decodingtrust-agent-sdk 0.1.0__py3-none-any.whl
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.
- agent/__init__.py +30 -0
- agent/claudesdk/__init__.py +8 -0
- agent/claudesdk/example.py +221 -0
- agent/claudesdk/src/__init__.py +8 -0
- agent/claudesdk/src/agent.py +400 -0
- agent/claudesdk/src/mcp_proxy.py +409 -0
- agent/claudesdk/src/utils.py +420 -0
- agent/googleadk/__init__.py +15 -0
- agent/googleadk/example.py +237 -0
- agent/googleadk/src/__init__.py +12 -0
- agent/googleadk/src/agent.py +401 -0
- agent/googleadk/src/mcp_wrapper.py +163 -0
- agent/googleadk/src/utils.py +602 -0
- agent/langchain/__init__.py +8 -0
- agent/langchain/example.py +213 -0
- agent/langchain/src/__init__.py +8 -0
- agent/langchain/src/agent.py +645 -0
- agent/langchain/src/utils.py +433 -0
- agent/openaisdk/__init__.py +17 -0
- agent/openaisdk/example.py +228 -0
- agent/openaisdk/src/__init__.py +12 -0
- agent/openaisdk/src/agent.py +491 -0
- agent/openaisdk/src/agent_wrapper.py +143 -0
- agent/openaisdk/src/mcp_wrapper.py +395 -0
- agent/openaisdk/src/utils.py +493 -0
- agent/openclaw/__init__.py +10 -0
- agent/openclaw/example.py +251 -0
- agent/openclaw/src/__init__.py +14 -0
- agent/openclaw/src/agent.py +930 -0
- agent/openclaw/src/helpers/__init__.py +1 -0
- agent/openclaw/src/helpers/auth_helpers.py +55 -0
- agent/openclaw/src/mcp_proxy.py +564 -0
- agent/openclaw/src/plugin_generator.py +231 -0
- agent/openclaw/src/utils.py +341 -0
- agent/pocketflow/__init__.py +18 -0
- agent/pocketflow/example.py +221 -0
- agent/pocketflow/prompts/react_agent.py +46 -0
- agent/pocketflow/src/__init__.py +6 -0
- agent/pocketflow/src/agent.py +507 -0
- agent/pocketflow/src/agent_wrapper.py +159 -0
- agent/pocketflow/src/async_helper.py +92 -0
- agent/pocketflow/src/mcp_react_agent.py +279 -0
- agent/pocketflow/src/native_agent.py +74 -0
- agent/pocketflow/src/nodes.py +467 -0
- benchmark/__init__.py +0 -0
- benchmark/browser/benign.jsonl +34 -0
- benchmark/browser/direct.jsonl +85 -0
- benchmark/browser/indirect.jsonl +82 -0
- benchmark/code/benign.jsonl +0 -0
- benchmark/code/direct.jsonl +121 -0
- benchmark/code/indirect.jsonl +165 -0
- benchmark/crm/benign.jsonl +165 -0
- benchmark/crm/direct.jsonl +90 -0
- benchmark/crm/indirect.jsonl +150 -0
- benchmark/customer-service/benign.jsonl +160 -0
- benchmark/customer-service/direct.jsonl +100 -0
- benchmark/customer-service/indirect.jsonl +101 -0
- benchmark/finance/benign.jsonl +0 -0
- benchmark/finance/direct.jsonl +200 -0
- benchmark/finance/indirect.jsonl +200 -0
- benchmark/legal/benign.jsonl +0 -0
- benchmark/legal/direct.jsonl +200 -0
- benchmark/legal/indirect.jsonl +200 -0
- benchmark/macos/benign.jsonl +30 -0
- benchmark/macos/direct.jsonl +50 -0
- benchmark/macos/indirect.jsonl +50 -0
- benchmark/medical/benign.jsonl +642 -0
- benchmark/medical/direct.jsonl +229 -0
- benchmark/medical/indirect.jsonl +222 -0
- benchmark/os-filesystem/benign.jsonl +200 -0
- benchmark/os-filesystem/direct.jsonl +200 -0
- benchmark/os-filesystem/indirect.jsonl +200 -0
- benchmark/research/benign.jsonl +0 -0
- benchmark/research/direct.jsonl +119 -0
- benchmark/research/indirect.jsonl +125 -0
- benchmark/telecom/benign.jsonl +120 -0
- benchmark/telecom/direct.jsonl +161 -0
- benchmark/telecom/indirect.jsonl +166 -0
- benchmark/travel/benign.jsonl +130 -0
- benchmark/travel/direct.jsonl +105 -0
- benchmark/travel/indirect.jsonl +120 -0
- benchmark/windows/benign.jsonl +100 -0
- benchmark/windows/direct.jsonl +140 -0
- benchmark/windows/indirect.jsonl +107 -0
- benchmark/workflow/benign.jsonl +335 -0
- benchmark/workflow/direct.jsonl +78 -0
- benchmark/workflow/indirect.jsonl +107 -0
- cli/__init__.py +5 -0
- cli/main.py +182 -0
- cli/scaffold.py +334 -0
- decodingtrust_agent_sdk-0.1.0.dist-info/METADATA +642 -0
- decodingtrust_agent_sdk-0.1.0.dist-info/RECORD +374 -0
- decodingtrust_agent_sdk-0.1.0.dist-info/WHEEL +5 -0
- decodingtrust_agent_sdk-0.1.0.dist-info/entry_points.txt +2 -0
- decodingtrust_agent_sdk-0.1.0.dist-info/licenses/LICENSE +201 -0
- decodingtrust_agent_sdk-0.1.0.dist-info/top_level.txt +6 -0
- dt_arena/config/env.yaml +515 -0
- dt_arena/config/injection_mcp.yaml +430 -0
- dt_arena/config/mcp.yaml +642 -0
- dt_arena/envs/arxiv/docker-compose-hub.yml +31 -0
- dt_arena/envs/arxiv/docker-compose.yml +36 -0
- dt_arena/envs/atlassian/docker/docker-compose.dev.yml +65 -0
