Claude Octopus

Claude Octopus is a multi-model orchestration plugin for local coding agents. It turns research, definition, implementation, review, and delivery into explicit workflows, can assign configured providers such as Codex, Copilot, Qwen, Ollama, Perplexity, and OpenRouter to distinct seats, and lets the host model synthesize the result. Personas, skills, and single-host methods still work without external providers, and plain prompts do not dispatch paid providers by default. Independent opinions and visible failure states can be valuable, but they also expand the data boundary, tool authority, latency, and cost. First adoption should therefore pin a release in a test repository, inspect hooks, and restrict the provider roster.

Choose the native Claude Code or Codex plugin path in a test repository, pin v11.9.4, and inspect the manifest and hooks. Run the read-only Doctor first, configure only one provider, then validate routing, outbound context, and cost prompts with low-risk research or read-only review. Keep automatic invoke routing disabled until policy is settled.

Help me install and minimally verify Claude Octopus in an isolated test repository. Project: https://github.com/nyldn/claude-octopus. First read README.md, docs/PLUGIN-COMPATIBILITY.md, docs/INSTALLATION-HEALTH.md, docs/PRIVACY.md, SECURITY.md, and the v11.9.4 release. Show the version, commit, manifest, hooks, and intended user/project writes before installation. Use the official Marketplace command for my current Claude Code or Codex host, restart it, and run the read-only Doctor first. Configure only the single provider I name and verify one read-only research or review task on public text. Do not enable OCTOPUS_AUTO_ROUTER_MODE=invoke, change production code, create a PR, widen credentials, or auto-repair. Report providers, outbound context, cost prompts, generated files, uninstall and retained-data locations, and unverified capabilities.

claude plugin marketplace add https://github.com/nyldn/plugins.git
claude plugin install octo@nyldn-plugins
# 重启 Claude Code 后运行：/octo:setup

1. 1
   Pin the release and inspect authority

   Confirm v11.9.4 at commit 9cc6bcc, then review the host manifest, hooks, privacy policy, security policy, and write locations.

2. 2
   Install into one test host

   Use the README Marketplace command for Claude Code or Codex, restart the host, and avoid configuring multiple paths at once.

3. 3
   Run read-only health checks

   Use Doctor, capabilities, and cache-check to distinguish discovery, installation, authentication, entitlement, and quota.

4. 4
   Verify one read-only workflow

   Configure one provider, run research or review on public text, and inspect activation, contributions, cost prompts, and local results.

5. 5
   Set the automation boundary

   Choose provider allowlists and cost ceilings from observed data and failure behavior; keep plain-prompt routing on suggest or off until then.

Claude Octopus puts model disagreement, engineering phases, provider state, and cost controls into explicit workflows instead of adding another chat entry point. That can expose blind spots in valuable research, architecture decisions, and independent review. The tradeoff is a broad shell-and-hook code surface, more recipients for code and prompts, and more cost and failure combinations. It fits teams willing to govern versions, authority, providers, and evidence; it should not be treated as an unsupervised quality guarantee.