import { spawn } from "node:child_process"; import path from "node:path"; import { fileURLToPath } from "node:url"; // Resolve the showcase root relative to THIS source file so the repro // drives the same checkout the test is being run from (worktree or // main). Hard-coding /Users/jpr5/proj/cpk/cpk/showcase would drive the // main checkout from inside a worktree and the runner edits from the // worktree would never execute. The test file lives at // showcase/harness/test/integration/bubble-race-repro.ts; the showcase // root is three directories up. const SHOWCASE_ROOT = path.resolve( path.dirname(fileURLToPath(import.meta.url)), "..", "..", "..", ); export interface BubbleRaceReproOpts { /** Harness target — `` or `:` per cli.ts:52. */ slug: string; /** Probe depth shorthand — one of "smoke"|"d4"|"d5"|"d6". */ level: "smoke" | "d4" | "d5" | "d6"; /** * User messages to drive into the existing canonical fixture for * `:`. The aimock match contract is content-based * (userMessage + context), so per-scenario behavior is selected by * sending the right message rather than swapping fixture files. */ messages: string[]; /** * Optional HTML injected via Playwright `addInitScript` BEFORE * navigation — used by defect 4 only. Forwarded to the harness via * the `BUBBLE_RACE_PRE_PAINT` env var; `helpers/init-scripts.ts` * reads it and calls `page.addInitScript` against the d5/d6 drivers' * pre-navigation hook (the first `page.goto` callsite for the path * the test drives). */ prePaint?: string; /** * Optional CSS selector whose matching nodes have their * `data-testid` attribute stripped via Playwright `addInitScript` * BEFORE navigation — used by defect 3 only when no natural * cascade-fallback-only demo exists. Forwarded to the harness via * the `BUBBLE_RACE_STRIP_SELECTOR` env var; `helpers/init-scripts.ts` * reads it and calls `page.addInitScript` to remove the testid from * any rendered node matching the selector, forcing the runner's * cascade to fall through to a non-canonical tier. */ prePaintStrip?: string; } export interface BubbleRaceTurnResult { turnIndex: number; assistantText: string; } export interface BubbleRaceReproResult { exitCode: number; turns: BubbleRaceTurnResult[]; /** Absolute path to the harness run-artifacts directory (parsed from stdout). */ runDir: string; stdout: string; stderr: string; } /** * Drives `bin/showcase test -- --verbose` (which in turn * runs `npx tsx harness/src/cli.ts test …` per cmd-test.sh:143). The * harness runs from source — no Docker image of the harness exists in * the local path. Parses per-turn assistant-text emissions out of * verbose stdout. * * The driver depends on TWO production log lines added in the SAME * commit that lands this driver: * 1. `[conversation-runner] turn N/total — settled text { turnNum, text }` * — added inside the runner immediately after the existing * `assistant settled` debug block (~line 461), matching the * surrounding "turn N/total — phrase" + structured-2nd-arg * style. Moves into `waitForTurnComplete` in Phase 3 and * survives Phase 5 cutover. * 2. `[harness] runDir=…` — the harness already prints the run * artifacts directory in --verbose mode; the driver captures it * out of stdout. If the existing log uses a different label, * adjust the parser accordingly (verified during Phase 0). * * Per-scenario behavior is configured via `messages` (the array of * user inputs the harness sends — wired into the d5/d6 probe drivers' * existing user-message channel) and optionally `prePaint` (defect 4 * only, forwarded via BUBBLE_RACE_PRE_PAINT env to init-scripts.ts). */ export async function runBubbleRaceRepro( opts: BubbleRaceReproOpts, ): Promise { const env = { ...process.env, BUBBLE_RACE_MESSAGES: JSON.stringify(opts.messages), ...(opts.prePaint ? { BUBBLE_RACE_PRE_PAINT: opts.prePaint } : {}), ...(opts.prePaintStrip ? { BUBBLE_RACE_STRIP_SELECTOR: opts.prePaintStrip } : {}), }; // `--direct` runs d5/d6 via the in-process driver in cli.ts rather than // the control-plane worker path. This is REQUIRED for the bubble-race // repros because the driver parses `[conversation-runner] turn N/total — // settled text …` lines out of the CLI subprocess's stdout; in // control-plane mode, those logs are emitted inside a separate worker // process whose stdout never reaches the CLI. The `--direct` switch is // exposed by cmd-test.sh (line 17/62) and the harness CLI (cli.ts:80). const proc = spawn( "bin/showcase", ["test", opts.slug, `--${opts.level}`, "--direct", "--verbose"], { cwd: SHOWCASE_ROOT, env }, ); let stdout = ""; let stderr = ""; proc.stdout.on("data", (b) => (stdout += b.toString())); proc.stderr.on("data", (b) => (stderr += b.toString())); proc.on("error", (err) => { console.warn( `[bubble-race-repro] spawn error: ${(err as Error).message ?? err}`, ); }); const exitCode: number = await new Promise((resolve) => proc.on("close", (code) => resolve(code ?? -1)), ); const turns: BubbleRaceTurnResult[] = []; // Match the structured-2nd-arg console.debug emission, e.g. // [conversation-runner] turn 1/3 — settled text { turnNum: 1, text: 'Bubbles' } // The 2nd-arg object is rendered by Node's util.inspect when the // shell captures stdout; the production emission shape is // `{ turnNum: N, text: }`. util.inspect picks the quote // character based on string content: single quotes by default, but // SWITCHES to double quotes when the string contains a `'` and no // `"` (and backticks when it contains both). The parser must accept // either single- or double-quoted form; the two-alternation regex // below mirrors util.inspect's standard escape semantics for each // (backslash-escape of the same quote + a literal backslash). The // PRE-text segment is anchored on the literal `turnNum: , // text:` key sequence rather than `[^}]*` so a stray `}` anywhere in // the line can't collapse the match boundary; the quoted-string // capture itself already terminates on the unescaped closing quote, // so by the time we're past it the only legal trailer is ` }`. // Also tolerates the multi-line pretty-printed inspect output // (`{\n turnNum: …,\n text: …\n}`) via the leading `\s*`. const re = /\[conversation-runner\] turn (\d+)\/\d+ — settled text \{\s*turnNum:\s*\d+,\s*text:\s*(?:'((?:[^'\\]|\\.)*)'|"((?:[^"\\]|\\.)*)")\s*\}/gm; let m: RegExpExecArray | null; while ((m = re.exec(stdout)) !== null) { // Un-escape the util.inspect output. The escape rules differ by // quote character — `\'` inside single-quoted, `\"` inside // double-quoted — but `\\` is the same in both. m[2] holds the // single-quoted capture; m[3] holds the double-quoted capture. const raw = m[2] ?? m[3] ?? ""; const escaped = m[2] !== undefined ? raw.replace(/\\'/g, "'") : raw.replace(/\\"/g, '"'); const text = escaped.replace(/\\\\/g, "\\"); turns.push({ turnIndex: Number(m[1]), assistantText: text }); } const runDirMatch = stdout.match(/\[harness\] runDir=(\S+)/); const runDir = runDirMatch ? runDirMatch[1] : ""; return { exitCode, turns, runDir, stdout, stderr }; }