rutherford-claude-plugin

Set up and verify Rutherford after installing this plugin: confirm the MCP server is connected, run doctor to see which coding agents actually drive, scaffold a config.toml, and install any missing npm ACP adapters. Use right after installing the plugin, when Rutherford tools are missing, or when the user asks to set up or get started with Rutherford.

Diagnose why Rutherford agents do not drive: read doctor per-agent states (not_installed, handshake_failed, no_answer, model_unavailable, error), run connect_only probes, fix PATH, auth, adapter, and Bedrock/Vertex model-id issues, and handle entitlement cases like Grok. Use when agents fail, doctor reports problems, or a panel returns failed voices.

Expand Rutherford's crew: discover installed ACP agents from the community registry, adopt a local Ollama or LM Studio model as a free voice, or define a new agent by hand in config. Use when the user wants more agents available to Rutherford or asks why an agent is missing.

Run a multi-round debate where several coding agents argue a question, see each other's positions, and revise across rounds on persistent sessions, with an optional judge and convergence tracking. Use when the user wants agents to argue it out, stress-test a decision, or reach consensus on a hard tradeoff.

Run long Rutherford work as a background job (mode=async) and manage it: list jobs, watch activity, poll status, fetch the result, cancel, continue a durable job, or persist a run to disk. Use when a delegation, consensus, or debate will take minutes, or the user wants to keep working while it runs.

Review a diff or a set of changed files across multiple coding agents at a senior or principal bar — a quick independent multi-model review, or a full review-then-debate-to-consensus flow. Use when the user wants code reviewed, a change checked for ship-readiness, or a multi-model review.

Configure Rutherford's config.toml defaults: safety posture, trusted workspaces, the enabled-agent allowlist, default effort and timeouts, persistence, and local-model auto-detection. Use when the user wants to change Rutherford's default behavior or write or edit its config file.

Create or edit saved Rutherford panels (reusable crews of agents) in panels.toon — seats, models, per-seat roles and stances, and the aggregation strategy — then hot-reload them with reload_panels. Use when the user wants to define, change, or list a named panel for consensus, debate, or review.

Set up Claude Code permissions so Rutherford's skills and tools stop prompting: allowlist the Rutherford MCP tools and read/write access to the .rutherford directories (project and home) in settings.json. Use during setup, or when the user is tired of approving every Rutherford config edit or tool call.

Author or override Rutherford role personas (reusable system prompts) under a role_dirs or .rutherford/roles directory, and apply the restart needed for roles to load. Use when the user wants a custom reviewer, architect, or other persona, or to add or edit a Rutherford role.

Hand one task to one coding agent through Rutherford and get a normalized result — read-only by default, with files in scope, a chosen model, an effort tier, a role, or a resumed session. Use when the user wants a single agent (Codex, Claude Code, Cursor, and so on) to read, explain, or analyze something.

Ask several coding agents the same question in parallel through Rutherford and compare their answers, or collapse them to one verdict with a strategy (unanimous, majority, plurality, weighted, or rank). Use when the user wants a second and third opinion, a cross-model answer, or a vote.

Let a coding agent actually edit code through Rutherford in write or yolo mode, behind the trusted-workspace gate and the git-worktree sandbox. Use when the user wants an agent to apply a fix, make changes, or implement something, not just read.