llm-switcher 1.3.7 → 1.4.0

package/tests/live-optimizer-interop.mjs

@@ -1,205 +1,205 @@
-// LIVE test: requires a running gateway (switch on) and a real upstream (costs tokens).
-// Offline test, no network needed: npm test
-const GATEWAY_PORT = Number(process.env.LLM_SWITCHER_PORT) || 3456;
-const BASE_URL = `http://127.0.0.1:${GATEWAY_PORT}`;
-
-console.log('================================================================');
-console.log('  LLM SWITCHER — TOKEN OPTIMIZER INTEROPERABILITY TEST SUITE   ');
-console.log('  Simulating failure modes from Headroom, RTK, and Ponytail    ');
-console.log('================================================================\n');
-
-let passed = 0;
-let failed = 0;
-
-async function runTest(name, fn) {
-  process.stdout.write(`[TEST] ${name}... `);
-  const t0 = Date.now();
-  try {
-    const detail = await fn();
-    console.log(`PASS (${Date.now() - t0}ms)`);
-    if (detail) console.log(`       ↳ ${detail}`);
-    passed++;
-  } catch (err) {
-    console.log(`FAIL (${Date.now() - t0}ms)`);
-    console.log(`       ↳ ERROR: ${err.message}`);
-    failed++;
-  }
-}
-
-// -----------------------------------------------------------------------------
-// TEST 1: Headroom Failure Mode — Orphaned tool_result
-// When Headroom prunes history to save tokens, it drops the assistant tool_use turn.
-// Anthropic strictly throws: 400 invalid_request_error: 'tool_use_id does not correspond to any tool_use'
-// LLM Switcher Healer Engine: Repairs orphaned result into context text -> 200 OK
-// -----------------------------------------------------------------------------
-await runTest('Headroom Simulation: Orphaned tool_result turn', async () => {
-  const payload = {
-    model: 'claude-opus-4-6',
-    max_tokens: 40,
-    messages: [
-      { role: 'user', content: 'Context turn before pruning' },
-      {
-        role: 'user',
-        content: [
-          { type: 'tool_result', tool_use_id: 'headroom_pruned_call_99', content: 'Database query result: 42 rows found' },
-          { type: 'text', text: 'Reply: healed successfully' }
-        ]
-      }
-    ],
-    stream: false
-  };
-
-  const res = await fetch(`${BASE_URL}/v1/messages`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json', 'anthropic-version': '2023-06-01' },
-    body: JSON.stringify(payload)
-  });
-
-  if (res.status !== 200) {
-    const err = await res.text();
-    throw new Error(`Expected HTTP 200, got ${res.status}: ${err}`);
-  }
-  const json = await res.json();
-  const text = json.content?.map(c => c.text).join('') || '';
-  return `Healed orphaned tool_result. Response HTTP 200: "${text.slice(0, 50).replace(/\n/g, ' ')}..."`;
-});
-
-// -----------------------------------------------------------------------------
-// TEST 2: Headroom Failure Mode — Consecutive User turns (Roles must alternate)
-// When Headroom collapses history, multiple user turns occur back-to-back.
-// Anthropic strictly throws: 400 invalid_request_error: 'roles must alternate'
-// LLM Switcher Healer Engine: Merges consecutive same-role turns -> 200 OK
-// -----------------------------------------------------------------------------
-await runTest('Headroom Simulation: Consecutive User turns (Role alternation violation)', async () => {
-  const payload = {
-    model: 'claude-opus-4-6',
-    max_tokens: 30,
-    messages: [
-      { role: 'user', content: 'User message turn 1' },
-      { role: 'user', content: 'User message turn 2 (no assistant in between)' },
-      { role: 'user', content: 'User message turn 3: reply pong' }
-    ],
-    stream: false
-  };
-
-  const res = await fetch(`${BASE_URL}/v1/messages`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json', 'anthropic-version': '2023-06-01' },
-    body: JSON.stringify(payload)
-  });
-
-  if (res.status !== 200) {
-    const err = await res.text();
-    throw new Error(`Expected HTTP 200, got ${res.status}: ${err}`);
-  }
-  const json = await res.json();
-  return `Merged consecutive turns seamlessly. Stop reason: ${json.stop_reason}`;
-});
-
-// -----------------------------------------------------------------------------
-// TEST 3: Headroom / RTK Failure Mode — Stripped Thinking Parameter
-// Optimizer stripped the 'thinking' object to reduce tokens.
