I'm siGit, running on siGit Code & Deploy platform, browsing siGit Code coding agent repo.
development
| 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 | use std::sync::atomic::{AtomicBool, Ordering}; |
| 23 | |
| 24 | use async_trait::async_trait; |
| 25 | use onde::inference::{ChatEngine, ChatMessage, ChatRole, EngineStatus, ToolDefinition}; |
| 26 | use serde::Deserialize; |
| 27 | use tokio::sync::Mutex; |
| 28 | |
| 29 | // ── Neutral types ─────────────────────────────────────────────────────────────── |
| 30 | |
| 31 | /// A tool the model may call, in a provider-neutral form. `parameters_schema` is |
| 32 | /// a JSON Schema encoded as a string (matching how siGit Code already declares |
| 33 | /// tools). |
| 34 | #[derive(Debug, Clone)] |
| 35 | pub struct ToolSpec { |
| 36 | pub name: String, |
| 37 | pub description: String, |
| 38 | pub parameters_schema: String, |
| 39 | } |
| 40 | |
| 41 | /// A tool call requested by the model. |
| 42 | #[derive(Debug, Clone)] |
| 43 | pub struct ToolCall { |
| 44 | pub id: String, |
| 45 | pub name: String, |
| 46 | /// Arguments as a JSON-encoded string. |
| 47 | pub arguments: String, |
| 48 | } |
| 49 | |
| 50 | /// The output of executing one tool call, fed back to the model. |
| 51 | #[derive(Debug, Clone)] |
| 52 | pub struct ToolResult { |
| 53 | pub tool_call_id: String, |
| 54 | pub content: String, |
| 55 | } |
| 56 | |
| 57 | /// The result of one assistant turn: free text and/or tool calls. |
| 58 | #[derive(Debug, Clone, Default)] |
| 59 | pub struct TurnResult { |
| 60 | pub text: String, |
| 61 | pub tool_calls: Vec<ToolCall>, |
| 62 | /// Why the model stopped. Only the last round of a turn decides how the |
| 63 | /// turn is reported; a round that goes on to run tools is not the end. |
| 64 | pub finish: FinishReason, |
| 65 | /// The context in use once this reply is in, as the endpoint counted it: |
| 66 | /// prompt plus completion tokens. `None` when nothing was reported, which |
| 67 | /// is always the case on-device. |
| 68 | pub context_tokens: Option<u64>, |
| 69 | } |
| 70 | |
| 71 | /// Why the model stopped generating, reduced to what the agent loop acts on. |
| 72 | /// Endpoints report it as `finish_reason`; without it a reply cut off at the |
| 73 | /// token limit is indistinguishable from one the model finished. |
| 74 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Default)] |
| 75 | pub enum FinishReason { |
| 76 | /// The model finished on its own: `stop`, `tool_calls`, or no reason given. |
| 77 | #[default] |
| 78 | Complete, |
| 79 | /// The output ran into the token limit, so the reply is cut short. |
| 80 | Length, |
| 81 | /// The endpoint withheld the reply instead of answering. |
| 82 | ContentFilter, |
| 83 | /// The harness cut the reply off because it had degenerated into |
| 84 | /// repetition (see `repetition`). |
| 85 | Repetition, |
| 86 | } |
| 87 | |
| 88 | impl FinishReason { |
| 89 | /// Map a wire `finish_reason`. OpenAI's spellings are `length` and |
| 90 | /// `content_filter`; gateways in front of other providers sometimes pass |
| 91 | /// the upstream's own word through, so those are accepted too. |
| 92 | fn from_wire(reason: Option<&str>) -> Self { |
| 93 | match reason { |
| 94 | Some("length" | "max_tokens") => Self::Length, |
| 95 | Some("content_filter" | "refusal") => Self::ContentFilter, |
| 96 | _ => Self::Complete, |
| 97 | } |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | /// Backend errors are plain strings. Callers map them to ACP errors. |
| 102 | pub type BackendError = String; |
| 103 | |
| 104 | /// Rough context budget for a conversation, in estimated tokens (see |
| 105 | /// [`estimate_tokens`]). When a snapshot exceeds this, the agent loops compact |
| 106 | /// history before the next tool round. |
| 107 | pub const DEFAULT_CONTEXT_TOKEN_BUDGET: usize = 24_000; |
| 108 | |
| 109 | /// How many trailing messages survive a compaction verbatim (the rest are |
| 110 | /// folded into the summary). |
| 111 | pub const COMPACT_KEEP_LAST: usize = 6; |
| 112 | |
| 113 | /// The summarization request sent to the model when compacting history. |
| 114 | const SUMMARIZE_PROMPT: &str = "Summarize this coding session so far: decisions made, \ |
| 115 | files touched, current state, open items. Be concise and factual."; |
| 116 | |
| 117 | /// Largest transcript (in estimated tokens, see [`estimate_tokens`]) sent to |
| 118 | /// the model for summarization in one go. |
| 119 | /// |
| 120 | /// Compaction runs because the conversation has outgrown the context budget, |
| 121 | /// so shipping the whole transcript verbatim asks the endpoint to process |
| 122 | /// exactly the payload that is already too big — on siGit Code Cloud that |
| 123 | /// request sits in inference until the gateway gives up (504), which is the |
| 124 | /// failure behind issue #125. Capping the transcript keeps the summarization |
| 125 | /// round inside the window the model actually has; the oldest and newest |
| 126 | /// exchanges carry the intent and the current state, while the truncated |
| 127 | /// middle is the part a summary can afford to lose. |
| 128 | const SUMMARY_TRANSCRIPT_TOKEN_CAP: usize = 12_000; |
| 129 | |
| 130 | /// Trim `transcript` to about `cap` estimated tokens by cutting whole lines |
| 131 | /// from the middle, keeping the opening and closing exchanges. |
| 132 | /// |
| 133 | /// A middle cut preserves the two ends that matter most to a summary: how the |
| 134 | /// session started (the original request) and where it stands now. Cutting by |
| 135 | /// whole lines keeps individual messages intact. A transcript already under |
| 136 | /// the cap is returned unchanged. |
| 137 | /// |
| 138 | /// Line boundaries are best-effort. A head with no newline in its half is |
| 139 | /// dropped and the tail keeps its own half. A tail with no newline in its half |
| 140 | /// (a single huge message, or a giant final tool result) can't be cut on a |
| 141 | /// line, so the newest `cap` worth of text is kept instead and the head goes: |
| 142 | /// the newest exchanges are the ones the summary can least afford to lose. |
| 143 | fn truncate_transcript_middle(transcript: &str, cap: usize) -> String { |
| 144 | let budget = cap * 4; // estimate_tokens is chars / 4; invert it. |
| 145 | if transcript.len() <= budget { |
| 146 | return transcript.to_string(); |
| 147 | } |
| 148 | |
| 149 | let total = transcript.len(); |
| 150 | // Each side gets half the budget, so the two together stay inside the cap. |
| 151 | // Walk each half to a line boundary. A head with no boundary in reach |
| 152 | // contributes nothing; a tail with none takes the hard cut below. |
| 153 | let half = budget / 2; |
| 154 | let head_end = { |
| 155 | let mut cut = half.min(total); |
| 156 | while cut < total && !transcript.is_char_boundary(cut) { |
| 157 | cut += 1; |
| 158 | } |
| 159 | cut = cut.min(total); |
| 160 | if transcript[..cut].contains('\n') { |
| 161 | transcript[..cut].rfind('\n').unwrap() + 1 |
| 162 | } else { |
| 163 | 0 |
| 164 | } |
| 165 | }; |
| 166 | let tail_start = { |
| 167 | let mut cut = total.saturating_sub(half); |
| 168 | while cut < total && !transcript.is_char_boundary(cut) { |
| 169 | cut += 1; |
| 170 | } |
| 171 | if transcript[cut..].contains('\n') { |
| 172 | transcript[cut..] |
| 173 | .find('\n') |
| 174 | .map(|offset| cut + offset + 1) |
| 175 | .unwrap() |
| 176 | } else { |
| 177 | total |
| 178 | } |
| 179 | }; |
| 180 | |
| 181 | if tail_start == total { |
| 182 | // The closing exchange has no line boundary in reach (a transcript of |
| 183 | // one huge message, or a giant newline-less tool result at the end): |
| 184 | // keep the newest budget. A head-only cut would invert the contract |
| 185 | // while the placeholder claims older messages were dropped. |
| 186 | let mut start = total.saturating_sub(budget); |
| 187 | while start > 0 && !transcript.is_char_boundary(start) { |
| 188 | start -= 1; |
| 189 | } |
| 190 | return format!( |
| 191 | "{}\n[…older messages omitted…]", |
| 192 | transcript[start..].trim_start() |
| 193 | ); |
| 194 | } |
| 195 | |
| 196 | if head_end >= tail_start { |
| 197 | // No line boundary in reach anywhere: keep the newest budget. |
| 198 | let mut start = total.saturating_sub(budget); |
| 199 | while start > 0 && !transcript.is_char_boundary(start) { |
| 200 | start -= 1; |
| 201 | } |
| 202 | return format!( |
| 203 | "{}\n[…older messages omitted…]", |
| 204 | transcript[start..].trim_start() |
| 205 | ); |
| 206 | } |
| 207 | |
| 208 | let omitted_chars = transcript[head_end..tail_start].chars().count(); |
| 209 | format!( |
| 210 | "{}\n[…{} characters of older messages omitted…]\n{}", |
| 211 | transcript[..head_end].trim_end(), |
| 212 | omitted_chars, |
| 213 | transcript[tail_start..].trim_start(), |
| 214 | ) |
| 215 | } |
| 216 | |
| 217 | /// Render a history snapshot as a plain-text transcript, with tool calls and |
| 218 | /// tool results spelled out as prose rather than left in their wire shapes. |
| 219 | /// |
| 220 | /// Compaction summarizes a conversation that is, by definition, thick with |
| 221 | /// `tool_calls` and `role: "tool"` messages — but the summarization round asks |
| 222 | /// for a plain answer and so sends no `tools` array. Forwarding the raw shapes |
| 223 | /// in that request produces tool blocks with no schema to validate against, |
| 224 | /// which strict endpoints reject outright: Anthropic answers 400, so every |
| 225 | /// compaction of a session that had ever run a tool failed, permanently, no |
| 226 | /// matter how small the history was. Flattening to text keeps everything the |
| 227 | /// summary actually needs and drops the shapes that only make sense alongside |
| 228 | /// a tool schema. It also sidesteps orphaned `tool_call_id`s and role- |
| 229 | /// alternation rules, neither of which a transcript can violate. |
| 230 | fn transcript_for_summary(history: &[serde_json::Value]) -> String { |
| 231 | let mut lines: Vec<String> = Vec::new(); |
| 232 | for message in history { |
| 233 | let role = message["role"].as_str().unwrap_or("user"); |
| 234 | // The system prompt is carried over verbatim, so it needn't be summarized. |
| 235 | if role == "system" { |
| 236 | continue; |
| 237 | } |
| 238 | |
| 239 | let mut parts: Vec<String> = Vec::new(); |
| 240 | let text = message_text(message); |
| 241 | if !text.trim().is_empty() { |
| 242 | parts.push(text); |
| 243 | } |
| 244 | // A summary cannot carry a picture, but it should say one was there. |
| 245 | match message_image_count(message) { |
| 246 | 0 => {} |
| 247 | 1 => parts.push("[attached an image]".to_string()), |
| 248 | count => parts.push(format!("[attached {count} images]")), |
| 249 | } |
| 250 | for call in message["tool_calls"].as_array().into_iter().flatten() { |
| 251 | parts.push(format!( |
| 252 | "called {}({})", |
| 253 | call["function"]["name"].as_str().unwrap_or("tool"), |
| 254 | call["function"]["arguments"].as_str().unwrap_or_default(), |
| 255 | )); |
| 256 | } |
| 257 | if parts.is_empty() { |
| 258 | continue; |
| 259 | } |
| 260 | |
| 261 | let label = if role == "tool" { "tool result" } else { role }; |
| 262 | lines.push(format!("{label}: {}", parts.join("\n"))); |
| 263 | } |
| 264 | lines.join("\n\n") |
| 265 | } |
| 266 | |
| 267 | /// Crude token estimate for a history snapshot: serialized characters / 4. |
| 268 | /// Deliberately model-agnostic — it only needs to be in the right ballpark to |
| 269 | /// decide when compaction is worth an extra inference round. |
| 270 | pub fn estimate_tokens(history: &[serde_json::Value]) -> usize { |
| 271 | let mut chars = 0; |
| 272 | let mut images = 0; |
| 273 | for message in history { |
| 274 | let image_count = message_image_count(message); |
| 275 | if image_count == 0 { |
| 276 | chars += message.to_string().chars().count(); |
| 277 | } else { |
| 278 | // An image travels as base64, megabytes of it, but a model bills |
| 279 | // it by resolution. Counting the payload as text would put every |
| 280 | // session with one screenshot permanently over budget. |
| 281 | chars += message_text(message).chars().count(); |
| 282 | images += image_count; |
| 283 | } |
| 284 | } |
| 285 | chars / 4 + images * IMAGE_TOKEN_ESTIMATE |
| 286 | } |
| 287 | |
| 288 | /// What one attached image is assumed to cost. Providers land between a few |
| 289 | /// hundred and a couple of thousand tokens depending on resolution; like the |
| 290 | /// rest of [`estimate_tokens`] this only has to be in the right ballpark. |
| 291 | const IMAGE_TOKEN_ESTIMATE: usize = 1_000; |
| 292 | |
| 293 | /// The text of a history message. `content` is a plain string for almost every |
| 294 | /// message; a user message that carries an image uses OpenAI's content-part |
| 295 | /// array instead, and its text parts are joined here. |
| 296 | pub fn message_text(message: &serde_json::Value) -> String { |
| 297 | match &message["content"] { |
| 298 | serde_json::Value::String(text) => text.clone(), |
| 299 | serde_json::Value::Array(parts) => parts |
| 300 | .iter() |
| 301 | .filter_map(|part| part["text"].as_str()) |
| 302 | .collect::<Vec<_>>() |
| 303 | .join("\n"), |
| 304 | _ => String::new(), |
| 305 | } |
| 306 | } |
| 307 | |
| 308 | /// How many content parts of `part_type` a history message carries. |
| 309 | fn message_part_count(message: &serde_json::Value, part_type: &str) -> usize { |
| 310 | message["content"] |
| 311 | .as_array() |
| 312 | .map(|parts| { |
| 313 | parts |
| 314 | .iter() |
| 315 | .filter(|part| part["type"] == part_type) |
| 316 | .count() |
| 317 | }) |
| 318 | .unwrap_or(0) |
| 319 | } |
| 320 | |
| 321 | /// How many images a history message carries as `image_url` content parts. |
| 322 | pub fn message_image_count(message: &serde_json::Value) -> usize { |
| 323 | message_part_count(message, "image_url") |
| 324 | } |
| 325 | |
| 326 | /// How many audio clips a history message carries as `input_audio` content |
| 327 | /// parts. |
| 328 | pub fn message_audio_count(message: &serde_json::Value) -> usize { |
| 329 | message_part_count(message, "input_audio") |
| 330 | } |
| 331 | |
| 332 | /// One streamed fragment, tagged by whether it is visible prose or the model's |
| 333 | /// reasoning. |
| 334 | /// |
| 335 | /// Reasoning reaches a backend two ways: inline `<think>…</think>` tags inside |
| 336 | /// the ordinary content stream (local models, and some OpenAI-compatible |
| 337 | /// endpoints), or a separate `reasoning_content` delta field (DeepSeek-style |
| 338 | /// endpoints). Inline tags can't be classified at the source — a tag may span |
| 339 | /// chunk boundaries — so they travel as [`TokenChunk::Visible`] and are split |
| 340 | /// out downstream by `StreamedReply`. The `reasoning_content` field *is* |
| 341 | /// unambiguous at the source, so the backend tags it [`TokenChunk::Reasoning`] |
| 342 | /// directly. |
| 343 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 344 | pub enum TokenChunk { |
| 345 | /// Content-stream text. May still contain inline `<think>` tags that a |
| 346 | /// downstream consumer splits into reasoning and prose. |
| 347 | Visible(String), |
| 348 | /// Text from a dedicated reasoning channel, already known to be reasoning. |
| 349 | Reasoning(String), |
| 350 | } |
| 351 | |
| 352 | /// A sink for streaming assistant fragments to the UI as they are produced. |
| 353 | /// |
| 354 | /// When a caller passes `Some(sink)`, a streaming-capable backend forwards each |
| 355 | /// fragment through it as the model emits it, tagged [`TokenChunk::Visible`] or |
| 356 | /// [`TokenChunk::Reasoning`]; the returned [`TurnResult`] still carries the |
| 357 | /// fully assembled (visible) text and any tool calls. When the sink is `None`, |
| 358 | /// the backend runs in non-streaming mode. Unbounded so the inference task |
| 359 | /// never blocks on a slow consumer. |
| 360 | pub type TokenSink = tokio::sync::mpsc::UnboundedSender<TokenChunk>; |
| 361 | |
| 362 | /// An image attached to a user message: base64 data and its media type, as an |
| 363 | /// ACP client sends it. |
| 364 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 365 | pub struct ImageInput { |
| 366 | pub mime_type: String, |
| 367 | pub data: String, |
| 368 | } |
| 369 | |
| 370 | /// An audio clip attached to a user message: base64 data and its media type, |
| 371 | /// as an ACP client sends it. |
| 372 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 373 | pub struct AudioInput { |
| 374 | pub mime_type: String, |
| 375 | pub data: String, |
| 376 | } |
| 377 | |
| 378 | impl AudioInput { |
| 379 | /// The `format` OpenAI's `input_audio` part expects (`wav`, `mp3`, ...), |
| 380 | /// from the MIME type an ACP client sends (`audio/wav`, `audio/mpeg`, ...). |
| 381 | fn format(&self) -> String { |
| 382 | let subtype = self |
| 383 | .mime_type |
| 384 | .split(';') |
| 385 | .next() |
| 386 | .unwrap_or_default() |
| 387 | .trim() |
| 388 | .to_lowercase(); |
| 389 | let subtype = subtype.strip_prefix("audio/").unwrap_or(&subtype); |
| 390 | match subtype.strip_prefix("x-").unwrap_or(subtype) { |
| 391 | "mpeg" | "mp3" => "mp3".to_string(), |
| 392 | "wave" | "wav" | "vnd.wave" => "wav".to_string(), |
| 393 | other => other.to_string(), |
| 394 | } |
| 395 | } |
| 396 | } |
| 397 | |
| 398 | /// What a model that cannot read images is shown in place of one. |
| 399 | const IMAGE_OMITTED_NOTE: &str = "[image omitted: this model cannot read images]"; |
| 400 | |
| 401 | /// What a model that cannot take audio is shown in place of a clip. |
| 402 | const AUDIO_OMITTED_NOTE: &str = "[audio omitted: this model cannot take audio]"; |
| 403 | |
| 404 | /// A user message in history form. Plain text keeps the string `content` every |
| 405 | /// other message uses; with attachments it becomes OpenAI's content-part |
| 406 | /// array, text first, each image as a base64 `data:` URL and each audio clip |
| 407 | /// as an `input_audio` part. |
| 408 | fn user_message(text: &str, images: &[ImageInput], audio: &[AudioInput]) -> serde_json::Value { |
| 409 | if images.is_empty() && audio.is_empty() { |
| 410 | return serde_json::json!({ "role": "user", "content": text }); |
| 411 | } |
| 412 | let mut parts = Vec::with_capacity(images.len() + audio.len() + 1); |
| 413 | if !text.is_empty() { |
| 414 | parts.push(serde_json::json!({ "type": "text", "text": text })); |
| 415 | } |
| 416 | for image in images { |
| 417 | parts.push(serde_json::json!({ |
| 418 | "type": "image_url", |
| 419 | "image_url": { |
| 420 | "url": format!("data:{};base64,{}", image.mime_type, image.data), |
| 421 | }, |
| 422 | })); |
| 423 | } |
| 424 | for clip in audio { |
| 425 | parts.push(serde_json::json!({ |
| 426 | "type": "input_audio", |
| 427 | "input_audio": { "data": clip.data, "format": clip.format() }, |
| 428 | })); |
| 429 | } |
| 430 | serde_json::json!({ "role": "user", "content": parts }) |
| 431 | } |
| 432 | |
| 433 | /// `history` with every content part of `part_type` replaced by `note`, for a |
| 434 | /// model that cannot take that kind of attachment. A message left with only |
| 435 | /// text is flattened back to a plain string. |
| 436 | /// |
| 437 | /// History itself keeps the attachments. A thread can move to a model that |
| 438 | /// reads them (or back to one), so what was attached is not thrown away just |
| 439 | /// because the model active right now cannot use it. |
| 440 | fn without_parts( |
| 441 | history: &[serde_json::Value], |
| 442 | part_type: &str, |
| 443 | note: &str, |
| 444 | ) -> Vec<serde_json::Value> { |
| 445 | history |
| 446 | .iter() |
| 447 | .map(|message| { |
