@linzumi/cli 1.0.275 → 1.0.276

package/scripts/diagnose-claude-probe.ts

@@ -1,198 +0,0 @@
-/*
-- Date: 2026-07-11
-  Spec: audit PR #3019, W17 (William seq 870: "SDK startup 15s timeout x2"
-  while `claude` ran fine when invoked directly).
-  Relationship: Side-by-side diagnostic for the Claude SDK startup probe.
-  Runs (1) a DIRECT invocation of the exact `claude` executable the probe's
-  fallback scan resolves and (2) the real availability probe
-  (probeClaudeCodeAvailabilityStatus - the same code path the runner runs at
-  startup), with IDENTICAL environment preparation
-  (prepareClaudeCodeSpawnEnvironment) and binary resolution, capturing
-  stderr, exit codes, and timing for both. When the probe fails while the
-  direct invocation succeeds, the report shows the probe's captured child
-  stderr tail - the evidence that used to be discarded (the SDK spawns with
-  stdio 'ignore' for stderr unless a callback is passed).
-
-  Usage:
-    npx tsx scripts/diagnose-claude-probe.ts [--claude-bin /path/to/claude]
-
-  Exit code 0 always (this is a reporter, not a gate); read the verdict.
-*/
-import { spawn } from 'node:child_process';
-import { existsSync } from 'node:fs';
-import { homedir } from 'node:os';
-import {
-  findClaudeCodeExecutableFallback,
-  prepareClaudeCodeSpawnEnvironment,
-  probeClaudeCodeAvailabilityStatus,
-} from '../src/claudeCodeSession';
-
-const DIRECT_TIMEOUT_MS = 15_000; // mirrors the probe's initializeTimeoutMs
-
-type DirectResult = {
-  readonly command: string;
-  readonly durationMs: number;
-  readonly exitCode: number | null;
-  readonly timedOut: boolean;
-  readonly stdoutTail: string;
-  readonly stderrTail: string;
-  readonly spawnError?: string;
-};
-
-function tail(value: string, max = 600): string {
-  const collapsed = value.replace(/\s+/g, ' ').trim();
-  return collapsed.slice(-max);
-}
-
-async function runDirect(
-  executable: string,
-  args: readonly string[],
-  env: NodeJS.ProcessEnv
-): Promise<DirectResult> {
-  const started = Date.now();
-  return await new Promise((resolvePromise) => {
-    let stdout = '';
-    let stderr = '';
-    let settled = false;
-    const child = spawn(executable, [...args], { env, windowsHide: true });
-    const settle = (partial: Partial<DirectResult>): void => {
-      if (settled) {
-        return;
-      }
-      settled = true;
-      clearTimeout(deadline);
-      resolvePromise({
-        command: `${executable} ${args.join(' ')}`,
-        durationMs: Date.now() - started,
-        exitCode: child.exitCode,
-        timedOut: false,
-        stdoutTail: tail(stdout),
-        stderrTail: tail(stderr),
-        ...partial,
-      });
-    };
-    const deadline = setTimeout(() => {
-      child.kill('SIGKILL');
-      settle({ timedOut: true, exitCode: null });
-    }, DIRECT_TIMEOUT_MS);
-    child.stdout?.on('data', (chunk: Buffer) => {
-      stdout += chunk.toString();
-    });
-    child.stderr?.on('data', (chunk: Buffer) => {
-      stderr += chunk.toString();
-    });
-    child.on('error', (error) => {
-      settle({ spawnError: error.message });
-    });
-    child.on('close', (code) => {
-      settle({ exitCode: code });
-    });
-  });
-}
-
-async function main(): Promise<void> {
-  const claudeBinFlagIndex = process.argv.indexOf('--claude-bin');
-  const claudeBin =
-    claudeBinFlagIndex === -1
-      ? undefined
-      : process.argv[claudeBinFlagIndex + 1];
-  // A bare `--claude-bin` with no path must fail loudly, not silently run
-  // as if the flag was never given (Greptile P2, PR #3035).
