Adversarial

Two-stage adversarial code review extension for swamp. Fans out across N configurable LLM backends as independent reviewers, then synthesises their findings via a separate judge agent.

How it works

1. Reviewer (@hivemq/adversarial/reviewer) — runs a two-phase plan→review loop for each configured model in parallel and bundles the HTML outputs into a single reviewerBundle resource.
2. Judge (@hivemq/adversarial/judge) — consumes the bundle, runs a consolidation pass, and emits a final judgeOutput resource.

Prompts are fetched from a promptsRepo you control at a pinned git sha, so prompt evolution is explicit and reproducible.

Quick start

Add both models and wire them with a workflow step:

swamp model add my-reviewer @hivemq/adversarial/reviewer
swamp model add my-judge @hivemq/adversarial/judge

steps:
  - name: review
    task:
      type: model_method
      modelIdOrName: my-reviewer
      methodName: run
      inputs:
        promptsRepo: "owner/prompts-repo"
  - name: judge
    task:
      type: model_method
      modelIdOrName: my-judge
      methodName: run
      inputs:
        promptsRepo: "owner/prompts-repo"
        reviewers: ${{ data.latest("my-reviewer", "main").attributes.outputs }}
    dependsOn:
      - step: review
        condition:
          type: succeeded

Reviewer inputs

Security

All subprocess env vars are filtered through an allowlist (PATH, HOME, locale, and ANTHROPIC_*/CLAUDE_*/PI_* prefixes). Diff content is blockquote-prefixed before prompt injection to mitigate attacker-controlled diff escaping the block. Large diffs are passed via stdin. Agents are killed after a 5-minute timeout.