backend-manager 5.6.4 → 5.7.0

package/src/manager/libraries/ai/providers/openai.js

@@ -335,12 +335,12 @@ function OpenAI(assistant, key) {
   const self = this;
 
   self.assistant = assistant;
-  self.Manager = assistant.Manager;
-  self.user = assistant.user;
+  self.Manager = assistant?.Manager;
+  self.user = assistant?.user;
   self.key = key
-    || self.Manager.config?.openai?.key
-    || self.Manager.config?.openai?.global
-    || self.Manager.config?.openai?.main
+    || self.Manager?.config?.openai?.key
+    || self.Manager?.config?.openai?.global
+    || self.Manager?.config?.openai?.main
     || process.env.OPENAI_API_KEY
     || process.env.BACKEND_MANAGER_OPENAI_API_KEY
 
@@ -451,12 +451,19 @@ OpenAI.prototype.request = function (options) {
     _log('Starting', options);
 
 
+    // Direct-messages mode: when a unified messages[] array is passed (incl.
+    // assistant toolCalls turns + role:'tool' results), it IS the full
+    // conversation — prompt/message/history are ignored and the array maps
+    // straight to the Responses API input (see formatMessages). The last user
+    // turn's text still feeds moderation.
+    const useMessages = Array.isArray(options.messages) && options.messages.length > 0;
+
     // Load prompt segments (one entry per role) and the user message
-    const promptSegments = options.prompt.map((segment) => ({
+    const promptSegments = useMessages ? [] : options.prompt.map((segment) => ({
       role: segment.role,
       content: loadContent(segment, _log),
     }));
-    const message = loadContent(options.message, _log);
+    const message = useMessages ? lastUserText(options.messages) : loadContent(options.message, _log);
     const user = options.user?.auth?.uid || assistant.request.geolocation.ip || 'unknown';
 
     // Log
@@ -477,9 +484,10 @@ OpenAI.prototype.request = function (options) {
       return reject(assistant.errorify(`Error loading message: ${message}`, {code: 400}));
     }
 
-    // Moderate if needed
+    // Moderate if needed (skipped in direct-messages mode when the last turn
+    // carries no user text — e.g. a tool-result continuation turn)
     let moderation = null;
-    if (options.moderate) {
+    if (options.moderate && !(useMessages && !message)) {
       moderation = await makeRequest('moderations', options, self, promptSegments, message, user, _log)
       .then(async (r) => {
         // {
@@ -883,6 +891,112 @@ function formatHistory(options, promptSegments, message, _log) {
   return formatted;
 }
 
+/**
+ * Map a unified messages[] array straight to the Responses API input array.
+ *
+ * Unified turn shapes:
+ *   - { role: 'system'|'developer'|'user'|'assistant', content: string }
+ *   - { role: 'assistant', content?, toolCalls: [{ id, name, arguments }] }
+ *     → message item (if content) + function_call items
+ *   - { role: 'tool', toolCallId, content } → function_call_output item
+ */
+function formatMessages(messages, _log) {
+  const input = [];
+
+  for (const m of messages) {
+    // Tool result turn → function_call_output item
+    if (m.role === 'tool') {
+      input.push({
+        type: 'function_call_output',
+        call_id: m.toolCallId,
+        output: typeof m.content === 'string' ? m.content : JSON.stringify(m.content || ''),
+      });
+      continue;
+    }
+
+    // Assistant turn with tool calls → message item (if text) + function_call items
+    if (m.role === 'assistant' && Array.isArray(m.toolCalls) && m.toolCalls.length) {
+      const text = typeof m.content === 'string' ? m.content.trim() : '';
+
+      if (text) {
+        input.push({
+          role: 'assistant',
+          content: formatMessageContent(text, [], _log, 'responses', 'assistant'),
+        });
+      }
+
+      for (const call of m.toolCalls) {
+        input.push({
+          type: 'function_call',
+          call_id: call.id,
+          name: call.name,
+          arguments: typeof call.arguments === 'string' ? call.arguments : JSON.stringify(call.arguments || {}),
+        });
+      }
+
+      continue;
+    }
+
+    // Plain text turn
+    const role = m.role || 'user';
+    const content = typeof m.content === 'string' ? m.content.trim() : String(m.content || '');
+
+    input.push({
+      role: role,
+      content: formatMessageContent(content, m.attachments, _log, 'responses', role),
+    });
+  }
+
+  return input;
+}
+
+// Last user turn's text — feeds moderation in direct-messages mode
+function lastUserText(messages) {
+  const lastUser = [...messages].reverse().find((m) => m.role === 'user' && typeof m.content === 'string');
+  return lastUser?.content || '';
+}
+
+// Normalized function tools ({ name, description, parameters }) get the
+// Responses API envelope; anything carrying another `type` passes verbatim
+// (hosted tools like { type: 'web_search' })
+function normalizeToolEntry(tool) {
+  if (tool && tool.name && (!tool.type || tool.type === 'function')) {
+    return {
+      type: 'function',
+      name: tool.name,
+      description: tool.description || '',
+      parameters: tool.parameters || { type: 'object', properties: {} },
+    };
+  }
+
+  return tool;
+}
+
+// 'auto' | 'required' | 'none' pass through; { name } → a specific function tool
+function normalizeToolChoice(choice) {
+  if (typeof choice === 'object' && choice?.name) {
+    return { type: 'function', name: choice.name };
+  }
+
+  return choice;
+}
+
+function parseArguments(args) {
+  if (args && typeof args === 'object') {
+    return args;
+  }
+
+  if (typeof args === 'string' && args.trim()) {
+    try {
+      return JSON5.parse(args);
+    } catch (e) {
+      return {};
+    }
+  }
+
+  return {};
+}
+
 function attemptRequest(options, self, promptSegments, message, user, moderation, attempt, assistant, resolve, reject, _log) {
   const retries = options.retries;
   const triggers = options.retryTriggers;
@@ -961,6 +1075,14 @@ function attemptRequest(options, self, promptSegments, message, user, moderation
       .join('\n')
       .trim();
 