-  if (claudeBinFlagIndex !== -1 && claudeBin === undefined) {
-    console.error('--claude-bin requires a path argument');
-    process.exit(1);
-  }
-
-  // IDENTICAL env preparation to both the probe and real sessions: the one
-  // shared function (merged login-shell PATH).
-  const env = prepareClaudeCodeSpawnEnvironment();
-
-  // IDENTICAL binary resolution to the probe's fallback scan.
-  const resolvedFallback = findClaudeCodeExecutableFallback({
-    env,
-    platform: process.platform,
-    fileExists: existsSync,
-    homeDir: homedir(),
-  });
-  const directExecutable = claudeBin ?? resolvedFallback;
-
-  console.log('=== W17 claude probe side-by-side diagnostic ===');
-  console.log(`cwd: ${process.cwd()}`);
-  console.log(
-    `resolved user-installed claude: ${resolvedFallback ?? '(none found on PATH or in ~/.local/bin, ~/.claude/local)'}`
-  );
-  console.log(`explicit --claude-bin: ${claudeBin ?? '(not set)'}`);
-  console.log('');
-
-  // --- Side A: direct invocation of the same binary ---
-  console.log('--- A: direct invocation ---');
-  let direct: DirectResult | undefined;
-  if (directExecutable === undefined) {
-    console.log('skipped: no claude executable resolved for a direct run');
-  } else {
-    direct = await runDirect(directExecutable, ['--version'], env);
-    console.log(
-      `direct \`${direct.command}\`: exit=${direct.exitCode} timedOut=${direct.timedOut} in ${direct.durationMs}ms`
-    );
-    if (direct.stdoutTail !== '') {
-      console.log(`  stdout: ${direct.stdoutTail}`);
-    }
-    if (direct.stderrTail !== '') {
-      console.log(`  stderr: ${direct.stderrTail}`);
-    }
-    if (direct.spawnError !== undefined) {
-      console.log(`  spawn error: ${direct.spawnError}`);
-    }
-  }
-  console.log('');
-
-  // --- Side B: the real SDK startup probe (same code path as the runner) ---
-  console.log(
-    '--- B: SDK startup probe (probeClaudeCodeAvailabilityStatus) ---'
-  );
-  const started = Date.now();
-  const outcome = await probeClaudeCodeAvailabilityStatus({
-    cwd: process.cwd(),
-    claudeBin,
-  });
-  const probeMs = Date.now() - started;
-  console.log(
-    `probe verdict: status=${outcome.status} transient=${outcome.transient} in ${probeMs}ms`
-  );
-  if (outcome.detail !== undefined) {
-    console.log(`probe failure detail (SDK error + child stderr tail):`);
-    console.log(`  ${outcome.detail}`);
-  }
-  console.log('');
-
-  console.log('--- verdict ---');
-  // The verdict branches on side A's ACTUAL result: "both failed" is a
-  // uniformly broken setup, not a probe-vs-direct divergence (Greptile P2,
-  // PR #3035).
-  const directSucceeded =
-    direct !== undefined &&
-    direct.exitCode === 0 &&
-    !direct.timedOut &&
-    direct.spawnError === undefined;
-  if (outcome.status === 'logged_in') {
-    console.log(
-      'probe and direct invocation agree: Claude Code is launchable.'
-    );
-  } else if (directSucceeded) {
-    console.log(
-      'DIVERGENCE: the direct invocation succeeded but the probe failed. The failure detail above is the evidence to attach to the issue.'
-    );
-  } else if (direct !== undefined) {
-    console.log(
-      'both the direct invocation AND the probe failed - a uniformly broken claude setup on this machine, not a probe divergence. Fix the direct invocation first (its stderr/exit code above).'
-    );
-  } else {
-    console.log(
-      'no user-installed claude and the probe failed - see detail above.'
-    );
-  }
-  // Reporter, not a gate.