| 448 | let Some(parts) = message["content"].as_array() else { |
| 449 | return message.clone(); |
| 450 | }; |
| 451 | if !parts.iter().any(|part| part["type"] == part_type) { |
| 452 | return message.clone(); |
| 453 | } |
| 454 | let mut kept: Vec<serde_json::Value> = Vec::with_capacity(parts.len()); |
| 455 | let mut notes = Vec::new(); |
| 456 | for part in parts { |
| 457 | if part["type"] == part_type { |
| 458 | notes.push(serde_json::json!({ "type": "text", "text": note })); |
| 459 | } else { |
| 460 | kept.push(part.clone()); |
| 461 | } |
| 462 | } |
| 463 | kept.extend(notes); |
| 464 | let mut stripped = message.clone(); |
| 465 | stripped["content"] = if kept.iter().all(|part| part["type"] == "text") { |
| 466 | let text = kept |
| 467 | .iter() |
| 468 | .filter_map(|part| part["text"].as_str()) |
| 469 | .filter(|text| !text.is_empty()) |
| 470 | .collect::<Vec<_>>() |
| 471 | .join("\n"); |
| 472 | serde_json::Value::String(text) |
| 473 | } else { |
| 474 | serde_json::Value::Array(kept) |
| 475 | }; |
| 476 | stripped |
| 477 | }) |
| 478 | .collect() |
| 479 | } |
| 480 | |
| 481 | /// `history` as a model that cannot read images has to receive it, with a |
| 482 | /// note where each image was. |
| 483 | fn without_images(history: &[serde_json::Value]) -> Vec<serde_json::Value> { |
| 484 | without_parts(history, "image_url", IMAGE_OMITTED_NOTE) |
| 485 | } |
| 486 | |
| 487 | /// `history` as a model that cannot take audio has to receive it, with a note |
| 488 | /// where each clip was. |
| 489 | fn without_audio(history: &[serde_json::Value]) -> Vec<serde_json::Value> { |
| 490 | without_parts(history, "input_audio", AUDIO_OMITTED_NOTE) |
| 491 | } |
| 492 | |
| 493 | // ── The trait ─────────────────────────────────────────────────────────────────── |
| 494 | |
| 495 | /// A swappable inference backend driving siGit Code's agent loop. |
| 496 | #[async_trait] |
| 497 | pub trait InferenceBackend: Send + Sync { |
| 498 | /// Start an assistant turn from a new user message, offering `tools`. |
| 499 | /// |
| 500 | /// If `sink` is `Some`, text is streamed through it as it is generated. A |
| 501 | /// backend may decline to stream a given round (for example, on-device |
| 502 | /// inference cannot stream while it is still deciding whether to call a |
| 503 | /// tool); in that case the text is delivered only via the returned result. |
| 504 | async fn send_message_with_tools( |
| 505 | &self, |
| 506 | text: &str, |
| 507 | tools: &[ToolSpec], |
| 508 | sink: Option<&TokenSink>, |
| 509 | ) -> Result<TurnResult, BackendError>; |
| 510 | |
| 511 | /// Continue the turn by returning tool results. `allow_tool_calls` controls |
| 512 | /// whether `tools` are offered again; the complete catalog remains available |
| 513 | /// so a disabled round can recognize and suppress tool-shaped model output. |
| 514 | /// `sink` streams assistant text when set. |
| 515 | async fn send_tool_results( |
| 516 | &self, |
| 517 | results: Vec<ToolResult>, |
| 518 | tools: &[ToolSpec], |
| 519 | allow_tool_calls: bool, |
| 520 | sink: Option<&TokenSink>, |
| 521 | ) -> Result<TurnResult, BackendError>; |
| 522 | |
| 523 | /// Record tool results in the conversation history *without* asking the |
| 524 | /// model to continue the turn. Used when a turn is abandoned mid-round |
| 525 | /// (the user cancelled at the permission gate): by then the assistant |
| 526 | /// message carrying the tool calls is already in history, and leaving them |
| 527 | /// unanswered makes strict OpenAI-compatible endpoints reject every later |
| 528 | /// request in the session. |
| 529 | async fn record_cancelled_tool_results(&self, results: Vec<ToolResult>); |
| 530 | |
| 531 | /// Start a turn from a user message that carries images or audio. |
| 532 | /// |
| 533 | /// The default drops the attachments and sends the text, which is right |
| 534 | /// for a backend that cannot read them. Callers check |
| 535 | /// [`Self::accepts_images`] and [`Self::accepts_audio`] first so the user |
| 536 | /// can be told, instead of the attachment vanishing. |
| 537 | async fn send_message_with_attachments( |
| 538 | &self, |
| 539 | text: &str, |
| 540 | images: &[ImageInput], |
| 541 | audio: &[AudioInput], |
| 542 | tools: &[ToolSpec], |
| 543 | sink: Option<&TokenSink>, |
| 544 | ) -> Result<TurnResult, BackendError> { |
| 545 | let _ = (images, audio); |
| 546 | self.send_message_with_tools(text, tools, sink).await |
| 547 | } |
| 548 | |
| 549 | /// Whether the model behind this backend reads images. On-device models do |
| 550 | /// not; a remote one answers from its model id. |
| 551 | fn accepts_images(&self) -> bool { |
| 552 | false |
| 553 | } |
| 554 | |
| 555 | /// Whether the model behind this backend takes audio. On-device models do |
| 556 | /// not; a remote one answers from its model id. |
| 557 | fn accepts_audio(&self) -> bool { |
| 558 | false |
| 559 | } |
| 560 | |
| 561 | /// Whether inference runs over the network (a configured provider) rather |
| 562 | /// than on-device. Drives UI labelling so the displayed model can't claim a |
| 563 | /// local model while requests actually go to the cloud. |
| 564 | fn is_remote(&self) -> bool; |
| 565 | |
| 566 | /// A backend for the same endpoint and model with an empty conversation, |
| 567 | /// or `None` when the conversation cannot be separated from the backend |
| 568 | /// (on-device, the engine holds it). An editor keeps several threads open |
| 569 | /// in one process; this is what lets each of them own its history instead |
| 570 | /// of taking turns on a shared one. |
| 571 | fn fresh(&self) -> Option<Arc<dyn InferenceBackend>> { |
| 572 | None |
| 573 | } |
| 574 | |
| 575 | /// A serializable snapshot of the conversation history, one JSON object per |
| 576 | /// message (`{"role": ..., "content": ...}` at minimum). The snapshot is |
| 577 | /// what the session store persists; it includes any seeded system message |
| 578 | /// so [`InferenceBackend::restore_history`] can replace state wholesale. |
| 579 | async fn history_snapshot(&self) -> Vec<serde_json::Value>; |
| 580 | |
| 581 | /// Replace the conversation history with a previously saved snapshot. |
| 582 | /// Backends that cannot represent every entry (e.g. on-device history has |
| 583 | /// no tool-call structure) flatten what they can and drop the rest. |
| 584 | async fn restore_history(&self, history: Vec<serde_json::Value>); |
| 585 | |
| 586 | /// Shrink the conversation history: summarize everything so far with one |
| 587 | /// extra (non-streaming) inference round, then rebuild history as |
| 588 | /// `[system message, summary, last keep_last non-system messages]`. When |
| 589 | /// the summarization round itself fails, a backend should fall back to a |
| 590 | /// deterministic shrink (dropping the oldest messages) rather than leave |
| 591 | /// an over-budget session stuck; an `Err` here means even that fallback |
| 592 | /// could not fit the history in the budget. |
| 593 | async fn compact_history(&self, keep_last: usize) -> Result<(), BackendError>; |
| 594 | } |
| 595 | |
| 596 | // ── Local backend (onde ChatEngine) ────────────────────────────────────────────── |
| 597 | |
| 598 | /// On-device inference. A thin adapter over `onde::ChatEngine`. |
| 599 | pub struct LocalBackend { |
| 600 | engine: Arc<ChatEngine>, |
| 601 | /// History restored before an on-device model is loaded. Onde ignores |
| 602 | /// `push_history` while unloaded, so keep the snapshot here until the |
| 603 | /// engine can accept it instead of silently dropping a loaded session. |
| 604 | pending_history: Mutex<Option<Vec<serde_json::Value>>>, |
| 605 | } |
| 606 | |
| 607 | impl LocalBackend { |
| 608 | pub fn new(engine: Arc<ChatEngine>) -> Self { |
| 609 | Self { |
| 610 | engine, |
| 611 | pending_history: Mutex::new(None), |
| 612 | } |
| 613 | } |
| 614 | |
| 615 | async fn apply_pending_history(&self) -> Result<(), BackendError> { |
| 616 | if self.pending_history.lock().await.is_none() { |
| 617 | return Ok(()); |
| 618 | } |
| 619 | let status = self.engine.info().await.status; |
| 620 | if status != EngineStatus::Ready { |
| 621 | return Err(format!( |
| 622 | "on-device model is not ready to restore session history (status: {status})" |
| 623 | )); |
| 624 | } |
| 625 | let Some(history) = self.pending_history.lock().await.take() else { |
| 626 | return Ok(()); |
| 627 | }; |
| 628 | |
| 629 | self.engine.clear_history().await; |
| 630 | for entry in history { |
| 631 | let role = entry["role"].as_str().unwrap_or(""); |
| 632 | // The on-device engine keeps text only, so an image attached |
| 633 | // while a cloud model was active stays behind here. |
| 634 | let content = message_text(&entry); |
| 635 | // Tool-call-only assistant entries and empty tool results carry no |
| 636 | // text a plain chat history can replay; drop them. |
| 637 | if content.is_empty() && role != "user" && role != "system" { |
| 638 | continue; |
| 639 | } |
| 640 | let message = match role { |
| 641 | "system" => ChatMessage::system(content), |
| 642 | "user" => ChatMessage::user(content), |
| 643 | "assistant" => ChatMessage::assistant(content), |
| 644 | // Tool results flatten to plain text (MVP; acceptable loss). |
| 645 | "tool" => ChatMessage::user(format!("[tool result]\n{content}")), |
| 646 | _ => continue, |
| 647 | }; |
| 648 | self.engine.push_history(message).await; |
| 649 | } |
| 650 | Ok(()) |
| 651 | } |
| 652 | } |
| 653 | |
| 654 | fn to_onde_tools(tools: &[ToolSpec]) -> Vec<ToolDefinition> { |
| 655 | tools |
| 656 | .iter() |
| 657 | .map(|tool| ToolDefinition { |
| 658 | name: tool.name.clone(), |
| 659 | description: tool.description.clone(), |
| 660 | parameters_schema: tool.parameters_schema.clone(), |
| 661 | }) |
| 662 | .collect() |
| 663 | } |
| 664 | |
| 665 | #[async_trait] |
| 666 | impl InferenceBackend for LocalBackend { |
| 667 | async fn send_message_with_tools( |
| 668 | &self, |
| 669 | text: &str, |
| 670 | tools: &[ToolSpec], |
| 671 | sink: Option<&TokenSink>, |
| 672 | ) -> Result<TurnResult, BackendError> { |
| 673 | self.apply_pending_history().await?; |
| 674 | // onde's tool-aware path is non-streaming: it has to buffer the whole |
| 675 | // reply to detect tool calls. We can only stream when no tools are on |
| 676 | // offer (a plain answer), which is exactly the tools-disabled case. |
| 677 | if let Some(sink) = sink |
| 678 | && tools.is_empty() |
| 679 | { |
| 680 | let rx = self |
| 681 | .engine |
| 682 | .stream_message(text) |
| 683 | .await |
| 684 | .map_err(|error| error.to_string())?; |
| 685 | return drain_onde_stream(rx, sink).await; |
| 686 | } |
| 687 | |
| 688 | let onde_tools = to_onde_tools(tools); |
| 689 | let result = self |
| 690 | .engine |
| 691 | .send_message_with_tools(text, &onde_tools) |
| 692 | .await |
| 693 | .map_err(|error| error.to_string())?; |
| 694 | Ok(onde_result_to_turn(result)) |
| 695 | } |
| 696 | |
| 697 | async fn send_tool_results( |
| 698 | &self, |
| 699 | results: Vec<ToolResult>, |
| 700 | tools: &[ToolSpec], |
| 701 | allow_tool_calls: bool, |
| 702 | sink: Option<&TokenSink>, |
| 703 | ) -> Result<TurnResult, BackendError> { |
| 704 | self.apply_pending_history().await?; |
| 705 | let onde_results: Vec<onde::inference::ToolResult> = results |
| 706 | .into_iter() |
| 707 | .map(|result| onde::inference::ToolResult { |
| 708 | tool_call_id: result.tool_call_id, |
| 709 | content: result.content, |
| 710 | }) |
| 711 | .collect(); |
| 712 | |
| 713 | // A forced-text round is the only round onde can stream, since no |
| 714 | // further tool calls are parsed. |
| 715 | if let Some(sink) = sink |
| 716 | && !allow_tool_calls |
| 717 | { |
| 718 | let rx = self |
| 719 | .engine |
| 720 | .stream_tool_results(onde_results, None) |
| 721 | .await |
| 722 | .map_err(|error| error.to_string())?; |
| 723 | return drain_onde_stream(rx, sink).await; |
| 724 | } |
| 725 | |
| 726 | let onde_tools = allow_tool_calls.then(|| to_onde_tools(tools)); |
| 727 | let result = self |
| 728 | .engine |
| 729 | .send_tool_results(onde_results, onde_tools.as_deref()) |
| 730 | .await |
| 731 | .map_err(|error| error.to_string())?; |
| 732 | Ok(onde_result_to_turn(result)) |
| 733 | } |
| 734 | |
| 735 | async fn record_cancelled_tool_results(&self, _results: Vec<ToolResult>) { |
| 736 | // onde's public API cannot append tool-result history entries without |
| 737 | // running another inference round, so the dangling tool call stays in |
| 738 | // its history. The chat template replays it as-is, which local models |
| 739 | // tolerate — worst case the model re-issues the call next turn. |
| 740 | } |
| 741 | |
| 742 | fn is_remote(&self) -> bool { |
| 743 | false |
| 744 | } |
| 745 | |
| 746 | async fn history_snapshot(&self) -> Vec<serde_json::Value> { |
| 747 | if let Some(history) = self.pending_history.lock().await.as_ref() { |
| 748 | return history.clone(); |
| 749 | } |
| 750 | // onde's `history()` already flattens tool entries: assistant tool |
| 751 | // calls become plain assistant text and tool results are omitted, so |
| 752 | // the snapshot is lossy for tool-heavy turns (acceptable in this MVP). |
| 753 | self.engine |
| 754 | .history() |
| 755 | .await |
| 756 | .iter() |
| 757 | .map(|message| { |
| 758 | serde_json::json!({ |
| 759 | "role": message.role.to_string(), |
| 760 | "content": message.content, |
| 761 | }) |
| 762 | }) |
| 763 | .collect() |
| 764 | } |
| 765 | |
| 766 | async fn restore_history(&self, history: Vec<serde_json::Value>) { |
| 767 | *self.pending_history.lock().await = Some(history); |
| 768 | if self.engine.info().await.status == EngineStatus::Ready { |
| 769 | // Ready was just observed, so failure here can only mean a status |
| 770 | // transition; keep the pending copy for the next inference call. |
| 771 | let _ = self.apply_pending_history().await; |
| 772 | } |
| 773 | } |
| 774 | |
| 775 | async fn compact_history(&self, keep_last: usize) -> Result<(), BackendError> { |
| 776 | self.apply_pending_history().await?; |
| 777 | let snapshot = self.engine.history().await; |
| 778 | |
| 779 | // Leading system messages carry session context; keep them all. |
| 780 | let system: Vec<ChatMessage> = snapshot |
| 781 | .iter() |
| 782 | .take_while(|message| message.role == ChatRole::System) |
| 783 | .cloned() |
| 784 | .collect(); |
| 785 | let non_system: Vec<ChatMessage> = snapshot |
| 786 | .into_iter() |
| 787 | .filter(|message| message.role != ChatRole::System) |
| 788 | .collect(); |
| 789 | let tail_start = non_system.len().saturating_sub(keep_last); |
| 790 | let tail: Vec<ChatMessage> = non_system[tail_start..].to_vec(); |
| 791 | |
| 792 | // One plain (tool-free) inference round produces the summary. Note the |
| 793 | // cap only bounds this request's own payload: the engine still replays |
| 794 | // its full (over-budget) history underneath it, so an already |
| 795 | // over-budget local session usually fails here and lands in the |
| 796 | // fallback below — the cap is about not making the doomed round worse. |
| 797 | let transcript = truncate_transcript_middle( |
| 798 | &chat_messages_transcript(&non_system), |
| 799 | SUMMARY_TRANSCRIPT_TOKEN_CAP, |
| 800 | ); |
| 801 | let result = self |
| 802 | .engine |
| 803 | .send_message(format!("{transcript}\n\n{SUMMARIZE_PROMPT}")) |
| 804 | .await |
| 805 | .map_err(|error| error.to_string()); |
| 806 | |
| 807 | let rebuild: Vec<ChatMessage> = match result { |
| 808 | Ok(result) => { |
| 809 | // Local models may reason in <think> blocks; keep only the visible part. |
| 810 | let (_think, summary) = crate::chat::strip_think_blocks(&result.text); |
| 811 | let mut rebuilt = system; |
| 812 | rebuilt.push(ChatMessage::user(format!( |
| 813 | "[Conversation summary]\n{summary}" |
| 814 | ))); |
| 815 | rebuilt.extend(tail); |
| 816 | rebuilt |
| 817 | } |
| 818 | Err(error) => { |
| 819 | log::warn!( |
| 820 | "local summarization round failed ({error}); falling back to plain truncation" |
| 821 | ); |
| 822 | // Same deterministic shrink as the remote path, on the plain |
| 823 | // chat shapes the engine holds (its history has no tool-call |
| 824 | // structure, so there are no orphaned tool results to guard). |
| 825 | match truncate_chat_messages_to_budget(system, tail) { |
| 826 | Some(rebuilt) => rebuilt, |
| 827 | // The system messages alone overflow the window: unrecoverable. |
| 828 | None => return Err(error), |
| 829 | } |
| 830 | } |
| 831 | }; |
| 832 | |
| 833 | self.engine.clear_history().await; |
| 834 | for message in rebuild { |
| 835 | self.engine.push_history(message).await; |
| 836 | } |
| 837 | Ok(()) |
| 838 | } |
| 839 | } |
| 840 | |
| 841 | /// The local counterpart of `truncate_history_to_budget`: system messages, the |
| 842 | /// truncation placeholder, and the newest of `tail` that fit the budget. |
| 843 | /// |
| 844 | /// The tail may drain to empty. A newest message that overflows the window on |
| 845 | /// its own (a giant tool result, the very thing that triggers compaction) is |
| 846 | /// dropped rather than failing the session, so `None` only means the system |
| 847 | /// messages alone don't fit. |
| 848 | fn truncate_chat_messages_to_budget( |
| 849 | system: Vec<ChatMessage>, |
| 850 | mut tail: Vec<ChatMessage>, |
| 851 | ) -> Option<Vec<ChatMessage>> { |
| 852 | let budget = DEFAULT_CONTEXT_TOKEN_BUDGET * 4; // estimate_tokens is chars / 4. |
| 853 | let fixed_chars: usize = system |
| 854 | .iter() |
| 855 | .map(|message| message.content.len()) |
| 856 | .sum::<usize>() |
| 857 | + TRUNCATION_PLACEHOLDER.len(); |
| 858 | if fixed_chars > budget { |
| 859 | return None; |
| 860 | } |
| 861 | |
| 862 | let mut tail_chars: usize = tail.iter().map(|message| message.content.len()).sum(); |
| 863 | let mut drop = 0; |
| 864 | while fixed_chars + tail_chars > budget { |
| 865 | tail_chars -= tail[drop].content.len(); |
| 866 | drop += 1; |
| 867 | } |
| 868 | tail.drain(..drop); |
| 869 | |
| 870 | let mut rebuilt = system; |
| 871 | rebuilt.push(ChatMessage::user(TRUNCATION_PLACEHOLDER.to_string())); |
| 872 | rebuilt.extend(tail); |
| 873 | Some(rebuilt) |
| 874 | } |
| 875 | |
| 876 | /// Flatten `onde` chat messages into the transcript shape summarization |
| 877 | /// expects (the local engine's history has no tool-call structure, so this is |
| 878 | /// a plain role/content rendering). |
| 879 | fn chat_messages_transcript(history: &[ChatMessage]) -> String { |
| 880 | history |
| 881 | .iter() |
| 882 | .filter(|message| message.role != ChatRole::System && !message.content.trim().is_empty()) |
| 883 | .map(|message| format!("{}: {}", message.role, message.content)) |
| 884 | .collect::<Vec<_>>() |
| 885 | .join("\n\n") |
| 886 | } |
| 887 | |
| 888 | /// Drain an onde streaming receiver, forwarding each token to `sink` and |
| 889 | /// assembling the full text. onde reports stream failures as a final chunk whose |
| 890 | /// `finish_reason` is `"error: …"`; surface those as a backend error. |
| 891 | async fn drain_onde_stream( |
| 892 | mut rx: tokio::sync::mpsc::Receiver<onde::inference::StreamChunk>, |
| 893 | sink: &TokenSink, |
| 894 | ) -> Result<TurnResult, BackendError> { |
| 895 | let mut text = String::new(); |
| 896 | let mut finish = FinishReason::default(); |
| 897 | while let Some(chunk) = rx.recv().await { |
| 898 | if !chunk.delta.is_empty() { |
| 899 | text.push_str(&chunk.delta); |
| 900 | // The receiver is the UI; if it's gone the turn is being cancelled, |
| 901 | // so stop assembling rather than spinning the model to completion. |
| 902 | // onde reports reasoning as inline `<think>` tags in `delta`, so it |
| 903 | // is visible content that a downstream consumer splits out. |
