The best way to control Claude Code, Codex and more from any computer or phone
| 1 | import { |
| 2 | IsoDateTime, |
| 3 | MessageId, |
| 4 | OrchestrationThreadSearchSource, |
| 5 | OrchestrationV2AppThreadJson, |
| 6 | OrchestrationV2ThreadOpens, |
| 7 | ThreadLinkedPullRequest, |
| 8 | ThreadPullRequestLink, |
| 9 | type OrchestrationSearchThreadsInput, |
| 10 | type OrchestrationSearchThreadsResult, |
| 11 | type OrchestrationV2SearchThreadsInput, |
| 12 | type OrchestrationV2SearchThreadsResult, |
| 13 | type OrchestrationV2ThreadShellSnapshot, |
| 14 | ProjectId, |
| 15 | ThreadId, |
| 16 | } from "@t3tools/contracts"; |
| 17 | import * as DateTime from "effect/DateTime"; |
| 18 | import * as Context from "effect/Context"; |
| 19 | import * as Effect from "effect/Effect"; |
| 20 | import * as Layer from "effect/Layer"; |
| 21 | import * as Option from "effect/Option"; |
| 22 | import * as Schema from "effect/Schema"; |
| 23 | import * as SqlClient from "effect/sql/SqlClient"; |
| 24 | import * as SqlSchema from "effect/sql/SqlSchema"; |
| 25 | import * as NodeSqliteClient from "@t3tools/shared/nodeSqliteClient"; |
| 26 | import { |
| 27 | messageSearchTokens, |
| 28 | foldSearchQuery, |
| 29 | rankMessageMatches, |
| 30 | rankSearchItems, |
| 31 | messageMatcher, |
| 32 | scoreThreadSearch, |
| 33 | type ThreadSearchSubject, |
| 34 | } from "@t3tools/shared/paletteSearch"; |
| 35 | import { threadSearchPullRequests } from "@t3tools/shared/threadPullRequests"; |
| 36 | import { |
| 37 | foldedMessageSearchText, |
| 38 | indexedMessageCondition, |
| 39 | messageSearchText, |
| 40 | nonPlaceholderMessageCondition, |
| 41 | } from "./messageSearchIndex.ts"; |
| 42 | import { loadOlderSettledThreadsPage } from "./ShellStream.ts"; |
| 43 | |
| 44 | // When the user last wrote in the thread aliased `threads`, as the clients |
| 45 | // rank recency; agent notices in the user role do not count. |
| 46 | const latestUserMessageAt = (sql: SqlClient.SqlClient) => sql`( |
| 47 | SELECT messages.updated_at |
| 48 | FROM orchestration_v2_projection_messages AS messages |
| 49 | WHERE messages.thread_id = threads.thread_id |
| 50 | AND messages.role = 'user' |
| 51 | AND COALESCE(json_extract(messages.payload_json, '$.createdBy'), 'user') = 'user' |
| 52 | ORDER BY messages.updated_at DESC, messages.message_id DESC |
| 53 | LIMIT 1 |
| 54 | )`; |
| 55 | |
| 56 | export class ThreadSearchError extends Schema.TaggedError<ThreadSearchError>()( |
| 57 | "ThreadSearchError", |
| 58 | { |
| 59 | operation: Schema.Literals(["query", "decode"]), |
| 60 | cause: Schema.Defect(), |
| 61 | }, |
| 62 | ) { |
| 63 | override get message(): string { |
| 64 | return `Thread search ${this.operation} failed.`; |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | // Token lists bind as JSON arrays, so any number of them is one parameter. |
| 69 | const TokenList = Schema.fromJsonString(Schema.Array(Schema.String)); |
| 70 | const SearchRequest = Schema.Struct({ |
| 71 | /** The FTS5 query, or null to scan messages instead. */ |
| 72 | match: Schema.NullOr(Schema.String), |
| 73 | /** Tokens every hit must contain, as typed or as their GLOB folding case and Latin accents. */ |
| 74 | substrings: Schema.fromJsonString( |
| 75 | Schema.Array(Schema.Struct({ token: Schema.String, accentInsensitive: Schema.String })), |
| 76 | ), |
| 77 | /** For each short word outside ASCII, the spellings a hit must contain one of. */ |
| 78 | shortSpellings: Schema.fromJsonString(Schema.Array(Schema.Array(Schema.String))), |
| 79 | projectId: Schema.NullOr(ProjectId), |
| 80 | /** Threads the caller already lists, left out before the hit limit. */ |
| 81 | excludedThreadIds: TokenList, |
| 82 | }); |
| 83 | const SearchRow = Schema.Struct({ |
| 84 | messageRowid: Schema.Number, |
| 85 | messageId: MessageId, |
| 86 | threadId: ThreadId, |
| 87 | projectId: ProjectId, |
| 88 | source: OrchestrationThreadSearchSource, |
| 89 | text: Schema.String, |
| 90 | messageCreatedAt: Schema.NullOr(IsoDateTime), |
| 91 | }); |
| 92 | |
| 93 | const ThreadOpensRow = Schema.Struct({ |
| 94 | threadId: ThreadId, |
| 95 | opens: Schema.NullOr(Schema.String), |
| 96 | latestUserMessageAt: Schema.NullOr(IsoDateTime), |
| 97 | }); |
| 98 | const decodeThreadOpens = Schema.decodeUnknownOption( |
| 99 | Schema.fromJsonString(OrchestrationV2ThreadOpens), |
| 100 | ); |
| 101 | |
| 102 | const SNIPPET_LENGTH = 240; |
| 103 | // The newest hits read per query. A common word stops here instead of reading |