-  process.exit(0);
-}
-
-void main();

package/scripts/differentialMutations.mjs

@@ -1,114 +0,0 @@
-// Spec: kandan/plans/2026-07-10-compute-nodes-rescript-commander-zero-ts-program.md (M2.6)
-// Relationship: The maintained seeded-mutation set for the differential
-// harness's anti-false-green mutation check. Each entry is a deliberate
-// behavior delta patched into the TS commander AT BUILD TIME
-// (scripts/build-differential-entry.mjs --mutation <name>); the harness's
-// MutationCheckConformance then asserts the differential suite goes RED on
-// the mutant (transcript diff for the shape/ordering classes, a named
-// ported-invariant violation for the invariant class). A seeded delta the
-// suite stays green on means the gate is lying - that conformance leg fails.
-//
-// This registry is the SINGLE source (derive, don't duplicate): the build
-// script applies `patches`, and the harness conformance imports this exact
-// module to drive and pin the set (closed classes, one mutation per class).
-//
-// MAINTENANCE RULES:
-//   - The set stays small (one mutation per class below) - every entry is a
-//     permanent CI cost, so each must earn its place by proving a DISTINCT
-//     sensitivity class of the differential canonical transcript.
-//   - `find` strings must match the current source EXACTLY ONCE; the build
-//     fails loudly on zero or multiple matches, so a refactor can never
-//     silently no-op a mutation (the anchor rot the exactly-once rule
-//     exists for).
-//   - Mutations must never remove an event a differential scenario AWAITS
-//     (the choreography synchronizes on milestones; deleting one turns a
-//     fast red into a slow timeout red).
-//   - No current-version literals (the commander-harness package rule).
-//
-// Classes (the M2.6 task's closed vocabulary):
-//   - transcript-shape: an event the commander silently stops emitting must
-//     change the canonical transcript bytes.
-//   - invariant: a behavior delta violating a ported ordering invariant
-//     must surface as that invariant's named violation.
-//   - ordering: a pure reorder of the commander's own send order (same
-//     event multiset) must change the canonical transcript bytes.
-
-export const differentialMutations = [
-  {
-    name: 'message-state-processing-drop',
-    class: 'transcript-shape',
-    scenario: 'harness-session-rebuild',
-    proves:
-      'a silently dropped message-state publish (the msg-state ladder silent-drop class: the user-visible "processing / starting turn" paint never reaches the server) changes the canonical transcript - the differential diff can see a suppressed notification type',
-    patches: [
-      {
-        file: 'src/channelSession.ts',
-        find: `  const startingConsensus = await publishQueuedMessageState(
-    args,
-    state,
-    next,
-    {
-      status: 'processing',
-      reason: 'starting turn',
-    },
-    'acked'
-  );`,
-        replace: `  // [seeded differential mutation: message-state-processing-drop] the
-  // starting-turn processing paint is silently dropped.
-  const startingConsensus = { kind: 'applied' } as const;`,
-      },
-    ],
-  },
-  {
-    name: 'join-roster-drop',
-    class: 'invariant',
-    expectInvariant: 'I319',
-    scenario: 'full-join-surfaces',
-    proves:
-      'a commander that silently drops the join-reply roster (the Commander-boots-into-empty-onboarding regression class the compute-nodes program exists to kill) under-delivers the display page and goes red as an I319 violation, not just a byte diff',
-    patches: [
-      {
-        file: 'src/runner.ts',
-        find: `  seedDashboardRosterFromReply(joinResponse, 'join');`,
-        replace: `  // [seeded differential mutation: join-roster-drop] the join-reply
-  // roster seed is silently dropped.`,
-      },
-    ],
-  },
-  {
-    name: 'handshake-initialized-reorder',
-    class: 'ordering',
-    scenario: 'harness-clean-turn',
-    proves:
-      "a pure reorder of the commander's own send order (the codex app-server handshake: `initialized` notified BEFORE the `initialize` request instead of after its result) changes the canonical transcript even though the event multiset is identical - per-party send order is preserved evidence, not normalized away",
-    patches: [
-      {
-        file: 'src/codexAppServer.ts',
-        // The early frame below is byte-for-byte the wire shape of the
-        // pristine `initialized` literal in codexAppServer.ts, so this
-        // mutation stays a PURE reorder. Drift teeth (Greptile finding on
-        // #3115): if the real notification ever gains fields, the early
-        // frame's content diverges and MutationCheckConformance's
-        // ordering-class assertion (equal sorted-line multiset) goes red
-        // until this patch is re-synced - the payload cannot silently rot.
