Operator	Example	Effect
`stars:>N`	`stars:>1000`	More than N stars
`stars:N..M`	`stars:100..500`	Between N and M
`language:X`	`language:python`	Primary language
`pushed:>DATE`	`pushed:>2025-06-01`	Commits after date — the key freshness signal
`created:>DATE`	`created:>2024-01-01`	Repo created after date
`topic:X`	`topic:cli`	Tagged with topic
`license:X`	`license:mit`	Specific license
`-X`	`-language:javascript`	Exclude (prefix any qualifier)
`archived:false`		Exclude archived repos
`is:public fork:false`		No forks
`in:name` / `in:readme`	`http in:name`	Restrict where term matches
`user:X` / `org:X`	`org:google`	Scope to owner
`"exact phrase"`	`"rate limiter"`	Phrase match
`NOT`	`redis NOT cache`	Exclude keyword (strings only)

High-Signal Query Templates


# Baseline: established + actively maintained
<problem> language:<lang> stars:>500 pushed:>2025-06-01 archived:false

# Find the dominant library (only a few results = clear winner)
<problem> language:python stars:>5000

# Hidden gems (newer, not yet famous, but active)
<problem> language:go stars:50..500 pushed:>2025-09-01 fork:false

# Curated lists — these exist for almost every topic
awesome <topic> in:name stars:>1000

# CLI tools
<task> topic:cli stars:>200 pushed:>2025-01-01

# Commercial-safe only
<problem> license:mit OR license:apache-2.0 stars:>500

# Boolean grouping
(language:rust OR language:go) <problem> stars:>1000

# Code search (different syntax — searches file contents)
path:**/*.py "from fastapi import" symbol:RateLimiter

Search Aggressively — webSearch Is Your Primary Tool

Use webSearch extensively. Do not rely on a single query or a single source. Every solution search should involve multiple rounds of web searching across different angles — GitHub, package registries, blog posts, Stack Overflow, and comparison articles. Cast a wide net before narrowing down.

GitHub searches

webSearch("site:github.com <problem> <language> stars")
webSearch("site:github.com awesome <topic>")
webSearch("site:github.com/issues <specific error message>")
webSearch("best <language> library for <problem> 2026")

Note: GitHub-specific qualifiers like language:, stars:>, and pushed:> only work on GitHub's own search engine. Through webSearch, use natural-language equivalents (e.g. "python" instead of language:python). For precise filtering, use gh search repos if the GitHub CLI is available (see below).

Package registry searches

webSearch("site:pypi.org <problem>")        # Python
webSearch("site:npmjs.com <problem>")        # Node
webSearch("site:crates.io <problem>")        # Rust
webSearch("site:pkg.go.dev <problem>")       # Go

Community and comparison searches

webSearch("<lib A> vs <lib B> <language>")
webSearch("<problem> <language> reddit")
webSearch("<problem> best library site:stackoverflow.com")
webSearch("<problem> comparison benchmark <language>")
webSearch("awesome <topic> list github")

Read what you find

webFetch every promising repo URL to read the README directly. Don't just rely on search result snippets — actually read the README, check the examples, and look at the API surface before recommending anything. For comparison posts and blog articles, webFetch the full content to extract specific benchmarks and tradeoffs.

GitHub CLI (if available)

gh search repos "rate limiter" --language=python --stars=">1000" \
  --sort=stars --limit=10 --json=name,stargazersCount,pushedAt,url,description

gh api repos/OWNER/REPO --jq '{stars:.stargazers_count, pushed:.pushed_at, issues:.open_issues_count, license:.license.spdx_id, archived:.archived}'

Health Evaluation — Check These Fast

Signal	Healthy	Walk away
Last commit	<3 months	>18 months
Stars	>1000 (lib), >100 (niche)	<20
Open/closed issue ratio	<0.3	>1.0 with no replies
Contributors	5+	1 (bus factor)
"Used by" (sidebar)	>1000	0
Releases	Tagged, semver, changelog	No tags
License	MIT, Apache-2.0, BSD	None, GPL/AGPL (if commercial)
CI badge	Green	Missing or red
`archived: true` banner	—	Instant no

Red flags in issues: Search the issue tracker for "memory leak", "abandoned", "unmaintained", "alternative". If maintainer hasn't replied to anything in 6 months, the project is effectively dead regardless of star count.

Download trend check:

Python: https://pypistats.org/packages/<name> — declining = dying
npm: https://npmtrends.com/<pkg1>-vs-<pkg2> — compare candidates head-to-head
Check bundle size: https://bundlephobia.com/package/<name> (frontend only)

License TL;DR

License	Commercial OK	Must open-source your code?
MIT, BSD, Apache-2.0, ISC	Yes	No
LGPL	Yes	Only if you modify the lib itself
GPL	Yes	Yes, if you distribute (viral)
AGPL	Yes	Yes, even for SaaS (network-viral)
No LICENSE file	No	default is all rights reserved

Awesome Lists (curated entry points)

sindresorhus/awesome — the root of all awesome lists. Then: awesome-python, awesome-go, awesome-rust, awesome-react, awesome-selfhosted, awesome-nodejs, free-for-dev, build-your-own-x (learn by reimplementing), public-apis.

Comparison Output Template


## pkg-name  [12.4k stars, pushed 2 weeks ago, MIT]
github.com/owner/pkg-name

**Does:** One-line pitch.
**Fit:** Why it matches this specific problem.
**Install:** `pip install pkg-name`

**Pro:** Active, typed, 89% test coverage.
**Con:** Pulls in 23 transitive deps; async-only API.

```python
from pkg import Thing
Thing().do(x)  # minimal working example
```

Decision Rules

Two libs within 2x stars of each other → pick the one pushed more recently
A lib with 50k stars but last commit 2023 → it's dead, find the fork (check "Forks" tab sorted by stars)
Lib does 10x more than needed → check if you can vendor the 200 lines you actually need (with attribution)
Can't find anything with >100 stars → problem may be too niche; search blog posts / Stack Overflow for how others solved it
Found 3+ viable options → npmtrends/pypistats comparison, then read the top 5 closed issues of each

Output

Always present key findings and recommendations as a plaintext summary in chat, even when also generating files. The user should be able to understand the results without opening any files.