| 104 | // every message, so time stays flat as history grows. |
| 105 | const MESSAGE_HIT_LIMIT = 2000; |
| 106 | // How many of the newest index hits the filters may pass over to fill that |
| 107 | // limit: a rare short word or hidden threads would otherwise read every hit. |
| 108 | const MESSAGE_CANDIDATE_LIMIT = 10_000; |
| 109 | // Without the index, how many of the newest messages a search scans. |
| 110 | const MESSAGE_SCAN_LIMIT = 20_000; |
| 111 | const NON_ASCII = /[\u0080-\uffff]/; |
| 112 | |
| 113 | /** SQLite's lower() folds ASCII only, so a short word is checked as each spelling it can match. */ |
| 114 | function shortTokenSpellings(token: string): ReadonlyArray<string> { |
| 115 | const forms = (char: string) => { |
| 116 | const plain = char.normalize("NFD").replace(/\p{M}/gu, ""); |
| 117 | return [char, char.toUpperCase(), plain, plain.toUpperCase()]; |
| 118 | }; |
| 119 | const spellings = Array.from(token).reduce<ReadonlyArray<string>>( |
| 120 | (prefixes, char) => prefixes.flatMap((prefix) => forms(char).map((form) => prefix + form)), |
| 121 | [""], |
| 122 | ); |
| 123 | return [...new Set(spellings)]; |
| 124 | } |
| 125 | |
| 126 | // Latin-1 Supplement and Extended-A/B/Additional letters that fold to ASCII, |
| 127 | // including Romanian and Vietnamese letters SQLite's lower() leaves alone. |
| 128 | const accentedForms = Map.groupBy( |
| 129 | [ |
| 130 | ...Array.from({ length: 0x250 - 0xc0 }, (_, offset) => String.fromCharCode(0xc0 + offset)), |
| 131 | ...Array.from({ length: 0x1f00 - 0x1e00 }, (_, offset) => String.fromCharCode(0x1e00 + offset)), |
| 132 | ].filter((char) => /^[a-z]$/.test(foldSearchQuery(char))), |
| 133 | (char) => foldSearchQuery(char), |
| 134 | ); |
| 135 | // Only text that may hold one of those letters (by UTF-8 lead byte) needs the GLOB: |
| 136 | // lower() already folds ASCII case. |
| 137 | const MESSAGE_HOLDS_ACCENTED_LATIN = [ |
| 138 | ...new Set( |
| 139 | [...accentedForms.values()].flat().map((char) => Buffer.from(char).toString("hex", 0, 1)), |
| 140 | ), |
| 141 | ] |
| 142 | .map((lead) => `instr(CAST(message.text AS BLOB), x'${lead}') > 0`) |
| 143 | .join(" OR "); |
| 144 | |
| 145 | /** A GLOB pattern for `token` anywhere, in any case and with or without Latin accents. */ |
| 146 | function accentInsensitiveGlob(token: string): string { |
| 147 | const pattern = Array.from(token, (char) => { |
| 148 | const letter = foldSearchQuery(char); |
| 149 | if (/^[a-z]$/.test(letter)) { |
| 150 | return `[${letter}${letter.toUpperCase()}${accentedForms.get(letter)?.join("") ?? ""}]`; |
| 151 | } |
| 152 | return /[*?[]/.test(char) ? `[${char}]` : char; |
| 153 | }); |
| 154 | return `*${pattern.join("")}*`; |
| 155 | } |
| 156 | |
| 157 | /** |
| 158 | * The SQL inputs for a message search, or null when the query has no token |
| 159 | * long enough to search. Tokens of three or more characters go to the |
| 160 | * trigram index as quoted phrases; shorter ones, and every token when the |
| 161 | * index cannot serve the query, are checked as substrings of the hits. |
| 162 | */ |
| 163 | function buildMessageSearch( |
| 164 | query: string, |
| 165 | hasIndex: boolean, |
| 166 | ): Omit<typeof SearchRequest.Type, "projectId" | "excludedThreadIds"> | null { |
| 167 | const { long, short } = messageSearchTokens(query); |
| 168 | if (long.length === 0) return null; |
| 169 | // Two-character CJK words are too short for trigrams. |
| 170 | const indexed = hasIndex && long.every((token) => Array.from(token).length >= 3); |
| 171 | const narrowing = short.filter((token) => !NON_ASCII.test(token)); |
| 172 | return { |
| 173 | match: indexed ? long.map((token) => `"${token.replaceAll('"', '""')}"`).join(" ") : null, |
| 174 | substrings: (indexed ? narrowing : [...long, ...narrowing]).map((token) => ({ |
| 175 | token, |
| 176 | accentInsensitive: accentInsensitiveGlob(token), |
| 177 | })), |
| 178 | shortSpellings: short.filter((token) => NON_ASCII.test(token)).map(shortTokenSpellings), |
| 179 | }; |
| 180 | } |
| 181 | |
| 182 | // A thread shows a link as its text, so in-thread find cannot reach a match |
| 183 | // in the address; messages are matched as they read. |
| 184 | const MARKDOWN_LINK = /!?\[([^\]\n]*)\]\([^)\s]*\)/g; |
| 185 | |
| 186 | // Result rows show one line, so the match must sit near the start to be seen. |
| 187 | const SNIPPET_LEAD = 24; |
| 188 | |
| 189 | const isLowSurrogate = (code: number) => code >= 0xdc00 && code <= 0xdfff; |
| 190 | |
| 191 | /** |
