CodeIssuesDiscussionsWikiPull RequestsProjectsCommitsActionsReleasesContributorsPulse● GatesSecuritySettingsDeploymentsPipelineInsightsAgents✨ Explain✨ Ask AI✨ Workspace✨ Spec✨ Tests▓ Debt Map✨ NL Search🏛 Archaeology
Blame · Line-by-line history
ci-autofix.ts
Each line is annotated with the commit that last touched it. Click any SHA to jump to that commit and see the surrounding change.
| 34e63b9 | 1 | /** |
| bc519ae | 2 | * CI Auto-Fix — when a gate run or workflow run fails on a PR, the repair |
| 3 | * flywheel cache is consulted first (Tier 0); on a miss Claude reads the | |
| 4 | * error logs, the failing test file, and the PR diff. Either way a | |
| 5 | * ready-to-apply patch is posted as a comment on the PR. | |
| 34e63b9 | 6 | * |
| 7 | * Entry points: | |
| 8 | * triggerCiAutofix(gateRunId) — fire-and-forget; call after a gate_run | |
| 9 | * row is written with status="failed". | |
| 10 | * applyAutofix(prCommentId, userId) — apply the patch from a comment onto | |
| bc519ae | 11 | * a new branch and return the branch name. Settles the flywheel entry |
| 12 | * (success/failed) so the cache's confidence accumulates. | |
| 34e63b9 | 13 | * |
| 14 | * Route wiring (src/routes/pulls.tsx or src/routes/api.ts): | |
| 15 | * POST /api/pr-comments/:commentId/apply-autofix → applyAutofix | |
| bc519ae | 16 | * |
| 17 | * Flywheel wiring (BUILD_BIBLE §7 finding 1): every failure is fingerprinted | |
| 18 | * via repair-flywheel.ts; a previously-successful patch with the same | |
| 19 | * signature and a good success rate is served WITHOUT an AI call. All served | |
| 20 | * fixes are recorded as 'pending' flywheel rows and settled on apply. | |
| 21 | * Flywheel/DB errors never break this path — a broken cache degrades to the | |
| 22 | * old always-call-AI behaviour. | |
| 34e63b9 | 23 | */ |
| 24 | ||
| 25 | import { and, eq } from "drizzle-orm"; | |
| 26 | import { mkdtemp, rm, writeFile } from "fs/promises"; | |
| 27 | import { join } from "path"; | |
| 28 | import { tmpdir } from "os"; | |
| 29 | import { db } from "../db"; | |
| bc519ae | 30 | import { |
| 31 | findCachedRepair, | |
| 32 | recordRepair, | |
| 33 | updateOutcome, | |
| 34 | type CachedRepair, | |
| 35 | type RecordRepairInput, | |
| 36 | } from "./repair-flywheel"; | |
| 34e63b9 | 37 | import { |
| 38 | gateRuns, | |
| 39 | pullRequests, | |
| 40 | prComments, | |
| 41 | repositories, | |
| 42 | users, | |
| 43 | repoCollaborators, | |
| 44 | } from "../db/schema"; | |
| 45 | import { getRepoPath } from "../git/repository"; | |
| 46 | import { getBotUserIdOrFallback } from "./bot-user"; | |
| 47 | import { | |
| 48 | getAnthropic, | |
| 49 | isAiAvailable, | |
| 50 | MODEL_SONNET, | |
| 51 | extractText, | |
| 52 | parseJsonResponse, | |
| 53 | } from "./ai-client"; | |
| 54 | ||
| 55 | // --------------------------------------------------------------------------- | |
| 56 | // Types | |
| 57 | // --------------------------------------------------------------------------- | |
| 58 | ||
| 59 | export interface AutofixResult { | |
| 60 | prNumber: number; | |
| 61 | repoId: string; | |
| 62 | gateRunId: string; | |
| 63 | patch: string; // unified diff format | |
| 64 | explanation: string; // 2-3 sentence explanation | |
| 65 | confidence: "high" | "medium" | "low"; | |
| 66 | affectedFiles: string[]; | |
| 67 | } | |
| 68 | ||
| bc519ae | 69 | export interface ClaudeAutofixResponse { |
| 34e63b9 | 70 | patch: string; |
| 71 | explanation: string; | |
| 72 | confidence: "high" | "medium" | "low"; | |
| 73 | affectedFiles: string[]; | |
| 74 | } | |
| 75 | ||
| bc519ae | 76 | /** |
| 77 | * DI seam for the repair-flywheel wiring. Production callers omit this and | |
| 78 | * get the real implementations; tests inject fakes so no DB is touched. | |
| 79 | */ | |
| 80 | export interface CiAutofixDeps { | |
| 81 | findCachedRepair?: typeof findCachedRepair; | |
| 82 | recordRepair?: typeof recordRepair; | |
| 83 | updateOutcome?: typeof updateOutcome; | |