- dt_arena/envs/atlassian/docker/docker-compose.yml +53 -0
- dt_arena/envs/atlassian/docker-compose-hub.yml +57 -0
- dt_arena/envs/atlassian/docker-compose.yml +72 -0
- dt_arena/envs/bigquery/docker-compose.yml +20 -0
- dt_arena/envs/booking/docker-compose.yml +59 -0
- dt_arena/envs/calendar/docker-compose-hub.yml +30 -0
- dt_arena/envs/calendar/docker-compose.yml +42 -0
- dt_arena/envs/custom-website/docker-compose.yml +6 -0
- dt_arena/envs/customer_service/docker-compose.yml +59 -0
- dt_arena/envs/databricks/docker-compose-hub.yml +47 -0
- dt_arena/envs/databricks/docker-compose.yml +51 -0
- dt_arena/envs/ecommerce/docker-compose.yml +6 -0
- dt_arena/envs/ers/docker-compose.yml +36 -0
- dt_arena/envs/ers/hrms/docker/docker-compose.yml +31 -0
- dt_arena/envs/finance/docker-compose.yml +23 -0
- dt_arena/envs/github/docker/docker-compose-hub.yml +50 -0
- dt_arena/envs/github/docker/docker-compose.yml +50 -0
- dt_arena/envs/gmail/docker-compose-hub.yml +51 -0
- dt_arena/envs/gmail/docker-compose.yml +65 -0
- dt_arena/envs/google-form/docker-compose-hub.yml +33 -0
- dt_arena/envs/google-form/docker-compose.yml +41 -0
- dt_arena/envs/googledocs/docker-compose-hub.yml +61 -0
- dt_arena/envs/googledocs/docker-compose.yml +78 -0
- dt_arena/envs/hospital/docker-compose-hub.yml +25 -0
- dt_arena/envs/hospital/docker-compose.yml +27 -0
- dt_arena/envs/legal/docker-compose.yml +22 -0
- dt_arena/envs/linkedin/docker-compose.yml +63 -0
- dt_arena/envs/macos/docker-compose.yml +79 -0
- dt_arena/envs/os-filesystem/docker-compose-hub.yml +16 -0
- dt_arena/envs/os-filesystem/docker-compose.yml +20 -0
- dt_arena/envs/paypal/docker-compose-hub.yml +48 -0
- dt_arena/envs/paypal/docker-compose.yml +63 -0
- dt_arena/envs/research/docker-compose-hub.yml +13 -0
- dt_arena/envs/research/docker-compose.yml +24 -0
- dt_arena/envs/salesforce_crm/docker-compose-hub.yaml +45 -0
- dt_arena/envs/salesforce_crm/docker-compose.yaml +49 -0
- dt_arena/envs/slack/docker-compose-hub.yml +28 -0
- dt_arena/envs/slack/docker-compose.yml +41 -0
- dt_arena/envs/snowflake/docker-compose-hub.yml +41 -0
- dt_arena/envs/snowflake/docker-compose.yml +44 -0
- dt_arena/envs/telecom/docker-compose-hub.yml +16 -0
- dt_arena/envs/telecom/docker-compose.yml +17 -0
- dt_arena/envs/telegram/docker-compose-hub.yml +57 -0
- dt_arena/envs/telegram/docker-compose.yml +62 -0
- dt_arena/envs/terminal/docker-compose-hub.yml +12 -0
- dt_arena/envs/terminal/docker-compose.yml +26 -0
- dt_arena/envs/travel/docker-compose-hub.yml +19 -0
- dt_arena/envs/travel/docker-compose.yml +19 -0
- dt_arena/envs/whatsapp/docker-compose-hub.yml +61 -0
- dt_arena/envs/whatsapp/docker-compose.yml +78 -0
- dt_arena/envs/windows/docker-compose.yml +71 -0
- dt_arena/envs/zoom/docker-compose-hub.yml +27 -0
- dt_arena/envs/zoom/docker-compose.yml +40 -0
- dt_arena/injection_mcp_server/atlassian/env_injection.py +134 -0
- dt_arena/injection_mcp_server/calendar/env_injection.py +217 -0
- dt_arena/injection_mcp_server/custom_website/env_injection.py +97 -0
- dt_arena/injection_mcp_server/customer_service/env_injection.py +659 -0
- dt_arena/injection_mcp_server/databricks/env_injection.py +255 -0
- dt_arena/injection_mcp_server/ecommerce/env_injection.py +110 -0
- dt_arena/injection_mcp_server/finance/env_injection.py +85 -0
- dt_arena/injection_mcp_server/github/env_injection.py +206 -0
- dt_arena/injection_mcp_server/gmail/env_injection.py +211 -0
- dt_arena/injection_mcp_server/google_form/env_injection.py +186 -0
- dt_arena/injection_mcp_server/googledocs/env_injection.py +44 -0
- dt_arena/injection_mcp_server/hospital/env_injection.py +43 -0
- dt_arena/injection_mcp_server/legal/env_injection.py +229 -0
- dt_arena/injection_mcp_server/macos/env_injection.py +272 -0
- dt_arena/injection_mcp_server/os-filesystem/env_injection.py +341 -0
- dt_arena/injection_mcp_server/paypal/env_injection.py +268 -0
- dt_arena/injection_mcp_server/research/env_injection.py +616 -0
- dt_arena/injection_mcp_server/salesforce/env_injection.py +514 -0
- dt_arena/injection_mcp_server/slack/env_injection.py +265 -0
- dt_arena/injection_mcp_server/snowflake/env_injection.py +230 -0
- dt_arena/injection_mcp_server/telecom/env_injection.py +503 -0
- dt_arena/injection_mcp_server/telegram/env_injection.py +171 -0
- dt_arena/injection_mcp_server/terminal/env_injection.py +523 -0
- dt_arena/injection_mcp_server/travel/env_injection.py +173 -0
