Module run_mode 

Run mode — how an eval is dispatched, independent of which harness runs it.

There are two dispatch mechanisms in the code today:

• DispatchMechanism::InSession — the runner hands tasks to in-session subagents (Claude Code’s Task tool). The reference is Claude Code.
• DispatchMechanism::Cli — each task is dispatched through a one-shot harness CLI subprocess (codex exec). The reference is Codex.

These two mechanisms underpin the three user-facing run modes documented in the README: fully-interactive rides on InSession; headless and hybrid both ride on Cli, differing only in whether a human/agent session drives the loop — not in how a single task reaches the harness.

This is distinct from the comparison Mode (new-skill / revision), which selects the two conditions being compared, not the dispatch path.

Structs

HarnessRunCapabilities

Run-option support for a harness’s currently wired dispatch mechanism.

Enums

DispatchMechanism

How a single dispatch is delivered to a harness. The primary code axis for run-mode concerns (next-steps guidance, transcript source).

RunMode

The user-facing run mode — who/what drives the loop plus which dispatch mechanism each task rides on. This is the parity vocabulary documented in the README (§Run modes); it maps down to a DispatchMechanism via RunMode::mechanism. hybrid and headless both ride on Cli and differ only in whether a session drives the loop — a distinction we persist (in conditions.json) even though it doesn’t change how a single task reaches the harness.

Functions

capabilities_for

The focused capability table for generic run option validation.

resolve_run_mode

Resolve the effective run mode for a harness, defaulting per harness when unspecified and rejecting unsupported (harness, mode) combinations. The Err string is operator-facing.