| 84 | aiAvailable?: () => boolean; | |
| 85 | } | |
| 86 | ||
| 87 | /** The fix triggerCiAutofix decided to post, plus its flywheel bookkeeping. */ | |
| 88 | export interface AutofixPlan { | |
| 89 | /** 'cache' = Tier-0 flywheel replay (no AI call); 'ai' = fresh Sonnet patch. */ | |
| 90 | source: "cache" | "ai"; | |
| 91 | fix: ClaudeAutofixResponse; | |
| 92 | /** Flywheel row recorded as 'pending' for this attempt; null if recording failed. */ | |
| 93 | flywheelEntryId: string | null; | |
| 94 | /** On a cache hit: the parent pattern that was replayed. */ | |
| 95 | cachedPatternId: string | null; | |
| 96 | } | |
| 97 | ||
| 34e63b9 | 98 | // --------------------------------------------------------------------------- |
| 99 | // Constants | |
| 100 | // --------------------------------------------------------------------------- | |
| 101 | ||
| 102 | /** Idempotency marker embedded in every autofix comment. */ | |
| 103 | export const CI_AUTOFIX_MARKER = "<!-- gluecron:ci-autofix:v1 -->"; | |
| 104 | ||
| bc519ae | 105 | /** |
| 106 | * Marker carrying the pending flywheel entry id, embedded in the autofix | |
| 107 | * comment so applyAutofix can settle the outcome (success/failed) later. | |
| 108 | */ | |
| 109 | export const FLYWHEEL_MARKER_PREFIX = "<!-- gluecron:ci-autofix:flywheel:"; | |
| 110 | ||
| 111 | /** | |
| 112 | * Minimum settled success rate before a cached pattern is replayed instead | |
| 113 | * of calling the AI. Below this the cache is considered unreliable for the | |
| 114 | * signature and we fall through to a fresh Sonnet patch. | |
| 115 | */ | |
| 116 | export const CACHE_MIN_SUCCESS_RATE = 0.5; | |
| 117 | ||
| 34e63b9 | 118 | /** Max bytes of PR diff sent to Claude. */ |
| 119 | const MAX_DIFF_BYTES = 80 * 1024; | |
| 120 | ||
| 121 | /** Max bytes of error log sent to Claude. */ | |
| 122 | const MAX_LOG_BYTES = 3 * 1024; | |
| 123 | ||
| 124 | /** Max bytes per test file read. */ | |
| 125 | const MAX_FILE_BYTES = 10 * 1024; | |
| 126 | ||
| 127 | /** Max number of failing test files to read. */ | |
| 128 | const MAX_TEST_FILES = 3; | |
| 129 | ||
| 130 | // --------------------------------------------------------------------------- | |
| 131 | // Helpers | |
| 132 | // --------------------------------------------------------------------------- | |
| 133 | ||
| 134 | async function spawnGit( | |
| 135 | args: string[], | |
| 136 | cwd: string | |
| 137 | ): Promise<{ stdout: string; stderr: string; exitCode: number }> { | |
| 138 | const proc = Bun.spawn(["git", ...args], { | |
| 139 | cwd, | |
| 140 | stdout: "pipe", | |
| 141 | stderr: "pipe", | |
| 142 | }); | |
| 143 | const [stdout, stderr] = await Promise.all([ | |
| 144 | new Response(proc.stdout).text(), | |
| 145 | new Response(proc.stderr).text(), | |
| 146 | ]); | |
| 147 | const exitCode = await proc.exited; | |
| 148 | return { stdout, stderr, exitCode }; | |
| 149 | } | |
| 150 | ||
| 151 | function truncate(s: string, maxBytes: number): string { | |
| 152 | const buf = Buffer.from(s, "utf8"); | |
| 153 | if (buf.length <= maxBytes) return s; | |
| 154 | return buf.slice(0, maxBytes).toString("utf8") + "\n[truncated]"; | |
| 155 | } | |
| 156 | ||
| 157 | /** | |
| 158 | * Extract failing test file paths, error message, and stack trace from a | |
| 159 | * raw CI error log string. Heuristic — good enough for most JS/TS test | |
| 160 | * runners (Jest, Vitest, Bun test) and Python pytest output. | |
| 161 | */ | |
| 162 | function parseErrorLog(errorLog: string): { | |
| 163 | testFiles: string[]; | |
| 164 | errorSummary: string; | |
| 165 | } { | |
| 166 | const lines = errorLog.split("\n"); | |
| 167 | ||
| 168 | // Collect candidate test file paths: | |
| 169 | // - lines mentioning .test.ts/.spec.ts/.test.js/.spec.js/.test.py paths | |
| 170 | // - lines with "FAIL <path>" (Jest pattern) | |
| 171 | const filePatterns = [ | |
| 172 | /(?:^|\s)([\w./\-]+\.(?:test|spec)\.[jt]sx?)/gm, | |
