IncidentPilot AI is an intelligent AWS incident troubleshooting agent powered by Strands Agents, Strands Decider 2B, and Amazon Bedrock.
概览
Read-only AWS incident investigation agent for the (category: Tool Selection, #strands-decider-2b #tool-selection). Incident → Strands Decider 2B picks the next diagnostic tool → read-only AWS diagnostic → Amazon Bedrock analysis → incident report → dashboard. Nothing in the UI is faked: unavailable components are shown as unavailable. Requirements: Python 3.11+ (tested 3.13), Node 20.9+ (tested 22). Or with Docker: docker compose up --build. - No credentials in the repo; use IAM roles / SSO. .env is git-ignored. - AWS access is read-only by design. Wildcard CORS is rejected. - Set INCIDENTPILOT_API_KEY to require X-API-Key on /api/v1/*.
README
以下为原文 README 节选，完整内容请查看原文。
IncidentPilot AI
From Cloud Chaos to Clarity.
Read-only AWS incident investigation agent for the AWS Strands Decider Agent
Hackathon (category: Tool Selection, #strands-decider-2b #tool-selection).
Pipeline: Incident → Strands Decider 2B picks the next diagnostic tool → read-only AWS diagnostic → Amazon Bedrock analysis → incident report → dashboard.
Status: Phase 4 (diagnostic tools, read-only AWS) complete, pending live Decider/Bedrock/AWS verification
| Area | State |
|---|---|
| FastAPI backend skeleton, config, schemas, API-key option, CORS | Done, tested |
| Next.js frontend: 5 routes, sidebar app shell, R3F 3D background, GSAP motion, mode badge, API client | Done, builds, screenshot-checked in headless Chromium |
| Simulation catalogue (5 scenarios, metadata only) | Done |
Decider 2B tool selection (backend/decision_patterns/tool_selection/) |
Implemented against the real strands-decider System One API; confidence-gated; reusable pattern with guide. Not yet run against the real model (see below) |
Bedrock / Strands agent (backend/app/agent/) |
Implemented: structured analysis, evidence-id grounding, report store, /api/v1/investigations. Tested with a Bedrock reply test double and the real offline path; no real Bedrock call has been made (see below) |
| Diagnostic tools (ALB, EC2, RDS, CloudWatch Logs) | Implemented: simulation backend (synthetic, labelled) and a live read-only Boto3 backend with an in-process allowlist; IAM templates in infra/. Tested with AWS reply test doubles; no real AWS account has been called |
| Reports (history, detail, JSON/Markdown export) | Done for simulation incidents |
Nothing in the UI is faked: unavailable components are shown as unavailable.
Workflow Diagram
Live mode (read-only AWS)
# .env
INCIDENTPILOT_MODE=live
INCIDENTPILOT_LIVE_TARGET_GROUP_ARN=arn:aws:elasticloadbalancing:...:targetgroup/...
INCIDENTPILOT_LIVE_DB_INSTANCE_ID=my-db # optional
INCIDENTPILOT_LIVE_LOG_GROUP=/app/web # optional
Live mode exposes one incident, live-current, built from CloudWatch alarms and headline
metrics for the configured targets. After Decider 2B picks a diagnostic with enough
confidence, that read-only tool runs and its findings join the evidence. Unconfigured
parts are reported “not measured”; with no target configured the API answers 503 and never
serves simulation data as live. Apply the least-privilege policy in infra/iam/; the app
also refuses any API call outside its allowlist. Not verified against a real AWS account.
Amazon Bedrock: setup and verification status
aws bedrock list-inference-profiles --region us-east-1 # pick one enabled in your account
# .env: INCIDENTPILOT_BEDROCK_MODEL_ID= AWS_REGION= (IAM role / SSO, no keys in files)推荐工具
换一个关键词，或者移除筛选条件。
安装
npx skillfish add kashifsayyad09/incident-pilot-agent