@illuma-ai/agents 1.1.20 → 1.1.22

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (246) hide show
  1. package/dist/cjs/graphs/Graph.cjs +12 -1
  2. package/dist/cjs/graphs/Graph.cjs.map +1 -1
  3. package/dist/cjs/graphs/MultiAgentGraph.cjs +85 -1
  4. package/dist/cjs/graphs/MultiAgentGraph.cjs.map +1 -1
  5. package/dist/cjs/llm/bedrock/index.cjs +14 -0
  6. package/dist/cjs/llm/bedrock/index.cjs.map +1 -1
  7. package/dist/cjs/run.cjs +20 -9
  8. package/dist/cjs/run.cjs.map +1 -1
  9. package/dist/esm/graphs/Graph.mjs +12 -1
  10. package/dist/esm/graphs/Graph.mjs.map +1 -1
  11. package/dist/esm/graphs/MultiAgentGraph.mjs +85 -1
  12. package/dist/esm/graphs/MultiAgentGraph.mjs.map +1 -1
  13. package/dist/esm/llm/bedrock/index.mjs +14 -0
  14. package/dist/esm/llm/bedrock/index.mjs.map +1 -1
  15. package/dist/esm/run.mjs +20 -9
  16. package/dist/esm/run.mjs.map +1 -1
  17. package/dist/types/graphs/MultiAgentGraph.d.ts +17 -0
  18. package/package.json +1 -1
  19. package/src/graphs/Graph.ts +12 -1
  20. package/src/graphs/MultiAgentGraph.ts +105 -1
  21. package/src/graphs/__tests__/multi-agent-delegate.test.ts +191 -0
  22. package/src/llm/bedrock/index.ts +17 -0
  23. package/src/run.ts +20 -11
  24. package/src/scripts/test-bedrock-handoff-autonomous.ts +231 -0
  25. package/src/agents/AgentContext.js +0 -782
  26. package/src/agents/AgentContext.test.js +0 -421
  27. package/src/agents/__tests__/AgentContext.test.js +0 -678
  28. package/src/agents/__tests__/resolveStructuredOutputMode.test.js +0 -117
  29. package/src/common/enum.js +0 -192
  30. package/src/common/index.js +0 -3
  31. package/src/events.js +0 -166
  32. package/src/graphs/Graph.js +0 -1857
  33. package/src/graphs/MultiAgentGraph.js +0 -1092
  34. package/src/graphs/__tests__/structured-output.integration.test.js +0 -624
  35. package/src/graphs/__tests__/structured-output.test.js +0 -144
  36. package/src/graphs/contextManagement.e2e.test.js +0 -718
  37. package/src/graphs/contextManagement.test.js +0 -485
  38. package/src/graphs/handoffValidation.test.js +0 -276
  39. package/src/graphs/index.js +0 -3
  40. package/src/index.js +0 -28
  41. package/src/instrumentation.js +0 -21
  42. package/src/llm/anthropic/index.js +0 -319
  43. package/src/llm/anthropic/types.js +0 -46
  44. package/src/llm/anthropic/utils/message_inputs.js +0 -627
  45. package/src/llm/anthropic/utils/message_outputs.js +0 -290
  46. package/src/llm/anthropic/utils/output_parsers.js +0 -89
  47. package/src/llm/anthropic/utils/tools.js +0 -25
  48. package/src/llm/bedrock/__tests__/bedrock-caching.test.js +0 -392
  49. package/src/llm/bedrock/index.js +0 -303
  50. package/src/llm/bedrock/types.js +0 -2
  51. package/src/llm/bedrock/utils/index.js +0 -6
  52. package/src/llm/bedrock/utils/message_inputs.js +0 -463
  53. package/src/llm/bedrock/utils/message_outputs.js +0 -269
  54. package/src/llm/fake.js +0 -92
  55. package/src/llm/google/index.js +0 -215
  56. package/src/llm/google/types.js +0 -12
  57. package/src/llm/google/utils/common.js +0 -670
  58. package/src/llm/google/utils/tools.js +0 -111
  59. package/src/llm/google/utils/zod_to_genai_parameters.js +0 -47
  60. package/src/llm/openai/index.js +0 -1033
  61. package/src/llm/openai/types.js +0 -2
  62. package/src/llm/openai/utils/index.js +0 -756
  63. package/src/llm/openai/utils/isReasoningModel.test.js +0 -79
  64. package/src/llm/openrouter/index.js +0 -261
  65. package/src/llm/openrouter/reasoning.test.js +0 -181
  66. package/src/llm/providers.js +0 -36
  67. package/src/llm/text.js +0 -65
  68. package/src/llm/vertexai/index.js +0 -402
  69. package/src/messages/__tests__/tools.test.js +0 -392
  70. package/src/messages/cache.js +0 -404
  71. package/src/messages/cache.test.js +0 -1167
  72. package/src/messages/content.js +0 -48
  73. package/src/messages/content.test.js +0 -314
  74. package/src/messages/core.js +0 -359
  75. package/src/messages/ensureThinkingBlock.test.js +0 -997
  76. package/src/messages/format.js +0 -973
  77. package/src/messages/formatAgentMessages.test.js +0 -2278
  78. package/src/messages/formatAgentMessages.tools.test.js +0 -362
  79. package/src/messages/formatMessage.test.js +0 -608
  80. package/src/messages/ids.js +0 -18
  81. package/src/messages/index.js +0 -9
  82. package/src/messages/labelContentByAgent.test.js +0 -725
  83. package/src/messages/prune.js +0 -438
  84. package/src/messages/reducer.js +0 -60
  85. package/src/messages/shiftIndexTokenCountMap.test.js +0 -63
  86. package/src/messages/summarize.js +0 -146
  87. package/src/messages/summarize.test.js +0 -332
  88. package/src/messages/tools.js +0 -90
  89. package/src/mockStream.js +0 -81
  90. package/src/prompts/collab.js +0 -7
  91. package/src/prompts/index.js +0 -3
  92. package/src/prompts/taskmanager.js +0 -58
  93. package/src/run.js +0 -427
  94. package/src/schemas/index.js +0 -3
  95. package/src/schemas/schema-preparation.test.js +0 -370
  96. package/src/schemas/validate.js +0 -314
  97. package/src/schemas/validate.test.js +0 -264
  98. package/src/scripts/abort.js +0 -127
  99. package/src/scripts/ant_web_search.js +0 -130
  100. package/src/scripts/ant_web_search_edge_case.js +0 -133
  101. package/src/scripts/ant_web_search_error_edge_case.js +0 -119
  102. package/src/scripts/args.js +0 -41
  103. package/src/scripts/bedrock-cache-debug.js +0 -186
  104. package/src/scripts/bedrock-content-aggregation-test.js +0 -195