| 192 | * Up to 240 characters starting a few words before the match at `matchStart`, |
| 193 | * whitespace collapsed, with ellipses where the text was cut. The client finds |
| 194 | * the highlights again in the result. |
| 195 | */ |
| 196 | export function buildSearchSnippet(text: string, matchStart: number): string { |
| 197 | let start = Math.max(0, matchStart - SNIPPET_LEAD); |
| 198 | const wordStart = text.slice(start, matchStart).search(/\s/); |
| 199 | if (start > 0 && wordStart !== -1) start += wordStart + 1; |
| 200 | // Cut between characters, not inside an emoji's surrogate pair. |
| 201 | if (isLowSurrogate(text.charCodeAt(start))) start++; |
| 202 | // Collapsing whitespace shortens the text, so read more of it. |
| 203 | const rawEnd = Math.min(text.length, start + SNIPPET_LENGTH * 2); |
| 204 | // Emphasis and code marks do not show in a thread either. |
| 205 | const shown = text |
| 206 | .slice(start, rawEnd) |
| 207 | .replace(/\*\*|`+/g, "") |
| 208 | .replace(/\s+/g, " ") |
| 209 | .trim(); |
| 210 | let end = Math.min(shown.length, SNIPPET_LENGTH - 2); |
| 211 | if (isLowSurrogate(shown.charCodeAt(end))) end--; |
| 212 | const cut = end < shown.length || rawEnd < text.length; |
| 213 | return `${start > 0 ? "…" : ""}${shown.slice(0, end)}${cut ? "…" : ""}`; |
| 214 | } |
| 215 | |
| 216 | // Filter hidden subagents and side chats before the hit limit; JS checks word starts and ranking. |
| 217 | // Both paths narrow ASCII tokens with Latin-1 and Extended-A/B/Additional accent variants. |
| 218 | // Without FTS5 trigrams, or for two-character CJK words, scan the same eligible messages; |
| 219 | // accented queries and non-Latin case can still miss: "fermée"/"fermee", "ОШИБКА"/"ошибка". |
| 220 | export const messageHitsQuery = ( |
| 221 | sql: SqlClient.SqlClient, |
| 222 | { match, substrings, shortSpellings, projectId, excludedThreadIds }: typeof SearchRequest.Encoded, |
| 223 | ) => sql` |
| 224 | WITH eligible AS MATERIALIZED ( |
| 225 | SELECT threads.thread_id, threads.project_id |
| 226 | FROM orchestration_v2_projection_threads AS threads |
| 227 | INNER JOIN projection_projects AS projects |
| 228 | ON projects.project_id = threads.project_id |
| 229 | WHERE threads.deleted_at IS NULL |
| 230 | AND threads.archived_at IS NULL |
| 231 | AND projects.deleted_at IS NULL |
| 232 | AND (${projectId} IS NULL OR threads.project_id = ${projectId}) |
| 233 | AND threads.thread_id NOT IN (SELECT value FROM json_each(${excludedThreadIds})) |
| 234 | AND json_extract(threads.payload_json, '$.lineage.relationshipToParent') |
| 235 | IS NOT 'subagent' |
| 236 | AND json_extract(threads.payload_json, '$.lineage.sideChat') IS NOT 1 |
| 237 | ) |
| 238 | ${ |
| 239 | match === null |
| 240 | ? sql` |
| 241 | SELECT |
| 242 | messages.rowid AS "messageRowid", |
| 243 | messages.message_id AS "messageId", |
| 244 | messages.thread_id AS "threadId", |
| 245 | eligible.project_id AS "projectId", |
| 246 | messages.role AS source, |
| 247 | ${messageSearchText(sql)} AS text, |
| 248 | messages.created_at AS "messageCreatedAt" |
| 249 | FROM orchestration_v2_projection_messages AS messages |
| 250 | INNER JOIN eligible ON eligible.thread_id = messages.thread_id |
| 251 | WHERE ${indexedMessageCondition(sql)} |
| 252 | -- A scan reads every message it passes, so it stops at the newest ones. |
| 253 | AND messages.rowid > ( |
| 254 | SELECT coalesce(max(rowid), 0) FROM orchestration_v2_projection_messages |
| 255 | ) - ${MESSAGE_SCAN_LIMIT} |
| 256 | ` |
| 257 | : sql` |
| 258 | SELECT |
| 259 | orchestration_v2_message_search.rowid AS "messageRowid", |
| 260 | messages.message_id AS "messageId", |
| 261 | messages.thread_id AS "threadId", |
| 262 | eligible.project_id AS "projectId", |
| 263 | messages.role AS source, |
| 264 | ${messageSearchText(sql)} AS text, |
| 265 | messages.created_at AS "messageCreatedAt" |
| 266 | FROM orchestration_v2_message_search |
| 267 | INNER JOIN orchestration_v2_projection_messages AS messages |
| 268 | ON messages.rowid = orchestration_v2_message_search.rowid |
| 269 | INNER JOIN eligible ON eligible.thread_id = messages.thread_id |
| 270 | WHERE orchestration_v2_message_search MATCH ${match} |
| 271 | AND orchestration_v2_message_search.rowid >= coalesce(( |
| 272 | SELECT orchestration_v2_message_search.rowid FROM orchestration_v2_message_search |
| 273 | WHERE orchestration_v2_message_search MATCH ${match} |
| 274 | ORDER BY orchestration_v2_message_search.rowid DESC |
