. An endpoint behind a proxy |
| 1202 | /// can answer with a full HTML page, and the whole thing ends up in the |
| 1203 | /// editor's error banner. |
| 1204 | const ERROR_BODY_LIMIT: usize = 500; |
| 1205 | |
| 1206 | /// Turn an error response into something worth showing a person. |
| 1207 | /// |
| 1208 | /// The endpoint's own message wins when there is one: it is written for the |
| 1209 | /// user, and the status code repeats what it already says. Anything else falls |
| 1210 | /// back to the status plus whatever the body held, which is all there is to go |
| 1211 | /// on. |
| 1212 | fn describe_api_error(status: reqwest::StatusCode, body: &str) -> String { |
| 1213 | if let Some(message) = api_error_message(body) { |
| 1214 | return message; |
| 1215 | } |
| 1216 | |
| 1217 | let body = body.trim(); |
| 1218 | if body.is_empty() { |
| 1219 | return format!("endpoint returned {status}"); |
| 1220 | } |
| 1221 | |
| 1222 | let mut detail = body; |
| 1223 | if detail.len() > ERROR_BODY_LIMIT { |
| 1224 | let mut cut = ERROR_BODY_LIMIT; |
| 1225 | while !detail.is_char_boundary(cut) { |
| 1226 | cut -= 1; |
| 1227 | } |
| 1228 | detail = &detail[..cut]; |
| 1229 | return format!("endpoint returned {status}: {detail}…"); |
| 1230 | } |
| 1231 | format!("endpoint returned {status}: {detail}") |
| 1232 | } |
| 1233 | |
| 1234 | /// The human-readable message out of an OpenAI error envelope, if the body is |
| 1235 | /// one and carries a non-empty message. |
| 1236 | fn api_error_message(body: &str) -> Option<String> { |
| 1237 | let envelope: ApiErrorEnvelope = serde_json::from_str(body).ok()?; |
| 1238 | let message = envelope.error.message?; |
| 1239 | let message = message.trim(); |
| 1240 | (!message.is_empty()).then(|| message.to_string()) |
| 1241 | } |
| 1242 | |
| 1243 | // ── OpenAI response shapes ──────────────────────────────────────────────────────── |
| 1244 | |
| 1245 | #[derive(Debug, Deserialize)] |
| 1246 | struct ChatCompletion { |
| 1247 | #[serde(default)] |
| 1248 | choices: Vec<CompletionChoice>, |
| 1249 | } |
| 1250 | |
| 1251 | #[derive(Debug, Deserialize)] |
| 1252 | struct CompletionChoice { |
| 1253 | message: ResponseMessage, |
| 1254 | } |
| 1255 | |
| 1256 | #[derive(Debug, Deserialize)] |
| 1257 | struct ResponseMessage { |
| 1258 | #[serde(default)] |
| 1259 | content: Option<String>, |
| 1260 | #[serde(default)] |
| 1261 | tool_calls: Option<Vec<ResponseToolCall>>, |
| 1262 | } |
| 1263 | |
| 1264 | impl ResponseMessage { |
| 1265 | /// Reconstruct the assistant message for replay in history, preserving any |
| 1266 | /// tool calls so the follow-up request is well-formed. |
| 1267 | fn into_history_value(self) -> serde_json::Value { |
| 1268 | let mut message = serde_json::json!({ "role": "assistant" }); |
| 1269 | message["content"] = match self.content { |
| 1270 | Some(text) => serde_json::Value::String(text), |
| 1271 | None => serde_json::Value::Null, |
| 1272 | }; |
| 1273 | if let Some(tool_calls) = self.tool_calls { |
| 1274 | message["tool_calls"] = serde_json::json!( |
| 1275 | tool_calls |
| 1276 | .into_iter() |
| 1277 | .map(|call| serde_json::json!({ |
| 1278 | "id": call.id, |
| 1279 | "type": "function", |
| 1280 | "function": { |
| 1281 | "name": call.function.name, |
| 1282 | "arguments": call.function.arguments, |
| 1283 | } |
| 1284 | })) |
| 1285 | .collect::<Vec<_>>() |
| 1286 | ); |
| 1287 | } |
| 1288 | message |
| 1289 | } |
| 1290 | } |
| 1291 | |
| 1292 | #[derive(Debug, Deserialize)] |
| 1293 | struct ResponseToolCall { |
| 1294 | id: String, |
| 1295 | function: ResponseFunction, |
| 1296 | } |
| 1297 | |
| 1298 | #[derive(Debug, Deserialize)] |
| 1299 | struct ResponseFunction { |
| 1300 | name: String, |
| 1301 | #[serde(default)] |
| 1302 | arguments: String, |
| 1303 | } |
| 1304 | |
| 1305 | // ── OpenAI streaming (SSE) chunk shapes ───────────────────────────────────────── |
| 1306 | |
| 1307 | #[derive(Debug, Deserialize)] |
| 1308 | struct StreamCompletion { |
| 1309 | #[serde(default)] |
| 1310 | choices: Vec<StreamChoice>, |
| 1311 | } |
| 1312 | |
| 1313 | #[derive(Debug, Deserialize)] |
| 1314 | struct StreamChoice { |
| 1315 | #[serde(default)] |
| 1316 | delta: StreamDelta, |
| 1317 | } |
| 1318 | |
| 1319 | #[derive(Debug, Default, Deserialize)] |
| 1320 | struct StreamDelta { |
| 1321 | #[serde(default)] |
| 1322 | content: Option<String>, |
| 1323 | #[serde(default)] |
| 1324 | tool_calls: Option<Vec<StreamToolCallDelta>>, |
