Flakiness Knowledge Graph Mcp

📊 flakiness-knowledge-graph-mcp

[npm version] [npm downloads] [CI] [License: MIT]

A Playwright custom reporter + MCP server that builds a local flakiness knowledge graph from your test run history. Ask your AI agent which tests are unreliable, on which browser, and whether they're getting worse.

🤔 The Problem

A single Playwright trace tells you what failed right now. It doesn't tell you whether this test has been silently flaking for two weeks, or only fails on Firefox in CI, or is getting slower with every release.

This tool fixes that by accumulating run history into a SQLite database and exposing it to AI agents via MCP.

🛠️ Tools

🚀 Setup

1. Install

npm install -g flakiness-knowledge-graph-mcp

Or build from source:

git clone https://github.com/vola-trebla/flakiness-knowledge-graph-mcp.gitcd flakiness-knowledge-graph-mcpnpm install && npm run build

2. Add the reporter to your Playwright project

// playwright.config.tsexport default defineConfig({  reporter: [["html"], ["flakiness-knowledge-graph-mcp/reporter", { dbPath: "./flakiness.db" }]],});

Run your tests normally — the reporter writes every result to flakiness.db automatically.

3. Add the MCP server to your editor

Cursor / VS Code (.cursor/mcp.json or .vscode/mcp.json)

{  "mcpServers": {    "flakiness-knowledge-graph": {      "command": "flakiness-knowledge-graph-mcp"    }  }}

Claude Code

claude mcp add flakiness-knowledge-graph flakiness-knowledge-graph-mcp

4. Try it with demo data

No Playwright project yet? Generate 30 days of realistic sample data:

npx flakiness-graph-seed ./demo.db

Then point your AI agent at ./demo.db to explore all 8 tools.

💬 Example usage

I've been running my Playwright suite for two weeks. The DB is at /my-project/flakiness.db.
1. get_flaky_tests — which tests are most unreliable? Show last 7 days only.2. get_test_history for the top flaky test — is it getting worse?3. get_flakiness_trend for the same test over 14 days — plot the daily rate.4. get_failure_patterns — does it only fail on a specific browser or OS?5. cluster_semantic_error_trees — are multiple tests failing with semantically identical errors?6. correlate_git_commit_flakiness — which commit introduced the flakiness?7. get_slow_tests — which tests should I optimize for CI speed?

Grouping errors that look different but aren't

get_error_groups clusters by raw string prefix — if the error contains a UUID or element ID it creates separate groups for what is really one root cause. cluster_semantic_error_trees strips dynamic values first:

{  "total_clusters": 2,  "clusters": [    {      "cluster_id": "cluster-1",      "canonical_message": "TimeoutError: locator.click: Timeout 30000ms exceeded\n  waiting for locator('#submit-btn')",      "normalized_message": "TimeoutError: locator.click: Timeout <num>ms exceeded waiting for locator",      "error_taxonomy": "TimeoutError",      "instance_count": 14,      "affected_tests": 3,      "sample_test_ids": ["checkout > submit order", "cart > add item", "checkout > apply coupon"]    },    {      "cluster_id": "cluster-2",      "canonical_message": "Error: 2 requests to https://api.example.com/orders/8f3a1c were made. Expected 1",      "normalized_message": "Error: <num> requests to <url> were made. Expected <num>",      "error_taxonomy": "AssertionError",      "instance_count": 6,      "affected_tests": 1,      "sample_test_ids": ["api-mock > intercept order"]    }  ]}

Finding the commit that broke a test

correlate_git_commit_flakiness uses a state machine — it looks for runs where a test was stable for ≥3 consecutive passes, then failed. The transition record includes the SHA from the CI environment:

{  "total_transitions": 1,  "transitions": [    {      "test_id": "auth > login > should redirect after login",      "title": "should redirect after login",      "transition_type": "stable_to_flaky",      "git_commit_sha": "a3f8c1d9e2b54f6a",      "git_branch": "main",      "git_author": "dev-handle",      "transition_date": "2025-04-14"    }  ]}

The reporter reads GITHUB_SHA / CI_COMMIT_SHA / CIRCLE_SHA1 / GIT_COMMIT automatically — no reporter config changes needed beyond upgrading to v0.2.0.

🔗 Works great with playwright-trace-decoder-mcp

These two MCP servers are designed to complement each other:

• flakiness-knowledge-graph-mcp answers "is this test flaky historically, and which commit caused it?"
• playwright-trace-decoder-mcp answers "what exactly failed in this specific run?"

Combined, an AI agent can diagnose whether a CI failure is a known flaky test or a new regression — without you opening a single file.

⚖️ Parallel Execution & CI Sharding

flakiness-knowledge-graph-mcp uses an in-process write queue to ensure that parallel Playwright workers within a single Node process do not corrupt the database.

However, if you run tests across multiple independent processes (e.g., parallel CI shards or separate machine runners) writing to the same shared network file:

1. Race Conditions: Standard file systems do not guarantee atomic writes for SQLite files across processes without OS-level locking.
2. Recommended Approach: Each CI shard should write to its own database file (e.g., flakiness-shard-1.db, flakiness-shard-2.db).
3. Merging: At the end of the CI pipeline, you can merge these files into a single master database using standard SQLite tools or by running a script that reads from one and inserts into the other.

For local development or single-machine CI runs, the default configuration is safe.

🏗️ Architecture

playwright.config.ts  └── FlakinessReporter → flakiness.db (SQLite via sql.js)
flakiness.db  └── test_runs table        id, test_id, title, suite, file,        status, duration_ms, browser, os,        timestamp, error, retry,        git_commit_sha, git_branch, git_author   ← added in v0.2.0
MCP server  └── reads flakiness.db on demand (in-process handle reuse)

sql.js is used instead of better-sqlite3 — pure JavaScript SQLite compiled to WebAssembly, no native compilation needed. The git columns are added via ALTER TABLE migration on first use — existing databases upgrade automatically.

📋 Scripts

npm run build        # compile TypeScript → dist/npm run lint         # ESLintnpm run format       # Prettier --writenpm run format:check # Prettier check (used in CI)npm run seed         # populate flakiness.db with 30 days of demo data

📄 License

MIT