-        find: `  const response = await sendRequest(websocket, pending, {
-    jsonrpc: '2.0',
-    id: 'initialize',
-    method: 'initialize',`,
-        replace: `  // [seeded differential mutation: handshake-initialized-reorder] the
-  // initialized notification jumps AHEAD of the initialize request.
-  websocket.send(JSON.stringify({ jsonrpc: '2.0', method: 'initialized' }));
-  const response = await sendRequest(websocket, pending, {
-    jsonrpc: '2.0',
-    id: 'initialize',
-    method: 'initialize',`,
-      },
-      {
-        file: 'src/codexAppServer.ts',
-        find: `  websocket.send(JSON.stringify(initialized));`,
-        replace: `  // [seeded differential mutation: handshake-initialized-reorder] the
-  // post-result send is suppressed (it already went out early above).`,
-      },
-    ],
-  },
-];

package/scripts/f5_live_cross_harness_failover_e2e.mjs

@@ -1,239 +0,0 @@
-#!/usr/bin/env node
-// F5 LIVE real-harness e2e (compute-nodes program, Workstream F): the
-// cross-harness failover RUNG PATH against the REAL claude-code harness -
-// no scripted double anywhere on the target side.
-//
-//   1. The failover ORACLE (the same bundle the conformance suite runs:
-//      scripts/build-cross-harness-failover-oracle.mjs) drives the REAL
-//      channelSession: a codex thread runs, its backend DIES mid-thread
-//      (the worker/IPC client closes - the precise F5 trigger), the fate
-//      ladder decides the cross_harness_failover rung, and the REAL
-//      transpilation bridge (T.2 exportPage -> T.4 rehydrateSession)
-//      seeds the REAL claude session store (~/.claude/projects/...).
-//   2. The REAL @anthropic-ai/claude-agent-sdk validates the seeded
-//      session (getSessionInfo), reads the transpiled history back
-//      (getSessionMessages), and RESUMES it (query with options.resume),
-//      completing a REAL follow-up turn whose answer must reference
-//      content that exists ONLY in the failed codex thread's transpiled
-//      activity (the coherence probe).
-//
-// This is a LOCAL, manual, credentialed run (it spends one real model
-// turn); it is NOT in any CI test list. Evidence lands in
-// /tmp/linzumi-evidence-f5/ and is mirrored into the committed evidence
-// doc (kandan/server_v2/web/docs/qa-evidence/2026-07-14-f5-failover/).
-//
-// Scope honesty (the T.4 precedent, extended): there is deliberately NO
-// deployed-preview leg - the failover's kandan/codex sides ride the
-// oracle's scripted doubles (killing a real production backend mid-thread
-// on the preview is not an exercisable CI act), while the LADDER, the
-// TRANSPILERS, the SESSION STORE, and the CLAUDE HARNESS are all real.
