01 / Problem
Reviewers either skim a diff and miss the dangerous line,
or read every line and burn an hour on a trivial one.
Linters and CI don't reason about what a diff means — a hardcoded secret, a quietly removed public function, a spec requirement nobody implemented.
02 / Solution
Release Gate
Five isolated subagents inspect one concern each, run in parallel, and hand back a summary — not raw noise. An aggregator turns five findings into one Readiness Score and one verdict.
Zero dependencies. Built-in fetch and fs only.
03 / Agents
Five subagents, one aggregator
Security & Risk
Hardcoded secrets, unsafe calls, dependency bumps
Test Coverage
Changed source without matching test changes
Breaking Change Trace
Exported symbols removed with no replacement
Spec Conformance
PR checked against a supplied spec doc
Changelog Draft
Informational entry, never blocks
04 / Demo story
One PR, four planted issues, one verdict
- Adds a hardcoded Stripe secret
- Removes a legacy function the spec says must stay
- Bumps a dependency with no compatibility note
- Touches no test files
node src/index.js demo → Readiness Score: 12/100 — BLOCK, all four caught.
05 / Why it's different
Deterministic, not another LLM reviewer
Every subagent is pattern-based on purpose: no per-PR model cost, no latency, no nondeterminism — and it still catches the classes of issue reviewers most often miss. The same { status, score, findings } contract makes it easy to add an LLM-backed subagent later for subtler judgment calls.
06 / Impact
Rough ROI, mid-size team
10 min
saved per PR when an issue is caught pre-review instead of during it
~$60k/yr
directional estimate — see IMPACT-METRICS.md for the framework
07 / Try it
Zero setup, one command
node src/index.js demo — no token, no network, no npm install.
Real PR: node src/index.js gate <pr> <owner> <repo> --post
Release Gate · IBM Bob 2.0 Hackathon · MIT License