zeta.rs

   1mod completion_diff_element;
   2mod init;
   3mod input_excerpt;
   4mod license_detection;
   5mod onboarding_modal;
   6mod onboarding_telemetry;
   7mod rate_completion_modal;
   8
   9pub(crate) use completion_diff_element::*;
  10use db::kvp::KEY_VALUE_STORE;
  11use http_client::http::{HeaderMap, HeaderValue};
  12pub use init::*;
  13use inline_completion::DataCollectionState;
  14use license_detection::LICENSE_FILES_TO_CHECK;
  15pub use license_detection::is_license_eligible_for_data_collection;
  16pub use rate_completion_modal::*;
  17
  18use anyhow::{Context as _, Result, anyhow};
  19use arrayvec::ArrayVec;
  20use client::{Client, UserStore};
  21use collections::{HashMap, HashSet, VecDeque};
  22use futures::AsyncReadExt;
  23use gpui::{
  24    App, AppContext as _, AsyncApp, Context, Entity, EntityId, Global, SemanticVersion,
  25    Subscription, Task, WeakEntity, actions,
  26};
  27use http_client::{HttpClient, Method};
  28use input_excerpt::excerpt_for_cursor_position;
  29use language::{
  30    Anchor, Buffer, BufferSnapshot, EditPreview, OffsetRangeExt, ToOffset, ToPoint, text_diff,
  31};
  32use language_model::{LlmApiToken, RefreshLlmTokenListener};
  33use postage::watch;
  34use project::Project;
  35use release_channel::AppVersion;
  36use settings::WorktreeId;
  37use std::str::FromStr;
  38use std::{
  39    borrow::Cow,
  40    cmp,
  41    fmt::Write,
  42    future::Future,
  43    mem,
  44    ops::Range,
  45    path::Path,
  46    rc::Rc,
  47    sync::Arc,
  48    time::{Duration, Instant},
  49};
  50use telemetry_events::InlineCompletionRating;
  51use thiserror::Error;
  52use util::ResultExt;
  53use uuid::Uuid;
  54use workspace::Workspace;
  55use workspace::notifications::{ErrorMessagePrompt, NotificationId};
  56use worktree::Worktree;
  57use zed_llm_client::{
  58    EDIT_PREDICTIONS_USAGE_AMOUNT_HEADER_NAME, EDIT_PREDICTIONS_USAGE_LIMIT_HEADER_NAME,
  59    EXPIRED_LLM_TOKEN_HEADER_NAME, MINIMUM_REQUIRED_VERSION_HEADER_NAME, PredictEditsBody,
  60    PredictEditsResponse, UsageLimit,
  61};
  62
  63const CURSOR_MARKER: &'static str = "<|user_cursor_is_here|>";
  64const START_OF_FILE_MARKER: &'static str = "<|start_of_file|>";
  65const EDITABLE_REGION_START_MARKER: &'static str = "<|editable_region_start|>";
  66const EDITABLE_REGION_END_MARKER: &'static str = "<|editable_region_end|>";
  67const BUFFER_CHANGE_GROUPING_INTERVAL: Duration = Duration::from_secs(1);
  68const ZED_PREDICT_DATA_COLLECTION_CHOICE: &str = "zed_predict_data_collection_choice";
  69
  70const MAX_CONTEXT_TOKENS: usize = 150;
  71const MAX_REWRITE_TOKENS: usize = 350;
  72const MAX_EVENT_TOKENS: usize = 500;
  73
  74/// Maximum number of events to track.
  75const MAX_EVENT_COUNT: usize = 16;
  76
  77actions!(edit_prediction, [ClearHistory]);
  78
  79#[derive(Debug, Clone, Copy)]
  80pub struct Usage {
  81    pub limit: UsageLimit,
  82    pub amount: i32,
  83}
  84
  85impl Usage {
  86    pub fn from_headers(headers: &HeaderMap<HeaderValue>) -> Result<Self> {
  87        let limit = headers
  88            .get(EDIT_PREDICTIONS_USAGE_LIMIT_HEADER_NAME)
  89            .ok_or_else(|| {
  90                anyhow!("missing {EDIT_PREDICTIONS_USAGE_LIMIT_HEADER_NAME:?} header")
  91            })?;
  92        let limit = UsageLimit::from_str(limit.to_str()?)?;
  93
  94        let amount = headers
  95            .get(EDIT_PREDICTIONS_USAGE_AMOUNT_HEADER_NAME)
  96            .ok_or_else(|| {
  97                anyhow!("missing {EDIT_PREDICTIONS_USAGE_AMOUNT_HEADER_NAME:?} header")
  98            })?;
  99        let amount = amount.to_str()?.parse::<i32>()?;
 100
 101        Ok(Self { limit, amount })
 102    }
 103}
 104
 105#[derive(Copy, Clone, Default, Debug, PartialEq, Eq, Hash)]
 106pub struct InlineCompletionId(Uuid);
 107
 108impl From<InlineCompletionId> for gpui::ElementId {
 109    fn from(value: InlineCompletionId) -> Self {
 110        gpui::ElementId::Uuid(value.0)
 111    }
 112}
 113
 114impl std::fmt::Display for InlineCompletionId {
 115    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
 116        write!(f, "{}", self.0)
 117    }
 118}
 119
 120#[derive(Clone)]
 121struct ZetaGlobal(Entity<Zeta>);
 122
 123impl Global for ZetaGlobal {}
 124
 125#[derive(Clone)]
 126pub struct InlineCompletion {
 127    id: InlineCompletionId,
 128    path: Arc<Path>,
 129    excerpt_range: Range<usize>,
 130    cursor_offset: usize,
 131    edits: Arc<[(Range<Anchor>, String)]>,
 132    snapshot: BufferSnapshot,
 133    edit_preview: EditPreview,
 134    input_outline: Arc<str>,
 135    input_events: Arc<str>,
 136    input_excerpt: Arc<str>,
 137    output_excerpt: Arc<str>,
 138    request_sent_at: Instant,
 139    response_received_at: Instant,
 140}
 141
 142impl InlineCompletion {
 143    fn latency(&self) -> Duration {
 144        self.response_received_at
 145            .duration_since(self.request_sent_at)
 146    }
 147
 148    fn interpolate(&self, new_snapshot: &BufferSnapshot) -> Option<Vec<(Range<Anchor>, String)>> {
 149        interpolate(&self.snapshot, new_snapshot, self.edits.clone())
 150    }
 151}
 152
 153fn interpolate(
 154    old_snapshot: &BufferSnapshot,
 155    new_snapshot: &BufferSnapshot,
 156    current_edits: Arc<[(Range<Anchor>, String)]>,
 157) -> Option<Vec<(Range<Anchor>, String)>> {
 158    let mut edits = Vec::new();
 159
 160    let mut model_edits = current_edits.into_iter().peekable();
 161    for user_edit in new_snapshot.edits_since::<usize>(&old_snapshot.version) {
 162        while let Some((model_old_range, _)) = model_edits.peek() {
 163            let model_old_range = model_old_range.to_offset(old_snapshot);
 164            if model_old_range.end < user_edit.old.start {
 165                let (model_old_range, model_new_text) = model_edits.next().unwrap();
 166                edits.push((model_old_range.clone(), model_new_text.clone()));
 167            } else {
 168                break;
 169            }
 170        }
 171
 172        if let Some((model_old_range, model_new_text)) = model_edits.peek() {
 173            let model_old_offset_range = model_old_range.to_offset(old_snapshot);
 174            if user_edit.old == model_old_offset_range {
 175                let user_new_text = new_snapshot
 176                    .text_for_range(user_edit.new.clone())
 177                    .collect::<String>();
 178
 179                if let Some(model_suffix) = model_new_text.strip_prefix(&user_new_text) {
 180                    if !model_suffix.is_empty() {
 181                        let anchor = old_snapshot.anchor_after(user_edit.old.end);
 182                        edits.push((anchor..anchor, model_suffix.to_string()));
 183                    }
 184
 185                    model_edits.next();
 186                    continue;
 187                }
 188            }
 189        }
 190
 191        return None;
 192    }
 193
 194    edits.extend(model_edits.cloned());
 195
 196    if edits.is_empty() { None } else { Some(edits) }
 197}
 198
 199impl std::fmt::Debug for InlineCompletion {
 200    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
 201        f.debug_struct("InlineCompletion")
 202            .field("id", &self.id)
 203            .field("path", &self.path)
 204            .field("edits", &self.edits)
 205            .finish_non_exhaustive()
 206    }
 207}
 208
 209pub struct Zeta {
 210    workspace: Option<WeakEntity<Workspace>>,
 211    client: Arc<Client>,
 212    events: VecDeque<Event>,
 213    registered_buffers: HashMap<gpui::EntityId, RegisteredBuffer>,
 214    shown_completions: VecDeque<InlineCompletion>,
 215    rated_completions: HashSet<InlineCompletionId>,
 216    data_collection_choice: Entity<DataCollectionChoice>,
 217    llm_token: LlmApiToken,
 218    _llm_token_subscription: Subscription,
 219    /// Whether the terms of service have been accepted.