- dt_arena/injection_mcp_server/whatsapp/env_injection.py +185 -0
- dt_arena/injection_mcp_server/windows/env_injection.py +943 -0
- dt_arena/injection_mcp_server/zoom/env_injection.py +216 -0
- dt_arena/mcp_server/atlassian/main.py +1554 -0
- dt_arena/mcp_server/atlassian/test_server.py +66 -0
- dt_arena/mcp_server/bigquery/main.py +333 -0
- dt_arena/mcp_server/booking/main.py +310 -0
- dt_arena/mcp_server/browser/main.py +1741 -0
- dt_arena/mcp_server/calendar/example_multi_user.py +162 -0
- dt_arena/mcp_server/calendar/main.py +792 -0
- dt_arena/mcp_server/calendar/test_mcp.py +135 -0
- dt_arena/mcp_server/customer_service/main.py +1063 -0
- dt_arena/mcp_server/databricks/main.py +566 -0
- dt_arena/mcp_server/databricks/probe.py +102 -0
- dt_arena/mcp_server/ers/main.py +845 -0
- dt_arena/mcp_server/finance/__init__.py +87 -0
- dt_arena/mcp_server/finance/core/__init__.py +12 -0
- dt_arena/mcp_server/finance/core/data_loader.py +558 -0
- dt_arena/mcp_server/finance/core/portfolio.py +565 -0
- dt_arena/mcp_server/finance/evaluation/__init__.py +20 -0
- dt_arena/mcp_server/finance/evaluation/evaluator.py +217 -0
- dt_arena/mcp_server/finance/evaluation/logger.py +137 -0
- dt_arena/mcp_server/finance/injection/__init__.py +66 -0
- dt_arena/mcp_server/finance/injection/config.py +176 -0
- dt_arena/mcp_server/finance/injection/content.py +755 -0
- dt_arena/mcp_server/finance/injection/html.py +409 -0
- dt_arena/mcp_server/finance/injection/locations.py +167 -0
- dt_arena/mcp_server/finance/injection/methods.py +193 -0
- dt_arena/mcp_server/finance/injection/presets.py +1023 -0
- dt_arena/mcp_server/finance/main.py +361 -0
- dt_arena/mcp_server/finance/run_mcp.py +21 -0
- dt_arena/mcp_server/finance/run_web.py +26 -0
- dt_arena/mcp_server/finance/server/__init__.py +41 -0
- dt_arena/mcp_server/finance/server/extractor.py +1453 -0
- dt_arena/mcp_server/finance/server/extractor_minimal.py +292 -0
- dt_arena/mcp_server/finance/server/extractor_simple.py +1164 -0
- dt_arena/mcp_server/finance/server/injection_mcp.py +865 -0
- dt_arena/mcp_server/finance/server/mcp.py +451 -0
- dt_arena/mcp_server/finance/server/tools/__init__.py +23 -0
- dt_arena/mcp_server/finance/server/tools/account.py +88 -0
- dt_arena/mcp_server/finance/server/tools/browsing.py +328 -0
- dt_arena/mcp_server/finance/server/tools/social.py +73 -0
- dt_arena/mcp_server/finance/server/tools/trading.py +242 -0
- dt_arena/mcp_server/finance/server/tools/utility.py +49 -0
- dt_arena/mcp_server/finance/server/web.py +2139 -0
- dt_arena/mcp_server/finance/tasks/benchmark/__init__.py +28 -0
- dt_arena/mcp_server/finance/tasks/benchmark/attack_pool.py +3026 -0
- dt_arena/mcp_server/finance/tasks/benchmark/attack_runner.py +1315 -0
- dt_arena/mcp_server/finance/tasks/benchmark/finra_requirements.py +1335 -0
- dt_arena/mcp_server/finance/tasks/benchmark/finra_tasks.py +3665 -0
- dt_arena/mcp_server/finance/tasks/benchmark/malicious_tasks.py +2673 -0
- dt_arena/mcp_server/finance/tasks/redteam_suite/run_redteam_suite.py +1713 -0
- dt_arena/mcp_server/finance/test_mcp_tools.py +476 -0
- dt_arena/mcp_server/github/main.py +441 -0
- dt_arena/mcp_server/gmail/main.py +1004 -0
- dt_arena/mcp_server/google_form/main.py +141 -0
- dt_arena/mcp_server/googledocs/main.py +458 -0
- dt_arena/mcp_server/hospital/mcp_server.py +458 -0
- dt_arena/mcp_server/legal/__init__.py +9 -0
- dt_arena/mcp_server/legal/core/__init__.py +14 -0
- dt_arena/mcp_server/legal/core/courtlistener_store.py +762 -0
- dt_arena/mcp_server/legal/core/data_loader.py +266 -0
- dt_arena/mcp_server/legal/core/document_store.py +197 -0
- dt_arena/mcp_server/legal/core/matter_manager.py +466 -0
- dt_arena/mcp_server/legal/main.py +89 -0
- dt_arena/mcp_server/legal/scripts/collect_data.py +988 -0
- dt_arena/mcp_server/legal/server/__init__.py +14 -0
- dt_arena/mcp_server/legal/server/mcp.py +2330 -0
- dt_arena/mcp_server/macos/client_test.py +270 -0
- dt_arena/mcp_server/macos/mcp_server.py +285 -0
- dt_arena/mcp_server/os-filesystem/main.py +1380 -0
- dt_arena/mcp_server/paypal/main.py +501 -0
- dt_arena/mcp_server/research/main.py +777 -0
- dt_arena/mcp_server/salesforce/main.py +2006 -0
- dt_arena/mcp_server/slack/main.py +318 -0
- dt_arena/mcp_server/snowflake/main.py +612 -0
- dt_arena/mcp_server/snowflake/probe.py +183 -0
- dt_arena/mcp_server/telecom/mcp_client.py +423 -0
- dt_arena/mcp_server/telecom/mcp_server.py +1059 -0
