OpenClaw

Returns completed attempts or carries normalized retry state into the next loop iteration.

Builds a runtime plan from model, harness, authentication, configuration, and workspace context.

Records blocked hook outcomes and stops denied tool calls before execution.

Starts execution with prepared bootstrap, tool, session, prompt, and catalog inputs.

Captures the run lifecycle generation and enters internal orchestration with the selected configuration.

Combines vector relevance with lexical ranking or path relevance using configured weights.

Queries bounded vector candidates within the selected source scope.

Loads the manager factory and forwards agent identity, request purpose, and optional host hooks.

Returns the built-in manager when the configured backend is not QMD.

Starts deadline-bounded backend resolution and manager setup for the active agent and purpose.

Hands a fully prepared agent request to the controlled execution loop.

Chooses a construction plan from model tool support, raw mode, disabling, and the effective allowlist.

Filters the current tool list through an expanded policy while recording audit details.

Selects base coding, shell, channel, OpenClaw, and plugin categories for the tool inventory.