-// LLM Switcher Thinking Guard: Detects reasoning model (ag/claude-opus-4-6-thinking)
-// and restores thinking.budget_tokens automatically -> thinking_delta emitted!
-// -----------------------------------------------------------------------------
-await runTest('Thinking Guard: Restoring stripped thinking parameter on reasoning models', async () => {
-  const payload = {
-    model: 'claude-opus-4-6',
-    max_tokens: 200,
-    // Note: NO thinking parameter included (simulating optimizer stripping it)
-    messages: [
-      { role: 'user', content: 'Solve step by step: what is 17 * 23? Show reasoning.' }
-    ],
-    stream: true
-  };
-
-  const res = await fetch(`${BASE_URL}/v1/messages`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json', 'anthropic-version': '2023-06-01' },
-    body: JSON.stringify(payload)
-  });
-
-  if (res.status !== 200) {
-    throw new Error(`Expected HTTP 200, got ${res.status}`);
-  }
-
-  const text = await res.text();
-  const hasThinkingDelta = text.includes('thinking_delta');
-  const hasSignatureDelta = text.includes('signature_delta');
-  const hasTextDelta = text.includes('text_delta');
-
-  if (!hasThinkingDelta) {
-    throw new Error('Thinking Guard failed: thinking_delta was NOT emitted in stream!');
-  }
-
-  return `Automatically restored thinking: thinking_delta=${hasThinkingDelta}, signature_delta=${hasSignatureDelta}, text_delta=${hasTextDelta}`;
-});
-
-// -----------------------------------------------------------------------------
-// TEST 4: RTK & Intermediary Headers Passthrough
-// RTK / tracing tools inject headers: x-rtk-version, traceparent, x-request-id.
-// LLM Switcher: Transparently preserves all tracking headers without rejection.
-// -----------------------------------------------------------------------------
-await runTest('RTK Intermediary: Custom headers and traceparent passthrough', async () => {
-  const payload = {
-    model: 'claude-opus-4-6',
-    max_tokens: 20,
-    messages: [{ role: 'user', content: 'ping' }],
-    stream: false
-  };
-
-  const res = await fetch(`${BASE_URL}/v1/messages`, {
-    method: 'POST',
-    headers: {
-      'Content-Type': 'application/json',
-      'anthropic-version': '2023-06-01',
-      'x-rtk-version': '0.37.2',
-      'x-optimizer-id': 'rtk-cli-hook',
-      'traceparent': '00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-01'
-    },
-    body: JSON.stringify(payload)
-  });
-
-  if (res.status !== 200) {
-    throw new Error(`Expected HTTP 200, got ${res.status}`);
-  }
-  return `Headers accepted cleanly with HTTP 200 OK`;
-});
-
-// -----------------------------------------------------------------------------
-// TEST 5: OpenAI Chat Healer Mode — Orphaned Tool Result in Chat Completions
-// Same test but over /v1/chat/completions: tool role message without prior tool_calls assistant message.
-// OpenAI API strictly throws: 400 'messages with role tool must be a response to a preceding message with tool_calls'
-// LLM Switcher: Converts orphaned tool to user text message -> 200 OK
-// -----------------------------------------------------------------------------
-await runTest('OpenAI Chat Healer: Orphaned role tool without preceding assistant tool_calls', async () => {
-  const payload = {
-    model: 'ag/gpt-oss-120b-medium',
-    max_tokens: 30,
-    messages: [
-      { role: 'user', content: 'Turn 1 context' },
-      { role: 'tool', tool_call_id: 'orphaned_chat_call_77', content: 'Tool execution logs: OK' },
-      { role: 'user', content: 'reply pong' }
-    ],
-    stream: false
-  };
-
-  const res = await fetch(`${BASE_URL}/v1/chat/completions`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify(payload)
-  });
-
-  if (res.status !== 200) {
-    const err = await res.text();
-    throw new Error(`Expected HTTP 200, got ${res.status}: ${err}`);
-  }
-  const json = await res.json();
-  return `Chat Healer rescued orphaned tool role. Stop reason: ${json.choices?.[0]?.finish_reason}`;
-});
-
-console.log('\n================================================================');
-console.log(`  TEST RESULTS: ${passed} PASSED / ${failed} FAILED               `);
-console.log('================================================================\n');
