The review shelf
Every repo gets put
to the test.
We run the workflow the README promises, then a detour the README did not plan for. Good tools get credit. Rough edges get daylight.
mattpocock/skills/grill-me is a skill that makes a coding agent interview you in numbered rounds of questions, each with a suggested answer, before it writes code. A prompt asking for follow-ups gets the same result; /grill-me saves you writing it.
Verdict: Try grill-me if you want a structured interview before your coding agent builds: numbered rounds of questions, each with a suggested answer you can accept with "yes". Users credit it with surfacing decisions they missed and complain of long sessions. Read each suggestion before you say yes. For a codebase, Matt Pocock recommends grill-with-docs.
tester-army/e2etester-army/e2e is a browser testing framework where each step is a plain-language goal. Recordings replay without a model; the checks at the end of each test are still yours to write.
Verdict: Try tester-army/e2e on a suite that finds controls by visible name and sees frequent renames. A trial costs cents in model calls. e2e trades selector edits for a re-record run after each rename or menu change, its replays took two to three times as long as Playwright on our ten tests, existing Playwright tests must be rewritten, and in default mode a goal can pass through a different button when the intended one is broken. If your suite uses test ids, renames are rare, or you are content to let an agent patch selectors, the trade is not worth it.
morluto/reaOct 9, 2026Verdict: adopt with care
rea gives a coding agent local tools to take apart a compiled program or a packaged app without its source. On small programs an agent with ordinary shell tools already explained the code almost perfectly, so rea added cost and no accuracy.
Verdict: Adopt rea with care: its decompiled code was readable and matched the source where we compared it, but on small programs the agent already explained the code almost perfectly with ordinary tools, so rea added cost and no accuracy.
nicobailon/visual-explainerOct 6, 2026Verdict: adopt with care
A Claude Code skill for browser pages with code diagrams. Full mode took about eight times as long as a plain-text answer in our comparison.
Verdict: Adopt visual-explainer with care for custom code diagrams you will inspect; use --quick for routine pages and stay with a plain-text answer when you only need the explanation.
coldteadotai/pr-lensOct 2, 2026Verdict: adopt with care
pr-lens is a pull request visualization tool for developers reviewing agent-written code. The diagrams offer a starting point; checking the changed files and relationships remains your job.
Verdict: Adopt pr-lens with care for supervised pull request orientation. Its references pointed to real files, but larger changes left most changed files undrawn and repeated command-line tool runs often selected different files. Compare the diagram with the full changed-file list, inspect relationships in the source, check the terminal for truncation, and choose the delivery destination explicitly.
dzhng/jevgrepSep 28, 2026Verdict: adopt with care
jevgrep is a semantic code search tool for coding agents: ask a plain-English question, get ranked files and excerpts. With its paid Jev model counted, making the agent search with jevgrep first raised our total cost by 19%.
Verdict: Adopt jevgrep with care for finding code in repositories your coding agent does not know. On six bug fixes in Claude Code, the ranked lists from its jg command always included the file each bug's official fix changed. Requiring jg first cost 19% more than grep once Jev is counted, and the advertised 30% saving leaves Jev out.
tamaratran/fast-jev-compactionSep 20, 2026Verdict: wait
fast-jev-compaction is a Claude Code plugin that hands compaction to Jev, TypeSafe's decision model, so raw tool output can survive a long session. The mechanism compacts; whether the model's selection is useful is the open question.
Verdict: Wait on fast-jev-compaction: the idea is right and the mechanism compacts, but the tested revision's selection signal did not survive independent measurement, the shipped cutoff keeps nothing, and there is no threshold we can support.
mostlygeek/llama-swapSep 14, 2026Verdict: adopt with care
llama-swap is a local LLM (large language model) router for developers who point one OpenAI-compatible client at several local models. Its profiles switch models with one API call; llama.cpp's own server already handles basic switching.
