auto_command.rs

  1use anyhow::{anyhow, Result};
  2use assistant_slash_command::{
  3    ArgumentCompletion, SlashCommand, SlashCommandOutput, SlashCommandOutputSection,
  4    SlashCommandResult,
  5};
  6use feature_flags::FeatureFlag;
  7use futures::StreamExt;
  8use gpui::{AppContext, AsyncAppContext, AsyncWindowContext, Task, WeakView, WindowContext};
  9use language::{CodeLabel, LspAdapterDelegate};
 10use language_model::{
 11    LanguageModelCompletionEvent, LanguageModelRegistry, LanguageModelRequest,
 12    LanguageModelRequestMessage, Role,
 13};
 14use semantic_index::{FileSummary, SemanticDb};
 15use smol::channel;
 16use std::sync::{atomic::AtomicBool, Arc};
 17use ui::{prelude::*, BorrowAppContext};
 18use util::ResultExt;
 19use workspace::Workspace;
 20
 21use crate::slash_command::create_label_for_command;
 22
 23pub struct AutoSlashCommandFeatureFlag;
 24
 25impl FeatureFlag for AutoSlashCommandFeatureFlag {
 26    const NAME: &'static str = "auto-slash-command";
 27}
 28
 29pub(crate) struct AutoCommand;
 30
 31impl SlashCommand for AutoCommand {
 32    fn name(&self) -> String {
 33        "auto".into()
 34    }
 35
 36    fn description(&self) -> String {
 37        "Automatically infer what context to add".into()
 38    }
 39
 40    fn icon(&self) -> IconName {
 41        IconName::Wand
 42    }
 43
 44    fn menu_text(&self) -> String {
 45        self.description()
 46    }
 47
 48    fn label(&self, cx: &AppContext) -> CodeLabel {
 49        create_label_for_command("auto", &["--prompt"], cx)
 50    }
 51
 52    fn complete_argument(
 53        self: Arc<Self>,
 54        _arguments: &[String],
 55        _cancel: Arc<AtomicBool>,
 56        workspace: Option<WeakView<Workspace>>,
 57        cx: &mut WindowContext,
 58    ) -> Task<Result<Vec<ArgumentCompletion>>> {
 59        // There's no autocomplete for a prompt, since it's arbitrary text.
 60        // However, we can use this opportunity to kick off a drain of the backlog.
 61        // That way, it can hopefully be done resummarizing by the time we've actually
 62        // typed out our prompt. This re-runs on every keystroke during autocomplete,
 63        // but in the future, we could instead do it only once, when /auto is first entered.
 64        let Some(workspace) = workspace.and_then(|ws| ws.upgrade()) else {
 65            log::warn!("workspace was dropped or unavailable during /auto autocomplete");
 66
 67            return Task::ready(Ok(Vec::new()));
 68        };
 69
 70        let project = workspace.read(cx).project().clone();
 71        let Some(project_index) =
 72            cx.update_global(|index: &mut SemanticDb, cx| index.project_index(project, cx))
 73        else {
 74            return Task::ready(Err(anyhow!("No project indexer, cannot use /auto")));
 75        };
 76
 77        let cx: &mut AppContext = cx;
 78
 79        cx.spawn(|cx: gpui::AsyncAppContext| async move {
 80            let task = project_index.read_with(&cx, |project_index, cx| {
 81                project_index.flush_summary_backlogs(cx)
 82            })?;
 83
 84            cx.background_executor().spawn(task).await;
 85
 86            anyhow::Ok(Vec::new())
 87        })
 88    }
 89
 90    fn requires_argument(&self) -> bool {
 91        true
 92    }
 93
 94    fn run(
 95        self: Arc<Self>,
 96        arguments: &[String],
 97        _context_slash_command_output_sections: &[SlashCommandOutputSection<language::Anchor>],
 98        _context_buffer: language::BufferSnapshot,
 99        workspace: WeakView<Workspace>,
100        _delegate: Option<Arc<dyn LspAdapterDelegate>>,
101        cx: &mut WindowContext,
102    ) -> Task<SlashCommandResult> {
103        let Some(workspace) = workspace.upgrade() else {
104            return Task::ready(Err(anyhow::anyhow!("workspace was dropped")));
105        };
106        if arguments.is_empty() {
107            return Task::ready(Err(anyhow!("missing prompt")));
108        };
109        let argument = arguments.join(" ");
110        let original_prompt = argument.to_string();
111        let project = workspace.read(cx).project().clone();
112        let Some(project_index) =
113            cx.update_global(|index: &mut SemanticDb, cx| index.project_index(project, cx))
114        else {
115            return Task::ready(Err(anyhow!("no project indexer")));
116        };
117
118        let task = cx.spawn(|cx: AsyncWindowContext| async move {
119            let summaries = project_index
120                .read_with(&cx, |project_index, cx| project_index.all_summaries(cx))?