| 275 | LIMIT 1 OFFSET ${MESSAGE_CANDIDATE_LIMIT - 1} |
| 276 | ), 0) |
| 277 | AND ${nonPlaceholderMessageCondition(sql)} |
| 278 | ` |
| 279 | } |
| 280 | -- Answers to an agent's question show inside its question card, where |
| 281 | -- neither find nor a jump can reach them. |
| 282 | AND messages.message_id NOT LIKE 'async-answer:%' |
| 283 | -- Nor can they reach a rolled-back turn, or a message of a stopped turn |
| 284 | -- that no shown item carries, such as a queued prompt that never ran. |
| 285 | AND NOT EXISTS ( |
| 286 | SELECT 1 FROM orchestration_v2_projection_runs AS runs |
| 287 | WHERE runs.run_id = messages.run_id |
| 288 | AND ( |
| 289 | runs.status = 'rolled_back' |
| 290 | OR ( |
| 291 | runs.status = 'cancelled' |
| 292 | AND NOT EXISTS ( |
| 293 | SELECT 1 FROM orchestration_v2_projection_turn_items AS items |
| 294 | WHERE items.run_id = runs.run_id |
| 295 | AND json_extract(items.payload_json, '$.messageId') = messages.message_id |
| 296 | AND json_extract(items.payload_json, '$.inputIntent') IS NOT 'queued_turn' |
| 297 | ) |
| 298 | ) |
| 299 | ) |
| 300 | ) |
| 301 | AND NOT EXISTS ( |
| 302 | -- The text read once, not once per check. |
| 303 | SELECT 1 FROM (SELECT ${foldedMessageSearchText(sql)} AS text) AS message, |
| 304 | json_each(${substrings}) AS substring |
| 305 | WHERE instr(lower(message.text), json_extract(substring.value, '$.token')) = 0 |
| 306 | AND NOT ( |
| 307 | (${sql.literal(MESSAGE_HOLDS_ACCENTED_LATIN)}) |
| 308 | AND message.text GLOB json_extract(substring.value, '$.accentInsensitive') |
| 309 | ) |
| 310 | ) |
| 311 | AND NOT EXISTS ( |
| 312 | SELECT 1 FROM json_each(${shortSpellings}) AS token |
| 313 | WHERE NOT EXISTS ( |
| 314 | SELECT 1 FROM json_each(token.value) AS spelling |
| 315 | WHERE instr(${foldedMessageSearchText(sql)}, spelling.value) > 0 |
| 316 | ) |
| 317 | ) |
| 318 | -- The index's own rowid, so FTS5 returns hits newest first and stops at |
| 319 | -- the limit; ordering by the messages table's rowid sorts every hit. |
| 320 | ORDER BY "messageRowid" DESC |
| 321 | LIMIT ${MESSAGE_HIT_LIMIT} |
| 322 | `; |
| 323 | |
| 324 | const TitleCandidateRow = Schema.Struct({ |
| 325 | thread: Schema.fromJsonString(OrchestrationV2AppThreadJson), |
| 326 | projectTitle: Schema.String, |
| 327 | latestUserMessageAt: Schema.NullOr(IsoDateTime), |
| 328 | }); |
| 329 | |
| 330 | export interface ThreadSearchHit { |
| 331 | readonly threadId: ThreadId; |
| 332 | readonly projectId: ProjectId; |
| 333 | readonly title: string; |
| 334 | readonly source: "title" | OrchestrationThreadSearchSource; |
| 335 | readonly snippet?: string; |
| 336 | readonly messageCreatedAt?: string | null; |
| 337 | } |
| 338 | |
| 339 | export class ThreadSearch extends Context.Service< |
| 340 | ThreadSearch, |
| 341 | { |
| 342 | /** Message text matches, for clients that rank their own thread list. */ |
| 343 | readonly search: ( |
| 344 | input: OrchestrationSearchThreadsInput, |
| 345 | ) => Effect.Effect<OrchestrationSearchThreadsResult, ThreadSearchError>; |
| 346 | /** |
| 347 | * Title, branch, pull request, and project matches ranked as the command |
| 348 | * palette ranks them, then message matches for the remaining threads. |
| 349 | */ |
| 350 | readonly searchThreads: ( |
| 351 | input: OrchestrationSearchThreadsInput & { readonly projectId?: ProjectId | undefined }, |
| 352 | ) => Effect.Effect<{ readonly matches: ReadonlyArray<ThreadSearchHit> }, ThreadSearchError>; |
| 353 | /** |
| 354 | * The settled threads before `before` that clients omitting older settled threads should |
| 355 | * receive for a query: the best title, branch, pull request and project matches, ranked as |
| 356 | * the palette ranks them, then the older ones among `messageThreadIds`, in their order. |
| 357 | */ |
| 358 | readonly searchOlderSettled: (input: { |
| 359 | readonly query: string; |
| 360 | readonly before: string; |
| 361 | readonly projectIds?: ReadonlyArray<ProjectId> | undefined; |
| 362 | /** The caller's name for this environment, which its palette also matches. */ |
| 363 | readonly environmentLabel?: string | undefined; |
| 364 | readonly messageThreadIds: ReadonlyArray<ThreadId>; |
| 365 | }) => Effect.Effect<ReadonlyArray<ThreadId>, ThreadSearchError>; |
| 366 | } |
| 367 | >()("t3/orchestration-v2/ThreadSearch") {} |
| 368 | |
| 369 | // Rows a client gets per search: enough for the shown results without sending every match. |
| 370 | const OLDER_SETTLED_TITLE_ROWS = 15; |
| 371 | const OLDER_SETTLED_MESSAGE_ROWS = 5; |
