Over the past few weeks I built four small, focused developer tools —
all sharing one idea: as more of our code gets written and reviewed by
AI, we need ways to check that work that don't themselves require an
LLM call. No API cost, no black-box judgment, just static analysis and
graph algorithms doing exactly what they're told.
Here's what each one does.
skillcheck
An ESLint-style linter for SKILL.md files — the instruction files
that Claude Code, Codex, and Cursor agents read to learn a skill. It
checks frontmatter completeness, catches broken relative links, flags
descriptions that are too short or too long, and enforces a token
budget so a skill file doesn't blow out an agent's context window
before it even starts working. Ships as a CLI and a GitHub Action.
github.com/DIYA73/skillcheck
mcp-schema-watch
MCP servers can change their tool schemas without warning, and if
you're depending on one, that breaking change shows up as a silent
failure in production. This polls the MCP servers you configure,
diffs each tool's schema against the last known snapshot, and tells
you whether a change is breaking or just additive — with Postgres
history, BullMQ-scheduled polling, a REST API, and Slack alerts on
breaking changes only.
github.com/DIYA73/mcp-schema-watch
pr-blast-radius
The deterministic version of "does this PR touch files it shouldn't."
Parses the real AST of every file in the repo, builds an import graph,
and checks which changed files are actually connected to each other —
versus which ones look like unrelated scope creep. Runs as a GitHub
Action that comments on the PR (and updates that comment on every push
instead of spamming new ones).
github.com/DIYA73/PR-Blast-Radius-
swarm-trace-viewer
The newest one, and still just the foundation layer. When an
orchestrator fans out into hundreds of subagents, finding out where
a run actually went wrong is hard — was this failure the real cause,
or just a downstream cascade from something else? This builds the
agent tree from a flat event list, classifies every failure as a root
cause or a cascade, and flags statistical outliers among sibling
agents (with high-fan-in "hub" nodes excluded so one shared dependency
doesn't make everything look connected). Includes a fully deterministic
synthetic trace generator for testing and demos, since there's no
public 1,000-agent orchestrator to record real traces from yet. Live
streaming (WebSockets, Redis pub/sub, Postgres history) and the actual
tree/timeline UI are next.
github.com/DIYA73/swarm-trace-viewer
All four are MIT licensed and open for issues or contributions. If
you're building with agents and one of these solves a problem you
have, I'd love to hear about it😁.
Top comments (0)