 220    tos_accepted: bool,
 221    /// Whether an update to a newer version of Zed is required to continue using Zeta.
 222    update_required: bool,
 223    _user_store_subscription: Subscription,
 224    license_detection_watchers: HashMap<WorktreeId, Rc<LicenseDetectionWatcher>>,
 225}
 226
 227impl Zeta {
 228    pub fn global(cx: &mut App) -> Option<Entity<Self>> {
 229        cx.try_global::<ZetaGlobal>().map(|global| global.0.clone())
 230    }
 231
 232    pub fn register(
 233        workspace: Option<WeakEntity<Workspace>>,
 234        worktree: Option<Entity<Worktree>>,
 235        client: Arc<Client>,
 236        user_store: Entity<UserStore>,
 237        cx: &mut App,
 238    ) -> Entity<Self> {
 239        let this = Self::global(cx).unwrap_or_else(|| {
 240            let entity = cx.new(|cx| Self::new(workspace, client, user_store, cx));
 241            cx.set_global(ZetaGlobal(entity.clone()));
 242            entity
 243        });
 244
 245        this.update(cx, move |this, cx| {
 246            if let Some(worktree) = worktree {
 247                worktree.update(cx, |worktree, cx| {
 248                    this.license_detection_watchers
 249                        .entry(worktree.id())
 250                        .or_insert_with(|| Rc::new(LicenseDetectionWatcher::new(worktree, cx)));
 251                });
 252            }
 253        });
 254
 255        this
 256    }
 257
 258    pub fn clear_history(&mut self) {
 259        self.events.clear();
 260    }
 261
 262    fn new(
 263        workspace: Option<WeakEntity<Workspace>>,
 264        client: Arc<Client>,
 265        user_store: Entity<UserStore>,
 266        cx: &mut Context<Self>,
 267    ) -> Self {
 268        let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
 269
 270        let data_collection_choice = Self::load_data_collection_choices();
 271        let data_collection_choice = cx.new(|_| data_collection_choice);
 272
 273        Self {
 274            workspace,
 275            client,
 276            events: VecDeque::new(),
 277            shown_completions: VecDeque::new(),
 278            rated_completions: HashSet::default(),
 279            registered_buffers: HashMap::default(),
 280            data_collection_choice,
 281            llm_token: LlmApiToken::default(),
 282            _llm_token_subscription: cx.subscribe(
 283                &refresh_llm_token_listener,
 284                |this, _listener, _event, cx| {
 285                    let client = this.client.clone();
 286                    let llm_token = this.llm_token.clone();
 287                    cx.spawn(async move |_this, _cx| {
 288                        llm_token.refresh(&client).await?;
 289                        anyhow::Ok(())
 290                    })
 291                    .detach_and_log_err(cx);
 292                },
 293            ),
 294            tos_accepted: user_store
 295                .read(cx)
 296                .current_user_has_accepted_terms()
 297                .unwrap_or(false),
 298            update_required: false,
 299            _user_store_subscription: cx.subscribe(&user_store, |this, user_store, event, cx| {
 300                match event {
 301                    client::user::Event::PrivateUserInfoUpdated => {
 302                        this.tos_accepted = user_store
 303                            .read(cx)
 304                            .current_user_has_accepted_terms()
 305                            .unwrap_or(false);
 306                    }
 307                    _ => {}
 308                }
 309            }),
 310            license_detection_watchers: HashMap::default(),
 311        }
 312    }
 313
 314    fn push_event(&mut self, event: Event) {
 315        if let Some(Event::BufferChange {
 316            new_snapshot: last_new_snapshot,
 317            timestamp: last_timestamp,
 318            ..
 319        }) = self.events.back_mut()
 320        {
 321            // Coalesce edits for the same buffer when they happen one after the other.
 322            let Event::BufferChange {
 323                old_snapshot,
 324                new_snapshot,
 325                timestamp,
 326            } = &event;
 327
 328            if timestamp.duration_since(*last_timestamp) <= BUFFER_CHANGE_GROUPING_INTERVAL
 329                && old_snapshot.remote_id() == last_new_snapshot.remote_id()
 330                && old_snapshot.version == last_new_snapshot.version
 331            {
 332                *last_new_snapshot = new_snapshot.clone();
 333                *last_timestamp = *timestamp;
 334                return;
 335            }
 336        }
 337
 338        self.events.push_back(event);
 339        if self.events.len() >= MAX_EVENT_COUNT {
 340            self.events.drain(..MAX_EVENT_COUNT / 2);
 341        }
 342    }
 343
 344    pub fn register_buffer(&mut self, buffer: &Entity<Buffer>, cx: &mut Context<Self>) {
 345        let buffer_id = buffer.entity_id();
 346        let weak_buffer = buffer.downgrade();
 347
 348        if let std::collections::hash_map::Entry::Vacant(entry) =
 349            self.registered_buffers.entry(buffer_id)
 350        {
 351            let snapshot = buffer.read(cx).snapshot();
 352
 353            entry.insert(RegisteredBuffer {
 354                snapshot,
 355                _subscriptions: [
 356                    cx.subscribe(buffer, move |this, buffer, event, cx| {
 357                        this.handle_buffer_event(buffer, event, cx);
 358                    }),
 359                    cx.observe_release(buffer, move |this, _buffer, _cx| {
 360                        this.registered_buffers.remove(&weak_buffer.entity_id());
 361                    }),
 362                ],
 363            });
 364        };
 365    }
 366
 367    fn handle_buffer_event(
 368        &mut self,
 369        buffer: Entity<Buffer>,
 370        event: &language::BufferEvent,
 371        cx: &mut Context<Self>,
 372    ) {
 373        if let language::BufferEvent::Edited = event {
 374            self.report_changes_for_buffer(&buffer, cx);
 375        }
 376    }
 377
 378    fn request_completion_impl<F, R>(
 379        &mut self,
 380        workspace: Option<Entity<Workspace>>,
 381        project: Option<&Entity<Project>>,
 382        buffer: &Entity<Buffer>,
 383        cursor: language::Anchor,
 384        can_collect_data: bool,
 385        cx: &mut Context<Self>,
 386        perform_predict_edits: F,
 387    ) -> Task<Result<Option<InlineCompletion>>>
 388    where
 389        F: FnOnce(PerformPredictEditsParams) -> R + 'static,
 390        R: Future<Output = Result<(PredictEditsResponse, Option<Usage>)>> + Send + 'static,
 391    {
 392        let snapshot = self.report_changes_for_buffer(&buffer, cx);
 393        let diagnostic_groups = snapshot.diagnostic_groups(None);
 394        let cursor_point = cursor.to_point(&snapshot);
 395        let cursor_offset = cursor_point.to_offset(&snapshot);
 396        let events = self.events.clone();
 397        let path: Arc<Path> = snapshot
 398            .file()
 399            .map(|f| Arc::from(f.full_path(cx).as_path()))
 400            .unwrap_or_else(|| Arc::from(Path::new("untitled")));
 401
 402        let zeta = cx.entity();
 403        let client = self.client.clone();
 404        let llm_token = self.llm_token.clone();
 405        let app_version = AppVersion::global(cx);
 406
 407        let buffer = buffer.clone();
 408
 409        let local_lsp_store =
 410            project.and_then(|project| project.read(cx).lsp_store().read(cx).as_local());
 411        let diagnostic_groups = if let Some(local_lsp_store) = local_lsp_store {
 412            Some(
 413                diagnostic_groups
 414                    .into_iter()
 415                    .filter_map(|(language_server_id, diagnostic_group)| {
 416                        let language_server =
 417                            local_lsp_store.running_language_server_for_id(language_server_id)?;
 418
 419                        Some((
 420                            language_server.name(),
 421                            diagnostic_group.resolve::<usize>(&snapshot),
 422                        ))
 423                    })
 424                    .collect::<Vec<_>>(),
 425            )
 426        } else {
 427            None
 428        };
 429
 430        cx.spawn(async move |_, cx| {
 431            let request_sent_at = Instant::now();
 432
 433            struct BackgroundValues {
 434                input_events: String,
 435                input_excerpt: String,
 436                speculated_output: String,
 437                editable_range: Range<usize>,
 438                input_outline: String,
 439            }
 440
 441            let values = cx
 442                .background_spawn({
 443                    let snapshot = snapshot.clone();
 444                    let path = path.clone();
 445                    async move {
 446                        let path = path.to_string_lossy();
 447                        let input_excerpt = excerpt_for_cursor_position(
 448                            cursor_point,
 449                            &path,
 450                            &snapshot,
 451                            MAX_REWRITE_TOKENS,
 452                            MAX_CONTEXT_TOKENS,
 453                        );
 454                        let input_events = prompt_for_events(&events, MAX_EVENT_TOKENS);
 455                        let input_outline = prompt_for_outline(&snapshot);
 456
 457                        anyhow::Ok(BackgroundValues {
 458                            input_events,
 459                            input_excerpt: input_excerpt.prompt,
 460                            speculated_output: input_excerpt.speculated_output,
 461                            editable_range: input_excerpt.editable_range.to_offset(&snapshot),
 462                            input_outline,
 463                        })
 464                    }
 465                })
 466                .await?;
 467
 468            log::debug!(
 469                "Events:\n{}\nExcerpt:\n{:?}",
 470                values.input_events,
 471                values.input_excerpt
 472            );
 473
 474            let body = PredictEditsBody {
 475                input_events: values.input_events.clone(),
 476                input_excerpt: values.input_excerpt.clone(),
 477                speculated_output: Some(values.speculated_output),
 478                outline: Some(values.input_outline.clone()),
 479                can_collect_data,
 480                diagnostic_groups: diagnostic_groups.and_then(|diagnostic_groups| {
 481                    diagnostic_groups
 482                        .into_iter()
