g1t/services/work/src/capture.rs
Pick any line to see why it is the way it is: the commit, the pull request and issue it came from, and what the agent was thinking.
| Agents get guardrails, run credentials, an audit log, a context hub, repository instructions and mentions; security upkeep; snake_case API | 1 | //! Memory that fills itself (see `g1t_contracts::capture`). |
| 2 | //! | |
| 3 | //! Candidates arrive from four places: | |
| 4 | //! | |
| 5 | //! - **Runs.** At the end of a run the harness asks the agent what it | |
| 6 | //! learned and reports it with the run's token (`report_learned`). | |
| 7 | //! - **Reviews.** A person's request for changes, or a comment that | |
| 8 | //! corrects the agent ("use the shared client instead"), on a pull | |
| 9 | //! request, becomes a convention candidate quoting the comment. | |
| 10 | //! - **Merges.** A merged pull request's title, why and files become a | |
| 11 | //! decision candidate. No model is asked: the pull request's own words | |
| 12 | //! are the decision, and a candidate waits for review anyway. | |
| 13 | //! - **Docs.** The context service reads a project's README, AGENTS.md, | |
| 14 | //! docs and manifests and sends what they say (`capture_memories`). | |
| 15 | //! | |
| 16 | //! The same thing said again, by an independent source, is one memory seen | |
| 17 | //! twice, which keeps it (`promotes`). A dismissed memory stays dismissed, | |
| 18 | //! so the same wording is never suggested again. Nothing that looks like a | |
| 19 | //! secret is stored, in the text or in its evidence. | |
| 20 | ||
| 21 | use std::collections::HashMap; | |
| 22 | ||
| 23 | use g1t_contracts::agents::*; | |
| 24 | use g1t_contracts::capture::*; | |
| 25 | use g1t_contracts::events::{Event, NewEvent, Publish}; | |
| 26 | use g1t_contracts::repos::{PathByIdArgs, RepoPath}; | |
| 27 | use g1t_contracts::time::rfc3339; | |
| 28 | use g1t_contracts::work::{Comment, CommentKind, Pull, PullStatus, Verdict}; | |
| 29 | use g1t_contracts::{FailureCode, Outcome, PrincipalKind, new_id}; | |
| 30 | use g1t_kit::now_ms; | |
| 31 | use serde::Deserialize; | |
| 32 | use worker::Result; | |
| 33 | use worker::wasm_bindgen::JsValue; | |
| 34 | ||
| 35 | use crate::Work; | |
| Thirteen MCP tools and classic token scopes; agents rate their confidence and can be put on an issue in one step | 36 | use crate::rows::PULL_COLUMNS; |
| Agents get guardrails, run credentials, an audit log, a context hub, repository instructions and mentions; security upkeep; snake_case API | 37 | use crate::checks::hash; |
| 38 | use crate::rows::{CommentRow, PullRow}; | |
| 39 | ||
| 40 | /// The most memories one project, or one workspace, keeps (as memory.rs). | |
| 41 | const MAX_PER_SCOPE: u32 = 500; | |
| 42 | /// How sure g1t is of what an agent says it learned. | |
| 43 | const RUN_CONFIDENCE: f64 = 0.6; | |
| 44 | /// Of a person's correction in a review. | |
| 45 | const REVIEW_CONFIDENCE: f64 = 0.7; | |
| 46 | /// Of a merged pull request's decision, summed up from its own words. | |
| 47 | const MERGE_CONFIDENCE: f64 = 0.5; | |
| 48 | /// The longest correction or decision kept, in characters. | |
| 49 | const MAX_SUMMARY_CHARS: usize = 400; | |
| 50 | /// The most files a decision names. | |
| 51 | const MAX_FILES_NAMED: usize = 6; | |
| 52 | /// Who people's and agents' names are not. | |
| 53 | const NOT_PEOPLE: [&str; 3] = ["g1t-agent", "g1t", "agent"]; | |
| 54 | ||
| 55 | /// Words that mark a sentence as telling the agent how things are done. | |
| 56 | const CORRECTIVE: [&str; 20] = [ | |
| 57 | "instead", | |
| 58 | "don't", | |
| 59 | "do not", | |
| 60 | "never", | |
| 61 | "always", | |
| 62 | "please use", | |
| 63 | "we use", | |
| 64 | "use the", | |
| 65 | "should", | |
| 66 | "shouldn't", | |
| 67 | "prefer", | |
| 68 | "avoid", | |
| 69 | "must", | |
| 70 | "rather than", | |
| 71 | "convention", | |
| 72 | "not how", | |
| 73 | "we don't", | |
| 74 | "make sure", | |
| 75 | "remember to", | |
| 76 | "the right way", | |
| 77 | ]; | |
| 78 | ||
| 79 | #[derive(Deserialize)] | |
| 80 | struct Existing { | |
| 81 | id: String, | |
| 82 | status: String, | |
| 83 | sources: String, | |
| 84 | confidence: Option<f64>, | |
| 85 | } | |
| 86 | ||
| 87 | #[derive(Deserialize)] | |
| 88 | struct RunTicket { | |
| 89 | id: String, | |
| 90 | workspace: String, | |
| 91 | repo_id: String, | |
| 92 | number: Option<u32>, | |
| 93 | token_hash: String, | |
| 94 | } | |
| 95 | ||
| 96 | /// `text` cut to `max` characters at a word, with an ellipsis if cut. | |
| 97 | fn clip(text: &str, max: usize) -> String { | |
| 98 | let text = text.split_whitespace().collect::<Vec<_>>().join(" "); | |
| 99 | if text.chars().count() <= max { | |
| 100 | return text; | |
| 101 | } | |