| 1325 | } |
| 1326 | |
| 1327 | #[derive(Debug, Deserialize)] |
| 1328 | struct StreamToolCallDelta { |
| 1329 | #[serde(default)] |
| 1330 | index: Option<u32>, |
| 1331 | #[serde(default)] |
| 1332 | id: Option<String>, |
| 1333 | #[serde(default)] |
| 1334 | function: Option<StreamFunctionDelta>, |
| 1335 | } |
| 1336 | |
| 1337 | #[derive(Debug, Deserialize)] |
| 1338 | struct StreamFunctionDelta { |
| 1339 | #[serde(default)] |
| 1340 | name: Option<String>, |
| 1341 | #[serde(default)] |
| 1342 | arguments: Option<String>, |
| 1343 | } |
| 1344 | |
| 1345 | // ── Carrying a conversation across a model switch ──────────────────────────── |
| 1346 | |
| 1347 | /// The part of a history snapshot that survives a model switch: every |
| 1348 | /// non-system message, with half-finished tool plumbing repaired. |
| 1349 | /// |
| 1350 | /// System messages are dropped because the backend being switched *to* seeds |
| 1351 | /// its own — a different model's prompt, or a freshly pushed session-context |
| 1352 | /// message. A switch can land mid-turn, after the assistant asked for a tool |
| 1353 | /// but before the results came back; a tool call with no matching result (or a |
| 1354 | /// result with no matching call) makes strict OpenAI-compatible endpoints |
| 1355 | /// reject every later request in the session, so those halves are stripped |
| 1356 | /// instead of carried. |
| 1357 | pub fn carryover_history(snapshot: Vec<serde_json::Value>) -> Vec<serde_json::Value> { |
| 1358 | let answered: std::collections::HashSet<String> = snapshot |
| 1359 | .iter() |
| 1360 | .filter(|message| message["role"] == "tool") |
| 1361 | .filter_map(|message| message["tool_call_id"].as_str().map(str::to_string)) |
| 1362 | .collect(); |
| 1363 | |
| 1364 | let mut carried: Vec<serde_json::Value> = Vec::with_capacity(snapshot.len()); |
| 1365 | let mut called: std::collections::HashSet<String> = std::collections::HashSet::new(); |
| 1366 | |
| 1367 | for mut message in snapshot { |
| 1368 | match message["role"].as_str().unwrap_or_default() { |
| 1369 | "system" => continue, |
| 1370 | "assistant" => { |
| 1371 | if let Some(calls) = message["tool_calls"].as_array() { |
| 1372 | let kept: Vec<serde_json::Value> = calls |
| 1373 | .iter() |
| 1374 | .filter(|call| call["id"].as_str().is_some_and(|id| answered.contains(id))) |
| 1375 | .cloned() |
| 1376 | .collect(); |
| 1377 | if kept.is_empty() { |
| 1378 | // Only unanswered requests: keep whatever text came |
| 1379 | // with them, drop the message if there was none. |
| 1380 | if let Some(object) = message.as_object_mut() { |
| 1381 | object.remove("tool_calls"); |
| 1382 | } |
| 1383 | if message["content"].as_str().unwrap_or_default().is_empty() { |
| 1384 | continue; |
| 1385 | } |
| 1386 | } else { |
| 1387 | for call in &kept { |
| 1388 | called.insert(call["id"].as_str().unwrap_or_default().to_string()); |
| 1389 | } |
| 1390 | message["tool_calls"] = serde_json::Value::Array(kept); |
| 1391 | } |
| 1392 | } |
| 1393 | carried.push(message); |
| 1394 | } |
| 1395 | "tool" => { |
| 1396 | let answers_a_kept_call = message["tool_call_id"] |
| 1397 | .as_str() |
| 1398 | .is_some_and(|id| called.contains(id)); |
| 1399 | if answers_a_kept_call { |
| 1400 | carried.push(message); |
| 1401 | } |
| 1402 | } |
| 1403 | _ => carried.push(message), |
| 1404 | } |
| 1405 | } |
| 1406 | |
| 1407 | // Session switches happen only after the serialized turn has ended. A |
| 1408 | // trailing user message therefore belongs to a cancelled/failed inference |
| 1409 | // request with no answer; carrying it would make the next activation feed |
| 1410 | // the model an orphaned prompt that the client considers cancelled. |
| 1411 | while carried |
| 1412 | .last() |
| 1413 | .is_some_and(|message| message["role"] == "user") |
| 1414 | { |
| 1415 | carried.pop(); |
| 1416 | } |
| 1417 | |
| 1418 | carried |
| 1419 | } |
| 1420 | |
| 1421 | /// Replay `carried` (from [`carryover_history`]) into `backend`, on top of the |
| 1422 | /// system messages `backend` seeded for itself. Used when a model switch |
| 1423 | /// installs a new backend — or reloads the on-device engine, which wipes its |
| 1424 | /// history — so the thread continues under the new model instead of restarting. |
| 1425 | pub async fn adopt_carryover(backend: &dyn InferenceBackend, carried: Vec<serde_json::Value>) { |
| 1426 | if carried.is_empty() { |
| 1427 | return; |
| 1428 | } |