 483                        .map(|(name, diagnostic_group)| {
 484                            Ok((name.to_string(), serde_json::to_value(diagnostic_group)?))
 485                        })
 486                        .collect::<Result<Vec<_>>>()
 487                        .log_err()
 488                }),
 489            };
 490
 491            let response = perform_predict_edits(PerformPredictEditsParams {
 492                client,
 493                llm_token,
 494                app_version,
 495                body,
 496            })
 497            .await;
 498            let (response, usage) = match response {
 499                Ok(response) => response,
 500                Err(err) => {
 501                    if err.is::<ZedUpdateRequiredError>() {
 502                        cx.update(|cx| {
 503                            zeta.update(cx, |zeta, _cx| {
 504                                zeta.update_required = true;
 505                            });
 506
 507                            if let Some(workspace) = workspace {
 508                                workspace.update(cx, |workspace, cx| {
 509                                    workspace.show_notification(
 510                                        NotificationId::unique::<ZedUpdateRequiredError>(),
 511                                        cx,
 512                                        |cx| {
 513                                            cx.new(|cx| {
 514                                                ErrorMessagePrompt::new(err.to_string(), cx)
 515                                                    .with_link_button(
 516                                                        "Update Zed",
 517                                                        "https://zed.dev/releases",
 518                                                    )
 519                                            })
 520                                        },
 521                                    );
 522                                });
 523                            }
 524                        })
 525                        .ok();
 526                    }
 527
 528                    return Err(err);
 529                }
 530            };
 531
 532            log::debug!("completion response: {}", &response.output_excerpt);
 533
 534            if let Some(usage) = usage {
 535                let limit = match usage.limit {
 536                    UsageLimit::Limited(limit) => limit.to_string(),
 537                    UsageLimit::Unlimited => "unlimited".to_string(),
 538                };
 539                log::info!("edit prediction usage: {} / {}", usage.amount, limit);
 540            }
 541
 542            Self::process_completion_response(
 543                response,
 544                buffer,
 545                &snapshot,
 546                values.editable_range,
 547                cursor_offset,
 548                path,
 549                values.input_outline,
 550                values.input_events,
 551                values.input_excerpt,
 552                request_sent_at,
 553                &cx,
 554            )
 555            .await
 556        })
 557    }
 558
 559    // Generates several example completions of various states to fill the Zeta completion modal
 560    #[cfg(any(test, feature = "test-support"))]
 561    pub fn fill_with_fake_completions(&mut self, cx: &mut Context<Self>) -> Task<()> {
 562        use language::Point;
 563
 564        let test_buffer_text = indoc::indoc! {r#"a longggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg line
 565            And maybe a short line
 566
 567            Then a few lines
 568
 569            and then another
 570            "#};
 571
 572        let project = None;
 573        let buffer = cx.new(|cx| Buffer::local(test_buffer_text, cx));
 574        let position = buffer.read(cx).anchor_before(Point::new(1, 0));
 575
 576        let completion_tasks = vec![
 577            self.fake_completion(
 578                project,
 579                &buffer,
 580                position,
 581                PredictEditsResponse {
 582                    request_id: Uuid::parse_str("e7861db5-0cea-4761-b1c5-ad083ac53a80").unwrap(),
 583                    output_excerpt: format!("{EDITABLE_REGION_START_MARKER}
 584a longggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg line
 585[here's an edit]
 586And maybe a short line
 587Then a few lines
 588and then another
 589{EDITABLE_REGION_END_MARKER}
 590                        ", ),
 591                },
 592                cx,
 593            ),
 594            self.fake_completion(
 595                project,
 596                &buffer,
 597                position,
 598                PredictEditsResponse {
 599                    request_id: Uuid::parse_str("077c556a-2c49-44e2-bbc6-dafc09032a5e").unwrap(),
 600                    output_excerpt: format!(r#"{EDITABLE_REGION_START_MARKER}
 601a longggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg line
 602And maybe a short line
 603[and another edit]
 604Then a few lines
 605and then another
 606{EDITABLE_REGION_END_MARKER}
 607                        "#),
 608                },
 609                cx,
 610            ),
 611            self.fake_completion(
 612                project,
 613                &buffer,
 614                position,
 615                PredictEditsResponse {
 616                    request_id: Uuid::parse_str("df8c7b23-3d1d-4f99-a306-1f6264a41277").unwrap(),
 617                    output_excerpt: format!(r#"{EDITABLE_REGION_START_MARKER}
 618a longggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg line
 619And maybe a short line
 620
 621Then a few lines
 622
 623and then another
 624{EDITABLE_REGION_END_MARKER}
 625                        "#),
 626                },
 627                cx,
 628            ),
 629            self.fake_completion(
 630                project,
 631                &buffer,
 632                position,
 633                PredictEditsResponse {
 634                    request_id: Uuid::parse_str("c743958d-e4d8-44a8-aa5b-eb1e305c5f5c").unwrap(),
 635                    output_excerpt: format!(r#"{EDITABLE_REGION_START_MARKER}
 636a longggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg line
 637And maybe a short line
 638
 639Then a few lines
 640
 641and then another
 642{EDITABLE_REGION_END_MARKER}
 643                        "#),
 644                },
 645                cx,
 646            ),
 647            self.fake_completion(
 648                project,
 649                &buffer,
 650                position,
 651                PredictEditsResponse {
 652                    request_id: Uuid::parse_str("ff5cd7ab-ad06-4808-986e-d3391e7b8355").unwrap(),
 653                    output_excerpt: format!(r#"{EDITABLE_REGION_START_MARKER}
 654a longggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg line
 655And maybe a short line
 656Then a few lines
 657[a third completion]
 658and then another
 659{EDITABLE_REGION_END_MARKER}
 660                        "#),
 661                },
 662                cx,
 663            ),
 664            self.fake_completion(
 665                project,
 666                &buffer,
 667                position,
 668                PredictEditsResponse {
 669                    request_id: Uuid::parse_str("83cafa55-cdba-4b27-8474-1865ea06be94").unwrap(),
 670                    output_excerpt: format!(r#"{EDITABLE_REGION_START_MARKER}
 671a longggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg line
 672And maybe a short line
 673and then another
 674[fourth completion example]
 675{EDITABLE_REGION_END_MARKER}
 676                        "#),
 677                },
 678                cx,
 679            ),
 680            self.fake_completion(
 681                project,
 682                &buffer,
 683                position,
 684                PredictEditsResponse {
 685                    request_id: Uuid::parse_str("d5bd3afd-8723-47c7-bd77-15a3a926867b").unwrap(),
 686                    output_excerpt: format!(r#"{EDITABLE_REGION_START_MARKER}
 687a longggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggggg line
 688And maybe a short line
 689Then a few lines
 690and then another
 691[fifth and final completion]
 692{EDITABLE_REGION_END_MARKER}
 693                        "#),
 694                },
 695                cx,
 696            ),
 697        ];
 698
 699        cx.spawn(async move |zeta, cx| {
 700            for task in completion_tasks {
 701                task.await.unwrap();
 702            }
 703
 704            zeta.update(cx, |zeta, _cx| {
 705                zeta.shown_completions.get_mut(2).unwrap().edits = Arc::new([]);
 706                zeta.shown_completions.get_mut(3).unwrap().edits = Arc::new([]);
 707            })
 708            .ok();
 709        })
 710    }
 711
 712    #[cfg(any(test, feature = "test-support"))]
 713    pub fn fake_completion(
 714        &mut self,
 715        project: Option<&Entity<Project>>,
 716        buffer: &Entity<Buffer>,
 717        position: language::Anchor,
 718        response: PredictEditsResponse,
 719        cx: &mut Context<Self>,
 720    ) -> Task<Result<Option<InlineCompletion>>> {
 721        use std::future::ready;
 722
 723        self.request_completion_impl(None, project, buffer, position, false, cx, |_params| {
 724            ready(Ok((response, None)))
 725        })
 726    }
 727
 728    pub fn request_completion(
 729        &mut self,
 730        project: Option<&Entity<Project>>,
 731        buffer: &Entity<Buffer>,
 732        position: language::Anchor,
 733        can_collect_data: bool,
 734        cx: &mut Context<Self>,
 735    ) -> Task<Result<Option<InlineCompletion>>> {
 736        let workspace = self
 737            .workspace
 738            .as_ref()
 739            .and_then(|workspace| workspace.upgrade());
 740        self.request_completion_impl(
 741            workspace,
 742            project,
 743            buffer,
 744            position,
 745            can_collect_data,
 746            cx,
 747            Self::perform_predict_edits,
 748        )
 749    }
 750
 751    fn perform_predict_edits(
 752        params: PerformPredictEditsParams,
 753    ) -> impl Future<Output = Result<(PredictEditsResponse, Option<Usage>)>> {
 754        async move {
 755            let PerformPredictEditsParams {
 756                client,
 757                llm_token,
 758                app_version,
 759                body,
 760                ..
