llm-switcher 1.2.6 → 1.2.8

package/tests/gateway.e2e.test.mjs

@@ -470,7 +470,7 @@ test('save-profile rejects a model name that could act as a command', async () =
   assert.equal(ok.status, 200);
 });
 
-test('Codex WS transport: upstream 429 becomes response.failed with rate_limit_exceeded', async () => {  const ws = new WebSocket(`ws://127.0.0.1:${proxyPort}/v1/responses`);
+test('Codex WS transport: upstream 429 becomes response.failed with rate_limit_exceeded', { skip: typeof WebSocket === 'undefined' && 'global WebSocket needs Node 22+' }, async () => {  const ws = new WebSocket(`ws://127.0.0.1:${proxyPort}/v1/responses`);
   await new Promise((resolve, reject) => {
     const t = setTimeout(() => reject(new Error('ws open timeout')), 5000);
     ws.addEventListener('open', () => { clearTimeout(t); resolve(); }, { once: true });

package/tests/helpers.mjs

@@ -1,24 +1,24 @@
-import assert from 'node:assert/strict';
-
-// Validate an Anthropic event sequence: indexes increase in start order, deltas only target an open block.
-export function assertValidAnthropicEvents(events) {
-  const open = new Set();
-  const seen = new Set();
-  let expectedNext = 0;
-  for (const { event, data } of events) {
-    if (event === 'content_block_start') {
-      assert.equal(data.index, expectedNext, `block index must be sequential (got ${data.index}, want ${expectedNext})`);
-      assert.ok(!seen.has(data.index), `index ${data.index} reused`);
-      seen.add(data.index);
-      open.add(data.index);
-      expectedNext++;
-    } else if (event === 'content_block_delta') {
-      assert.ok(open.has(data.index), `delta for block ${data.index} which is not open`);
-    } else if (event === 'content_block_stop') {
-      assert.ok(open.has(data.index), `stop for block ${data.index} which is not open`);
-      open.delete(data.index);
-    } else if (event === 'message_stop') {
-      assert.equal(open.size, 0, 'all blocks must be closed before message_stop');
-    }
-  }
-}
+import assert from 'node:assert/strict';
+
+// Validate an Anthropic event sequence: indexes increase in start order, deltas only target an open block.
+export function assertValidAnthropicEvents(events) {
+  const open = new Set();
+  const seen = new Set();
+  let expectedNext = 0;
+  for (const { event, data } of events) {
+    if (event === 'content_block_start') {
+      assert.equal(data.index, expectedNext, `block index must be sequential (got ${data.index}, want ${expectedNext})`);
+      assert.ok(!seen.has(data.index), `index ${data.index} reused`);
+      seen.add(data.index);
+      open.add(data.index);
+      expectedNext++;
+    } else if (event === 'content_block_delta') {
+      assert.ok(open.has(data.index), `delta for block ${data.index} which is not open`);
+    } else if (event === 'content_block_stop') {
+      assert.ok(open.has(data.index), `stop for block ${data.index} which is not open`);
+      open.delete(data.index);
+    } else if (event === 'message_stop') {
+      assert.equal(open.size, 0, 'all blocks must be closed before message_stop');
+    }
+  }
+}

