Umby's Jev Stack
September 20, 2026 · View on GitHub
A skill tree of TypeSafe Jev agent skills over HTTP. Install the whole tree or grab one skill.
Repo: https://github.com/Umbylicus/umby-jev-stack · Tree index: TREE.md
Quick start (4 steps)
- Get a key — console.typesafe.ai → create a TypeSafe API key.
- Install — whole tree or one skill (see Install).
- Set
JEV_API_KEY— exact name, case-sensitive. - Prompt your agent — e.g. "Run umby-jev-review on this repo. Use frozen questions. Compile every finding."
Skill tree (11 skills)
| # | Skill | Use when |
|---|---|---|
| 1 | umby-jev-review | Full repo review — six independent Nouls, executable defects only (v2.0.2) |
| 2 | umby-jev-diff-gate | Same hunts on git diff / PR patch only |
| 3 | umby-jev-secrets | Stricter secret + PII sweep |
| 4 | umby-jev-breaking-change | API / schema / export breaks |
| 5 | umby-jev-test-gap | Behavior change without tests |
| 6 | umby-jev-agent-safety | Destructive command risk before run |
| 7 | umby-jev-prompt-screen | Jailbreak / injection in user text |
| 8 | umby-jev-issue-triage | Issue type + urgency + independent security Noul |
| 9 | umby-jev-docs-drift | Docs claims to verify later |
| 10 | umby-jev-browser | Browser Harness/Chrome — indexed operation/target heads + independent error Nouls |
| 11 | Speedy Jev | Incremental clutter/performance candidates; ≥ 0.70 list; look-only proposals |
See TREE.md for install details and how to add skill 12+.
Install
Whole tree
npx plugins add Umbylicus/umby-jev-stack --target cursor
Select umby-jev-stack. Works with --target claude-code or --target codex.
One skill
Same command — select the skill name when prompted (e.g. umby-jev-secrets).
Manual: cp -r skills/<skill-id>/ ~/.cursor/skills/<skill-id>/
API key (JEV_API_KEY)
| Where | How |
|---|---|
| Shell | export JEV_API_KEY="your-key" |
| Cursor desktop | Settings → MCP/plugin env, or launchctl setenv JEV_API_KEY "$JEV_API_KEY" + quit/reopen from Dock |
| Cursor cloud agents | Runtime Secret named exactly JEV_API_KEY |
Verify: test -n "$JEV_API_KEY" && echo ok — do not commit keys.
How every skill works
- HTTP-first —
POST https://api.typesafe.ai/v1/systemone, modeljev-latest - Independent Nouls (or separate Choice + Score where documented) — every positive flag gets a compile row
- Never a single highest-priority Choice that hides co-occurring hits
- Jev classifies only; confirm/compile agents are look-only and explain possible remedies without editing application code. A human reviews and decides.
jev-review MCP is optional — not required.
License
MIT — see LICENSE.
Tree 1.6.0
- Speedy Jev 1.0.0 adds a local candidate index (path/hash/kind/last-seen), incremental scoring, and four independent Nouls. Only Speedy lists scores ≥ 0.70.
- Browser 2.0.0 follows jev-ultrafast: operation plus compatible speculative target heads in one request; text helper only for TYPE_TEXT; every error Noul ≥ 0.5 and click/type path are compiled.
- Diff-gate 2.0.0 ports review v2 structured hunts and ignore lists, scoped strictly to diff/PR changes.
- Review 2.0.2 clarifies look-only confirmation and the oversized-function chunk cap; Fable's six questions remain byte-for-byte unchanged.
state.content≤ 10,000 characters, repo-sized keep-alive HTTPS pool (concurrency = snippets), retry/backoff only as needed, and compile ≥ 0.5 remain intact. - Siblings use structured instructions/criteria and explicit evidence boundaries. Issue-triage 2.0.0 corrects Score criteria to an ordered array and retains security independently of category.
All skills use POST https://api.typesafe.ai/v1/systemone, model jev-latest, and JEV_API_KEY (SDK examples may use TYPESAFE_API_KEY). No application fixes or automatic cleanup. MIT; owner Billy Lewis / Umbylicus.