-
-process.exit(failed ? 1 : 0);
+// LIVE test: requires a running gateway (switch on) and a real upstream (costs tokens).
+// Offline test, no network needed: npm test
+const GATEWAY_PORT = Number(process.env.LLM_SWITCHER_PORT) || 3456;
+const BASE_URL = `http://127.0.0.1:${GATEWAY_PORT}`;
+
+console.log('================================================================');
+console.log('  LLM SWITCHER — TOKEN OPTIMIZER INTEROPERABILITY TEST SUITE   ');
+console.log('  Simulating failure modes from Headroom, RTK, and Ponytail    ');
+console.log('================================================================\n');
+
+let passed = 0;
+let failed = 0;
+
+async function runTest(name, fn) {
+  process.stdout.write(`[TEST] ${name}... `);
+  const t0 = Date.now();
+  try {
+    const detail = await fn();
+    console.log(`PASS (${Date.now() - t0}ms)`);
+    if (detail) console.log(`       ↳ ${detail}`);
+    passed++;
+  } catch (err) {
+    console.log(`FAIL (${Date.now() - t0}ms)`);
+    console.log(`       ↳ ERROR: ${err.message}`);
+    failed++;
+  }
+}
+
+// -----------------------------------------------------------------------------
+// TEST 1: Headroom Failure Mode — Orphaned tool_result
+// When Headroom prunes history to save tokens, it drops the assistant tool_use turn.
+// Anthropic strictly throws: 400 invalid_request_error: 'tool_use_id does not correspond to any tool_use'
+// LLM Switcher Healer Engine: Repairs orphaned result into context text -> 200 OK
+// -----------------------------------------------------------------------------
+await runTest('Headroom Simulation: Orphaned tool_result turn', async () => {
+  const payload = {
+    model: 'claude-opus-4-6',
+    max_tokens: 40,
+    messages: [
+      { role: 'user', content: 'Context turn before pruning' },
+      {
+        role: 'user',
+        content: [
+          { type: 'tool_result', tool_use_id: 'headroom_pruned_call_99', content: 'Database query result: 42 rows found' },
+          { type: 'text', text: 'Reply: healed successfully' }
+        ]
+      }
+    ],
+    stream: false
+  };
+
+  const res = await fetch(`${BASE_URL}/v1/messages`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json', 'anthropic-version': '2023-06-01' },
+    body: JSON.stringify(payload)
+  });
+
+  if (res.status !== 200) {
+    const err = await res.text();
+    throw new Error(`Expected HTTP 200, got ${res.status}: ${err}`);
+  }
+  const json = await res.json();
+  const text = json.content?.map(c => c.text).join('') || '';
+  return `Healed orphaned tool_result. Response HTTP 200: "${text.slice(0, 50).replace(/\n/g, ' ')}..."`;
+});
+
+// -----------------------------------------------------------------------------
+// TEST 2: Headroom Failure Mode — Consecutive User turns (Roles must alternate)
+// When Headroom collapses history, multiple user turns occur back-to-back.
+// Anthropic strictly throws: 400 invalid_request_error: 'roles must alternate'
+// LLM Switcher Healer Engine: Merges consecutive same-role turns -> 200 OK
+// -----------------------------------------------------------------------------
+await runTest('Headroom Simulation: Consecutive User turns (Role alternation violation)', async () => {
+  const payload = {
+    model: 'claude-opus-4-6',
+    max_tokens: 30,
+    messages: [
+      { role: 'user', content: 'User message turn 1' },
+      { role: 'user', content: 'User message turn 2 (no assistant in between)' },
+      { role: 'user', content: 'User message turn 3: reply pong' }
+    ],
+    stream: false
+  };
+
+  const res = await fetch(`${BASE_URL}/v1/messages`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json', 'anthropic-version': '2023-06-01' },
+    body: JSON.stringify(payload)
+  });
+
+  if (res.status !== 200) {
+    const err = await res.text();
+    throw new Error(`Expected HTTP 200, got ${res.status}: ${err}`);
+  }
+  const json = await res.json();
+  return `Merged consecutive turns seamlessly. Stop reason: ${json.stop_reason}`;
+});
+
+// -----------------------------------------------------------------------------
+// TEST 3: Headroom / RTK Failure Mode — Stripped Thinking Parameter
+// Optimizer stripped the 'thinking' object to reduce tokens.