- dt_arena/mcp_server/telegram/main.py +338 -0
- dt_arena/mcp_server/terminal/main.py +163 -0
- dt_arena/mcp_server/travel/client_test.py +16 -0
- dt_arena/mcp_server/travel/mcp_server.py +404 -0
- dt_arena/mcp_server/whatsapp/main.py +318 -0
- dt_arena/mcp_server/windows/client_test.py +270 -0
- dt_arena/mcp_server/windows/mcp_server.py +218 -0
- dt_arena/mcp_server/zoom/main.py +466 -0
- dt_arena/src/__init__.py +0 -0
- dt_arena/src/hooks/__init__.py +0 -0
- dt_arena/src/hooks/audit_log.py +30 -0
- dt_arena/src/hooks/hooks.json +3 -0
- dt_arena/src/run_benign.py +142 -0
- dt_arena/src/types/__init__.py +0 -0
- dt_arena/src/types/agent.py +441 -0
- dt_arena/src/types/attacks.py +2 -0
- dt_arena/src/types/environment.py +2 -0
- dt_arena/src/types/hooks.py +174 -0
- dt_arena/src/types/judge.py +52 -0
- dt_arena/src/types/red_teaming_trajectory.py +385 -0
- dt_arena/src/types/task.py +260 -0
- dt_arena/src/types/trajectory.py +315 -0
- dt_arena/utils/__init__.py +1 -0
- dt_arena/utils/atlassian/__init__.py +27 -0
- dt_arena/utils/atlassian/helpers.py +520 -0
- dt_arena/utils/bigquery/__init__.py +1 -0
- dt_arena/utils/bigquery/helpers.py +246 -0
- dt_arena/utils/calendar/__init__.py +1 -0
- dt_arena/utils/calendar/helpers.py +87 -0
- dt_arena/utils/customer_service/__init__.py +17 -0
- dt_arena/utils/customer_service/cs_env_client.py +940 -0
- dt_arena/utils/customer_service/helpers.py +339 -0
- dt_arena/utils/customer_service/judges/__init__.py +20 -0
- dt_arena/utils/customer_service/judges/abusive_comms.py +149 -0
- dt_arena/utils/customer_service/judges/bulk_destructive.py +202 -0
- dt_arena/utils/customer_service/judges/cross_customer.py +481 -0
- dt_arena/utils/customer_service/judges/email_exfiltration.py +320 -0
- dt_arena/utils/customer_service/judges/financial_fraud.py +221 -0
- dt_arena/utils/customer_service/judges/llm_utils.py +264 -0
- dt_arena/utils/customer_service/judges/pii_disclosure.py +837 -0
- dt_arena/utils/customer_service/judges/policy_misrep.py +44 -0
- dt_arena/utils/customer_service/judges/text_utils.py +21 -0
- dt_arena/utils/databricks/__init__.py +2 -0
- dt_arena/utils/databricks/helpers.py +210 -0
- dt_arena/utils/finance/__init__.py +0 -0
- dt_arena/utils/finance/helpers.py +263 -0
- dt_arena/utils/github/__init__.py +1 -0
- dt_arena/utils/github/helpers.py +249 -0
- dt_arena/utils/gmail/__init__.py +1 -0
- dt_arena/utils/gmail/helpers.py +344 -0
- dt_arena/utils/google_form/__init__.py +2 -0
- dt_arena/utils/google_form/helpers.py +133 -0
- dt_arena/utils/legal/__init__.py +0 -0
- dt_arena/utils/legal/helpers.py +228 -0
- dt_arena/utils/macos/__init__.py +0 -0
- dt_arena/utils/macos/env_setup.py +215 -0
- dt_arena/utils/macos/helpers.py +61 -0
- dt_arena/utils/os_filesystem/__init__.py +1 -0
- dt_arena/utils/os_filesystem/helpers.py +366 -0
- dt_arena/utils/paypal/__init__.py +1 -0
- dt_arena/utils/paypal/helpers.py +178 -0
- dt_arena/utils/port_allocator.py +266 -0
- dt_arena/utils/research/__init__.py +0 -0
- dt_arena/utils/research/helpers.py +251 -0
- dt_arena/utils/salesforce/__init__.py +1 -0
- dt_arena/utils/salesforce/helpers.py +719 -0
- dt_arena/utils/slack/__init__.py +1 -0
- dt_arena/utils/slack/helpers.py +176 -0
- dt_arena/utils/snowflake/__init__.py +1 -0
- dt_arena/utils/snowflake/helpers.py +166 -0
- dt_arena/utils/telecom/__init__.py +1 -0
- dt_arena/utils/telecom/helpers.py +760 -0
- dt_arena/utils/telegram/__init__.py +0 -0
- dt_arena/utils/telegram/helpers.py +174 -0
- dt_arena/utils/terminal/__init__.py +0 -0
- dt_arena/utils/terminal/helpers.py +20 -0
- dt_arena/utils/travel/__init__.py +0 -0
- dt_arena/utils/travel/env_client.py +537 -0
- dt_arena/utils/travel/llm_judge.py +137 -0
- dt_arena/utils/travel/prompts.py +64 -0
- dt_arena/utils/utils/__init__.py +122 -0
- dt_arena/utils/whatsapp/__init__.py +0 -0
- dt_arena/utils/whatsapp/helpers.py +226 -0
- dt_arena/utils/windows/__init__.py +0 -0
- dt_arena/utils/windows/env_reset.py +224 -0
- dt_arena/utils/windows/env_setup.py +280 -0
- dt_arena/utils/windows/exfil_helpers.py +170 -0
- dt_arena/utils/windows/helpers.py +74 -0
- dt_arena/utils/zoom/__init__.py +1 -0
- dt_arena/utils/zoom/helpers.py +70 -0
- eval/__init__.py +1 -0
- eval/evaluation.py +426 -0
- eval/task_runner.py +449 -0
- utils/__init__.py +148 -0
- utils/agent_helpers.py +308 -0
- utils/agent_wrapper.py +189 -0
- utils/compose_utils.py +135 -0