 761            } = params;
 762
 763            let http_client = client.http_client();
 764            let mut token = llm_token.acquire(&client).await?;
 765            let mut did_retry = false;
 766
 767            loop {
 768                let request_builder = http_client::Request::builder().method(Method::POST);
 769                let request_builder =
 770                    if let Ok(predict_edits_url) = std::env::var("ZED_PREDICT_EDITS_URL") {
 771                        request_builder.uri(predict_edits_url)
 772                    } else {
 773                        request_builder.uri(
 774                            http_client
 775                                .build_zed_llm_url("/predict_edits/v2", &[])?
 776                                .as_ref(),
 777                        )
 778                    };
 779                let request = request_builder
 780                    .header("Content-Type", "application/json")
 781                    .header("Authorization", format!("Bearer {}", token))
 782                    .body(serde_json::to_string(&body)?.into())?;
 783
 784                let mut response = http_client.send(request).await?;
 785
 786                if let Some(minimum_required_version) = response
 787                    .headers()
 788                    .get(MINIMUM_REQUIRED_VERSION_HEADER_NAME)
 789                    .and_then(|version| SemanticVersion::from_str(version.to_str().ok()?).ok())
 790                {
 791                    if app_version < minimum_required_version {
 792                        return Err(anyhow!(ZedUpdateRequiredError {
 793                            minimum_version: minimum_required_version
 794                        }));
 795                    }
 796                }
 797
 798                if response.status().is_success() {
 799                    let usage = Usage::from_headers(response.headers()).ok();
 800
 801                    let mut body = String::new();
 802                    response.body_mut().read_to_string(&mut body).await?;
 803                    return Ok((serde_json::from_str(&body)?, usage));
 804                } else if !did_retry
 805                    && response
 806                        .headers()
 807                        .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
 808                        .is_some()
 809                {
 810                    did_retry = true;
 811                    token = llm_token.refresh(&client).await?;
 812                } else {
 813                    let mut body = String::new();
 814                    response.body_mut().read_to_string(&mut body).await?;
 815                    return Err(anyhow!(
 816                        "error predicting edits.\nStatus: {:?}\nBody: {}",
 817                        response.status(),
 818                        body
 819                    ));
 820                }
 821            }
 822        }
 823    }
 824
 825    fn process_completion_response(
 826        prediction_response: PredictEditsResponse,
 827        buffer: Entity<Buffer>,
 828        snapshot: &BufferSnapshot,
 829        editable_range: Range<usize>,
 830        cursor_offset: usize,
 831        path: Arc<Path>,
 832        input_outline: String,
 833        input_events: String,
 834        input_excerpt: String,
 835        request_sent_at: Instant,
 836        cx: &AsyncApp,
 837    ) -> Task<Result<Option<InlineCompletion>>> {
 838        let snapshot = snapshot.clone();
 839        let request_id = prediction_response.request_id;
 840        let output_excerpt = prediction_response.output_excerpt;
 841        cx.spawn(async move |cx| {
 842            let output_excerpt: Arc<str> = output_excerpt.into();
 843
 844            let edits: Arc<[(Range<Anchor>, String)]> = cx
 845                .background_spawn({
 846                    let output_excerpt = output_excerpt.clone();
 847                    let editable_range = editable_range.clone();
 848                    let snapshot = snapshot.clone();
 849                    async move { Self::parse_edits(output_excerpt, editable_range, &snapshot) }
 850                })
 851                .await?
 852                .into();
 853
 854            let Some((edits, snapshot, edit_preview)) = buffer.read_with(cx, {
 855                let edits = edits.clone();
 856                |buffer, cx| {
 857                    let new_snapshot = buffer.snapshot();
 858                    let edits: Arc<[(Range<Anchor>, String)]> =
 859                        interpolate(&snapshot, &new_snapshot, edits)?.into();
 860                    Some((edits.clone(), new_snapshot, buffer.preview_edits(edits, cx)))
 861                }
 862            })?
 863            else {
 864                return anyhow::Ok(None);
 865            };
 866
 867            let edit_preview = edit_preview.await;
 868
 869            Ok(Some(InlineCompletion {
 870                id: InlineCompletionId(request_id),
 871                path,
 872                excerpt_range: editable_range,
 873                cursor_offset,
 874                edits,
 875                edit_preview,
 876                snapshot,
 877                input_outline: input_outline.into(),
 878                input_events: input_events.into(),
 879                input_excerpt: input_excerpt.into(),
 880                output_excerpt,
 881                request_sent_at,
 882                response_received_at: Instant::now(),
 883            }))
 884        })
 885    }
 886
 887    fn parse_edits(
 888        output_excerpt: Arc<str>,
 889        editable_range: Range<usize>,
 890        snapshot: &BufferSnapshot,
 891    ) -> Result<Vec<(Range<Anchor>, String)>> {
 892        let content = output_excerpt.replace(CURSOR_MARKER, "");
 893
 894        let start_markers = content
 895            .match_indices(EDITABLE_REGION_START_MARKER)
 896            .collect::<Vec<_>>();
 897        anyhow::ensure!(
 898            start_markers.len() == 1,
 899            "expected exactly one start marker, found {}",
 900            start_markers.len()
 901        );
 902
 903        let end_markers = content
 904            .match_indices(EDITABLE_REGION_END_MARKER)
 905            .collect::<Vec<_>>();
 906        anyhow::ensure!(
 907            end_markers.len() == 1,
 908            "expected exactly one end marker, found {}",
 909            end_markers.len()
 910        );
 911
 912        let sof_markers = content
 913            .match_indices(START_OF_FILE_MARKER)
 914            .collect::<Vec<_>>();
 915        anyhow::ensure!(
 916            sof_markers.len() <= 1,
 917            "expected at most one start-of-file marker, found {}",
 918            sof_markers.len()
 919        );
 920
 921        let codefence_start = start_markers[0].0;
 922        let content = &content[codefence_start..];
 923
 924        let newline_ix = content.find('\n').context("could not find newline")?;
 925        let content = &content[newline_ix + 1..];
 926
 927        let codefence_end = content
 928            .rfind(&format!("\n{EDITABLE_REGION_END_MARKER}"))
 929            .context("could not find end marker")?;
 930        let new_text = &content[..codefence_end];
 931
 932        let old_text = snapshot
 933            .text_for_range(editable_range.clone())
 934            .collect::<String>();
 935
 936        Ok(Self::compute_edits(
 937            old_text,
 938            new_text,
 939            editable_range.start,
 940            &snapshot,
 941        ))
 942    }
 943
 944    pub fn compute_edits(
 945        old_text: String,
 946        new_text: &str,
 947        offset: usize,
 948        snapshot: &BufferSnapshot,
 949    ) -> Vec<(Range<Anchor>, String)> {
 950        text_diff(&old_text, &new_text)
 951            .into_iter()
 952            .map(|(mut old_range, new_text)| {
 953                old_range.start += offset;
 954                old_range.end += offset;
 955
 956                let prefix_len = common_prefix(
 957                    snapshot.chars_for_range(old_range.clone()),
 958                    new_text.chars(),
 959                );
 960                old_range.start += prefix_len;
 961
 962                let suffix_len = common_prefix(
 963                    snapshot.reversed_chars_for_range(old_range.clone()),
 964                    new_text[prefix_len..].chars().rev(),