Verdict: Adopt llama-swap with care if you switch one client between local models often. llama.cpp's own router already covers occasional switching, and a request that hits a crashed llama-swap model server gets an empty 502, so turn on your client's retry.
teamchong/pxpipeSep 14, 2026Verdict: adopt with care
pxpipe is a local proxy that converts Claude Code's repeated context into images to reduce input tokens. Cached text is already cheap, and reading it from images can lose detail.
Verdict: Use pxpipe with care when you can check its answers and measure whole-session usage. Our recommendation is for Opus 5, enabled explicitly. Skip it for long Fable sessions or exact values you cannot verify. Use pxpipe export if you only need files rendered into images.
ast-grep/ast-grepSep 4, 2026Verdict: adopt with care
ast-grep is a syntax-aware code search and rewrite tool for maintainers and coding agents. It can target exact syntax before a broad refactor; writing the rule is still work.
Verdict: Adopt ast-grep with care as a precision layer for structural audits and reviewed repository-wide refactors; keep built-in search or ripgrep for ordinary navigation because rule discovery adds substantial time and tool traffic.
kucherenko/jscpdjscpd is a local copy/paste detector that points maintainers to repeated source-code blocks and can enforce a CI threshold. The findings are useful; a zero exit code still needs supervision.
Verdict: A useful local duplicate-code shortlist with verified locations and broad report coverage. CI users should also verify the artifact and analyzed-source count instead of trusting exit code zero alone.
ccusage/ccusageccusage is a local Claude Code usage tracker and multi-agent reporting CLI for tokens, estimated API costs, and terminal status. Use Anthropic's own tools for plan limits and billing.
Verdict: A useful local Claude Code usage tracker and multi-agent reporting CLI whose controlled totals matched and whose real Claude Code statusline worked; use provider tools for authoritative billing and limits.
jackwener/OpenCLIAug 20, 2026Verdict: adopt with care
OpenCLI is an open-source browser CLI that lets agents query and operate websites through a logged-in Chrome session. The reach is broad; the permission boundary is yours to enforce.
Verdict: A capable browser CLI for supervised local research and custom read-only adapters, provided the operator owns the permission and maintenance boundaries.
unclecode/crawl4aiAug 16, 2026Verdict: adopt with care
crawl4ai is an open-source Python scraper that renders a page in a real browser and returns it as Markdown for RAG pipelines and agents. The rendering is solid; deciding what to keep is a setting you have to find.
Verdict: A capable way to get web pages into text a language model can read, but turn on a content filter and check the output rather than the status code, because the defaults return the whole page and the server reports success for settings it ignored.
Panniantong/Agent-ReachAug 10, 2026Verdict: adopt with care
Agent-Reach routes coding agents' web and social research across 15 platforms, choosing the specialized CLI, API, or browser session each job needs.
Verdict: A useful maintained routing layer for technically capable local-agent users, but not a uniform API or a maintenance-free service.
zohaibbashir/Google-Maps-ScrapperAug 7, 2026Verdict: adopt with care
This open-source Google Maps scraper drives a browser and exports local listings to CSV. The code is compact; keeping the dataset trustworthy is your job.
Verdict: Use this revision for supervised, overwrite-only small batches that you inspect immediately; do not trust append, the default one-result path, or a success log as proof of a valid dataset.
every-app/open-seoopen-seo is a self-hosted, open-source SEO suite for keyword research, audits, rank tracking, and agents. The app is free; live DataForSEO data is not.
Verdict: Try it for intermittent research if you accept DataForSEO pricing and want a self-hosted human and agent workspace; adopt recurring or critical operations with care.
crawlseo/crawlseoAug 7, 2026Verdict: adopt with care
crawlseo puts Search Console data and a site crawl in one self-hosted dashboard. The convenience is real; so are the operator chores.
Verdict: A credible small-site SEO workbench for technical operators, but own the database and OAuth setup, set crawl limits outside the visible control when needed, and inspect critical findings before acting on them.