| 904 | if sink.send(TokenChunk::Visible(chunk.delta)).is_err() { |
| 905 | break; |
| 906 | } |
| 907 | } |
| 908 | if chunk.done { |
| 909 | if let Some(reason) = chunk.finish_reason.as_deref() |
| 910 | && let Some(message) = reason.strip_prefix("error: ") |
| 911 | { |
| 912 | return Err(message.to_string()); |
| 913 | } |
| 914 | finish = FinishReason::from_wire(chunk.finish_reason.as_deref()); |
| 915 | break; |
| 916 | } |
| 917 | } |
| 918 | Ok(TurnResult { |
| 919 | text, |
| 920 | tool_calls: Vec::new(), |
| 921 | finish, |
| 922 | context_tokens: None, |
| 923 | }) |
| 924 | } |
| 925 | |
| 926 | /// Convert an `onde` tool-aware result into the neutral [`TurnResult`]. |
| 927 | fn onde_result_to_turn(result: onde::inference::ToolAwareResult) -> TurnResult { |
| 928 | TurnResult { |
| 929 | finish: FinishReason::from_wire(Some(&result.finish_reason)), |
| 930 | context_tokens: None, |
| 931 | text: result.text, |
| 932 | tool_calls: result |
| 933 | .tool_calls |
| 934 | .into_iter() |
| 935 | .map(|call| ToolCall { |
| 936 | id: call.id, |
| 937 | name: call.function_name, |
| 938 | arguments: call.arguments, |
| 939 | }) |
| 940 | .collect(), |
| 941 | } |
| 942 | } |
| 943 | |
| 944 | // ── OpenAI-compatible backend ───────────────────────────────────────────────────── |
| 945 | |
| 946 | /// Inference against any OpenAI-compatible Chat Completions endpoint. |
| 947 | /// |
| 948 | /// Conversation state is held client-side and replayed on every request, so the |
| 949 | /// endpoint can be stateless. Standard OpenAI function-calling is used end to |
| 950 | /// end (`tools`, `choices[].message.tool_calls`, `role: "tool"` follow-ups). |
| 951 | pub struct OpenAiBackend { |
| 952 | base_url: String, |
| 953 | api_key: String, |
| 954 | /// Read the bearer from the stored session on each request instead of |
| 955 | /// using `api_key`; see `bearer`. |
| 956 | session_token: bool, |
| 957 | model: String, |
| 958 | http: reqwest::Client, |
| 959 | /// The full message list sent on each request (system + turns + tool results). |
| 960 | history: Mutex<Vec<serde_json::Value>>, |
| 961 | /// Whether `model` reads images (see `provider::model_accepts_images`). |
| 962 | /// When it does not, images already in `history` are left out of each |
| 963 | /// request rather than sent to an endpoint that would refuse them. |
| 964 | accepts_images: bool, |
| 965 | /// Whether `model` takes audio (see `provider::model_accepts_audio`), with |
| 966 | /// the same effect on audio already in `history`. |
| 967 | accepts_audio: bool, |
| 968 | /// Set once the endpoint has refused `stream_options`, so streamed |
| 969 | /// requests stop asking it for token usage. |
| 970 | stream_usage_refused: Arc<AtomicBool>, |
| 971 | } |
| 972 | |
| 973 | impl OpenAiBackend { |
| 974 | /// Build a backend for `{base_url, api_key, model}`, seeding the optional |
| 975 | /// system prompt. `base_url` should include the API root (e.g. ending in |
| 976 | /// `/v1`); the chat path is appended. |
| 977 | pub fn new( |
| 978 | base_url: impl Into<String>, |
| 979 | api_key: impl Into<String>, |
| 980 | model: impl Into<String>, |
| 981 | system_prompt: Option<String>, |
| 982 | ) -> Self { |
| 983 | let mut history = Vec::new(); |
| 984 | if let Some(prompt) = system_prompt { |
| 985 | history.push(serde_json::json!({ "role": "system", "content": prompt })); |
| 986 | } |
| 987 | let model = model.into(); |
| 988 | Self { |
| 989 | base_url: base_url.into(), |
| 990 | api_key: api_key.into(), |
| 991 | session_token: false, |
| 992 | accepts_images: crate::provider::model_accepts_images(&model), |
| 993 | accepts_audio: crate::provider::model_accepts_audio(&model), |
| 994 | model, |
| 995 | http: reqwest::Client::new(), |
| 996 | history: Mutex::new(history), |
| 997 | stream_usage_refused: Arc::default(), |
| 998 | } |
| 999 | } |
| 1000 | |
| 1001 | /// Have the backend send the stored siGit Code Cloud session token as its |
| 1002 | /// bearer, read on every request. A token captured at build time goes |
| 1003 | /// stale when the user signs in as someone else; `api_key` is then only |
| 1004 | /// the fallback for a signed-out request. |
| 1005 | pub fn with_session_token(mut self, on: bool) -> Self { |
| 1006 | self.session_token = on; |
| 1007 | self |
| 1008 | } |
| 1009 | |
| 1010 | /// The bearer for the next request. |
| 1011 | fn bearer(&self) -> String { |
| 1012 | if self.session_token |
| 1013 | && let Some(token) = crate::credentials::load_token() |
| 1014 | { |
| 1015 | return token; |
| 1016 | } |
| 1017 | self.api_key.clone() |
| 1018 | } |
| 1019 | |
| 1020 | fn tools_json(tools: &[ToolSpec]) -> Vec<serde_json::Value> { |
| 1021 | tools |
| 1022 | .iter() |
| 1023 | .map(|tool| { |
| 1024 | // parameters_schema is a JSON string; parse it, defaulting to an |
| 1025 | // empty object schema if malformed. |
| 1026 | let parameters: serde_json::Value = serde_json::from_str(&tool.parameters_schema) |
| 1027 | .unwrap_or_else(|_| serde_json::json!({ "type": "object", "properties": {} })); |
| 1028 | serde_json::json!({ |
| 1029 | "type": "function", |
| 1030 | "function": { |
| 1031 | "name": tool.name, |
| 1032 | "description": tool.description, |
| 1033 | "parameters": parameters, |
| 1034 | } |
| 1035 | }) |
| 1036 | }) |
| 1037 | .collect() |
| 1038 | } |
| 1039 | |
| 1040 | /// POST the current history and apply the assistant reply to history. |
| 1041 | /// `tools` is always the known catalog, while `allow_tool_calls` determines |
| 1042 | /// whether it is advertised to the model and whether returned calls may run. |
| 1043 | /// Streams via SSE when `sink` is set; otherwise reads a single JSON response. |
| 1044 | /// |
| 1045 | /// A reply whose only attempt at a tool call was a block that couldn't be |
| 1046 | /// parsed gets one retry: the model is told the call didn't run and asked |
| 1047 | /// to make it again through the structured interface. Without that the |
| 1048 | /// turn ends on whatever prose preceded the broken block, and the call is |
| 1049 | /// silently lost. |
| 1050 | async fn complete( |
| 1051 | &self, |
| 1052 | tools: &[ToolSpec], |
| 1053 | allow_tool_calls: bool, |
| 1054 | sink: Option<&TokenSink>, |
| 1055 | ) -> Result<TurnResult, BackendError> { |
| 1056 | let (result, malformed) = self.request(tools, allow_tool_calls, sink).await?; |
| 1057 | let retry_note = if !allow_tool_calls || !result.tool_calls.is_empty() { |
| 1058 | None |
| 1059 | } else if malformed > 0 { |
| 1060 | log::warn!( |
| 1061 | "dropped {malformed} unparseable inline tool call(s); asking the model to retry" |
| 1062 | ); |
| 1063 | Some(MALFORMED_TOOL_CALL_RETRY) |
| 1064 | } else if result.finish == FinishReason::Repetition { |
| 1065 | log::warn!("reply was cut off as a repetition loop; asking the model to retry"); |
| 1066 | Some(REPETITION_RETRY) |
| 1067 | } else { |
| 1068 | None |
| 1069 | }; |
| 1070 | let Some(retry_note) = retry_note else { |
| 1071 | return Ok(result); |
| 1072 | }; |
| 1073 | |
| 1074 | self.history.lock().await.push(serde_json::json!({ |
| 1075 | "role": "user", |
| 1076 | "content": retry_note, |
| 1077 | })); |
| 1078 | if let Some(sink) = sink |
| 1079 | && !result.text.trim().is_empty() |
| 1080 | { |
| 1081 | // The retry's text follows what already streamed; keep it from |
| 1082 | // running on into the previous sentence. |
| 1083 | let _ = sink.send(TokenChunk::Visible("\n\n".to_string())); |
| 1084 | } |
| 1085 | let (retry, retry_malformed) = self.request(tools, allow_tool_calls, sink).await?; |
| 1086 | if retry_malformed > 0 && retry.tool_calls.is_empty() { |
| 1087 | log::warn!( |
| 1088 | "the retry also wrote {retry_malformed} unparseable inline tool call(s); \ |
| 1089 | ending the turn without them" |
| 1090 | ); |
| 1091 | } |
| 1092 | Ok(TurnResult { |
| 1093 | text: join_reply_text(&result.text, &retry.text), |
| 1094 | tool_calls: retry.tool_calls, |
| 1095 | finish: retry.finish, |
| 1096 | context_tokens: retry.context_tokens, |
| 1097 | }) |
| 1098 | } |
| 1099 | |
| 1100 | /// One chat-completion round trip. Returns the turn plus how many inline |
| 1101 | /// tool-call blocks were dropped because they couldn't be parsed. |
| 1102 | async fn request( |
| 1103 | &self, |
| 1104 | tools: &[ToolSpec], |
| 1105 | allow_tool_calls: bool, |
| 1106 | sink: Option<&TokenSink>, |
| 1107 | ) -> Result<(TurnResult, usize), BackendError> { |
| 1108 | let url = format!("{}/chat/completions", self.base_url.trim_end_matches('/')); |
| 1109 | let streaming = sink.is_some(); |
| 1110 | |
| 1111 | let messages = { |
| 1112 | let history = self.history.lock().await; |
| 1113 | let messages = if self.accepts_images { |
| 1114 | history.clone() |
| 1115 | } else { |
| 1116 | without_images(&history) |
| 1117 | }; |
| 1118 | if self.accepts_audio { |
| 1119 | messages |
| 1120 | } else { |
| 1121 | without_audio(&messages) |
| 1122 | } |
| 1123 | }; |
| 1124 | let mut body = serde_json::json!({ |
| 1125 | "model": self.model, |
| 1126 | "messages": messages, |
| 1127 | "stream": streaming, |
| 1128 | }); |
| 1129 | // A stream reports no token usage unless asked to. It comes in a last |
| 1130 | // chunk with no choices, which feeds the client's context meter. |
| 1131 | let ask_for_usage = streaming && !self.stream_usage_refused.load(Ordering::Relaxed); |
| 1132 | if ask_for_usage { |
| 1133 | body["stream_options"] = serde_json::json!({ "include_usage": true }); |
| 1134 | } |
| 1135 | if allow_tool_calls && !tools.is_empty() { |
| 1136 | body["tools"] = serde_json::Value::Array(Self::tools_json(tools)); |
| 1137 | // OpenAI specifies `auto` as the default when tools are present, |
| 1138 | // but not every OpenAI-compatible gateway implements that default. |
| 1139 | // Sending it explicitly is important for the cloud path: otherwise |
| 1140 | // a model can describe the action it intends to take as text and |
| 1141 | // end the turn without returning a tool call at all. |
| 1142 | body["tool_choice"] = serde_json::Value::String("auto".to_string()); |
| 1143 | } |
| 1144 | |
| 1145 | let mut response = self |
| 1146 | .http |
| 1147 | .post(&url) |
| 1148 | .bearer_auth(self.bearer()) |
| 1149 | .json(&body) |
| 1150 | .send() |
| 1151 | .await |
| 1152 | .map_err(|error| format!("request to {url} failed: {error}"))?; |
| 1153 | |
| 1154 | if !response.status().is_success() { |
| 1155 | let status = response.status(); |
| 1156 | let error_body = response.text().await.unwrap_or_default(); |
| 1157 | // Usage is a nicety; an endpoint that does not know the option is |
| 1158 | // asked again without it, and never asked for it again. |
| 1159 | if !(ask_for_usage && refuses_stream_options(status, &error_body)) { |
| 1160 | return Err(describe_api_error(status, &error_body)); |
| 1161 | } |
| 1162 | log::info!("{url} refused stream_options; streaming without token usage"); |
| 1163 | self.stream_usage_refused.store(true, Ordering::Relaxed); |
| 1164 | if let Some(fields) = body.as_object_mut() { |
| 1165 | fields.remove("stream_options"); |
| 1166 | } |
| 1167 | response = self |
| 1168 | .http |
| 1169 | .post(&url) |
| 1170 | .bearer_auth(self.bearer()) |
| 1171 | .json(&body) |
| 1172 | .send() |
| 1173 | .await |
| 1174 | .map_err(|error| format!("request to {url} failed: {error}"))?; |
| 1175 | if !response.status().is_success() { |
| 1176 | let status = response.status(); |
| 1177 | let error_body = response.text().await.unwrap_or_default(); |
| 1178 | return Err(describe_api_error(status, &error_body)); |
| 1179 | } |
| 1180 | } |
| 1181 | |
| 1182 | if let Some(sink) = sink { |
| 1183 | self.consume_stream(response, sink, tools, allow_tool_calls) |
| 1184 | .await |
| 1185 | } else { |
| 1186 | self.consume_json(response, tools, allow_tool_calls).await |
| 1187 | } |
| 1188 | } |
| 1189 | |
| 1190 | /// One self-contained summarization round: summarize `transcript` without |
| 1191 | /// touching the live history. |
| 1192 | /// |
| 1193 | /// Unlike [`complete`], which replays and mutates `self.history`, this |
| 1194 | /// sends its own one-off message list. Compaction only rewrites history |
| 1195 | /// once it has a summary in hand, so a failed round leaves the session |
| 1196 | /// exactly as it was — and the caller can fall back to plain truncation |
| 1197 | /// instead of failing the session (issue #125). |
| 1198 | async fn summarize( |
| 1199 | &self, |
| 1200 | system: &[serde_json::Value], |
| 1201 | transcript: &str, |
| 1202 | ) -> Result<String, BackendError> { |
| 1203 | let url = format!("{}/chat/completions", self.base_url.trim_end_matches('/')); |
| 1204 | let mut messages = Vec::with_capacity(system.len() + 1); |
| 1205 | messages.extend(system.iter().cloned()); |
| 1206 | messages.push(serde_json::json!({ |
| 1207 | "role": "user", |
| 1208 | "content": format!("{transcript}\n\n{SUMMARIZE_PROMPT}"), |
| 1209 | })); |
| 1210 | let body = serde_json::json!({ |
| 1211 | "model": self.model, |
| 1212 | "messages": messages, |
| 1213 | "stream": false, |
| 1214 | }); |
| 1215 | |
| 1216 | let response = self |
| 1217 | .http |
| 1218 | .post(&url) |
| 1219 | .bearer_auth(self.bearer()) |
| 1220 | .json(&body) |
| 1221 | .send() |
| 1222 | .await |
| 1223 | .map_err(|error| format!("request to {url} failed: {error}"))?; |
| 1224 | |
| 1225 | if !response.status().is_success() { |
| 1226 | let status = response.status(); |
| 1227 | let body = response.text().await.unwrap_or_default(); |
| 1228 | return Err(describe_api_error(status, &body)); |
| 1229 | } |
| 1230 | |
| 1231 | let parsed: ChatCompletion = response |
| 1232 | .json() |
| 1233 | .await |
| 1234 | .map_err(|error| format!("response parse error: {error}"))?; |
| 1235 | parsed |
| 1236 | .choices |
| 1237 | .into_iter() |
| 1238 | .next() |
| 1239 | .and_then(|choice| choice.message.content) |
| 1240 | .filter(|text| !text.trim().is_empty()) |
| 1241 | .ok_or_else(|| "endpoint returned no summary".to_string()) |
| 1242 | } |
| 1243 | |
| 1244 | /// Parse a single non-streaming chat-completion response. |
| 1245 | async fn consume_json( |
| 1246 | &self, |
| 1247 | response: reqwest::Response, |
| 1248 | tools: &[ToolSpec], |
| 1249 | allow_tool_calls: bool, |
| 1250 | ) -> Result<(TurnResult, usize), BackendError> { |
| 1251 | let parsed: ChatCompletion = response |
| 1252 | .json() |
| 1253 | .await |
| 1254 | .map_err(|error| format!("response parse error: {error}"))?; |
| 1255 | let context_tokens = parsed.usage.as_ref().map(Usage::context_tokens); |
| 1256 | |
| 1257 | let choice = parsed |
| 1258 | .choices |
| 1259 | .into_iter() |
| 1260 | .next() |
| 1261 | .ok_or_else(|| "endpoint returned no choices".to_string())?; |
| 1262 | let finish = FinishReason::from_wire(choice.finish_reason.as_deref()); |
| 1263 | let mut message = choice.message; |
| 1264 | |
| 1265 | let text = message.content.clone().unwrap_or_default(); |
| 1266 | let mut tool_calls: Vec<ToolCall> = message |
| 1267 | .tool_calls |
| 1268 | .iter() |
| 1269 | .flatten() |
| 1270 | .map(|call| ToolCall { |
| 1271 | id: call.id.clone(), |
| 1272 | name: call.function.name.clone(), |
| 1273 | arguments: call.function.arguments.clone(), |
| 1274 | }) |
| 1275 | .collect(); |
| 1276 | |
| 1277 | let mut extracted = crate::inline_tool_calls::extract(&text, tools); |
| 1278 | // Markup the model made up in the harness's shape goes too, from the |
| 1279 | // reply and the history alike. See `harness_markup`. |
| 1280 | extracted.text = crate::harness_markup::strip(&extracted.text); |
| 1281 | let malformed = extracted.malformed; |
| 1282 | if malformed > 0 { |
| 1283 | log::warn!("dropped {malformed} unparseable inline tool call(s) from the reply"); |
| 1284 | } |
| 1285 | |
| 1286 | if !allow_tool_calls { |
| 1287 | let recovered = extracted.calls; |
| 1288 | let text = extracted.text; |
| 1289 | let suppressed = tool_calls.len() + recovered.len(); |
| 1290 | if suppressed > 0 { |
| 1291 | let names = tool_calls |
| 1292 | .iter() |
| 1293 | .map(|call| call.name.as_str()) |
| 1294 | .chain(recovered.iter().map(|call| call.name.as_str())) |
| 1295 | .collect::<Vec<_>>() |
| 1296 | .join(", "); |
| 1297 | log::warn!( |
| 1298 | "suppressed {suppressed} tool call(s) from a forced-text response: {names}" |
| 1299 | ); |
| 1300 | } |
| 1301 | self.history |
| 1302 | .lock() |
| 1303 | .await |
| 1304 | .push(streamed_assistant_history(&text, &[])); |
| 1305 | return Ok(( |
| 1306 | TurnResult { |
| 1307 | text, |
| 1308 | tool_calls: Vec::new(), |
| 1309 | finish, |
| 1310 | context_tokens, |
| 1311 | }, |
| 1312 | malformed, |
| 1313 | )); |
| 1314 | } |
| 1315 | |
| 1316 | // Some models write a tool call out as literal `<tool_call>` text |
| 1317 | // instead of using the structured field (see `inline_tool_calls`). |
| 1318 | // Recover it, or the turn ends with the tag rendered as prose and |
| 1319 | // whatever the model meant to do is dropped. |
| 1320 | if tool_calls.is_empty() && (malformed > 0 || !extracted.calls.is_empty()) { |
| 1321 | let cleaned = extracted.text; |
| 1322 | let recovered = extracted.calls; |
| 1323 | if !recovered.is_empty() { |
| 1324 | log::warn!( |
| 1325 | "recovered {} tool call(s) the model emitted as text instead of a structured call", |
| 1326 | recovered.len() |
| 1327 | ); |
| 1328 | let tool_calls: Vec<ToolCall> = recovered |
| 1329 | .into_iter() |
| 1330 | .enumerate() |
| 1331 | .map(|(index, call)| ToolCall { |
| 1332 | id: format!("call_recovered_{index}"), |
| 1333 | name: call.name, |
| 1334 | arguments: call.arguments, |
| 1335 | }) |
| 1336 | .collect(); |
| 1337 | // Record the recovered shape, not the raw tag: the tool |
| 1338 | // results that follow have to answer an assistant message |
| 1339 | // that actually carries these calls, or the next request is |
| 1340 | // rejected for orphaned tool results. |
| 1341 | self.history |
| 1342 | .lock() |
| 1343 | .await |
| 1344 | .push(streamed_assistant_history(&cleaned, &tool_calls)); |
| 1345 | return Ok(( |
| 1346 | TurnResult { |
| 1347 | text: cleaned, |
| 1348 | tool_calls, |
| 1349 | finish, |
| 1350 | context_tokens, |
| 1351 | }, |
| 1352 | malformed, |
| 1353 | )); |
| 1354 | } |
| 1355 | // Nothing recovered, so what was found is a broken block. Keep it |
| 1356 | // out of history as well as the reply; see `inline_tool_calls` for |
| 1357 | // why it poisons later turns. |
| 1358 | self.push_reply_without_calls(&cleaned).await; |
| 1359 | return Ok(( |
| 1360 | TurnResult { |
| 1361 | text: cleaned, |
| 1362 | tool_calls, |
| 1363 | finish, |
| 1364 | context_tokens, |
| 1365 | }, |
| 1366 | malformed, |
| 1367 | )); |
| 1368 | } else if malformed > 0 || !extracted.calls.is_empty() { |
| 1369 | // Structured calls arrived with inline blocks beside them. As on |
| 1370 | // the streaming path, inline calls that parse run alongside the |
| 1371 | // structured ones, broken blocks are dropped, and neither kind of |
| 1372 | // block is left in the reply or the history. |
| 1373 | let offset = tool_calls.len(); |
| 1374 | tool_calls.extend( |
| 1375 | extracted |
| 1376 | .calls |
| 1377 | .into_iter() |
| 1378 | .enumerate() |
| 1379 | .map(|(index, call)| ToolCall { |
| 1380 | id: format!("call_recovered_{}", offset + index), |
| 1381 | name: call.name, |
| 1382 | arguments: call.arguments, |
| 1383 | }), |
| 1384 | ); |
| 1385 | self.history |
| 1386 | .lock() |
| 1387 | .await |
| 1388 | .push(streamed_assistant_history(&extracted.text, &tool_calls)); |
| 1389 | return Ok(( |
| 1390 | TurnResult { |
| 1391 | text: extracted.text, |
| 1392 | tool_calls, |
| 1393 | finish, |
| 1394 | context_tokens, |
| 1395 | }, |
| 1396 | malformed, |
| 1397 | )); |
| 1398 | } |
| 1399 | |