- utils/config.py +77 -0
- utils/env_helpers.py +104 -0
- utils/eval_stats.py +88 -0
- utils/injection_helpers.py +429 -0
- utils/injection_mcp_helpers.py +152 -0
- utils/judge_helpers.py +181 -0
- utils/judge_utils.py +472 -0
- utils/llm.py +196 -0
- utils/logging.py +45 -0
- utils/mcp_helpers.py +232 -0
- utils/mcp_manager.py +235 -0
- utils/memory_guard.py +18 -0
- utils/red_teaming_sandbox.py +476 -0
- utils/reset_helpers.py +318 -0
- utils/resource_manager.py +370 -0
- utils/skill_helpers.py +447 -0
- utils/task_executor.py +904 -0
- utils/task_helpers.py +270 -0
- utils/template_helpers.py +179 -0
|
@@ -0,0 +1,476 @@
|
|
|
1
|
+
#!/usr/bin/env python3
|
|
2
|
+
"""
|
|
3
|
+
Comprehensive MCP Tools Test Suite
|
|
4
|
+
|
|
5
|
+
This script tests each MCP tool and verifies:
|
|
6
|
+
1. HTML pages are rendered correctly
|
|
7
|
+
2. Extraction captures all relevant content
|
|
8
|
+
3. Trading and comment APIs work correctly
|
|
9
|
+
4. Parameters are diverse and support different functions
|
|
10
|
+
|
|
11
|
+
Run with web server already running on http://localhost:5000
|
|
12
|
+
"""
|
|
13
|
+
|
|
14
|
+
import sys
|
|
15
|
+
from pathlib import Path
|
|
16
|
+
|
|
17
|
+
# Add paths
|
|
18
|
+
sys.path.insert(0, str(Path(__file__).parent))
|
|
19
|
+
sys.path.insert(0, str(Path(__file__).parent.parent.parent.parent))
|
|
20
|
+
|
|
21
|
+
import httpx
|
|
22
|
+
import json
|
|
23
|
+
from server.extractor import HtmlExtractor
|
|
24
|
+
|
|
25
|
+
# Test configuration
|
|
26
|
+
WEB_URL = "http://localhost:5000"
|
|
27
|
+
client = httpx.Client(timeout=30.0)
|
|
28
|
+
extractor = HtmlExtractor()
|
|
29
|
+
|
|
30
|
+
# Tool parameter documentation
|
|
31
|
+
MCP_TOOLS_DOCUMENTATION = """
|
|
32
|
+
================================================================================
|
|
33
|
+
FINANCE MCP TOOLS - COMPLETE REFERENCE
|
|
34
|
+
================================================================================
|
|
35
|
+
|
|
36
|
+
1. browse_stock(symbol, section="quote")
|
|
37
|
+
- symbol: Stock ticker (e.g., "AAPL", "NVDA", "MSFT", "TSLA")
|
|
38
|
+
- section: "quote", "news", "chart", "analysis", "profile"
|
|
39
|
+
|
|
40
|
+
2. browse_news(symbol="", limit=10, month="", date="")
|
|
41
|
+
- symbol: Empty for global news, or ticker for stock-specific news
|
|
42
|
+
- limit: 1-50, default 10
|
|
43
|
+
- month: "YYYY-MM" format filter
|
|
44
|
+
- date: "YYYY-MM-DD" exact date filter
|
|
45
|
+
|
|
46
|
+
3. browse_article(symbol, article_id)
|
|
47
|
+
- symbol: Stock ticker
|
|
48
|
+
- article_id: Numeric article ID from browse_news results
|
|
49
|
+
|
|
50
|
+
4. browse_portfolio(section="holdings", symbol="")
|
|
51
|
+
- section: "holdings", "trade", "orders"
|
|
52
|
+
- symbol: Pre-fill symbol on trade page
|
|
53
|
+
|
|
54
|
+
5. browse_options(symbol, expiration="")
|
|
55
|
+
- symbol: Stock ticker
|
|
56
|
+
- expiration: "YYYY-MM-DD" expiration date filter
|
|
57
|
+
|
|
58
|
+
6. browse_markets()
|
|
59
|
+
- No parameters, returns market overview
|
|
60
|
+
|
|
61
|
+
7. browse_stock_list()
|
|
62
|
+
- No parameters, returns all tradeable stocks
|
|
63
|
+
|
|
64
|
+
8. search(query, search_type="all")
|
|
65
|
+
- query: Search term
|
|
66
|
+
- search_type: "all", "stocks", "news"
|
|
67
|
+
|
|
68
|
+
9. trade_stock(action, symbol, quantity)
|
|
69
|
+
- action: "buy" or "sell"
|
|
70
|
+
- symbol: Stock ticker
|
|
71
|
+
- quantity: Number of shares (positive integer)
|
|
72
|
+
|
|
73
|
+
10. trade_option(action, symbol, option_type, strike, expiration, quantity, premium)
|
|
74
|
+
- action: "buy" or "sell"
|
|
75
|
+
- symbol: Underlying stock ticker
|
|
76
|
+
- option_type: "call" or "put"
|
|
77
|
+
- strike: Strike price (float)
|
|
78
|
+
- expiration: "YYYY-MM-DD" format
|
|
79
|
+
- quantity: Number of contracts
|
|
80
|
+
- premium: Premium per share (float)
|
|
81
|
+
|
|
82
|
+
11. post_comment(symbol, article_id, content, sentiment="neutral")
|
|
83
|
+
- symbol: Stock ticker
|
|
84
|
+
- article_id: Article ID
|
|
85
|
+
- content: Comment text
|
|
86
|
+
- sentiment: "bullish", "bearish", or "neutral"
|
|
87
|
+
================================================================================
|
|
88
|
+
"""
|
|
89
|
+
|
|
90
|
+
def test_tool(name: str, url: str, page_type: str, checks: list):
|
|
91
|
+
"""Test a single tool extraction."""
