OmniAtlas
Demo WorkspaceRepositories / OpenClaw
Focused tour of agent execution, memory search, and tool orchestration; not a whole-project map.
OpenClaw is an agent runtime with explicit orchestration, tool-policy, and memory boundaries. This curated snapshot focuses on the path that executes an embedded agent, the construction and enforcement of its tool surface, and the memory plugin's bounded hybrid retrieval path.
Agent requests enter through the embedded runner, become a prepared runtime, and move through a retry-aware attempt loop before harness dispatch. Each attempt constructs a capability-aware tool inventory, applies ordered policy layers, and wraps allowed calls with pre-execution enforcement. The memory-core plugin contributes memory_search, resolves an agent-scoped backend, queries built-in keyword and vector indexes, and merges their results through hybrid ranking.
Execution paths
Traces an embedded agent request from lifecycle-scoped orchestration through the prepared loop, runtime-plan construction, and prepared attempt execution.
Follows memory_search from deadline-bounded manager setup through agent-scoped resolution, the built-in vector path, and hybrid score construction.
Shows how one attempt chooses a tool construction plan, selects tool categories, filters them through policy, and blocks denied calls before execution.
Change surface
Defines the memory_search tool and coordinates bounded retrieval, visibility filtering, ranking, and citations.
Builds the base tool set for an attempt from model, workspace, sandbox, and capability constraints.
Public embedded-agent entry point that resolves runtime configuration and coordinates one prepared run.
Assembles core, shell, channel, plugin, and searchable tools into the effective agent tool surface.
Working conventions
Runs built-in memory retrieval across vector and keyword indexes before combining ranked results.
Runtime boundaries
Defines the memory_search tool and coordinates bounded retrieval, visibility filtering, ranking, and citations.
Public embedded-agent entry point that resolves runtime configuration and coordinates one prepared run.
Builds the base tool set for an attempt from model, workspace, sandbox, and capability constraints.
9 named architectural layers represented in this snapshot.