| 102 | let cut: String = text.chars().take(max - 1).collect(); | |
| 103 | let cut = cut.rsplit_once(' ').map_or(cut.as_str(), |(head, _)| head).to_owned(); | |
| 104 | format!("{cut}…") | |
| 105 | } | |
| 106 | ||
| 107 | /// Markdown reduced to plain prose: no code blocks, quotes, headings or | |
| 108 | /// list markers. | |
| 109 | fn prose(markdown: &str) -> Vec<String> { | |
| 110 | let mut lines = Vec::new(); | |
| 111 | let mut fenced = false; | |
| 112 | for line in markdown.lines() { | |
| 113 | let trimmed = line.trim(); | |
| 114 | if trimmed.starts_with("```") || trimmed.starts_with("~~~") { | |
| 115 | fenced = !fenced; | |
| 116 | continue; | |
| 117 | } | |
| 118 | if fenced || trimmed.starts_with('>') || trimmed.starts_with('#') || trimmed.starts_with("<!--") { | |
| 119 | lines.push(String::new()); | |
| 120 | continue; | |
| 121 | } | |
| 122 | let item = trimmed.trim_start_matches(['-', '*', '+']).trim_start(); | |
| 123 | lines.push(item.to_owned()); | |
| 124 | } | |
| 125 | lines | |
| 126 | } | |
| 127 | ||
| 128 | /// The sentences of a comment, as plain prose. | |
| 129 | fn sentences(body: &str) -> Vec<String> { | |
| 130 | let mut out = Vec::new(); | |
| 131 | for line in prose(body) { | |
| 132 | let mut current = String::new(); | |
| 133 | for c in line.chars() { | |
| 134 | current.push(c); | |
| 135 | if matches!(c, '.' | '!' | '?') { | |
| 136 | let sentence = current.trim().to_owned(); | |
| 137 | if !sentence.is_empty() { | |
| 138 | out.push(sentence); | |
| 139 | } | |
| 140 | current.clear(); | |
| 141 | } | |
| 142 | } | |
| 143 | let rest = current.trim(); | |
| 144 | if !rest.is_empty() { | |
| 145 | out.push(rest.to_owned()); | |
| 146 | } | |
| 147 | } | |
| 148 | out | |
| 149 | } | |
| 150 | ||
| 151 | /// Whether a sentence tells the agent how something is done here. | |
| 152 | fn corrective(sentence: &str) -> bool { | |
| 153 | let lower = format!(" {} ", sentence.to_lowercase()); | |
| 154 | !sentence.trim_end().ends_with('?') && CORRECTIVE.iter().any(|word| lower.contains(word)) | |
| 155 | } | |
| 156 | ||
| 157 | /// What a person's review comment teaches, as one convention: its first | |
| 158 | /// sentence that says how things are done, or, for a request for changes, | |
| 159 | /// its first real sentence. None when it teaches nothing reusable: a | |
| 160 | /// question, a thank-you, a nit too short to mean anything. | |
| 161 | pub(crate) fn correction(body: &str, requested_changes: bool) -> Option<String> { | |
| 162 | let all = sentences(body); | |
| 163 | let found = all | |
| 164 | .iter() | |
| 165 | .find(|sentence| sentence.chars().count() >= 15 && corrective(sentence)) | |
| 166 | .or_else(|| { | |
| 167 | requested_changes.then(|| { | |
| 168 | all.iter() | |
| 169 | .find(|sentence| sentence.chars().count() >= 20 && !sentence.ends_with('?')) | |
| 170 | })? | |
| 171 | })?; | |
| 172 | Some(clip(found, MAX_SUMMARY_CHARS)) | |
| 173 | } | |
| 174 | ||
| 175 | /// A merged pull request's decision, from its own words: its title, the | |
| 176 | /// first paragraph of its description (why), and the files it changed. | |
| 177 | /// None when it says no more than its title. | |
| 178 | pub(crate) fn decision(number: u32, title: &str, body: Option<&str>, files: &[String]) -> Option<String> { | |
| 179 | let body = body.unwrap_or_default(); | |
| 180 | let mut paragraph = String::new(); | |
| 181 | for line in prose(body) { | |
| 182 | if line.trim().is_empty() { | |
| 183 | if !paragraph.is_empty() { | |
| 184 | break; | |
| 185 | } | |
| 186 | continue; | |
| 187 | } | |
| 188 | // "Summary:" and the like say nothing by themselves. | |
| 189 | if paragraph.is_empty() && line.trim_end_matches(':').split_whitespace().count() <= 1 { | |
| 190 | continue; | |
| 191 | } | |
| 192 | if !paragraph.is_empty() { | |
| 193 | paragraph.push(' '); | |
| 194 | } | |
| 195 | paragraph.push_str(line.trim()); | |
| 196 | } | |
| 197 | if paragraph.chars().count() < 30 { | |
| 198 | return None; | |
| 199 | } | |
| 200 | let why = clip(¶graph, MAX_SUMMARY_CHARS); | |
| 201 | let mut text = format!("Decided in #{number}, \"{}\": {why}", clip(title, 120)); | |
| 202 | if !files.is_empty() { | |
| 203 | let named: Vec<&str> = files.iter().take(MAX_FILES_NAMED).map(String::as_str).collect(); | |
| 204 | let more = files.len().saturating_sub(MAX_FILES_NAMED); | |
| 205 | text.push_str(&format!( | |
| 206 | " Changed {}{}.", | |
| 207 | named.join(", "), | |
| 208 | if more > 0 { format!(" and {more} more") } else { String::new() } | |
| 209 | )); | |
| 210 | } | |
| 211 | Some(clip(&text, MAX_MEMORY_CHARS)) | |
| 212 | } | |
| 213 | ||
| 214 | /// The JSON list of sources, with `reference` added once. | |
| 215 | fn with_source(sources: &str, reference: &str) -> Vec<String> { | |