|
|
92
|
+
print(f"\n{'='*60}")
|
|
93
|
+
print(f"Testing: {name}")
|
|
94
|
+
print(f"URL: {url}")
|
|
95
|
+
print(f"Page Type: {page_type}")
|
|
96
|
+
print(f"{'='*60}")
|
|
97
|
+
|
|
98
|
+
try:
|
|
99
|
+
response = client.get(f"{WEB_URL}{url}")
|
|
100
|
+
if response.status_code != 200:
|
|
101
|
+
print(f"❌ HTTP Error: {response.status_code}")
|
|
102
|
+
return False
|
|
103
|
+
|
|
104
|
+
html = response.text
|
|
105
|
+
extracted = extractor.extract(html, page_type)
|
|
106
|
+
|
|
107
|
+
print(f"\n--- Extracted Content (first 1500 chars) ---")
|
|
108
|
+
print(extracted[:1500])
|
|
109
|
+
print(f"\n--- Checks ---")
|
|
110
|
+
|
|
111
|
+
all_passed = True
|
|
112
|
+
for check_name, check_pattern in checks:
|
|
113
|
+
if check_pattern.lower() in extracted.lower():
|
|
114
|
+
print(f" ✅ {check_name}: Found")
|
|
115
|
+
else:
|
|
116
|
+
print(f" ❌ {check_name}: NOT FOUND")
|
|
117
|
+
all_passed = False
|
|
118
|
+
|
|
119
|
+
return all_passed
|
|
120
|
+
except Exception as e:
|
|
121
|
+
print(f"❌ Error: {e}")
|
|
122
|
+
return False
|
|
123
|
+
|
|
124
|
+
|
|
125
|
+
def main():
|
|
126
|
+
print("=" * 70)
|
|
127
|
+
print("Finance MCP Tools Test Suite")
|
|
128
|
+
print("=" * 70)
|
|
129
|
+
|
|
130
|
+
results = {}
|
|
131
|
+
|
|
132
|
+
# ===================================================================
|
|
133
|
+
# 1. browse_stock(symbol, section="quote")
|
|
134
|
+
# ===================================================================
|
|
135
|
+
results["browse_stock(quote)"] = test_tool(
|
|
136
|
+
"browse_stock('AAPL', 'quote')",
|
|
137
|
+
"/quote/AAPL/",
|
|
138
|
+
"quote",
|
|
139
|
+
[
|
|
140
|
+
("Company Name", "Apple"),
|
|
141
|
+
("Stock Symbol", "AAPL"),
|
|
142
|
+
("Price", "278"),
|
|
143
|
+
("Key Statistics", "Statistics"),
|
|
144
|
+
("Trade Panel", "Cash"),
|
|
145
|
+
]
|
|
146
|
+
)
|
|
147
|
+
|
|
148
|
+
# ===================================================================
|
|
149
|
+
# 2. browse_stock(symbol, section="news")
|
|
150
|
+
# ===================================================================
|
|
151
|
+
results["browse_stock(news)"] = test_tool(
|
|
152
|
+
"browse_stock('AAPL', 'news')",
|
|
153
|
+
"/quote/AAPL/news/",
|
|
154
|
+
"news",
|
|
155
|
+
[
|
|
156
|
+
("News Header", "News"),
|
|
157
|
+
("Article ID", "Article ID"),
|
|
158
|
+
("Headline", "📌"),
|
|
159
|
+
("Source info", "Source"),
|
|
160
|
+
]
|
|
161
|
+
)
|
|
162
|
+
|
|
163
|
+
# ===================================================================
|
|
164
|
+
# 3. browse_stock(symbol, section="chart")
|
|
165
|
+
# ===================================================================
|
|
166
|
+
results["browse_stock(chart)"] = test_tool(
|
|
167
|
+
"browse_stock('AAPL', 'chart')",
|
|
168
|
+
"/quote/AAPL/chart/",
|
|
169
|
+
"chart",
|
|
170
|
+
[
|
|
171
|
+
("Chart Title", "Chart"),
|
|
172
|
+
("Symbol", "AAPL"),
|
|
173
|
+
("Price", "Price"),
|
|
174
|
+
]
|
|
175
|
+
)
|
|
176
|
+
|
|
177
|
+
# ===================================================================
|
|
178
|
+
# 4. browse_stock(symbol, section="analysis")
|
|
179
|
+
# ===================================================================
|
|
180
|
+
results["browse_stock(analysis)"] = test_tool(
|
|
181
|
+
"browse_stock('AAPL', 'analysis')",
|
|
182
|
+
"/quote/AAPL/analysis/",
|
|
183
|
+
"analysis",
|
|
184
|
+
[
|
|
185
|
+
("Analysis Title", "Analysis"),
|
|
186
|
+
("Symbol", "AAPL"),
|
|
187
|
+
]
|
|
188
|
+
)
|
|
189
|
+
|
|
190
|
+
# ===================================================================
|
|
191
|
+
# 5. browse_stock(symbol, section="profile")
|
|
192
|
+
# ===================================================================
|
|
193
|
+
results["browse_stock(profile)"] = test_tool(
|
|
194
|
+
"browse_stock('AAPL', 'profile')",
|
|
195
|
+
"/quote/AAPL/profile/",
|
|
196
|
+
"profile",
|
|
197
|
+
[
|
|
198
|
+
("Profile Title", "Profile"),
|
|
199
|
+
("Company Name", "Apple"),
|
|
200
|
+
("Sector/Industry", "Technology"),
|
|
201
|
+
]
|
|
202
|
+
)
|
|
203
|
+
|
|
204
|
+
# ===================================================================
|
|
205
|
+
# 6. browse_news() - global news
|
|
206
|
+
# ===================================================================
|
|
207
|
+
results["browse_news()"] = test_tool(
|
|
208
|
+
"browse_news()",
|
|
209
|
+
"/news/?limit=10",
|
|
210
|
+
"news_center",
|
|
211
|
+
[
|
|
212
|
+
("News Center", "News"),
|
|
213
|
+
("Trending", "Trending"),
|
|
214
|
+
("Article entries", "["),
|
|
215
|
+
]
|
|
216
|
+
)
|
|
217
|
+
|
|
218
|
+
# ===================================================================
|
|
219
|
+
# 7. browse_news(symbol) - stock-specific news
|
|
220
|
+
# ===================================================================
|
|
221
|
+
results["browse_news(AAPL)"] = test_tool(
|
|
222
|
+
"browse_news('AAPL', limit=5)",
|
|
223
|
+
"/quote/AAPL/news/?limit=5",
|
|
224
|
+
"news",
|
|
225
|
+
[
|
|
226
|
+
("News for Symbol", "News"),
|
|
227
|
+
("Article ID", "Article ID"),
|
|
228
|
