Bun

Bun skill index — bunjs v0.2.2

You are holding the index, not the content. Its whole job is to tell you which one or two files to open, so you pay for the guidance a task needs and nothing else.

Read only what the task calls for. Opening all eight is ~4,000 lines and defeats the point of an index. Two is normal. Five means the task should be split.

The eight

Each path is relative to the directory holding this file — the other skills are its sibling directories, which is true both when this ships as a plugin (plugins/bunjs/skills/<name>/) and when it is installed as a project skill (.claude/skills/<name>/).

Relative paths in a SKILL.md are resolved against that file's own directory — MEASURED in benches/skill-router/ (RTR-1, --repeat 8, Sonnet 5) over 74 skill-file reads, with no counterexamples. There is no reader for whom "the plugin's own directory" is a location; the only thing an agent knows is where the file it just read lives. So ../<name>/ is not a preference here, it is the only spelling that lands. A path that misses sends the reader into a dead end, and in that run the ones who did not recover were every Recall failure.

Routing

Match on what the task will make you write, not on the words the user used. Someone who says "add login" is asking for security; someone who says "it's slow" is asking for performance even though neither said the skill's name.

Most real tasks need a short ordered chain. Build code first, then harden it:

Chains stop where the task stops. "Create a todo app" is project-setup → http-service → errors; add testing only once there is something to test, security only once it has users, production only when it is being shipped. Reading all six upfront buys nothing you can act on yet.

Two rules that come from the skills themselves

Copy the shipped assets/, never retype them. Six of the eight ship tested code — error hierarchy, security guards, logger, test harness, env parser, benchmark harness. Each skill has the exact cp line. Retyped versions drop the subtle parts: the enumeration-timing burn, the full-jitter backoff, the cycle-safe cause walk.

Every skill has an Acceptance section. It is the definition of done, and it is the part most likely to be skipped. bun test and tsc --noEmit are the floor in all eight — bun run strips types without checking them, so a type error never surfaces at runtime.

Scope

Only Bun is tested. Node and Deno at your own risk.

If the task is not Bun or TypeScript, say so and stop — none of these eight apply, and guessing from them produces confident advice about the wrong runtime.