121                .await?;
122
123            commands_for_summaries(&summaries, &original_prompt, &cx).await
124        });
125
126        // As a convenience, append /auto's argument to the end of the prompt
127        // so you don't have to write it again.
128        let original_prompt = argument.to_string();
129
130        cx.background_executor().spawn(async move {
131            let commands = task.await?;
132            let mut prompt = String::new();
133
134            log::info!(
135                "Translating this response into slash-commands: {:?}",
136                commands
137            );
138
139            for command in commands {
140                prompt.push('/');
141                prompt.push_str(&command.name);
142                prompt.push(' ');
143                prompt.push_str(&command.arg);
144                prompt.push('\n');
145            }
146
147            prompt.push('\n');
148            prompt.push_str(&original_prompt);
149
150            Ok(SlashCommandOutput {
151                text: prompt,
152                sections: Vec::new(),
153                run_commands_in_text: true,
154            }
155            .to_event_stream())
156        })
157    }
158}
159
160const PROMPT_INSTRUCTIONS_BEFORE_SUMMARY: &str = include_str!("prompt_before_summary.txt");
161const PROMPT_INSTRUCTIONS_AFTER_SUMMARY: &str = include_str!("prompt_after_summary.txt");
162
163fn summaries_prompt(summaries: &[FileSummary], original_prompt: &str) -> String {
164    let json_summaries = serde_json::to_string(summaries).unwrap();
165
166    format!("{PROMPT_INSTRUCTIONS_BEFORE_SUMMARY}\n{json_summaries}\n{PROMPT_INSTRUCTIONS_AFTER_SUMMARY}\n{original_prompt}")
167}
168
169/// The slash commands that the model is told about, and which we look for in the inference response.
170const SUPPORTED_SLASH_COMMANDS: &[&str] = &["search", "file"];
171
172#[derive(Debug, Clone)]
173struct CommandToRun {
174    name: String,
175    arg: String,
176}
177
178/// Given the pre-indexed file summaries for this project, as well as the original prompt
179/// string passed to `/auto`, get a list of slash commands to run, along with their arguments.
180///
181/// The prompt's output does not include the slashes (to reduce the chance that it makes a mistake),
182/// so taking one of these returned Strings and turning it into a real slash-command-with-argument
183/// involves prepending a slash to it.
184///
185/// This function will validate that each of the returned lines begins with one of SUPPORTED_SLASH_COMMANDS.
186/// Any other lines it encounters will be discarded, with a warning logged.
187async fn commands_for_summaries(
188    summaries: &[FileSummary],
189    original_prompt: &str,
190    cx: &AsyncAppContext,
191) -> Result<Vec<CommandToRun>> {
192    if summaries.is_empty() {
193        log::warn!("Inferring no context because there were no summaries available.");
194        return Ok(Vec::new());
195    }
196
197    // Use the globally configured model to translate the summaries into slash-commands,
198    // because Qwen2-7B-Instruct has not done a good job at that task.
199    let Some(model) = cx.update(|cx| LanguageModelRegistry::read_global(cx).active_model())? else {
200        log::warn!("Can't infer context because there's no active model.");
201        return Ok(Vec::new());
202    };
203    // Only go up to 90% of the actual max token count, to reduce chances of
204    // exceeding the token count due to inaccuracies in the token counting heuristic.
205    let max_token_count = (model.max_token_count() * 9) / 10;
206
207    // Rather than recursing (which would require this async function use a pinned box),
208    // we use an explicit stack of arguments and answers for when we need to "recurse."
209    let mut stack = vec![summaries];
210    let mut final_response = Vec::new();
211    let mut prompts = Vec::new();
212
213    // TODO We only need to create multiple Requests because we currently
214    // don't have the ability to tell if a CompletionProvider::complete response
215    // was a "too many tokens in this request" error. If we had that, then
216    // we could try the request once, instead of having to make separate requests
217    // to check the token count and then afterwards to run the actual prompt.
218    let make_request = |prompt: String| LanguageModelRequest {
219        messages: vec![LanguageModelRequestMessage {
220            role: Role::User,
221            content: vec![prompt.into()],
222            // Nothing in here will benefit from caching
223            cache: false,
224        }],
225        tools: Vec::new(),
226        stop: Vec::new(),
227        temperature: None,
228    };
229
230    while let Some(current_summaries) = stack.pop() {
231        // The split can result in one slice being empty and the other having one element.
232        // Whenever that happens, skip the empty one.