+// LLM Switcher Thinking Guard: Detects reasoning model (ag/claude-opus-4-6-thinking)
+// and restores thinking.budget_tokens automatically -> thinking_delta emitted!
+// -----------------------------------------------------------------------------
+await runTest('Thinking Guard: Restoring stripped thinking parameter on reasoning models', async () => {
+  const payload = {
+    model: 'claude-opus-4-6',
+    max_tokens: 200,
+    // Note: NO thinking parameter included (simulating optimizer stripping it)
+    messages: [
+      { role: 'user', content: 'Solve step by step: what is 17 * 23? Show reasoning.' }
+    ],
+    stream: true
+  };
+
+  const res = await fetch(`${BASE_URL}/v1/messages`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json', 'anthropic-version': '2023-06-01' },
+    body: JSON.stringify(payload)
+  });
+
+  if (res.status !== 200) {
+    throw new Error(`Expected HTTP 200, got ${res.status}`);
+  }
+
+  const text = await res.text();
+  const hasThinkingDelta = text.includes('thinking_delta');
+  const hasSignatureDelta = text.includes('signature_delta');
+  const hasTextDelta = text.includes('text_delta');
+
+  if (!hasThinkingDelta) {
+    throw new Error('Thinking Guard failed: thinking_delta was NOT emitted in stream!');
+  }
+
+  return `Automatically restored thinking: thinking_delta=${hasThinkingDelta}, signature_delta=${hasSignatureDelta}, text_delta=${hasTextDelta}`;
+});
+
+// -----------------------------------------------------------------------------
+// TEST 4: RTK & Intermediary Headers Passthrough
+// RTK / tracing tools inject headers: x-rtk-version, traceparent, x-request-id.
+// LLM Switcher: Transparently preserves all tracking headers without rejection.
+// -----------------------------------------------------------------------------
+await runTest('RTK Intermediary: Custom headers and traceparent passthrough', async () => {
+  const payload = {
+    model: 'claude-opus-4-6',
+    max_tokens: 20,
+    messages: [{ role: 'user', content: 'ping' }],
+    stream: false
+  };
+
+  const res = await fetch(`${BASE_URL}/v1/messages`, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      'anthropic-version': '2023-06-01',
+      'x-rtk-version': '0.37.2',
+      'x-optimizer-id': 'rtk-cli-hook',
+      'traceparent': '00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-01'
+    },
+    body: JSON.stringify(payload)
+  });
+
+  if (res.status !== 200) {
+    throw new Error(`Expected HTTP 200, got ${res.status}`);
+  }
+  return `Headers accepted cleanly with HTTP 200 OK`;
+});
+
+// -----------------------------------------------------------------------------
+// TEST 5: OpenAI Chat Healer Mode — Orphaned Tool Result in Chat Completions
+// Same test but over /v1/chat/completions: tool role message without prior tool_calls assistant message.
+// OpenAI API strictly throws: 400 'messages with role tool must be a response to a preceding message with tool_calls'
+// LLM Switcher: Converts orphaned tool to user text message -> 200 OK
+// -----------------------------------------------------------------------------
+await runTest('OpenAI Chat Healer: Orphaned role tool without preceding assistant tool_calls', async () => {
+  const payload = {
+    model: 'ag/gpt-oss-120b-medium',
+    max_tokens: 30,
+    messages: [
+      { role: 'user', content: 'Turn 1 context' },
+      { role: 'tool', tool_call_id: 'orphaned_chat_call_77', content: 'Tool execution logs: OK' },
+      { role: 'user', content: 'reply pong' }
+    ],
+    stream: false
+  };
+
+  const res = await fetch(`${BASE_URL}/v1/chat/completions`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify(payload)
+  });
+
+  if (res.status !== 200) {
+    const err = await res.text();
+    throw new Error(`Expected HTTP 200, got ${res.status}: ${err}`);
+  }
+  const json = await res.json();
+  return `Chat Healer rescued orphaned tool role. Stop reason: ${json.choices?.[0]?.finish_reason}`;
+});
+
+console.log('\n================================================================');
+console.log(`  TEST RESULTS: ${passed} PASSED / ${failed} FAILED               `);
+console.log('================================================================\n');
+
+process.exit(failed ? 1 : 0);