| 372 | |
| 373 | // Only the fields the palette scorer reads, so an interactive search does not decode payloads. |
| 374 | const OlderSettledTitleRow = Schema.Struct({ |
| 375 | id: ThreadId, |
| 376 | title: Schema.String, |
| 377 | branch: Schema.NullOr(Schema.String), |
| 378 | pullRequests: Schema.NullOr(Schema.fromJsonString(Schema.Array(ThreadPullRequestLink))), |
| 379 | linkedPullRequest: Schema.NullOr(Schema.fromJsonString(ThreadLinkedPullRequest)), |
| 380 | branchPullRequest: Schema.NullOr(Schema.fromJsonString(ThreadLinkedPullRequest)), |
| 381 | opens: Schema.NullOr(Schema.String), |
| 382 | createdAt: IsoDateTime, |
| 383 | projectTitle: Schema.String, |
| 384 | latestUserMessageAt: Schema.NullOr(IsoDateTime), |
| 385 | }); |
| 386 | |
| 387 | /** @public */ |
| 388 | export const make = Effect.gen(function* () { |
| 389 | const sql = yield* SqlClient.SqlClient; |
| 390 | |
| 391 | const searchRows = SqlSchema.findAll({ |
| 392 | Request: SearchRequest, |
| 393 | Result: SearchRow, |
| 394 | execute: (request) => messageHitsQuery(sql, request), |
| 395 | }); |
| 396 | |
| 397 | const threadOpens = SqlSchema.findAll({ |
| 398 | Request: Schema.Struct({ threadIds: TokenList }), |
| 399 | Result: ThreadOpensRow, |
| 400 | execute: ({ threadIds }) => sql` |
| 401 | SELECT |
| 402 | threads.thread_id AS "threadId", |
| 403 | json_extract(threads.payload_json, '$.opens') AS opens, |
| 404 | ${latestUserMessageAt(sql)} AS "latestUserMessageAt" |
| 405 | FROM orchestration_v2_projection_threads AS threads |
| 406 | WHERE threads.thread_id IN (SELECT value FROM json_each(${threadIds})) |
| 407 | `, |
| 408 | }); |
| 409 | |
| 410 | const titleCandidates = SqlSchema.findAll({ |
| 411 | Request: Schema.Struct({ projectId: Schema.NullOr(ProjectId) }), |
| 412 | Result: TitleCandidateRow, |
| 413 | execute: ({ projectId }) => sql` |
| 414 | SELECT |
| 415 | threads.payload_json AS thread, |
| 416 | projects.title AS "projectTitle", |
| 417 | ${latestUserMessageAt(sql)} AS "latestUserMessageAt" |
| 418 | FROM orchestration_v2_projection_threads AS threads |
| 419 | INNER JOIN projection_projects AS projects |
| 420 | ON projects.project_id = threads.project_id |
| 421 | WHERE threads.deleted_at IS NULL |
| 422 | AND threads.archived_at IS NULL |
| 423 | AND projects.deleted_at IS NULL |
| 424 | AND json_extract(threads.payload_json, '$.lineage.relationshipToParent') |
| 425 | IS NOT 'subagent' |
| 426 | AND json_extract(threads.payload_json, '$.lineage.sideChat') IS NOT 1 |
| 427 | AND (${projectId} IS NULL OR threads.project_id = ${projectId}) |
| 428 | `, |
| 429 | }); |
| 430 | |
| 431 | const olderSettledTitleCandidates = SqlSchema.findAll({ |
| 432 | Request: Schema.Struct({ |
| 433 | before: Schema.String, |
| 434 | projectIds: Schema.NullOr(TokenList), |
| 435 | }), |
| 436 | Result: OlderSettledTitleRow, |
| 437 | execute: ({ before, projectIds }) => sql` |
| 438 | SELECT |
| 439 | threads.thread_id AS id, |
| 440 | json_extract(threads.payload_json, '$.title') AS title, |
| 441 | json_extract(threads.payload_json, '$.branch') AS branch, |
| 442 | json_extract(threads.payload_json, '$.pullRequests') AS "pullRequests", |
| 443 | json_extract(threads.payload_json, '$.linkedPullRequest') AS "linkedPullRequest", |
| 444 | json_extract(threads.payload_json, '$.branchPullRequest') AS "branchPullRequest", |
| 445 | json_extract(threads.payload_json, '$.opens') AS opens, |
| 446 | json_extract(threads.payload_json, '$.createdAt') AS "createdAt", |
| 447 | projects.title AS "projectTitle", |
| 448 | ${latestUserMessageAt(sql)} AS "latestUserMessageAt" |
| 449 | FROM orchestration_v2_projection_threads AS threads |
| 450 | INNER JOIN projection_projects AS projects |
| 451 | ON projects.project_id = threads.project_id |
| 452 | WHERE threads.deleted_at IS NULL |
| 453 | AND threads.archived_at IS NULL |
| 454 | AND projects.deleted_at IS NULL |
| 455 | AND json_extract(threads.payload_json, '$.lineage.relationshipToParent') |
| 456 | IS NOT 'subagent' |
| 457 | AND json_extract(threads.payload_json, '$.lineage.sideChat') IS NOT 1 |
| 458 | AND json_extract(threads.payload_json, '$.settledOverride') = 'settled' |
| 459 | AND json_extract(threads.payload_json, '$.settledAt') < ${before} |
| 460 | AND (${projectIds} IS NULL |
| 461 | OR threads.project_id IN (SELECT value FROM json_each(${projectIds}))) |
| 462 | `, |
| 463 | }); |
| 464 | |
| 465 | const olderSettledIds = SqlSchema.findAll({ |
| 466 | Request: Schema.Struct({ |