| 1429 | let mut rebuilt: Vec<serde_json::Value> = backend |
| 1430 | .history_snapshot() |
| 1431 | .await |
| 1432 | .into_iter() |
| 1433 | .take_while(|message| message["role"] == "system") |
| 1434 | .collect(); |
| 1435 | rebuilt.extend(carried); |
| 1436 | backend.restore_history(rebuilt).await; |
| 1437 | } |
| 1438 | |
| 1439 | #[cfg(test)] |
| 1440 | mod tests { |
| 1441 | use super::*; |
| 1442 | |
| 1443 | /// The case from the issue: the endpoint says something a person can act |
| 1444 | /// on, and the status code and JSON around it are noise. |
| 1445 | #[test] |
| 1446 | fn an_endpoints_own_message_is_what_the_user_sees() { |
| 1447 | let body = r#"{"error":{"message":"Monthly siGit Code Cloud allowance reached. It resets at the start of your next billing period.","type":"server_error"}}"#; |
| 1448 | |
| 1449 | let described = describe_api_error(reqwest::StatusCode::TOO_MANY_REQUESTS, body); |
| 1450 | |
| 1451 | assert_eq!( |
| 1452 | described, |
| 1453 | "Monthly siGit Code Cloud allowance reached. \ |
| 1454 | It resets at the start of your next billing period." |
| 1455 | ); |
| 1456 | } |
| 1457 | |
| 1458 | #[test] |
| 1459 | fn a_body_that_is_not_an_error_envelope_keeps_the_status() { |
| 1460 | let described = |
| 1461 | describe_api_error(reqwest::StatusCode::BAD_GATEWAY, "<html>bad gateway</html>"); |
| 1462 | |
| 1463 | assert!(described.contains("502"), "{described}"); |
| 1464 | assert!(described.contains("bad gateway"), "{described}"); |
| 1465 | } |
| 1466 | |
| 1467 | #[test] |
| 1468 | fn an_empty_body_still_says_something() { |
| 1469 | let described = describe_api_error(reqwest::StatusCode::INTERNAL_SERVER_ERROR, " "); |
| 1470 | |
| 1471 | assert_eq!(described, "endpoint returned 500 Internal Server Error"); |
| 1472 | } |
| 1473 | |
| 1474 | #[test] |
| 1475 | fn an_oversized_body_is_cut_down() { |
| 1476 | let body = "x".repeat(ERROR_BODY_LIMIT * 3); |
| 1477 | |
| 1478 | let described = describe_api_error(reqwest::StatusCode::BAD_GATEWAY, &body); |
| 1479 | |
| 1480 | assert!(described.len() < ERROR_BODY_LIMIT + 100, "{described}"); |
| 1481 | assert!(described.ends_with('…'), "{described}"); |
| 1482 | } |
| 1483 | |
| 1484 | /// An envelope with nothing useful in it must not shadow the status, which |
| 1485 | /// would leave the user with an empty error. |
| 1486 | #[test] |
| 1487 | fn an_envelope_with_a_blank_message_falls_back() { |
| 1488 | let described = describe_api_error( |
| 1489 | reqwest::StatusCode::FORBIDDEN, |
| 1490 | r#"{"error":{"message":" "}}"#, |
| 1491 | ); |
| 1492 | |
| 1493 | assert!(described.contains("403"), "{described}"); |
| 1494 | } |
| 1495 | |
| 1496 | #[test] |
| 1497 | fn tools_json_wraps_function_schema() { |
| 1498 | let tools = vec![ToolSpec { |
| 1499 | name: "read_file".to_string(), |
| 1500 | description: "Read a file".to_string(), |
| 1501 | parameters_schema: r#"{"type":"object","properties":{"path":{"type":"string"}}}"# |
| 1502 | .to_string(), |
| 1503 | }]; |
| 1504 | let json = OpenAiBackend::tools_json(&tools); |
| 1505 | assert_eq!(json[0]["type"], "function"); |
| 1506 | assert_eq!(json[0]["function"]["name"], "read_file"); |
| 1507 | assert_eq!( |
| 1508 | json[0]["function"]["parameters"]["properties"]["path"]["type"], |
| 1509 | "string" |
| 1510 | ); |
| 1511 | } |
| 1512 | |
| 1513 | #[test] |
| 1514 | fn malformed_schema_falls_back_to_empty_object() { |
| 1515 | let tools = vec![ToolSpec { |
| 1516 | name: "x".to_string(), |
| 1517 | description: String::new(), |
| 1518 | parameters_schema: "not json".to_string(), |
| 1519 | }]; |
| 1520 | let json = OpenAiBackend::tools_json(&tools); |
| 1521 | assert_eq!(json[0]["function"]["parameters"]["type"], "object"); |
| 1522 | } |
| 1523 | |
| 1524 | #[test] |
| 1525 | fn streamed_assistant_history_omits_empty_tool_calls() { |
| 1526 | let value = streamed_assistant_history("hello", &[]); |
| 1527 | assert_eq!(value["role"], "assistant"); |
| 1528 | assert_eq!(value["content"], "hello"); |
| 1529 | assert!(value.get("tool_calls").is_none()); |
| 1530 | } |
| 1531 | |
| 1532 | #[test] |
| 1533 | fn streamed_assistant_history_preserves_tool_calls() { |
| 1534 | let calls = vec![ToolCall { |
| 1535 | id: "call_0".to_string(), |
| 1536 | name: "read_file".to_string(), |
| 1537 | arguments: r#"{"path":"a.rs"}"#.to_string(), |
| 1538 | }]; |
| 1539 | let value = streamed_assistant_history("", &calls); |
| 1540 | assert!(value["content"].is_null()); |
| 1541 | assert_eq!(value["tool_calls"][0]["id"], "call_0"); |