| 216 | let mut list: Vec<String> = serde_json::from_str(sources).unwrap_or_default(); | |
| 217 | if !list.iter().any(|seen| seen == reference) { | |
| 218 | list.push(reference.to_owned()); | |
| 219 | } | |
| 220 | list | |
| 221 | } | |
| 222 | ||
| 223 | /// An item tidied, or why it cannot be kept. | |
| 224 | fn checked(item: &CaptureItem) -> std::result::Result<(String, Option<String>), &'static str> { | |
| 225 | let text = item.text.split_whitespace().collect::<Vec<_>>().join(" "); | |
| 226 | if text.chars().count() < 8 { | |
| 227 | return Err("too short"); | |
| 228 | } | |
| 229 | if text.chars().count() > MAX_MEMORY_CHARS { | |
| 230 | return Err("too long"); | |
| 231 | } | |
| 232 | if secret_in(&text).is_some() { | |
| 233 | return Err("looks like a secret"); | |
| 234 | } | |
| 235 | let evidence = item | |
| 236 | .evidence | |
| 237 | .as_deref() | |
| 238 | .map(|evidence| clip(evidence, MAX_EVIDENCE_CHARS)) | |
| 239 | .filter(|evidence| !evidence.is_empty()); | |
| 240 | // Evidence quoting a key would put the key in memory all the same. | |
| 241 | if evidence.as_deref().is_some_and(|evidence| secret_in(evidence).is_some()) { | |
| 242 | return Err("looks like a secret"); | |
| 243 | } | |
| 244 | Ok((text, evidence)) | |
| 245 | } | |
| 246 | ||
| 247 | impl Work { | |
| 248 | /// The `owner/name` of a repository, by id, remembered for one capture. | |
| 249 | async fn path_of(&self, cache: &mut HashMap<String, Option<RepoPath>>, repo_id: &str) -> Result<Option<RepoPath>> { | |
| 250 | if let Some(found) = cache.get(repo_id) { | |
| 251 | return Ok(found.clone()); | |
| 252 | } | |
| 253 | let found: Option<RepoPath> = | |
| 254 | g1t_kit::call(&self.repos, "path_by_id", &PathByIdArgs { id: repo_id.to_owned() }).await?; | |
| 255 | cache.insert(repo_id.to_owned(), found.clone()); | |
| 256 | Ok(found) | |
| 257 | } | |
| 258 | ||
| 259 | /// Tells subscribers (the context service's search) that a memory | |
| 260 | /// changed. Never fails what changed it. | |
| 261 | pub(crate) async fn memory_changed(&self, id: &str, workspace: &str, status: &str, repo_id: Option<&str>) { | |
| 262 | let event = NewEvent { | |
| 263 | kind: "memory.changed", | |
| 264 | source: "work", | |
| 265 | repo_id: repo_id.map(str::to_owned), | |
| 266 | actor: None, | |
| 267 | data: MemoryChanged { | |
| 268 | memory_id: id.to_owned(), | |
| 269 | workspace: workspace.to_owned(), | |
| 270 | status: status.to_owned(), | |
| 271 | }, | |
| 272 | }; | |
| 273 | let sent: Result<()> = g1t_kit::call(&self.events, "publish", &Publish { events: vec![event] }).await; | |
| 274 | if let Err(error) = sent { | |
| 275 | worker::console_error!("memory.changed for {id} was not published: {error}"); | |
| 276 | } | |
| 277 | } | |
| 278 | ||
| 279 | /// Adds candidates, or counts them as another sighting of what is there. | |
| 280 | pub(crate) async fn capture(&self, workspace: &str, items: &[CaptureItem], by: &str) -> Result<Captured> { | |
| 281 | let workspace = workspace.to_lowercase(); | |
| 282 | let mut done = Captured::default(); | |
| 283 | let mut paths = HashMap::new(); | |
| 284 | for item in items.iter().take(MAX_CAPTURE) { | |
| 285 | let Ok((text, evidence)) = checked(item) else { | |
| 286 | done.refused += 1; | |
| 287 | continue; | |
| 288 | }; | |
| 289 | let (key, repo) = match (item.scope, &item.repo_id) { | |
| 290 | (MemoryScope::Workspace, _) => (workspace.clone(), item.repo_id.clone()), | |
| 291 | (MemoryScope::Project, Some(repo_id)) => (repo_id.clone(), Some(repo_id.clone())), | |
| 292 | (MemoryScope::Project, None) => { | |
| 293 | done.refused += 1; | |
| 294 | continue; | |
| 295 | } | |
| 296 | }; | |
| 297 | let path = match &repo { | |
| 298 | Some(repo_id) => self.path_of(&mut paths, repo_id).await?, | |
| 299 | None => None, | |
| 300 | }; | |
| 301 | // A project's memory belongs to a repository of this workspace. | |
| 302 | if item.scope == MemoryScope::Project | |
| 303 | && !path.as_ref().is_some_and(|path| path.namespace.to_lowercase() == workspace) | |
| 304 | { | |
| 305 | done.refused += 1; | |
| 306 | continue; | |
| 307 | } | |
| 308 | let print = fingerprint(&text); | |
| 309 | let now = rfc3339(now_ms()); | |
| 310 | let existing = self | |
| 311 | .db | |
| 312 | .prepare( | |
| 313 | "SELECT id, status, sources, confidence FROM memories | |
| 314 | WHERE scope = ? AND scope_key = ? AND (fingerprint = ? OR text = ?) LIMIT 1", | |
| 315 | ) | |
| 316 | .bind(&[item.scope.as_str().into(), key.as_str().into(), print.as_str().into(), text.as_str().into()])? | |
| 317 | .first::<Existing>(None) | |
| 318 | .await?; | |
| 319 | if let Some(existing) = existing { | |
| 320 | done.merged += 1; | |
| 321 | if existing.status == MemoryStatus::Dismissed.as_str() { | |
| 322 | continue; | |
| 323 | } | |
