Why add a resolver instead of giving the agent skills files?

Skills leave the agent reading several files and carrying them through the task. JCR moves lookup into a resolver with beam search over names and descriptions, returning item paths and context text while trails and scores stay in diagnostics.

How does it handle a request needing Stripe plus Slack steps?

Jev classifies single versus compound requests, an OpenAI model splits compound ones preserving details, then each step is searched in parallel with up to three paths kept at 60% of the best score. Single-step requests skip decomposition.

What will it cost me to try the audit versus a real resolve?

The audit needs no keys and validates index and items structure, but resolving and running agents call model APIs and incur charges. You need Node 22 or newer, and JCR, Claude and Codex modes each need their own key combination.

Can I reuse the capability format without Jev or npm install?

The format is independent of Jev and the root package is private and not published as an npm library, so you run it from a checkout. The catalog tree uses index and items files with unlimited nesting and a 16-round budget per step.

How does the resolver search the capability tree?

At each node Jev ranks direct children using their names and descriptions, with a no-match option, and questions for open paths in a step's search round are batched. A beam search scores paths by the geometric mean of routing probabilities.

How many paths does the beam search keep?

By default up to three paths whose score is at least 60% of the best, so close alternatives can be explored further. The resolver reports an explicit no-match or depth-limit outcome when appropriate.

How big is the bundled capability catalog?

11 groups, 960 nodes, and 11,360 items, with a maximum node depth of 6 and a largest direct child set of 175. npm run capabilities:audit prints updated counts and needs no API keys.

CommunityJev Agents & MCP

Jcr

Capability resolver that uses Jev to search a nested command catalog and return selected instruction context to an agent through one MCP tool, keeping the search outside agent context.

See highlights

Open github.com

Stars: 19
Language: JavaScript
License: MIT
Published by: Community
Author: NiazMorshed2007
Last verified: Oct 2026

NiazMorshed2007/jcr

User repository on GitHub

View on GitHub

A Jev-powered resolver for agent harnesses to find deterministic commands and their context in a nested capability tree.

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

JCR gives an agent one tool to find the documented commands it needs for a task. It uses Jev to search a nested capability tree and return the context attached to selected operations. Skills made it easier to load instructions when they are needed. But an agent can still read several files to find a few commands, then carry those files through the rest of the task. JCR moves that lookup into a resolver. The main agent gets the selected instructions while the search stays outside its context.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

Jev ranks child nodes by name and description with explicit no-match and depth-limit results.
Beam search keeps close alternatives and returns several matches or an ambiguity response.
Bundled catalog tree with audit rules for ids, names, descriptions and context.
Claude and Codex harnesses compare skills mode against resolver mode with metrics.

Watch out

The resolver only returns documentation and never executes commands, and depth-limit, no-match and ambiguity outcomes require the caller to narrow the request. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-sift

MCP server that screens files, public webpages, or inline text against a query and returns compact relevance probabilities, so the main agent opens only promising items.