1use crate::example::{ActualExcerpt, NamedExample};
2use crate::headless::ZetaCliAppState;
3use crate::paths::{CACHE_DIR, LATEST_EXAMPLE_RUN_DIR, RUN_DIR, print_run_data_dir};
4use crate::{
5 CacheMode, PredictArguments, PredictionOptions, PredictionProvider, PredictionsOutputFormat,
6};
7use ::serde::Serialize;
8use anyhow::{Context, Result, anyhow};
9use cloud_zeta2_prompt::{CURSOR_MARKER, write_codeblock};
10use futures::StreamExt as _;
11use gpui::{AppContext, AsyncApp, Entity};
12use project::Project;
13use project::buffer_store::BufferStoreEvent;
14use serde::Deserialize;
15use std::fs;
16use std::io::{IsTerminal, Write};
17use std::path::PathBuf;
18use std::sync::Arc;
19use std::sync::Mutex;
20use std::time::{Duration, Instant};
21use zeta::{EvalCache, EvalCacheEntryKind, EvalCacheKey, Zeta};
22
23pub async fn run_predict(
24 args: PredictArguments,
25 app_state: &Arc<ZetaCliAppState>,
26 cx: &mut AsyncApp,
27) {
28 let example = NamedExample::load(args.example_path).unwrap();
29 let project = example.setup_project(app_state, cx).await.unwrap();
30 let zeta = setup_zeta(args.options.provider, &project, app_state, cx).unwrap();
31 let _edited_buffers = example.apply_edit_history(&project, cx).await.unwrap();
32 let result = perform_predict(example, project, zeta, None, args.options, cx)
33 .await
34 .unwrap();
35 result.write(args.format, std::io::stdout()).unwrap();
36
37 print_run_data_dir(true, std::io::stdout().is_terminal());
38}
39
40pub fn setup_zeta(
41 provider: PredictionProvider,
42 project: &Entity<Project>,
43 app_state: &Arc<ZetaCliAppState>,
44 cx: &mut AsyncApp,
45) -> Result<Entity<Zeta>> {
46 let zeta =
47 cx.new(|cx| zeta::Zeta::new(app_state.client.clone(), app_state.user_store.clone(), cx))?;
48
49 zeta.update(cx, |zeta, _cx| {
50 let model = match provider {
51 PredictionProvider::Zeta1 => zeta::ZetaEditPredictionModel::Zeta1,
52 PredictionProvider::Zeta2 => zeta::ZetaEditPredictionModel::Zeta2,
53 PredictionProvider::Sweep => zeta::ZetaEditPredictionModel::Sweep,
54 };
55 zeta.set_edit_prediction_model(model);
56 })?;
57
58 let buffer_store = project.read_with(cx, |project, _| project.buffer_store().clone())?;
59
60 cx.subscribe(&buffer_store, {
61 let project = project.clone();
62 let zeta = zeta.clone();
63 move |_, event, cx| match event {
64 BufferStoreEvent::BufferAdded(buffer) => {
65 zeta.update(cx, |zeta, cx| zeta.register_buffer(&buffer, &project, cx));
66 }
67 _ => {}
68 }
69 })?
70 .detach();
71
72 anyhow::Ok(zeta)
73}
74
75pub async fn perform_predict(
76 example: NamedExample,
77 project: Entity<Project>,
78 zeta: Entity<Zeta>,
79 repetition_ix: Option<u16>,
80 options: PredictionOptions,
81 cx: &mut AsyncApp,
82) -> Result<PredictionDetails> {
83 let mut cache_mode = options.cache;
84 if repetition_ix.is_some() {
85 if cache_mode != CacheMode::Auto && cache_mode != CacheMode::Skip {
86 panic!("Repetitions are not supported in Auto cache mode");
87 } else {
88 cache_mode = CacheMode::Skip;
89 }
90 } else if cache_mode == CacheMode::Auto {
91 cache_mode = CacheMode::Requests;
92 }
93
94 let mut example_run_dir = RUN_DIR.join(&example.file_name());
95 if let Some(repetition_ix) = repetition_ix {
96 example_run_dir = example_run_dir.join(format!("{:03}", repetition_ix));
97 }
98 fs::create_dir_all(&example_run_dir)?;
99 if LATEST_EXAMPLE_RUN_DIR.is_symlink() {
100 fs::remove_file(&*LATEST_EXAMPLE_RUN_DIR)?;
101 }
102
103 #[cfg(unix)]
104 std::os::unix::fs::symlink(&example_run_dir, &*LATEST_EXAMPLE_RUN_DIR)
105 .context("creating latest link")?;
106
107 #[cfg(windows)]
108 std::os::windows::fs::symlink_dir(&example_run_dir, &*LATEST_EXAMPLE_RUN_DIR)
109 .context("creating latest link")?;
110
111 zeta.update(cx, |zeta, _cx| {
112 zeta.with_eval_cache(Arc::new(RunCache {
113 example_run_dir: example_run_dir.clone(),
