CHAPTER 04
How Capabilities Run
English summary
A capability moves through availability, authorization, visibility, execution, return, and delivery. DeerFlow applies catalog and runtime policy, executes side effects in a sandbox, delegates only bounded work, and keeps artifact declaration separate from run settlement.
01Candidate catalog and two authorization layers
02Deferred MCP discovery and skill activation
03Sandboxed side effects
04Subagent delegation and artifact delivery
Read the implementation
This edition deliberately stops at a concise overview. The links below open the exact DeerFlow source files used by the complete Chinese chapter.
tools/tools.py · candidate aggregation and name prioritytools/builtins/tool_search.py · catalog-scoped promotionskill_tool_policy_middleware.py · visible schema and execution policysandbox/middleware.py · persist lazy sandbox identitytools/builtins/task_tool.py · selective inheritance and no nestingsubagents/executor.py · fresh graph without child checkpointpresent_file_tool.py · artifact update