| 173 | /(?:^|\s)([\w./\-]+_test\.py)/gm, | |
| 174 | /(?:^FAIL\s+)([\w./\-]+)/gm, | |
| 175 | ]; | |
| 176 | ||
| 177 | const filesSet = new Set<string>(); | |
| 178 | for (const pattern of filePatterns) { | |
| 179 | let m: RegExpExecArray | null; | |
| 180 | pattern.lastIndex = 0; | |
| 181 | while ((m = pattern.exec(errorLog)) !== null) { | |
| 182 | const path = m[1].trim(); | |
| 183 | if (path && !path.startsWith("-") && !path.startsWith("+")) { | |
| 184 | filesSet.add(path); | |
| 185 | } | |
| 186 | } | |
| 187 | } | |
| 188 | ||
| 189 | // Error summary: first 3KB of the log (most runners put the error first). | |
| 190 | const errorSummary = truncate(errorLog, MAX_LOG_BYTES); | |
| 191 | ||
| 192 | return { | |
| 193 | testFiles: Array.from(filesSet).slice(0, MAX_TEST_FILES), | |
| 194 | errorSummary, | |
| 195 | }; | |
| 196 | } | |
| 197 | ||
| bc519ae | 198 | // --------------------------------------------------------------------------- |
| 199 | // Repair flywheel wiring (Tier 0) | |
| 200 | // --------------------------------------------------------------------------- | |
| 201 | ||
| 202 | /** Matches the id inside a FLYWHEEL_MARKER_PREFIX comment marker. */ | |
| 203 | const FLYWHEEL_MARKER_RE = /<!-- gluecron:ci-autofix:flywheel:([0-9a-fA-F-]{8,64}) -->/; | |
| 204 | ||
| 205 | /** Parse the flywheel entry id out of an autofix comment body, if present. */ | |
| 206 | export function extractFlywheelEntryId(commentBody: string): string | null { | |
| 207 | const m = commentBody.match(FLYWHEEL_MARKER_RE); | |
| 208 | return m ? m[1] : null; | |
| 209 | } | |
| 210 | ||
| 211 | /** | |
| 212 | * Record a 'pending' flywheel row for a fix we're about to post. Never | |
| 213 | * throws — a flywheel write failure must not stop the fix from shipping, | |
| 214 | * it only means this attempt won't contribute to the cache's learning. | |
| 215 | */ | |
| 216 | async function safeRecordRepair( | |
| 217 | input: RecordRepairInput, | |
| 218 | deps: CiAutofixDeps | |
| 219 | ): Promise<string | null> { | |
| 220 | const record = deps.recordRepair ?? recordRepair; | |
| 221 | try { | |
| 222 | return await record(input); | |
| 223 | } catch (err) { | |
| 224 | console.warn( | |
| 225 | "[ci-autofix] flywheel record failed (fix still served):", | |
| 226 | err instanceof Error ? err.message : err | |
| 227 | ); | |
| 228 | return null; | |
| 229 | } | |
| 230 | } | |
| 231 | ||
| 232 | /** | |
| 233 | * Settle the flywheel entry referenced by an autofix comment (if any) so | |
| 234 | * the pattern's success rate accumulates. Never throws — flywheel | |
| 235 | * bookkeeping must not break the apply path. | |
| 236 | */ | |
| 237 | export async function recordAutofixOutcome( | |
| 238 | commentBody: string, | |
| 239 | outcome: "success" | "failed", | |
| 240 | deps: CiAutofixDeps = {} | |
| 241 | ): Promise<void> { | |
| 242 | const entryId = extractFlywheelEntryId(commentBody); | |
| 243 | if (!entryId) return; | |
| 244 | const settle = deps.updateOutcome ?? updateOutcome; | |
| 245 | try { | |
| 246 | await settle(entryId, outcome); | |
| 247 | } catch (err) { | |
| 248 | console.warn( | |
| 249 | "[ci-autofix] flywheel outcome update failed:", | |
| 250 | err instanceof Error ? err.message : err | |
| 251 | ); | |
| 252 | } | |
| 253 | } | |
| 254 | ||
| 255 | /** | |
| 256 | * Tier-0-then-AI resolution for a CI failure (BUILD_BIBLE §7 finding 1). | |
| 257 | * | |
| 258 | * 1. Tier 0 — consult the repair flywheel for a previously-successful fix | |
| 259 | * with the same failure signature. On a usable hit (success rate ≥ | |
| 260 | * CACHE_MIN_SUCCESS_RATE and a stored patch) the cached patch is | |
| 261 | * served directly: no Anthropic call at all. | |
| 262 | * 2. Tier 2 — fall through to a fresh Claude Sonnet patch via the | |
| 263 | * `generateAiFix` thunk (guarded by isAiAvailable(); the thunk also | |
| 264 | * keeps the expensive git context-gathering off the cache-hit path). | |
| 265 | * | |
