Question 1

Does this actually change which skill Claude Code picks?

Accepted Answer

Not reliably, by the author's account. Claude Code still shows the model every skill's description and the model still chooses, so the router is a second opinion from a weaker judge. In his week-long run, 28 of 539 suggestions were followed by a call of the suggested skill, and reading a random sample of unused suggestions found most were off target, several reacting to text an agent had written rather than to a typed prompt.

Question 2

How does this differ from TypeSafe's skill suggestion cookbook?

Accepted Answer

The cookbook's measurement used claude-haiku-4-5 reading a skill index cut to 60 characters per skill, so its second reader had less to add. This port ranks on the whole description, then sends the full SKILL.md body for the top three, so nothing is truncated. The question text and thresholds still come from the cookbook; what this repository adds is the plumbing and the way to measure it.

Question 3

Can I tell whether a threshold change made the router better or worse?

Accepted Answer

Only partly. Every row hashes the question wording and every threshold, so rows written after a change do not pool with older ones, and the prompt is stored as a hash and a length rather than as text. But the per-candidate fit question is not in the hash yet, and inside Claude Code the log cannot separate Jev's pick from how the agent chose its next step.

Question 4

What happens if the API is slow or unreachable?

Accepted Answer

The client never retries, because a hook in front of a prompt has a few seconds and a retried timeout spends them twice. In shadow mode the work goes to a detached child process and returns at once, so nothing waits on a decision nobody reads. Every failure exits 0, which keeps a broken call from blocking a prompt but also hides it from the user.

Question 5

How does it decide which skill to suggest?

Accepted Answer

Two requests. The wide one ranks the whole roster by name and full description with a choice question and asks three noul gate questions about the prompt; if the gate mean is under 0.30 it stops. The narrow one sends the top three with their descriptions and full SKILL.md bodies, picks one by choice, and asks one fit noul per candidate.

Question 6

How is the skill roster built?

Accepted Answer

From your user skills, the skills of enabled plugins, and project skills from the session's working directory up to the git root. Skills marked disable-model-invocation are left out. Nothing is truncated, because ranking on a prefix would rank on less than Claude already sees.

Question 7

What does shadow mode actually do?

Accepted Answer

It hands the judgment to a detached child process and returns at once, so nobody waits on a decision nobody reads, and every failure exits 0. The default mode only logs the answer; inject mode also gives Claude a one-line hint. A broken API call then shows up as a missing log line rather than an error.

Jev Skill Router

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community