  105. package/src/scripts/bedrock-merge-test.js +0 -80
  106. package/src/scripts/bedrock-parallel-tools-test.js +0 -150
  107. package/src/scripts/caching.js +0 -106
  108. package/src/scripts/cli.js +0 -152
  109. package/src/scripts/cli2.js +0 -119
  110. package/src/scripts/cli3.js +0 -163
  111. package/src/scripts/cli4.js +0 -165
  112. package/src/scripts/cli5.js +0 -165
  113. package/src/scripts/code_exec.js +0 -171
  114. package/src/scripts/code_exec_files.js +0 -180
  115. package/src/scripts/code_exec_multi_session.js +0 -185
  116. package/src/scripts/code_exec_ptc.js +0 -265
  117. package/src/scripts/code_exec_session.js +0 -217
  118. package/src/scripts/code_exec_simple.js +0 -120
  119. package/src/scripts/content.js +0 -111
  120. package/src/scripts/empty_input.js +0 -125
  121. package/src/scripts/handoff-test.js +0 -96
  122. package/src/scripts/image.js +0 -138
  123. package/src/scripts/memory.js +0 -83
  124. package/src/scripts/multi-agent-chain.js +0 -271
  125. package/src/scripts/multi-agent-conditional.js +0 -185
  126. package/src/scripts/multi-agent-document-review-chain.js +0 -171
  127. package/src/scripts/multi-agent-hybrid-flow.js +0 -264
  128. package/src/scripts/multi-agent-parallel-start.js +0 -214
  129. package/src/scripts/multi-agent-parallel.js +0 -346
  130. package/src/scripts/multi-agent-sequence.js +0 -184
  131. package/src/scripts/multi-agent-supervisor.js +0 -324
  132. package/src/scripts/multi-agent-test.js +0 -147
  133. package/src/scripts/parallel-asymmetric-tools-test.js +0 -202
  134. package/src/scripts/parallel-full-metadata-test.js +0 -176
  135. package/src/scripts/parallel-tools-test.js +0 -256
  136. package/src/scripts/programmatic_exec.js +0 -277
  137. package/src/scripts/programmatic_exec_agent.js +0 -168
  138. package/src/scripts/search.js +0 -118
  139. package/src/scripts/sequential-full-metadata-test.js +0 -143
  140. package/src/scripts/simple.js +0 -174
  141. package/src/scripts/single-agent-metadata-test.js +0 -152
  142. package/src/scripts/stream.js +0 -113
  143. package/src/scripts/test-custom-prompt-key.js +0 -132
  144. package/src/scripts/test-handoff-input.js +0 -143
  145. package/src/scripts/test-handoff-preamble.js +0 -227
  146. package/src/scripts/test-handoff-steering.js +0 -353
  147. package/src/scripts/test-multi-agent-list-handoff.js +0 -318
  148. package/src/scripts/test-parallel-agent-labeling.js +0 -253
  149. package/src/scripts/test-parallel-handoffs.js +0 -229
  150. package/src/scripts/test-thinking-handoff-bedrock.js +0 -132
  151. package/src/scripts/test-thinking-handoff.js +0 -132
  152. package/src/scripts/test-thinking-to-thinking-handoff-bedrock.js +0 -140
  153. package/src/scripts/test-tool-before-handoff-role-order.js +0 -223
  154. package/src/scripts/test-tools-before-handoff.js +0 -187
  155. package/src/scripts/test_code_api.js +0 -263
  156. package/src/scripts/thinking-bedrock.js +0 -128
  157. package/src/scripts/thinking-vertexai.js +0 -130
  158. package/src/scripts/thinking.js +0 -134
  159. package/src/scripts/tool_search.js +0 -114
  160. package/src/scripts/tools.js +0 -125
  161. package/src/specs/agent-handoffs-bedrock.integration.test.js +0 -280
  162. package/src/specs/agent-handoffs.test.js +0 -924
  163. package/src/specs/anthropic.simple.test.js +0 -287
  164. package/src/specs/azure.simple.test.js +0 -381
  165. package/src/specs/cache.simple.test.js +0 -282
  166. package/src/specs/custom-event-await.test.js +0 -148
  167. package/src/specs/deepseek.simple.test.js +0 -189
  168. package/src/specs/emergency-prune.test.js +0 -308
  169. package/src/specs/moonshot.simple.test.js +0 -237
  170. package/src/specs/observability.integration.test.js +0 -1337
  171. package/src/specs/openai.simple.test.js +0 -233
  172. package/src/specs/openrouter.simple.test.js +0 -202
  173. package/src/specs/prune.test.js +0 -733
  174. package/src/specs/reasoning.test.js +0 -144
  175. package/src/specs/spec.utils.js +0 -4
  176. package/src/specs/thinking-handoff.test.js +0 -486
  177. package/src/specs/thinking-prune.test.js +0 -600
  178. package/src/specs/token-distribution-edge-case.test.js +0 -246
  179. package/src/specs/token-memoization.test.js +0 -32
  180. package/src/specs/tokens.test.js +0 -49
  181. package/src/specs/tool-error.test.js +0 -139
  182. package/src/splitStream.js +0 -204
  183. package/src/splitStream.test.js +0 -504
  184. package/src/stream.js +0 -650
  185. package/src/stream.test.js +0 -225
  186. package/src/test/mockTools.js +0 -340
  187. package/src/tools/BrowserTools.js +0 -245
  188. package/src/tools/Calculator.js +0 -38
  189. package/src/tools/Calculator.test.js +0 -225
  190. package/src/tools/CodeExecutor.js +0 -233
  191. package/src/tools/ProgrammaticToolCalling.js +0 -602
  192. package/src/tools/StreamingToolCallBuffer.js +0 -179
  193. package/src/tools/ToolNode.js +0 -930
  194. package/src/tools/ToolSearch.js +0 -904
  195. package/src/tools/__tests__/BrowserTools.test.js +0 -306
  196. package/src/tools/__tests__/ProgrammaticToolCalling.integration.test.js +0 -276
  197. package/src/tools/__tests__/ProgrammaticToolCalling.test.js +0 -807
  198. package/src/tools/__tests__/StreamingToolCallBuffer.test.js +0 -175
  199. package/src/tools/__tests__/ToolApproval.test.js +0 -675
  200. package/src/tools/__tests__/ToolNode.recovery.test.js +0 -200
  201. package/src/tools/__tests__/ToolNode.session.test.js +0 -319
