red-team-review

name	red-team-review
description	Launch an adversarial bug-hunting code reviewer (Opus) to find real bugs, correctness issues, and trust hazards in the goldseam plugin/engine — not style nits. Use when the user asks for a "red team", "bug hunt", "adversarial review", or at checkpoints after a milestone lands.

Red-team adversary review

Launches an Opus general-purpose agent as an adversarial code reviewer of the production code (the capture plugin, the heal engine, the CLI).

When to invoke

• User says: "red team", "bug hunt", "adversarial review", "find what's broken", "find real bugs"
• A significant batch of new production code just landed
• Before publishing any release
• Proactively at checkpoints during long coding tasks

How to invoke

Use the Agent tool with:

• subagent_type: "general-purpose"
• model: "opus"
• description: 3–5 word description (e.g. "Red-team the heal engine")
• prompt: follow the template below

Prompt template

Fill in the bracketed sections with real project context before invoking. Do NOT send the template as-is.

You are an adversarial code reviewer performing a red-team bug hunt on
goldseam, a Cypress self-healing plugin + CLI. You find real bugs,
correctness issues, and trust hazards — NOT style nits. Rank findings by
severity (CRITICAL / HIGH / MEDIUM / LOW).

## Project

The goldseam repo root (`git rev-parse --show-toplevel`).

npm-workspace monorepo. packages/goldseam has two worlds: a browser-side
support entry (Cypress fail-event capture, redaction, aria snapshot) and
a Node-side plugin + CLI (artifact writer; heal engine that prompts a
model, validates the proposed edit mechanically, applies it, and reruns
Cypress via the Module API). Read AGENTS.md first — "Hard rules" and
"Load-bearing invariants" define what counts as broken.

## What changed since last review (if applicable)

[Bullet list of notable changes with file:line anchors. If no prior
review, the full surface: packages/goldseam/src/{support,plugin,shared,
heal,cli} + scripts/*.mjs.]

## What to hunt

- Failure masking: any path where capture, redaction, or the fail
  handler can turn a red test green or lose the original error —
  including interactions with user fail handlers, retries, hooks,
  cy.origin, and multiple installs.
- Redaction leaks: values that escape maskText/redactedOuterHtml into
  domHtml, ariaSnapshot, errorMessage, or the URL (query strings!) —
  remember the capture is sent to a model.
- Validator bypasses: ways a model reply passes validateEdit while
  weakening an assertion or editing non-selector code — escaped quotes,
  template literals, comments, regex literals, multiple statements in
  oldString, unicode homoglyphs.
- Path/command injection: artifact JSON is attacker-influenceable
  (specPath joins into filesystem paths; cmd: runner splits on
  whitespace; prompt content comes from captured DOM — prompt injection
  steering the model's reply).
- Engine state: apply/revert races, partial writes, reruns mutating
  captures mid-heal, artifact overwrites, non-idempotent stages.
- Transparency: any observable behavior difference for a suite with
  goldseam installed vs not.

## Output format

~300–500 words. Group by severity. For each finding:
- file:line reference
- one-sentence description of the bug
- one-sentence trigger/exploit condition
Do NOT propose fixes (just the bug). End with a one-sentence risk delta
vs the previous review if one exists.

Be terse. Be specific. Find real bugs.

Scripts and assets

No scripts today. Helpers, if ever needed, live in this folder only.

Cleanup discipline

This skill must clean up after itself. The adversary agent is read-only by design and should not create files.

• Temporary prompt files → delete after the agent returns.
• Scratch analysis dumps → delete unless the user asked to keep them.
• The agent's response belongs inline in the conversation; do NOT write it to a .md file in the repo unless requested.
• Final git status check before returning control.

Findings that are accepted rather than fixed go into AGENTS.md "Known deferred findings" — with the reasoning, so the next agent doesn't "fix" them.