Slice Runner

Execute an approved plan by bounded Codex-native slices. Do not design the plan here.

Contract

Treat the main Codex thread as architect, integrator, QA, and committer.
Use native subagents only: spawn_agent, wait_agent, send_input, and close_agent.
Never call codex exec, agents-bridge, shell wrappers, or external agent CLIs for the slice work.
Execute one slice at a time unless the plan explicitly marks write scopes as independent.
Keep worker prompts self-contained and strict. Do not let workers choose public APIs, broad architecture, or scope.
Run gates yourself. A worker's "green" claim is not evidence.
Commit locally only when the user or plan asks for commits. Never push without explicit user request.