| 266 | * Every served fix is recorded as a 'pending' flywheel row; applyAutofix | |
| 267 | * settles it via recordAutofixOutcome. Flywheel/DB failures NEVER break | |
| 268 | * this path — a broken cache degrades to the old always-call-AI behaviour. | |
| 269 | */ | |
| 270 | export async function resolveAutofix(args: { | |
| 271 | repositoryId: string; | |
| 272 | failureText: string; | |
| 273 | generateAiFix: () => Promise<ClaudeAutofixResponse | null>; | |
| 274 | deps?: CiAutofixDeps; | |
| 275 | }): Promise<AutofixPlan | null> { | |
| 276 | const deps = args.deps ?? {}; | |
| 277 | const lookup = deps.findCachedRepair ?? findCachedRepair; | |
| 278 | const aiOk = deps.aiAvailable ?? isAiAvailable; | |
| 279 | ||
| 280 | // ── Tier 0: flywheel cache. Fail open — any error counts as a miss. | |
| 281 | let cached: CachedRepair | null = null; | |
| 282 | if (args.failureText.trim()) { | |
| 283 | try { | |
| 284 | cached = await lookup(args.repositoryId, args.failureText); | |
| 285 | } catch (err) { | |
| 286 | console.warn( | |
| 287 | "[ci-autofix] flywheel lookup failed (falling through to AI):", | |
| 288 | err instanceof Error ? err.message : err | |
| 289 | ); | |
| 290 | } | |
| 291 | } | |
| 292 | ||
| 293 | // A usable hit needs a stored full patch (pre-0105 rows have none) AND a | |
| 294 | // good settled success rate; anything else falls through to the AI tier. | |
| 295 | if ( | |
| 296 | cached && | |
| 297 | cached.patch?.trim() && | |
| 298 | cached.successRate >= CACHE_MIN_SUCCESS_RATE | |
| 299 | ) { | |
| 300 | const fix: ClaudeAutofixResponse = { | |
| 301 | patch: cached.patch, | |
| 302 | explanation: | |
| 303 | cached.patchSummary || | |
| 304 | "Replayed a previously-successful repair for this failure signature.", | |
| 305 | // A pattern that keeps working is high confidence; anything that has | |
| 306 | // failed at least occasionally is medium. | |
| 307 | confidence: cached.successRate >= 0.9 ? "high" : "medium", | |
| 308 | affectedFiles: cached.filesChanged, | |
| 309 | }; | |
| 310 | const entryId = await safeRecordRepair( | |
| 311 | { | |
| 312 | repositoryId: args.repositoryId, | |
| 313 | failureText: args.failureText, | |
| 314 | classification: cached.classification, | |
| 315 | tier: "cached", | |
| 316 | patchSummary: fix.explanation, | |
| 317 | // Carry the patch onto the new row so it is itself replayable once | |
| 318 | // it settles (findCachedRepair prefers the most recent success). | |
| 319 | patch: cached.patch, | |
| 320 | filesChanged: fix.affectedFiles, | |
| 321 | commitSha: null, | |
| 322 | parentPatternId: cached.id, | |
| 323 | }, | |
| 324 | deps | |
| 325 | ); | |
| 326 | // Audit the saved AI call — this line is the flywheel's whole point. | |
| 327 | console.log( | |
| 328 | `[ci-autofix] flywheel cache HIT — pattern ${cached.id} (success rate ${(cached.successRate * 100).toFixed(0)}%, ${cached.hitCount} prior hits) served without an AI call` | |
| 329 | ); | |
| 330 | return { source: "cache", fix, flywheelEntryId: entryId, cachedPatternId: cached.id }; | |
| 331 | } | |
| 332 | ||
| 333 | // ── Tier 2: fresh AI patch. All AI features degrade gracefully when no | |
| 334 | // API key is configured — cached fixes above still work without one. | |
| 335 | if (!aiOk()) return null; | |
| 336 | ||
| 337 | const fix = await args.generateAiFix(); | |
| 338 | if (!fix || !fix.patch || !fix.explanation) return null; | |
| 339 | if (fix.confidence === "low") return null; | |
| 340 | ||
| 341 | const entryId = await safeRecordRepair( | |
| 342 | { | |
| 343 | repositoryId: args.repositoryId, | |
| 344 | failureText: args.failureText, | |
| 345 | classification: null, | |
| 346 | tier: "ai-sonnet", | |
| 347 | patchSummary: fix.explanation, | |
| 348 | // Stored so the entry is replayable by the Tier-0 cache once it | |
| 349 | // settles to 'success'. | |
| 350 | patch: fix.patch, | |
| 351 | filesChanged: fix.affectedFiles, | |
| 352 | commitSha: null, | |
| 353 | }, | |