 965                );
 966                old_range.end = old_range.end.saturating_sub(suffix_len);
 967
 968                let new_text = new_text[prefix_len..new_text.len() - suffix_len].to_string();
 969                let range = if old_range.is_empty() {
 970                    let anchor = snapshot.anchor_after(old_range.start);
 971                    anchor..anchor
 972                } else {
 973                    snapshot.anchor_after(old_range.start)..snapshot.anchor_before(old_range.end)
 974                };
 975                (range, new_text)
 976            })
 977            .collect()
 978    }
 979
 980    pub fn is_completion_rated(&self, completion_id: InlineCompletionId) -> bool {
 981        self.rated_completions.contains(&completion_id)
 982    }
 983
 984    pub fn completion_shown(&mut self, completion: &InlineCompletion, cx: &mut Context<Self>) {
 985        self.shown_completions.push_front(completion.clone());
 986        if self.shown_completions.len() > 50 {
 987            let completion = self.shown_completions.pop_back().unwrap();
 988            self.rated_completions.remove(&completion.id);
 989        }
 990        cx.notify();
 991    }
 992
 993    pub fn rate_completion(
 994        &mut self,
 995        completion: &InlineCompletion,
 996        rating: InlineCompletionRating,
 997        feedback: String,
 998        cx: &mut Context<Self>,
 999    ) {
1000        self.rated_completions.insert(completion.id);
1001        telemetry::event!(
1002            "Edit Prediction Rated",
1003            rating,
1004            input_events = completion.input_events,
1005            input_excerpt = completion.input_excerpt,
1006            input_outline = completion.input_outline,
1007            output_excerpt = completion.output_excerpt,
1008            feedback
1009        );
1010        self.client.telemetry().flush_events();
1011        cx.notify();
1012    }
1013
1014    pub fn shown_completions(&self) -> impl DoubleEndedIterator<Item = &InlineCompletion> {
1015        self.shown_completions.iter()
1016    }
1017
1018    pub fn shown_completions_len(&self) -> usize {
1019        self.shown_completions.len()
1020    }
1021
1022    fn report_changes_for_buffer(
1023        &mut self,
1024        buffer: &Entity<Buffer>,
1025        cx: &mut Context<Self>,
1026    ) -> BufferSnapshot {
1027        self.register_buffer(buffer, cx);
1028
1029        let registered_buffer = self
1030            .registered_buffers
1031            .get_mut(&buffer.entity_id())
1032            .unwrap();
1033        let new_snapshot = buffer.read(cx).snapshot();
1034
1035        if new_snapshot.version != registered_buffer.snapshot.version {
1036            let old_snapshot = mem::replace(&mut registered_buffer.snapshot, new_snapshot.clone());
1037            self.push_event(Event::BufferChange {
1038                old_snapshot,
1039                new_snapshot: new_snapshot.clone(),
1040                timestamp: Instant::now(),
1041            });
1042        }
1043
1044        new_snapshot
1045    }
1046
1047    fn load_data_collection_choices() -> DataCollectionChoice {
1048        let choice = KEY_VALUE_STORE
1049            .read_kvp(ZED_PREDICT_DATA_COLLECTION_CHOICE)
1050            .log_err()
1051            .flatten();
1052
1053        match choice.as_deref() {
1054            Some("true") => DataCollectionChoice::Enabled,
1055            Some("false") => DataCollectionChoice::Disabled,
1056            Some(_) => {
1057                log::error!("unknown value in '{ZED_PREDICT_DATA_COLLECTION_CHOICE}'");
1058                DataCollectionChoice::NotAnswered
1059            }
1060            None => DataCollectionChoice::NotAnswered,
1061        }
1062    }
1063}
1064
1065struct PerformPredictEditsParams {
1066    pub client: Arc<Client>,
1067    pub llm_token: LlmApiToken,
1068    pub app_version: SemanticVersion,
1069    pub body: PredictEditsBody,
1070}
1071
1072#[derive(Error, Debug)]
1073#[error(
1074    "You must update to Zed version {minimum_version} or higher to continue using edit predictions."
1075)]
1076pub struct ZedUpdateRequiredError {
1077    minimum_version: SemanticVersion,
1078}
1079
1080struct LicenseDetectionWatcher {
1081    is_open_source_rx: watch::Receiver<bool>,
1082    _is_open_source_task: Task<()>,
1083}
1084
1085impl LicenseDetectionWatcher {
1086    pub fn new(worktree: &Worktree, cx: &mut Context<Worktree>) -> Self {
1087        let (mut is_open_source_tx, is_open_source_rx) = watch::channel_with::<bool>(false);
1088
1089        // Check if worktree is a single file, if so we do not need to check for a LICENSE file
1090        let task = if worktree.abs_path().is_file() {
1091            Task::ready(())
1092        } else {
1093            let loaded_files = LICENSE_FILES_TO_CHECK
1094                .iter()
1095                .map(Path::new)
1096                .map(|file| worktree.load_file(file, cx))
1097                .collect::<ArrayVec<_, { LICENSE_FILES_TO_CHECK.len() }>>();
1098
1099            cx.background_spawn(async move {
1100                for loaded_file in loaded_files.into_iter() {
1101                    let Ok(loaded_file) = loaded_file.await else {
1102                        continue;
1103                    };
1104
1105                    let path = &loaded_file.file.path;
1106                    if is_license_eligible_for_data_collection(&loaded_file.text) {
1107                        log::info!("detected '{path:?}' as open source license");
1108                        *is_open_source_tx.borrow_mut() = true;
1109                    } else {
1110                        log::info!("didn't detect '{path:?}' as open source license");
1111                    }
1112
1113                    // stop on the first license that successfully read
1114                    return;
1115                }
1116
1117                log::debug!("didn't find a license file to check, assuming closed source");
1118            })
1119        };
1120
1121        Self {
1122            is_open_source_rx,
1123            _is_open_source_task: task,
1124        }
1125    }
1126
1127    /// Answers false until we find out it's open source
1128    pub fn is_project_open_source(&self) -> bool {
1129        *self.is_open_source_rx.borrow()
1130    }
1131}
1132
1133fn common_prefix<T1: Iterator<Item = char>, T2: Iterator<Item = char>>(a: T1, b: T2) -> usize {
1134    a.zip(b)
1135        .take_while(|(a, b)| a == b)
1136        .map(|(a, _)| a.len_utf8())
1137        .sum()
1138}
1139
1140fn prompt_for_outline(snapshot: &BufferSnapshot) -> String {
1141    let mut input_outline = String::new();
1142
1143    writeln!(
1144        input_outline,
1145        "```{}",
1146        snapshot
1147            .file()
1148            .map_or(Cow::Borrowed("untitled"), |file| file
1149                .path()
1150                .to_string_lossy())
1151    )
1152    .unwrap();
1153
1154    if let Some(outline) = snapshot.outline(None) {
1155        for item in &outline.items {
1156            let spacing = " ".repeat(item.depth);
1157            writeln!(input_outline, "{}{}", spacing, item.text).unwrap();
1158        }
1159    }
1160
1161    writeln!(input_outline, "```").unwrap();
1162
1163    input_outline
1164}
1165
1166fn prompt_for_events(events: &VecDeque<Event>, mut remaining_tokens: usize) -> String {
1167    let mut result = String::new();
1168    for event in events.iter().rev() {
1169        let event_string = event.to_prompt();
1170        let event_tokens = tokens_for_bytes(event_string.len());
1171        if event_tokens > remaining_tokens {
1172            break;
1173        }
1174
1175        if !result.is_empty() {
1176            result.insert_str(0, "\n\n");
1177        }
1178        result.insert_str(0, &event_string);
1179        remaining_tokens -= event_tokens;
1180    }
1181    result
1182}
1183
1184struct RegisteredBuffer {
1185    snapshot: BufferSnapshot,
1186    _subscriptions: [gpui::Subscription; 2],
1187}
1188
1189#[derive(Clone)]
1190enum Event {
1191    BufferChange {
1192        old_snapshot: BufferSnapshot,
1193        new_snapshot: BufferSnapshot,
1194        timestamp: Instant,
1195    },
1196}
1197
1198impl Event {
1199    fn to_prompt(&self) -> String {
1200        match self {
1201            Event::BufferChange {
1202                old_snapshot,
1203                new_snapshot,
1204                ..