  202. package/src/tools/__tests__/ToolSearch.integration.test.js +0 -125
  203. package/src/tools/__tests__/ToolSearch.test.js +0 -812
  204. package/src/tools/__tests__/handlers.test.js +0 -799
  205. package/src/tools/__tests__/truncation-recovery.integration.test.js +0 -362
  206. package/src/tools/handlers.js +0 -306
  207. package/src/tools/schema.js +0 -25
  208. package/src/tools/search/anthropic.js +0 -34
  209. package/src/tools/search/content.js +0 -116
  210. package/src/tools/search/content.test.js +0 -133
  211. package/src/tools/search/firecrawl.js +0 -173
  212. package/src/tools/search/format.js +0 -198
  213. package/src/tools/search/highlights.js +0 -241
  214. package/src/tools/search/index.js +0 -3
  215. package/src/tools/search/jina-reranker.test.js +0 -106
  216. package/src/tools/search/rerankers.js +0 -165
  217. package/src/tools/search/schema.js +0 -102
  218. package/src/tools/search/search.js +0 -561
  219. package/src/tools/search/serper-scraper.js +0 -126
  220. package/src/tools/search/test.js +0 -129
  221. package/src/tools/search/tool.js +0 -453
  222. package/src/tools/search/types.js +0 -2
  223. package/src/tools/search/utils.js +0 -59
  224. package/src/types/graph.js +0 -24
  225. package/src/types/graph.test.js +0 -192
  226. package/src/types/index.js +0 -7
  227. package/src/types/llm.js +0 -2
  228. package/src/types/messages.js +0 -2
  229. package/src/types/run.js +0 -2
  230. package/src/types/stream.js +0 -2
  231. package/src/types/tools.js +0 -2
  232. package/src/utils/contextAnalytics.js +0 -79
  233. package/src/utils/contextAnalytics.test.js +0 -166
  234. package/src/utils/events.js +0 -26
  235. package/src/utils/graph.js +0 -11
  236. package/src/utils/handlers.js +0 -65
  237. package/src/utils/index.js +0 -10
  238. package/src/utils/llm.js +0 -21
  239. package/src/utils/llmConfig.js +0 -205
  240. package/src/utils/logging.js +0 -37
  241. package/src/utils/misc.js +0 -51
  242. package/src/utils/run.js +0 -69
  243. package/src/utils/schema.js +0 -21
  244. package/src/utils/title.js +0 -119
  245. package/src/utils/tokens.js +0 -92
  246. package/src/utils/toonFormat.js +0 -379
@@ -1,807 +0,0 @@
1
- // src/tools/__tests__/ProgrammaticToolCalling.test.ts
2
- /**
3
- * Unit tests for Programmatic Tool Calling.
4
- * Tests manual invocation with mock tools and Code API responses.
5
- */
6
- import { describe, it, expect, beforeEach } from '@jest/globals';
7
- import { createProgrammaticToolCallingTool, formatCompletedResponse, extractUsedToolNames, filterToolsByUsage, executeTools, normalizeToPythonIdentifier, unwrapToolResponse, } from '../ProgrammaticToolCalling';
8
- import { createProgrammaticToolRegistry, createGetTeamMembersTool, createGetExpensesTool, createGetWeatherTool, createCalculatorTool, } from '@/test/mockTools';
9
- describe('ProgrammaticToolCalling', () => {
10
- describe('executeTools', () => {
11
- let toolMap;
12
- beforeEach(() => {
13
- const tools = [
14
- createGetTeamMembersTool(),
15
- createGetExpensesTool(),
16
- createGetWeatherTool(),
17
- createCalculatorTool(),
18
- ];
19
- toolMap = new Map(tools.map((t) => [t.name, t]));
20
- });
21
- it('executes a single tool successfully', async () => {
22
- const toolCalls = [
23
- {
24
- id: 'call_001',
25
- name: 'get_weather',
26
- input: { city: 'San Francisco' },
27
- },
28
- ];
29
- const results = await executeTools(toolCalls, toolMap);
30
- expect(results).toHaveLength(1);
31
- expect(results[0].call_id).toBe('call_001');
32
- expect(results[0].is_error).toBe(false);
33
- expect(results[0].result).toEqual({
34
- temperature: 65,
35
- condition: 'Foggy',
36
- });
37
- });
38
- it('executes multiple tools in parallel', async () => {
39
- const toolCalls = [
40
- {
41
- id: 'call_001',
42
- name: 'get_weather',
43
- input: { city: 'San Francisco' },
44
- },
45
- {
46
- id: 'call_002',
47
- name: 'get_weather',
48
- input: { city: 'New York' },
49
- },
50
- {
51
- id: 'call_003',
52
- name: 'get_weather',
53
- input: { city: 'London' },
54
- },
55
- ];
56
- const startTime = Date.now();
57
- const results = await executeTools(toolCalls, toolMap);
58
- const duration = Date.now() - startTime;
59
- // Should execute in parallel (< 150ms total, not 120ms sequential)
60
- expect(duration).toBeLessThan(150);
61
- expect(results).toHaveLength(3);
62
- expect(results[0].is_error).toBe(false);
63
- expect(results[1].is_error).toBe(false);
64
- expect(results[2].is_error).toBe(false);
65
- expect(results[0].result.temperature).toBe(65);
66
- expect(results[1].result.temperature).toBe(75);
67
- expect(results[2].result.temperature).toBe(55);
68
- });
69
- it('handles tool not found error', async () => {
70
- const toolCalls = [
71
- {
72
- id: 'call_001',
73
- name: 'nonexistent_tool',
74
- input: {},
75
- },
76
- ];
77
- const results = await executeTools(toolCalls, toolMap);
78
- expect(results).toHaveLength(1);
79
- expect(results[0].call_id).toBe('call_001');
80
- expect(results[0].is_error).toBe(true);
81
- expect(results[0].error_message).toContain('nonexistent_tool');
82
- expect(results[0].error_message).toContain('Available tools:');
83
- });
84
- it('handles tool execution error', async () => {
85
- const toolCalls = [
86
- {
87
- id: 'call_001',
88
- name: 'get_weather',
89
- input: { city: 'InvalidCity' },
90
- },
91
- ];
92
- const results = await executeTools(toolCalls, toolMap);
93
- expect(results).toHaveLength(1);
94
- expect(results[0].call_id).toBe('call_001');