| 467 | before: Schema.String, |
| 468 | threadIds: TokenList, |
| 469 | projectIds: Schema.NullOr(TokenList), |
| 470 | }), |
| 471 | Result: Schema.Struct({ id: ThreadId }), |
| 472 | execute: ({ before, threadIds, projectIds }) => sql` |
| 473 | SELECT threads.thread_id AS id |
| 474 | FROM orchestration_v2_projection_threads AS threads |
| 475 | WHERE threads.thread_id IN (SELECT value FROM json_each(${threadIds})) |
| 476 | AND threads.deleted_at IS NULL |
| 477 | AND threads.archived_at IS NULL |
| 478 | AND json_extract(threads.payload_json, '$.lineage.relationshipToParent') |
| 479 | IS NOT 'subagent' |
| 480 | AND json_extract(threads.payload_json, '$.lineage.sideChat') IS NOT 1 |
| 481 | AND json_extract(threads.payload_json, '$.settledOverride') = 'settled' |
| 482 | AND json_extract(threads.payload_json, '$.settledAt') < ${before} |
| 483 | AND (${projectIds} IS NULL |
| 484 | OR threads.project_id IN (SELECT value FROM json_each(${projectIds}))) |
| 485 | `, |
| 486 | }); |
| 487 | |
| 488 | const read = <A, E, R>(effect: Effect.Effect<A, E, R>) => |
| 489 | sql.withTransaction(effect).pipe( |
| 490 | NodeSqliteClient.readOnly, |
| 491 | Effect.mapError( |
| 492 | (cause) => |
| 493 | new ThreadSearchError({ |
| 494 | operation: Schema.isSchemaError(cause) ? "decode" : "query", |
| 495 | cause, |
| 496 | }), |
| 497 | ), |
| 498 | ); |
| 499 | |
| 500 | // A project scope applies before ranking, so other projects' stronger hits |
| 501 | // cannot use up the limit. |
| 502 | const searchMessages = Effect.fn("ThreadSearch.search")(function* ( |
| 503 | input: OrchestrationSearchThreadsInput & { |
| 504 | readonly projectId: ProjectId | null; |
| 505 | /** Threads the caller already lists, which must not use up the limit. */ |
| 506 | readonly excludedThreadIds?: ReadonlySet<ThreadId>; |
| 507 | /** Off for an agent's search, which has no use for the user's picks. */ |
| 508 | readonly picks?: boolean; |
| 509 | }, |
| 510 | ) { |
| 511 | // One message per thread: the best word match, then what the user wrote, |
| 512 | // since that is what they remember asking, then the newest by when it was |
| 513 | // written. Row order is not age: imported histories arrive in one batch. |
| 514 | // Text relevance (bm25) would cost a pass over every hit in history. |
| 515 | type Hit = { |
| 516 | readonly row: typeof SearchRow.Type; |
| 517 | /** The message as it reads, links as their text. */ |
| 518 | readonly text: string; |
| 519 | readonly quality: number; |
| 520 | readonly offset: number; |
| 521 | }; |
| 522 | const newer = (left: Hit, right: Hit) => |
| 523 | (right.row.messageCreatedAt ?? "").localeCompare(left.row.messageCreatedAt ?? ""); |
| 524 | const compareHits = (left: Hit, right: Hit) => |
| 525 | right.quality - left.quality || |
| 526 | Number(left.row.source !== "user") - Number(right.row.source !== "user") || |
| 527 | newer(left, right); |
| 528 | const rows = yield* read( |
| 529 | Effect.gen(function* () { |
| 530 | const index = yield* sql` |
| 531 | SELECT 1 FROM sqlite_master WHERE name = 'orchestration_v2_message_search' |
| 532 | `; |
| 533 | const messageSearch = buildMessageSearch(input.query, index.length > 0); |
| 534 | return messageSearch === null |
| 535 | ? [] |
| 536 | : yield* searchRows({ |
| 537 | ...messageSearch, |
| 538 | projectId: input.projectId, |
| 539 | excludedThreadIds: [...(input.excludedThreadIds ?? [])], |
| 540 | }); |
| 541 | }), |
| 542 | ); |
| 543 | const best = new Map<ThreadId, Hit>(); |
| 544 | // Matching messages per thread at its best quality: a thread that keeps |
| 545 | // using the word ranks above a passing mention, and weaker hits ("Trust" |
| 546 | // for "rust") never add to a whole word's count. |
| 547 | const hitCounts = new Map<ThreadId, number>(); |
| 548 | const hitCount = (threadId: ThreadId) => hitCounts.get(threadId) ?? 1; |
| 549 | const matchMessage = messageMatcher(input.query); |
| 550 | for (const row of rows) { |
| 551 | const text = row.text.replace(MARKDOWN_LINK, "$1"); |
| 552 | const match = matchMessage(text); |
| 553 | if (match === null) continue; |
| 554 | const hit = { row, text, ...match }; |
| 555 | const current = best.get(row.threadId); |
| 556 | if (current === undefined || hit.quality > current.quality) { |
| 557 | hitCounts.set(row.threadId, 1); |
| 558 | } else if (hit.quality === current.quality) { |
| 559 | hitCounts.set(row.threadId, hitCount(row.threadId) + 1); |
| 560 | } |