-
-import { execFileSync } from 'node:child_process';
-import { randomUUID } from 'node:crypto';
-import { mkdirSync, readFileSync, realpathSync, writeFileSync } from 'node:fs';
-import { homedir } from 'node:os';
-import { dirname, join, resolve } from 'node:path';
-import { fileURLToPath } from 'node:url';
-
-const packageRoot = resolve(dirname(fileURLToPath(import.meta.url)), '..');
-const evidenceDir = '/tmp/linzumi-evidence-f5';
-mkdirSync(evidenceDir, { recursive: true });
-
-function fail(message) {
-  console.error(`F5-LIVE-E2E FAIL: ${message}`);
-  process.exit(1);
-}
-
-// --- 1. The REAL rung path: oracle -> ladder -> bridge -> REAL store ---------
-
-execFileSync(
-  process.execPath,
-  [join(packageRoot, 'scripts', 'build-cross-harness-failover-oracle.mjs')],
-  { stdio: 'inherit' }
-);
-
-const cwdRaw = join('/tmp', `linzumi-f5-live-${Date.now()}`);
-mkdirSync(cwdRaw, { recursive: true });
-const cwd = realpathSync(cwdRaw);
-
-const reportRaw = execFileSync(
-  process.execPath,
-  [join(packageRoot, '.differential', 'cross-harness-failover-oracle.mjs')],
-  {
-    env: {
-      ...process.env,
-      // The REAL store: the bridge seeds ~/.claude/projects/<munged-cwd>/.
-      F5_ORACLE_HOME: homedir(),
-      F5_ORACLE_CWD: cwd,
-      F5_ORACLE_EXECUTOR: 'bridge',
-    },
-    encoding: 'utf8',
-  }
-);
-const report = JSON.parse(reportRaw.trim());
-
-const failoverStep = (report.ladderSteps ?? []).find(
-  (step) => step.rung === 'cross_harness_failover'
-);
-if (!failoverStep) fail('the ladder never decided the failover rung');
-if (failoverStep.fate !== 'transient_local') {
-  fail(`unexpected fate: ${failoverStep.fate}`);
-}
-const dispatchedLog = (report.failoverLogs ?? []).find(
-  (log) => log.event === 'codex.fate_ladder_cross_harness_failover_dispatched'
-);
-if (!dispatchedLog) fail('no failover_dispatched receipt log');
-if (dispatchedLog.payload.model !== 'oracle-model-slug') {
-  fail('the receipts did not carry the explicitly-passed model (D39)');
-}
-if (!report.seeded) fail('the bridge seeded no session store');
-const { targetSessionId, sessionFile, transpiledItemCount } = report.seeded;
-console.log('[f5-live] rung decided + REAL store seeded:', sessionFile);
-console.log(
-  '[f5-live] transpiled items:',
-  transpiledItemCount,
-  '(redacted reasoning:',
-  report.seeded.redactedReasoningCount + ')'
-);
-
-writeFileSync(
-  join(evidenceDir, 'f5-live-oracle-report.json'),
-  JSON.stringify(report, null, 2) + '\n'
-);
-
-// The seeded store must contain the recap marker + the journal content.
-const storeLines = readFileSync(sessionFile, 'utf8')
-  .trim()
-  .split('\n')
-  .map((line) => JSON.parse(line));
-const storeText = JSON.stringify(storeLines);
-if (!storeText.includes('[linzumi cross-harness resume]')) {
-  fail('the seeded store carries no cross-harness resume marker');
-}
-if (!storeText.includes('renamed the column and updated 3 call sites')) {
-  fail('the seeded store lost the transpiled codex activity');
-}
-
-// --- 2. The REAL SDK: validate, read back, RESUME, follow-up turn ------------
-
-const sdk = await import('@anthropic-ai/claude-agent-sdk');
-
-// Never run the harness as a nested claude-code session.
-const env = { ...process.env };
-delete env.CLAUDECODE;
-delete env.CLAUDE_CODE_ENTRYPOINT;
-delete env.CLAUDE_CODE_SSE_PORT;
-
-const info = await sdk.getSessionInfo(targetSessionId, { dir: cwd });
-if (!info) fail('the real SDK store did not find the seeded session');
-console.log('[f5-live] getSessionInfo found the failover-seeded session');
-
-const storedMessages = await sdk.getSessionMessages(targetSessionId, {
-  dir: cwd,
-});
-if (storedMessages.length !== storeLines.length) {
-  fail(
-    `store read-back count ${storedMessages.length} != seeded ${storeLines.length}`
-  );
-}
-console.log(
-  '[f5-live] getSessionMessages returned the seeded history:',
-  storedMessages.length,
-  'messages'
-);
-
-// The coherence probe: answerable ONLY from the transpiled codex activity
-// that crossed with the failover.