95
- expect(results[0].is_error).toBe(true);
96
- expect(results[0].error_message).toContain('Weather data not available');
97
- });
98
- it('handles mix of successful and failed tool calls', async () => {
99
- const toolCalls = [
100
- {
101
- id: 'call_001',
102
- name: 'get_weather',
103
- input: { city: 'San Francisco' },
104
- },
105
- {
106
- id: 'call_002',
107
- name: 'get_weather',
108
- input: { city: 'InvalidCity' },
109
- },
110
- {
111
- id: 'call_003',
112
- name: 'get_weather',
113
- input: { city: 'New York' },
114
- },
115
- ];
116
- const results = await executeTools(toolCalls, toolMap);
117
- expect(results).toHaveLength(3);
118
- expect(results[0].is_error).toBe(false);
119
- expect(results[1].is_error).toBe(true);
120
- expect(results[2].is_error).toBe(false);
121
- });
122
- it('executes tools with different parameters', async () => {
123
- const toolCalls = [
124
- {
125
- id: 'call_001',
126
- name: 'get_team_members',
127
- input: {},
128
- },
129
- {
130
- id: 'call_002',
131
- name: 'get_expenses',
132
- input: { user_id: 'u1' },
133
- },
134
- {
135
- id: 'call_003',
136
- name: 'calculator',
137
- input: { expression: '2 + 2 * 3' },
138
- },
139
- ];
140
- const results = await executeTools(toolCalls, toolMap);
141
- expect(results).toHaveLength(3);
142
- expect(results[0].is_error).toBe(false);
143
- expect(results[1].is_error).toBe(false);
144
- expect(results[2].is_error).toBe(false);
145
- expect(Array.isArray(results[0].result)).toBe(true);
146
- expect(results[0].result).toHaveLength(3);
147
- expect(Array.isArray(results[1].result)).toBe(true);
148
- expect(results[2].result.result).toBe(8);
149
- });
150
- });
151
- describe('normalizeToPythonIdentifier', () => {
152
- it('converts hyphens to underscores', () => {
153
- expect(normalizeToPythonIdentifier('my-tool-name')).toBe('my_tool_name');
154
- });
155
- it('converts spaces to underscores', () => {
156
- expect(normalizeToPythonIdentifier('my tool name')).toBe('my_tool_name');
157
- });
158
- it('leaves underscores unchanged', () => {
159
- expect(normalizeToPythonIdentifier('my_tool_name')).toBe('my_tool_name');
160
- });
161
- it('handles mixed hyphens and underscores', () => {
162
- expect(normalizeToPythonIdentifier('my-tool_name-v2')).toBe('my_tool_name_v2');
163
- });
164
- it('handles MCP-style names with hyphens', () => {
165
- expect(normalizeToPythonIdentifier('create_spreadsheet_mcp_Google-Workspace')).toBe('create_spreadsheet_mcp_Google_Workspace');
166
- });
167
- it('removes invalid characters', () => {
168
- expect(normalizeToPythonIdentifier('tool@name!v2')).toBe('toolnamev2');
169
- expect(normalizeToPythonIdentifier('get.data.v2')).toBe('getdatav2');
170
- });
171
- it('prefixes with underscore if starts with number', () => {
172
- expect(normalizeToPythonIdentifier('123tool')).toBe('_123tool');
173
- expect(normalizeToPythonIdentifier('1-tool')).toBe('_1_tool');
174
- });
175
- it('appends _tool suffix for Python keywords', () => {
176
- expect(normalizeToPythonIdentifier('return')).toBe('return_tool');
177
- expect(normalizeToPythonIdentifier('async')).toBe('async_tool');
178
- expect(normalizeToPythonIdentifier('import')).toBe('import_tool');
179
- });
180
- });
181
- describe('unwrapToolResponse', () => {
182
- describe('non-MCP tools', () => {
183
- it('returns result as-is for non-MCP tools', () => {
184
- const result = { temperature: 65, condition: 'Foggy' };
185
- expect(unwrapToolResponse(result, false)).toEqual(result);
186
- });
187
- it('returns string as-is for non-MCP tools', () => {
188
- expect(unwrapToolResponse('plain string', false)).toBe('plain string');
189
- });
190
- it('returns array as-is for non-MCP tools', () => {
191
- const result = [1, 2, 3];
192
- expect(unwrapToolResponse(result, false)).toEqual(result);
193
- });
194
- });
195
- describe('MCP tools - tuple format [content, artifacts]', () => {
196
- it('extracts string content from tuple', () => {
197
- const result = ['Hello world', { artifacts: [] }];
198
- expect(unwrapToolResponse(result, true)).toBe('Hello world');
199
- });
200
- it('parses JSON string content from tuple', () => {
201
- const result = ['{"temperature": 65}', { artifacts: [] }];
202
- expect(unwrapToolResponse(result, true)).toEqual({ temperature: 65 });
203
- });
204
- it('parses JSON array string content from tuple', () => {
205
- const result = ['[1, 2, 3]', { artifacts: [] }];
206
- expect(unwrapToolResponse(result, true)).toEqual([1, 2, 3]);
207
- });
208
- it('extracts text from single content block in tuple', () => {
209
- const result = [{ type: 'text', text: 'Spreadsheet info here' }, {}];
210
- expect(unwrapToolResponse(result, true)).toBe('Spreadsheet info here');
211
- });
212
- it('extracts and parses JSON from single content block in tuple', () => {
213
- const result = [
214
- { type: 'text', text: '{"id": "123", "name": "Test"}' },
215
- {},
216
- ];
217
- expect(unwrapToolResponse(result, true)).toEqual({
218
- id: '123',
219
- name: 'Test',
220
- });
221
- });
222
- it('extracts text from array of content blocks in tuple', () => {
223
- const result = [
224
- [
225
- { type: 'text', text: 'Line 1' },
226
- { type: 'text', text: 'Line 2' },
227
- ],
228
- {},
229
- ];
230
- expect(unwrapToolResponse(result, true)).toBe('Line 1\nLine 2');
231
- });
232
- it('returns object content as-is when not a text block', () => {
233
- const result = [{ temperature: 65, condition: 'Foggy' }, {}];
234