| 1400 | // Record the assistant turn so later tool results have context. No |
| 1401 | // inline block was found, so the only change `extracted.text` can |
| 1402 | // carry is the stripped markup. |
| 1403 | let text = extracted.text; |
| 1404 | if message.content.is_some() { |
| 1405 | message.content = Some(text.clone()); |
| 1406 | } |
| 1407 | self.history.lock().await.push(message.into_history_value()); |
| 1408 | |
| 1409 | Ok(( |
| 1410 | TurnResult { |
| 1411 | text, |
| 1412 | tool_calls, |
| 1413 | finish, |
| 1414 | context_tokens, |
| 1415 | }, |
| 1416 | 0, |
| 1417 | )) |
| 1418 | } |
| 1419 | |
| 1420 | /// Record a text-only assistant reply. An empty one is skipped: strict |
| 1421 | /// endpoints reject an assistant message with neither content nor tool |
| 1422 | /// calls, and there is nothing in it worth replaying. |
| 1423 | async fn push_reply_without_calls(&self, text: &str) { |
| 1424 | if !text.is_empty() { |
| 1425 | self.history |
| 1426 | .lock() |
| 1427 | .await |
| 1428 | .push(streamed_assistant_history(text, &[])); |
| 1429 | } |
| 1430 | } |
| 1431 | |
| 1432 | /// Consume an OpenAI Server-Sent Events stream, forwarding content deltas to |
| 1433 | /// `sink` and reassembling any tool calls (which arrive fragmented across |
| 1434 | /// chunks, keyed by `index`). |
| 1435 | async fn consume_stream( |
| 1436 | &self, |
| 1437 | response: reqwest::Response, |
| 1438 | sink: &TokenSink, |
| 1439 | tools: &[ToolSpec], |
| 1440 | allow_tool_calls: bool, |
| 1441 | ) -> Result<(TurnResult, usize), BackendError> { |
| 1442 | use futures::StreamExt; |
| 1443 | |
| 1444 | let mut stream = response.bytes_stream(); |
| 1445 | // Newlines are ASCII, so splitting raw bytes on `\n` never bisects a |
| 1446 | // multibyte UTF-8 sequence; we only lossily decode whole lines. |
| 1447 | let mut buffer: Vec<u8> = Vec::new(); |
| 1448 | let mut text = String::new(); |
| 1449 | let mut tool_accum: Vec<StreamingToolCall> = Vec::new(); |
| 1450 | let mut done = false; |
| 1451 | // Arrives on a late chunk, usually one with an empty delta. |
| 1452 | let mut finish = FinishReason::default(); |
| 1453 | // Arrives on the last chunk, which has no choices, when asked for. |
| 1454 | let mut context_tokens = None; |
| 1455 | // Recovers a tool call the model wrote as literal `<tool_call>` text |
| 1456 | // instead of a structured delta. Scanning here (rather than after the |
| 1457 | // stream) keeps the tag off the UI: content goes straight to `sink` as |
| 1458 | // it arrives, so by the time a whole turn is assembled the tag has |
| 1459 | // already been rendered. See `inline_tool_calls`. |
| 1460 | let mut scanner = crate::inline_tool_calls::StreamScanner::new(tools); |
| 1461 | // What it lets through then loses any markup the model made up in the |
| 1462 | // harness's shape, like a `<system_warning>` block. See |
| 1463 | // `harness_markup`. |
| 1464 | let mut markup = crate::harness_markup::MarkupFilter::new(); |
| 1465 | let mut recovered: Vec<ToolCall> = Vec::new(); |
| 1466 | let mut malformed = 0usize; |
| 1467 | // Stops a reply that falls into a repetition loop. See `repetition`. |
| 1468 | let mut guard = crate::repetition::RepetitionGuard::new(); |
| 1469 | let mut degenerate = false; |
| 1470 | |
| 1471 | while let Some(item) = stream.next().await { |
| 1472 | let bytes = item.map_err(|error| format!("stream read error: {error}"))?; |
| 1473 | buffer.extend_from_slice(&bytes); |
| 1474 | |
| 1475 | while let Some(pos) = buffer.iter().position(|&b| b == b'\n') { |
| 1476 | let line: Vec<u8> = buffer.drain(..=pos).collect(); |
| 1477 | let line = String::from_utf8_lossy(&line); |
| 1478 | let line = line.trim(); |
| 1479 | |
| 1480 | let Some(data) = line.strip_prefix("data:") else { |
| 1481 | continue; |
| 1482 | }; |
| 1483 | let data = data.trim(); |
| 1484 | if data == "[DONE]" { |
| 1485 | done = true; |
| 1486 | break; |
| 1487 | } |
| 1488 | if data.is_empty() { |
| 1489 | continue; |
| 1490 | } |
| 1491 | |
| 1492 | // An endpoint that fails after the status line is already sent |
| 1493 | // reports it in the stream instead, as an ordinary `data:` |
| 1494 | // frame holding an error envelope. siGit Code Cloud does this |
| 1495 | // when the upstream fails mid-turn. It has no `choices`, so |
| 1496 | // without this it parses as an empty chunk and is skipped, and |
| 1497 | // the turn ends looking like the model simply said nothing. |
| 1498 | if let Some(message) = api_error_message(data) { |
| 1499 | return Err(message); |
| 1500 | } |
| 1501 | |
| 1502 | let chunk: StreamCompletion = match serde_json::from_str(data) { |
| 1503 | Ok(chunk) => chunk, |
| 1504 | // Skip keep-alive comments and anything we can't parse rather |
| 1505 | // than aborting a turn over one malformed frame. |
| 1506 | Err(_) => continue, |
| 1507 | }; |
| 1508 | |
| 1509 | if let Some(usage) = &chunk.usage { |
| 1510 | context_tokens = Some(usage.context_tokens()); |
| 1511 | } |
| 1512 | let Some(choice) = chunk.choices.into_iter().next() else { |
| 1513 | continue; |
| 1514 | }; |
| 1515 | if let Some(reason) = choice.finish_reason.as_deref() { |
| 1516 | finish = FinishReason::from_wire(Some(reason)); |
| 1517 | } |
| 1518 | // Reasoning streamed in its own field is tagged at the source; |
| 1519 | // it bypasses the inline-tag scanner and is never kept in the |
| 1520 | // assistant message text. |
| 1521 | if let Some(reasoning) = choice.delta.reasoning_content |
| 1522 | && !reasoning.is_empty() |
| 1523 | && sink.send(TokenChunk::Reasoning(reasoning)).is_err() |
| 1524 | { |
| 1525 | // Consumer dropped (turn cancelled). |
| 1526 | done = true; |
| 1527 | break; |
| 1528 | } |
| 1529 | if let Some(content) = choice.delta.content |
| 1530 | && !content.is_empty() |
| 1531 | { |
| 1532 | let mut cancelled = false; |
| 1533 | for event in scanner.push(&content) { |
| 1534 | match event { |
| 1535 | crate::inline_tool_calls::ScanEvent::Text(chunk) => { |
| 1536 | let chunk = markup.push(&chunk); |
| 1537 | if chunk.is_empty() { |
| 1538 | continue; |
| 1539 | } |
| 1540 | text.push_str(&chunk); |
| 1541 | let looping = guard.push(&chunk); |
| 1542 | if sink.send(TokenChunk::Visible(chunk)).is_err() { |
| 1543 | // Consumer dropped (turn cancelled). |
| 1544 | cancelled = true; |
| 1545 | break; |
| 1546 | } |
| 1547 | if looping { |
| 1548 | degenerate = true; |
| 1549 | break; |
| 1550 | } |
| 1551 | } |
| 1552 | crate::inline_tool_calls::ScanEvent::ToolCall(call) => { |
| 1553 | if allow_tool_calls { |
| 1554 | log::warn!( |
| 1555 | "recovered tool call '{}' the model emitted as text instead of a structured call", |
| 1556 | call.name |
| 1557 | ); |
| 1558 | } |
| 1559 | recovered.push(ToolCall { |
| 1560 | id: format!("call_recovered_{}", recovered.len()), |
| 1561 | name: call.name, |
| 1562 | arguments: call.arguments, |
| 1563 | }); |
| 1564 | } |
| 1565 | crate::inline_tool_calls::ScanEvent::Malformed(block) => { |
| 1566 | log_malformed_block(&block); |
| 1567 | malformed += 1; |
| 1568 | } |
| 1569 | } |
| 1570 | } |
| 1571 | if cancelled || degenerate { |
| 1572 | done = true; |
| 1573 | break; |
| 1574 | } |
| 1575 | } |
| 1576 | for delta in choice.delta.tool_calls.into_iter().flatten() { |
| 1577 | let index = delta.index.unwrap_or(0) as usize; |
| 1578 | if tool_accum.len() <= index { |
| 1579 | tool_accum.resize_with(index + 1, StreamingToolCall::default); |
| 1580 | } |
| 1581 | let slot = &mut tool_accum[index]; |
| 1582 | if let Some(id) = delta.id { |
| 1583 | slot.id = id; |
| 1584 | } |
| 1585 | if let Some(function) = delta.function { |
| 1586 | if let Some(name) = function.name { |
| 1587 | slot.name = name; |
| 1588 | } |
| 1589 | if let Some(arguments) = function.arguments { |
| 1590 | slot.arguments.push_str(&arguments); |
| 1591 | } |
| 1592 | } |
| 1593 | } |
| 1594 | } |
| 1595 | |
| 1596 | if done { |
| 1597 | break; |
| 1598 | } |
| 1599 | } |
| 1600 | |
| 1601 | // Flush what was held back. A partial marker is just text; a |
| 1602 | // tool-call block that never closed is dropped like any other |
| 1603 | // unparseable one. |
| 1604 | let mut leftover = String::new(); |
| 1605 | if degenerate { |
| 1606 | // Dropping the stream below closes the connection, so the endpoint |
| 1607 | // stops generating. Nothing held back is worth flushing. |
| 1608 | let keep = guard.loop_start().unwrap_or(text.len()).min(text.len()); |
| 1609 | log::warn!( |
| 1610 | "reply degenerated into repetition; cut {} of {} bytes", |
| 1611 | text.len() - keep, |
| 1612 | text.len() |
| 1613 | ); |
| 1614 | text.truncate(keep); |
| 1615 | text.truncate(text.trim_end().len()); |
| 1616 | finish = FinishReason::Repetition; |
| 1617 | let _ = sink.send(TokenChunk::Visible(REPETITION_NOTICE.to_string())); |
| 1618 | } else { |
| 1619 | match scanner.finish() { |
| 1620 | Some(crate::inline_tool_calls::ScanEvent::Text(rest)) => { |
| 1621 | leftover = markup.push(&rest); |
| 1622 | } |
| 1623 | Some(crate::inline_tool_calls::ScanEvent::Malformed(block)) => { |
| 1624 | log_malformed_block(&block); |
| 1625 | malformed += 1; |
| 1626 | } |
| 1627 | Some(crate::inline_tool_calls::ScanEvent::ToolCall(_)) | None => {} |
| 1628 | } |
| 1629 | leftover.push_str(&markup.finish()); |
| 1630 | if !leftover.is_empty() { |
| 1631 | text.push_str(&leftover); |
| 1632 | let _ = sink.send(TokenChunk::Visible(leftover)); |
| 1633 | } |
| 1634 | } |
| 1635 | |
| 1636 | let mut tool_calls: Vec<ToolCall> = tool_accum |
| 1637 | .iter() |
| 1638 | .filter(|call| !call.name.is_empty()) |
| 1639 | .enumerate() |
| 1640 | .map(|(index, call)| ToolCall { |
| 1641 | id: if call.id.is_empty() { |
| 1642 | format!("call_{index}") |
| 1643 | } else { |
| 1644 | call.id.clone() |
| 1645 | }, |
| 1646 | name: call.name.clone(), |
| 1647 | arguments: call.arguments.clone(), |
| 1648 | }) |
| 1649 | .collect(); |
| 1650 | tool_calls.extend(recovered); |
| 1651 | |
| 1652 | if !allow_tool_calls && !tool_calls.is_empty() { |
| 1653 | let names = tool_calls |
| 1654 | .iter() |
| 1655 | .map(|call| call.name.as_str()) |
| 1656 | .collect::<Vec<_>>() |
| 1657 | .join(", "); |
| 1658 | log::warn!( |
| 1659 | "suppressed {} tool call(s) from a forced-text response: {names}", |
| 1660 | tool_calls.len() |
| 1661 | ); |
| 1662 | tool_calls.clear(); |
| 1663 | } |
| 1664 | |
| 1665 | // Record the assistant turn so later tool results have context. |
| 1666 | if (malformed > 0 || degenerate) && tool_calls.is_empty() { |
| 1667 | self.push_reply_without_calls(&text).await; |
| 1668 | } else { |
| 1669 | self.history |
| 1670 | .lock() |
| 1671 | .await |
| 1672 | .push(streamed_assistant_history(&text, &tool_calls)); |
| 1673 | } |
| 1674 | |
| 1675 | Ok(( |
| 1676 | TurnResult { |
| 1677 | text, |
| 1678 | tool_calls, |
| 1679 | finish, |
| 1680 | context_tokens, |
| 1681 | }, |
| 1682 | malformed, |
| 1683 | )) |
| 1684 | } |
| 1685 | } |
| 1686 | |
| 1687 | /// Sent as a user turn after a reply whose tool call couldn't be parsed. |
| 1688 | const MALFORMED_TOOL_CALL_RETRY: &str = "[siGit Code] Your last reply tried to call a tool by \ |
| 1689 | writing the call out as text, and it could not be parsed, so nothing ran and the user did \ |
| 1690 | not see it. Do not write tool calls or tool results as text. Make the call again using the \ |
| 1691 | tool-calling interface, or answer in plain prose if no tool is needed."; |
| 1692 | |
| 1693 | /// Sent as a user turn after a reply was cut off as a repetition loop. |
| 1694 | const REPETITION_RETRY: &str = "[siGit Code] Your last reply started repeating the same few \ |
| 1695 | words and was cut off. Do not continue it. Make the next tool call you need, or answer \ |
| 1696 | briefly in plain prose."; |
| 1697 | |
| 1698 | /// Shown to the user, who has already watched the loop stream. Not kept in |
| 1699 | /// history. |
| 1700 | const REPETITION_NOTICE: &str = |
| 1701 | "\n\n[siGit Code stopped this reply because the model started repeating itself.]"; |
| 1702 | |
| 1703 | fn log_malformed_block(block: &str) { |
| 1704 | log::warn!( |
| 1705 | "dropped an unparseable inline tool call ({} chars): {}", |
| 1706 | block.len(), |
| 1707 | block.chars().take(200).collect::<String>() |
| 1708 | ); |
| 1709 | } |
| 1710 | |
| 1711 | /// Join the visible text of a reply and its retry the way the stream showed |
| 1712 | /// them: as separate paragraphs. |
| 1713 | fn join_reply_text(first: &str, retry: &str) -> String { |
| 1714 | match (first.trim().is_empty(), retry.trim().is_empty()) { |
| 1715 | (true, _) => retry.to_string(), |
| 1716 | (false, true) => first.to_string(), |
| 1717 | (false, false) => format!("{first}\n\n{retry}"), |
| 1718 | } |
| 1719 | } |
| 1720 | |
| 1721 | /// One tool call being reassembled from streamed deltas. |
| 1722 | #[derive(Default)] |
| 1723 | struct StreamingToolCall { |
| 1724 | id: String, |
| 1725 | name: String, |
| 1726 | arguments: String, |
| 1727 | } |
| 1728 | |
| 1729 | /// Rebuild the assistant message for replay in history after a streamed turn, |
| 1730 | /// preserving any tool calls so the follow-up request is well-formed. Mirrors |
| 1731 | /// [`ResponseMessage::into_history_value`] for the non-streaming path. |
| 1732 | fn streamed_assistant_history(text: &str, tool_calls: &[ToolCall]) -> serde_json::Value { |
| 1733 | let mut message = serde_json::json!({ "role": "assistant" }); |
| 1734 | message["content"] = if text.is_empty() { |
| 1735 | serde_json::Value::Null |
| 1736 | } else { |
| 1737 | serde_json::Value::String(text.to_string()) |
| 1738 | }; |
| 1739 | if !tool_calls.is_empty() { |
| 1740 | message["tool_calls"] = serde_json::json!( |
| 1741 | tool_calls |
| 1742 | .iter() |
| 1743 | .map(|call| serde_json::json!({ |
| 1744 | "id": call.id, |
| 1745 | "type": "function", |
| 1746 | "function": { |
| 1747 | "name": call.name, |
| 1748 | "arguments": call.arguments, |
| 1749 | } |
| 1750 | })) |
| 1751 | .collect::<Vec<_>>() |
| 1752 | ); |
| 1753 | } |
| 1754 | message |
| 1755 | } |
| 1756 | |
| 1757 | #[async_trait] |
| 1758 | impl InferenceBackend for OpenAiBackend { |
| 1759 | async fn send_message_with_tools( |
| 1760 | &self, |
| 1761 | text: &str, |
| 1762 | tools: &[ToolSpec], |
| 1763 | sink: Option<&TokenSink>, |
| 1764 | ) -> Result<TurnResult, BackendError> { |
| 1765 | self.history |
| 1766 | .lock() |
| 1767 | .await |
| 1768 | .push(serde_json::json!({ "role": "user", "content": text })); |
| 1769 | self.complete(tools, true, sink).await |
| 1770 | } |
| 1771 | |
| 1772 | async fn send_tool_results( |
| 1773 | &self, |
| 1774 | results: Vec<ToolResult>, |
| 1775 | tools: &[ToolSpec], |
| 1776 | allow_tool_calls: bool, |
| 1777 | sink: Option<&TokenSink>, |
| 1778 | ) -> Result<TurnResult, BackendError> { |
| 1779 | { |
| 1780 | let mut history = self.history.lock().await; |
| 1781 | for result in results { |
| 1782 | history.push(serde_json::json!({ |
| 1783 | "role": "tool", |
| 1784 | "tool_call_id": result.tool_call_id, |
| 1785 | "content": result.content, |
| 1786 | })); |
| 1787 | } |
| 1788 | } |
| 1789 | self.complete(tools, allow_tool_calls, sink).await |
| 1790 | } |
| 1791 | |
| 1792 | async fn record_cancelled_tool_results(&self, results: Vec<ToolResult>) { |
| 1793 | let mut history = self.history.lock().await; |
| 1794 | for result in results { |
| 1795 | // Cancellation races the in-flight continuation request. That |
| 1796 | // request appends its tool results before awaiting HTTP, so cleanup |
| 1797 | // must be safe whether cancellation won just before or just after |
| 1798 | // the append. |
| 1799 | let matching_call = history.iter().rposition(|message| { |
| 1800 | message["role"] == "assistant" |
| 1801 | && message["tool_calls"].as_array().is_some_and(|calls| { |
| 1802 | calls.iter().any(|call| call["id"] == result.tool_call_id) |
| 1803 | }) |
| 1804 | }); |
| 1805 | let already_recorded = matching_call.is_some_and(|call_index| { |
| 1806 | history[call_index + 1..].iter().any(|message| { |
| 1807 | message["role"] == "tool" && message["tool_call_id"] == result.tool_call_id |
| 1808 | }) |
| 1809 | }); |
| 1810 | if already_recorded { |
| 1811 | continue; |
| 1812 | } |
| 1813 | history.push(serde_json::json!({ |
| 1814 | "role": "tool", |
| 1815 | "tool_call_id": result.tool_call_id, |
| 1816 | "content": result.content, |
| 1817 | })); |
| 1818 | } |
| 1819 | } |
| 1820 | |
| 1821 | async fn send_message_with_attachments( |
| 1822 | &self, |
| 1823 | text: &str, |
| 1824 | images: &[ImageInput], |
| 1825 | audio: &[AudioInput], |
| 1826 | tools: &[ToolSpec], |
| 1827 | sink: Option<&TokenSink>, |
| 1828 | ) -> Result<TurnResult, BackendError> { |
| 1829 | self.history |
| 1830 | .lock() |
| 1831 | .await |
| 1832 | .push(user_message(text, images, audio)); |
| 1833 | self.complete(tools, true, sink).await |
| 1834 | } |
| 1835 | |
| 1836 | fn accepts_images(&self) -> bool { |
| 1837 | self.accepts_images |
| 1838 | } |
| 1839 | |
| 1840 | fn accepts_audio(&self) -> bool { |
| 1841 | self.accepts_audio |
| 1842 | } |
| 1843 | |
| 1844 | fn is_remote(&self) -> bool { |
| 1845 | true |
| 1846 | } |
| 1847 | |
| 1848 | fn fresh(&self) -> Option<Arc<dyn InferenceBackend>> { |
| 1849 | Some(Arc::new(Self { |
| 1850 | base_url: self.base_url.clone(), |
| 1851 | api_key: self.api_key.clone(), |
| 1852 | session_token: self.session_token, |
| 1853 | model: self.model.clone(), |
| 1854 | // The client is a handle on one connection pool; sharing it keeps |
| 1855 | // connections to the endpoint warm across sessions. |
| 1856 | http: self.http.clone(), |
| 1857 | history: Mutex::new(Vec::new()), |
| 1858 | accepts_images: self.accepts_images, |
| 1859 | accepts_audio: self.accepts_audio, |
| 1860 | // It is the same endpoint, so what it refused it refuses here. |
| 1861 | stream_usage_refused: Arc::clone(&self.stream_usage_refused), |
| 1862 | })) |
| 1863 | } |
| 1864 | |
| 1865 | async fn history_snapshot(&self) -> Vec<serde_json::Value> { |
| 1866 | self.history.lock().await.clone() |
| 1867 | } |
| 1868 | |
| 1869 | async fn restore_history(&self, history: Vec<serde_json::Value>) { |
| 1870 | // The snapshot includes the seeded system message, so a wholesale |
| 1871 | // replacement restores exactly what was saved. |
| 1872 | *self.history.lock().await = history; |
| 1873 | } |
| 1874 | |
| 1875 | async fn compact_history(&self, keep_last: usize) -> Result<(), BackendError> { |
| 1876 | let snapshot: Vec<serde_json::Value> = self.history.lock().await.clone(); |
| 1877 | |
| 1878 | // Leading system messages carry session context (the seeded prompt, |
| 1879 | // pushed context); keep them all, like the summarization-success path. |
| 1880 | let system: Vec<serde_json::Value> = snapshot |
| 1881 | .iter() |
| 1882 | .take_while(|message| message["role"] == "system") |
| 1883 | .cloned() |
| 1884 | .collect(); |
| 1885 | |
| 1886 | let non_system: Vec<serde_json::Value> = snapshot |
| 1887 | .iter() |
| 1888 | .filter(|message| message["role"] != "system") |
| 1889 | .cloned() |
| 1890 | .collect(); |
| 1891 | let tail_start = non_system.len().saturating_sub(keep_last); |
| 1892 | let mut tail = non_system[tail_start..].to_vec(); |
| 1893 | // Drop leading tool results whose assistant tool-call message was |
| 1894 | // summarized away — strict endpoints reject orphaned `role: "tool"` |
| 1895 | // entries on the very next request. |
| 1896 | while tail |
| 1897 | .first() |
| 1898 | .is_some_and(|message| message["role"] == "tool") |
| 1899 | { |
| 1900 | tail.remove(0); |
| 1901 | } |
| 1902 | |
| 1903 | // Ask the endpoint for a summary of the conversation so far, through a |
| 1904 | // direct completion call. The request carries the conversation as a |
| 1905 | // flattened transcript in a single user message rather than the live |
