Inspiration

It's 2 a.m. Production is down. You furiously fix the outage with panic commits named "fix", "asdf", and "please work".

Once the panic subsides, reality sets in: your team, tech lead, or client expects a calm, polite, professional pull request before the morning standup.

We built Commit Confessional to bridge the gap between late-night debugging reality and corporate engineering diplomacy. Instead of another generic AI code explainer, it serves two critical purposes:

An emotional mirror that confesses the blunt, unvarnished truth of what happened.
A pragmatic workflow utility that writes the clean GitHub PR description for you in seconds.

"Honest first. Polite second."

What it does

Commit Confessional takes messy git diffs or raw commit logs and produces two distinct documents side-by-side from the exact same input:

Honest Incident Report (Left Panel): A 5-line deadpan, monospace autopsy stating the cold facts: Severity, Event, Root Cause, Fix, and Lessons Learned.
Diplomatic Pull Request (Right Panel): A structured, neutral markdown description featuring a summary, itemized changes, and step-by-step verification instructions ready to copy into GitHub.
Panic Score Gauge (0–100): A diagnostic index measuring code anxiety based on 5 code signals (rushed commit messages, shouting/mashing, repeated file edits, debug log leftovers, and diff size) with an AI adjustment strictly capped at $\pm 15$ points.

How we built it

Frontend & Framework: Built on Next.js 15 (App Router) and TypeScript with Tailwind CSS.
AI Engine: Powered by Google Gemini 3.5 Flash using strict structured JSON schemas via Zod.
Deterministic Heuristic Engine: The Panic Score base calculation runs completely in deterministic TypeScript code: $$\text{Panic Score} = \text{clamp}{[0, 100]}\left( S{\text{base}} + \Delta_{\text{AI}} \right)$$ where $S_{\text{base}} \in [0, 85]$ and $\Delta_{\text{AI}} \in [-15, +15]$.
Integrity Guardrail: Every file path mentioned by Gemini in the generated output is checked against the actual input diff to eliminate hallucinated filenames.
Production Infrastructure: Self-hosted Linux environment with Nginx HTTPS reverse proxy, PM2 process supervisor, and Cloudflare Tunnel.

Challenges we ran into

Dual-Persona Tone Discipline: Ensuring Gemini delivered dry, deadpan humor on the left while staying strictly corporate, polite, and neutral on the right without tone bleed.
Hard Real-Time Latency Budgets: Enforcing a 28-second hard deadline with automatic exponential retries on HTTP 503/429 so the user interface never freezes or hangs.
Deterministic Transparency: Ensuring the AI never has the unmonitored final say over numerical scores. All heuristic weights and observed evidence are exposed via the "Why this score?" popover.

Accomplishments that we're proud of

Sub-10 Second User Loop: Users can load a preset, compute the panic score, and copy a production-ready PR description in under 10 seconds.
Zero Hallucinated Files: 100% verification rate preventing invented file paths.
100% Test Coverage on Core Logic: 59 unit tests passing across parsing, heuristic scoring, schema contracts, and API routes.
Real Production Deployment: Live and accessible at bwai.rofihosted.space without relying on mocked APIs or dummy responses.

What we learned

A great AI developer tool isn't just an unconstrained prompt in a text box. What makes generative AI truly reliable and trustworthy are the deterministic guardrails built around the model: schema validation, output verification, defensive timeout budgets, and algorithmic transparency.

What's next for Commit Confessional

CLI Extension (git confess): Generate PR descriptions and incident postmortems straight from the developer terminal before pushing.
GitHub Action Integration: Automatically comment panic metrics and summaries directly on incoming PRs.
Team Panic Analytics: Longitudinal dashboards tracking late-night hotfix trends across sprint cycles.