Sentry Auto-Triage: Autonomous Construct Agent Workplace

An intelligent, fully autonomous business workflow agent built for the CONSTRUCT × Techfluence 2026 App Store Bounty Track.

This is not a simple notification wrapper or a chat bot. It is a true, persistent, autonomous edge application running on Cloudflare Workers that equips a Construct OS Agent to handle the entire lifecycle of an incident exactly like a Human DevOps/SRE Engineer would.

---

🏆 Bounty Submission Details

1. Business Impact: Real, End-to-End Automation

When a Sentry alert fires, resolving it requires context switching across 4-5 different web apps, taking 15-30 minutes of human time. This app completely eliminates that operational overhead:

1. Listens (Sentry): Receives incident webhooks natively.
2. Analyzes (Intelligent Agent): Parses the stack trace, identifies the precise culprit file and line, and calculates an algorithmic priority score (e.g. fatal + spike = P0, warning + old = P3).
3. Investigates (GitHub): Executes a GraphQL git-blame on the exact culprit line to find the offending commit message and author.
4. Documents (Linear): Formats a rich Markdown ticket containing the stack trace, regression evidence, and code blame, directly assigning priority natively.
5. Alerts (Slack): Posts a clean, actionable summary to the engineering team's #incidents channel with ticket links attached.

2. Workflow Novelty

Our application implements complex logic typically missing from simple AI agents:

• Stateful Deduplication: Ensures that an incident is only triaged once an hour to protect the team from alert fatigue and duplicate tickets.
• Micro-Analysis: Sentry provides long messy traces; our triage_issue tool filters out framework internals and focuses strictly on the precise application code that threw the error.
• Git Blame Accuracy: The agent executes a GraphQL call into GitHub to find the actual code owner of the exact line, not just whoever touched the file last.

3. Technical Quality

Built natively on Cloudflare Workers using @construct-computer/app-sdk.

• Completely Serverless & Edge Optimized.
• Bypasses slow agent prompting by performing data manipulation natively in TypeScript edge functions before returning pure, actionable data to the Construct LLM agent.

---

🌐 Live Deployment

The app is deployed LIVE on Cloudflare Workers:

https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev

Open this in your browser to see the App UI directly.

---

👨‍⚖️ FOR THE JUDGES: How to Test

There are two ways to test — with or without credentials.

---

✅ Option 1 — No Credentials Required (Instant curl tests)

All tools are exposed via the MCP JSON-RPC protocol at the /mcp endpoint.

Windows users: Replace ' with " and escape inner quotes accordingly, or use Git Bash / WSL.

---

🔵 Health Check

Verify the worker is live:

curl https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev/health

Expected: ok

---

🔵 List All 5 Tools

See all registered tools and their input schemas:

curl -X POST https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/list","id":1}'

Expected: JSON list of all 5 tools with descriptions.

---

🔴 Tool 1: triage_issue

Parses a Sentry issue, assigns P0-P3 priority, detects regression — all in one call:

curl -X POST https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"triage_issue","arguments":{"issue_json":"{\"level\":\"fatal\",\"count\":200,\"userCount\":80,\"title\":\"TypeError: Cannot read properties of null\",\"firstSeen\":\"2026-04-22T00:00:00Z\",\"lastSeen\":\"2026-04-22T01:00:00Z\"}","event_json":"{}"}},"id":1}'

Expected: priority: P0, severityScore: 210, action: "Page on-call immediately. All hands on deck.", isSpike: true

---

🟡 Tool 2: blame_code_line

Uses GitHub GraphQL to find the exact author of a culprit line. Replace YOUR_GHP_TOKEN with a real GitHub token:

curl -X POST https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"blame_code_line","arguments":{"repo":"Botcoder757/sentry-triage-final","file_path":"server.ts","line_number":10,"github_token":"YOUR_GHP_TOKEN","branch":"main"}},"id":2}'

Expected: author name, email, commit SHA, commit message for line 10 of server.ts.