| 561 | if (current === undefined || compareHits(hit, current) < 0) best.set(row.threadId, hit); |
| 562 | } |
| 563 | // Ranked as clients rank them, with picks and use, before the limit cuts: |
| 564 | // a thread picked for this query must come back even when it only |
| 565 | // matches in a few old messages. |
| 566 | const opens = best.size === 0 ? [] : yield* read(threadOpens({ threadIds: [...best.keys()] })); |
| 567 | const subjectById = new Map( |
| 568 | opens.map((row) => [ |
| 569 | row.threadId, |
| 570 | { |
| 571 | title: "", |
| 572 | activityAt: row.latestUserMessageAt === null ? null : Date.parse(row.latestUserMessageAt), |
| 573 | opens: Option.getOrNull(decodeThreadOpens(row.opens)), |
| 574 | }, |
| 575 | ]), |
| 576 | ); |
| 577 | const now = DateTime.toEpochMillis(yield* DateTime.now); |
| 578 | const ranked = rankMessageMatches( |
| 579 | [...best.values()].map((hit) => ({ |
| 580 | ...hit, |
| 581 | hits: hitCount(hit.row.threadId), |
| 582 | messageCreatedAt: hit.row.messageCreatedAt, |
| 583 | })), |
| 584 | (hit) => subjectById.get(hit.row.threadId), |
| 585 | foldSearchQuery(input.query), |
| 586 | now, |
| 587 | { picks: input.picks ?? true }, |
| 588 | ); |
| 589 | return { |
| 590 | matches: ranked.slice(0, input.limit ?? 50).map(({ row, text, quality, offset }) => ({ |
| 591 | hits: hitCount(row.threadId), |
| 592 | threadId: row.threadId, |
| 593 | messageId: row.messageId, |
| 594 | projectId: row.projectId, |
| 595 | source: row.source, |
| 596 | snippet: buildSearchSnippet(text, offset), |
| 597 | messageCreatedAt: row.messageCreatedAt, |
| 598 | quality, |
| 599 | })), |
| 600 | }; |
| 601 | }); |
| 602 | |
| 603 | const search: ThreadSearch["Service"]["search"] = (input) => |
| 604 | searchMessages({ ...input, projectId: null }); |
| 605 | |
| 606 | const searchThreads: ThreadSearch["Service"]["searchThreads"] = Effect.fn( |
| 607 | "ThreadSearch.searchThreads", |
| 608 | )(function* (input) { |
| 609 | const limit = input.limit ?? 50; |
| 610 | const normalizedQuery = foldSearchQuery(input.query); |
| 611 | const now = DateTime.toEpochMillis(yield* DateTime.now); |
| 612 | // Agents call this rarely, and a project has hundreds of threads, not |
| 613 | // millions, so ranking every active thread in memory stays cheap. |
| 614 | const candidates = yield* read(titleCandidates({ projectId: input.projectId ?? null })); |
| 615 | const subjectById = new Map( |
| 616 | candidates.map((row): [ThreadId, ThreadSearchSubject] => [ |
| 617 | row.thread.id, |
| 618 | { |
| 619 | id: row.thread.id, |
| 620 | title: row.thread.title, |
| 621 | branch: row.thread.branch, |
| 622 | projectTitle: row.projectTitle, |
| 623 | pullRequests: threadSearchPullRequests(row.thread), |
| 624 | // Like the clients: what the user last wrote, else when the thread began. |
| 625 | activityAt: Date.parse( |
| 626 | row.latestUserMessageAt ?? DateTime.formatIso(row.thread.createdAt), |
| 627 | ), |
| 628 | opens: row.thread.opens ?? null, |
| 629 | }, |
| 630 | ]), |
| 631 | ); |
| 632 | const titleHits: ThreadSearchHit[] = rankSearchItems(candidates, (row, options) => |
| 633 | // An agent's search has no use for the user's palette picks. |
| 634 | scoreThreadSearch(normalizedQuery, subjectById.get(row.thread.id)!, now, { |
| 635 | ...options, |
| 636 | picks: false, |
| 637 | }), |
| 638 | ) |
| 639 | .slice(0, limit) |
| 640 | .map(({ item }) => ({ |
| 641 | threadId: item.thread.id, |
| 642 | projectId: item.thread.projectId, |
| 643 | title: item.thread.title, |
| 644 | source: "title", |
| 645 | })); |
| 646 | const shown = new Set(titleHits.map((hit) => hit.threadId)); |
| 647 | const { matches } = yield* searchMessages({ |
| 648 | query: input.query, |
| 649 | limit: 50, |
| 650 | projectId: input.projectId ?? null, |
| 651 | excludedThreadIds: shown, |
| 652 | picks: false, |
| 653 | }); |
| 654 | const messageHits = matches |
| 655 | .filter((match) => subjectById.has(match.threadId)) |
| 656 | .map(({ quality: _quality, hits: _hits, messageId: _messageId, ...match }) => ({ |
| 657 | ...match, |
| 658 | title: subjectById.get(match.threadId)!.title, |
| 659 | })); |
| 660 | // Titles come first, but a fifth of the slots stay for message matches, |
| 661 | // as the palette keeps its "In messages" section. |
| 662 | const messageSlots = Math.min(messageHits.length, Math.floor(limit / 5)); |
| 663 | return { |
| 664 | matches: [...titleHits.slice(0, limit - messageSlots), ...messageHits].slice(0, limit), |
| 665 | }; |
| 666 | }); |
| 667 | |