| 324 | let sources = with_source(&existing.sources, &item.reference); | |
| 325 | let confidence = match (existing.confidence, item.confidence) { | |
| 326 | (Some(a), Some(b)) => Some(a.max(b)), | |
| 327 | (a, b) => a.or(b), | |
| 328 | }; | |
| 329 | let keep = existing.status == MemoryStatus::Candidate.as_str() | |
| 330 | && promotes(sources.len(), item.source, confidence); | |
| 331 | if keep { | |
| 332 | done.kept += 1; | |
| 333 | } | |
| 334 | let status = if keep { MemoryStatus::Kept.as_str() } else { existing.status.as_str() }; | |
| 335 | self.db | |
| 336 | .prepare( | |
| 337 | "UPDATE memories SET sources = ?, confidence = ?, status = ?, fingerprint = ?, updated_at = ? | |
| 338 | WHERE id = ?", | |
| 339 | ) | |
| 340 | .bind(&[ | |
| 341 | serde_json::to_string(&sources)?.into(), | |
| 342 | confidence.map_or(JsValue::NULL, JsValue::from), | |
| 343 | status.into(), | |
| 344 | print.as_str().into(), | |
| 345 | now.as_str().into(), | |
| 346 | existing.id.as_str().into(), | |
| 347 | ])? | |
| 348 | .run() | |
| 349 | .await?; | |
| 350 | if keep { | |
| 351 | self.memory_changed(&existing.id, &workspace, status, repo.as_deref()).await; | |
| 352 | } | |
| 353 | continue; | |
| 354 | } | |
| 355 | let count = self | |
| 356 | .db | |
| 357 | .prepare("SELECT count(*) AS value FROM memories WHERE scope = ? AND scope_key = ? AND status != 'dismissed'") | |
| 358 | .bind(&[item.scope.as_str().into(), key.as_str().into()])? | |
| 359 | .first::<u32>(Some("value")) | |
| 360 | .await? | |
| 361 | .unwrap_or_default(); | |
| 362 | if count >= MAX_PER_SCOPE { | |
| 363 | done.refused += 1; | |
| 364 | continue; | |
| 365 | } | |
| 366 | let keep = promotes(1, item.source, item.confidence); | |
| 367 | let status = if keep { MemoryStatus::Kept } else { MemoryStatus::Candidate }; | |
| 368 | let id = new_id("mem", now_ms()); | |
| 369 | let repo_text = path.as_ref().map(|path| format!("{}/{}", path.namespace, path.name)); | |
| 370 | self.db | |
| 371 | .prepare( | |
| 372 | "INSERT INTO memories | |
| 373 | (id, scope, scope_key, workspace, repo, text, kind, source_kind, source_run, | |
| 374 | source_repo, source_number, created_by, pinned, created_at, updated_at, | |
| 375 | status, confidence, source_ref, evidence, fingerprint, sources) | |
| 376 | VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 0, ?, ?, ?, ?, ?, ?, ?, ?)", | |
| 377 | ) | |
| 378 | .bind(&[ | |
| 379 | id.as_str().into(), | |
| 380 | item.scope.as_str().into(), | |
| 381 | key.as_str().into(), | |
| 382 | workspace.as_str().into(), | |
| 383 | if item.scope == MemoryScope::Project { repo_text.clone().map_or(JsValue::NULL, JsValue::from) } else { JsValue::NULL }, | |
| 384 | text.as_str().into(), | |
| 385 | item.kind.as_str().into(), | |
| 386 | item.source.as_str().into(), | |
| 387 | item.run_id.as_deref().map_or(JsValue::NULL, JsValue::from), | |
| 388 | repo_text.map_or(JsValue::NULL, JsValue::from), | |
| 389 | item.number.map_or(JsValue::NULL, JsValue::from), | |
| 390 | by.into(), | |
| 391 | now.as_str().into(), | |
| 392 | now.as_str().into(), | |
| 393 | status.as_str().into(), | |
| 394 | item.confidence.map_or(JsValue::NULL, JsValue::from), | |
| 395 | item.reference.as_str().into(), | |
| 396 | evidence.map_or(JsValue::NULL, JsValue::from), | |
| 397 | print.as_str().into(), | |
| 398 | serde_json::to_string(&[&item.reference])?.into(), | |
| 399 | ])? | |
| 400 | .run() | |
| 401 | .await?; | |
| 402 | done.added += 1; | |
| 403 | if keep { | |
| 404 | done.kept += 1; | |
| 405 | } | |
| 406 | self.memory_changed(&id, &workspace, status.as_str(), repo.as_deref()).await; | |
| 407 | } | |
| 408 | Ok(done) | |
| 409 | } | |
| 410 | ||
| 411 | pub(crate) async fn capture_memories(&self, a: CaptureMemoriesArgs) -> Result<Captured> { | |
| 412 | self.capture(&a.workspace, &a.items, a.by.as_deref().unwrap_or("g1t")).await | |
| 413 | } | |
| 414 | ||
| 415 | pub(crate) async fn report_learned(&self, a: ReportLearnedArgs) -> Result<Outcome<Captured>> { | |
| 416 | let run = self | |
| 417 | .db | |
| 418 | .prepare("SELECT id, workspace, repo_id, number, token_hash FROM agent_runs WHERE id = ?") | |
| 419 | .bind(&[a.run_id.as_str().into()])? | |
| 420 | .first::<RunTicket>(None) | |
| 421 | .await? | |
| 422 | .filter(|run| !a.token.is_empty() && run.token_hash == hash(&a.token)); | |
| 423 | let Some(run) = run else { | |
| 424 | return Ok(Outcome::fail(FailureCode::NotFound, "Run not found.")); | |
| 425 | }; | |
| 426 | let items: Vec<CaptureItem> = a | |
| 427 | .items | |
| 428 | .iter() | |
| 429 | .take(MAX_LEARNED) | |
| 430 | .map(|learned| CaptureItem { | |
| 431 | scope: if learned.scope.as_deref() == Some("workspace") { | |
| 432 | MemoryScope::Workspace | |
| 433 | } else { | |
| 434 | MemoryScope::Project | |
| 435 | }, | |