| 1906 | // history: this round offers no tools, and a tool-shaped history sent |
| 1907 | // without a tool schema is rejected upstream (see |
| 1908 | // `transcript_for_summary`). The transcript is capped so the request |
| 1909 | // cannot be as large as the over-budget conversation that triggered |
| 1910 | // compaction (see `SUMMARY_TRANSCRIPT_TOKEN_CAP`). |
| 1911 | let transcript = truncate_transcript_middle( |
| 1912 | &transcript_for_summary(&snapshot), |
| 1913 | SUMMARY_TRANSCRIPT_TOKEN_CAP, |
| 1914 | ); |
| 1915 | let summary = match self.summarize(&system, &transcript).await { |
| 1916 | Ok(summary) => summary, |
| 1917 | Err(error) => { |
| 1918 | log::warn!( |
| 1919 | "summarization round failed ({error}); falling back to plain truncation" |
| 1920 | ); |
| 1921 | // Compaction must not fail the session: fold the overflow into |
| 1922 | // a placeholder and drop the oldest messages until the rebuild |
| 1923 | // fits the budget. Needs no endpoint, so it cannot fail the way |
| 1924 | // the inference round just did (issue #125). |
| 1925 | let rebuilt = truncate_history_to_budget(&system, tail.clone()).ok_or(error)?; |
| 1926 | *self.history.lock().await = rebuilt; |
| 1927 | return Ok(()); |
| 1928 | } |
| 1929 | }; |
| 1930 | |
| 1931 | let mut rebuilt = Vec::with_capacity(system.len() + tail.len() + 1); |
| 1932 | rebuilt.extend(system); |
| 1933 | rebuilt.push(serde_json::json!({ |
| 1934 | "role": "user", |
| 1935 | "content": format!("[Conversation summary]\n{summary}"), |
| 1936 | })); |
| 1937 | rebuilt.extend(tail); |
| 1938 | *self.history.lock().await = rebuilt; |
| 1939 | Ok(()) |
| 1940 | } |
| 1941 | } |
| 1942 | |
| 1943 | /// The placeholder message replacing dropped history when compaction falls |
| 1944 | /// back to plain truncation (the model reads it, so it explains the gap). |
| 1945 | const TRUNCATION_PLACEHOLDER: &str = "[Older conversation omitted: the session outgrew \ |
| 1946 | the context window and the summarization service was unavailable. Earlier \ |
| 1947 | messages were dropped, so answers may lack that context.]"; |
| 1948 | |
| 1949 | /// Last-resort compaction when the summarization round cannot run: rebuild the |
| 1950 | /// history as `[system, placeholder, newest messages that fit the budget]`. |
| 1951 | /// |
| 1952 | /// Returns `None` when even the smallest rebuild exceeds the budget (a |
| 1953 | /// pathological case — e.g. one message larger than the whole window); the |
| 1954 | /// caller then reports the original summarization error and leaves history |
| 1955 | /// untouched. |
| 1956 | fn truncate_history_to_budget( |
| 1957 | system: &[serde_json::Value], |
| 1958 | mut tail: Vec<serde_json::Value>, |
| 1959 | ) -> Option<Vec<serde_json::Value>> { |
| 1960 | /// A `role: "tool"` entry only makes sense behind the assistant message |
| 1961 | /// that requested it; dropping that message strands it. |
| 1962 | fn drop_orphaned_tool_results(tail: &mut Vec<serde_json::Value>) { |
| 1963 | while tail |
| 1964 | .first() |
| 1965 | .is_some_and(|message| message["role"] == "tool") |
| 1966 | { |
| 1967 | tail.remove(0); |
| 1968 | } |
| 1969 | } |
| 1970 | |
| 1971 | // A caller splitting mid-round can hand in a tail that starts with tool |
| 1972 | // results whose call message already fell outside the kept window. |
| 1973 | drop_orphaned_tool_results(&mut tail); |
| 1974 | |
| 1975 | loop { |
| 1976 | let mut candidate = Vec::with_capacity(tail.len() + system.len() + 1); |
| 1977 | candidate.extend(system.iter().cloned()); |
| 1978 | candidate.push(serde_json::json!({ |
| 1979 | "role": "user", |
| 1980 | "content": TRUNCATION_PLACEHOLDER, |
| 1981 | })); |
| 1982 | candidate.extend(tail.iter().cloned()); |
| 1983 | if estimate_tokens(&candidate) <= DEFAULT_CONTEXT_TOKEN_BUDGET { |
| 1984 | return Some(candidate); |
| 1985 | } |
| 1986 | |
| 1987 | // Nothing left to drop: even the minimal rebuild (system + placeholder, |
| 1988 | // no conversation) overflows — the session is genuinely unrecoverable. |
| 1989 | if tail.is_empty() { |
| 1990 | return None; |
| 1991 | } |
| 1992 | if tail[0]["role"] == "assistant" |
| 1993 | && tail[0]["tool_calls"] |
| 1994 | .as_array() |
| 1995 | .is_some_and(|calls| !calls.is_empty()) |
| 1996 | && tail.get(1).is_some_and(|message| message["role"] == "tool") |
| 1997 | { |
| 1998 | // Dropping an assistant tool-call message would orphan its |
| 1999 | // results; drop the whole round (call + results) together. |
| 2000 | let mut end = 1; |
| 2001 | while tail |
| 2002 | .get(end) |
| 2003 | .is_some_and(|message| message["role"] == "tool") |
| 2004 | { |
| 2005 | end += 1; |
| 2006 | } |
| 2007 | tail.drain(..end); |
| 2008 | } else { |
| 2009 | tail.remove(0); |
| 2010 | drop_orphaned_tool_results(&mut tail); |
| 2011 | } |
| 2012 | } |
| 2013 | } |
| 2014 | |
| 2015 | // ── OpenAI error shape ──────────────────────────────────────────────────────── |
| 2016 | |
| 2017 | /// The OpenAI error envelope: `{"error": {"message": ..., "type": ...}}`. |
| 2018 | /// |
| 2019 | /// Every endpoint siGit talks to speaks it, including siGit Code Cloud, whose |
| 2020 | /// messages are written for the person reading them ("Monthly siGit Code Cloud |
| 2021 | /// allowance reached..."). Worth unwrapping rather than pasting the raw body |
| 2022 | /// into the editor's error banner. |
| 2023 | #[derive(Debug, Deserialize)] |
| 2024 | struct ApiErrorEnvelope { |
| 2025 | error: ApiErrorBody, |
| 2026 | } |
| 2027 | |
| 2028 | #[derive(Debug, Deserialize)] |
| 2029 | struct ApiErrorBody { |
| 2030 | #[serde(default)] |
| 2031 | message: Option<String>, |
| 2032 | } |
| 2033 | |
| 2034 | /// How much of an unparseable error body to keep. An endpoint behind a proxy |
| 2035 | /// can answer with a full HTML page, and the whole thing ends up in the |
| 2036 | /// editor's error banner. |
| 2037 | const ERROR_BODY_LIMIT: usize = 500; |
| 2038 | |
| 2039 | /// Turn an error response into something worth showing a person. |
| 2040 | /// |
| 2041 | /// The endpoint's own message wins when there is one: it is written for the |
| 2042 | /// user, and the status code repeats what it already says. Anything else falls |
| 2043 | /// back to the status plus whatever the body held, which is all there is to go |
| 2044 | /// on — except an HTML page, which is a proxy or Rails error page and tells |
| 2045 | /// the person nothing the status doesn't (pasting it into chat is the dump |
| 2046 | /// seen in issue #125's screenshots). |
| 2047 | fn describe_api_error(status: reqwest::StatusCode, body: &str) -> String { |
| 2048 | if let Some(message) = api_error_message(body) { |
| 2049 | return message; |
| 2050 | } |
| 2051 | |
| 2052 | let body = body.trim(); |
| 2053 | if body.is_empty() || is_probable_html(body) { |
| 2054 | return format!("endpoint returned {status}"); |
| 2055 | } |
| 2056 | |
| 2057 | let mut detail = body; |
| 2058 | if detail.len() > ERROR_BODY_LIMIT { |
| 2059 | let mut cut = ERROR_BODY_LIMIT; |
| 2060 | while !detail.is_char_boundary(cut) { |
| 2061 | cut -= 1; |
| 2062 | } |
| 2063 | detail = &detail[..cut]; |
| 2064 | return format!("endpoint returned {status}: {detail}…"); |
| 2065 | } |
| 2066 | format!("endpoint returned {status}: {detail}") |
| 2067 | } |
| 2068 | |
| 2069 | /// The human-readable message out of an OpenAI error envelope, if the body is |
| 2070 | /// one and carries a non-empty message. |
| 2071 | fn api_error_message(body: &str) -> Option<String> { |
| 2072 | let envelope: ApiErrorEnvelope = serde_json::from_str(body).ok()?; |
| 2073 | let message = envelope.error.message?; |
| 2074 | let message = message.trim(); |
| 2075 | (!message.is_empty()).then(|| message.to_string()) |
| 2076 | } |
| 2077 | |
| 2078 | /// Whether `body` looks like an HTML page rather than prose: a doctype or an |
| 2079 | /// opening tag in the first chunk. Gateways and Rails apps answer failures |
| 2080 | /// with full pages; none of that markup is worth showing. |
| 2081 | fn is_probable_html(body: &str) -> bool { |
| 2082 | let head: String = body.chars().take(512).collect(); |
| 2083 | let head = head.trim_start().to_ascii_lowercase(); |
| 2084 | head.starts_with("<!doctype") || head.starts_with("<html") || head.starts_with("<head") |
| 2085 | } |
| 2086 | |
| 2087 | // ── OpenAI response shapes ──────────────────────────────────────────────────────── |
| 2088 | |
| 2089 | #[derive(Debug, Deserialize)] |
| 2090 | struct ChatCompletion { |
| 2091 | #[serde(default)] |
| 2092 | choices: Vec<CompletionChoice>, |
| 2093 | #[serde(default)] |
| 2094 | usage: Option<Usage>, |
| 2095 | } |
| 2096 | |
| 2097 | /// Token usage as an OpenAI-compatible endpoint reports it. |
| 2098 | #[derive(Debug, Deserialize)] |
| 2099 | struct Usage { |
| 2100 | #[serde(default)] |
| 2101 | prompt_tokens: u64, |
| 2102 | #[serde(default)] |
| 2103 | completion_tokens: u64, |
| 2104 | #[serde(default)] |
| 2105 | total_tokens: Option<u64>, |
| 2106 | } |
| 2107 | |
| 2108 | impl Usage { |
| 2109 | /// The context in use once the reply is in. |
| 2110 | fn context_tokens(&self) -> u64 { |
| 2111 | self.total_tokens |
| 2112 | .unwrap_or(self.prompt_tokens + self.completion_tokens) |
| 2113 | } |
| 2114 | } |
| 2115 | |
| 2116 | /// Whether a failed request was turned down for carrying `stream_options`, |
| 2117 | /// which some OpenAI-compatible servers predate and reject as unknown. |
| 2118 | fn refuses_stream_options(status: reqwest::StatusCode, body: &str) -> bool { |
| 2119 | matches!(status.as_u16(), 400 | 422) && body.contains("stream_options") |
| 2120 | } |
| 2121 | |
| 2122 | #[derive(Debug, Deserialize)] |
| 2123 | struct CompletionChoice { |
| 2124 | message: ResponseMessage, |
| 2125 | #[serde(default)] |
| 2126 | finish_reason: Option<String>, |
| 2127 | } |
| 2128 | |
| 2129 | #[derive(Debug, Deserialize)] |
| 2130 | struct ResponseMessage { |
| 2131 | #[serde(default)] |
| 2132 | content: Option<String>, |
| 2133 | #[serde(default)] |
| 2134 | tool_calls: Option<Vec<ResponseToolCall>>, |
| 2135 | } |
| 2136 | |
| 2137 | impl ResponseMessage { |
| 2138 | /// Reconstruct the assistant message for replay in history, preserving any |
| 2139 | /// tool calls so the follow-up request is well-formed. |
| 2140 | fn into_history_value(self) -> serde_json::Value { |
| 2141 | let mut message = serde_json::json!({ "role": "assistant" }); |
| 2142 | message["content"] = match self.content { |
| 2143 | Some(text) => serde_json::Value::String(text), |
| 2144 | None => serde_json::Value::Null, |
| 2145 | }; |
| 2146 | if let Some(tool_calls) = self.tool_calls { |
| 2147 | message["tool_calls"] = serde_json::json!( |
| 2148 | tool_calls |
| 2149 | .into_iter() |
| 2150 | .map(|call| serde_json::json!({ |
| 2151 | "id": call.id, |
| 2152 | "type": "function", |
| 2153 | "function": { |
| 2154 | "name": call.function.name, |
| 2155 | "arguments": call.function.arguments, |
| 2156 | } |
| 2157 | })) |
| 2158 | .collect::<Vec<_>>() |
| 2159 | ); |
| 2160 | } |
| 2161 | message |
| 2162 | } |
| 2163 | } |
| 2164 | |
| 2165 | #[derive(Debug, Deserialize)] |
| 2166 | struct ResponseToolCall { |
| 2167 | id: String, |
| 2168 | function: ResponseFunction, |
| 2169 | } |
| 2170 | |
| 2171 | #[derive(Debug, Deserialize)] |
| 2172 | struct ResponseFunction { |
| 2173 | name: String, |
| 2174 | #[serde(default)] |
| 2175 | arguments: String, |
| 2176 | } |
| 2177 | |
| 2178 | // ── OpenAI streaming (SSE) chunk shapes ───────────────────────────────────────── |
| 2179 | |
| 2180 | #[derive(Debug, Deserialize)] |
| 2181 | struct StreamCompletion { |
| 2182 | #[serde(default)] |
| 2183 | choices: Vec<StreamChoice>, |
| 2184 | #[serde(default)] |
| 2185 | usage: Option<Usage>, |
| 2186 | } |
| 2187 | |
| 2188 | #[derive(Debug, Deserialize)] |
| 2189 | struct StreamChoice { |
| 2190 | #[serde(default)] |
| 2191 | delta: StreamDelta, |
| 2192 | #[serde(default)] |
| 2193 | finish_reason: Option<String>, |
| 2194 | } |
| 2195 | |
| 2196 | #[derive(Debug, Default, Deserialize)] |
| 2197 | struct StreamDelta { |
| 2198 | #[serde(default)] |
| 2199 | content: Option<String>, |
| 2200 | // Some endpoints stream model reasoning in a dedicated field rather than |
| 2201 | // inline `<think>` tags. It is unambiguously reasoning at the source, so it |
| 2202 | // travels as `TokenChunk::Reasoning` and never enters the content scanner. |
| 2203 | #[serde(default, alias = "reasoning")] |
| 2204 | reasoning_content: Option<String>, |
| 2205 | #[serde(default)] |
| 2206 | tool_calls: Option<Vec<StreamToolCallDelta>>, |
| 2207 | } |
| 2208 | |
| 2209 | #[derive(Debug, Deserialize)] |
| 2210 | struct StreamToolCallDelta { |
| 2211 | #[serde(default)] |
| 2212 | index: Option<u32>, |
| 2213 | #[serde(default)] |
| 2214 | id: Option<String>, |
| 2215 | #[serde(default)] |
| 2216 | function: Option<StreamFunctionDelta>, |
| 2217 | } |
| 2218 | |
| 2219 | #[derive(Debug, Deserialize)] |
| 2220 | struct StreamFunctionDelta { |
| 2221 | #[serde(default)] |
| 2222 | name: Option<String>, |
| 2223 | #[serde(default)] |
| 2224 | arguments: Option<String>, |
| 2225 | } |
| 2226 | |
| 2227 | // ── Carrying a conversation across a model switch ──────────────────────────── |
| 2228 | |
| 2229 | /// The part of a history snapshot that survives a model switch: every |
| 2230 | /// non-system message, with half-finished tool plumbing repaired. |
| 2231 | /// |
| 2232 | /// System messages are dropped because the backend being switched *to* seeds |
| 2233 | /// its own — a different model's prompt, or a freshly pushed session-context |
| 2234 | /// message. A switch can land mid-turn, after the assistant asked for a tool |
| 2235 | /// but before the results came back; a tool call with no matching result (or a |
| 2236 | /// result with no matching call) makes strict OpenAI-compatible endpoints |
| 2237 | /// reject every later request in the session, so those halves are stripped |
| 2238 | /// instead of carried. |
| 2239 | pub fn carryover_history(snapshot: Vec<serde_json::Value>) -> Vec<serde_json::Value> { |
| 2240 | let answered: std::collections::HashSet<String> = snapshot |
| 2241 | .iter() |
| 2242 | .filter(|message| message["role"] == "tool") |
| 2243 | .filter_map(|message| message["tool_call_id"].as_str().map(str::to_string)) |
| 2244 | .collect(); |
| 2245 | |
| 2246 | let mut carried: Vec<serde_json::Value> = Vec::with_capacity(snapshot.len()); |
| 2247 | let mut called: std::collections::HashSet<String> = std::collections::HashSet::new(); |
| 2248 | |
| 2249 | for mut message in snapshot { |
| 2250 | match message["role"].as_str().unwrap_or_default() { |
| 2251 | "system" => continue, |
| 2252 | "assistant" => { |
| 2253 | if let Some(calls) = message["tool_calls"].as_array() { |
| 2254 | let kept: Vec<serde_json::Value> = calls |
| 2255 | .iter() |
| 2256 | .filter(|call| call["id"].as_str().is_some_and(|id| answered.contains(id))) |
| 2257 | .cloned() |
| 2258 | .collect(); |
| 2259 | if kept.is_empty() { |
| 2260 | // Only unanswered requests: keep whatever text came |
| 2261 | // with them, drop the message if there was none. |
| 2262 | if let Some(object) = message.as_object_mut() { |
| 2263 | object.remove("tool_calls"); |
| 2264 | } |
| 2265 | if message_text(&message).is_empty() { |
| 2266 | continue; |
| 2267 | } |
| 2268 | } else { |
| 2269 | for call in &kept { |
| 2270 | called.insert(call["id"].as_str().unwrap_or_default().to_string()); |
| 2271 | } |
| 2272 | message["tool_calls"] = serde_json::Value::Array(kept); |
| 2273 | } |
| 2274 | } |
| 2275 | carried.push(message); |
| 2276 | } |
| 2277 | "tool" => { |
| 2278 | let answers_a_kept_call = message["tool_call_id"] |
| 2279 | .as_str() |
| 2280 | .is_some_and(|id| called.contains(id)); |
| 2281 | if answers_a_kept_call { |
| 2282 | carried.push(message); |
| 2283 | } |
| 2284 | } |
| 2285 | _ => carried.push(message), |
| 2286 | } |
| 2287 | } |
| 2288 | |
| 2289 | carried |
| 2290 | } |
| 2291 | |
| 2292 | /// Take a cancelled prompt's user message back out of the live history. |
| 2293 | /// |
| 2294 | /// The client drops a cancelled prompt from its thread, so the model must not |
| 2295 | /// keep answering it. Only the first inference round can cancel with the user |
| 2296 | /// message as the last entry; later rounds end on tool or assistant output. |
| 2297 | /// A prompt that fails with an error is not taken back: the client keeps it, |
| 2298 | /// so the history keeps it too and a model switch carries it over. |
| 2299 | pub async fn forget_trailing_user_message(backend: &dyn InferenceBackend) { |
| 2300 | let mut history = backend.history_snapshot().await; |
| 2301 | if history |
| 2302 | .last() |
| 2303 | .is_some_and(|message| message["role"] == "user") |
| 2304 | { |
| 2305 | history.pop(); |
| 2306 | backend.restore_history(history).await; |
| 2307 | } |
| 2308 | } |
| 2309 | |
| 2310 | /// Replay `carried` (from [`carryover_history`]) into `backend`, on top of the |
| 2311 | /// system messages `backend` seeded for itself. Used when a model switch |
| 2312 | /// installs a new backend — or reloads the on-device engine, which wipes its |
| 2313 | /// history — so the thread continues under the new model instead of restarting. |
| 2314 | pub async fn adopt_carryover(backend: &dyn InferenceBackend, carried: Vec<serde_json::Value>) { |
| 2315 | if carried.is_empty() { |
| 2316 | return; |
| 2317 | } |
| 2318 | let mut rebuilt: Vec<serde_json::Value> = backend |
| 2319 | .history_snapshot() |
| 2320 | .await |
| 2321 | .into_iter() |
| 2322 | .take_while(|message| message["role"] == "system") |
| 2323 | .collect(); |
| 2324 | rebuilt.extend(carried); |
| 2325 | backend.restore_history(rebuilt).await; |
| 2326 | } |
| 2327 | |
| 2328 | #[cfg(test)] |
| 2329 | mod tests { |
| 2330 | use super::*; |
| 2331 | |
| 2332 | /// The case from the issue: the endpoint says something a person can act |
| 2333 | /// on, and the status code and JSON around it are noise. |
| 2334 | #[test] |
| 2335 | fn an_endpoints_own_message_is_what_the_user_sees() { |
| 2336 | let body = r#"{"error":{"message":"Monthly siGit Code Cloud allowance reached. It resets at the start of your next billing period.","type":"server_error"}}"#; |
| 2337 | |
| 2338 | let described = describe_api_error(reqwest::StatusCode::TOO_MANY_REQUESTS, body); |
| 2339 | |
| 2340 | assert_eq!( |
| 2341 | described, |
| 2342 | "Monthly siGit Code Cloud allowance reached. \ |
| 2343 | It resets at the start of your next billing period." |
| 2344 | ); |
| 2345 | } |
| 2346 | |
| 2347 | #[test] |
| 2348 | fn a_body_that_is_not_an_error_envelope_keeps_the_status() { |
| 2349 | let described = describe_api_error(reqwest::StatusCode::BAD_GATEWAY, "upstream said no"); |
| 2350 | |
| 2351 | assert!(described.contains("502"), "{described}"); |
| 2352 | assert!(described.contains("upstream said no"), "{described}"); |
| 2353 | } |
| 2354 | |
| 2355 | /// The issue-#125 screenshots: a gateway's HTML error page pasted into |
| 2356 | /// chat tells the person nothing the status doesn't already say. |
| 2357 | #[test] |
| 2358 | fn an_html_error_page_is_not_pasted_into_the_message() { |
| 2359 | let page = "<!doctype html>\n<html lang=\"en\">\n<head>\n\ |
| 2360 | <title>We're sorry, but something went wrong (500 Internal Server Error)</title>\n\ |
| 2361 | </head>\n<body>…</body>\n</html>"; |
| 2362 | |
| 2363 | let described = describe_api_error(reqwest::StatusCode::GATEWAY_TIMEOUT, page); |
| 2364 | |
| 2365 | assert_eq!(described, "endpoint returned 504 Gateway Timeout"); |
| 2366 | } |
| 2367 | |
| 2368 | #[test] |
| 2369 | fn an_empty_body_still_says_something() { |
| 2370 | let described = describe_api_error(reqwest::StatusCode::INTERNAL_SERVER_ERROR, " "); |