---

🟢 Tool 3: create_ticket_and_notify

Generates a formatted Linear ticket markdown AND Slack message in one call — no credentials needed to generate the formatted output:

curl -X POST https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"create_ticket_and_notify","arguments":{"title":"TypeError: Cannot read properties of null","level":"fatal","sentry_url":"https://sentry.io/issues/123","priority":"P0","priority_num":1,"action":"Page on-call immediately.","event_count":200,"culprit_file":"src/api/auth.ts","culprit_line":42,"stack_summary":"src/api/auth.ts:42 in getUser","error_type":"TypeError","is_regression":false,"regression_evidence":"Fresh issue","author":"Prathamesh","author_email":"prathamesh@example.com","commit_sha":"abc1234","commit_message":"fix: handle null user","commit_url":"https://github.com/Botcoder757/sentry-triage-final/commit/abc1234","repo":"Botcoder757/sentry-triage-final"}},"id":3}'

Expected: linearMd (full markdown ticket ready for Linear) and slackText (formatted Slack message) in response.

---

🔵 Tool 4: deduplicate_check

Checks if an issue was already triaged in the last hour:

curl -X POST https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"deduplicate_check","arguments":{"issue_id":"abc123"}},"id":4}'

Expected: {"already_triaged": false} on first call.

---

🔵 Tool 5: record_triage

Records a completed triage to prevent duplicate tickets for 1 hour:

curl -X POST https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","method":"tools/call","params":{"name":"record_triage","arguments":{"issue_id":"abc123","ticket_id":"ENG-42","ticket_url":"https://linear.app/team/ENG-42"}},"id":5}'

Expected: {"recorded": true, "issue_id": "abc123"}

Now run deduplicate_check again with the same issue_id — it will return already_triaged: true ✅

---

🔵 Webhook Endpoint

Simulate a Sentry alert firing:

curl -X POST https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev/webhook \
  -H "Content-Type: application/json" \
  -d '{"data": {"issue": {"id": "abc123"}}}'

Expected: {"received": true, "issueId": "abc123"}

Check the pending queue (the Construct agent polls this every 3 seconds):

curl https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev/pending

Expected: {"pending": ["abc123"]}

---

🖥️ Option 2 — Full End-to-End via Construct Desktop

This is the intended production flow — the Construct Agent orchestrates all 5 tools automatically across Sentry, GitHub, Linear, and Slack.

Prerequisites: Connect these via Composio inside Construct Desktop:

1. Sentry (Project Issues enabled)
2. Linear (Ticket creation permissions)
3. Slack (Message post permissions)
4. GitHub Personal Access Token (ghp_...)

Step 1. Link to Construct Desktop

1. Open Construct Desktop App → Settings → Developer
2. Toggle Developer Mode ON
3. Under Connect Dev Server, paste:

   https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev
   

4. Click Connect — Construct loads the UI automatically

Step 2. Configure the App

1. Click ⚙️ Project Configuration in the UI
2. Fill in your Sentry Org, Project, GitHub Repo, Branch, Slack Channel, Linear Team ID, and GitHub Token
3. Click Save Configuration

Step 3. Trigger the Pipeline

• Click "Scan Sentry & Run Full Triage" — the agent fetches the last 3 unresolved Sentry issues and runs the full 6-step pipeline autonomously
• OR paste the webhook URL into a Sentry Alert Rule for fully automatic triggering:

  https://construct-app-sentry-auto-agent.naikprathamesh782.workers.dev/webhook
  

Watch the Construct Chat Panel — the agent declares each step as it fetches, parses, blames, files the Linear ticket, and posts the Slack alert.

---

🛠️ Architecture & Tools

Tool	Capability
triage_issue	Parses stack trace, assigns P0-P3 priority, detects regression in one call
blame_code_line	GitHub GraphQL git blame — finds exact author of the culprit line
create_ticket_and_notify	Generates Linear ticket markdown + Slack message in one call
deduplicate_check	Checks if issue was already triaged in last hour
record_triage	Records completed triage to prevent duplicate tickets for 1 hour

The "All-in-One" Edge Compute Design Pattern

A common mistake when building agents is providing hundreds of granular micro-tools (parse_JSON, regex_match, timestamp_convert) which drains tokens and forces models to hallucinate intermediate steps.

Our server.ts combines data manipulation, analytics, and cleansing into rigid edge-compute abstractions — leaving the LLM free to do what it does best: orchestrate business decisions based on parsed output.

---

💻 Local Development

# Clone
git clone https://github.com/Botcoder757/sentry-triage-final.git
cd sentry-triage-final

# Install dependencies
npm install

# Start dev server (runs on localhost:8787)
npm run dev

Connect your Construct Desktop to http://localhost:8787 in Developer Mode.