| 436 | repo_id: Some(run.repo_id.clone()), | |
| 437 | kind: learned.kind.as_deref().and_then(MemoryKind::parse).unwrap_or_default(), | |
| 438 | text: learned.text.clone(), | |
| 439 | confidence: Some(RUN_CONFIDENCE), | |
| 440 | source: CaptureSource::Run, | |
| 441 | reference: format!("run:{}", run.id), | |
| 442 | evidence: learned.evidence.clone(), | |
| 443 | number: run.number, | |
| 444 | run_id: Some(run.id.clone()), | |
| 445 | }) | |
| 446 | .collect(); | |
| 447 | Ok(Outcome::Ok(self.capture(&run.workspace, &items, "g1t-agent").await?)) | |
| 448 | } | |
| 449 | ||
| 450 | pub(crate) async fn list_candidates(&self, a: ListCandidatesArgs) -> Result<Outcome<Vec<Memory>>> { | |
| 451 | let workspace = a.workspace.to_lowercase(); | |
| 452 | if !a.viewer.as_ref().is_some_and(|viewer| viewer.is_member(&workspace)) { | |
| 453 | return Ok(Outcome::fail(FailureCode::Forbidden, "Memory is for members of the workspace and its agents.")); | |
| 454 | } | |
| 455 | let statement = match &a.repo { | |
| 456 | Some(path) => { | |
| 457 | let repo = match self.repo(path, &a.viewer).await? { | |
| 458 | Outcome::Ok(repo) => repo, | |
| 459 | Outcome::Fail(failure) => return Ok(Outcome::Fail(failure)), | |
| 460 | }; | |
| 461 | // The project's own, and the workspace's learned there. | |
| 462 | self.db | |
| 463 | .prepare( | |
| 464 | "SELECT * FROM memories WHERE workspace = ? AND status = 'candidate' | |
| 465 | AND (scope_key = ? OR (scope = 'workspace' AND source_repo = ?)) | |
| 466 | ORDER BY updated_at DESC LIMIT 200", | |
| 467 | ) | |
| 468 | .bind(&[ | |
| 469 | workspace.as_str().into(), | |
| 470 | repo.id.as_str().into(), | |
| 471 | format!("{}/{}", repo.namespace, repo.name).into(), | |
| 472 | ])? | |
| 473 | } | |
| 474 | None => self | |
| 475 | .db | |
| 476 | .prepare("SELECT * FROM memories WHERE workspace = ? AND status = 'candidate' ORDER BY updated_at DESC LIMIT 200") | |
| 477 | .bind(&[workspace.as_str().into()])?, | |
| 478 | }; | |
| 479 | let rows = statement.all().await?.results::<crate::memory::MemoryRow>()?; | |
| 480 | Ok(Outcome::Ok(rows.into_iter().map(Memory::from).collect())) | |
| 481 | } | |
| 482 | ||
| 483 | pub(crate) async fn review_memory(&self, a: ReviewMemoryArgs) -> Result<Outcome<Memory>> { | |
| 484 | let workspace = a.workspace.to_lowercase(); | |
| 485 | if !a.actor.is_member(&workspace) || a.actor.kind != PrincipalKind::User || !a.actor.verified { | |
| 486 | return Ok(Outcome::fail(FailureCode::Forbidden, "Only members of the workspace can review its memory.")); | |
| 487 | } | |
| 488 | let Some(row) = self | |
| 489 | .db | |
| 490 | .prepare("SELECT * FROM memories WHERE id = ? AND workspace = ?") | |
| 491 | .bind(&[a.id.as_str().into(), workspace.as_str().into()])? | |
| 492 | .first::<crate::memory::MemoryRow>(None) | |
| 493 | .await? | |
| 494 | else { | |
| 495 | return Ok(Outcome::fail(FailureCode::NotFound, "Memory not found.")); | |
| 496 | }; | |
| 497 | let text = match a.text.as_deref().map(str::trim).filter(|text| !text.is_empty()) { | |
| 498 | Some(text) => { | |
| 499 | if text.chars().count() > MAX_MEMORY_CHARS { | |
| 500 | return Ok(Outcome::fail(FailureCode::Invalid, format!("Keep a memory to {MAX_MEMORY_CHARS} characters."))); | |
| 501 | } | |
| 502 | if let Some(what) = secret_in(text) { | |
| 503 | return Ok(Outcome::fail( | |
| 504 | FailureCode::Invalid, | |
| 505 | format!("That looks like {what}. Memory never holds secrets; say where the secret lives instead."), | |
| 506 | )); | |
| 507 | } | |
| 508 | Some(text.to_owned()) | |
| 509 | } | |
| 510 | None => None, | |
| 511 | }; | |
| 512 | let status = match a.decision { | |
| 513 | ReviewDecision::Keep => MemoryStatus::Kept, | |
| 514 | ReviewDecision::Dismiss => MemoryStatus::Dismissed, | |
| 515 | }; | |
| 516 | let now = rfc3339(now_ms()); | |
| 517 | let print = text.as_deref().map(fingerprint); | |
| 518 | self.db | |
| 519 | .prepare( | |
| 520 | "UPDATE memories SET status = ?, text = COALESCE(?, text), fingerprint = COALESCE(?, fingerprint), | |
| 521 | kind = COALESCE(?, kind), reviewed_by = ?, reviewed_at = ?, updated_at = ? | |
| 522 | WHERE id = ? AND workspace = ?", | |
| 523 | ) | |
| 524 | .bind(&[ | |
| 525 | status.as_str().into(), | |
| 526 | text.map_or(JsValue::NULL, JsValue::from), | |
| 527 | print.map_or(JsValue::NULL, JsValue::from), | |
| 528 | a.kind.map_or(JsValue::NULL, |kind| kind.as_str().into()), | |
| 529 | a.actor.username.as_str().into(), | |
| 530 | now.as_str().into(), | |
| 531 | now.as_str().into(), | |
| 532 | a.id.as_str().into(), | |
| 533 | workspace.as_str().into(), | |
| 534 | ])? | |
| 535 | .run() | |
| 536 | .await?; | |
| 537 | let repo_id = (row.scope == "project").then(|| row.scope_key.clone()); | |
| 538 | self.memory_changed(&a.id, &workspace, status.as_str(), repo_id.as_deref()).await; | |
