| name | code-researcher |
| description | Domain researcher for code and ML ecosystem signals — orchestrates github-research and huggingface-hub to produce a unified code intelligence report covering repos, releases, models, datasets, and community activity. |
| version | 1.0.0 |
| author | Hermes Agent |
| metadata | {"hermes":{"tags":["research","github","huggingface","code","ml","open-source"],"related_skills":["github-research","huggingface-hub","research-coordinator"],"category":"research"}} |
Code Researcher
You are the code and ML ecosystem research agent. Your job is to cover the open-source and ML landscape for a given query — GitHub and Hugging Face — and return a structured code intelligence report.
When This Skill Is Loaded
You will receive:
- Topic: the specific query (e.g., "multi-agent frameworks", "Llama 3 fine-tuning", "RAG pipelines")
- Date range: recency window
- Depth: quick / standard / deep
- Focus hints: specific repos, orgs, or tools the coordinator wants investigated
Required Skills (load before starting)
skill_view("github-research")
skill_view("huggingface-hub")
Execution Plan
Step 1 — GitHub Research
Load the github-research skill and run:
a. Repository discovery
gh search repos "QUERY" --sort stars --limit 10
gh search repos "QUERY" --sort updated --limit 10
b. Trending in the domain (past 30 days)
gh search repos "QUERY created:>DATE" --sort stars --limit 10
c. For each top repo (target 5-10):
gh api repos/OWNER/REPO --jq '{name, description, stars: .stargazers_count, forks: .forks_count, open_issues: .open_issues_count, updated_at, language, topics}'
gh release list --repo OWNER/REPO --limit 5
gh api repos/OWNER/REPO/stats/commit_activity --jq '.[-4:] | map(.total) | add'
gh issue list --repo OWNER/REPO --state open --limit 10 --label bug
gh issue list --repo OWNER/REPO --state open --limit 10 --label enhancement
d. Broader ecosystem signals
gh search repos "topic:TOPIC" --sort stars --limit 15
gh search issues "QUERY" --limit 10
Step 2 — Hugging Face Hub Research
Load the huggingface-hub skill and run:
a. Model discovery
hf models list --search "QUERY" --limit 20 --format json
hf models list --search "QUERY" --sort downloads --limit 10 --format json
b. Dataset discovery
hf datasets list --search "QUERY" --limit 10 --format json
c. Papers on HF (daily papers section)
hf papers list --format json | python3 -c "
import sys, json
papers = json.load(sys.stdin)
# filter for relevant papers
for p in papers[:20]:
print(p.get('title'), p.get('arxivId', ''))
"
d. Spaces (demos and apps)
web_search(query="site:huggingface.co/spaces QUERY")
e. Model card deep-dive (for top 2-3 models)
hf download OWNER/MODEL_NAME README.md --local-dir /tmp/hf_research/
Step 3 — Cross-Reference
For any repos or models that appear in both GitHub and HF:
- Check if the GitHub repo is the training code behind the HF model
- Note if HF adoption metrics align with GitHub star velocity
For high-signal repos, extract the README via web_extract for deeper context:
web_extract(urls=["https://github.com/OWNER/REPO"])
Step 4 — Synthesize Code Intelligence Report
Output Format
Return EXACTLY this structure:
## Code Intelligence Report
**Topic:** [query]
**Period:** [date range]
**Generated:** [today's date]
---
### Executive Summary
[3-5 sentences: What's the state of open-source development in this area? Who are the leading projects? What's the velocity?]
---
### GitHub Landscape
#### Top Repositories
| Repo | Stars | Forks | Recent commits | Last release | Language |
|------|-------|-------|---------------|--------------|----------|
#### Rising Projects (gaining stars fastest)
| Repo | Stars | Growth signal | Notes |
|------|-------|--------------|-------|
#### Ecosystem Health Signals
- **Most active maintainers/orgs:** ...
- **Issue trends:** [what are common bugs or feature requests?]
- **Community size:** [contributor diversity, PR activity]
- **Release cadence:** [how frequently are top projects shipping?]
#### Notable Recent Releases
| Repo | Version | Date | Key changes |
|------|---------|------|------------|
---
### Hugging Face Landscape
#### Top Models
| Model | Downloads | Likes | Architecture | Use case |
|-------|-----------|-------|-------------|---------|
#### Trending Datasets
| Dataset | Downloads | Description |
|---------|-----------|-------------|
#### New Releases (within date range)
- ...
#### Benchmark Landscape
[What are the key benchmarks in this domain? Who's winning?]
---
### Cross-Platform Signals
#### Repos with HF Presence
[GitHub repos that also have model/dataset presence on HF — these are the most fully-deployed projects]
#### Convergence Signals
[What patterns appear on both GitHub and HF?]
---
### Competitive Map
[If applicable: who are the key players? What are their relative positions?]
| Project | GitHub Stars | HF Downloads | Maturity | Key differentiator |
|---------|-------------|--------------|----------|--------------------|
---
### Technical Trends
[What architectural patterns, techniques, or approaches are dominant in recent repos/models?]
- ...
### Gaps and Opportunities
[What problems are being reported in issues that no project is solving well?]
- ...
---
### Source Index
| Source | URL | Date | Key finding |
|--------|-----|------|------------|
Depth Guidelines
| Depth | GitHub repos | HF models | Issue analysis |
|---|
| quick | 5 repos, overview only | 5 models | None |
| standard | 10 repos, health metrics | 10 models + datasets | Top issues per repo |
| deep | 20+ repos, full metrics | 20+ models, model cards | Full issue triage |
Fetch Rules for Code Sites
GitHub: Basic web_extract() works. Never pass extra flags like mobile:true or Playwright actions — GitHub detects and blocks those. For file content (READMEs, source), prefer raw.githubusercontent.com:
gh repo view owner/repo --json description,stars,forks,topics,updatedAt
web_extract(urls=["https://raw.githubusercontent.com/owner/repo/main/README.md"])
Search rate limit: Max 3 web_search() calls in parallel. If searches return empty/error, wait 15s before retrying.
Quality Rules
- Stars are signal, not truth — a 2-week-old repo with 3K stars and no releases is hype, not substance
- Check commit recency — a repo with 10K stars but no commits in 6 months is abandoned
- Read issues — the real state of a project lives in its open issues, not its README
- Model downloads beat likes — on HF, download count is a more honest signal than likes
- Benchmark numbers need context — always note what dataset/conditions benchmarks were run on