1205            } => {
1206                let mut prompt = String::new();
1207
1208                let old_path = old_snapshot
1209                    .file()
1210                    .map(|f| f.path().as_ref())
1211                    .unwrap_or(Path::new("untitled"));
1212                let new_path = new_snapshot
1213                    .file()
1214                    .map(|f| f.path().as_ref())
1215                    .unwrap_or(Path::new("untitled"));
1216                if old_path != new_path {
1217                    writeln!(prompt, "User renamed {:?} to {:?}\n", old_path, new_path).unwrap();
1218                }
1219
1220                let diff = language::unified_diff(&old_snapshot.text(), &new_snapshot.text());
1221                if !diff.is_empty() {
1222                    write!(
1223                        prompt,
1224                        "User edited {:?}:\n```diff\n{}\n```",
1225                        new_path, diff
1226                    )
1227                    .unwrap();
1228                }
1229
1230                prompt
1231            }
1232        }
1233    }
1234}
1235
1236#[derive(Debug, Clone)]
1237struct CurrentInlineCompletion {
1238    buffer_id: EntityId,
1239    completion: InlineCompletion,
1240}
1241
1242impl CurrentInlineCompletion {
1243    fn should_replace_completion(&self, old_completion: &Self, snapshot: &BufferSnapshot) -> bool {
1244        if self.buffer_id != old_completion.buffer_id {
1245            return true;
1246        }
1247
1248        let Some(old_edits) = old_completion.completion.interpolate(&snapshot) else {
1249            return true;
1250        };
1251        let Some(new_edits) = self.completion.interpolate(&snapshot) else {
1252            return false;
1253        };
1254
1255        if old_edits.len() == 1 && new_edits.len() == 1 {
1256            let (old_range, old_text) = &old_edits[0];
1257            let (new_range, new_text) = &new_edits[0];
1258            new_range == old_range && new_text.starts_with(old_text)
1259        } else {
1260            true
1261        }
1262    }
1263}
1264
1265struct PendingCompletion {
1266    id: usize,
1267    _task: Task<()>,
1268}
1269
1270#[derive(Debug, Clone, Copy)]
1271pub enum DataCollectionChoice {
1272    NotAnswered,
1273    Enabled,
1274    Disabled,
1275}
1276
1277impl DataCollectionChoice {
1278    pub fn is_enabled(self) -> bool {
1279        match self {
1280            Self::Enabled => true,
1281            Self::NotAnswered | Self::Disabled => false,
1282        }
1283    }
1284
1285    pub fn is_answered(self) -> bool {
1286        match self {
1287            Self::Enabled | Self::Disabled => true,
1288            Self::NotAnswered => false,
1289        }
1290    }
1291
1292    pub fn toggle(&self) -> DataCollectionChoice {
1293        match self {
1294            Self::Enabled => Self::Disabled,
1295            Self::Disabled => Self::Enabled,
1296            Self::NotAnswered => Self::Enabled,
1297        }
1298    }
1299}
1300
1301impl From<bool> for DataCollectionChoice {
1302    fn from(value: bool) -> Self {
1303        match value {
1304            true => DataCollectionChoice::Enabled,
1305            false => DataCollectionChoice::Disabled,
1306        }
1307    }
1308}
1309
1310pub struct ProviderDataCollection {
1311    /// When set to None, data collection is not possible in the provider buffer
1312    choice: Option<Entity<DataCollectionChoice>>,
1313    license_detection_watcher: Option<Rc<LicenseDetectionWatcher>>,
1314}
1315
1316impl ProviderDataCollection {
1317    pub fn new(zeta: Entity<Zeta>, buffer: Option<Entity<Buffer>>, cx: &mut App) -> Self {
1318        let choice_and_watcher = buffer.and_then(|buffer| {
1319            let file = buffer.read(cx).file()?;
1320
1321            if !file.is_local() || file.is_private() {
1322                return None;
1323            }
1324
1325            let zeta = zeta.read(cx);
1326            let choice = zeta.data_collection_choice.clone();
1327
1328            let license_detection_watcher = zeta
1329                .license_detection_watchers
1330                .get(&file.worktree_id(cx))
1331                .cloned()?;
1332
1333            Some((choice, license_detection_watcher))
1334        });
1335
1336        if let Some((choice, watcher)) = choice_and_watcher {
1337            ProviderDataCollection {
1338                choice: Some(choice),
1339                license_detection_watcher: Some(watcher),
1340            }
1341        } else {
1342            ProviderDataCollection {
1343                choice: None,
1344                license_detection_watcher: None,
1345            }
1346        }
1347    }
1348
1349    pub fn can_collect_data(&self, cx: &App) -> bool {
1350        self.is_data_collection_enabled(cx) && self.is_project_open_source()
1351    }
1352
1353    pub fn is_data_collection_enabled(&self, cx: &App) -> bool {
1354        self.choice
1355            .as_ref()
1356            .is_some_and(|choice| choice.read(cx).is_enabled())
1357    }
1358
1359    fn is_project_open_source(&self) -> bool {
1360        self.license_detection_watcher
1361            .as_ref()
1362            .is_some_and(|watcher| watcher.is_project_open_source())
1363    }
1364
1365    pub fn toggle(&mut self, cx: &mut App) {
1366        if let Some(choice) = self.choice.as_mut() {
1367            let new_choice = choice.update(cx, |choice, _cx| {
1368                let new_choice = choice.toggle();
1369                *choice = new_choice;
1370                new_choice
1371            });
1372
1373            db::write_and_log(cx, move || {
1374                KEY_VALUE_STORE.write_kvp(
1375                    ZED_PREDICT_DATA_COLLECTION_CHOICE.into(),
1376                    new_choice.is_enabled().to_string(),
1377                )
1378            });
1379        }
1380    }
1381}
1382
1383pub struct ZetaInlineCompletionProvider {
1384    zeta: Entity<Zeta>,
1385    pending_completions: ArrayVec<PendingCompletion, 2>,
1386    next_pending_completion_id: usize,
1387    current_completion: Option<CurrentInlineCompletion>,
1388    /// None if this is entirely disabled for this provider
1389    provider_data_collection: ProviderDataCollection,
1390    last_request_timestamp: Instant,
1391}
1392
1393impl ZetaInlineCompletionProvider {
1394    pub const THROTTLE_TIMEOUT: Duration = Duration::from_millis(300);
1395
1396    pub fn new(zeta: Entity<Zeta>, provider_data_collection: ProviderDataCollection) -> Self {
1397        Self {
1398            zeta,
1399            pending_completions: ArrayVec::new(),
1400            next_pending_completion_id: 0,
1401            current_completion: None,
1402            provider_data_collection,
1403            last_request_timestamp: Instant::now(),
1404        }
1405    }
1406}
1407
1408impl inline_completion::EditPredictionProvider for ZetaInlineCompletionProvider {
1409    fn name() -> &'static str {
1410        "zed-predict"
1411    }
1412
1413    fn display_name() -> &'static str {
1414        "Zed's Edit Predictions"
1415    }
1416
1417    fn show_completions_in_menu() -> bool {
1418        true
1419    }
1420
1421    fn show_tab_accept_marker() -> bool {
1422        true
1423    }
1424
1425    fn data_collection_state(&self, cx: &App) -> DataCollectionState {
1426        let is_project_open_source = self.provider_data_collection.is_project_open_source();
1427
1428        if self.provider_data_collection.is_data_collection_enabled(cx) {
1429            DataCollectionState::Enabled {
1430                is_project_open_source,
1431            }
1432        } else {
1433            DataCollectionState::Disabled {
1434                is_project_open_source,
1435            }
1436        }
1437    }
1438
1439    fn toggle_data_collection(&mut self, cx: &mut App) {
1440        self.provider_data_collection.toggle(cx);
1441    }
1442
1443    fn is_enabled(
1444        &self,
1445        _buffer: &Entity<Buffer>,
1446        _cursor_position: language::Anchor,
1447        _cx: &App,
1448    ) -> bool {
1449        true
1450    }
1451
1452    fn needs_terms_acceptance(&self, cx: &App) -> bool {
1453        !self.zeta.read(cx).tos_accepted
1454    }
1455
1456    fn is_refreshing(&self) -> bool {
1457        !self.pending_completions.is_empty()
1458    }
1459
1460    fn refresh(
1461        &mut self,
1462        project: Option<Entity<Project>>,
1463        buffer: Entity<Buffer>,
1464        position: language::Anchor,
1465        _debounce: bool,
1466        cx: &mut Context<Self>,
1467    ) {
1468        if !self.zeta.read(cx).tos_accepted {
1469            return;
1470        }
1471
1472        if self.zeta.read(cx).update_required {
1473            return;
1474        }
1475
1476        if let Some(current_completion) = self.current_completion.as_ref() {
1477            let snapshot = buffer.read(cx).snapshot();
1478            if current_completion
1479                .completion
1480                .interpolate(&snapshot)
1481                .is_some()
1482            {
1483                return;
1484            }
1485        }
1486
1487        let pending_completion_id = self.next_pending_completion_id;
1488        self.next_pending_completion_id += 1;
1489        let can_collect_data = self.provider_data_collection.can_collect_data(cx);
1490        let last_request_timestamp = self.last_request_timestamp;
1491
1492        let task = cx.spawn(async move |this, cx| {
1493            if let Some(timeout) = (last_request_timestamp + Self::THROTTLE_TIMEOUT)
1494                .checked_duration_since(Instant::now())
1495            {
1496                cx.background_executor().timer(timeout).await;
1497            }
1498
1499            let completion_request = this.update(cx, |this, cx| {
1500                this.last_request_timestamp = Instant::now();
1501                this.zeta.update(cx, |zeta, cx| {
1502                    zeta.request_completion(
1503                        project.as_ref(),
1504                        &buffer,
1505                        position,
1506                        can_collect_data,
1507                        cx,
1508                    )
1509                })
1510            });
1511
1512            let completion = match completion_request {
1513                Ok(completion_request) => {
1514                    let completion_request = completion_request.await;
1515                    completion_request.map(|c| {
1516                        c.map(|completion| CurrentInlineCompletion {
1517                            buffer_id: buffer.entity_id(),
1518                            completion,
1519                        })
1520                    })
1521                }
1522                Err(error) => Err(error),
1523            };
1524            let Some(new_completion) = completion
1525                .context("edit prediction failed")
1526                .log_err()
1527                .flatten()
1528            else {
1529                this.update(cx, |this, cx| {
1530                    if this.pending_completions[0].id == pending_completion_id {
1531                        this.pending_completions.remove(0);
1532                    } else {
1533                        this.pending_completions.clear();
1534                    }
1535
1536                    cx.notify();
1537                })
1538                .ok();
1539                return;
1540            };
1541
1542            this.update(cx, |this, cx| {
1543                if this.pending_completions[0].id == pending_completion_id {
1544                    this.pending_completions.remove(0);
1545                } else {
1546                    this.pending_completions.clear();
1547                }
1548
1549                if let Some(old_completion) = this.current_completion.as_ref() {
1550                    let snapshot = buffer.read(cx).snapshot();
1551                    if new_completion.should_replace_completion(&old_completion, &snapshot) {
1552                        this.zeta.update(cx, |zeta, cx| {
1553                            zeta.completion_shown(&new_completion.completion, cx);
1554                        });
1555                        this.current_completion = Some(new_completion);
1556                    }
1557                } else {
1558                    this.zeta.update(cx, |zeta, cx| {
1559                        zeta.completion_shown(&new_completion.completion, cx);
1560                    });
1561                    this.current_completion = Some(new_completion);
1562                }
1563
1564                cx.notify();
1565            })
1566            .ok();
1567        });
1568
1569        // We always maintain at most two pending completions. When we already
1570        // have two, we replace the newest one.