233        if current_summaries.is_empty() {
234            continue;
235        }
236
237        log::info!(
238            "Inferring prompt context using {} file summaries",
239            current_summaries.len()
240        );
241
242        let prompt = summaries_prompt(&current_summaries, original_prompt);
243        let start = std::time::Instant::now();
244        // Per OpenAI, 1 token ~= 4 chars in English (we go with 4.5 to overestimate a bit, because failed API requests cost a lot of perf)
245        // Verifying this against an actual model.count_tokens() confirms that it's usually within ~5% of the correct answer, whereas
246        // getting the correct answer from tiktoken takes hundreds of milliseconds (compared to this arithmetic being ~free).
247        // source: https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them
248        let token_estimate = prompt.len() * 2 / 9;
249        let duration = start.elapsed();
250        log::info!(
251            "Time taken to count tokens for prompt of length {:?}B: {:?}",
252            prompt.len(),
253            duration
254        );
255
256        if token_estimate < max_token_count {
257            prompts.push(prompt);
258        } else if current_summaries.len() == 1 {
259            log::warn!("Inferring context for a single file's summary failed because the prompt's token length exceeded the model's token limit.");
260        } else {
261            log::info!(
262                "Context inference using file summaries resulted in a prompt containing {token_estimate} tokens, which exceeded the model's max of {max_token_count}. Retrying as two separate prompts, each including half the number of summaries.",
263            );
264            let (left, right) = current_summaries.split_at(current_summaries.len() / 2);
265            stack.push(right);
266            stack.push(left);
267        }
268    }
269
270    let all_start = std::time::Instant::now();
271
272    let (tx, rx) = channel::bounded(1024);
273
274    let completion_streams = prompts
275        .into_iter()
276        .map(|prompt| {
277            let request = make_request(prompt.clone());
278            let model = model.clone();
279            let tx = tx.clone();
280            let stream = model.stream_completion(request, &cx);
281
282            (stream, tx)
283        })
284        .collect::<Vec<_>>();
285
286    cx.background_executor()
287        .spawn(async move {
288            let futures = completion_streams
289                .into_iter()
290                .enumerate()
291                .map(|(ix, (stream, tx))| async move {
292                    let start = std::time::Instant::now();
293                    let events = stream.await?;
294                    log::info!("Time taken for awaiting /await chunk stream #{ix}: {:?}", start.elapsed());
295
296                    let completion: String = events
297                        .filter_map(|event| async {
298                            if let Ok(LanguageModelCompletionEvent::Text(text)) = event {
299                                Some(text)
300                            } else {
301                                None
302                            }
303                        })
304                        .collect()
305                        .await;
306
307                    log::info!("Time taken for all /auto chunks to come back for #{ix}: {:?}", start.elapsed());
308
309                    for line in completion.split('\n') {
310                        if let Some(first_space) = line.find(' ') {
311                            let command = &line[..first_space].trim();
312                            let arg = &line[first_space..].trim();
313
314                            tx.send(CommandToRun {
315                                name: command.to_string(),
316                                arg: arg.to_string(),
317                            })
318                            .await?;
319                        } else if !line.trim().is_empty() {
320                            // All slash-commands currently supported in context inference need a space for the argument.
321                            log::warn!(
322                                "Context inference returned a non-blank line that contained no spaces (meaning no argument for the slash command): {:?}",
323                                line
324                            );
325                        }
326                    }
327
328                    anyhow::Ok(())
329                })
330                .collect::<Vec<_>>();
331
332            let _ = futures::future::try_join_all(futures).await.log_err();
333
334            let duration = all_start.elapsed();
335            eprintln!("All futures completed in {:?}", duration);
336        })
337        .await;
338
339    drop(tx); // Close the channel so that rx.collect() won't hang. This is safe because all futures have completed.
340    let results = rx.collect::<Vec<_>>().await;
341    eprintln!(
342        "Finished collecting from the channel with {} results",
343        results.len()
344    );
345    for command in results {
346        // Don't return empty or duplicate commands
347        if !command.name.is_empty()
348            && !final_response
349                .iter()
350                .any(|cmd: &CommandToRun| cmd.name == command.name && cmd.arg == command.arg)
351        {
352            if SUPPORTED_SLASH_COMMANDS
353                .iter()
354                .any(|supported| &command.name == supported)
355            {
356                final_response.push(command);
357            } else {
358                log::warn!(
359                    "Context inference returned an unrecognized slash command: {:?}",
360                    command
361                );
362            }
363        }
364    }
365
366    // Sort the commands by name (reversed just so that /search appears before /file)
367    final_response.sort_by(|cmd1, cmd2| cmd1.name.cmp(&cmd2.name).reverse());
368
369    Ok(final_response)
370}