| 1542 | assert_eq!(value["tool_calls"][0]["type"], "function"); |
| 1543 | assert_eq!(value["tool_calls"][0]["function"]["name"], "read_file"); |
| 1544 | assert_eq!( |
| 1545 | value["tool_calls"][0]["function"]["arguments"], |
| 1546 | r#"{"path":"a.rs"}"# |
| 1547 | ); |
| 1548 | } |
| 1549 | |
| 1550 | #[tokio::test] |
| 1551 | async fn cancelled_tool_results_close_out_history() { |
| 1552 | let backend = OpenAiBackend::new("http://localhost", "", "test-model", None); |
| 1553 | backend |
| 1554 | .history |
| 1555 | .lock() |
| 1556 | .await |
| 1557 | .push(streamed_assistant_history( |
| 1558 | "", |
| 1559 | &[ToolCall { |
| 1560 | id: "call_9".to_string(), |
| 1561 | name: "run_command".to_string(), |
| 1562 | arguments: r#"{"command":"ls"}"#.to_string(), |
| 1563 | }], |
| 1564 | )); |
| 1565 | |
| 1566 | backend |
| 1567 | .record_cancelled_tool_results(vec![ToolResult { |
| 1568 | tool_call_id: "call_9".to_string(), |
| 1569 | content: "cancelled by the user".to_string(), |
| 1570 | }]) |
| 1571 | .await; |
| 1572 | |
| 1573 | let history = backend.history.lock().await; |
| 1574 | let last = history.last().unwrap(); |
| 1575 | assert_eq!(last["role"], "tool"); |
| 1576 | assert_eq!(last["tool_call_id"], "call_9"); |
| 1577 | assert_eq!(last["content"], "cancelled by the user"); |
| 1578 | |
| 1579 | drop(history); |
| 1580 | backend |
| 1581 | .record_cancelled_tool_results(vec![ToolResult { |
| 1582 | tool_call_id: "call_9".to_string(), |
| 1583 | content: "duplicate cleanup".to_string(), |
| 1584 | }]) |
| 1585 | .await; |
| 1586 | let history = backend.history.lock().await; |
| 1587 | assert_eq!( |
| 1588 | history |
| 1589 | .iter() |
| 1590 | .filter(|message| message["tool_call_id"] == "call_9") |
| 1591 | .count(), |
| 1592 | 1 |
| 1593 | ); |
| 1594 | } |
| 1595 | |
| 1596 | #[test] |
| 1597 | fn carryover_drops_system_messages_and_keeps_the_turns() { |
| 1598 | let snapshot = vec![ |
| 1599 | serde_json::json!({ "role": "system", "content": "old prompt" }), |
| 1600 | serde_json::json!({ "role": "user", "content": "hello" }), |
| 1601 | serde_json::json!({ "role": "assistant", "content": "hi" }), |
| 1602 | ]; |
| 1603 | |
| 1604 | let carried = carryover_history(snapshot); |
| 1605 | |
| 1606 | assert_eq!(carried.len(), 2); |
| 1607 | assert_eq!(carried[0]["role"], "user"); |
| 1608 | assert_eq!(carried[1]["content"], "hi"); |
| 1609 | } |
| 1610 | |
| 1611 | #[test] |
| 1612 | fn carryover_drops_a_cancelled_trailing_user_message() { |
| 1613 | let carried = carryover_history(vec![ |
| 1614 | serde_json::json!({ "role": "system", "content": "prompt" }), |
| 1615 | serde_json::json!({ "role": "user", "content": "completed question" }), |
| 1616 | serde_json::json!({ "role": "assistant", "content": "completed answer" }), |
| 1617 | serde_json::json!({ "role": "user", "content": "cancelled question" }), |
| 1618 | ]); |
| 1619 | |
| 1620 | assert_eq!(carried.len(), 2, "{carried:#?}"); |
| 1621 | assert_eq!(carried.last().unwrap()["content"], "completed answer"); |
| 1622 | } |
| 1623 | |
| 1624 | #[test] |
| 1625 | fn carryover_strips_a_tool_call_that_never_got_a_result() { |
| 1626 | // Switching mid-turn: the assistant asked for a tool, nothing answered. |
| 1627 | let snapshot = vec![ |
| 1628 | serde_json::json!({ "role": "user", "content": "read a.rs" }), |
| 1629 | streamed_assistant_history( |
| 1630 | "on it", |
| 1631 | &[ToolCall { |
| 1632 | id: "call_1".to_string(), |
| 1633 | name: "read_file".to_string(), |
| 1634 | arguments: "{}".to_string(), |
| 1635 | }], |
| 1636 | ), |
| 1637 | ]; |
| 1638 | |
| 1639 | let carried = carryover_history(snapshot); |
| 1640 | |
| 1641 | assert_eq!(carried.len(), 2); |
| 1642 | assert_eq!(carried[1]["content"], "on it"); |
| 1643 | assert!( |
| 1644 | carried[1].get("tool_calls").is_none(), |
| 1645 | "an unanswered tool call must not survive the switch" |
| 1646 | ); |
| 1647 | } |
| 1648 | |
| 1649 | #[test] |
| 1650 | fn carryover_drops_a_textless_unanswered_tool_call_and_its_late_result() { |
| 1651 | let snapshot = vec![ |
| 1652 | streamed_assistant_history( |
| 1653 | "", |
| 1654 | &[ToolCall { |
| 1655 | id: "call_1".to_string(), |
| 1656 | name: "read_file".to_string(), |
| 1657 | arguments: "{}".to_string(), |
| 1658 | }], |
| 1659 | ), |
| 1660 | // An orphan: its assistant message is gone with the line above. |
| 1661 | serde_json::json!({ |
| 1662 | "role": "tool", |
| 1663 | "tool_call_id": "call_other", |
