claude-fleets

Reusable adversarial fleet workflows for Claude Code — drop them into any project to get multi-agent review and audit with independent verification.

A "fleet" fans out one agent per dimension (correctness, security, tests, …), collects findings, then independently re-checks every finding with a skeptic that defaults to rejecting it. Only findings that survive verification come back. This is the pattern that kills plausible-but-wrong findings a single reviewer emits.

Workflows

Workflow	What it does
fleet-review	Adversarial multi-dimension code review: per-dimension find → independent verify → confirmed findings (sorted by severity).
plan-audit	Re-audit a plan / design doc against the real system: per-section audit → cross-check high-severity staleness claims.

Install

Per project (committed with the repo) or globally (~/.claude/workflows/, available in every project):

# one project
mkdir -p .claude/workflows && curl -fsSL \
  https://raw.githubusercontent.com/junaidtitan/claude-fleets/main/workflows/fleet-review.js \
  -o .claude/workflows/fleet-review.js

# all projects (global)
mkdir -p ~/.claude/workflows && cp workflows/*.js ~/.claude/workflows/

Or clone and run ./install.sh ~/.claude/workflows (global) or ./install.sh /path/to/project/.claude/workflows.

Use

Once a workflow file is in .claude/workflows/, invoke it by name (Claude Code's Workflow tool resolves it):

Workflow({ name: "fleet-review", args: {
  base: "/abs/path/to/repo",
  context: "Invariants the reviewers must respect; what's intentional vs a bug.",
  target: "run `git diff origin/main` to see the changes under review",
  dimensions: [                              // optional — sensible defaults if omitted
    { key: "correctness", prompt: "Real bugs only: logic, edge cases, None/empty, races." },
    { key: "security",    prompt: "Injection, path traversal, secret/PII leakage, escaping." },
    { key: "tests",       prompt: "Real coverage? Tautological/over-fitted? Masked bugs?" }
  ],
}})

Workflow({ name: "plan-audit", args: {
  base: "/abs/path/to/repo",
  context: "Plan at docs/plan.md. Real system under src/. Built since the plan: X, Y.",
  dimensions: [{ key: "schema", prompt: "Audit the schema section against real fixtures." }],
}})

You can also wire a slash command (.claude/commands/review.md) that calls the workflow, or just ask Claude Code to "run a fleet-review on this diff".

Args (both workflows)

• base — absolute repo path (default ".").
• context — ground truth the agents must respect (invariants, what's intentional). The most important arg — a good context paragraph is what makes findings real instead of noisy.
• target (fleet-review) — what to review (a diff command or file list).
• dimensions — [{ key, prompt }] review lenses. Scale to the task: a few for a quick check, many for an audit. Sensible defaults if omitted.
• model — optional per-agent model override (e.g. "sonnet" for cheap stages, "opus" for the hardest).

Notes

• Requires Claude Code with the Workflow tool (multi-agent orchestration). The fleet runs in the background; watch progress with /workflows.
• These are plain JS workflow scripts (not TypeScript) — see the comments in each file for the exact args contract.
• Tune dimensions and context per task; that's where the leverage is.