+
]
|
|
229
|
+
)
|
|
230
|
+
|
|
231
|
+
# ===================================================================
|
|
232
|
+
# 8. browse_article(symbol, article_id)
|
|
233
|
+
# ===================================================================
|
|
234
|
+
# First get an article ID from news
|
|
235
|
+
try:
|
|
236
|
+
news_response = client.get(f"{WEB_URL}/news/?limit=1")
|
|
237
|
+
# Just test with a known article structure
|
|
238
|
+
results["browse_article()"] = test_tool(
|
|
239
|
+
"browse_article('AAPL', 137649646)",
|
|
240
|
+
"/news/AAPL/article/137649646/",
|
|
241
|
+
"article",
|
|
242
|
+
[
|
|
243
|
+
("Article Title", "📰"),
|
|
244
|
+
("Article Content", "Content"),
|
|
245
|
+
("Comments Section", "Comment"),
|
|
246
|
+
]
|
|
247
|
+
)
|
|
248
|
+
except:
|
|
249
|
+
results["browse_article()"] = False
|
|
250
|
+
|
|
251
|
+
# ===================================================================
|
|
252
|
+
# 9. browse_portfolio(section="holdings")
|
|
253
|
+
# ===================================================================
|
|
254
|
+
results["browse_portfolio(holdings)"] = test_tool(
|
|
255
|
+
"browse_portfolio('holdings')",
|
|
256
|
+
"/portfolios/",
|
|
257
|
+
"portfolio",
|
|
258
|
+
[
|
|
259
|
+
("Portfolio Title", "Portfolio"),
|
|
260
|
+
("Cash Balance", "Cash"),
|
|
261
|
+
("Holdings", "Holding"),
|
|
262
|
+
]
|
|
263
|
+
)
|
|
264
|
+
|
|
265
|
+
# ===================================================================
|
|
266
|
+
# 10. browse_portfolio(section="trade")
|
|
267
|
+
# ===================================================================
|
|
268
|
+
results["browse_portfolio(trade)"] = test_tool(
|
|
269
|
+
"browse_portfolio('trade', 'AAPL')",
|
|
270
|
+
"/portfolios/trade/?symbol=AAPL",
|
|
271
|
+
"trade",
|
|
272
|
+
[
|
|
273
|
+
("Trading Page", "Trading"),
|
|
274
|
+
]
|
|
275
|
+
)
|
|
276
|
+
|
|
277
|
+
# ===================================================================
|
|
278
|
+
# 11. browse_portfolio(section="orders")
|
|
279
|
+
# ===================================================================
|
|
280
|
+
results["browse_portfolio(orders)"] = test_tool(
|
|
281
|
+
"browse_portfolio('orders')",
|
|
282
|
+
"/portfolios/orders/",
|
|
283
|
+
"orders",
|
|
284
|
+
[
|
|
285
|
+
("Order History", "Order"),
|
|
286
|
+
]
|
|
287
|
+
)
|
|
288
|
+
|
|
289
|
+
# ===================================================================
|
|
290
|
+
# 12. browse_options(symbol)
|
|
291
|
+
# ===================================================================
|
|
292
|
+
results["browse_options()"] = test_tool(
|
|
293
|
+
"browse_options('AAPL')",
|
|
294
|
+
"/portfolios/options/AAPL/",
|
|
295
|
+
"options",
|
|
296
|
+
[
|
|
297
|
+
("Options Chain", "Options"),
|
|
298
|
+
]
|
|
299
|
+
)
|
|
300
|
+
|
|
301
|
+
# ===================================================================
|
|
302
|
+
# 13. browse_markets()
|
|
303
|
+
# ===================================================================
|
|
304
|
+
results["browse_markets()"] = test_tool(
|
|
305
|
+
"browse_markets()",
|
|
306
|
+
"/markets/",
|
|
307
|
+
"markets",
|
|
308
|
+
[
|
|
309
|
+
("Market Overview", "Market"),
|
|
310
|
+
("Indices", "NASDAQ"),
|
|
311
|
+
("Gainers or Losers", "Gainer"),
|
|
312
|
+
("Trending", "Trending"),
|
|
313
|
+
]
|
|
314
|
+
)
|
|
315
|
+
|
|
316
|
+
# ===================================================================
|
|
317
|
+
# 14. browse_stock_list()
|
|
318
|
+
# ===================================================================
|
|
319
|
+
results["browse_stock_list()"] = test_tool(
|
|
320
|
+
"browse_stock_list()",
|
|
321
|
+
"/markets/stocks/",
|
|
322
|
+
"stock_list",
|
|
323
|
+
[
|
|
324
|
+
("Stock List", "Stock"),
|
|
325
|
+
("Multiple Symbols", "AAPL"),
|
|
326
|
+
]
|
|
327
|
+
)
|
|
328
|
+
|
|
329
|
+
# ===================================================================
|
|
330
|
+
# 15. search(query)
|
|
331
|
+
# ===================================================================
|
|
332
|
+
results["search()"] = test_tool(
|
|
333
|
+
"search('apple')",
|
|
334
|
+
"/search/?q=apple&type=all",
|
|
335
|
+
"search",
|
|
336
|
+
[
|
|
337
|
+
("Search Results", "Search"),
|
|
338
|
+
]
|
|
339
|
+
)
|
|
340
|
+
|
|
341
|
+
# ===================================================================
|
|
342
|
+
# Summary
|
|
343
|
+
# ===================================================================
|
|
344
|
+
print("\n" + "=" * 70)
|
|
345
|
+
print("TEST SUMMARY")
|
|
346
|
+
print("=" * 70)
|
|
347
|
+
|
|
348
|
+
passed = sum(1 for v in results.values() if v)
|
|
349
|
+
total = len(results)
|
|
350
|
+
|
|
351
|
+
for name, result in results.items():
|
|
352
|
+
status = "✅ PASS" if result else "❌ FAIL"
|
|
353
|
+
print(f" {status}: {name}")
|
|
354
|
+
|
|
355
|
+
print(f"\nTotal: {passed}/{total} tests passed")
|
|
356
|
+
|
|
357
|
+
if passed < total:
|
|
358
|
+
print("\n⚠️ Some tests failed. Check extractor or templates.")
|
|
359
|
+
else:
|
|
360
|
+
print("\n✅ All extraction tests passed!")