- expect(unwrapToolResponse(result, true)).toEqual({
235
- temperature: 65,
236
- condition: 'Foggy',
237
- });
238
- });
239
- });
240
- describe('MCP tools - single content block (not in tuple)', () => {
241
- it('extracts text from single content block object', () => {
242
- const result = { type: 'text', text: 'No data found in range' };
243
- expect(unwrapToolResponse(result, true)).toBe('No data found in range');
244
- });
245
- it('extracts and parses JSON from single content block object', () => {
246
- const result = {
247
- type: 'text',
248
- text: '{"sheets": [{"name": "raw_data"}]}',
249
- };
250
- expect(unwrapToolResponse(result, true)).toEqual({
251
- sheets: [{ name: 'raw_data' }],
252
- });
253
- });
254
- it('handles real-world MCP spreadsheet response', () => {
255
- const result = {
256
- type: 'text',
257
- text: 'Spreadsheet: "NYC Taxi - Top Pickup Neighborhoods" (ID: abc123)\nSheets (2):\n - "raw_data" (ID: 123) | Size: 1000x26',
258
- };
259
- expect(unwrapToolResponse(result, true)).toBe('Spreadsheet: "NYC Taxi - Top Pickup Neighborhoods" (ID: abc123)\nSheets (2):\n - "raw_data" (ID: 123) | Size: 1000x26');
260
- });
261
- it('handles real-world MCP no data response', () => {
262
- const result = {
263
- type: 'text',
264
- text: 'No data found in range \'raw_data!A1:D25\' for user@example.com.',
265
- };
266
- expect(unwrapToolResponse(result, true)).toBe('No data found in range \'raw_data!A1:D25\' for user@example.com.');
267
- });
268
- });
269
- describe('MCP tools - array of content blocks (not in tuple)', () => {
270
- it('extracts text from array of content blocks', () => {
271
- const result = [
272
- { type: 'text', text: 'First block' },
273
- { type: 'text', text: 'Second block' },
274
- ];
275
- expect(unwrapToolResponse(result, true)).toBe('First block\nSecond block');
276
- });
277
- it('filters out non-text blocks', () => {
278
- const result = [
279
- { type: 'text', text: 'Text content' },
280
- { type: 'image', data: 'base64...' },
281
- { type: 'text', text: 'More text' },
282
- ];
283
- expect(unwrapToolResponse(result, true)).toBe('Text content\nMore text');
284
- });
285
- });
286
- describe('edge cases', () => {
287
- it('returns non-text block object as-is', () => {
288
- const result = { type: 'image', data: 'base64...' };
289
- expect(unwrapToolResponse(result, true)).toEqual(result);
290
- });
291
- it('handles empty array', () => {
292
- expect(unwrapToolResponse([], true)).toEqual([]);
293
- });
294
- it('handles malformed JSON in text block gracefully', () => {
295
- const result = { type: 'text', text: '{ invalid json }' };
296
- expect(unwrapToolResponse(result, true)).toBe('{ invalid json }');
297
- });
298
- it('handles null', () => {
299
- expect(unwrapToolResponse(null, true)).toBe(null);
300
- });
301
- it('handles undefined', () => {
302
- expect(unwrapToolResponse(undefined, true)).toBe(undefined);
303
- });
304
- });
305
- });
306
- describe('extractUsedToolNames', () => {
307
- const createToolMap = (names) => {
308
- const map = new Map();
309
- for (const name of names) {
310
- map.set(normalizeToPythonIdentifier(name), name);
311
- }
312
- return map;
313
- };
314
- const availableTools = createToolMap([
315
- 'get_weather',
316
- 'get_team_members',
317
- 'get_expenses',
318
- 'calculator',
319
- 'search_docs',
320
- ]);
321
- it('extracts single tool name from simple code', () => {
322
- const code = `result = await get_weather(city="SF")
323
- print(result)`;
324
- const used = extractUsedToolNames(code, availableTools);
325
- expect(used.size).toBe(1);
326
- expect(used.has('get_weather')).toBe(true);
327
- });
328
- it('extracts multiple tool names from code', () => {
329
- const code = `team = await get_team_members()
330
- for member in team:
331
- expenses = await get_expenses(user_id=member['id'])
332
- print(f"{member['name']}: {sum(e['amount'] for e in expenses)}")`;
333
- const used = extractUsedToolNames(code, availableTools);
334
- expect(used.size).toBe(2);
335
- expect(used.has('get_team_members')).toBe(true);
336
- expect(used.has('get_expenses')).toBe(true);
337
- });
338
- it('extracts tools from asyncio.gather calls', () => {
339
- const code = `results = await asyncio.gather(
340
- get_weather(city="SF"),
341
- get_weather(city="NYC"),
342
- get_expenses(user_id="u1")
343
- )`;
344
- const used = extractUsedToolNames(code, availableTools);
345
- expect(used.size).toBe(2);
346
- expect(used.has('get_weather')).toBe(true);
347
- expect(used.has('get_expenses')).toBe(true);
348
- });
349
- it('does not match partial tool names', () => {
350
- const code = `# Using get_weather_data instead
351
- result = await get_weather_data(city="SF")`;
352
- const used = extractUsedToolNames(code, availableTools);
353
- expect(used.has('get_weather')).toBe(false);
354
- });
355
- it('matches tool names in different contexts', () => {
356
- const code = `# direct call
357
- x = await calculator(expression="1+1")
358
- # in list comprehension
359
- results = [await get_weather(city=c) for c in cities]
360
- # conditional
361
- if condition:
362
- await get_team_members()`;
363
- const used = extractUsedToolNames(code, availableTools);
364
- expect(used.size).toBe(3);
365
- expect(used.has('calculator')).toBe(true);
366
- expect(used.has('get_weather')).toBe(true);
367
- expect(used.has('get_team_members')).toBe(true);
368
- });
369
- it('returns empty set when no tools are used', () => {
370
- const code = `print("Hello, World!")
