Multi-model adversarial code review and spec-to-software pipeline with Claude Octopus

What they did

They installed Claude Octopus alongside Claude Code and kept the ordinary Claude-native path for routine work, escalating only when they wanted a second opinion. For a risky refactor they ran an adversarial review: the host Claude Code produced a plan and diff, then Octopus fanned the same artifact out to several external providers (Codex, Copilot, Qwen, Ollama, Perplexity) and collected each model's findings. A 75% consensus gate flagged where the models disagreed, so the human reviewer looked at contested hunks rather than every line. They also tried Dark Factory mode on a smaller feature: they wrote a spec, let the pipeline run research, define, develop, and deliver phases with quality gates between them, then reviewed the output instead of each step. Persistent memory via claude-mem kept prior decisions and research context alive across sessions, so later reviews referenced earlier rationale instead of re-litigating it.

How it went

The consensus gate surfaced two real disagreements — one provider flagged a race condition the host had missed, another disputed an error-handling choice — which the human resolved before merge. Dark Factory produced a working draft that needed moderate polish, mostly naming and test coverage. Setup was incremental: they started with zero external providers and added them one at a time as each was detected.

Worth knowing

Requires Claude Code v2.1.14+ and explicit /octo:* invocation; Octopus stays dormant otherwise. External providers need their own auth — Qwen's free OAuth tier ended, so it now needs API keys. Consensus gates flag disagreement but do not resolve it; a human still adjudicates. Dark Factory output is a draft, not shippable code, and multi-provider runs cost more time and tokens than a single-model review.

Wiring

Install the Octopus plugin into Claude Code (v2.1.14+). Keep Claude-native /init, /review, and /security-review as the default path; Octopus activates only on explicit /octo:* commands. Add providers one at a time — each is detected from existing local auth or subscription (Codex, Antigravity CLI, Copilot, Cursor CLI) or from a local Ollama instance; Qwen needs an API key or Coding-Plan auth. For memory, wire claude-mem (or agentmemory/deja-vu) so decisions persist across sessions. For reviews, pass the diff as the artifact so each run gets its own artifactId and proof packet; set OCTOPUS_PROOF_RUN_ID to a unique value per run if overriding defaults. Use /octo:embrace for the Discover → Define → Develop → Deliver pipeline with quality gates between phases.

Try it yourself

Set up Claude Octopus in this repo for an adversarial review workflow. First, detect which external providers are already authenticated (Codex, Copilot, Cursor CLI, Ollama) and report which are available without new credentials. Then take the current uncommitted diff, run it through an Octopus review across the available providers, and produce a table of findings grouped by file and line, marking each finding as unanimous, majority, or contested under the 75% consensus gate. Do not modify code yet — just report, and call out any contested finding that touches concurrency, error handling, or auth.