| 354 | deps | |
| 355 | ); | |
| 356 | return { source: "ai", fix, flywheelEntryId: entryId, cachedPatternId: null }; | |
| 357 | } | |
| 358 | ||
| 34e63b9 | 359 | // --------------------------------------------------------------------------- |
| 360 | // Main entry point | |
| 361 | // --------------------------------------------------------------------------- | |
| 362 | ||
| 363 | /** | |
| 364 | * Fire-and-forget entry point called after a gate_run row is set to 'failed'. | |
| 365 | * Never throws — all errors are swallowed after logging. | |
| bc519ae | 366 | * |
| 367 | * Note: no isAiAvailable() gate here — the Tier-0 flywheel cache needs no | |
| 368 | * API key, so cached fixes still flow when AI is unconfigured. The actual | |
| 369 | * Anthropic call is guarded inside resolveAutofix (graceful degradation). | |
| 34e63b9 | 370 | */ |
| bc519ae | 371 | export async function triggerCiAutofix( |
| 372 | gateRunId: string, | |
| 373 | deps: CiAutofixDeps = {} | |
| 374 | ): Promise<void> { | |
| 34e63b9 | 375 | try { |
| bc519ae | 376 | await _runAutofix(gateRunId, deps); |
| 34e63b9 | 377 | } catch (err) { |
| 378 | console.error( | |
| 379 | "[ci-autofix] crashed:", | |
| 380 | err instanceof Error ? err.message : err | |
| 381 | ); | |
| 382 | } | |
| 383 | } | |
| 384 | ||
| bc519ae | 385 | async function _runAutofix( |
| 386 | gateRunId: string, | |
| 387 | deps: CiAutofixDeps = {} | |
| 388 | ): Promise<void> { | |
| 34e63b9 | 389 | // 1. Load the gate run |
| 390 | const [gateRun] = await db | |
| 391 | .select() | |
| 392 | .from(gateRuns) | |
| 393 | .where(eq(gateRuns.id, gateRunId)) | |
| 394 | .limit(1); | |
| 395 | ||
| 396 | if (!gateRun) return; | |
| 397 | if (gateRun.status !== "failed") return; | |
| 398 | if (!gateRun.pullRequestId) return; | |
| 399 | ||
| 400 | // 2. Load the PR row | |
| 401 | const [pr] = await db | |
| 402 | .select() | |
| 403 | .from(pullRequests) | |
| 404 | .where(eq(pullRequests.id, gateRun.pullRequestId)) | |
| 405 | .limit(1); | |
| 406 | ||
| 407 | if (!pr) return; | |
| 408 | ||
| 409 | // 3. Load repo (owner/name) | |
| 410 | const [repoRow] = await db | |
| 411 | .select({ | |
| 412 | id: repositories.id, | |
| 413 | name: repositories.name, | |
| 414 | diskPath: repositories.diskPath, | |
| 415 | ownerUsername: users.username, | |
| 416 | }) | |
| 417 | .from(repositories) | |
| 418 | .innerJoin(users, eq(repositories.ownerId, users.id)) | |
| 419 | .where(eq(repositories.id, gateRun.repositoryId)) | |
| 420 | .limit(1); | |
| 421 | ||
| 422 | if (!repoRow) return; | |
| 423 | ||
| 424 | // 4. Check idempotency — skip if already posted for this gateRunId | |
| 425 | const idempotencyMarker = `<!-- gluecron:ci-autofix:run:${gateRunId} -->`; | |
| 426 | const existing = await db | |
| 427 | .select({ id: prComments.id }) | |
| 428 | .from(prComments) | |
| 429 | .where( | |
| 430 | and( | |
| 431 | eq(prComments.pullRequestId, gateRun.pullRequestId), | |
| 432 | // We check by looking for comments with the autofix marker. | |
| 433 | // drizzle doesn't have LIKE with dynamic params easily, but | |
| 434 | // we query all AI comments and filter client-side (there won't be many). | |
| 435 | eq(prComments.isAiReview, true) | |
| 436 | ) | |
| 437 | ) | |
| 438 | .limit(50); | |
| 439 | ||
| 440 | for (const row of existing) { | |
| 441 | // Load the body to check idempotency marker | |
| 442 | const [full] = await db | |
| 443 | .select({ body: prComments.body }) | |
| 444 | .from(prComments) | |
| 445 | .where(eq(prComments.id, row.id)) | |
| 446 | .limit(1); | |
| 447 | if (full?.body?.includes(idempotencyMarker)) return; | |
| 448 | } | |
| 449 | ||
| 450 | const repoDir = getRepoPath(repoRow.ownerUsername, repoRow.name); | |
| 451 | ||
| bc519ae | 452 | // 5. Failure text — drives both the flywheel signature and the AI prompt. |
| 34e63b9 | 453 | const errorLog = gateRun.summary || gateRun.details || ""; |
| bc519ae | 454 | const failureText = |
