Overwatch

AI-assisted incident triage for backend services.

Overwatch watches a local backend process, captures crash context, analyzes likely root cause, estimates blast radius, and generates a postmortem with verification and rollback guidance.

Demo video

Why This Exists

Backend incidents are rarely blocked by typing speed. They are blocked by context recovery:

• What failed first?
• Which stack frame matters?
• What files and configs are related?
• Is retry safe?
• What is the likely blast radius?
• What should be verified before a fix is trusted?

Overwatch automates the first-pass triage workflow while keeping code changes human-approved.

What It Does

• Runs a target backend process under a watchdog
• Captures stdout, stderr, exit codes, timestamps, and crash metadata
• Parses stack traces and classifies failure scenarios
• Scans nearby workspace files for relevant source context
• Scores severity, blast radius, retry safety, regression risk, and confidence
• Builds an AI-ready incident prompt for Cursor or another local agent command
• Falls back to deterministic offline analysis when no AI provider is configured
• Generates a durable POST_MORTEM.md with root cause, patch proposal, rollback plan, and verification steps

Architecture

flowchart TD
    A["Backend process"] --> B["Watchdog runner"]
    B --> C["Crash parser"]
    C --> D["Workspace scanner"]
    D --> E["Risk scoring"]
    E --> F["AI/offline analyzer"]
    F --> G["Postmortem + verification plan"]

Loading

Tech Stack

TypeScript · Node.js · Express · Docker · Cursor-compatible agent commands · Local deterministic fallback

Quickstart

git clone https://github.com/Build4mBottom/overwatch.git
cd overwatch
npm install
cp .env.example .env
npm run demo:offline

Run the live watchdog demo:

npm run start:watchdog

Try a specific failure scenario:

SCENARIO=api-contract npm run start:watchdog

Supported demo scenarios include malformed JSON, async promise failures, invalid environment configuration, API contract errors, dependency resolution, database connection failures, timeouts, worker crashes, memory pressure, event-loop starvation, and retry storms.

Safety Model

Overwatch is read-only by default.

• It observes and analyzes failures.
• It proposes remediation steps.
• It does not automatically modify production code.
• Human review remains the final approval point.

Repository Structure

src/
  runner.ts                 # process watchdog
  crash_parser.ts           # stack trace and failure parsing
  workspace_scanner.ts      # local context recovery
  severity_classifier.ts    # severity classification
  blast_radius.ts           # impact estimation
  retry_safety.ts           # retry/rollback guidance
  agent.ts                  # AI/offline analysis path
  postmortem_generator.ts   # incident report generation
examples/
docs/

What This Demonstrates

• Backend failure analysis
• Runtime-aware developer tooling
• AI-assisted operational workflows
• Safe automation boundaries
• TypeScript service design
• Incident documentation and verification discipline

Status

Prototype/demo project. Built to demonstrate incident-triage architecture and developer-tooling product thinking.

License

MIT