|
|
361
|
+
|
|
362
|
+
# ===================================================================
|
|
363
|
+
# API TESTS (Trading & Comments)
|
|
364
|
+
# ===================================================================
|
|
365
|
+
print("\n" + "=" * 70)
|
|
366
|
+
print("API TESTS (Trading & Comments)")
|
|
367
|
+
print("=" * 70)
|
|
368
|
+
|
|
369
|
+
api_results = {}
|
|
370
|
+
|
|
371
|
+
# Test buy stock
|
|
372
|
+
print("\n📈 Test: trade_stock('buy', 'NVDA', 5)")
|
|
373
|
+
try:
|
|
374
|
+
response = client.post(f"{WEB_URL}/api/portfolio/buy", json={"symbol": "NVDA", "shares": 5})
|
|
375
|
+
result = response.json()
|
|
376
|
+
if result.get("success"):
|
|
377
|
+
print(f" ✅ Buy successful: ${result.get('transaction', {}).get('total', 0):.2f}")
|
|
378
|
+
api_results["trade_stock(buy)"] = True
|
|
379
|
+
else:
|
|
380
|
+
print(f" ❌ Buy failed: {result.get('error')}")
|
|
381
|
+
api_results["trade_stock(buy)"] = False
|
|
382
|
+
except Exception as e:
|
|
383
|
+
print(f" ❌ Error: {e}")
|
|
384
|
+
api_results["trade_stock(buy)"] = False
|
|
385
|
+
|
|
386
|
+
# Test sell stock
|
|
387
|
+
print("\n📉 Test: trade_stock('sell', 'NVDA', 2)")
|
|
388
|
+
try:
|
|
389
|
+
response = client.post(f"{WEB_URL}/api/portfolio/sell", json={"symbol": "NVDA", "shares": 2})
|
|
390
|
+
result = response.json()
|
|
391
|
+
if result.get("success"):
|
|
392
|
+
print(f" ✅ Sell successful: ${result.get('transaction', {}).get('total', 0):.2f}")
|
|
393
|
+
api_results["trade_stock(sell)"] = True
|
|
394
|
+
else:
|
|
395
|
+
print(f" ❌ Sell failed: {result.get('error')}")
|
|
396
|
+
api_results["trade_stock(sell)"] = False
|
|
397
|
+
except Exception as e:
|
|
398
|
+
print(f" ❌ Error: {e}")
|
|
399
|
+
api_results["trade_stock(sell)"] = False
|
|
400
|
+
|
|
401
|
+
# Test post comment
|
|
402
|
+
print("\n💬 Test: post_comment('AAPL', 137649646, 'Great article!', 'bullish')")
|
|
403
|
+
try:
|
|
404
|
+
response = client.post(f"{WEB_URL}/api/comment", json={
|
|
405
|
+
"symbol": "AAPL",
|
|
406
|
+
"article_id": 137649646,
|
|
407
|
+
"content": "Testing the comment system!",
|
|
408
|
+
"sentiment": "bullish",
|
|
409
|
+
"user": "TestBot"
|
|
410
|
+
})
|
|
411
|
+
result = response.json()
|
|
412
|
+
if result.get("success"):
|
|
413
|
+
print(f" ✅ Comment posted: {result.get('comment', {}).get('content', '')[:50]}")
|
|
414
|
+
api_results["post_comment()"] = True
|
|
415
|
+
else:
|
|
416
|
+
print(f" ❌ Comment failed: {result.get('error')}")
|
|
417
|
+
api_results["post_comment()"] = False
|
|
418
|
+
except Exception as e:
|
|
419
|
+
print(f" ❌ Error: {e}")
|
|
420
|
+
api_results["post_comment()"] = False
|
|
421
|
+
|
|
422
|
+
# Test buy option
|
|
423
|
+
print("\n📊 Test: trade_option('buy', 'AAPL', 'call', 280, '2025-12-19', 1, 5.50)")
|
|
424
|
+
try:
|
|
425
|
+
response = client.post(f"{WEB_URL}/api/portfolio/buy_option", json={
|
|
426
|
+
"symbol": "AAPL",
|
|
427
|
+
"option_type": "call",
|
|
428
|
+
"strike": 280.0,
|
|
429
|
+
"expiration": "2025-12-19",
|
|
430
|
+
"quantity": 1,
|
|
431
|
+
"premium": 5.50
|
|
432
|
+
})
|
|
433
|
+
result = response.json()
|
|
434
|
+
if result.get("success"):
|
|
435
|
+
print(f" ✅ Option bought: ${result.get('transaction', {}).get('total', 0):.2f}")
|
|
436
|
+
api_results["trade_option(buy)"] = True
|
|
437
|
+
else:
|
|
438
|
+
print(f" ❌ Option buy failed: {result.get('error')}")
|
|
439
|
+
api_results["trade_option(buy)"] = False
|
|
440
|
+
except Exception as e:
|
|
441
|
+
print(f" ❌ Error: {e}")
|
|
442
|
+
api_results["trade_option(buy)"] = False
|
|
443
|
+
|
|
444
|
+
# ===================================================================
|
|
445
|
+
# FINAL SUMMARY
|
|
446
|
+
# ===================================================================
|
|
447
|
+
print("\n" + "=" * 70)
|
|
448
|
+
print("FINAL SUMMARY")
|
|
449
|
+
print("=" * 70)
|
|
450
|
+
|
|
451
|
+
print("\n📋 Extraction Tests:")
|
|
452
|
+
for name, result in results.items():
|
|
453
|
+
status = "✅" if result else "❌"
|
|
454
|
+
print(f" {status} {name}")
|
|
455
|
+
|
|
456
|
+
print("\n📋 API Tests:")
|
|
457
|
+
for name, result in api_results.items():
|
|
458
|
+
status = "✅" if result else "❌"
|
|
459
|
+
print(f" {status} {name}")
|
|
460
|
+
|
|
461
|
+
extraction_passed = sum(1 for v in results.values() if v)
|
|
462
|
+
api_passed = sum(1 for v in api_results.values() if v)
|
|
463
|
+
total_passed = extraction_passed + api_passed
|
|
464
|
+
total_tests = len(results) + len(api_results)
|
|
465
|
+
|
|
466
|
+
print(f"\n📊 Total: {total_passed}/{total_tests} tests passed")
|
|
467
|
+
|
|
468
|
+
# Print tool documentation
|
|
469
|
+
print(MCP_TOOLS_DOCUMENTATION)
|
|
470
|
+
|
|
471
|
+
return 0 if total_passed == total_tests else 1
|
|
472
|
+
|
|
473
|
+
|
|
474
|
+
if __name__ == "__main__":
|
|
475
|
+
sys.exit(main())
|
|
476
|
+
|