| 455 | typeof errorLog === "string" ? errorLog : JSON.stringify(errorLog); | |
| 456 | ||
| 457 | // 6. Tier 0 (flywheel cache) first, then the AI fallback. The expensive | |
| 458 | // context gathering (PR diff + failing test files + Sonnet call) lives | |
| 459 | // inside the thunk so a cache hit never touches git or the API. | |
| 460 | const plan = await resolveAutofix({ | |
| 461 | repositoryId: repoRow.id, | |
| 462 | failureText, | |
| 463 | deps, | |
| 464 | generateAiFix: async () => { | |
| 465 | // 6a. Get the PR diff (max 80KB) | |
| 466 | const diffResult = await spawnGit( | |
| 467 | ["diff", `${pr.baseBranch}...${pr.headBranch}`], | |
| 468 | repoDir | |
| 469 | ); | |
| 470 | const prDiff = truncate(diffResult.stdout, MAX_DIFF_BYTES); | |
| 471 | ||
| 472 | if (!prDiff.trim()) return null; // nothing to work with | |
| 473 | ||
| 474 | // 6b. Parse errorLog to extract test files + error summary | |
| 475 | const { testFiles, errorSummary } = parseErrorLog(failureText); | |
| 476 | ||
| 477 | // 6c. Read failing test files via git show HEAD:path | |
| 478 | let testFileContent = ""; | |
| 479 | for (const filePath of testFiles) { | |
| 480 | const showResult = await spawnGit( | |
| 481 | ["show", `${pr.headBranch}:${filePath}`], | |
| 482 | repoDir | |
| 483 | ); | |
| 484 | if (showResult.exitCode === 0 && showResult.stdout) { | |
| 485 | const content = truncate(showResult.stdout, MAX_FILE_BYTES); | |
| 486 | testFileContent += `\n\n--- ${filePath} ---\n${content}`; | |
| 487 | } | |
| 488 | } | |
| 34e63b9 | 489 | |
| bc519ae | 490 | // 6d. Call Claude Sonnet 4.6 |
| 491 | const client = getAnthropic(); | |
| 492 | const prompt = `You are a senior engineer fixing a CI failure. | |
| 34e63b9 | 493 | |
| 494 | PR diff (what changed): | |
| 495 | ${prDiff} | |
| 496 | ||
| 497 | Failing test output: | |
| 498 | ${errorSummary} | |
| 499 | ||
| 500 | Test file content:${testFileContent || "\n(no test files detected)"} | |
| 501 | ||
| 502 | Produce a minimal unified diff patch that fixes the CI failure. The patch must: | |
| 503 | 1. Be valid unified diff format (--- a/file, +++ b/file, @@ lines) | |
| 504 | 2. Fix only what's needed — no refactoring | |
| 505 | 3. Not modify the test itself unless the test expectation is genuinely wrong | |
| 506 | ||
| 507 | Return JSON: {"patch": "...", "explanation": "...", "confidence": "high|medium|low", "affectedFiles": ["..."]}`; | |
| 508 | ||
| bc519ae | 509 | const message = await client.messages.create({ |
| 510 | model: MODEL_SONNET, | |
| 511 | max_tokens: 4096, | |
| 512 | messages: [{ role: "user", content: prompt }], | |
| 513 | }); | |
| 34e63b9 | 514 | |
| bc519ae | 515 | const rawText = extractText(message); |
| 516 | return parseJsonResponse<ClaudeAutofixResponse>(rawText); | |
| 517 | }, | |
| 518 | }); | |
| 34e63b9 | 519 | |
| bc519ae | 520 | // No usable fix (cache miss + AI declined/low-confidence/unavailable). |
| 521 | if (!plan) return; | |
| 34e63b9 | 522 | |
| bc519ae | 523 | // 7. Build and post the comment |
| 524 | const commentBody = buildAutofixComment(plan, idempotencyMarker, gateRunId); | |
| 34e63b9 | 525 | |
| 526 | const botAuthorId = await getBotUserIdOrFallback(repoRow.id); | |
| 527 | if (!botAuthorId) return; | |
| 528 | ||
| 529 | await db.insert(prComments).values({ | |
| 530 | pullRequestId: gateRun.pullRequestId, | |
| 531 | authorId: botAuthorId, | |
| 532 | body: commentBody, | |
| 533 | isAiReview: true, | |
| 534 | }); | |
| 535 | } | |
| 536 | ||
| 537 | function buildAutofixComment( | |
| bc519ae | 538 | plan: AutofixPlan, |
| 34e63b9 | 539 | idempotencyMarker: string, |
| 540 | gateRunId: string | |
| 541 | ): string { | |
| bc519ae | 542 | const result = plan.fix; |
| 34e63b9 | 543 | const confidenceBadge = |
| 544 | result.confidence === "high" | |
| 545 | ? "🟢 High confidence" | |
| 546 | : result.confidence === "medium" | |
| 547 | ? "🟡 Medium confidence" | |
| 548 | : "🔴 Low confidence"; | |
| 549 | ||