-const probe =
-  'Answer from this conversation only, in one short sentence: before the ' +
-  'harness failover, how many call sites did the agent say it had already ' +
-  'updated after renaming the column?';
-
-const transcript = [];
-const query = sdk.query({
-  prompt: probe,
-  options: {
-    cwd,
-    resume: targetSessionId,
-    model: 'haiku',
-    permissionMode: 'default',
-    allowedTools: [],
-    maxTurns: 1,
-    persistSession: true,
-    env,
-  },
-});
-
-let announcedSessionId = null;
-let resultMessage = null;
-let assistantText = '';
-for await (const message of query) {
-  transcript.push(message);
-  if (message.type === 'system' && message.subtype === 'init') {
-    announcedSessionId = message.session_id;
-    console.log('[f5-live] system:init announced session:', announcedSessionId);
-  }
-  if (message.type === 'assistant') {
-    const blocks = message.message?.content ?? [];
-    for (const block of blocks) {
-      if (block.type === 'text') assistantText += block.text + '\n';
-    }
-  }
-  if (message.type === 'result') {
-    resultMessage = message;
-    break;
-  }
-}
-
-writeFileSync(
-  join(evidenceDir, 'f5-live-resume-transcript.ndjson'),
-  transcript.map((message) => JSON.stringify(message)).join('\n') + '\n'
-);
-
-if (!resultMessage) fail('the follow-up turn produced no result terminal');
-if (resultMessage.is_error) {
-  fail(`the follow-up turn errored: ${JSON.stringify(resultMessage)}`);
-}
-console.log('[f5-live] follow-up turn completed:', resultMessage.subtype);
-console.log('[f5-live] follow-up answer:', assistantText.trim());
-
-const coherent = /\b(3|three)\b/i.test(assistantText);
-if (!coherent) {
-  fail(
-    'the follow-up answer does not reference the transpiled codex activity ' +
-      '(expected the 3 updated call sites) - the resume did not carry the ' +
-      'failed thread across the harness boundary'
-  );
-}
-console.log(
-  '[f5-live] COHERENCE PROVED: the answer references the 3 updated call ' +
-    'sites, which exist only in the transpiled codex activity'
-);
-
-const summary = {
-  ran_at: new Date().toISOString(),
-  trigger:
-    'transient_local infra kill on a DEAD codex worker (the oracle-scripted backend), redispatch/worker_rebuild unavailable',
-  flag: 'LINZUMI_CROSS_HARNESS_FAILOVER=1 (default OFF)',
-  ladder_decision: failoverStep,
-  failover_receipt: dispatchedLog.payload,
-  source_harness: 'codex',
-  target_harness: 'claude_code',
-  target_session_id: targetSessionId,
-  announced_session_id: announcedSessionId,
-  session_file: sessionFile,
-  transpiled_item_count: transpiledItemCount,
-  redacted_reasoning_count: report.seeded.redactedReasoningCount,
-  follow_up_result_subtype: resultMessage.subtype,
-  follow_up_answer: assistantText.trim(),
-  coherence_probe: probe,
-  coherence_proved: true,
-  scope_honesty:
-    'kandan/codex sides scripted by the oracle doubles; the ladder, the T transpilers, the session store, and the claude harness (SDK resume + real model turn) are all real. No deployed-preview leg (killing a real backend mid-thread on the preview is not an exercisable CI act).',
-};
-writeFileSync(
-  join(evidenceDir, 'f5-live-failover-summary.json'),
-  JSON.stringify(summary, null, 2) + '\n'
-);
-console.log('[f5-live] PASS - evidence in', evidenceDir);
-process.exit(0);