| 539 | let reviewed = self | |
| 540 | .db | |
| 541 | .prepare("SELECT * FROM memories WHERE id = ?") | |
| 542 | .bind(&[a.id.as_str().into()])? | |
| 543 | .first::<crate::memory::MemoryRow>(None) | |
| 544 | .await?; | |
| 545 | Ok(match reviewed { | |
| 546 | Some(row) => Outcome::Ok(Memory::from(row)), | |
| 547 | None => Outcome::fail(FailureCode::NotFound, "Memory not found."), | |
| 548 | }) | |
| 549 | } | |
| 550 | ||
| 551 | pub(crate) async fn memories_by_id(&self, a: MemoriesByIdArgs) -> Result<Vec<Memory>> { | |
| 552 | let ids: Vec<&String> = a.ids.iter().take(100).collect(); | |
| 553 | let rows = self | |
| 554 | .db | |
| 555 | .prepare("SELECT * FROM memories WHERE workspace = ? AND id IN (SELECT value FROM json_each(?))") | |
| 556 | .bind(&[a.workspace.to_lowercase().into(), serde_json::to_string(&ids)?.into()])? | |
| 557 | .all() | |
| 558 | .await? | |
| 559 | .results::<crate::memory::MemoryRow>()?; | |
| 560 | Ok(rows.into_iter().map(Memory::from).collect()) | |
| 561 | } | |
| 562 | ||
| 563 | pub(crate) async fn search_memories(&self, a: SearchMemoriesArgs) -> Result<Vec<Memory>> { | |
| 564 | let status = a.status.unwrap_or(MemoryStatus::Kept); | |
| 565 | let limit = a.limit.unwrap_or(20).clamp(1, 100) as usize; | |
| 566 | let statement = match &a.repo_ids { | |
| 567 | Some(ids) => self | |
| 568 | .db | |
| 569 | .prepare( | |
| 570 | "SELECT * FROM memories WHERE workspace = ? AND status = ? | |
| 571 | AND (scope = 'workspace' OR scope_key IN (SELECT value FROM json_each(?))) | |
| 572 | ORDER BY pinned DESC, COALESCE(last_used_at, updated_at) DESC LIMIT 1000", | |
| 573 | ) | |
| 574 | .bind(&[a.workspace.to_lowercase().into(), status.as_str().into(), serde_json::to_string(ids)?.into()])?, | |
| 575 | None => self | |
| 576 | .db | |
| 577 | .prepare( | |
| 578 | "SELECT * FROM memories WHERE workspace = ? AND status = ? | |
| 579 | ORDER BY pinned DESC, COALESCE(last_used_at, updated_at) DESC LIMIT 1000", | |
| 580 | ) | |
| 581 | .bind(&[a.workspace.to_lowercase().into(), status.as_str().into()])?, | |
| 582 | }; | |
| 583 | let words: Vec<String> = a | |
| 584 | .query | |
| 585 | .as_deref() | |
| 586 | .unwrap_or_default() | |
| 587 | .split_whitespace() | |
| 588 | .map(str::to_lowercase) | |
| 589 | .collect(); | |
| 590 | Ok(statement | |
| 591 | .all() | |
| 592 | .await? | |
| 593 | .results::<crate::memory::MemoryRow>()? | |
| 594 | .into_iter() | |
| 595 | .map(Memory::from) | |
| 596 | .filter(|memory| { | |
| 597 | let text = memory.text.to_lowercase(); | |
| 598 | words.iter().all(|word| text.contains(word.as_str())) | |
| 599 | }) | |
| 600 | .take(limit) | |
| 601 | .collect()) | |
| 602 | } | |
| 603 | ||
| 604 | /// What a merged pull request decided, as a candidate. | |
| 605 | fn decision_item(pull: &Pull) -> Option<CaptureItem> { | |
| 606 | let files: Vec<String> = pull.files.iter().map(|file| file.path.clone()).collect(); | |
| 607 | let text = decision(pull.number, &pull.title, pull.body.as_deref(), &files)?; | |
| 608 | Some(CaptureItem { | |
| 609 | scope: MemoryScope::Project, | |
| 610 | repo_id: Some(pull.repo_id.clone()), | |
| 611 | kind: MemoryKind::Decision, | |
| 612 | text, | |
| 613 | confidence: Some(MERGE_CONFIDENCE), | |
| 614 | source: CaptureSource::Pr, | |
| 615 | reference: format!("pull:{}#{}", pull.repo_id, pull.number), | |
| 616 | evidence: Some(clip(&pull.title, MAX_EVIDENCE_CHARS)), | |
| 617 | number: Some(pull.number), | |
| 618 | run_id: None, | |
| 619 | }) | |
| 620 | } | |
| 621 | ||
| 622 | /// What a person's comment on a pull request corrected, as a candidate. | |
| 623 | fn correction_item(pull: &Pull, comment: &Comment) -> Option<CaptureItem> { | |
| 624 | if comment.kind != CommentKind::Comment | |
| 625 | || comment.author.kind == PrincipalKind::Agent | |
| 626 | || NOT_PEOPLE.contains(&comment.author.username.as_str()) | |
| 627 | { | |
| 628 | return None; | |
| 629 | } | |
| 630 | let text = correction(&comment.body, comment.verdict == Some(Verdict::RequestChanges))?; | |
| 631 | Some(CaptureItem { | |
| 632 | scope: MemoryScope::Project, | |
| 633 | repo_id: Some(pull.repo_id.clone()), | |
| 634 | kind: MemoryKind::Convention, | |
| 635 | text, | |
| 636 | confidence: Some(REVIEW_CONFIDENCE), | |
| 637 | source: CaptureSource::Review, | |
| 638 | reference: format!("comment:{}", comment.id), | |
| 639 | evidence: Some(format!( | |
| 640 | "{} on #{}{}: {}", | |
| 641 | comment.author.username, | |
| 642 | pull.number, | |
| 643 | comment.path.as_deref().map(|path| format!(" ({path})")).unwrap_or_default(), | |
| 644 | clip(&comment.body, MAX_EVIDENCE_CHARS - 60) | |
| 645 | )), | |
| 646 | number: Some(pull.number), | |
| 647 | run_id: None, | |
| 648 | }) | |
| 649 | } | |