| 668 | const searchOlderSettled: ThreadSearch["Service"]["searchOlderSettled"] = Effect.fn( |
| 669 | "ThreadSearch.searchOlderSettled", |
| 670 | )(function* ({ query, before, projectIds, environmentLabel, messageThreadIds }) { |
| 671 | const normalizedQuery = foldSearchQuery(query); |
| 672 | const now = DateTime.toEpochMillis(yield* DateTime.now); |
| 673 | const candidates = yield* read( |
| 674 | olderSettledTitleCandidates({ before, projectIds: projectIds ?? null }), |
| 675 | ); |
| 676 | const subjectOf = (row: typeof OlderSettledTitleRow.Type): ThreadSearchSubject => ({ |
| 677 | id: row.id, |
| 678 | title: row.title, |
| 679 | branch: row.branch, |
| 680 | projectTitle: row.projectTitle, |
| 681 | environmentLabel: environmentLabel ?? null, |
| 682 | pullRequests: threadSearchPullRequests({ |
| 683 | pullRequests: row.pullRequests ?? undefined, |
| 684 | linkedPullRequest: row.linkedPullRequest, |
| 685 | branchPullRequest: row.branchPullRequest, |
| 686 | }), |
| 687 | // Like the clients: what the user last wrote, else when the thread began. |
| 688 | activityAt: Date.parse(row.latestUserMessageAt ?? row.createdAt), |
| 689 | opens: Option.getOrNull(decodeThreadOpens(row.opens)), |
| 690 | }); |
| 691 | // Picks on, as in the command palette: a thread picked for this query must make the cut. |
| 692 | const titleIds = rankSearchItems(candidates, (row, options) => |
| 693 | scoreThreadSearch(normalizedQuery, subjectOf(row), now, { ...options, picks: true }), |
| 694 | ) |
| 695 | .slice(0, OLDER_SETTLED_TITLE_ROWS) |
| 696 | .map(({ item }) => item.id); |
| 697 | const shown = new Set(titleIds); |
| 698 | const messageCandidates = [...new Set(messageThreadIds)].filter((id) => !shown.has(id)); |
| 699 | const older = |
| 700 | messageCandidates.length === 0 |
| 701 | ? new Set<ThreadId>() |
| 702 | : new Set( |
| 703 | (yield* read( |
| 704 | olderSettledIds({ |
| 705 | before, |
| 706 | threadIds: messageCandidates, |
| 707 | projectIds: projectIds ?? null, |
| 708 | }), |
| 709 | )).map((row) => row.id), |
| 710 | ); |
| 711 | return [ |
| 712 | ...titleIds, |
| 713 | ...messageCandidates.filter((id) => older.has(id)).slice(0, OLDER_SETTLED_MESSAGE_ROWS), |
| 714 | ]; |
| 715 | }); |
| 716 | |
| 717 | return ThreadSearch.of({ search, searchThreads, searchOlderSettled }); |
| 718 | }); |
| 719 | |
| 720 | export const layer = Layer.effect(ThreadSearch, make); |
| 721 | |
| 722 | /** |
| 723 | * The client search: message hits, and with `olderSettledBefore`, the rows of the best older |
| 724 | * settled matches, which shells that omit them do not hold. |
| 725 | */ |
| 726 | export const searchThreadsWithOlderSettledRows = Effect.fn( |
| 727 | "ThreadSearch.searchThreadsWithOlderSettledRows", |
| 728 | )(function* <E1, E2, E3, E4, E5>(input: { |
| 729 | readonly request: OrchestrationV2SearchThreadsInput; |
| 730 | readonly search: ( |
| 731 | input: OrchestrationSearchThreadsInput, |
| 732 | ) => Effect.Effect<OrchestrationSearchThreadsResult, E1>; |
| 733 | readonly searchOlderSettled: ( |
| 734 | input: Parameters<ThreadSearch["Service"]["searchOlderSettled"]>[0], |
| 735 | ) => Effect.Effect<ReadonlyArray<ThreadId>, E2>; |
| 736 | readonly sql: SqlClient.SqlClient; |
| 737 | readonly readThreads: ( |
| 738 | threadIds: ReadonlyArray<ThreadId>, |
| 739 | ) => Effect.Effect<Effect.Effect<OrchestrationV2ThreadShellSnapshot, E4>, E3>; |
| 740 | readonly latestSequence: Effect.Effect<number, E5>; |
| 741 | }) { |
| 742 | const { olderSettledBefore, projectIds, environmentLabel, titlesOnly, ...request } = |
| 743 | input.request; |
| 744 | const messages: OrchestrationV2SearchThreadsResult = |
| 745 | titlesOnly === true ? { matches: [] } : yield* input.search(request); |
| 746 | if (olderSettledBefore === undefined) return messages; |
| 747 | // Ranked before the rows' transaction, so scoring never holds the shared reader. |
| 748 | const threadIds = yield* input.searchOlderSettled({ |
| 749 | query: request.query, |
| 750 | before: olderSettledBefore, |
| 751 | projectIds, |
| 752 | environmentLabel, |
| 753 | messageThreadIds: messages.matches.map((match) => match.threadId), |
| 754 | }); |
| 755 | const older = yield* loadOlderSettledThreadsPage({ |
| 756 | sql: input.sql, |
| 757 | listThreads: Effect.succeed({ threadIds, counts: [] }), |
| 758 | readThreads: input.readThreads, |
| 759 | latestSequence: input.latestSequence, |
| 760 | }); |
| 761 | return { ...messages, snapshotSequence: older.snapshotSequence, threads: older.threads }; |
| 762 | }); |