1571        if self.pending_completions.len() <= 1 {
1572            self.pending_completions.push(PendingCompletion {
1573                id: pending_completion_id,
1574                _task: task,
1575            });
1576        } else if self.pending_completions.len() == 2 {
1577            self.pending_completions.pop();
1578            self.pending_completions.push(PendingCompletion {
1579                id: pending_completion_id,
1580                _task: task,
1581            });
1582        }
1583    }
1584
1585    fn cycle(
1586        &mut self,
1587        _buffer: Entity<Buffer>,
1588        _cursor_position: language::Anchor,
1589        _direction: inline_completion::Direction,
1590        _cx: &mut Context<Self>,
1591    ) {
1592        // Right now we don't support cycling.
1593    }
1594
1595    fn accept(&mut self, _cx: &mut Context<Self>) {
1596        self.pending_completions.clear();
1597    }
1598
1599    fn discard(&mut self, _cx: &mut Context<Self>) {
1600        self.pending_completions.clear();
1601        self.current_completion.take();
1602    }
1603
1604    fn suggest(
1605        &mut self,
1606        buffer: &Entity<Buffer>,
1607        cursor_position: language::Anchor,
1608        cx: &mut Context<Self>,
1609    ) -> Option<inline_completion::InlineCompletion> {
1610        let CurrentInlineCompletion {
1611            buffer_id,
1612            completion,
1613            ..
1614        } = self.current_completion.as_mut()?;
1615
1616        // Invalidate previous completion if it was generated for a different buffer.
1617        if *buffer_id != buffer.entity_id() {
1618            self.current_completion.take();
1619            return None;
1620        }
1621
1622        let buffer = buffer.read(cx);
1623        let Some(edits) = completion.interpolate(&buffer.snapshot()) else {
1624            self.current_completion.take();
1625            return None;
1626        };
1627
1628        let cursor_row = cursor_position.to_point(buffer).row;
1629        let (closest_edit_ix, (closest_edit_range, _)) =
1630            edits.iter().enumerate().min_by_key(|(_, (range, _))| {
1631                let distance_from_start = cursor_row.abs_diff(range.start.to_point(buffer).row);
1632                let distance_from_end = cursor_row.abs_diff(range.end.to_point(buffer).row);
1633                cmp::min(distance_from_start, distance_from_end)
1634            })?;
1635
1636        let mut edit_start_ix = closest_edit_ix;
1637        for (range, _) in edits[..edit_start_ix].iter().rev() {
1638            let distance_from_closest_edit =
1639                closest_edit_range.start.to_point(buffer).row - range.end.to_point(buffer).row;
1640            if distance_from_closest_edit <= 1 {
1641                edit_start_ix -= 1;
1642            } else {
1643                break;
1644            }
1645        }
1646
1647        let mut edit_end_ix = closest_edit_ix + 1;
1648        for (range, _) in &edits[edit_end_ix..] {
1649            let distance_from_closest_edit =
1650                range.start.to_point(buffer).row - closest_edit_range.end.to_point(buffer).row;
1651            if distance_from_closest_edit <= 1 {
1652                edit_end_ix += 1;
1653            } else {
1654                break;
1655            }
1656        }
1657
1658        Some(inline_completion::InlineCompletion {
1659            id: Some(completion.id.to_string().into()),
1660            edits: edits[edit_start_ix..edit_end_ix].to_vec(),
1661            edit_preview: Some(completion.edit_preview.clone()),
1662        })
1663    }
1664}
1665
1666fn tokens_for_bytes(bytes: usize) -> usize {
1667    /// Typical number of string bytes per token for the purposes of limiting model input. This is
1668    /// intentionally low to err on the side of underestimating limits.
1669    const BYTES_PER_TOKEN_GUESS: usize = 3;
1670    bytes / BYTES_PER_TOKEN_GUESS
1671}
1672
1673#[cfg(test)]
1674mod tests {
1675    use client::test::FakeServer;
1676    use clock::FakeSystemClock;
1677    use gpui::TestAppContext;
1678    use http_client::FakeHttpClient;
1679    use indoc::indoc;
1680    use language::Point;
1681    use rpc::proto;
1682    use settings::SettingsStore;
1683
1684    use super::*;
1685
1686    #[gpui::test]
1687    async fn test_inline_completion_basic_interpolation(cx: &mut TestAppContext) {
1688        let buffer = cx.new(|cx| Buffer::local("Lorem ipsum dolor", cx));
1689        let edits: Arc<[(Range<Anchor>, String)]> = cx.update(|cx| {
1690            to_completion_edits(
1691                [(2..5, "REM".to_string()), (9..11, "".to_string())],
1692                &buffer,
1693                cx,
1694            )
1695            .into()
1696        });
1697
1698        let edit_preview = cx
1699            .read(|cx| buffer.read(cx).preview_edits(edits.clone(), cx))
1700            .await;
1701
1702        let completion = InlineCompletion {
1703            edits,
1704            edit_preview,
1705            path: Path::new("").into(),
1706            snapshot: cx.read(|cx| buffer.read(cx).snapshot()),
1707            id: InlineCompletionId(Uuid::new_v4()),
1708            excerpt_range: 0..0,
1709            cursor_offset: 0,
1710            input_outline: "".into(),
1711            input_events: "".into(),
1712            input_excerpt: "".into(),
1713            output_excerpt: "".into(),
1714            request_sent_at: Instant::now(),
1715            response_received_at: Instant::now(),
1716        };
1717
1718        cx.update(|cx| {
1719            assert_eq!(
1720                from_completion_edits(
1721                    &completion.interpolate(&buffer.read(cx).snapshot()).unwrap(),
1722                    &buffer,
1723                    cx
1724                ),
1725                vec![(2..5, "REM".to_string()), (9..11, "".to_string())]
1726            );
1727
1728            buffer.update(cx, |buffer, cx| buffer.edit([(2..5, "")], None, cx));
1729            assert_eq!(
1730                from_completion_edits(
1731                    &completion.interpolate(&buffer.read(cx).snapshot()).unwrap(),
1732                    &buffer,
1733                    cx
1734                ),
1735                vec![(2..2, "REM".to_string()), (6..8, "".to_string())]
1736            );
1737
1738            buffer.update(cx, |buffer, cx| buffer.undo(cx));
1739            assert_eq!(
1740                from_completion_edits(
1741                    &completion.interpolate(&buffer.read(cx).snapshot()).unwrap(),
1742                    &buffer,
1743                    cx
1744                ),
1745                vec![(2..5, "REM".to_string()), (9..11, "".to_string())]
1746            );
1747
1748            buffer.update(cx, |buffer, cx| buffer.edit([(2..5, "R")], None, cx));
1749            assert_eq!(
1750                from_completion_edits(
1751                    &completion.interpolate(&buffer.read(cx).snapshot()).unwrap(),
1752                    &buffer,
1753                    cx
1754                ),
1755                vec![(3..3, "EM".to_string()), (7..9, "".to_string())]
1756            );
1757
1758            buffer.update(cx, |buffer, cx| buffer.edit([(3..3, "E")], None, cx));
1759            assert_eq!(
1760                from_completion_edits(
1761                    &completion.interpolate(&buffer.read(cx).snapshot()).unwrap(),
1762                    &buffer,
1763                    cx
1764                ),
1765                vec![(4..4, "M".to_string()), (8..10, "".to_string())]
1766            );
1767
1768            buffer.update(cx, |buffer, cx| buffer.edit([(4..4, "M")], None, cx));
1769            assert_eq!(
1770                from_completion_edits(
1771                    &completion.interpolate(&buffer.read(cx).snapshot()).unwrap(),
1772                    &buffer,
1773                    cx
1774                ),
1775                vec![(9..11, "".to_string())]
1776            );
1777
1778            buffer.update(cx, |buffer, cx| buffer.edit([(4..5, "")], None, cx));
1779            assert_eq!(
1780                from_completion_edits(
1781                    &completion.interpolate(&buffer.read(cx).snapshot()).unwrap(),
1782                    &buffer,
1783                    cx
1784                ),
1785                vec![(4..4, "M".to_string()), (8..10, "".to_string())]
1786            );
1787
1788            buffer.update(cx, |buffer, cx| buffer.edit([(8..10, "")], None, cx));
1789            assert_eq!(
1790                from_completion_edits(
1791                    &completion.interpolate(&buffer.read(cx).snapshot()).unwrap(),