371
- x = 1 + 2`;
372
- const used = extractUsedToolNames(code, availableTools);
373
- expect(used.size).toBe(0);
374
- });
375
- it('handles tool names with special characters via normalization', () => {
376
- const specialTools = createToolMap(['get_data.v2', 'calc+plus']);
377
- const code = `await get_datav2()
378
- await calcplus()`;
379
- const used = extractUsedToolNames(code, specialTools);
380
- expect(used.has('get_data.v2')).toBe(true);
381
- expect(used.has('calc+plus')).toBe(true);
382
- });
383
- it('matches hyphenated tool names using underscore in code', () => {
384
- const mcpTools = createToolMap([
385
- 'create_spreadsheet_mcp_Google-Workspace',
386
- 'search_gmail_mcp_Google-Workspace',
387
- ]);
388
- const code = `result = await create_spreadsheet_mcp_Google_Workspace(title="Test")
389
- print(result)`;
390
- const used = extractUsedToolNames(code, mcpTools);
391
- expect(used.size).toBe(1);
392
- expect(used.has('create_spreadsheet_mcp_Google-Workspace')).toBe(true);
393
- });
394
- });
395
- describe('filterToolsByUsage', () => {
396
- const allToolDefs = [
397
- {
398
- name: 'get_weather',
399
- description: 'Get weather for a city',
400
- parameters: {
401
- type: 'object',
402
- properties: { city: { type: 'string' } },
403
- },
404
- },
405
- {
406
- name: 'get_team_members',
407
- description: 'Get team members',
408
- parameters: { type: 'object', properties: {} },
409
- },
410
- {
411
- name: 'get_expenses',
412
- description: 'Get expenses for a user',
413
- parameters: {
414
- type: 'object',
415
- properties: { user_id: { type: 'string' } },
416
- },
417
- },
418
- {
419
- name: 'calculator',
420
- description: 'Evaluate an expression',
421
- parameters: {
422
- type: 'object',
423
- properties: { expression: { type: 'string' } },
424
- },
425
- },
426
- ];
427
- it('filters to only used tools', () => {
428
- const code = `result = await get_weather(city="SF")
429
- print(result)`;
430
- const filtered = filterToolsByUsage(allToolDefs, code);
431
- expect(filtered).toHaveLength(1);
432
- expect(filtered[0].name).toBe('get_weather');
433
- });
434
- it('filters to multiple used tools', () => {
435
- const code = `team = await get_team_members()
436
- for member in team:
437
- expenses = await get_expenses(user_id=member['id'])`;
438
- const filtered = filterToolsByUsage(allToolDefs, code);
439
- expect(filtered).toHaveLength(2);
440
- expect(filtered.map((t) => t.name).sort()).toEqual([
441
- 'get_expenses',
442
- 'get_team_members',
443
- ]);
444
- });
445
- it('returns all tools when no tools are detected', () => {
446
- const code = 'print("Hello, World!")';
447
- const filtered = filterToolsByUsage(allToolDefs, code);
448
- expect(filtered).toHaveLength(4);
449
- });
450
- it('preserves tool definition structure', () => {
451
- const code = 'await calculator(expression="2+2")';
452
- const filtered = filterToolsByUsage(allToolDefs, code);
453
- expect(filtered).toHaveLength(1);
454
- expect(filtered[0]).toEqual(allToolDefs[3]);
455
- expect(filtered[0].parameters).toBeDefined();
456
- expect(filtered[0].description).toBe('Evaluate an expression');
457
- });
458
- it('handles empty tool definitions', () => {
459
- const code = 'await get_weather(city="SF")';
460
- const filtered = filterToolsByUsage([], code);
461
- expect(filtered).toHaveLength(0);
462
- });
463
- });
464
- describe('formatCompletedResponse', () => {
465
- it('formats response with stdout', () => {
466
- const response = {
467
- status: 'completed',
468
- stdout: 'Hello, World!\n',
469
- stderr: '',
470
- files: [],
471
- session_id: 'sess_abc123',
472
- };
473
- const [output, artifact] = formatCompletedResponse(response);
474
- expect(output).toContain('stdout:\nHello, World!');
475
- expect(artifact.session_id).toBe('sess_abc123');
476
- expect(artifact.files).toEqual([]);
477
- });
478
- it('shows empty output message when no stdout', () => {
479
- const response = {
480
- status: 'completed',
481
- stdout: '',
482
- stderr: '',
483
- files: [],
484
- session_id: 'sess_abc123',
485
- };
486
- const [output] = formatCompletedResponse(response);
487
- expect(output).toContain('stdout: Empty. Ensure you\'re writing output explicitly');
488
- });
489
- it('includes stderr when present', () => {
490
- const response = {
491
- status: 'completed',
492
- stdout: 'Output\n',
493
- stderr: 'Warning: deprecated function\n',
494
- files: [],
495
- session_id: 'sess_abc123',
496
- };
497
- const [output] = formatCompletedResponse(response);
498
- expect(output).toContain('stdout:\nOutput');
499
- expect(output).toContain('stderr:\nWarning: deprecated function');
500
- });
501
- it('formats file information correctly', () => {
502
- const response = {
503
- status: 'completed',
504
- stdout: 'Generated report\n',
505
- stderr: '',
506
- files: [
507
- { id: '1', name: 'report.pdf' },
508
- { id: '2', name: 'data.csv' },
509
- ],
510
- session_id: 'sess_abc123',
511
- };
512
- const [output, artifact] = formatCompletedResponse(response);
513
- expect(output).toContain('Generated files:');
514
- expect(output).toContain('report.pdf');
515
- expect(output).toContain('data.csv');
516
- expect(artifact.files).toHaveLength(2);
