A Claude Code plugin that does a senior-engineer-grade code review of a repo. Five phases, parallel sub-agents, adversarial validation. Writes findings.md.
Install
Inside Claude Code:
/plugin marketplace add Farfield-Dev/deep-review
/plugin install deep-review@deep-review
/reload-plugins
(/reload-plugins registers the freshly installed skill in the current session. Skip it and /deep-review:run errors with "Unknown command".)
Auto-updates from this repo's main.
Run
In the repo you want reviewed:
--fast skips the team-intent and product-scan phases (cheaper, faster, narrower).
Output lands in ./findings.md with intermediate artifacts under ./.deep-review/ (gitignore it).
What you get
Each finding has a severity, an impact category (REVENUE_LEAK, SUPPORT_BURDEN, DATA_LOSS, SECURITY_BREACH, COMPLIANCE, PROD_INCIDENT, BRAND_DAMAGE), a file + line, a one-paragraph root cause, a verification artifact, and a fix direction. Findings that don't map to one of the impact categories get dropped, and every finding passes a 100%-confidence adversarial check before it lands in the report. Expect 0–10 findings on a medium backend, not 50.
Cost
On your own Anthropic key, $5–$25 per cold review on a 10–50k LOC backend. Sonnet on the cheap phases (architecture map, team intent), Opus on the heavy ones (action traces, validation). --fast is about a third of that.
Limitations
- Cold-start. No memory between runs.
- Best on backends with real workflows. Static sites and pure UI repos don't surface much.
- Opinionated, not exhaustive. Prefers 3 ironclad findings to 30 maybe-bugs.
License
MIT. See LICENSE.
By Farfield.