| 650 | ||
| 651 | async fn workspace_of(&self, repo_id: &str) -> Result<Option<String>> { | |
| 652 | let mut cache = HashMap::new(); | |
| 653 | Ok(self.path_of(&mut cache, repo_id).await?.map(|path| path.namespace.to_lowercase())) | |
| 654 | } | |
| 655 | ||
| 656 | pub(crate) async fn seed_from_pulls(&self, a: SeedFromPullsArgs) -> Result<Captured> { | |
| 657 | let Some(workspace) = self.workspace_of(&a.repo_id).await? else { | |
| 658 | return Ok(Captured::default()); | |
| 659 | }; | |
| 660 | let limit = a.limit.unwrap_or(20).clamp(1, 50); | |
| 661 | let pulls: Vec<Pull> = self | |
| 662 | .db | |
| Thirteen MCP tools and classic token scopes; agents rate their confidence and can be put on an issue in one step | 663 | .prepare(format!("SELECT {PULL_COLUMNS} FROM pulls WHERE repo_id = ? AND status = 'merged' ORDER BY merged_at DESC LIMIT ?")) |
| Agents get guardrails, run credentials, an audit log, a context hub, repository instructions and mentions; security upkeep; snake_case API | 664 | .bind(&[a.repo_id.as_str().into(), limit.into()])? |
| 665 | .all() | |
| 666 | .await? | |
| 667 | .results::<PullRow>()? | |
| 668 | .into_iter() | |
| 669 | .map(Pull::from) | |
| 670 | .collect(); | |
| 671 | let mut items = Vec::new(); | |
| 672 | for pull in &pulls { | |
| 673 | items.extend(Self::decision_item(pull)); | |
| 674 | let comments = self | |
| 675 | .db | |
| 676 | .prepare("SELECT * FROM comments WHERE repo_id = ? AND number = ? ORDER BY id LIMIT 100") | |
| 677 | .bind(&[pull.repo_id.as_str().into(), pull.number.into()])? | |
| 678 | .all() | |
| 679 | .await? | |
| 680 | .results::<CommentRow>()?; | |
| 681 | items.extend( | |
| 682 | comments | |
| 683 | .into_iter() | |
| 684 | .map(Comment::from) | |
| 685 | .filter_map(|comment| Self::correction_item(pull, &comment)), | |
| 686 | ); | |
| 687 | } | |
| 688 | let mut done = Captured::default(); | |
| 689 | for chunk in items.chunks(MAX_CAPTURE) { | |
| 690 | let part = self.capture(&workspace, chunk, "g1t").await?; | |
| 691 | done.added += part.added; | |
| 692 | done.merged += part.merged; | |
| 693 | done.kept += part.kept; | |
| 694 | done.refused += part.refused; | |
| 695 | } | |
| 696 | Ok(done) | |
| 697 | } | |
| 698 | ||
| 699 | /// Captures from the bus: corrections in review, decisions on merge. | |
| 700 | async fn capture_event(&self, event: &Event) -> Result<()> { | |
| 701 | let field = |name: &str| event.data[name].as_str().map(str::to_owned); | |
| 702 | match event.kind.as_str() { | |
| 703 | "pull.merged" => { | |
| 704 | let Some(pull) = (match field("pullId") { | |
| 705 | Some(id) => self.pull_by_id(&id).await?, | |
| 706 | None => None, | |
| 707 | }) else { | |
| 708 | return Ok(()); | |
| 709 | }; | |
| 710 | if pull.status != PullStatus::Merged { | |
| 711 | return Ok(()); | |
| 712 | } | |
| 713 | if let (Some(item), Some(workspace)) = (Self::decision_item(&pull), self.workspace_of(&pull.repo_id).await?) { | |
| 714 | self.capture(&workspace, &[item], "g1t").await?; | |
| 715 | } | |
| 716 | } | |
| 717 | "comment.created" => { | |
| 718 | let (Some(comment_id), Some(pull_id)) = (field("commentId"), field("pullId")) else { | |
| 719 | return Ok(()); | |
| 720 | }; | |
| 721 | let Some(pull) = self.pull_by_id(&pull_id).await? else { | |
| 722 | return Ok(()); | |
| 723 | }; | |
| 724 | let Some(comment) = self | |
| 725 | .db | |
| 726 | .prepare("SELECT * FROM comments WHERE id = ?") | |
| 727 | .bind(&[comment_id.as_str().into()])? | |
| 728 | .first::<CommentRow>(None) | |
| 729 | .await? | |
| 730 | .map(Comment::from) | |
| 731 | else { | |
| 732 | return Ok(()); | |
| 733 | }; | |
| 734 | if let (Some(item), Some(workspace)) = | |
| 735 | (Self::correction_item(&pull, &comment), self.workspace_of(&pull.repo_id).await?) | |
| 736 | { | |
| 737 | self.capture(&workspace, &[item], "g1t").await?; | |
| 738 | } | |
| 739 | } | |
| 740 | _ => {} | |
| 741 | } | |
| 742 | Ok(()) | |
| 743 | } | |
| 744 | } | |
| 745 | ||
| 746 | /// The methods this module serves at `POST /rpc/<method>`. | |
| 747 | pub(crate) const METHODS: [&str; 7] = [ | |
| 748 | "capture_memories", | |
| 749 | "report_learned", | |
| 750 | "list_candidates", | |
| 751 | "review_memory", | |
| 752 | "memories_by_id", | |
| 753 | "search_memories", | |
| 754 | "seed_from_pulls", | |
| 755 | ]; | |
| 756 | ||
| 757 | pub(crate) async fn dispatch(work: &Work, method: &str, body: serde_json::Value) -> Result<worker::Response> { | |
| 758 | use g1t_kit::{args, reply}; | |
| 759 | match method { | |
| 760 | "capture_memories" => reply(&work.capture_memories(args(body)?).await?), | |
| 761 | "report_learned" => reply(&work.report_learned(args(body)?).await?), | |
| 762 | "list_candidates" => reply(&work.list_candidates(args(body)?).await?), | |