+    // Normalized tool calls (Responses API function_call items) + stop reason
+    const toolCalls = output
+      .filter((o) => o.type === 'function_call')
+      .map((o) => ({ id: o.call_id, name: o.name, arguments: parseArguments(o.arguments) }));
+    const stopReason = toolCalls.length
+      ? 'tool_use'
+      : (r.status === 'incomplete' && r.incomplete_details?.reason === 'max_output_tokens' ? 'max_tokens' : 'end');
+
     // Get model configuration
     const modelConfig = getModelConfig(options.model);
 
@@ -979,9 +1101,10 @@ function attemptRequest(options, self, promptSegments, message, user, moderation
     _log('Response', outputText.length, typeof outputText, outputText);
     _log('Tokens', self.tokens);
 
-    // Try to parse JSON response if needed
+    // Try to parse JSON response if needed — never on a tool-call turn, where
+    // empty text is the normal intermediate state (the caller continues the loop)
     try {
-      const parsed = options.response === 'json' ? JSON5.parse(outputText) : outputText;
+      const parsed = options.response === 'json' && !toolCalls.length ? JSON5.parse(outputText) : outputText;
 
       // Return
       return resolve({
@@ -989,6 +1112,9 @@ function attemptRequest(options, self, promptSegments, message, user, moderation
         content: parsed,
         tokens: self.tokens,
         moderation: moderation,
+        raw: r,
+        toolCalls: toolCalls,
+        stopReason: stopReason,
       })
     } catch (e) {
       assistant.error('Error parsing response', r, e);
@@ -1055,8 +1181,12 @@ function makeRequest(mode, options, self, promptSegments, message, user, _log) {
         user: user,
       }
     } else if (mode === 'responses') {
-      // Format history for responses API
-      const history = formatHistory(options, promptSegments, message, _log);
+      // Format input for the Responses API — direct-messages mode maps the
+      // unified messages[] straight through; legacy mode builds from
+      // prompt segments + history + message
+      const history = Array.isArray(options.messages) && options.messages.length
+        ? formatMessages(options.messages, _log)
+        : formatHistory(options, promptSegments, message, _log);
 
       // Set request
       request.url = 'https://api.openai.com/v1/responses';
@@ -1080,15 +1210,17 @@ function makeRequest(mode, options, self, promptSegments, message, user, _log) {
         request.body.reasoning = reasoning;
       }
 
