devlica

devlica

Clone a developer persona from GitHub activity into Cursor skills.

devlica crawls GitHub data (repos, commits, pull requests, reviews, issues, gists, stars, and more), analyzes patterns with an LLM, benchmarks persona quality against held-out review comments, and writes Cursor skill files to disk.

Current Status

• LLM providers: anthropic (default), openai, ollama
• Anthropic auth modes:
  • API key (ANTHROPIC_API_KEY)
  • Google Vertex AI (Claude Code style env vars + ADC)
• Multi-token GitHub crawling supported via GITHUB_TOKEN, GITHUB_TOKEN_1, GITHUB_TOKEN_2, ...
• Optional private token supported via GITHUB_PRIVATE_TOKEN
• Exhaustive crawling mode available via --exhaustive (subject to GitHub upstream limits)

Build

go build -o devlica .

Usage

./devlica [flags] <github-username>

Required Environment

GitHub tokens

At least one GitHub token is required:

export GITHUB_TOKEN=ghp_...

Optional extra tokens for pool rotation:

export GITHUB_TOKEN_1=ghp_...
export GITHUB_TOKEN_2=ghp_...

Optional private token:

export GITHUB_PRIVATE_TOKEN=ghp_...

Anthropic (default provider) with API key

export ANTHROPIC_API_KEY=sk-ant-...
./devlica drpaneas

Anthropic with Vertex AI (Claude Code style)

devlica supports the same Vertex AI style variables used by Claude Code.

# Enable ADC credentials first (outside devlica)
gcloud auth application-default login
gcloud auth application-default set-quota-project cloudability-it-gemini

# Claude Code style Vertex env
export CLAUDE_CODE_USE_VERTEX=1
export CLOUD_ML_REGION=us-east5
export ANTHROPIC_VERTEX_PROJECT_ID=<your-project-id>

# Run devlica (provider defaults to anthropic)
./devlica drpaneas

Notes:

• Vertex mode is enabled only when CLAUDE_CODE_USE_VERTEX=1.
• In Vertex mode, both CLOUD_ML_REGION and project ID are required.
• Project ID is read from ANTHROPIC_VERTEX_PROJECT_ID first, then falls back to GCLOUD_PROJECT or GOOGLE_CLOUD_PROJECT.
• If Vertex mode is not enabled, ANTHROPIC_API_KEY is required for Anthropic.

OpenAI

export OPENAI_API_KEY=sk-...
./devlica -provider openai drpaneas

Ollama

# Optional override, defaults to http://localhost:11434
export OLLAMA_HOST=http://localhost:11434
./devlica -provider ollama drpaneas

Flags

-provider string    LLM provider: openai, anthropic, ollama (default "anthropic")
-model string       LLM model (default: per-provider)
-output string      Output directory for generated skills (default "./output")
-max-repos int      Maximum repositories to deep-crawl (default 10)
-exhaustive         Crawl exhaustive public GitHub activity data (disables sampling caps)
-verbose            Enable verbose logging

Default Models

• anthropic: claude-opus-4-6
• openai: gpt-4o
• ollama: llama3

Use -model to override.

How It Works

1. Crawl GitHub activity and code/review context.
2. Analyze style and behavior in parallel LLM passes.
3. Benchmark persona quality against held-out review comments, and refine when needed.
4. Generate Cursor skill files in the output directory.

GitHub Upstream Limits

Even in --exhaustive mode, some data sources are capped by GitHub:

• Public events are limited to recent activity windows.
• Search API queries are capped per query.

devlica logs warnings when collected counts look truncated by those limits.

Output

Generated skills:

output/
  <username>-coding-style/SKILL.md
  <username>-code-reviewer/SKILL.md
  <username>-developer-profile/SKILL.md