| 2371 | |
| 2372 | assert_eq!(described, "endpoint returned 500 Internal Server Error"); |
| 2373 | } |
| 2374 | |
| 2375 | #[test] |
| 2376 | fn an_oversized_body_is_cut_down() { |
| 2377 | let body = "x".repeat(ERROR_BODY_LIMIT * 3); |
| 2378 | |
| 2379 | let described = describe_api_error(reqwest::StatusCode::BAD_GATEWAY, &body); |
| 2380 | |
| 2381 | assert!(described.len() < ERROR_BODY_LIMIT + 100, "{described}"); |
| 2382 | assert!(described.ends_with('…'), "{described}"); |
| 2383 | } |
| 2384 | |
| 2385 | /// An envelope with nothing useful in it must not shadow the status, which |
| 2386 | /// would leave the user with an empty error. |
| 2387 | #[test] |
| 2388 | fn an_envelope_with_a_blank_message_falls_back() { |
| 2389 | let described = describe_api_error( |
| 2390 | reqwest::StatusCode::FORBIDDEN, |
| 2391 | r#"{"error":{"message":" "}}"#, |
| 2392 | ); |
| 2393 | |
| 2394 | assert!(described.contains("403"), "{described}"); |
| 2395 | } |
| 2396 | |
| 2397 | #[test] |
| 2398 | fn tools_json_wraps_function_schema() { |
| 2399 | let tools = vec![ToolSpec { |
| 2400 | name: "read_file".to_string(), |
| 2401 | description: "Read a file".to_string(), |
| 2402 | parameters_schema: r#"{"type":"object","properties":{"path":{"type":"string"}}}"# |
| 2403 | .to_string(), |
| 2404 | }]; |
| 2405 | let json = OpenAiBackend::tools_json(&tools); |
| 2406 | assert_eq!(json[0]["type"], "function"); |
| 2407 | assert_eq!(json[0]["function"]["name"], "read_file"); |
| 2408 | assert_eq!( |
| 2409 | json[0]["function"]["parameters"]["properties"]["path"]["type"], |
| 2410 | "string" |
| 2411 | ); |
| 2412 | } |
| 2413 | |
| 2414 | #[test] |
| 2415 | fn malformed_schema_falls_back_to_empty_object() { |
| 2416 | let tools = vec![ToolSpec { |
| 2417 | name: "x".to_string(), |
| 2418 | description: String::new(), |
| 2419 | parameters_schema: "not json".to_string(), |
| 2420 | }]; |
| 2421 | let json = OpenAiBackend::tools_json(&tools); |
| 2422 | assert_eq!(json[0]["function"]["parameters"]["type"], "object"); |
| 2423 | } |
| 2424 | |
| 2425 | #[test] |
| 2426 | fn streamed_assistant_history_omits_empty_tool_calls() { |
| 2427 | let value = streamed_assistant_history("hello", &[]); |
| 2428 | assert_eq!(value["role"], "assistant"); |
| 2429 | assert_eq!(value["content"], "hello"); |
| 2430 | assert!(value.get("tool_calls").is_none()); |
| 2431 | } |
| 2432 | |
| 2433 | #[test] |
| 2434 | fn streamed_assistant_history_preserves_tool_calls() { |
| 2435 | let calls = vec![ToolCall { |
| 2436 | id: "call_0".to_string(), |
| 2437 | name: "read_file".to_string(), |
| 2438 | arguments: r#"{"path":"a.rs"}"#.to_string(), |
| 2439 | }]; |
| 2440 | let value = streamed_assistant_history("", &calls); |
| 2441 | assert!(value["content"].is_null()); |
| 2442 | assert_eq!(value["tool_calls"][0]["id"], "call_0"); |
| 2443 | assert_eq!(value["tool_calls"][0]["type"], "function"); |
| 2444 | assert_eq!(value["tool_calls"][0]["function"]["name"], "read_file"); |
| 2445 | assert_eq!( |
| 2446 | value["tool_calls"][0]["function"]["arguments"], |
| 2447 | r#"{"path":"a.rs"}"# |
| 2448 | ); |
| 2449 | } |
| 2450 | |
| 2451 | #[tokio::test] |
| 2452 | async fn cancelled_tool_results_close_out_history() { |
| 2453 | let backend = OpenAiBackend::new("http://localhost", "", "test-model", None); |
| 2454 | backend |
| 2455 | .history |
| 2456 | .lock() |
| 2457 | .await |
| 2458 | .push(streamed_assistant_history( |
| 2459 | "", |
| 2460 | &[ToolCall { |
| 2461 | id: "call_9".to_string(), |
| 2462 | name: "run_command".to_string(), |
| 2463 | arguments: r#"{"command":"ls"}"#.to_string(), |
| 2464 | }], |
| 2465 | )); |
| 2466 | |
| 2467 | backend |
| 2468 | .record_cancelled_tool_results(vec![ToolResult { |
| 2469 | tool_call_id: "call_9".to_string(), |
| 2470 | content: "cancelled by the user".to_string(), |
| 2471 | }]) |
| 2472 | .await; |
| 2473 | |
| 2474 | let history = backend.history.lock().await; |
| 2475 | let last = history.last().unwrap(); |
| 2476 | assert_eq!(last["role"], "tool"); |
| 2477 | assert_eq!(last["tool_call_id"], "call_9"); |
| 2478 | assert_eq!(last["content"], "cancelled by the user"); |
| 2479 | |
| 2480 | drop(history); |
| 2481 | backend |
| 2482 | .record_cancelled_tool_results(vec![ToolResult { |
| 2483 | tool_call_id: "call_9".to_string(), |
| 2484 | content: "duplicate cleanup".to_string(), |
| 2485 | }]) |
| 2486 | .await; |
| 2487 | let history = backend.history.lock().await; |
| 2488 | assert_eq!( |
| 2489 | history |
| 2490 | .iter() |
| 2491 | .filter(|message| message["tool_call_id"] == "call_9") |
| 2492 | .count(), |
| 2493 | 1 |
| 2494 | ); |
| 2495 | } |
| 2496 | |
| 2497 | #[test] |
| 2498 | fn carryover_drops_system_messages_and_keeps_the_turns() { |
| 2499 | let snapshot = vec![ |
| 2500 | serde_json::json!({ "role": "system", "content": "old prompt" }), |
| 2501 | serde_json::json!({ "role": "user", "content": "hello" }), |
| 2502 | serde_json::json!({ "role": "assistant", "content": "hi" }), |
| 2503 | ]; |
| 2504 | |
| 2505 | let carried = carryover_history(snapshot); |
| 2506 | |
| 2507 | assert_eq!(carried.len(), 2); |
| 2508 | assert_eq!(carried[0]["role"], "user"); |
| 2509 | assert_eq!(carried[1]["content"], "hi"); |
| 2510 | } |
| 2511 | |
| 2512 | #[test] |
| 2513 | fn carryover_keeps_a_trailing_user_message_from_a_failed_turn() { |
| 2514 | // The client keeps a prompt whose inference errored, so a switch must |
| 2515 | // not drop it (issue #124). |
| 2516 | let carried = carryover_history(vec![ |
| 2517 | serde_json::json!({ "role": "system", "content": "prompt" }), |
| 2518 | serde_json::json!({ "role": "user", "content": "completed question" }), |
| 2519 | serde_json::json!({ "role": "assistant", "content": "completed answer" }), |
| 2520 | serde_json::json!({ "role": "user", "content": "failed question" }), |
| 2521 | ]); |
| 2522 | |
| 2523 | assert_eq!(carried.len(), 3, "{carried:#?}"); |
| 2524 | assert_eq!(carried.last().unwrap()["content"], "failed question"); |
| 2525 | } |
| 2526 | |
| 2527 | #[tokio::test] |
| 2528 | async fn forget_trailing_user_message_drops_only_a_cancelled_prompt() { |
| 2529 | let backend = OpenAiBackend::new("http://localhost", "", "m", None); |
| 2530 | backend |
| 2531 | .restore_history(vec![ |
| 2532 | serde_json::json!({ "role": "user", "content": "completed question" }), |
| 2533 | serde_json::json!({ "role": "assistant", "content": "completed answer" }), |
| 2534 | serde_json::json!({ "role": "user", "content": "cancelled question" }), |
| 2535 | ]) |
| 2536 | .await; |
| 2537 | |
| 2538 | forget_trailing_user_message(&backend).await; |
| 2539 | |
| 2540 | let history = backend.history_snapshot().await; |
| 2541 | assert_eq!(history.len(), 2, "{history:#?}"); |
| 2542 | assert_eq!(history[1]["content"], "completed answer"); |
| 2543 | |
| 2544 | forget_trailing_user_message(&backend).await; |
| 2545 | assert_eq!(backend.history_snapshot().await.len(), 2); |
| 2546 | } |
| 2547 | |
| 2548 | #[test] |
| 2549 | fn carryover_strips_a_tool_call_that_never_got_a_result() { |
| 2550 | // Switching mid-turn: the assistant asked for a tool, nothing answered. |
| 2551 | let snapshot = vec![ |
| 2552 | serde_json::json!({ "role": "user", "content": "read a.rs" }), |
| 2553 | streamed_assistant_history( |
| 2554 | "on it", |
| 2555 | &[ToolCall { |
| 2556 | id: "call_1".to_string(), |
| 2557 | name: "read_file".to_string(), |
| 2558 | arguments: "{}".to_string(), |
| 2559 | }], |
| 2560 | ), |
| 2561 | ]; |
| 2562 | |
| 2563 | let carried = carryover_history(snapshot); |
| 2564 | |
| 2565 | assert_eq!(carried.len(), 2); |
| 2566 | assert_eq!(carried[1]["content"], "on it"); |
| 2567 | assert!( |
| 2568 | carried[1].get("tool_calls").is_none(), |
| 2569 | "an unanswered tool call must not survive the switch" |
| 2570 | ); |
| 2571 | } |
| 2572 | |
| 2573 | #[test] |
| 2574 | fn carryover_drops_a_textless_unanswered_tool_call_and_its_late_result() { |
| 2575 | let snapshot = vec![ |
| 2576 | streamed_assistant_history( |
| 2577 | "", |
| 2578 | &[ToolCall { |
| 2579 | id: "call_1".to_string(), |
| 2580 | name: "read_file".to_string(), |
| 2581 | arguments: "{}".to_string(), |
| 2582 | }], |
| 2583 | ), |
| 2584 | // An orphan: its assistant message is gone with the line above. |
| 2585 | serde_json::json!({ |
| 2586 | "role": "tool", |
| 2587 | "tool_call_id": "call_other", |
| 2588 | "content": "file contents", |
| 2589 | }), |
| 2590 | ]; |
| 2591 | |
| 2592 | assert!(carryover_history(snapshot).is_empty()); |
| 2593 | } |
| 2594 | |
| 2595 | #[test] |
| 2596 | fn carryover_keeps_an_answered_tool_call_paired_with_its_result() { |
| 2597 | let snapshot = vec![ |
| 2598 | streamed_assistant_history( |
| 2599 | "", |
| 2600 | &[ToolCall { |
| 2601 | id: "call_1".to_string(), |
| 2602 | name: "read_file".to_string(), |
| 2603 | arguments: "{}".to_string(), |
| 2604 | }], |
| 2605 | ), |
| 2606 | serde_json::json!({ |
| 2607 | "role": "tool", |
| 2608 | "tool_call_id": "call_1", |
| 2609 | "content": "file contents", |
| 2610 | }), |
| 2611 | ]; |
| 2612 | |
| 2613 | let carried = carryover_history(snapshot); |
| 2614 | |
| 2615 | assert_eq!(carried.len(), 2); |
| 2616 | assert_eq!(carried[0]["tool_calls"][0]["id"], "call_1"); |
| 2617 | assert_eq!(carried[1]["tool_call_id"], "call_1"); |
| 2618 | } |
| 2619 | |
| 2620 | #[tokio::test] |
| 2621 | async fn adopt_carryover_replays_the_thread_under_the_new_system_prompt() { |
| 2622 | let new_backend = |
| 2623 | OpenAiBackend::new("http://localhost", "", "m", Some("new prompt".into())); |
| 2624 | |
| 2625 | let carried = carryover_history(vec![ |
| 2626 | serde_json::json!({ "role": "system", "content": "old prompt" }), |
| 2627 | serde_json::json!({ "role": "user", "content": "hello" }), |
| 2628 | serde_json::json!({ "role": "assistant", "content": "hi" }), |
| 2629 | ]); |
| 2630 | adopt_carryover(&new_backend, carried).await; |
| 2631 | |
| 2632 | let history = new_backend.history_snapshot().await; |
| 2633 | assert_eq!(history.len(), 3); |
| 2634 | assert_eq!(history[0]["role"], "system"); |
| 2635 | assert_eq!(history[0]["content"], "new prompt"); |
| 2636 | assert_eq!(history[1]["content"], "hello"); |
| 2637 | assert_eq!(history[2]["content"], "hi"); |
| 2638 | } |
| 2639 | |
| 2640 | #[tokio::test] |
| 2641 | async fn adopt_carryover_leaves_a_fresh_backend_alone() { |
| 2642 | let new_backend = |
| 2643 | OpenAiBackend::new("http://localhost", "", "m", Some("new prompt".into())); |
| 2644 | |
| 2645 | adopt_carryover(&new_backend, Vec::new()).await; |
| 2646 | |
| 2647 | let history = new_backend.history_snapshot().await; |
| 2648 | assert_eq!(history.len(), 1); |
| 2649 | assert_eq!(history[0]["content"], "new prompt"); |
| 2650 | } |
| 2651 | |
| 2652 | fn user_with_image(text: &str, payload_chars: usize) -> serde_json::Value { |
| 2653 | serde_json::json!({ |
| 2654 | "role": "user", |
| 2655 | "content": [ |
| 2656 | { "type": "text", "text": text }, |
| 2657 | { "type": "image_url", "image_url": { |
| 2658 | "url": format!("data:image/png;base64,{}", "A".repeat(payload_chars)), |
| 2659 | }}, |
| 2660 | ], |
| 2661 | }) |
| 2662 | } |
| 2663 | |
| 2664 | #[test] |
| 2665 | fn user_message_uses_content_parts_only_when_there_is_an_image() { |
| 2666 | assert_eq!( |
| 2667 | user_message("hello", &[], &[]), |
| 2668 | serde_json::json!({ "role": "user", "content": "hello" }) |
| 2669 | ); |
| 2670 | |
| 2671 | let image = ImageInput { |
| 2672 | mime_type: "image/png".to_string(), |
| 2673 | data: "AAAA".to_string(), |
| 2674 | }; |
| 2675 | assert_eq!( |
| 2676 | user_message("what is this?", std::slice::from_ref(&image), &[]), |
| 2677 | serde_json::json!({ |
| 2678 | "role": "user", |
| 2679 | "content": [ |
| 2680 | { "type": "text", "text": "what is this?" }, |
| 2681 | { "type": "image_url", "image_url": { "url": "data:image/png;base64,AAAA" } }, |
| 2682 | ], |
| 2683 | }) |
| 2684 | ); |
| 2685 | // An image with no text sends no empty text part. |
| 2686 | let only_image = user_message("", &[image], &[]); |
| 2687 | assert_eq!(only_image["content"].as_array().unwrap().len(), 1); |
| 2688 | } |
| 2689 | |
| 2690 | #[test] |
| 2691 | fn audio_is_sent_as_an_input_audio_part() { |
| 2692 | let clip = AudioInput { |
| 2693 | mime_type: "audio/mpeg".to_string(), |
| 2694 | data: "BBBB".to_string(), |
| 2695 | }; |
| 2696 | assert_eq!( |
| 2697 | user_message("transcribe this", &[], &[clip]), |
| 2698 | serde_json::json!({ |
| 2699 | "role": "user", |
| 2700 | "content": [ |
| 2701 | { "type": "text", "text": "transcribe this" }, |
| 2702 | { "type": "input_audio", "input_audio": { "data": "BBBB", "format": "mp3" } }, |
| 2703 | ], |
| 2704 | }) |
| 2705 | ); |
| 2706 | } |
| 2707 | |
| 2708 | #[test] |
| 2709 | fn audio_format_follows_the_mime_type() { |
| 2710 | let format = |mime: &str| { |
| 2711 | AudioInput { |
| 2712 | mime_type: mime.to_string(), |
| 2713 | data: String::new(), |
| 2714 | } |
| 2715 | .format() |
| 2716 | }; |
| 2717 | assert_eq!(format("audio/wav"), "wav"); |
| 2718 | assert_eq!(format("audio/x-wav"), "wav"); |
| 2719 | assert_eq!(format("audio/wave"), "wav"); |
| 2720 | assert_eq!(format("audio/mpeg"), "mp3"); |
| 2721 | assert_eq!(format("Audio/MP3; codecs=mp3"), "mp3"); |
| 2722 | assert_eq!(format("audio/flac"), "flac"); |
| 2723 | } |
| 2724 | |
| 2725 | #[test] |
| 2726 | fn a_model_without_audio_gets_a_note_and_keeps_the_image() { |
| 2727 | let history = vec![serde_json::json!({ |
| 2728 | "role": "user", |
| 2729 | "content": [ |
| 2730 | { "type": "text", "text": "what is this?" }, |
| 2731 | { "type": "image_url", "image_url": { "url": "data:image/png;base64,AAAA" } }, |
| 2732 | { "type": "input_audio", "input_audio": { "data": "BBBB", "format": "wav" } }, |
| 2733 | ], |
| 2734 | })]; |
| 2735 | |
| 2736 | let sent = without_audio(&history); |
| 2737 | assert_eq!( |
| 2738 | sent[0]["content"], |
| 2739 | serde_json::json!([ |
| 2740 | { "type": "text", "text": "what is this?" }, |
| 2741 | { "type": "image_url", "image_url": { "url": "data:image/png;base64,AAAA" } }, |
| 2742 | { "type": "text", "text": AUDIO_OMITTED_NOTE }, |
| 2743 | ]) |
| 2744 | ); |
| 2745 | // With neither, the message is plain text again. |
| 2746 | assert_eq!( |
| 2747 | without_images(&sent)[0]["content"], |
| 2748 | format!("what is this?\n{AUDIO_OMITTED_NOTE}\n{IMAGE_OMITTED_NOTE}") |
| 2749 | ); |
| 2750 | // History keeps the clip for a model that can take it later. |
| 2751 | assert_eq!(message_audio_count(&history[0]), 1); |
| 2752 | } |
| 2753 | |
| 2754 | #[test] |
| 2755 | fn a_text_only_model_is_sent_a_note_in_place_of_each_image() { |
| 2756 | let history = vec![ |
| 2757 | serde_json::json!({ "role": "system", "content": "prompt" }), |
| 2758 | user_with_image("what is this?", 16), |
| 2759 | serde_json::json!({ "role": "assistant", "content": "a cat" }), |
| 2760 | ]; |
| 2761 | let sent = without_images(&history); |
| 2762 | |
| 2763 | assert_eq!(sent[0], history[0]); |
| 2764 | assert_eq!(sent[2], history[2]); |
| 2765 | assert_eq!( |
| 2766 | sent[1]["content"], |
| 2767 | format!("what is this?\n{IMAGE_OMITTED_NOTE}") |
| 2768 | ); |
| 2769 | // History keeps the image for a model that can read it later. |
| 2770 | assert_eq!(message_image_count(&history[1]), 1); |
| 2771 | } |
| 2772 | |
| 2773 | #[test] |
| 2774 | fn a_remote_backend_reads_images_by_model_id() { |
| 2775 | let own_endpoint = OpenAiBackend::new("http://localhost", "", "gpt-4o-mini", None); |
| 2776 | assert!(own_endpoint.accepts_images()); |
| 2777 | let text_tier = OpenAiBackend::new("http://localhost", "", "onde-nova", None); |
| 2778 | assert!(!text_tier.accepts_images()); |
| 2779 | let image_tier = OpenAiBackend::new("http://localhost", "", "onde-large", None); |
| 2780 | assert!(image_tier.accepts_images()); |
| 2781 | } |
| 2782 | |
| 2783 | #[test] |
| 2784 | fn a_remote_backend_takes_audio_only_on_the_users_own_endpoint() { |
| 2785 | let own_endpoint = OpenAiBackend::new("http://localhost", "", "gpt-4o-audio", None); |
| 2786 | assert!(own_endpoint.accepts_audio()); |
| 2787 | let cloud_tier = OpenAiBackend::new("http://localhost", "", "onde-large", None); |
| 2788 | assert!(!cloud_tier.accepts_audio()); |
| 2789 | } |
| 2790 | |
| 2791 | #[test] |
| 2792 | fn message_text_reads_both_content_shapes() { |
| 2793 | assert_eq!( |
| 2794 | message_text(&serde_json::json!({ "role": "user", "content": "hello" })), |
| 2795 | "hello" |
| 2796 | ); |
| 2797 | let with_image = user_with_image("what is this?", 16); |
| 2798 | assert_eq!(message_text(&with_image), "what is this?"); |
| 2799 | assert_eq!(message_image_count(&with_image), 1); |
| 2800 | // A tool-call-only assistant message has no content at all. |
| 2801 | let no_content = serde_json::json!({ "role": "assistant", "content": null }); |
| 2802 | assert_eq!(message_text(&no_content), ""); |
| 2803 | assert_eq!(message_image_count(&no_content), 0); |
| 2804 | } |
| 2805 | |
| 2806 | #[test] |
| 2807 | fn estimate_tokens_does_not_count_an_image_payload_as_text() { |
| 2808 | // Two megabytes of base64 would read as half a million tokens. |
| 2809 | let history = vec![user_with_image("what is this?", 2_000_000)]; |
| 2810 | let estimate = estimate_tokens(&history); |
| 2811 | assert!( |
| 2812 | (IMAGE_TOKEN_ESTIMATE..IMAGE_TOKEN_ESTIMATE + 100).contains(&estimate), |
| 2813 | "{estimate}" |
| 2814 | ); |
| 2815 | } |
| 2816 | |
| 2817 | #[test] |
| 2818 | fn summary_transcript_keeps_the_text_of_a_message_with_an_image() { |
| 2819 | let transcript = transcript_for_summary(&[user_with_image("what is this?", 64)]); |
| 2820 | assert!(transcript.contains("what is this?"), "{transcript}"); |
| 2821 | assert!(transcript.contains("[attached an image]"), "{transcript}"); |
| 2822 | assert!(!transcript.contains("base64"), "{transcript}"); |
| 2823 | } |
| 2824 | |
| 2825 | #[test] |
| 2826 | fn carryover_keeps_a_user_message_that_carries_an_image() { |
| 2827 | let carried = carryover_history(vec![ |
| 2828 | serde_json::json!({ "role": "system", "content": "prompt" }), |
| 2829 | user_with_image("what is this?", 16), |
| 2830 | serde_json::json!({ "role": "assistant", "content": "a cat" }), |
| 2831 | ]); |
| 2832 | assert_eq!(carried.len(), 2); |
| 2833 | assert_eq!(message_image_count(&carried[0]), 1); |
| 2834 | } |
| 2835 | |
| 2836 | #[test] |
| 2837 | fn estimate_tokens_scales_with_serialized_size() { |
| 2838 | assert_eq!(estimate_tokens(&[]), 0); |
| 2839 | |
| 2840 | let short = vec![serde_json::json!({ "role": "user", "content": "hi" })]; |
| 2841 | let long = vec![serde_json::json!({ "role": "user", "content": "x".repeat(4_000) })]; |
| 2842 | let short_estimate = estimate_tokens(&short); |
| 2843 | let long_estimate = estimate_tokens(&long); |
| 2844 | |
| 2845 | assert!(short_estimate > 0, "non-empty history estimates > 0 tokens"); |
| 2846 | assert!(long_estimate > short_estimate, "longer history costs more"); |
| 2847 | // 4,000 content chars / 4 ≈ 1,000 tokens, plus a little JSON framing. |
| 2848 | assert!((1_000..1_100).contains(&long_estimate), "{long_estimate}"); |
| 2849 | } |
| 2850 | |
| 2851 | #[tokio::test] |
| 2852 | async fn openai_snapshot_restore_round_trips_exactly() { |
| 2853 | let backend = OpenAiBackend::new("http://localhost", "", "m", Some("be helpful".into())); |
| 2854 | { |
| 2855 | let mut history = backend.history.lock().await; |
| 2856 | history.push(serde_json::json!({ "role": "user", "content": "hello" })); |
| 2857 | history.push(streamed_assistant_history( |
| 2858 | "", |
| 2859 | &[ToolCall { |
| 2860 | id: "call_1".to_string(), |
| 2861 | name: "read_file".to_string(), |
| 2862 | arguments: r#"{"path":"a.rs"}"#.to_string(), |
| 2863 | }], |
| 2864 | )); |