package/tests/hook-status.test.mjs

@@ -0,0 +1,119 @@
+// The launch hook: what a coding tool shows the person about the switcher at the start of a session.
+//
+// The shim toast fires only when the shim runs, and it needs the shim directory first on PATH. A
+// hook runs inside the tool itself, so it reports even when the launcher is the person's own script.
+// Both tools read one JSON object from stdout and show `systemMessage` to the person.
+import { test } from 'node:test';
+import assert from 'node:assert/strict';
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+import http from 'node:http';
+import { execFileSync } from 'node:child_process';
+import { fileURLToPath, pathToFileURL } from 'node:url';
+
+const ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..');
+const HOOK = path.join(ROOT, 'hook-status.mjs');
+const { identityProof } = await import(pathToFileURL(path.join(ROOT, 'state.mjs')).href);
+
+function tmpState(t) {
+  const dir = fs.mkdtempSync(path.join(os.tmpdir(), 'llmsw-hook-'));
+  t.after(() => fs.rmSync(dir, { recursive: true, force: true }));
+  // The proof is an HMAC over the admin token, so the fake gateway needs the same file.
+  fs.writeFileSync(path.join(dir, 'admin.token'), 'tok-for-the-hook-test', { mode: 0o600 });
+  return dir;
+}
+
+function writeConfig(dir, port) {
+  fs.writeFileSync(path.join(dir, 'config.json'), JSON.stringify({
+    port,
+    activeProfiles: { claude: 'cl', codex: null },
+    profiles: {
+      cl: {
+        name: 'Intact Claude', baseURL: 'https://intact.example.io/v1', apiKey: 'sk-secret',
+        defaultModels: { opus: 'gemini-3.8-flash' }, model1M: { opus: true }
+      }
+    }
+  }), { mode: 0o600 });
+}
+
+// A gateway of this build: it answers /health with a proof over the challenge.
+function fakeGateway(t, port) {
+  const srv = http.createServer((req, res) => {
+    const nonce = new URL(req.url, 'http://127.0.0.1').searchParams.get('challenge') || '';
+    const body = { proxy: 'llm-switcher', port, pid: process.pid };
+    body.proof = identityProof(nonce, { role: 'gateway', port, pid: body.pid }, 'tok-for-the-hook-test');
+    res.writeHead(200, { 'content-type': 'application/json' });
+    res.end(JSON.stringify(body));
+  });
+  return new Promise((resolve) => srv.listen(port, '127.0.0.1', () => {
+    t.after(() => new Promise(r => srv.close(r)));
+    resolve(srv);
+  }));
+}
+
+function freePort() {
+  return new Promise((resolve) => {
+    const s = http.createServer();
+    s.listen(0, '127.0.0.1', () => { const p = s.address().port; s.close(() => resolve(p)); });
+  });
+}
+
+function runHook(dir, tool) {
+  const out = execFileSync(process.execPath, [HOOK, tool], {
+    encoding: 'utf8', timeout: 20000,
+    env: {
+      ...process.env,
+      LLM_SWITCHER_STATE_DIR: dir,
+      LLM_SWITCHER_CONFIG: path.join(dir, 'config.json'),
+      LLM_SWITCHER_PORT: ''
+    }
+  });
+  return { raw: out, json: JSON.parse(out.trim() || '{}') };
+}
+
+test('a tool that is not routed gets no message at all', async (t) => {
+  const dir = tmpState(t);
+  writeConfig(dir, await freePort());
+  // codex has no profile, so its route file is empty. This is the state after `switch off codex`.
+  fs.writeFileSync(path.join(dir, 'active.flag'), 'active');
+  fs.writeFileSync(path.join(dir, 'route-codex.txt'), '');
+  const r = runHook(dir, 'codex');
+  assert.equal(r.json.systemMessage, undefined, 'a tool on its official endpoint says nothing');
+});
+
+test('a routed tool with a live gateway reports where its traffic goes', async (t) => {
+  const dir = tmpState(t);
+  const port = await freePort();
+  writeConfig(dir, port);
+  await fakeGateway(t, port);
+  fs.writeFileSync(path.join(dir, 'active.flag'), 'active');
+  fs.writeFileSync(path.join(dir, 'route-claude.txt'), 'claude -> cl | intact.example.io | gemini-3.8-flash | 1M\n');
+  const r = runHook(dir, 'claude');
+  assert.match(r.json.systemMessage, /LLM Switcher/, 'the message names the switcher');
+  assert.match(r.json.systemMessage, /intact\.example\.io/, 'and the host the traffic goes to');
+  assert.match(r.json.systemMessage, /gemini-3\.8-flash/, 'and the model');
+  assert.doesNotMatch(r.json.systemMessage, /sk-secret/, 'and never the API key');
+});
+
+test('a routed tool with a dead gateway is warned that it cannot reach the provider', async (t) => {
+  const dir = tmpState(t);
+  const port = await freePort();   // nothing listens on it
+  writeConfig(dir, port);
+  fs.writeFileSync(path.join(dir, 'active.flag'), 'active');
+  fs.writeFileSync(path.join(dir, 'route-claude.txt'), 'claude -> cl | intact.example.io | gemini-3.8-flash | 1M\n');
+  const r = runHook(dir, 'claude');
+  assert.match(r.json.systemMessage, /does not answer|not running/i, 'the message says the gateway is down');
+  assert.match(r.json.systemMessage, new RegExp(String(port)), 'and names the port');
+  assert.match(r.json.systemMessage, /switch on/, 'and names the command that fixes it');
+});
+
+test('a broken state never fails the session: exit 0 and parseable output', async (t) => {
+  const dir = tmpState(t);
+  fs.writeFileSync(path.join(dir, 'config.json'), '{ this is not json', { mode: 0o600 });
+  fs.writeFileSync(path.join(dir, 'active.flag'), 'active');
+  const r = runHook(dir, 'claude');
+  assert.equal(typeof r.json, 'object', 'the tool still gets one JSON object');
+  const bad = runHook(dir, 'not-a-tool');
+  assert.equal(bad.json.systemMessage, undefined, 'an unknown tool name says nothing');
+});

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);