sessionlisted

# Drive agent sessions Request: `$ARGUMENTS` The runtime lives in the project, under `scripts/ai-harness/`. This skill drives it; it does not reimplement it. Long-form background, including what was measured against the real CLI, is in `${CLAUDE_PLUGIN_ROOT}/references/agent-sessions.md` — read it when a decision here is not obvious. This skill declares no pre-approved tools. Each command goes through the normal permission flow, deliberately: these commands dispatch agents. ## Use this only for what the Agent tool cannot do A separate CLI session is the expensive dispatch. It pays a cold start, a filesystem return channel, and a sweep afterwards to catch what it left running. Three properties justify that, and nothing else does: - the lane needs its own **worktree**, - it must **outlive** the session that started it, - it **writes concurrently** with another lane. Isolation alone is not on that list. The Agent tool already runs a subagent in an isolated context, in-process, in parallel, and returns the conclusion directly - so reconnaissance, review, and verification belong there, not here. Routing them through this skill buys nothing and costs a round trip per dispatch. ## Preconditions If `scripts/ai-harness/harness_session.py` does not exist, stop. Either the harness is Lite, which has no generated agents and therefore nothing to manage, or no harness is installed. Say which, and stop. ## Resolve the Python interpreter ```bash python3 --version ``` If that fai