1792                    &buffer,
1793                    cx
1794                ),
1795                vec![(4..4, "M".to_string())]
1796            );
1797
1798            buffer.update(cx, |buffer, cx| buffer.edit([(4..6, "")], None, cx));
1799            assert_eq!(completion.interpolate(&buffer.read(cx).snapshot()), None);
1800        })
1801    }
1802
1803    #[gpui::test]
1804    async fn test_clean_up_diff(cx: &mut TestAppContext) {
1805        cx.update(|cx| {
1806            let settings_store = SettingsStore::test(cx);
1807            cx.set_global(settings_store);
1808            client::init_settings(cx);
1809        });
1810
1811        let edits = edits_for_prediction(
1812            indoc! {"
1813                fn main() {
1814                    let word_1 = \"lorem\";
1815                    let range = word.len()..word.len();
1816                }
1817            "},
1818            indoc! {"
1819                <|editable_region_start|>
1820                fn main() {
1821                    let word_1 = \"lorem\";
1822                    let range = word_1.len()..word_1.len();
1823                }
1824
1825                <|editable_region_end|>
1826            "},
1827            cx,
1828        )
1829        .await;
1830        assert_eq!(
1831            edits,
1832            [
1833                (Point::new(2, 20)..Point::new(2, 20), "_1".to_string()),
1834                (Point::new(2, 32)..Point::new(2, 32), "_1".to_string()),
1835            ]
1836        );
1837
1838        let edits = edits_for_prediction(
1839            indoc! {"
1840                fn main() {
1841                    let story = \"the quick\"
1842                }
1843            "},
1844            indoc! {"
1845                <|editable_region_start|>
1846                fn main() {
1847                    let story = \"the quick brown fox jumps over the lazy dog\";
1848                }
1849
1850                <|editable_region_end|>
1851            "},
1852            cx,
1853        )
1854        .await;
1855        assert_eq!(
1856            edits,
1857            [
1858                (
1859                    Point::new(1, 26)..Point::new(1, 26),
1860                    " brown fox jumps over the lazy dog".to_string()
1861                ),
1862                (Point::new(1, 27)..Point::new(1, 27), ";".to_string()),
1863            ]
1864        );
1865    }
1866
1867    #[gpui::test]
1868    async fn test_inline_completion_end_of_buffer(cx: &mut TestAppContext) {
1869        cx.update(|cx| {
1870            let settings_store = SettingsStore::test(cx);
1871            cx.set_global(settings_store);
1872            client::init_settings(cx);
1873        });
1874
1875        let buffer_content = "lorem\n";
1876        let completion_response = indoc! {"
1877            ```animals.js
1878            <|start_of_file|>
1879            <|editable_region_start|>
1880            lorem
1881            ipsum
1882            <|editable_region_end|>
1883            ```"};
1884
1885        let http_client = FakeHttpClient::create(move |_| async move {
1886            Ok(http_client::Response::builder()
1887                .status(200)
1888                .body(
1889                    serde_json::to_string(&PredictEditsResponse {
1890                        request_id: Uuid::parse_str("7e86480f-3536-4d2c-9334-8213e3445d45")
1891                            .unwrap(),
1892                        output_excerpt: completion_response.to_string(),
1893                    })
1894                    .unwrap()
1895                    .into(),
1896                )
1897                .unwrap())
1898        });
1899
1900        let client = cx.update(|cx| Client::new(Arc::new(FakeSystemClock::new()), http_client, cx));
1901        cx.update(|cx| {
1902            RefreshLlmTokenListener::register(client.clone(), cx);
1903        });
1904        let server = FakeServer::for_client(42, &client, cx).await;
1905        let user_store = cx.new(|cx| UserStore::new(client.clone(), cx));
1906        let zeta = cx.new(|cx| Zeta::new(None, client, user_store, cx));
1907
1908        let buffer = cx.new(|cx| Buffer::local(buffer_content, cx));
1909        let cursor = buffer.read_with(cx, |buffer, _| buffer.anchor_before(Point::new(1, 0)));
1910        let completion_task = zeta.update(cx, |zeta, cx| {
1911            zeta.request_completion(None, &buffer, cursor, false, cx)
1912        });
1913
1914        let token_request = server.receive::<proto::GetLlmToken>().await.unwrap();
1915        server.respond(
1916            token_request.receipt(),
1917            proto::GetLlmTokenResponse { token: "".into() },
1918        );
1919
1920        let completion = completion_task.await.unwrap().unwrap();
1921        buffer.update(cx, |buffer, cx| {
1922            buffer.edit(completion.edits.iter().cloned(), None, cx)
1923        });
1924        assert_eq!(
1925            buffer.read_with(cx, |buffer, _| buffer.text()),
1926            "lorem\nipsum"
1927        );
1928    }
1929
1930    async fn edits_for_prediction(
1931        buffer_content: &str,
1932        completion_response: &str,
1933        cx: &mut TestAppContext,
1934    ) -> Vec<(Range<Point>, String)> {
1935        let completion_response = completion_response.to_string();
1936        let http_client = FakeHttpClient::create(move |_| {
1937            let completion = completion_response.clone();
1938            async move {
1939                Ok(http_client::Response::builder()
1940                    .status(200)
1941                    .body(
1942                        serde_json::to_string(&PredictEditsResponse {
1943                            request_id: Uuid::new_v4(),
1944                            output_excerpt: completion,
1945                        })
1946                        .unwrap()
1947                        .into(),
1948                    )
1949                    .unwrap())
1950            }
1951        });
1952
1953        let client = cx.update(|cx| Client::new(Arc::new(FakeSystemClock::new()), http_client, cx));
1954        cx.update(|cx| {
1955            RefreshLlmTokenListener::register(client.clone(), cx);
1956        });
1957        let server = FakeServer::for_client(42, &client, cx).await;
1958        let user_store = cx.new(|cx| UserStore::new(client.clone(), cx));
1959        let zeta = cx.new(|cx| Zeta::new(None, client, user_store, cx));
1960
1961        let buffer = cx.new(|cx| Buffer::local(buffer_content, cx));
1962        let snapshot = buffer.read_with(cx, |buffer, _| buffer.snapshot());
1963        let cursor = buffer.read_with(cx, |buffer, _| buffer.anchor_before(Point::new(1, 0)));
1964        let completion_task = zeta.update(cx, |zeta, cx| {
1965            zeta.request_completion(None, &buffer, cursor, false, cx)
1966        });
1967
1968        let token_request = server.receive::<proto::GetLlmToken>().await.unwrap();
1969        server.respond(
1970            token_request.receipt(),
1971            proto::GetLlmTokenResponse { token: "".into() },
1972        );
1973
1974        let completion = completion_task.await.unwrap().unwrap();
1975        completion
1976            .edits
1977            .into_iter()
1978            .map(|(old_range, new_text)| (old_range.to_point(&snapshot), new_text.clone()))
1979            .collect::<Vec<_>>()
1980    }
1981
1982    fn to_completion_edits(
1983        iterator: impl IntoIterator<Item = (Range<usize>, String)>,
1984        buffer: &Entity<Buffer>,
1985        cx: &App,
1986    ) -> Vec<(Range<Anchor>, String)> {
1987        let buffer = buffer.read(cx);
1988        iterator
1989            .into_iter()
1990            .map(|(range, text)| {
1991                (
1992                    buffer.anchor_after(range.start)..buffer.anchor_before(range.end),
1993                    text,
1994                )
1995            })
1996            .collect()
1997    }
1998
1999    fn from_completion_edits(
2000        editor_edits: &[(Range<Anchor>, String)],
2001        buffer: &Entity<Buffer>,
2002        cx: &App,
2003    ) -> Vec<(Range<usize>, String)> {
2004        let buffer = buffer.read(cx);
2005        editor_edits
2006            .iter()
2007            .map(|(range, text)| {
2008                (
2009                    range.start.to_offset(buffer)..range.end.to_offset(buffer),
2010                    text.clone(),
2011                )
2012            })
2013            .collect()
2014    }
2015
2016    #[ctor::ctor]
2017    fn init_logger() {
2018        if std::env::var("RUST_LOG").is_ok() {
2019            env_logger::init();
2020        }
2021    }
2022}