114 cache_mode,
115 }));
116 })?;
117
118 let (cursor_buffer, cursor_anchor) = example.cursor_position(&project, cx).await?;
119
120 let result = Arc::new(Mutex::new(PredictionDetails::new(example_run_dir.clone())));
121
122 let prompt_format = options.zeta2.prompt_format;
123
124 zeta.update(cx, |zeta, _cx| {
125 let mut options = zeta.options().clone();
126 options.prompt_format = prompt_format.into();
127 zeta.set_options(options);
128 })?;
129
130 let mut debug_task = gpui::Task::ready(Ok(()));
131
132 if options.provider == crate::PredictionProvider::Zeta2 {
133 let mut debug_rx = zeta.update(cx, |zeta, _| zeta.debug_info())?;
134
135 debug_task = cx.background_spawn({
136 let result = result.clone();
137 async move {
138 let mut start_time = None;
139 let mut search_queries_generated_at = None;
140 let mut search_queries_executed_at = None;
141 while let Some(event) = debug_rx.next().await {
142 match event {
143 zeta::ZetaDebugInfo::ContextRetrievalStarted(info) => {
144 start_time = Some(info.timestamp);
145 fs::write(
146 example_run_dir.join("search_prompt.md"),
147 &info.search_prompt,
148 )?;
149 }
150 zeta::ZetaDebugInfo::SearchQueriesGenerated(info) => {
151 search_queries_generated_at = Some(info.timestamp);
152 fs::write(
153 example_run_dir.join("search_queries.json"),
154 serde_json::to_string_pretty(&info.search_queries).unwrap(),
155 )?;
156 }
157 zeta::ZetaDebugInfo::SearchQueriesExecuted(info) => {
158 search_queries_executed_at = Some(info.timestamp);
159 }
160 zeta::ZetaDebugInfo::ContextRetrievalFinished(_info) => {}
161 zeta::ZetaDebugInfo::EditPredictionRequested(request) => {
162 let prediction_started_at = Instant::now();
163 start_time.get_or_insert(prediction_started_at);
164 let prompt = request.local_prompt.unwrap_or_default();
165 fs::write(example_run_dir.join("prediction_prompt.md"), &prompt)?;
166
167 {
168 let mut result = result.lock().unwrap();
169 result.prompt_len = prompt.chars().count();
170
171 for included_file in request.inputs.included_files {
172 let insertions =
173 vec![(request.inputs.cursor_point, CURSOR_MARKER)];
174 result.excerpts.extend(included_file.excerpts.iter().map(
175 |excerpt| ActualExcerpt {
176 path: included_file.path.components().skip(1).collect(),
177 text: String::from(excerpt.text.as_ref()),
178 },
179 ));
180 write_codeblock(
181 &included_file.path,
182 included_file.excerpts.iter(),
183 if included_file.path == request.inputs.cursor_path {
184 &insertions
185 } else {
186 &[]
187 },
188 included_file.max_row,
189 false,
190 &mut result.excerpts_text,
191 );
192 }
193 }
194
195 let response =
196 request.response_rx.await?.0.map_err(|err| anyhow!(err))?;
197 let response = zeta::text_from_response(response).unwrap_or_default();
198 let prediction_finished_at = Instant::now();
199 fs::write(example_run_dir.join("prediction_response.md"), &response)?;
200
201 let mut result = result.lock().unwrap();
202 result.generated_len = response.chars().count();
203
204 result.planning_search_time =
205 Some(search_queries_generated_at.unwrap() - start_time.unwrap());
206 result.running_search_time = Some(
207 search_queries_executed_at.unwrap()
208 - search_queries_generated_at.unwrap(),
209 );
210 result.prediction_time = prediction_finished_at - prediction_started_at;
211 result.total_time = prediction_finished_at - start_time.unwrap();
212
213 break;
214 }
215 }
216 }
217 anyhow::Ok(())
218 }
219 });
220
221 zeta.update(cx, |zeta, cx| {
222 zeta.refresh_context(project.clone(), cursor_buffer.clone(), cursor_anchor, cx)
223 })?
224 .await?;
225 }
226
227 let prediction = zeta
228 .update(cx, |zeta, cx| {
229 zeta.request_prediction(
230 &project,
231 &cursor_buffer,
232 cursor_anchor,
233 cloud_llm_client::PredictEditsRequestTrigger::Cli,
234 cx,
235 )
236 })?
