action_runtime

Action Runtime Examples

These examples are written for the Action-based runtime. Every numbered example creates a request-scoped turn, passes turn.prompt into agent.get_action_result(...) to inspect intermediate ActionResult records first, then uses turn.get_result() to produce the final DeepSeek reply through OpenAICompatible.

Before running them, set:

• DEEPSEEK_API_KEY
• DEEPSEEK_BASE_URL (optional, defaults to https://api.deepseek.com/v1)
• DEEPSEEK_DEFAULT_MODEL (optional, defaults to deepseek-chat)

Example groups:

• Function actions
  • 1_1_function_action_func_deepseek.py
  • 1_2_function_register_action_deepseek.py
• MCP actions
  • 2_1_mcp_stdio_action_deepseek.py
  • 2_2_mcp_http_action_deepseek.py
  • _calculator_mcp_server.py is the shared local MCP server
• Built-in action packages
  • See examples/builtin_actions/ for Search/Browse package mounting and local browse examples.
• Sandbox actions
  • 3_1_python_sandbox_action_deepseek.py
  • 3_2_bash_sandbox_action_deepseek.py
  • 3_3_third_party_sandlock_action_deepseek.py
  • 3_4_third_party_docker_sandbox_action_deepseek.py
  • 3_5_action_execution_recall_local.py
• Plugin customization examples
  • 4_1_custom_action_executor_plugin_local.py
  • 4_2_custom_action_runtime_plugin_local.py
  • 4_3_custom_action_flow_plugin_local.py
• Cookbook patterns
  • See examples/cookbook/ for model-backed Action loop, router, concurrent todo, reflection, and safe shell policy patterns adapted from practical app-development training material.

Shared helper:

• _shared.py loads .env, configures DeepSeek, prints intermediate action records, and prints final reply logs from result.full_result_data["extra"]["action_logs"]

Notes:

• Every numbered example registers or imports actions, mounts them on an agent, runs a real prompt, prints intermediate action records, and then prints the final reply plus extra.action_logs.
• Future action examples must be runnable in their declared environment and must include an Expected key output comment in the file. For model-backed examples, the comment should describe the stable action/result shape rather than an exact model sentence.
• Cookbook examples must call DeepSeek or local Ollama for planner/classifier/evaluator/reviser steps. Local functions are acceptable only as the business capability being called by an Action or workflow step, not as a model-decision substitute.
• By default, agent.get_action_result(prompt=turn.prompt) stores action_results on that turn prompt so the following turn.get_result() can reuse those intermediate results instead of executing the action loop again. Pass store_for_reply=False when you only want isolated inspection.
• Instruction-heavy actions expose compact model_digest data to later model context and keep full raw input/output behind artifact_refs; 3_5_action_execution_recall_local.py shows explicit artifact recall through agent.action.read_action_artifact(...).
• 3_3_third_party_sandlock_action_deepseek.py demonstrates a Linux SandLock third-party sandbox executor registered through the new ActionExecutor plugin type.
• 3_4_third_party_docker_sandbox_action_deepseek.py demonstrates a local Docker third-party sandbox executor registered through the new ActionExecutor plugin type.
• 4_1 to 4_3 focus on extension points for ActionExecutor, ActionRuntime, and ActionFlow, and also run through an agent with DeepSeek for the final reply.
• The SandLock example requires Linux 6.7+ and pip install sandlock.
• The Docker sandbox example requires a local Docker daemon that is running and not paused. It auto-pulls alpine:3.20 when the image is not available locally.