-      // Only include tools if `tools.list` is a non-empty array. When present, the
-      // response output may contain tool-call items (e.g. web_search_call)
-      // alongside the message — the message extractor below already ignores
-      // non-message items, so this is purely additive.
+      // Only include tools if `tools.list` is a non-empty array. Normalized
+      // function tools ({ name, description, parameters }) get the Responses
+      // `type: 'function'` envelope; hosted tools (web_search, code_interpreter)
+      // pass verbatim. When present, the response output may contain tool-call
+      // items alongside the message — function_call items are extracted into
+      // the normalized `toolCalls` return field.
       if (Array.isArray(options.tools?.list) && options.tools.list.length) {
-        request.body.tools = options.tools.list;
+        request.body.tools = options.tools.list.map(normalizeToolEntry);
 
         if (options.tools.choice) {
-          request.body.tool_choice = options.tools.choice;
+          request.body.tool_choice = normalizeToolChoice(options.tools.choice);
         }
       }
     }
@@ -1186,5 +1318,8 @@ module.exports = OpenAI;
 module.exports._internals = {
   normalizePrompt,
   formatHistory,
+  formatMessages,
+  normalizeToolEntry,
+  normalizeToolChoice,
   VALID_PROMPT_ROLES,
 };

package/src/manager/libraries/ai/providers/test.js

@@ -0,0 +1,242 @@
+/**
+ * Test provider — a deterministic, first-class AI provider for test suites
+ * (the `test` payment-processor precedent: a real implementation of the
+ * provider seam, not a mock injected into callers).
+ *
+ * REFUSES to run in production — only development/testing environments.
+ *
+ * Behavior is scripted by directives embedded in the LAST user message. The
+ * directives form a SEQUENCE consumed across the turns of a tool loop: call N
+ * executes directive N-1 (indexed by how many assistant turns follow the last
+ * user turn in options.messages). Directive content must not contain `]]`.
+ *
+ *   [[tool:check_order {"orderNumber":"123"}]]   — one tool call this step
+ *   [[tools:[{"name":"a","arguments":{}},{"name":"b","arguments":{}}]]
+ *                                                 — parallel tool calls this step
+ *   [[reply:{"message":"done"}]]                  — final reply (JSON or text)
+ *   [[delay:500]]                                 — modifier: delay the NEXT step
+ *   [[error:boom]]                                — throw at this step
+ *
+ * No directives (or script exhausted) → echo reply: `Echo: <text>` (wrapped as
+ * { message } when options.response === 'json').
+ *
+ * Returns the same shape as the real providers:
+ *   { content, output, tokens, raw, toolCalls, stopReason }
+ */
+const JSON5 = require('json5');
+
+// The closing `]]` must not be followed by another `]` so directive values may
+// END with a JSON `]` (e.g. [[tools:[...]]]); `]]` strictly INSIDE a value is
+// still unsupported
+const DIRECTIVE_REGEX = /\[\[(tool|tools|reply|delay|error)(?::([\s\S]*?))?\]\](?!\])/g;
+
+function TestProvider(assistant, key) {
+  const self = this;
+
+  self.assistant = assistant;
+  self.Manager = assistant?.Manager;
+  self.key = key || 'test';
+
+  self.tokens = {
+    total:  { count: 0, price: 0 },
+    input:  { count: 0, price: 0 },
+    output: { count: 0, price: 0 },
+  };
+
+  return self;
+}
+
+TestProvider.prototype.request = async function (options) {
+  const self = this;
+
+  assertAllowedEnvironment(self.Manager);
+
+  options = options || {};
+
+  const messages = Array.isArray(options.messages) ? options.messages : [];
+  const lastUserMessage = [...messages].reverse().find((m) => m.role === 'user' && typeof m.content === 'string');
+  const scriptSource = lastUserMessage?.content || stringifyLoose(options.message?.content) || '';
+
+  const { steps, cleanText } = parseScript(scriptSource);
+
+  // Which step of the script is this call? One assistant turn is appended per
+  // loop iteration, so call N sees N-1 assistant turns after the last user turn.
+  const lastUserIdx = messages.lastIndexOf(lastUserMessage);
+  const stepIndex = messages.slice(lastUserIdx + 1).filter((m) => m.role === 'assistant').length;
+
+  const step = steps[stepIndex] || { type: 'echo' };
+
+  // Simulated token accounting so usage/cost paths execute
+  const inputChars = messages.reduce((n, m) => n + stringifyLoose(m.content).length, 0) || scriptSource.length;
+  self.tokens.input.count += Math.ceil(inputChars / 4);