517
- expect(artifact.files).toEqual(response.files);
518
- });
519
- it('handles image files with special message', () => {
520
- const response = {
521
- status: 'completed',
522
- stdout: '',
523
- stderr: '',
524
- files: [
525
- { id: '1', name: 'chart.png' },
526
- { id: '2', name: 'photo.jpg' },
527
- ],
528
- session_id: 'sess_abc123',
529
- };
530
- const [output] = formatCompletedResponse(response);
531
- expect(output).toContain('chart.png');
532
- expect(output).toContain('Image is already displayed to the user');
533
- });
534
- });
535
- describe('createProgrammaticToolCallingTool - Manual Invocation', () => {
536
- let ptcTool;
537
- let toolMap;
538
- let toolDefinitions;
539
- beforeEach(() => {
540
- const tools = [
541
- createGetTeamMembersTool(),
542
- createGetExpensesTool(),
543
- createGetWeatherTool(),
544
- ];
545
- toolMap = new Map(tools.map((t) => [t.name, t]));
546
- toolDefinitions = Array.from(createProgrammaticToolRegistry().values()).filter((t) => ['get_team_members', 'get_expenses', 'get_weather'].includes(t.name));
547
- ptcTool = createProgrammaticToolCallingTool({
548
- apiKey: 'test-key',
549
- baseUrl: 'http://mock-api',
550
- });
551
- });
552
- it('throws error when no toolMap provided', async () => {
553
- await expect(ptcTool.invoke({
554
- code: 'result = await get_weather(city="SF")\nprint(result)',
555
- tools: toolDefinitions,
556
- toolMap,
557
- })).rejects.toThrow('No toolMap provided');
558
- });
559
- it('throws error when toolMap is empty', async () => {
560
- const args = {
561
- code: 'result = await get_weather(city="SF")\nprint(result)',
562
- tools: toolDefinitions,
563
- toolMap: new Map(),
564
- };
565
- const toolCall = {
566
- name: 'programmatic_tool_calling',
567
- args,
568
- };
569
- await expect(ptcTool.invoke(args, {
570
- toolCall,
571
- })).rejects.toThrow('No toolMap provided');
572
- });
573
- it('throws error when no tool definitions provided', async () => {
574
- const args = {
575
- code: 'result = await get_weather(city="SF")\nprint(result)',
576
- // No tools
577
- };
578
- const toolCall = {
579
- name: 'programmatic_code_execution',
580
- args,
581
- toolMap,
582
- // No `toolDefs`
583
- };
584
- await expect(ptcTool.invoke(args, { toolCall })).rejects.toThrow('No tool definitions provided');
585
- });
586
- it('uses toolDefs from config when tools not provided', async () => {
587
- // Skip this test - requires mocking fetch which has complex typing
588
- // This functionality is tested in the live script tests instead
589
- });
590
- });
591
- describe('Tool Classification', () => {
592
- it('filters tools by allowed_callers', () => {
593
- const registry = createProgrammaticToolRegistry();
594
- const codeExecutionTools = Array.from(registry.values()).filter((t) => (t.allowed_callers ?? ['direct']).includes('code_execution'));
595
- // get_team_members, get_expenses, calculator: code_execution only
596
- const codeOnlyTools = codeExecutionTools.filter((t) => !(t.allowed_callers?.includes('direct') === true));
597
- expect(codeOnlyTools.length).toBeGreaterThanOrEqual(3);
598
- // get_weather: both direct and code_execution
599
- const bothTools = Array.from(registry.values()).filter((t) => t.allowed_callers?.includes('direct') === true &&
600
- t.allowed_callers.includes('code_execution'));
601
- expect(bothTools.length).toBeGreaterThanOrEqual(1);
602
- expect(bothTools.some((t) => t.name === 'get_weather')).toBe(true);
603
- });
604
- });
605
- describe('Error Handling', () => {
606
- let toolMap;
607
- beforeEach(() => {
608
- const tools = [createGetWeatherTool()];
609
- toolMap = new Map(tools.map((t) => [t.name, t]));
610
- });
611
- it('returns error for invalid city without throwing', async () => {
612
- const toolCalls = [
613
- {
614
- id: 'call_001',
615
- name: 'get_weather',
616
- input: { city: 'InvalidCity' },
617
- },
618
- ];
619
- const results = await executeTools(toolCalls, toolMap);
620
- expect(results).toHaveLength(1);
621
- expect(results[0].is_error).toBe(true);
622
- expect(results[0].result).toBeNull();
623
- expect(results[0].error_message).toContain('Weather data not available');
624
- });
625
- it('continues execution when one tool fails', async () => {
626
- const toolCalls = [
627
- {
628
- id: 'call_001',
629
- name: 'get_weather',
630
- input: { city: 'San Francisco' },
631
- },
632
- {
633
- id: 'call_002',
634
- name: 'get_weather',
635
- input: { city: 'InvalidCity' },
636
- },
637
- {
638
- id: 'call_003',
639
- name: 'get_weather',
640
- input: { city: 'London' },
641
- },
642
- ];
643
- const results = await executeTools(toolCalls, toolMap);
644
- expect(results).toHaveLength(3);
645
- expect(results[0].is_error).toBe(false);
646
- expect(results[1].is_error).toBe(true);
647
- expect(results[2].is_error).toBe(false);
648
- });
649
- });
650
- describe('Parallel Execution Performance', () => {
651
- let toolMap;
652
- beforeEach(() => {
653
- const tools = [createGetExpensesTool()];
654
- toolMap = new Map(tools.map((t) => [t.name, t]));
655
- });
656
- it('executes tools in parallel, not sequentially', async () => {
657
- const toolCalls = [
658
