AI Coding Task Dataset Builder

Turn GitHub issues into structured, agent-ready coding tasks at scale.

The Actor scans repositories or entire GitHub organizations, removes pull requests and low-value issue noise, detects actionable development work, enriches each accepted issue with repository context, and produces a clean dataset for coding agents, evaluations, benchmarks, internal automation, and task-routing pipelines.

What you get

Each accepted issue becomes one dataset item with:

• repository and issue metadata
• task type and 0–100 task score
• estimated complexity
• primary programming language
• detected frameworks and package managers
• detected ORM/database technologies
• likely affected files with ranking reasons
• build, test, and lint commands when detectable
• agent/contributor instruction file paths
• compact issue + repository context
• a ready-to-use prompt for Codex, Cursor, Claude Code, and similar coding agents
• the repository commit SHA used for the scan

The Actor does not execute repository code.

Quick start

Try it without credentials

Run the Actor with its default input.

The default mode uses a built-in demo repository and four demo issues. It performs no GitHub requests and outputs three realistic coding tasks while filtering one support/question issue.

Scan repositories

Set Run mode to Live GitHub scan and add repositories such as:

vercel/next.js
https://github.com/expressjs/express
owner/private-repository

For private repositories, add a GitHub token with read access.

Scan an organization

Add an organization such as:

microsoft
vercel
sveltejs

The Actor expands the organization into repositories, filters forks and archived repositories by default, and scans issues in each selected repository.

You can combine explicit repositories, organizations, and an Apify source dataset in the same run.

Best use cases

Coding-agent task queues

Create a continuously refreshed pool of real GitHub work for autonomous or human-assisted coding agents.

Evaluation and benchmark datasets

Use open or closed issues to create reproducible task datasets with stack context, likely files, and build/test commands.

Internal engineering automation

Convert issue backlogs into normalized task records for routing, prioritization, triage, or downstream agents.

Repository intelligence at scale

Scan many repositories with one consistent output format instead of maintaining repository-specific extraction logic.

Task filtering

The Actor uses an explainable heuristic score from 0 to 100 to distinguish coding tasks from questions, duplicate reports, support requests, discussions, and other low-actionability issues.

Signals include:

• bug, enhancement, feature, test, security, performance, dependency, CI, and similar labels
• implementation/fix/refactor language
• acceptance criteria and reproduction steps
• technical/code context
• issue detail quality
• negative signals such as question, duplicate, invalid, support, or discussion labels

The default minimum score is 55. Increase it for a smaller, cleaner dataset; lower it for higher recall.

Every result includes taskScore and taskScoreReasons so the decision is inspectable.

Task types

Issues are categorized as one of:

• bugfix
• feature
• refactor
• tests
• performance
• security
• documentation
• dependencies
• devops
• maintenance
• other

Relevant-file detection

filesLikelyAffected is generated from the repository tree using signals such as:

• paths explicitly mentioned in the issue
• filename/path keyword matches
• task-type-specific repository paths
• implementation, test, dependency, CI, documentation, and security conventions

relevantFiles also contains the score and reason for each suggestion.

These paths are guidance for an agent, not a guaranteed ground-truth patch set. The generated prompt explicitly tells agents to inspect the repository before making changes.

Repository context

For each repository, the Actor reads a small set of high-value files instead of cloning or executing the project. These can include:

• package.json
• lockfiles and workspace files
• pyproject.toml and requirements files
• go.mod, Cargo.toml, pom.xml, Gradle files
• framework configuration
• Prisma/Drizzle configuration
• Docker/deployment configuration
• CI workflows
• AGENTS.md
• CLAUDE.md
• Cursor/Copilot-style instruction files

This context is used to detect frameworks, package managers, data layers, and common build/test commands.

Security and secrets

The Actor is designed for repository analysis without executing untrusted code.

• repository code is never executed
• .env, credential files, private keys, and similar secret-like files are not fetched
• .env.example and other safe templates can be read for structural context
• the GitHub token is configured as a secret Actor input
• generated prompts warn coding agents to treat issue text and linked external content as untrusted task data

For private repositories, output naturally contains information derived from repositories the supplied token can access. Protect the resulting dataset and key-value store according to your own access requirements.

GitHub token

A token is optional for very small public scans but strongly recommended for production workloads because unauthenticated GitHub API limits are low.

Use the least-privileged token that can read the repositories you want to analyze. The Actor only performs read requests.

Inputs

Important inputs include:

Input	Purpose	Default
mode	Safe demo or live GitHub scan	demo
repositories	Explicit GitHub repositories	empty
organizations	GitHub organizations to expand	empty
sourceDatasetId	Optional Apify dataset containing repositories	empty
githubToken	GitHub API token	empty
issueState	Open, closed, or all issues	open
updatedSinceDays	Only issues updated within this window; 0 disables	365
minTaskScore	Minimum coding-task confidence	55
maxRepositories	Maximum repositories	100
maxIssuesPerRepository	Maximum issues scanned per repository	500
maxTasks	Maximum accepted output tasks	5000
relevantFilesPerTask	Likely files retained per task	8
generateAgentPrompt	Generate a coding-agent prompt	true

The Apify input UI exposes additional filters for labels, assignment state, forks, archived repositories, concurrency, and context depth.

Output

One accepted task equals one default-dataset item.

Example:

{
  "type": "coding_task",
  "taskId": "example/acme-dashboard#142",
  "repository": "example/acme-dashboard",
  "issueNumber": 142,
  "issueTitle": "Fix user table crash when email is missing",
  "taskType": "bugfix",
  "taskScore": 79,
  "complexity": "medium",
  "language": "TypeScript",
  "frameworks": ["Next.js", "React"],
  "filesLikelyAffected": [
    "src/components/UserTable.tsx",
    "src/components/UserTable.test.tsx"
  ],
  "testCommand": "pnpm test",
  "buildCommand": "pnpm build",
  "agentPrompt": "You are working in the GitHub repository ..."
}

The run key-value store also contains SUMMARY, including repository counts, scanned issues, filtered issues, accepted tasks, API request count, and repository/input errors.

If Save repository contexts is enabled, compact REPO_CONTEXT_*.json files are stored there as well.

Large runs

For large organization scans:

• provide a GitHub token
• keep repository concurrency conservative
• narrow Updated within days when you only need recent work
• use label filters when your repositories have consistent issue taxonomy
• lower Repository context files if API budget matters more than enrichment depth
• set maxTasks to control output size

The Actor paginates GitHub issues and organization repositories automatically and retries temporary API/rate-limit responses.

Notes for benchmark datasets

When scanning closed issues, the Actor intentionally does not fetch issue comments or merged pull-request patches. This reduces accidental solution leakage into the generated task prompt.

commitSha represents the repository snapshot analyzed at scan time. It is not claimed to be the historical commit that existed when an older issue was opened.

Output consistency

Task rows are structured for downstream automation and can be consumed directly from the default dataset as JSON, CSV, or through the Apify API.