| 2865 | history.push(serde_json::json!({ |
| 2866 | "role": "tool", "tool_call_id": "call_1", "content": "fn main() {}", |
| 2867 | })); |
| 2868 | history.push(serde_json::json!({ "role": "assistant", "content": "done" })); |
| 2869 | } |
| 2870 | let snapshot = backend.history_snapshot().await; |
| 2871 | assert_eq!( |
| 2872 | snapshot[0]["role"], "system", |
| 2873 | "snapshot keeps the system message" |
| 2874 | ); |
| 2875 | |
| 2876 | // Restoring into a backend seeded with a *different* system prompt must |
| 2877 | // replace everything, including that seed. |
| 2878 | let restored = OpenAiBackend::new("http://localhost", "", "m", Some("other seed".into())); |
| 2879 | restored.restore_history(snapshot.clone()).await; |
| 2880 | assert_eq!(restored.history_snapshot().await, snapshot); |
| 2881 | } |
| 2882 | |
| 2883 | #[tokio::test] |
| 2884 | async fn local_restore_replaces_engine_history_and_system_context() { |
| 2885 | let engine = Arc::new(ChatEngine::new()); |
| 2886 | let backend = LocalBackend::new(Arc::clone(&engine)); |
| 2887 | backend |
| 2888 | .restore_history(vec![ |
| 2889 | serde_json::json!({ "role": "system", "content": "thread A context" }), |
| 2890 | serde_json::json!({ "role": "user", "content": "thread A question" }), |
| 2891 | ]) |
| 2892 | .await; |
| 2893 | |
| 2894 | backend |
| 2895 | .restore_history(vec![ |
| 2896 | serde_json::json!({ "role": "system", "content": "thread B context" }), |
| 2897 | serde_json::json!({ "role": "user", "content": "thread B question" }), |
| 2898 | ]) |
| 2899 | .await; |
| 2900 | |
| 2901 | let restored = backend.history_snapshot().await; |
| 2902 | assert_eq!(restored.len(), 2, "{restored:#?}"); |
| 2903 | assert_eq!(restored[0]["content"], "thread B context"); |
| 2904 | assert_eq!(restored[1]["content"], "thread B question"); |
| 2905 | assert!( |
| 2906 | restored.iter().all(|message| !message["content"] |
| 2907 | .as_str() |
| 2908 | .unwrap_or_default() |
| 2909 | .contains("thread A")), |
| 2910 | "restore_history must replace, not append to, the local engine: {restored:#?}" |
| 2911 | ); |
| 2912 | |
| 2913 | let error = backend |
| 2914 | .send_message_with_tools("must not run", &[], None) |
| 2915 | .await |
| 2916 | .unwrap_err(); |
| 2917 | assert!(error.contains("not ready"), "{error}"); |
| 2918 | assert_eq!(backend.history_snapshot().await, restored); |
| 2919 | } |
| 2920 | |
| 2921 | /// Minimal scripted OpenAI-compatible endpoint: accepts one HTTP request on |
| 2922 | /// a std listener and answers with a fixed non-streaming completion. The |
| 2923 | /// receiver yields the request body the backend actually put on the wire. |
| 2924 | fn spawn_completion_stub( |
| 2925 | summary: &str, |
| 2926 | ) -> (std::net::SocketAddr, std::sync::mpsc::Receiver<String>) { |
| 2927 | spawn_message_stub(serde_json::json!({ "role": "assistant", "content": summary })) |
| 2928 | } |
| 2929 | |
| 2930 | /// Like [`spawn_completion_stub`], answering with an arbitrary message. |
| 2931 | fn spawn_message_stub( |
| 2932 | message: serde_json::Value, |
| 2933 | ) -> (std::net::SocketAddr, std::sync::mpsc::Receiver<String>) { |
| 2934 | spawn_choice_stub(serde_json::json!({ "message": message })) |
| 2935 | } |
| 2936 | |
| 2937 | /// Like [`spawn_message_stub`], answering with a whole choice, so a test |
| 2938 | /// can set what sits beside the message. |
| 2939 | fn spawn_choice_stub( |
| 2940 | choice: serde_json::Value, |
| 2941 | ) -> (std::net::SocketAddr, std::sync::mpsc::Receiver<String>) { |
| 2942 | use std::io::{Read, Write}; |
| 2943 | |
| 2944 | let listener = std::net::TcpListener::bind("127.0.0.1:0").unwrap(); |
| 2945 | let addr = listener.local_addr().unwrap(); |
| 2946 | let (sender, receiver) = std::sync::mpsc::channel(); |
| 2947 | let body = serde_json::json!({ "choices": [choice] }).to_string(); |
| 2948 | std::thread::spawn(move || { |
| 2949 | let (mut stream, _) = listener.accept().unwrap(); |
| 2950 | // Read until the full request (headers + content-length body) is in. |
| 2951 | let mut request = Vec::new(); |
| 2952 | let mut chunk = [0u8; 4096]; |
| 2953 | loop { |
| 2954 | let n = stream.read(&mut chunk).unwrap_or(0); |
| 2955 | if n == 0 { |
| 2956 | break; |
| 2957 | } |
| 2958 | request.extend_from_slice(&chunk[..n]); |
| 2959 | if let Some(headers_end) = |
| 2960 | request.windows(4).position(|window| window == b"\r\n\r\n") |
| 2961 | { |
| 2962 | let headers = String::from_utf8_lossy(&request[..headers_end]); |
| 2963 | let content_length = headers |
| 2964 | .lines() |
| 2965 | .find_map(|line| { |
| 2966 | line.to_ascii_lowercase() |
| 2967 | .strip_prefix("content-length:") |
| 2968 | .map(|value| value.trim().parse::<usize>().unwrap_or(0)) |
| 2969 | }) |
| 2970 | .unwrap_or(0); |
| 2971 | if request.len() >= headers_end + 4 + content_length { |
| 2972 | let _ = sender.send( |
| 2973 | String::from_utf8_lossy(&request[headers_end + 4..]).into_owned(), |
| 2974 | ); |
| 2975 | break; |
| 2976 | } |
| 2977 | } |
| 2978 | } |
| 2979 | let response = format!( |
| 2980 | "HTTP/1.1 200 OK\r\ncontent-type: application/json\r\n\ |
| 2981 | content-length: {}\r\nconnection: close\r\n\r\n{}", |
| 2982 | body.len(), |
| 2983 | body |
| 2984 | ); |
| 2985 | let _ = stream.write_all(response.as_bytes()); |
| 2986 | }); |
| 2987 | (addr, receiver) |
| 2988 | } |
| 2989 | |
| 2990 | /// Scripted SSE endpoint: one response per connection, in order. The |
| 2991 | /// receiver yields each request body. |
| 2992 | fn spawn_sse_sequence_stub( |
| 2993 | replies: Vec<Vec<String>>, |
| 2994 | ) -> (std::net::SocketAddr, std::sync::mpsc::Receiver<String>) { |
| 2995 | use std::io::{Read, Write}; |
| 2996 | |
| 2997 | let listener = std::net::TcpListener::bind("127.0.0.1:0").unwrap(); |
| 2998 | let addr = listener.local_addr().unwrap(); |
| 2999 | let (sender, receiver) = std::sync::mpsc::channel(); |
| 3000 | std::thread::spawn(move || { |
| 3001 | for deltas in replies { |
| 3002 | let (mut stream, _) = listener.accept().unwrap(); |
| 3003 | let mut request = Vec::new(); |
| 3004 | let mut buf = [0u8; 4096]; |
| 3005 | loop { |
| 3006 | let n = stream.read(&mut buf).unwrap_or(0); |
| 3007 | if n == 0 { |
| 3008 | break; |
| 3009 | } |
| 3010 | request.extend_from_slice(&buf[..n]); |
| 3011 | if let Some(end) = request.windows(4).position(|w| w == b"\r\n\r\n") { |
| 3012 | let headers = String::from_utf8_lossy(&request[..end]).to_ascii_lowercase(); |
| 3013 | let length = headers |
| 3014 | .lines() |
| 3015 | .find_map(|l| l.strip_prefix("content-length:")?.trim().parse().ok()) |
| 3016 | .unwrap_or(0usize); |
| 3017 | if request.len() >= end + 4 + length { |
| 3018 | let _ = sender |
| 3019 | .send(String::from_utf8_lossy(&request[end + 4..]).into_owned()); |
| 3020 | break; |
| 3021 | } |
| 3022 | } |
| 3023 | } |
| 3024 | let mut body = String::new(); |
| 3025 | for delta in deltas { |
| 3026 | let frame = serde_json::json!({ |
| 3027 | "choices": [{ "delta": { "content": delta } }] |
| 3028 | }); |
| 3029 | body.push_str(&format!("data: {frame}\n\n")); |
| 3030 | } |
| 3031 | body.push_str("data: [DONE]\n\n"); |
| 3032 | let response = format!( |
| 3033 | "HTTP/1.1 200 OK\r\ncontent-type: text/event-stream\r\n\ |
| 3034 | content-length: {}\r\nconnection: close\r\n\r\n{body}", |
| 3035 | body.len() |
| 3036 | ); |
| 3037 | let _ = stream.write_all(response.as_bytes()); |
| 3038 | } |
| 3039 | }); |
| 3040 | (addr, receiver) |
| 3041 | } |
| 3042 | |
| 3043 | /// Issue #123: a streamed reply that degenerates into a repetition loop is |
| 3044 | /// cut off, history keeps only what came before the loop, and the model |
| 3045 | /// gets one retry. |
| 3046 | #[tokio::test] |
| 3047 | async fn a_repetition_loop_is_cut_and_retried_once() { |
| 3048 | let intro = "Let me look at the commits. "; |
| 3049 | let looped: Vec<String> = std::iter::once(intro.to_string()) |
| 3050 | .chain( |
| 3051 | crate::repetition::issue_loop_text() |
| 3052 | .as_bytes() |
| 3053 | .chunks(40) |
| 3054 | .map(|c| String::from_utf8_lossy(c).into_owned()), |
| 3055 | ) |
| 3056 | .collect(); |
| 3057 | let (addr, requests) = |
| 3058 | spawn_sse_sequence_stub(vec![looped, vec!["Here is the answer.".to_string()]]); |
| 3059 | let backend = |
| 3060 | OpenAiBackend::new(format!("http://{addr}/v1"), "test-key", "test-model", None); |
| 3061 | let (sink, mut tokens) = tokio::sync::mpsc::unbounded_channel(); |
| 3062 | let tools = vec![ToolSpec { |
| 3063 | name: "run_command".to_string(), |
| 3064 | description: "Run a command".to_string(), |
| 3065 | parameters_schema: r#"{"type":"object","properties":{}}"#.to_string(), |
| 3066 | }]; |
| 3067 | |
| 3068 | let result = backend |
| 3069 | .send_message_with_tools("what changed", &tools, Some(&sink)) |
| 3070 | .await |
| 3071 | .unwrap(); |
| 3072 | |
| 3073 | assert_eq!(result.finish, FinishReason::Complete); |
| 3074 | assert!(result.text.starts_with(intro.trim_end()), "{}", result.text); |
| 3075 | assert!(result.text.ends_with("Here is the answer.")); |
| 3076 | assert!( |
| 3077 | result.text.len() < 600, |
| 3078 | "junk kept: {} bytes", |
| 3079 | result.text.len() |
| 3080 | ); |
| 3081 | |
| 3082 | let first = requests.recv().unwrap(); |
| 3083 | assert!(!first.contains(REPETITION_RETRY)); |
| 3084 | let second = requests.recv().unwrap(); |
| 3085 | assert!(second.contains(REPETITION_RETRY)); |
| 3086 | |
| 3087 | let history = backend.history_snapshot().await; |
| 3088 | let kept = history |
| 3089 | .iter() |
| 3090 | .find(|m| m["role"] == "assistant") |
| 3091 | .and_then(|m| m["content"].as_str()) |
| 3092 | .unwrap(); |
| 3093 | assert!( |
| 3094 | kept.len() < 600, |
| 3095 | "history kept the loop: {} bytes", |
| 3096 | kept.len() |
| 3097 | ); |
| 3098 | |
| 3099 | drop(sink); |
| 3100 | let mut shown = String::new(); |
| 3101 | while let Ok(chunk) = tokens.try_recv() { |
| 3102 | if let TokenChunk::Visible(text) = chunk { |
| 3103 | shown.push_str(&text); |
| 3104 | } |
| 3105 | } |
| 3106 | assert!(shown.contains("stopped this reply"), "{shown}"); |
| 3107 | } |
| 3108 | |
| 3109 | /// Serves `responses` in order, one per connection, as `(status, |
| 3110 | /// content type, body)`, and hands back each request body. |
| 3111 | fn spawn_scripted_stub( |
| 3112 | responses: Vec<(u16, &'static str, String)>, |
| 3113 | ) -> (std::net::SocketAddr, std::sync::mpsc::Receiver<String>) { |
| 3114 | use std::io::{Read, Write}; |
| 3115 | |
| 3116 | let listener = std::net::TcpListener::bind("127.0.0.1:0").unwrap(); |
| 3117 | let addr = listener.local_addr().unwrap(); |
| 3118 | let (sender, receiver) = std::sync::mpsc::channel(); |
| 3119 | std::thread::spawn(move || { |
| 3120 | for (status, content_type, body) in responses { |
| 3121 | let Ok((mut stream, _)) = listener.accept() else { |
| 3122 | return; |
| 3123 | }; |
| 3124 | let mut request = Vec::new(); |
| 3125 | let mut chunk = [0u8; 4096]; |
| 3126 | loop { |
| 3127 | let n = stream.read(&mut chunk).unwrap_or(0); |
| 3128 | if n == 0 { |
| 3129 | break; |
| 3130 | } |
| 3131 | request.extend_from_slice(&chunk[..n]); |
| 3132 | if let Some(headers_end) = |
| 3133 | request.windows(4).position(|window| window == b"\r\n\r\n") |
| 3134 | { |
| 3135 | let headers = String::from_utf8_lossy(&request[..headers_end]); |
| 3136 | let content_length = headers |
| 3137 | .lines() |
| 3138 | .find_map(|line| { |
| 3139 | line.to_ascii_lowercase() |
| 3140 | .strip_prefix("content-length:") |
| 3141 | .map(|value| value.trim().parse::<usize>().unwrap_or(0)) |
| 3142 | }) |
| 3143 | .unwrap_or(0); |
| 3144 | if request.len() >= headers_end + 4 + content_length { |
| 3145 | let _ = sender.send( |
| 3146 | String::from_utf8_lossy(&request[headers_end + 4..]).into_owned(), |
| 3147 | ); |
| 3148 | break; |
| 3149 | } |
| 3150 | } |
| 3151 | } |
| 3152 | let response = format!( |
| 3153 | "HTTP/1.1 {status} X\r\ncontent-type: {content_type}\r\n\ |
| 3154 | content-length: {}\r\nconnection: close\r\n\r\n{}", |
| 3155 | body.len(), |
| 3156 | body |
| 3157 | ); |
| 3158 | let _ = stream.write_all(response.as_bytes()); |
| 3159 | } |
| 3160 | }); |
| 3161 | (addr, receiver) |
| 3162 | } |
| 3163 | |
| 3164 | fn sse(events: &[serde_json::Value]) -> String { |
| 3165 | let mut body: String = events |
| 3166 | .iter() |
| 3167 | .map(|event| format!("data: {event}\n\n")) |
| 3168 | .collect(); |
| 3169 | body.push_str("data: [DONE]\n\n"); |
| 3170 | body |
| 3171 | } |
| 3172 | |
| 3173 | /// A stream asks for token usage and reads it off the last chunk, the |
| 3174 | /// one with no choices, for the client's context meter (issue #194). |
| 3175 | #[tokio::test] |
| 3176 | async fn a_stream_reports_the_usage_it_was_asked_for() { |
| 3177 | let (addr, requests) = spawn_scripted_stub(vec![( |
| 3178 | 200, |
| 3179 | "text/event-stream", |
| 3180 | sse(&[ |
| 3181 | serde_json::json!({"choices": [{"delta": {"content": "Hi."}}]}), |
| 3182 | serde_json::json!({ |
| 3183 | "choices": [], |
| 3184 | "usage": {"prompt_tokens": 1200, "completion_tokens": 34, "total_tokens": 1234}, |
| 3185 | }), |
| 3186 | ]), |
| 3187 | )]); |
| 3188 | let backend = |
| 3189 | OpenAiBackend::new(format!("http://{addr}/v1"), "test-key", "test-model", None); |
| 3190 | let (sink, _chunks) = tokio::sync::mpsc::unbounded_channel(); |
| 3191 | |
| 3192 | let result = backend |
| 3193 | .send_message_with_tools("hello", &[], Some(&sink)) |
| 3194 | .await |
| 3195 | .unwrap(); |
| 3196 | |
| 3197 | assert_eq!(result.text, "Hi."); |
| 3198 | assert_eq!(result.context_tokens, Some(1234)); |
| 3199 | let request: serde_json::Value = serde_json::from_str(&requests.recv().unwrap()).unwrap(); |
| 3200 | assert_eq!(request["stream_options"]["include_usage"], true); |
| 3201 | } |
| 3202 | |
| 3203 | #[tokio::test] |
| 3204 | async fn a_json_reply_reports_its_usage() { |
| 3205 | let body = serde_json::json!({ |
| 3206 | "choices": [{"message": {"role": "assistant", "content": "Hi."}}], |
| 3207 | "usage": {"prompt_tokens": 90, "completion_tokens": 10}, |
| 3208 | }) |
| 3209 | .to_string(); |
| 3210 | let (addr, requests) = spawn_scripted_stub(vec![(200, "application/json", body)]); |
| 3211 | let backend = |
| 3212 | OpenAiBackend::new(format!("http://{addr}/v1"), "test-key", "test-model", None); |
| 3213 | |
| 3214 | let result = backend |
| 3215 | .send_message_with_tools("hello", &[], None) |
| 3216 | .await |
| 3217 | .unwrap(); |
| 3218 | |
| 3219 | // No total given: prompt plus completion. |
| 3220 | assert_eq!(result.context_tokens, Some(100)); |
| 3221 | let request: serde_json::Value = serde_json::from_str(&requests.recv().unwrap()).unwrap(); |
| 3222 | assert!(request.get("stream_options").is_none(), "{request}"); |
| 3223 | } |
| 3224 | |
| 3225 | /// An endpoint that predates `stream_options` is asked again without it, |
| 3226 | /// and not asked for usage again after that. |
| 3227 | #[tokio::test] |
| 3228 | async fn an_endpoint_that_refuses_stream_options_still_streams() { |
| 3229 | let refusal = serde_json::json!({ |
| 3230 | "error": {"message": "Unrecognized request argument supplied: stream_options"}, |
| 3231 | }) |
| 3232 | .to_string(); |
| 3233 | let reply = sse(&[serde_json::json!({"choices": [{"delta": {"content": "Hi."}}]})]); |
| 3234 | let (addr, requests) = spawn_scripted_stub(vec![ |
| 3235 | (400, "application/json", refusal), |
| 3236 | (200, "text/event-stream", reply.clone()), |
| 3237 | (200, "text/event-stream", reply), |
| 3238 | ]); |
| 3239 | let backend = |
| 3240 | OpenAiBackend::new(format!("http://{addr}/v1"), "test-key", "test-model", None); |
| 3241 | let (sink, _chunks) = tokio::sync::mpsc::unbounded_channel(); |
| 3242 | |
| 3243 | let first = backend |
| 3244 | .send_message_with_tools("hello", &[], Some(&sink)) |
| 3245 | .await |
| 3246 | .unwrap(); |
| 3247 | assert_eq!(first.text, "Hi."); |
| 3248 | assert_eq!(first.context_tokens, None); |
| 3249 | let second = backend |
| 3250 | .send_message_with_tools("again", &[], Some(&sink)) |
| 3251 | .await |
| 3252 | .unwrap(); |
| 3253 | assert_eq!(second.text, "Hi."); |
| 3254 | |
| 3255 | let bodies: Vec<serde_json::Value> = requests |
| 3256 | .try_iter() |
| 3257 | .map(|body| serde_json::from_str(&body).unwrap()) |
| 3258 | .collect(); |
| 3259 | assert_eq!(bodies.len(), 3); |
| 3260 | assert!(bodies[0].get("stream_options").is_some()); |
| 3261 | assert!(bodies[1].get("stream_options").is_none()); |
| 3262 | assert!(bodies[2].get("stream_options").is_none()); |
| 3263 | } |
| 3264 | |
| 3265 | #[test] |
| 3266 | fn finish_reasons_map_from_the_wire() { |
| 3267 | for reason in [None, Some("stop"), Some("tool_calls"), Some("unheard-of")] { |
| 3268 | assert_eq!(FinishReason::from_wire(reason), FinishReason::Complete); |
| 3269 | } |
| 3270 | assert_eq!( |
| 3271 | FinishReason::from_wire(Some("length")), |
| 3272 | FinishReason::Length |
| 3273 | ); |
| 3274 | assert_eq!( |
| 3275 | FinishReason::from_wire(Some("content_filter")), |
| 3276 | FinishReason::ContentFilter |
| 3277 | ); |
| 3278 | } |
| 3279 | |
| 3280 | /// The non-streaming path (headless runs, subagents) reads the reason off |
| 3281 | /// the choice; a reply cut off at the token limit must not look finished. |
| 3282 | #[tokio::test] |
| 3283 | async fn a_truncated_json_reply_reports_length() { |
| 3284 | let (addr, _requests) = spawn_choice_stub(serde_json::json!({ |
| 3285 | "message": { "role": "assistant", "content": "The answer is" }, |
| 3286 | "finish_reason": "length", |
| 3287 | })); |
| 3288 | let backend = |
| 3289 | OpenAiBackend::new(format!("http://{addr}/v1"), "test-key", "test-model", None); |
| 3290 | |
| 3291 | let result = backend |
| 3292 | .send_message_with_tools("explain", &[], None) |
| 3293 | .await |
| 3294 | .unwrap(); |
| 3295 | |
| 3296 | assert_eq!(result.text, "The answer is"); |
| 3297 | assert_eq!(result.finish, FinishReason::Length); |
| 3298 | } |
| 3299 | |
| 3300 | /// A structured call can come back with a broken inline block in the same |
| 3301 | /// message. The call runs; the block reaches neither the reply nor the |
| 3302 | /// history, where the model would read back a call that never ran. |
| 3303 | #[tokio::test] |
| 3304 | async fn a_broken_inline_block_beside_a_structured_call_is_dropped() { |
| 3305 | let (addr, _requests) = spawn_message_stub(serde_json::json!({ |
| 3306 | "role": "assistant", |
| 3307 | "content": "Checking.<tool_call>run_command CheckStatus=nope</tool_call>", |
| 3308 | "tool_calls": [{ |
| 3309 | "id": "call_1", |
| 3310 | "type": "function", |
| 3311 | "function": { "name": "run_command", "arguments": "{\"command\":\"pwd\"}" }, |
| 3312 | }], |
| 3313 | })); |
| 3314 | let backend = |
| 3315 | OpenAiBackend::new(format!("http://{addr}/v1"), "test-key", "test-model", None); |
| 3316 | let tools = vec![ToolSpec { |
| 3317 | name: "run_command".to_string(), |
| 3318 | description: "Run a command".to_string(), |
| 3319 | parameters_schema: r#"{"type":"object","properties":{"command":{"type":"string"}}}"# |
| 3320 | .to_string(), |
| 3321 | }]; |
| 3322 | |
| 3323 | let result = backend |
| 3324 | .send_message_with_tools("where am I", &tools, None) |
| 3325 | .await |
| 3326 | .unwrap(); |
| 3327 | |
| 3328 | assert_eq!(result.text, "Checking."); |
| 3329 | assert_eq!(result.tool_calls.len(), 1); |
| 3330 | assert_eq!(result.tool_calls[0].id, "call_1"); |
| 3331 | let history = backend.history_snapshot().await; |
| 3332 | let reply = history.last().unwrap(); |
| 3333 | assert_eq!(reply["content"], "Checking."); |
| 3334 | assert_eq!(reply["tool_calls"][0]["id"], "call_1"); |
| 3335 | } |
| 3336 | |
| 3337 | /// A structured call plus a well-formed inline block: both run, and the |
| 3338 | /// inline markup leaves the reply and history, as on the streaming path. |