+
+  if (step.delay) {
+    await new Promise((resolve) => setTimeout(resolve, Math.min(step.delay, 30000)));
+  }
+
+  if (step.type === 'error') {
+    throw new Error(step.message || 'Test provider scripted error');
+  }
+
+  if (step.type === 'tools') {
+    const toolCalls = step.calls.map((call, i) => ({
+      id: `call_test_${stepIndex}_${i}`,
+      name: call.name,
+      arguments: call.arguments || {},
+    }));
+
+    self.tokens.output.count += 10 * toolCalls.length;
+    finalizeTokens(self.tokens);
+
+    return {
+      content: '',
+      output: [],
+      tokens: self.tokens,
+      raw: { provider: 'test', step: stepIndex, toolCalls },
+      toolCalls,
+      stopReason: 'tool_use',
+    };
+  }
+
+  // Final reply — scripted or echo
+  let text;
+
+  if (step.type === 'reply') {
+    text = step.content;
+  } else {
+    text = `Echo: ${cleanText || '(empty)'}`;
+
+    if (options.response === 'json') {
+      text = JSON.stringify({ message: text });
+    }
+  }
+
+  self.tokens.output.count += Math.ceil(text.length / 4);
+  finalizeTokens(self.tokens);
+
+  let parsed = text;
+
+  if (options.response === 'json') {
+    try {
+      parsed = JSON5.parse(text);
+    } catch (e) {
+      // Loose by design — a scripted plain-text reply stays a string
+    }
+  }
+
+  return {
+    content: parsed,
+    output: [{ type: 'text', text }],
+    tokens: self.tokens,
+    raw: { provider: 'test', step: stepIndex },
+    toolCalls: [],
+    stopReason: 'end',
+  };
+};
+
+/**
+ * Parse the directive script out of a message. Returns the ordered steps and
+ * the message text with directives stripped (the echo source).
+ */
+function parseScript(rawSource) {
+  // Consumers may markdown-escape user input before it reaches the provider
+  // (\[\[tool:...\]\]) — unescape so directives still parse
+  const source = String(rawSource || '').replace(/\\([\\`*_{}[\]()#+\-.!~|>])/g, '$1');
+
+  const steps = [];
+  let pendingDelay = 0;
+  let match;
+
+  DIRECTIVE_REGEX.lastIndex = 0;
+
+  while ((match = DIRECTIVE_REGEX.exec(source)) !== null) {
+    const [, type, value] = match;
+
+    if (type === 'delay') {
+      pendingDelay = parseInt(value, 10) || 0;
+      continue;
+    }
+
+    const step = buildStep(type, value);
+    step.delay = pendingDelay;
+    pendingDelay = 0;
+
+    steps.push(step);
+  }
+
+  // Trailing delay with no following directive → delay the default echo
+  if (pendingDelay) {
+    steps.push({ type: 'echo', delay: pendingDelay });
+  }
+
+  const cleanText = source.replace(DIRECTIVE_REGEX, '').trim();
+
+  return { steps, cleanText };
+}
+
+function buildStep(type, value) {
+  if (type === 'error') {
+    return { type: 'error', message: (value || '').trim() };
+  }
+
+  if (type === 'reply') {
+    return { type: 'reply', content: (value || '').trim() };
+  }
+
+  if (type === 'tools') {
+    return { type: 'tools', calls: JSON5.parse(value) };
+  }
+
+  // tool:name {json}
+  const trimmed = (value || '').trim();
+  const spaceIdx = trimmed.search(/\s/);
+  const name = spaceIdx < 0 ? trimmed : trimmed.slice(0, spaceIdx);
+  const argsRaw = spaceIdx < 0 ? '' : trimmed.slice(spaceIdx).trim();
+
+  return {
+    type: 'tools',
+    calls: [{ name, arguments: argsRaw ? JSON5.parse(argsRaw) : {} }],
+  };
+}
+
+function assertAllowedEnvironment(Manager) {
+  // The Manager's environment detection is the SSOT when in scope
+  if (Manager && typeof Manager.isDevelopment === 'function' && typeof Manager.isTesting === 'function') {
+    if (Manager.isDevelopment() || Manager.isTesting()) {
+      return;
+    }
+
+    throw new Error('AI test provider is only available in development or testing environments');
+  }
+
+  // No Manager (pure unit tests) — allow only under explicit test/emulator signals
+  if (process.env.BEM_TESTING === 'true' || process.env.FUNCTIONS_EMULATOR) {
+    return;
+  }
+
+  throw new Error('AI test provider is only available in development or testing environments');
+}
+
+function finalizeTokens(tokens) {
+  tokens.total.count = tokens.input.count + tokens.output.count;
+}
+
+function stringifyLoose(content) {
+  if (typeof content === 'string') {
+    return content;
+  }
+
+  if (Array.isArray(content)) {
+    return content.map((c) => c?.text || '').join('\n');
+  }
+
+  return content ? String(content) : '';
+}
+
+module.exports = TestProvider;
+
+// Exposed for unit tests. Not part of the public API.
+module.exports._internals = {
+  parseScript,
+  buildStep,
+};