Claude Session Driver

Why it's ranked here

The tool offers a credible orchestration layer rather than a thin process launcher. Shared commands across three agent harnesses reduce workflow branching, while structured event logs make parallel work observable. The central concern is deliberate: workers bypass permission prompts, and monitoring records actions without preventing them.

What's good

It supports delegation, fan-out, pipelines, supervised conversations, and human handoff through the same command model. Prompt delivery waits for an explicit submission event and retries Enter when interactive input is slow. Worker status distinguishes active, terminated, missing, and not-yet-registered sessions. Process execution uses argument arrays instead of a shell, reducing quoting and injection hazards.

Tradeoffs

Every worker runs without interactive permission approval, so task scope and working directories need careful control. Event collection is observation-only and cannot block a dangerous tool call. The setup depends on tmux, Node, and at least one supported agent command. Existing-session adoption works only with Claude because Codex and Pi assign their own session identifiers.

How to use it well

Use it when one coding agent should coordinate several independent or staged engineering tasks, especially when you want machine-readable progress and occasional human takeover. Start with bounded assignments, inspect events, then combine worker results in the controller. It does not replace sandboxing, policy enforcement, or approval gates for untrusted work.

Observed

License

MIT

Primary language

TypeScript

Packaging

Private pnpm package using ECMAScript modules and tsup

Install surface

Claude plugin distributed through the superpowers marketplace

Interface

Single command-line interface with top-level and per-worker commands

Harness support

Claude Code, Codex, and Pi

Runtime requirements

Node 22.12 or newer, tmux, and at least one supported harness command