| 1664 | "content": "file contents", |
| 1665 | }), |
| 1666 | ]; |
| 1667 | |
| 1668 | assert!(carryover_history(snapshot).is_empty()); |
| 1669 | } |
| 1670 | |
| 1671 | #[test] |
| 1672 | fn carryover_keeps_an_answered_tool_call_paired_with_its_result() { |
| 1673 | let snapshot = vec![ |
| 1674 | streamed_assistant_history( |
| 1675 | "", |
| 1676 | &[ToolCall { |
| 1677 | id: "call_1".to_string(), |
| 1678 | name: "read_file".to_string(), |
| 1679 | arguments: "{}".to_string(), |
| 1680 | }], |
| 1681 | ), |
| 1682 | serde_json::json!({ |
| 1683 | "role": "tool", |
| 1684 | "tool_call_id": "call_1", |
| 1685 | "content": "file contents", |
| 1686 | }), |
| 1687 | ]; |
| 1688 | |
| 1689 | let carried = carryover_history(snapshot); |
| 1690 | |
| 1691 | assert_eq!(carried.len(), 2); |
| 1692 | assert_eq!(carried[0]["tool_calls"][0]["id"], "call_1"); |
| 1693 | assert_eq!(carried[1]["tool_call_id"], "call_1"); |
| 1694 | } |
| 1695 | |
| 1696 | #[tokio::test] |
| 1697 | async fn adopt_carryover_replays_the_thread_under_the_new_system_prompt() { |
| 1698 | let new_backend = |
| 1699 | OpenAiBackend::new("http://localhost", "", "m", Some("new prompt".into())); |
| 1700 | |
| 1701 | let carried = carryover_history(vec![ |
| 1702 | serde_json::json!({ "role": "system", "content": "old prompt" }), |
| 1703 | serde_json::json!({ "role": "user", "content": "hello" }), |
| 1704 | serde_json::json!({ "role": "assistant", "content": "hi" }), |
| 1705 | ]); |
| 1706 | adopt_carryover(&new_backend, carried).await; |
| 1707 | |
| 1708 | let history = new_backend.history_snapshot().await; |
| 1709 | assert_eq!(history.len(), 3); |
| 1710 | assert_eq!(history[0]["role"], "system"); |
| 1711 | assert_eq!(history[0]["content"], "new prompt"); |
| 1712 | assert_eq!(history[1]["content"], "hello"); |
| 1713 | assert_eq!(history[2]["content"], "hi"); |
| 1714 | } |
| 1715 | |
| 1716 | #[tokio::test] |
| 1717 | async fn adopt_carryover_leaves_a_fresh_backend_alone() { |
| 1718 | let new_backend = |
| 1719 | OpenAiBackend::new("http://localhost", "", "m", Some("new prompt".into())); |
| 1720 | |
| 1721 | adopt_carryover(&new_backend, Vec::new()).await; |
| 1722 | |
| 1723 | let history = new_backend.history_snapshot().await; |
| 1724 | assert_eq!(history.len(), 1); |
| 1725 | assert_eq!(history[0]["content"], "new prompt"); |
| 1726 | } |
| 1727 | |
| 1728 | #[test] |
| 1729 | fn estimate_tokens_scales_with_serialized_size() { |
| 1730 | assert_eq!(estimate_tokens(&[]), 0); |
| 1731 | |
| 1732 | let short = vec![serde_json::json!({ "role": "user", "content": "hi" })]; |
| 1733 | let long = vec![serde_json::json!({ "role": "user", "content": "x".repeat(4_000) })]; |
| 1734 | let short_estimate = estimate_tokens(&short); |
| 1735 | let long_estimate = estimate_tokens(&long); |
| 1736 | |
| 1737 | assert!(short_estimate > 0, "non-empty history estimates > 0 tokens"); |
| 1738 | assert!(long_estimate > short_estimate, "longer history costs more"); |
| 1739 | // 4,000 content chars / 4 ≈ 1,000 tokens, plus a little JSON framing. |
| 1740 | assert!((1_000..1_100).contains(&long_estimate), "{long_estimate}"); |
| 1741 | } |
| 1742 | |
| 1743 | #[tokio::test] |
| 1744 | async fn openai_snapshot_restore_round_trips_exactly() { |
| 1745 | let backend = OpenAiBackend::new("http://localhost", "", "m", Some("be helpful".into())); |
| 1746 | { |
| 1747 | let mut history = backend.history.lock().await; |
| 1748 | history.push(serde_json::json!({ "role": "user", "content": "hello" })); |
| 1749 | history.push(streamed_assistant_history( |
| 1750 | "", |
| 1751 | &[ToolCall { |
| 1752 | id: "call_1".to_string(), |
| 1753 | name: "read_file".to_string(), |
| 1754 | arguments: r#"{"path":"a.rs"}"#.to_string(), |
| 1755 | }], |
| 1756 | )); |
| 1757 | history.push(serde_json::json!({ |
| 1758 | "role": "tool", "tool_call_id": "call_1", "content": "fn main() {}", |
| 1759 | })); |
| 1760 | history.push(serde_json::json!({ "role": "assistant", "content": "done" })); |
| 1761 | } |
| 1762 | let snapshot = backend.history_snapshot().await; |
| 1763 | assert_eq!( |
| 1764 | snapshot[0]["role"], "system", |
| 1765 | "snapshot keeps the system message" |
| 1766 | ); |
| 1767 | |
| 1768 | // Restoring into a backend seeded with a *different* system prompt must |
| 1769 | // replace everything, including that seed. |
| 1770 | let restored = OpenAiBackend::new("http://localhost", "", "m", Some("other seed".into())); |
| 1771 | restored.restore_history(snapshot.clone()).await; |