237 .await?;
238
239 debug_task.await?;
240
241 let mut result = Arc::into_inner(result).unwrap().into_inner().unwrap();
242
243 result.diff = prediction
244 .and_then(|prediction| {
245 let prediction = prediction.prediction.ok()?;
246 prediction.edit_preview.as_unified_diff(&prediction.edits)
247 })
248 .unwrap_or_default();
249
250 anyhow::Ok(result)
251}
252
253struct RunCache {
254 cache_mode: CacheMode,
255 example_run_dir: PathBuf,
256}
257
258impl RunCache {
259 fn output_cache_path((kind, key): &EvalCacheKey) -> PathBuf {
260 CACHE_DIR.join(format!("{kind}_out_{key:x}.json",))
261 }
262
263 fn input_cache_path((kind, key): &EvalCacheKey) -> PathBuf {
264 CACHE_DIR.join(format!("{kind}_in_{key:x}.json",))
265 }
266
267 fn link_to_run(&self, key: &EvalCacheKey) {
268 let output_link_path = self.example_run_dir.join(format!("{}_out.json", key.0));
269 fs::hard_link(Self::output_cache_path(key), &output_link_path).unwrap();
270
271 let input_link_path = self.example_run_dir.join(format!("{}_in.json", key.0));
272 fs::hard_link(Self::input_cache_path(key), &input_link_path).unwrap();
273 }
274}
275
276impl EvalCache for RunCache {
277 fn read(&self, key: EvalCacheKey) -> Option<String> {
278 let path = RunCache::output_cache_path(&key);
279
280 if path.exists() {
281 let use_cache = match key.0 {
282 EvalCacheEntryKind::Search => self.cache_mode.use_cached_search_results(),
283 EvalCacheEntryKind::Context | EvalCacheEntryKind::Prediction => {
284 self.cache_mode.use_cached_llm_responses()
285 }
286 };
287 if use_cache {
288 log::info!("Using cache entry: {}", path.display());
289 self.link_to_run(&key);
290 Some(fs::read_to_string(path).unwrap())
291 } else {
292 log::trace!("Skipping cached entry: {}", path.display());
293 None
294 }
295 } else if matches!(self.cache_mode, CacheMode::Force) {
296 panic!(
297 "No cached entry found for {:?}. Run without `--cache force` at least once.",
298 key.0
299 );
300 } else {
301 None
302 }
303 }
304
305 fn write(&self, key: EvalCacheKey, input: &str, output: &str) {
306 fs::create_dir_all(&*CACHE_DIR).unwrap();
307
308 let input_path = RunCache::input_cache_path(&key);
309 fs::write(&input_path, input).unwrap();
310
311 let output_path = RunCache::output_cache_path(&key);
312 log::trace!("Writing cache entry: {}", output_path.display());
313 fs::write(&output_path, output).unwrap();
314
315 self.link_to_run(&key);
316 }
317}
318
319#[derive(Clone, Debug, Serialize, Deserialize)]
320pub struct PredictionDetails {
321 pub diff: String,
322 pub excerpts: Vec<ActualExcerpt>,
323 pub excerpts_text: String, // TODO: contains the worktree root path. Drop this field and compute it on the fly
324 pub planning_search_time: Option<Duration>,
325 pub running_search_time: Option<Duration>,
326 pub prediction_time: Duration,
327 pub total_time: Duration,
328 pub run_example_dir: PathBuf,
329 pub prompt_len: usize,
330 pub generated_len: usize,
331}
332
333impl PredictionDetails {
334 pub fn new(run_example_dir: PathBuf) -> Self {
335 Self {
336 diff: Default::default(),
337 excerpts: Default::default(),
338 excerpts_text: Default::default(),
339 planning_search_time: Default::default(),
340 running_search_time: Default::default(),
341 prediction_time: Default::default(),
342 total_time: Default::default(),
343 run_example_dir,
344 prompt_len: 0,
345 generated_len: 0,
346 }
347 }
348
349 pub fn write(&self, format: PredictionsOutputFormat, mut out: impl Write) -> Result<()> {
350 let formatted = match format {
351 PredictionsOutputFormat::Md => self.to_markdown(),
352 PredictionsOutputFormat::Json => serde_json::to_string_pretty(self)?,
353 PredictionsOutputFormat::Diff => self.diff.clone(),
354 };
355
356 Ok(out.write_all(formatted.as_bytes())?)
357 }
358
359 pub fn to_markdown(&self) -> String {
360 let inference_time = self.planning_search_time.unwrap_or_default() + self.prediction_time;
361
362 format!(
363 "## Excerpts\n\n\
364 {}\n\n\
365 ## Prediction\n\n\
366 {}\n\n\
367 ## Time\n\n\
368 Planning searches: {}ms\n\
369 Running searches: {}ms\n\
370 Making Prediction: {}ms\n\n\
371 -------------------\n\n\
372 Total: {}ms\n\
373 Inference: {}ms ({:.2}%)\n",
374 self.excerpts_text,
375 self.diff,
376 self.planning_search_time.unwrap_or_default().as_millis(),
377 self.running_search_time.unwrap_or_default().as_millis(),
378 self.prediction_time.as_millis(),
379 self.total_time.as_millis(),
380 inference_time.as_millis(),
381 (inference_time.as_millis() as f64 / self.total_time.as_millis() as f64) * 100.
382 )
383 }
384}