package/tests/service.test.mjs

@@ -28,7 +28,7 @@ test('systemd unit quotes the command and carries the environment', () => {
 });
 
 const plist = () => launchdPlist({ nodeBin: '/n', script: '/a&b/proxy.mjs', port: 4001, logPath: '/l<og>', env: [['CLAUDE_CONFIG_DIR', '/x&y']] });
-const taskXml = () => scheduledTaskXml({ nodeBin: 'C:\\Program Files\\nodejs\\node.exe', script: 'C:\\Users\\A & B\\llm-switcher\\proxy.mjs', port: 4002, userId: 'PC\\a&b' });
+const taskXml = () => scheduledTaskXml({ nodeBin: 'C:\\Program Files\\nodejs\\node.exe', script: 'C:\\Users\\A & B\\llm-switcher\\proxy.mjs', port: 4002, userId: 'PC\\a&b', systemRoot: 'C:\\Windows' });
 
 test('the plist and the task XML are well-formed', { skip: !HAS_XMLLINT && 'needs xmllint' }, () => {
   execFileSync('xmllint', ['--noout', '-'], { input: plist() });
@@ -43,8 +43,10 @@ test('launchd plist escapes values and carries the environment', () => {
 
 test('Windows task XML keeps the command and its arguments apart and has no run-time limit', () => {
   const xml = taskXml();
-  assert.match(xml, /<Command>C:\\Program Files\\nodejs\\node.exe<\/Command>/);
-  assert.match(xml, /<Arguments>&quot;C:\\Users\\A &amp; B\\llm-switcher\\proxy.mjs&quot; --port 4002<\/Arguments>/);
+  // node.exe is a console program: started directly at logon it opens a terminal window, and closing
+  // that window stops the gateway. conhost --headless gives it a console that nobody sees.
+  assert.match(xml, /<Command>C:\\Windows\\System32\\conhost.exe<\/Command>/);
+  assert.match(xml, /<Arguments>--headless &quot;C:\\Program Files\\nodejs\\node.exe&quot; &quot;C:\\Users\\A &amp; B\\llm-switcher\\proxy.mjs&quot; --port 4002<\/Arguments>/);
   assert.match(xml, /<ExecutionTimeLimit>PT0S<\/ExecutionTimeLimit>/);
   assert.match(xml, /<LogonTrigger>\s*<Enabled>true<\/Enabled>\s*<UserId>PC\\a&amp;b<\/UserId>/);
   assert.equal(portFromServiceText(xml), 4002);
@@ -60,7 +62,7 @@ test('the port is read back from UTF-16 console output, as schtasks /Query /XML
 