| 763 | "review_memory" => reply(&work.review_memory(args(body)?).await?), | |
| 764 | "memories_by_id" => reply(&work.memories_by_id(args(body)?).await?), | |
| 765 | "search_memories" => reply(&work.search_memories(args(body)?).await?), | |
| 766 | "seed_from_pulls" => reply(&work.seed_from_pulls(args(body)?).await?), | |
| 767 | _ => worker::Response::error("Unknown method", 404), | |
| 768 | } | |
| 769 | } | |
| 770 | ||
| 771 | /// Captures what an event teaches. Never holds up the queue: a failure is | |
| 772 | /// logged and the event goes on to its other handlers. | |
| 773 | pub(crate) async fn on_event(work: &Work, event: &Event) { | |
| 774 | if let Err(error) = work.capture_event(event).await { | |
| 775 | worker::console_error!("capturing memory from {} {} failed: {error}", event.kind, event.id); | |
| 776 | } | |
| 777 | } | |
| 778 | ||
| 779 | #[cfg(test)] | |
| 780 | mod tests { | |
| 781 | use super::*; | |
| 782 | ||
| 783 | #[test] | |
| 784 | fn a_correction_keeps_the_sentence_that_says_how() { | |
| 785 | let body = "Thanks for this! We use the shared `api` client instead of calling fetch directly. Otherwise looks fine."; | |
| 786 | assert_eq!( | |
| 787 | correction(body, false).as_deref(), | |
| 788 | Some("We use the shared `api` client instead of calling fetch directly.") | |
| 789 | ); | |
| 790 | } | |
| 791 | ||
| 792 | #[test] | |
| 793 | fn questions_and_praise_teach_nothing() { | |
| 794 | assert_eq!(correction("Why not use the cache here?", false), None); | |
| 795 | assert_eq!(correction("LGTM, nice work.", false), None); | |
| 796 | assert_eq!(correction("> never do this\n\nok", false), None); | |
| 797 | assert_eq!(correction("```\nnever run this\n```", false), None); | |
| 798 | } | |
| 799 | ||
| 800 | #[test] | |
| 801 | fn a_request_for_changes_without_a_marker_keeps_its_first_sentence() { | |
| 802 | assert_eq!( | |
| 803 | correction("The migration has to be idempotent for D1 replays.", true).as_deref(), | |
| 804 | Some("The migration has to be idempotent for D1 replays.") | |
| 805 | ); | |
| 806 | assert_eq!(correction("The migration has to be idempotent for D1 replays.", false), None); | |
| 807 | } | |
| 808 | ||
| 809 | #[test] | |
| 810 | fn a_decision_names_why_and_the_files() { | |
| 811 | let files: Vec<String> = (1..=8).map(|n| format!("src/f{n}.rs")).collect(); | |
| 812 | let text = decision( | |
| 813 | 12, | |
| 814 | "Keep the v1 webhook payload", | |
| 815 | Some("## Summary\n\nTwo customers still parse the v1 payload, so it stays alongside v2.\n\nMore detail."), | |
| 816 | &files, | |
| 817 | ) | |
| 818 | .unwrap(); | |
| 819 | assert!(text.starts_with("Decided in #12, \"Keep the v1 webhook payload\": Two customers still parse")); | |
| 820 | assert!(text.contains("src/f1.rs") && text.contains("and 2 more")); | |
| 821 | assert!(!text.contains("More detail")); | |
| 822 | } | |
| 823 | ||
| 824 | #[test] | |
| 825 | fn a_pull_request_that_says_only_its_title_decides_nothing() { | |
| 826 | assert_eq!(decision(3, "Fix typo", Some("Fix typo."), &[]), None); | |
| 827 | assert_eq!(decision(3, "Fix typo", None, &[]), None); | |
| 828 | } | |
| 829 | ||
| 830 | #[test] | |
| 831 | fn sources_are_counted_once_each() { | |
| 832 | assert_eq!(with_source("[\"run:a\"]", "run:a"), ["run:a"]); | |
| 833 | assert_eq!(with_source("[\"run:a\"]", "comment:b"), ["run:a", "comment:b"]); | |
| 834 | assert_eq!(with_source("not json", "doc:x"), ["doc:x"]); | |
| 835 | } | |
| 836 | ||
| 837 | fn item(text: &str, evidence: Option<&str>) -> CaptureItem { | |
| 838 | CaptureItem { | |
| 839 | scope: MemoryScope::Project, | |
| 840 | repo_id: Some("rep_1".into()), | |
| 841 | kind: MemoryKind::Gotcha, | |
| 842 | text: text.into(), | |
| 843 | confidence: Some(0.6), | |
| 844 | source: CaptureSource::Run, | |
| 845 | reference: "run:1".into(), | |
| 846 | evidence: evidence.map(str::to_owned), | |
| 847 | number: None, | |
| 848 | run_id: None, | |
| 849 | } | |
| 850 | } | |
| 851 | ||
| 852 | #[test] | |
| 853 | fn secrets_are_refused_on_capture_in_text_and_evidence() { | |
| 854 | // Key-shaped values are joined at run time, as in the contracts' tests. | |
| 855 | let key = format!("{}{}", "gh", "p_abcdefghijklmnopqrstuvwxyz0123456789"); | |
| 856 | assert!(checked(&item(&format!("Clone with {key} for the mirror."), None)).is_err()); | |
| 857 | assert!(checked(&item("The mirror needs a token from the MIRROR_TOKEN secret.", Some(&format!("export T={key}")))).is_err()); | |
| 858 | let (text, evidence) = checked(&item(" Tests need TZ=UTC. ", Some("cargo test failed on dates"))).unwrap(); | |
| 859 | assert_eq!(text, "Tests need TZ=UTC."); | |
| 860 | assert_eq!(evidence.as_deref(), Some("cargo test failed on dates")); | |
| 861 | assert!(checked(&item("short", None)).is_err()); | |
| 862 | } | |
| 863 | } |