execution-modes.md

Execution Modes

Back to README · See also evaluation.md and openclaw-setup.md

Mode summary

Mode	Behavior	Typical use
mock	simulates openclaw * commands in process	unit tests and narrow backend checks
multi	routes each command family to its local backend or skill	default benchmark evaluation
real	runs openclaw * through the real CLI subprocess path while keeping routed local skills for the rest	integration testing
hybrid	runs a live OpenClaw gateway and the same routed skill stack, with optional online providers enabled	realistic end-to-end evaluation

multi is the default benchmark mode because it preserves interactive state across command families without requiring a full online deployment.

What changes by mode

The task schema and evaluator stay the same across modes. What changes is backend realization.

• In mock, the environment uses a narrow in-process mock path.
• In multi, commands are routed through the local backend and skill stack.
• In real, openclaw * is executed through the real CLI subprocess path.
• In hybrid, openclaw runs against a live gateway while the rest of the routed stack remains in place.

Routing vs provider realization

Execution routing and provider realization are separate concerns.

In real and hybrid, command families such as calendar, email, weather, tasks, and file still execute through the normal routed path and therefore appear in traces, effects, and evaluator-visible state. The distinction is whether those skills stay local or trigger online provider behavior when configured.

Practical guidance

• Use mock for unit tests and narrow runtime debugging.
• Use multi for benchmark development, reproducible evaluation, and most reported runs.
• Use real when you specifically need the real openclaw CLI subprocess path.
• Use hybrid when you want a live OpenClaw-backed run and accept the extra setup and provider variability.