| bc519ae | 550 | // Flywheel bookkeeping marker — applyAutofix parses this to settle the |
| 551 | // pending entry's outcome. Absent when the flywheel write failed. | |
| 552 | const flywheelMarker = plan.flywheelEntryId | |
| 553 | ? `\n${FLYWHEEL_MARKER_PREFIX}${plan.flywheelEntryId} -->` | |
| 554 | : ""; | |
| 555 | ||
| 556 | const sourceNote = | |
| 557 | plan.source === "cache" | |
| 558 | ? `\n\n♻️ **Served from the repair cache** — this failure signature was fixed successfully before; no AI call was made.` | |
| 559 | : ""; | |
| 560 | ||
| 34e63b9 | 561 | return `${CI_AUTOFIX_MARKER} |
| bc519ae | 562 | ${idempotencyMarker}${flywheelMarker} |
| 34e63b9 | 563 | |
| 564 | ## 🔧 AI Auto-Fix | |
| 565 | ||
| bc519ae | 566 | ${result.explanation}${sourceNote} |
| 34e63b9 | 567 | |
| 568 | **Confidence:** ${confidenceBadge} | |
| 569 | ||
| 570 | \`\`\`diff | |
| 571 | ${result.patch} | |
| 572 | \`\`\` | |
| 573 | ||
| 574 | <details><summary>Apply this fix</summary> | |
| 575 | ||
| 576 | Copy the patch above or click **Apply Fix** to commit it automatically. | |
| 577 | ||
| 578 | <form method="post" action="/api/pr-comments/COMMENT_ID/apply-autofix" style="display:inline"> | |
| 579 | <button type="submit" style="margin-top:8px;padding:6px 14px;background:#6c63ff;color:#fff;border:none;border-radius:6px;cursor:pointer"> | |
| 580 | ⚡ Apply Fix | |
| 581 | </button> | |
| 582 | </form> | |
| 583 | ||
| 584 | </details> | |
| 585 | ||
| 586 | <sub>Gate run: <code>${gateRunId}</code> · Affected files: ${result.affectedFiles.join(", ") || "see patch above"}</sub>`; | |
| 587 | } | |
| 588 | ||
| 589 | // --------------------------------------------------------------------------- | |
| 590 | // Apply autofix | |
| 591 | // --------------------------------------------------------------------------- | |
| 592 | ||
| 593 | /** | |
| 594 | * Applies the patch from a PR comment onto a new branch. | |
| 595 | * Returns the new branch name so the caller can redirect to compare view. | |
| bc519ae | 596 | * |
| 597 | * Also settles the comment's pending flywheel entry: a clean apply+commit | |
| 598 | * records 'success' (the pattern becomes replayable by the Tier-0 cache), | |
| 599 | * an apply failure records 'failed' so unreliable patterns lose confidence. | |
| 34e63b9 | 600 | */ |
| 601 | export async function applyAutofix( | |
| 602 | prCommentId: string, | |
| bc519ae | 603 | userId: string, |
| 604 | deps: CiAutofixDeps = {} | |
| 34e63b9 | 605 | ): Promise<{ branchName: string }> { |
| 606 | // 1. Load the comment | |
| 607 | const [comment] = await db | |
| 608 | .select({ | |
| 609 | id: prComments.id, | |
| 610 | pullRequestId: prComments.pullRequestId, | |
| 611 | body: prComments.body, | |
| 612 | isAiReview: prComments.isAiReview, | |
| 613 | }) | |
| 614 | .from(prComments) | |
| 615 | .where(eq(prComments.id, prCommentId)) | |
| 616 | .limit(1); | |
| 617 | ||
| 618 | if (!comment) throw new Error("Comment not found"); | |
| 619 | if (!comment.body.includes(CI_AUTOFIX_MARKER)) { | |
| 620 | throw new Error("Not an autofix comment"); | |
| 621 | } | |
| 622 | ||
| 623 | // 2. Load PR + repo for access check | |
| 624 | const [pr] = await db | |
| 625 | .select() | |
| 626 | .from(pullRequests) | |
| 627 | .where(eq(pullRequests.id, comment.pullRequestId)) | |
| 628 | .limit(1); | |
| 629 | ||
| 630 | if (!pr) throw new Error("PR not found"); | |
| 631 | ||
| 632 | const [repoRow] = await db | |
| 633 | .select({ | |
| 634 | id: repositories.id, | |
| 635 | name: repositories.name, | |
| 636 | ownerId: repositories.ownerId, | |
| 637 | ownerUsername: users.username, | |
| 638 | }) | |
| 639 | .from(repositories) | |
| 640 | .innerJoin(users, eq(repositories.ownerId, users.id)) | |
| 641 | .where(eq(repositories.id, pr.repositoryId)) | |
| 642 | .limit(1); | |
| 643 | ||
| 644 | if (!repoRow) throw new Error("Repository not found"); | |
| 645 | ||
| 646 | // Verify write access: must be repo owner or collaborator | |
