ham-zax’s JevShelf keeps a large library of agent instructions behind small, domain-specific router skills. Skillshare manages installation and distribution; JevShelf selects a specialist; the coding agent separately loads its instructions and decides how to apply them.

I inspected commit 8cdd151469ae643b1379d9a00a240471990882d3 and the pinned router dependency. Using the documented bootstrap on Python 3.13.5, I ran the offline suite: 68 tests passed. A local demo selected summarize-notes and successfully loaded a bounded page with its expected digest. I made no live Jev calls and did not test cross-host Skillshare synchronization; this is not a security audit.

Search is not a full instruction dump

  • Local mode uses exact-name lookup or BM25, a keyword-ranking method; it is explicitly labeled lexical, not Jev.
  • Hosted mode sends task/context, collection metadata and bounded skill excerpts to TypeSafe’s actual /v1/systemone API.
  • The code configures verification of up to six candidates per ranking batch, returning at most three selected skills; these are implementation limits, not measured optimal settings.
  • Applicability questions distinguish a useful method for the current step from a merely related topic.
  • Excerpt selection reserves space for capabilities and exclusions, but a bounded sample can still omit decisive instructions.

Load separately, check continuity

Search returns metadata and a SHA-256 digest, a content fingerprint. load can check that the selected file has not changed, then return it in bounded pages. The underlying reader rejects symlink paths, directory escapes and unsupported files.

The first load’s digest check is optional in code; continuation pages require one. A digest establishes content continuity—not that the skill is trustworthy or that Jev chose correctly. Typed output is not a correct decision.

Useful boundaries, unmeasured savings

  • Native discovery must expose the router skill, and the host model must choose to invoke it; this is not universal prompt interception.
  • Jev uncertainty and errors do not silently become successful lexical results.
  • Hosted ranking is not wholly local: task text and excerpts cross a service boundary.
  • Plain package installation failed here because the pinned router was unavailable in the package registry; the documented Git-based bootstrap resolved it.

The project offers an inspectable way to reduce unnecessary instruction loading. It does not establish a token-saving percentage, a routing-accuracy result or a latency advantage. Those need a representative library and real tasks, not an offline fixture suite.