| 1772 | assert_eq!(restored.history_snapshot().await, snapshot); |
| 1773 | } |
| 1774 | |
| 1775 | #[tokio::test] |
| 1776 | async fn local_restore_replaces_engine_history_and_system_context() { |
| 1777 | let engine = Arc::new(ChatEngine::new()); |
| 1778 | let backend = LocalBackend::new(Arc::clone(&engine)); |
| 1779 | backend |
| 1780 | .restore_history(vec![ |
| 1781 | serde_json::json!({ "role": "system", "content": "thread A context" }), |
| 1782 | serde_json::json!({ "role": "user", "content": "thread A question" }), |
| 1783 | ]) |
| 1784 | .await; |
| 1785 | |
| 1786 | backend |
| 1787 | .restore_history(vec![ |
| 1788 | serde_json::json!({ "role": "system", "content": "thread B context" }), |
| 1789 | serde_json::json!({ "role": "user", "content": "thread B question" }), |
| 1790 | ]) |
| 1791 | .await; |
| 1792 | |
| 1793 | let restored = backend.history_snapshot().await; |
| 1794 | assert_eq!(restored.len(), 2, "{restored:#?}"); |
| 1795 | assert_eq!(restored[0]["content"], "thread B context"); |
| 1796 | assert_eq!(restored[1]["content"], "thread B question"); |
| 1797 | assert!( |
| 1798 | restored.iter().all(|message| !message["content"] |
| 1799 | .as_str() |
| 1800 | .unwrap_or_default() |
| 1801 | .contains("thread A")), |
| 1802 | "restore_history must replace, not append to, the local engine: {restored:#?}" |
| 1803 | ); |
| 1804 | |
| 1805 | let error = backend |
| 1806 | .send_message_with_tools("must not run", &[], None) |
| 1807 | .await |
| 1808 | .unwrap_err(); |
| 1809 | assert!(error.contains("not ready"), "{error}"); |
| 1810 | assert_eq!(backend.history_snapshot().await, restored); |
| 1811 | } |
| 1812 | |
| 1813 | /// Minimal scripted OpenAI-compatible endpoint: accepts one HTTP request on |
| 1814 | /// a std listener and answers with a fixed non-streaming completion. The |
| 1815 | /// receiver yields the request body the backend actually put on the wire. |
| 1816 | fn spawn_completion_stub( |
| 1817 | summary: &str, |
| 1818 | ) -> (std::net::SocketAddr, std::sync::mpsc::Receiver<String>) { |
| 1819 | spawn_message_stub(serde_json::json!({ "role": "assistant", "content": summary })) |
| 1820 | } |
| 1821 | |
| 1822 | /// Like [`spawn_completion_stub`], answering with an arbitrary message. |
| 1823 | fn spawn_message_stub( |
| 1824 | message: serde_json::Value, |
| 1825 | ) -> (std::net::SocketAddr, std::sync::mpsc::Receiver<String>) { |
| 1826 | use std::io::{Read, Write}; |
| 1827 | |
| 1828 | let listener = std::net::TcpListener::bind("127.0.0.1:0").unwrap(); |
| 1829 | let addr = listener.local_addr().unwrap(); |
| 1830 | let (sender, receiver) = std::sync::mpsc::channel(); |
| 1831 | let body = serde_json::json!({ "choices": [{ "message": message }] }).to_string(); |
| 1832 | std::thread::spawn(move || { |
| 1833 | let (mut stream, _) = listener.accept().unwrap(); |
| 1834 | // Read until the full request (headers + content-length body) is in. |
| 1835 | let mut request = Vec::new(); |
| 1836 | let mut chunk = [0u8; 4096]; |
| 1837 | loop { |
| 1838 | let n = stream.read(&mut chunk).unwrap_or(0); |
| 1839 | if n == 0 { |
| 1840 | break; |
| 1841 | } |
| 1842 | request.extend_from_slice(&chunk[..n]); |
| 1843 | if let Some(headers_end) = |
| 1844 | request.windows(4).position(|window| window == b"\r\n\r\n") |
| 1845 | { |
| 1846 | let headers = String::from_utf8_lossy(&request[..headers_end]); |
| 1847 | let content_length = headers |
| 1848 | .lines() |
| 1849 | .find_map(|line| { |
| 1850 | line.to_ascii_lowercase() |
| 1851 | .strip_prefix("content-length:") |
| 1852 | .map(|value| value.trim().parse::<usize>().unwrap_or(0)) |
| 1853 | }) |
| 1854 | .unwrap_or(0); |
| 1855 | if request.len() >= headers_end + 4 + content_length { |
| 1856 | let _ = sender.send( |
| 1857 | String::from_utf8_lossy(&request[headers_end + 4..]).into_owned(), |
| 1858 | ); |
| 1859 | break; |
| 1860 | } |
| 1861 | } |
| 1862 | } |
| 1863 | let response = format!( |
| 1864 | "HTTP/1.1 200 OK\r\ncontent-type: application/json\r\n\ |
| 1865 | content-length: {}\r\nconnection: close\r\n\r\n{}", |
| 1866 | body.len(), |
| 1867 | body |
| 1868 | ); |
| 1869 | let _ = stream.write_all(response.as_bytes()); |
| 1870 | }); |
| 1871 | (addr, receiver) |
| 1872 | } |
| 1873 | |
| 1874 | /// A structured call can come back with a broken inline block in the same |
| 1875 | /// message. The call runs; the block reaches neither the reply nor the |
