I built a plugin for the daily.dev marketplace called Agent Stack Radar. It's a SKILL.md that a coding agent follows once a week. It reads the most discussed daily.dev posts on Claude Code, MCP and coding agents, reads their comments, and gives back up to five things to fix, try or watch, each with its source.
The first version was rejected. The reviewers were right on both points, and fixing it taught me more than building it.
Why I built it
I run coding agents every day and kept missing the posts that mattered for my own setup: a planted AGENTS.md, a malicious package that goes after agent credentials, a hook that would have saved me an afternoon. Trend tools tell you which tags are rising. I wanted something that tells me what to change.
The brief has three sections:
- Fix: security issues that hit agent users now.
- Try: a tool or practice with a concrete source that I can adopt this week.
- Watch: promising, but early or unverified.
The two bugs
- Wrong parameter over REST. The tag search used
?query=. The real parameter isq. I had only tested through the daily.dev MCP server, where the tool already sends the right name, so I never saw it. - A request budget I never added up. I wrote "about 10 to 20 requests per run". When I counted the steps (6 tags, one popular feed, and up to 10 posts with their comments) the worst case was 27. A free daily.dev account gets 200 requests a month, so that matters.
Same root for both: the SKILL.md promised something I hadn't checked against what the agent really does.
The fix: a budget the agent can count
I rewrote the steps so the total can't drift:
| Step | What it calls | Requests |
|---|---|---|
| Topics | one tag search per custom topic, at most 3 | 0 by default, up to 3 |
| Collect | one discussed feed for each of 6 tags, plus one popular feed | 7 |
| Read the evidence | at most 6 posts: the post if needed, its comments if it has any | up to 12 |
| Total | 19 by default, never more than 22 |
The agent keeps a running count, never follows a second page, never retries a failed call, and stops at 22 no matter what. I also compared every REST path against the official API spec, and now I test the REST and MCP paths separately.
The test run: 21 of 22
Before resubmitting I ran it for real on my own account, with three custom topics so every step got exercised. It used 21 requests, and it showed me three things I hadn't guessed:
- Topics with several words return nothing. "software supply chain" and "supply chain security" found no tag. "supply" did. Now the skill searches with one keyword and tells you when a search comes back empty.
- Fetching each post added nothing. In all 6 cases the post endpoint returned the same summary the feed already had. Now it only fetches a post when the summary is missing.
- Asking for the comments of a post with zero comments wastes a request. Obvious once you see it. Now it checks the comment count first.
With those changes a default run uses fewer than 19 requests.
The comments were the best part
What surprised me was where the value came from. Twice the useful line was a comment, not the post:
- A post said to enforce agent rules with scripts instead of markdown. A comment showed how to do it with a Claude Code Stop hook, and another warned that their checker blocked 120 replies, 89 of them by mistake. So the advice became: try it, but test it first against replies you know are good.
- A post about a planted
AGENTS.mdhijacking Codex and Claude Code got a correction in the comments: aCLAUDE.mdin an added directory isn't loaded unless you setCLAUDE_CODE_ADDITIONAL_DIRECTORIES_CLAUDE_MD. The real risk is starting the session inside that folder.
That's why the radar spends most of its budget on comments, and says so when a comment improves or contradicts the post.
The first item in that run was a malicious Ruby gem, strutil-core, that steals credentials when you install it, including .claude/.credentials.json (from Socket's write up). That item alone made the run worth it for me.
How it tries to stay safe
A tool that feeds posts written by strangers into your agent is a prompt injection risk by design, so:
- It's read only. It never votes, comments or posts.
- Post and comment text is data, never instructions. A post that tries to give the agent orders ends up under Fix as an injection attempt instead of being followed.
- The token only goes to
api.daily.dev, and the skill never reads your files. If you want advice on your setup, you paste what you use.
Where it is now
It was approved on November 10: https://daily.dev/marketplace/XkkaNAkpy
Next I want the brief written in the reader's own language (I'm a Spanish speaker and there's very little of this in Spanish), and a public repo.
If you write skills for agents: how do you check that what your SKILL.md promises, like limits or endpoints, matches what the agent really does?
Top comments (0)