 test('autoupdate flag is placed into task XML, systemd unit, and launchd plist', () => {
   const xml = scheduledTaskXml({ nodeBin: 'node.exe', script: 'proxy.mjs', port: 3456, userId: 'u', autoupdate: true });
-  assert.match(xml, /<Arguments>&quot;proxy\.mjs&quot; --port 3456 --autoupdate<\/Arguments>/);
+  assert.match(xml, /<Arguments>--headless &quot;node\.exe&quot; &quot;proxy\.mjs&quot; --port 3456 --autoupdate<\/Arguments>/);
   assert.equal(portFromServiceText(xml), 3456);
 
   const unit = systemdUnit({ nodeBin: '/usr/bin/node', script: '/opt/proxy.mjs', port: 3456, autoupdate: true });

package/tests/state.test.mjs

@@ -797,6 +797,7 @@ test('A5: migration splits shared auto profile into -claude and -codex halves wi
         name: 'Shared Profile',
         apiKey: 'secret-key',
         defaultModels: { opus: 'opus-model', sonnet: 'sonnet-model', subagent: 'sub-model' },
+        // The retired 1M flags are in an old config: the split must not carry them into a half.
         model1M: { opus: true, sonnet: false }
       }
     }
@@ -819,10 +820,8 @@ test('A5: migration splits shared auto profile into -claude and -codex halves wi
   assert.equal(codex.defaultModels.main, 'opus-model');
   assert.equal(codex.defaultModels.review, 'sonnet-model');
   assert.equal(codex.defaultModels.subagent, 'sub-model');
-  assert.equal(claude.model1M.opus, true);
-  assert.equal(claude.model1M.sonnet, false);
-  assert.equal(codex.model1M.main, true);
-  assert.equal(codex.model1M.review, false);
+  assert.equal(claude.model1M, undefined, 'the retired 1M flags do not survive the split');
+  assert.equal(codex.model1M, undefined);
 });
 
 test('A5: R7 row 4 splits profile with Claude/no-tool fields active for responses with slot fallbacks', () => {
@@ -881,7 +880,7 @@ test('A5: migration preserves legacy "default" key on both halves during split',
   assert.equal(res.config.profiles['combo-codex'].defaultModels.default, 'chat-default');
 });
 
-test('A5: migration never fills slot whose value is empty string or false', () => {
+test('A5: migration never fills slot whose value is empty string', () => {
   const raw = {
     activeProfile: 'combo',
     profiles: {
@@ -895,7 +894,7 @@ test('A5: migration never fills slot whose value is empty string or false', () =
   const res = migrateConfigInMemory(raw);
   const codex = res.config.profiles['combo-codex'];
   assert.equal(codex.defaultModels.main, '');
-  assert.equal(codex.model1M.main, false);
+  assert.equal(codex.model1M, undefined, 'and the retired flags do not come along');
 });
 
 test('A5: migration splits renamed command-word profile as <key>-profile-claude and <key>-profile-codex', () => {
@@ -1610,7 +1609,7 @@ test('each routed tool gets a route line; a tool that is off gets an empty one',
   assert.match(r.on.claude, /\bcl\b/, 'the line names the profile that serves claude');
   assert.match(r.on.claude, /intact\.example\.io/, 'and the host the traffic goes to');
   assert.match(r.on.claude, /gemini-3\.8-flash/, 'and the model it maps to');
-  assert.match(r.on.claude, /1M/, 'and the 1M window, because that changes what the session costs');
+  assert.doesNotMatch(r.on.claude, /1M/, 'and no window: the tool sizes the session from the model');
   assert.match(r.on.codex, /\bcx\b/);
   assert.match(r.on.codex, /other\.example\.io/);
   assert.doesNotMatch(r.on.codex, /intact\.example\.io/, 'one tool never shows the other tool route');