| 1876 | /// history, where the model would read back a call that never ran. |
| 1877 | #[tokio::test] |
| 1878 | async fn a_broken_inline_block_beside_a_structured_call_is_dropped() { |
| 1879 | let (addr, _requests) = spawn_message_stub(serde_json::json!({ |
| 1880 | "role": "assistant", |
| 1881 | "content": "Checking.<tool_call>run_command CheckStatus=nope</tool_call>", |
| 1882 | "tool_calls": [{ |
| 1883 | "id": "call_1", |
| 1884 | "type": "function", |
| 1885 | "function": { "name": "run_command", "arguments": "{\"command\":\"pwd\"}" }, |
| 1886 | }], |
| 1887 | })); |
| 1888 | let backend = |
| 1889 | OpenAiBackend::new(format!("http://{addr}/v1"), "test-key", "test-model", None); |
| 1890 | let tools = vec![ToolSpec { |
| 1891 | name: "run_command".to_string(), |
| 1892 | description: "Run a command".to_string(), |
| 1893 | parameters_schema: r#"{"type":"object","properties":{"command":{"type":"string"}}}"# |
| 1894 | .to_string(), |
| 1895 | }]; |
| 1896 | |
| 1897 | let result = backend |
| 1898 | .send_message_with_tools("where am I", &tools, None) |
| 1899 | .await |
| 1900 | .unwrap(); |
| 1901 | |
| 1902 | assert_eq!(result.text, "Checking."); |
| 1903 | assert_eq!(result.tool_calls.len(), 1); |
| 1904 | assert_eq!(result.tool_calls[0].id, "call_1"); |
| 1905 | let history = backend.history_snapshot().await; |
| 1906 | let reply = history.last().unwrap(); |
| 1907 | assert_eq!(reply["content"], "Checking."); |
| 1908 | assert_eq!(reply["tool_calls"][0]["id"], "call_1"); |
| 1909 | } |
| 1910 | |
| 1911 | /// A structured call plus a well-formed inline block: both run, and the |
| 1912 | /// inline markup leaves the reply and history, as on the streaming path. |
| 1913 | #[tokio::test] |
| 1914 | async fn a_well_formed_inline_block_beside_a_structured_call_also_runs() { |
| 1915 | let (addr, _requests) = spawn_message_stub(serde_json::json!({ |
| 1916 | "role": "assistant", |
| 1917 | "content": "Checking.<tool_call>run_command<arg_key>command</arg_key><arg_value>ls</arg_value></tool_call>", |
| 1918 | "tool_calls": [{ |
| 1919 | "id": "call_1", |
| 1920 | "type": "function", |
| 1921 | "function": { "name": "run_command", "arguments": "{\"command\":\"pwd\"}" }, |
| 1922 | }], |
| 1923 | })); |
| 1924 | let backend = |
| 1925 | OpenAiBackend::new(format!("http://{addr}/v1"), "test-key", "test-model", None); |
| 1926 | let tools = vec![ToolSpec { |
| 1927 | name: "run_command".to_string(), |
| 1928 | description: "Run a command".to_string(), |
| 1929 | parameters_schema: r#"{"type":"object","properties":{"command":{"type":"string"}}}"# |
| 1930 | .to_string(), |
| 1931 | }]; |
| 1932 | |
| 1933 | let result = backend |
| 1934 | .send_message_with_tools("where am I", &tools, None) |
| 1935 | .await |
| 1936 | .unwrap(); |
| 1937 | |
| 1938 | assert_eq!(result.text, "Checking."); |
| 1939 | let ids: Vec<&str> = result.tool_calls.iter().map(|c| c.id.as_str()).collect(); |
| 1940 | assert_eq!(ids, ["call_1", "call_recovered_1"]); |
| 1941 | assert_eq!(result.tool_calls[1].arguments, r#"{"command":"ls"}"#); |
| 1942 | let history = backend.history_snapshot().await; |
| 1943 | let reply = history.last().unwrap(); |
| 1944 | assert_eq!(reply["content"], "Checking."); |
| 1945 | assert_eq!(reply["tool_calls"].as_array().unwrap().len(), 2); |
| 1946 | } |
| 1947 | |
| 1948 | #[tokio::test] |
| 1949 | async fn compact_history_rebuilds_system_summary_and_tail() { |
| 1950 | let (addr, _requests) = spawn_completion_stub("We refactored backend.rs; tests pass."); |
| 1951 | let backend = OpenAiBackend::new( |
| 1952 | format!("http://{addr}/v1"), |
| 1953 | "test-key", |
| 1954 | "test-model", |
| 1955 | Some("be helpful".into()), |
| 1956 | ); |
| 1957 | { |
| 1958 | let mut history = backend.history.lock().await; |
| 1959 | for i in 0..5 { |
| 1960 | let role = if i % 2 == 0 { "user" } else { "assistant" }; |
| 1961 | history.push(serde_json::json!({ |
| 1962 | "role": role, "content": format!("message {i}"), |
| 1963 | })); |
| 1964 | } |
| 1965 | } |
| 1966 | |
| 1967 | backend.compact_history(2).await.unwrap(); |
| 1968 | |
| 1969 | let history = backend.history_snapshot().await; |
| 1970 | assert_eq!(history.len(), 4, "system + summary + last 2: {history:?}"); |
| 1971 | assert_eq!(history[0]["role"], "system"); |
| 1972 | assert_eq!(history[0]["content"], "be helpful"); |
| 1973 | assert_eq!(history[1]["role"], "user"); |
| 1974 | let summary_text = history[1]["content"].as_str().unwrap(); |
| 1975 | assert!(summary_text.starts_with("[Conversation summary]\n")); |