| 647 | const isOwner = repoRow.ownerId === userId; | |
| 648 | if (!isOwner) { | |
| 649 | const [collab] = await db | |
| 650 | .select({ id: repoCollaborators.id }) | |
| 651 | .from(repoCollaborators) | |
| 652 | .where( | |
| 653 | and( | |
| 654 | eq(repoCollaborators.repositoryId, repoRow.id), | |
| 655 | eq(repoCollaborators.userId, userId) | |
| 656 | ) | |
| 657 | ) | |
| 658 | .limit(1); | |
| 659 | if (!collab) throw new Error("Forbidden: no write access"); | |
| 660 | } | |
| 661 | ||
| 662 | // 3. Extract patch from comment body | |
| 663 | const patchMatch = comment.body.match(/```diff\n([\s\S]*?)```/); | |
| 664 | if (!patchMatch) throw new Error("No patch found in comment"); | |
| 665 | const patch = patchMatch[1]; | |
| 666 | ||
| 667 | // 4. Create a new branch from the PR's head | |
| 668 | const branchName = `fix/autofix-${Date.now()}`; | |
| 669 | const repoDir = getRepoPath(repoRow.ownerUsername, repoRow.name); | |
| 670 | ||
| 671 | // Create the branch at the PR head SHA | |
| 672 | const headSha = await spawnGit( | |
| 673 | ["rev-parse", pr.headBranch], | |
| 674 | repoDir | |
| 675 | ); | |
| 676 | if (headSha.exitCode !== 0) throw new Error("Cannot resolve head branch"); | |
| 677 | ||
| 678 | await spawnGit( | |
| 679 | ["branch", branchName, headSha.stdout.trim()], | |
| 680 | repoDir | |
| 681 | ); | |
| 682 | ||
| 683 | // 5. Apply the patch via git apply in a temp worktree | |
| 684 | const tmpDir = await mkdtemp(join(tmpdir(), "autofix-")); | |
| 685 | try { | |
| 686 | // Add worktree for the new branch | |
| 687 | const wtResult = await spawnGit( | |
| 688 | ["worktree", "add", tmpDir, branchName], | |
| 689 | repoDir | |
| 690 | ); | |
| 691 | if (wtResult.exitCode !== 0) { | |
| 692 | throw new Error(`git worktree add failed: ${wtResult.stderr}`); | |
| 693 | } | |
| 694 | ||
| 695 | // Write the patch to a temp file | |
| 696 | const patchFile = join(tmpDir, "autofix.patch"); | |
| 697 | await writeFile(patchFile, patch, "utf8"); | |
| 698 | ||
| 699 | // Apply the patch | |
| 700 | const applyResult = await spawnGit( | |
| 701 | ["apply", "--index", patchFile], | |
| 702 | tmpDir | |
| 703 | ); | |
| 704 | if (applyResult.exitCode !== 0) { | |
| 705 | throw new Error(`git apply failed: ${applyResult.stderr}`); | |
| 706 | } | |
| 707 | ||
| 708 | // 6. Commit | |
| 709 | const commitResult = await spawnGit( | |
| 710 | [ | |
| 711 | "commit", | |
| 712 | "-m", | |
| 713 | "fix: apply AI autofix for CI failure", | |
| 714 | "--author", | |
| 715 | "gluecron[bot] <bot@gluecron.com>", | |
| 716 | ], | |
| 717 | tmpDir | |
| 718 | ); | |
| 719 | if (commitResult.exitCode !== 0) { | |
| 720 | throw new Error(`git commit failed: ${commitResult.stderr}`); | |
| 721 | } | |
| 722 | ||
| 723 | // 7. Push the branch back to the bare repo | |
| 724 | // In a bare-repo + worktree setup the push target is the bare repo itself. | |
| 725 | await spawnGit( | |
| 726 | ["push", repoDir, `HEAD:refs/heads/${branchName}`], | |
| 727 | tmpDir | |
| 728 | ); | |
| bc519ae | 729 | |
| 730 | // 8. The repair landed — settle the flywheel entry so this pattern's | |
| 731 | // success rate climbs and future identical failures hit the Tier-0 cache. | |
| 732 | await recordAutofixOutcome(comment.body, "success", deps); | |
| 733 | } catch (err) { | |
| 734 | // The patch failed to apply/commit/push — settle as 'failed' so the | |
| 735 | // flywheel learns this pattern is unreliable. Best-effort: the original | |
| 736 | // error is always rethrown for the route to surface. | |
| 737 | await recordAutofixOutcome(comment.body, "failed", deps); | |
| 738 | throw err; | |
| 34e63b9 | 739 | } finally { |
| 740 | // Cleanup worktree | |
| 741 | await spawnGit(["worktree", "remove", "--force", tmpDir], repoDir).catch( | |
| 742 | () => {} | |
| 743 | ); | |
| 744 | await rm(tmpDir, { recursive: true, force: true }).catch(() => {}); | |
| 745 | } | |
| 746 | ||
| 747 | return { branchName }; | |
| 748 | } |