• ## Inspiration

with the reliability of deterministic evidence. The goal: no made‑up claims, only proof. ## What it does

Code Roast scans a repo, extracts evidence‑bound issues (long functions, duplication, cycles, test presence), and then uses Gemini 3 to narrate the findings without inventing facts. It can also generate Fix‑It previews that are guarded by evidence line ranges, plus a proof‑locked apply flow that only succeeds if tests pass on a new git branch. It outputs a judge‑ready summary (action items + architecture map + impact preview).

## How we built it

We designed a strict, multi‑agent pipeline with clear responsibility boundaries:

  • Repo Scanner → Code Analysis → Insight Aggregator → Evidence Guard → Fix‑It → Fix‑Apply → Roast Narrator → Output Formatter
  • Deterministic agents do the facts; Gemini 3 only narrates what it’s allowed to say
  • Fix‑It patches are validated and verified before being shown or applied
  • A demo script (demo:judge) produces a consistent, presentation‑ready report

We also added an architecture map and an impact summary. Internally, we think of the “proof wall” as:

[ \text{Allowed Output} = \text{Narration} \cap \text{Evidence} ]

If the evidence isn’t there, we output “not enough data.”

## Challenges we ran into

  • Gemini sometimes returns invalid diffs, which required strict templates, retries, and debug output.
  • Rate limits and quotas during demos required a “Fix‑It limit” and graceful fallback.
  • Balancing readable output with enough evidence detail without overwhelming users.

## Accomplishments that we’re proud of

  • Hallucination‑resistant narration with evidence‑bound constraints.
  • Proof‑locked apply that creates a clean branch and runs tests before declaring success.
  • A judge‑ready demo mode that outputs architecture maps and actionable summaries.

## What we learned

  • Reliability matters more than cleverness in developer tools.
  • Clear evidence constraints make LLMs safer and more useful.
  • A good demo is as important as code quality for hackathons.

## What’s next for Code Roast

  • Richer signals (god files, layering violations, dependency direction).
  • A “before/after” scorecard across commits.
  • CI integration so Code Roast runs automatically on PRs.

Built With

  • 3)
  • api
  • apis:
  • cli
  • databases/cloud:
  • eslint
  • gemini
  • git/github
  • google
  • javascript-runtime:-node.js-frameworks/libraries:-@google/genai
  • languages:-typescript
  • local
  • none
  • platforms:
  • typescript-compiler
Share this project:

Updates