| 1976 | assert!(summary_text.contains("We refactored backend.rs; tests pass.")); |
| 1977 | assert_eq!( |
| 1978 | history[2], |
| 1979 | serde_json::json!({ "role": "assistant", "content": "message 3" }) |
| 1980 | ); |
| 1981 | assert_eq!( |
| 1982 | history[3], |
| 1983 | serde_json::json!({ "role": "user", "content": "message 4" }) |
| 1984 | ); |
| 1985 | } |
| 1986 | |
| 1987 | /// Compacting a tool-heavy session must not put tool shapes on the wire. |
| 1988 | /// The summarization round offers no `tools`, and endpoints reject tool |
| 1989 | /// calls and tool results that arrive without a schema — which used to make |
| 1990 | /// compaction fail forever in any session that had run a single tool. |
| 1991 | #[tokio::test] |
| 1992 | async fn compact_history_sends_no_tool_artifacts() { |
| 1993 | let (addr, requests) = spawn_completion_stub("Ran git status on main."); |
| 1994 | let backend = OpenAiBackend::new( |
| 1995 | format!("http://{addr}/v1"), |
| 1996 | "test-key", |
| 1997 | "test-model", |
| 1998 | Some("be helpful".into()), |
| 1999 | ); |
| 2000 | { |
| 2001 | let mut history = backend.history.lock().await; |
| 2002 | history.push(serde_json::json!({ "role": "user", "content": "check the repo" })); |
| 2003 | history.push(serde_json::json!({ |
| 2004 | "role": "assistant", |
| 2005 | "content": null, |
| 2006 | "tool_calls": [{ |
| 2007 | "id": "call_1", |
| 2008 | "type": "function", |
| 2009 | "function": { "name": "run_command", "arguments": "{\"command\":\"git status\"}" }, |
| 2010 | }], |
| 2011 | })); |
| 2012 | history.push(serde_json::json!({ |
| 2013 | "role": "tool", |
| 2014 | "tool_call_id": "call_1", |
| 2015 | "content": "on branch main", |
| 2016 | })); |
| 2017 | } |
| 2018 | |
| 2019 | backend.compact_history(2).await.unwrap(); |
| 2020 | |
| 2021 | let body: serde_json::Value = |
| 2022 | serde_json::from_str(&requests.recv().unwrap()).expect("request body is JSON"); |
| 2023 | assert!( |
| 2024 | body.get("tools").is_none(), |
| 2025 | "summarization offers no tools: {body}" |
| 2026 | ); |
| 2027 | for message in body["messages"].as_array().unwrap() { |
| 2028 | assert!( |
| 2029 | message.get("tool_calls").is_none(), |
| 2030 | "no tool_calls may be sent without a schema: {message}" |
| 2031 | ); |
| 2032 | assert_ne!( |
| 2033 | message["role"], "tool", |
| 2034 | "no tool results may be sent without a schema: {message}" |
| 2035 | ); |
| 2036 | } |
| 2037 | |
| 2038 | // The tool round still has to survive into the summary request as prose, |
| 2039 | // or the summary loses the work the session actually did. |
| 2040 | let transcript = body["messages"].as_array().unwrap().last().unwrap()["content"] |
| 2041 | .as_str() |
| 2042 | .unwrap(); |
| 2043 | assert!(transcript.contains("called run_command({\"command\":\"git status\"})")); |
| 2044 | assert!(transcript.contains("tool result: on branch main")); |
| 2045 | } |
| 2046 | |
| 2047 | #[tokio::test] |
| 2048 | async fn compact_history_failure_leaves_history_intact() { |
| 2049 | // No listener at this address: the summarization request fails, and |
| 2050 | // history must roll back to exactly what it was. |
| 2051 | let backend = |
| 2052 | OpenAiBackend::new("http://127.0.0.1:9", "", "test-model", Some("sys".into())); |
| 2053 | backend |
| 2054 | .history |
| 2055 | .lock() |
| 2056 | .await |
| 2057 | .push(serde_json::json!({ "role": "user", "content": "hello" })); |
| 2058 | let before = backend.history_snapshot().await; |
| 2059 | |
| 2060 | assert!(backend.compact_history(2).await.is_err()); |
| 2061 | assert_eq!(backend.history_snapshot().await, before); |
| 2062 | } |
| 2063 | |
| 2064 | #[test] |
| 2065 | fn assistant_message_with_tool_calls_round_trips() { |
| 2066 | let message = ResponseMessage { |
| 2067 | content: None, |
| 2068 | tool_calls: Some(vec![ResponseToolCall { |
| 2069 | id: "call_1".to_string(), |
| 2070 | function: ResponseFunction { |
| 2071 | name: "read_file".to_string(), |
| 2072 | arguments: r#"{"path":"a.rs"}"#.to_string(), |
| 2073 | }, |
| 2074 | }]), |
| 2075 | }; |
| 2076 | let value = message.into_history_value(); |
| 2077 | assert_eq!(value["role"], "assistant"); |
| 2078 | assert!(value["content"].is_null()); |
| 2079 | assert_eq!(value["tool_calls"][0]["id"], "call_1"); |
| 2080 | assert_eq!(value["tool_calls"][0]["type"], "function"); |
| 2081 | assert_eq!(value["tool_calls"][0]["function"]["name"], "read_file"); |
| 2082 | } |
| 2083 | } |