| 3339 | #[tokio::test] |
| 3340 | async fn a_well_formed_inline_block_beside_a_structured_call_also_runs() { |
| 3341 | let (addr, _requests) = spawn_message_stub(serde_json::json!({ |
| 3342 | "role": "assistant", |
| 3343 | "content": "Checking.<tool_call>run_command<arg_key>command</arg_key><arg_value>ls</arg_value></tool_call>", |
| 3344 | "tool_calls": [{ |
| 3345 | "id": "call_1", |
| 3346 | "type": "function", |
| 3347 | "function": { "name": "run_command", "arguments": "{\"command\":\"pwd\"}" }, |
| 3348 | }], |
| 3349 | })); |
| 3350 | let backend = |
| 3351 | OpenAiBackend::new(format!("http://{addr}/v1"), "test-key", "test-model", None); |
| 3352 | let tools = vec![ToolSpec { |
| 3353 | name: "run_command".to_string(), |
| 3354 | description: "Run a command".to_string(), |
| 3355 | parameters_schema: r#"{"type":"object","properties":{"command":{"type":"string"}}}"# |
| 3356 | .to_string(), |
| 3357 | }]; |
| 3358 | |
| 3359 | let result = backend |
| 3360 | .send_message_with_tools("where am I", &tools, None) |
| 3361 | .await |
| 3362 | .unwrap(); |
| 3363 | |
| 3364 | assert_eq!(result.text, "Checking."); |
| 3365 | let ids: Vec<&str> = result.tool_calls.iter().map(|c| c.id.as_str()).collect(); |
| 3366 | assert_eq!(ids, ["call_1", "call_recovered_1"]); |
| 3367 | assert_eq!(result.tool_calls[1].arguments, r#"{"command":"ls"}"#); |
| 3368 | let history = backend.history_snapshot().await; |
| 3369 | let reply = history.last().unwrap(); |
| 3370 | assert_eq!(reply["content"], "Checking."); |
| 3371 | assert_eq!(reply["tool_calls"].as_array().unwrap().len(), 2); |
| 3372 | } |
| 3373 | |
| 3374 | #[tokio::test] |
| 3375 | async fn compact_history_rebuilds_system_summary_and_tail() { |
| 3376 | let (addr, _requests) = spawn_completion_stub("We refactored backend.rs; tests pass."); |
| 3377 | let backend = OpenAiBackend::new( |
| 3378 | format!("http://{addr}/v1"), |
| 3379 | "test-key", |
| 3380 | "test-model", |
| 3381 | Some("be helpful".into()), |
| 3382 | ); |
| 3383 | { |
| 3384 | let mut history = backend.history.lock().await; |
| 3385 | for i in 0..5 { |
| 3386 | let role = if i % 2 == 0 { "user" } else { "assistant" }; |
| 3387 | history.push(serde_json::json!({ |
| 3388 | "role": role, "content": format!("message {i}"), |
| 3389 | })); |
| 3390 | } |
| 3391 | } |
| 3392 | |
| 3393 | backend.compact_history(2).await.unwrap(); |
| 3394 | |
| 3395 | let history = backend.history_snapshot().await; |
| 3396 | assert_eq!(history.len(), 4, "system + summary + last 2: {history:?}"); |
| 3397 | assert_eq!(history[0]["role"], "system"); |
| 3398 | assert_eq!(history[0]["content"], "be helpful"); |
| 3399 | assert_eq!(history[1]["role"], "user"); |
| 3400 | let summary_text = history[1]["content"].as_str().unwrap(); |
| 3401 | assert!(summary_text.starts_with("[Conversation summary]\n")); |
| 3402 | assert!(summary_text.contains("We refactored backend.rs; tests pass.")); |
| 3403 | assert_eq!( |
| 3404 | history[2], |
| 3405 | serde_json::json!({ "role": "assistant", "content": "message 3" }) |
| 3406 | ); |
| 3407 | assert_eq!( |
| 3408 | history[3], |
| 3409 | serde_json::json!({ "role": "user", "content": "message 4" }) |
| 3410 | ); |
| 3411 | } |
| 3412 | |
| 3413 | /// Compacting a tool-heavy session must not put tool shapes on the wire. |
| 3414 | /// The summarization round offers no `tools`, and endpoints reject tool |
| 3415 | /// calls and tool results that arrive without a schema — which used to make |
| 3416 | /// compaction fail forever in any session that had run a single tool. |
| 3417 | #[tokio::test] |
| 3418 | async fn compact_history_sends_no_tool_artifacts() { |
| 3419 | let (addr, requests) = spawn_completion_stub("Ran git status on main."); |
| 3420 | let backend = OpenAiBackend::new( |
| 3421 | format!("http://{addr}/v1"), |
| 3422 | "test-key", |
| 3423 | "test-model", |
| 3424 | Some("be helpful".into()), |
| 3425 | ); |
| 3426 | { |
| 3427 | let mut history = backend.history.lock().await; |
| 3428 | history.push(serde_json::json!({ "role": "user", "content": "check the repo" })); |
| 3429 | history.push(serde_json::json!({ |
| 3430 | "role": "assistant", |
| 3431 | "content": null, |
| 3432 | "tool_calls": [{ |
| 3433 | "id": "call_1", |
| 3434 | "type": "function", |
| 3435 | "function": { "name": "run_command", "arguments": "{\"command\":\"git status\"}" }, |
| 3436 | }], |
| 3437 | })); |
| 3438 | history.push(serde_json::json!({ |
| 3439 | "role": "tool", |
| 3440 | "tool_call_id": "call_1", |
| 3441 | "content": "on branch main", |
| 3442 | })); |
| 3443 | } |
| 3444 | |
| 3445 | backend.compact_history(2).await.unwrap(); |
| 3446 | |
| 3447 | let body: serde_json::Value = |
| 3448 | serde_json::from_str(&requests.recv().unwrap()).expect("request body is JSON"); |
| 3449 | assert!( |
| 3450 | body.get("tools").is_none(), |
| 3451 | "summarization offers no tools: {body}" |
| 3452 | ); |
| 3453 | for message in body["messages"].as_array().unwrap() { |
| 3454 | assert!( |
| 3455 | message.get("tool_calls").is_none(), |
| 3456 | "no tool_calls may be sent without a schema: {message}" |
| 3457 | ); |
| 3458 | assert_ne!( |
| 3459 | message["role"], "tool", |
| 3460 | "no tool results may be sent without a schema: {message}" |
| 3461 | ); |
| 3462 | } |
| 3463 | |
| 3464 | // The tool round still has to survive into the summary request as prose, |
| 3465 | // or the summary loses the work the session actually did. |
| 3466 | let transcript = body["messages"].as_array().unwrap().last().unwrap()["content"] |
| 3467 | .as_str() |
| 3468 | .unwrap(); |
| 3469 | assert!(transcript.contains("called run_command({\"command\":\"git status\"})")); |
| 3470 | assert!(transcript.contains("tool result: on branch main")); |
| 3471 | } |
| 3472 | |
| 3473 | /// Compaction only fails when even the truncation fallback cannot fit the |
| 3474 | /// history in the budget — here the newest message alone is bigger than |
| 3475 | /// the whole window. |
| 3476 | #[tokio::test] |
| 3477 | async fn compact_history_failure_leaves_history_intact() { |
| 3478 | // No listener at this address, so summarization fails; the system |
| 3479 | // prompt alone is bigger than the window, so even the minimal |
| 3480 | // fallback rebuild cannot fit — the genuinely unrecoverable case. |
| 3481 | let backend = OpenAiBackend::new( |
| 3482 | "http://127.0.0.1:9", |
| 3483 | "", |
| 3484 | "test-model", |
| 3485 | Some("s".repeat(DEFAULT_CONTEXT_TOKEN_BUDGET * 4 + 1000)), |
| 3486 | ); |
| 3487 | backend |
| 3488 | .history |
| 3489 | .lock() |
| 3490 | .await |
| 3491 | .push(serde_json::json!({ "role": "user", "content": "hello" })); |
| 3492 | let before = backend.history_snapshot().await; |
| 3493 | |
| 3494 | assert!(backend.compact_history(2).await.is_err()); |
| 3495 | assert_eq!(backend.history_snapshot().await, before); |
| 3496 | } |
| 3497 | |
| 3498 | /// The issue-#125 path: the summarization endpoint is down, and the |
| 3499 | /// session must still shrink rather than die with "start a new thread". |
| 3500 | #[tokio::test] |
| 3501 | async fn failed_summarization_falls_back_to_truncation() { |
| 3502 | // Nothing listens here, so the summarization round always fails. |
| 3503 | let backend = |
| 3504 | OpenAiBackend::new("http://127.0.0.1:9", "", "test-model", Some("sys".into())); |
| 3505 | { |
| 3506 | let mut history = backend.history.lock().await; |
| 3507 | for i in 0..20 { |
| 3508 | history.push(serde_json::json!({ |
| 3509 | "role": "user", |
| 3510 | "content": format!("message {i}: {}", "padding ".repeat(200)), |
| 3511 | })); |
| 3512 | } |
| 3513 | } |
| 3514 | |
| 3515 | backend.compact_history(4).await.unwrap(); |
| 3516 | |
| 3517 | let history = backend.history_snapshot().await; |
| 3518 | assert_eq!(history[0]["role"], "system", "{history:?}"); |
| 3519 | assert_eq!(history[1]["content"], TRUNCATION_PLACEHOLDER, "{history:?}"); |
| 3520 | // The newest messages survive; the oldest were dropped to fit. |
| 3521 | assert_eq!( |
| 3522 | history.last().unwrap()["content"] |
| 3523 | .as_str() |
| 3524 | .unwrap() |
| 3525 | .split(':') |
| 3526 | .next() |
| 3527 | .unwrap(), |
| 3528 | "message 19", |
| 3529 | "{history:?}" |
| 3530 | ); |
| 3531 | assert!( |
| 3532 | estimate_tokens(&history) <= DEFAULT_CONTEXT_TOKEN_BUDGET, |
| 3533 | "fallback result must fit the budget: ~{} tokens", |
| 3534 | estimate_tokens(&history) |
| 3535 | ); |
| 3536 | } |
| 3537 | |
| 3538 | /// A history small enough to begin with keeps its whole tail: the |
| 3539 | /// fallback drops nothing beyond what `keep_last` already folded away. |
| 3540 | #[tokio::test] |
| 3541 | async fn truncation_fallback_keeps_the_newest_messages_that_fit() { |
| 3542 | let system = vec![serde_json::json!({ "role": "system", "content": "sys" })]; |
| 3543 | let tail: Vec<serde_json::Value> = (0..4) |
| 3544 | .map(|i| serde_json::json!({ "role": "user", "content": format!("m{i}") })) |
| 3545 | .collect(); |
| 3546 | |
| 3547 | let rebuilt = truncate_history_to_budget(&system, tail).unwrap(); |
| 3548 | |
| 3549 | assert_eq!(rebuilt[0]["role"], "system"); |
| 3550 | assert_eq!(rebuilt[1]["content"], TRUNCATION_PLACEHOLDER); |
| 3551 | assert_eq!(rebuilt.len(), 6, "{rebuilt:?}"); |
| 3552 | assert_eq!(rebuilt[2]["content"], "m0"); |
| 3553 | assert_eq!(rebuilt[5]["content"], "m3"); |
| 3554 | } |
| 3555 | |
| 3556 | /// A multi-system-message session keeps every leading system message in |
| 3557 | /// the fallback, exactly like the summarization-success path. |
| 3558 | #[test] |
| 3559 | fn truncation_fallback_keeps_all_leading_system_messages() { |
| 3560 | let system = vec![ |
| 3561 | serde_json::json!({ "role": "system", "content": "seed prompt" }), |
| 3562 | serde_json::json!({ "role": "system", "content": "session context" }), |
| 3563 | ]; |
| 3564 | let tail = vec![serde_json::json!({ "role": "user", "content": "hi" })]; |
| 3565 | |
| 3566 | let rebuilt = truncate_history_to_budget(&system, tail).unwrap(); |
| 3567 | |
| 3568 | assert_eq!( |
| 3569 | rebuilt[0]["content"], "seed prompt", |
| 3570 | "first system message kept: {rebuilt:?}" |
| 3571 | ); |
| 3572 | assert_eq!( |
| 3573 | rebuilt[1]["content"], "session context", |
| 3574 | "second system message kept: {rebuilt:?}" |
| 3575 | ); |
| 3576 | } |
| 3577 | |
| 3578 | #[test] |
| 3579 | fn truncation_fallback_drops_orphaned_tool_results_with_their_call() { |
| 3580 | // A caller that split mid-round hands in a tail starting with tool |
| 3581 | // results whose assistant call message already fell outside the kept |
| 3582 | // window; those orphans must go before anything else, or strict |
| 3583 | // endpoints reject the next request. |
| 3584 | let tail = vec![ |
| 3585 | serde_json::json!({ "role": "tool", "tool_call_id": "call_1", "content": "ok" }), |
| 3586 | serde_json::json!({ "role": "user", "content": "thanks" }), |
| 3587 | serde_json::json!({ "role": "assistant", "content": "done" }), |
| 3588 | ]; |
| 3589 | |
| 3590 | let rebuilt = truncate_history_to_budget(&[], tail).unwrap(); |
| 3591 | |
| 3592 | let roles: Vec<&str> = rebuilt.iter().filter_map(|m| m["role"].as_str()).collect(); |
| 3593 | assert_eq!( |
| 3594 | roles, |
| 3595 | ["user", "user", "assistant"], |
| 3596 | "placeholder + the surviving exchange, orphan gone: {rebuilt:?}" |
| 3597 | ); |
| 3598 | } |
| 3599 | |
| 3600 | /// Review follow-up: a tail that drains to empty (one oversized tool |
| 3601 | /// round) must still test the minimal rebuild — system + placeholder fits, |
| 3602 | /// so the session is recoverable, not "unfixable". |
| 3603 | #[test] |
| 3604 | fn truncation_fallback_recovers_when_the_whole_tail_was_one_oversized_round() { |
| 3605 | let system = vec![serde_json::json!({ "role": "system", "content": "sys" })]; |
| 3606 | let tail = vec![ |
| 3607 | serde_json::json!({ |
| 3608 | "role": "assistant", |
| 3609 | "content": null, |
| 3610 | "tool_calls": [{ |
| 3611 | "id": "call_1", "type": "function", |
| 3612 | "function": { "name": "read_file", "arguments": "{}" }, |
| 3613 | }], |
| 3614 | }), |
| 3615 | serde_json::json!({ |
| 3616 | "role": "tool", |
| 3617 | "tool_call_id": "call_1", |
| 3618 | "content": "y".repeat(DEFAULT_CONTEXT_TOKEN_BUDGET * 4 + 1000), |
| 3619 | }), |
| 3620 | ]; |
| 3621 | |
| 3622 | let rebuilt = truncate_history_to_budget(&system, tail).unwrap(); |
| 3623 | |
| 3624 | assert_eq!(rebuilt.len(), 2, "{rebuilt:?}"); |
| 3625 | assert_eq!(rebuilt[0]["role"], "system"); |
| 3626 | assert_eq!(rebuilt[1]["content"], TRUNCATION_PLACEHOLDER); |
| 3627 | } |
| 3628 | |
| 3629 | #[test] |
| 3630 | fn truncation_fallback_reports_when_nothing_fits() { |
| 3631 | // Even the minimal rebuild (system + placeholder, no conversation) |
| 3632 | // overflows: the system messages alone are bigger than the window. |
| 3633 | let system = vec![serde_json::json!({ |
| 3634 | "role": "system", |
| 3635 | "content": "s".repeat(DEFAULT_CONTEXT_TOKEN_BUDGET * 4 + 1000), |
| 3636 | })]; |
| 3637 | |
| 3638 | assert!(truncate_history_to_budget(&system, Vec::new()).is_none()); |
| 3639 | } |
| 3640 | |
| 3641 | /// Review follow-up: the local fallback recovers from a newest message |
| 3642 | /// that overflows the window on its own, the same way the remote one does. |
| 3643 | #[test] |
| 3644 | fn local_truncation_fallback_drops_an_oversized_final_message() { |
| 3645 | let system = vec![ChatMessage::system("sys")]; |
| 3646 | let tail = vec![ |
| 3647 | ChatMessage::user("hi"), |
| 3648 | ChatMessage::user("y".repeat(DEFAULT_CONTEXT_TOKEN_BUDGET * 4 + 1000)), |
| 3649 | ]; |
| 3650 | |
| 3651 | let rebuilt = truncate_chat_messages_to_budget(system, tail).unwrap(); |
| 3652 | |
| 3653 | assert_eq!(rebuilt.len(), 2, "{rebuilt:?}"); |
| 3654 | assert_eq!(rebuilt[0].content, "sys"); |
| 3655 | assert_eq!(rebuilt[1].content, TRUNCATION_PLACEHOLDER); |
| 3656 | } |
| 3657 | |
| 3658 | #[test] |
| 3659 | fn local_truncation_fallback_keeps_the_newest_messages_that_fit() { |
| 3660 | let big = DEFAULT_CONTEXT_TOKEN_BUDGET * 4 / 2; |
| 3661 | let tail = vec![ |
| 3662 | ChatMessage::user("a".repeat(big)), |
| 3663 | ChatMessage::user("b".repeat(big)), |
| 3664 | ChatMessage::user("newest"), |
| 3665 | ]; |
| 3666 | |
| 3667 | let rebuilt = truncate_chat_messages_to_budget(Vec::new(), tail).unwrap(); |
| 3668 | |
| 3669 | assert_eq!(rebuilt.len(), 3, "oldest dropped, rest kept"); |
| 3670 | assert_eq!(rebuilt[0].content, TRUNCATION_PLACEHOLDER); |
| 3671 | assert!(rebuilt[1].content.starts_with('b')); |
| 3672 | assert_eq!(rebuilt[2].content, "newest"); |
| 3673 | } |
| 3674 | |
| 3675 | #[test] |
| 3676 | fn local_truncation_fallback_reports_when_the_system_messages_overflow() { |
| 3677 | let system = vec![ChatMessage::system( |
| 3678 | "s".repeat(DEFAULT_CONTEXT_TOKEN_BUDGET * 4 + 1000), |
| 3679 | )]; |
| 3680 | |
| 3681 | assert!(truncate_chat_messages_to_budget(system, Vec::new()).is_none()); |
| 3682 | } |
| 3683 | |
| 3684 | /// Review follow-up: head and tail share the budget, so the trimmed |
| 3685 | /// transcript stays near the cap instead of reaching twice it. |
| 3686 | #[test] |
| 3687 | fn truncate_transcript_middle_stays_within_the_cap() { |
| 3688 | let lines: Vec<String> = (0..2_000) |
| 3689 | .map(|i| format!("user: message {i} with some padding text")) |
| 3690 | .collect(); |
| 3691 | let transcript = lines.join("\n"); |
| 3692 | |
| 3693 | let trimmed = truncate_transcript_middle(&transcript, 200); // ~800 chars |
| 3694 | |
| 3695 | assert!(trimmed.contains("message 0 "), "opening kept: {trimmed}"); |
| 3696 | assert!(trimmed.contains("message 1999"), "closing kept: {trimmed}"); |
| 3697 | assert!( |
| 3698 | trimmed.len() <= 800 + 80, |
| 3699 | "cap plus the omission marker, got {} chars", |
| 3700 | trimmed.len() |
| 3701 | ); |
| 3702 | } |
| 3703 | |
| 3704 | #[test] |
| 3705 | fn truncate_transcript_middle_keeps_short_transcripts_verbatim() { |
| 3706 | let transcript = "user: hello\n\nassistant: hi"; |
| 3707 | |
| 3708 | assert_eq!(truncate_transcript_middle(transcript, 1_000), transcript); |
| 3709 | } |
| 3710 | |
| 3711 | /// Review follow-up: a huge final message with no newline in the tail |
| 3712 | /// half must not be silently dropped — the closing exchange is what the |
| 3713 | /// summary most needs. The head gives up its share instead. |
| 3714 | #[test] |
| 3715 | fn truncate_transcript_middle_keeps_a_newlineless_final_message() { |
| 3716 | let mut lines: Vec<String> = (0..100).map(|i| format!("user: message {i}")).collect(); |
| 3717 | lines.push(format!( |
| 3718 | "user: {}current state: pending review", |
| 3719 | "noise ".repeat(600) |
| 3720 | )); |
| 3721 | let transcript = lines.join("\n"); |
| 3722 | |
| 3723 | let trimmed = truncate_transcript_middle(&transcript, 400); // ~1600 chars |
| 3724 | |
| 3725 | assert!( |
| 3726 | trimmed.contains("current state: pending review"), |
| 3727 | "the closing exchange survives: {trimmed}" |
| 3728 | ); |
| 3729 | assert!( |
| 3730 | trimmed.contains("older messages omitted"), |
| 3731 | "the gap is marked: {trimmed}" |
| 3732 | ); |
| 3733 | assert!( |
| 3734 | !trimmed.contains("message 0"), |
| 3735 | "the head gives up its share to the unbreakable tail: {trimmed}" |
| 3736 | ); |
| 3737 | } |
| 3738 | |
| 3739 | /// Review follow-up: no newline anywhere in reach — keep the newest |
| 3740 | /// budget, not the oldest (which would invert the contract while the |
| 3741 | /// placeholder claims older messages were dropped). |
| 3742 | #[test] |
| 3743 | fn truncate_transcript_middle_without_any_newline_keeps_the_newest() { |
| 3744 | let transcript = format!("old decisions {} current state", "x".repeat(4_000)); |
| 3745 | |
| 3746 | let trimmed = truncate_transcript_middle(&transcript, 200); // ~800 chars |
| 3747 | |
| 3748 | assert!( |
| 3749 | trimmed.contains("current state"), |
| 3750 | "the newest content survives: {trimmed}" |
| 3751 | ); |
| 3752 | assert!( |
| 3753 | !trimmed.contains("old decisions"), |
| 3754 | "the oldest goes when only one side can survive: {trimmed}" |
| 3755 | ); |
| 3756 | } |
| 3757 | |
| 3758 | #[test] |
| 3759 | fn truncate_transcript_middle_cuts_the_middle_not_the_ends() { |
| 3760 | let mut lines: Vec<String> = Vec::new(); |
| 3761 | for i in 0..200 { |
| 3762 | lines.push(format!("user: message {i} with some padding text")); |
| 3763 | } |
| 3764 | let transcript = lines.join("\n"); |
| 3765 | |
| 3766 | let trimmed = truncate_transcript_middle(&transcript, 200); // ~800 chars |
| 3767 | |
| 3768 | assert!(trimmed.contains("message 0"), "opening kept: {trimmed}"); |
| 3769 | assert!(trimmed.contains("message 199"), "closing kept: {trimmed}"); |
| 3770 | assert!( |
| 3771 | trimmed.contains("older messages omitted"), |
| 3772 | "gap is marked: {trimmed}" |
| 3773 | ); |
| 3774 | assert!( |
| 3775 | trimmed.len() <= transcript.len(), |
| 3776 | "never grows the transcript" |
| 3777 | ); |
| 3778 | } |
| 3779 | |
| 3780 | #[test] |
| 3781 | fn assistant_message_with_tool_calls_round_trips() { |
| 3782 | let message = ResponseMessage { |
| 3783 | content: None, |
| 3784 | tool_calls: Some(vec![ResponseToolCall { |
| 3785 | id: "call_1".to_string(), |
| 3786 | function: ResponseFunction { |
| 3787 | name: "read_file".to_string(), |
| 3788 | arguments: r#"{"path":"a.rs"}"#.to_string(), |
| 3789 | }, |
| 3790 | }]), |
| 3791 | }; |
| 3792 | let value = message.into_history_value(); |
| 3793 | assert_eq!(value["role"], "assistant"); |
| 3794 | assert!(value["content"].is_null()); |
| 3795 | assert_eq!(value["tool_calls"][0]["id"], "call_1"); |
| 3796 | assert_eq!(value["tool_calls"][0]["type"], "function"); |
| 3797 | assert_eq!(value["tool_calls"][0]["function"]["name"], "read_file"); |
| 3798 | } |
| 3799 | } |