- { id: 'call_001', name: 'get_expenses', input: { user_id: 'u1' } },
659
- { id: 'call_002', name: 'get_expenses', input: { user_id: 'u2' } },
660
- { id: 'call_003', name: 'get_expenses', input: { user_id: 'u3' } },
661
- ];
662
- const startTime = Date.now();
663
- const results = await executeTools(toolCalls, toolMap);
664
- const duration = Date.now() - startTime;
665
- // Each tool has 30ms delay
666
- // Sequential would be ~90ms, parallel should be ~30-50ms
667
- expect(duration).toBeLessThan(80);
668
- expect(results).toHaveLength(3);
669
- expect(results.every((r) => r.is_error === false)).toBe(true);
670
- });
671
- });
672
- describe('Response Formatting', () => {
673
- it('formats stdout-only response', () => {
674
- const response = {
675
- status: 'completed',
676
- stdout: 'Team size: 3\n- Alice\n- Bob\n- Charlie\n',
677
- stderr: '',
678
- files: [],
679
- session_id: 'sess_xyz',
680
- };
681
- const [output, artifact] = formatCompletedResponse(response);
682
- expect(output).toBe('stdout:\nTeam size: 3\n- Alice\n- Bob\n- Charlie');
683
- expect(artifact).toEqual({
684
- session_id: 'sess_xyz',
685
- files: [],
686
- });
687
- });
688
- it('formats response with files', () => {
689
- const response = {
690
- status: 'completed',
691
- stdout: 'Report generated\n',
692
- stderr: '',
693
- files: [
694
- { id: '1', name: 'report.csv' },
695
- { id: '2', name: 'chart.png' },
696
- ],
697
- session_id: 'sess_xyz',
698
- };
699
- const [output, artifact] = formatCompletedResponse(response);
700
- expect(output).toContain('Generated files:');
701
- expect(output).toContain('report.csv');
702
- expect(output).toContain('chart.png');
703
- expect(output).toContain('File is already downloaded');
704
- expect(output).toContain('Image is already displayed');
705
- expect(artifact.files).toHaveLength(2);
706
- });
707
- it('handles multiple files with correct separators', () => {
708
- const response = {
709
- status: 'completed',
710
- stdout: 'Done\n',
711
- stderr: '',
712
- files: [
713
- { id: '1', name: 'file1.txt' },
714
- { id: '2', name: 'file2.txt' },
715
- ],
716
- session_id: 'sess_xyz',
717
- };
718
- const [output] = formatCompletedResponse(response);
719
- // 2 files format: "- /mnt/data/file1.txt | ..., - /mnt/data/file2.txt | ..."
720
- expect(output).toContain('file1.txt');
721
- expect(output).toContain('file2.txt');
722
- expect(output).toContain('- /mnt/data/file1.txt');
723
- expect(output).toContain('- /mnt/data/file2.txt');
724
- });
725
- it('handles many files with newline separators', () => {
726
- const response = {
727
- status: 'completed',
728
- stdout: 'Done\n',
729
- stderr: '',
730
- files: [
731
- { id: '1', name: 'file1.txt' },
732
- { id: '2', name: 'file2.txt' },
733
- { id: '3', name: 'file3.txt' },
734
- { id: '4', name: 'file4.txt' },
735
- ],
736
- session_id: 'sess_xyz',
737
- };
738
- const [output] = formatCompletedResponse(response);
739
- // More than 3 files should use newline separators
740
- expect(output).toContain('file1.txt');
741
- expect(output).toContain('file4.txt');
742
- expect(output.match(/,\n/g)?.length).toBeGreaterThanOrEqual(2);
743
- });
744
- });
745
- describe('Tool Data Extraction', () => {
746
- let toolMap;
747
- beforeEach(() => {
748
- const tools = [
749
- createGetTeamMembersTool(),
750
- createGetExpensesTool(),
751
- createCalculatorTool(),
752
- ];
753
- toolMap = new Map(tools.map((t) => [t.name, t]));
754
- });
755
- it('extracts correct data from team members tool', async () => {
756
- const toolCalls = [
757
- { id: 'call_001', name: 'get_team_members', input: {} },
758
- ];
759
- const results = await executeTools(toolCalls, toolMap);
760
- expect(results[0].result).toEqual([
761
- { id: 'u1', name: 'Alice', department: 'Engineering' },
762
- { id: 'u2', name: 'Bob', department: 'Marketing' },
763
- { id: 'u3', name: 'Charlie', department: 'Engineering' },
764
- ]);
765
- });
766
- it('extracts correct data from expenses tool', async () => {
767
- const toolCalls = [
768
- { id: 'call_001', name: 'get_expenses', input: { user_id: 'u1' } },
769
- ];
770
- const results = await executeTools(toolCalls, toolMap);
771
- expect(results[0].result).toEqual([
772
- { amount: 150.0, category: 'travel' },
773
- { amount: 75.5, category: 'meals' },
774
- ]);
775
- });
776
- it('handles empty expense data', async () => {
777
- const toolCalls = [
778
- {
779
- id: 'call_001',
780
- name: 'get_expenses',
781
- input: { user_id: 'nonexistent' },
782
- },
783
- ];
784
- const results = await executeTools(toolCalls, toolMap);
785
- expect(results[0].is_error).toBe(false);
786
- expect(results[0].result).toEqual([]);
787
- });
788
- it('calculates correct result', async () => {
789
- const toolCalls = [
790
- {
791
- id: 'call_001',
792
- name: 'calculator',
793
- input: { expression: '2 + 2 * 3' },
794
- },
795
- {
796
- id: 'call_002',
797
- name: 'calculator',
798
- input: { expression: '(10 + 5) / 3' },
799
- },
800
- ];
801
- const results = await executeTools(toolCalls, toolMap);
802
- expect(results[0].result.result).toBe(8);
803
- expect(results[1].result.result).toBe(5);
804
- });
805
- });
806
- });
807
- //# sourceMappingURL=ProgrammaticToolCalling.test.js.map