Sugendran Reviews

sugendran-reviews

Deep pull-request review via specialist subagents. A navigator reads the change first and writes a reading guide; six focused reviewers run in parallel with that guide as shared context; a synthesiser ranks every finding into one action plan.

Install

/plugin marketplace add sugendran/sugendran-codes
/plugin install sugendran-reviews@sugendran-codes

Use

/sugendran-reviews:review                 # everything applicable to the diff
/sugendran-reviews:review security tests  # only those aspects
/sugendran-reviews:review types           # only type/interface design

The command scopes the current change (working tree, staged, or an open PR via gh), runs the navigator, fans out the reviewers, and presents the synthesiser's ranked plan.

Pipeline

sr-change-navigator  →  reading guide (intent vs. churn, domain, read order, hotspots)
        │  (guide fed to every reviewer)
        ▼
┌──── parallel reviewers ────┐
│  sr-code-reviewer          │  quality, CLAUDE.md compliance, bugs, comment rot
│  sr-silent-failure-hunter  │  swallowed/masked errors, bad fallbacks
│  sr-test-analyzer          │  coverage gaps, tautologies, domain meaning, human-verify gap
│  sr-type-design-analyzer   │  types + interfaces; SOLID; Vogels' 6 API rules
│  sr-security-reviewer      │  injection, authz, secrets, validation, crypto
│  sr-production-safety      │  rolling deploy, migrations, idempotency, blast radius, observability
└─────────────────────────────┘
        │  (findings, one severity scale)
        ▼
sr-synthesis  →  deduped, scope-checked, ranked plan (Critical → Low) + merge verdict

Conventions

• Scope snapshots — the orchestrator diffs once into filtered views (source / tests / config); each reviewer sees only the slice it can act on. Docs and generated files are stripped to names. Small diffs are reviewed inline with no fan-out at all.
• Scope discipline — reviewers report only what this change causes or endangers; pre-existing issues nearby are out of scope. Synthesis drops the rest.
• Severity — every finding is Critical / High / Medium / Low + a confidence.
• Reports, never edits — this toolkit advises; it does not change code.
• Agents are prefixed sr- so they coexist with any other review plugins installed.

Licence

MIT