Ten levels of Jev
Ten levels from one Noul smart-if to a pi agent that writes its own Jev questions, with a live Vue lab and a YouTube walkthrough.
Overview
Ten levels of Jev (github.com/disler/ten-levels-of-jev, MIT) is IndyDevDan's curriculum repo: a live Vue lab, a terminal runner, real pi agent sessions, and a hyper-jev skill you can hand an agent. The README frames Jev as JSON in, decisions out. You send a state (string, object, or array) and a map of questions. About 300 ms later you get typed answers with probabilities. The repo arranges 30 concrete uses into 10 levels, 3 options each. Levels 1 to 5 call Jev from plain code (just jev1 through just jev5). Levels 6 to 10 put Jev inside the pi coding agent (pi 0.85.1+): hooks the agent never sees, then tools the agent chooses, until level 10's ask_jev tool lets the agent write the questions. Prereqs in the README: Node 24, bun, just, and an OpenRouter key for live runs. just test runs 201 offline tests on a deterministic mock. just web builds the lab and opens http://127.0.0.1:4399. The walkthrough video is https://youtu.be/_U-O5lYhJ7Q. Public counts were 147 stars and 45 forks when this listing was drafted.
Problem: Agent code is full of small judgments, and teams either keep patching a regex or spend a chat model call to get prose they still have to parse.
Built for: Agentic engineers who want a ladder of 30 runnable Jev examples, from one Noul in plain TypeScript up through live pi sessions where the agent writes the questions.
First indexed on Jev Directory: 2026-10-02
Creator and team
- Name
- IndyDevDan
- Organization
- IndyDevDan
- Handle
- @disler
“Ten levels of Jev, from one smart if statement to a coding agent that reaches for Jev on its own.”
How Jev is used
- Role in the product flow
- Pedagogical ladder: code calls Jev in levels 1 to 5, then pi hooks and tools call it in levels 6 to 10, ending with the agent authoring the questions
- Primitives
- ChoiceScoreNoul
- State in
- A state string, object, or array, plus questions keyed by ids you choose. Noul is yes or no, with optional true and false criteria. Choice is one of up to 255 declared options. Score is a position on 2 to 10 levels you describe. From level 8 on, code can attach file content, a glob, or a command the extension runs before the call.
- Decision out
- Typed answers with probabilities the option file thresholds in code. A Choice answer is one of your options. Level 4 uses confidence as a second gate: below 0.5 a person decides, above 0.9 a destructive action can skip confirmation.
- just install, or the manual path: bun install in apps/ten-levels/web, copy .env.sample, set OPENROUTER_API_KEY, install pi for levels 6 to 10
- just test for 201 offline tests on the mock. Levels 1 to 5 still run with no key. Levels 6 to 10 need pi and the key
- just web opens the Vue lab on http://127.0.0.1:4399. just jev1 through just jev10 run one level in the terminal. just jev4 b runs one option
- Levels 6 and 7 hook tool_call, tool_result, and turn end. The agent does not ask. Levels 8 and 9 expose file tools. Level 10 exposes one ask_jev tool
- Each option is one TypeScript file that holds its questions and thresholds together, which the README says is the part a human reviews
This repo is a ladder you can run, plus a skill that copies the pattern. dbreunig-building-with-jev-skill is Drew Breunig's SKILL.md for question design, state shape, and thresholds on jev-1.13. It does not ship 30 option files or a pi session runner. y0usaf-pi-jev is a Pi extension that batches gate nouls on bash, write, and edit. Level 6 here is a teaching hook inside this lab's sandbox, with its own jev-guard.ts. tianyucodings-jevharness lets a model write a harness and then calls Jev for fast task decisions. jkudish-jev-mcp exposes 10 typed MCP tools. Ten levels keeps the client in apps/ten-levels/src/core/client.ts, which the folder map describes as mock, OpenRouter, and TypeSafe. .claude/skills/hyper-jev/ packages that client, the 30 examples, and a cookbook. The README's own rule set: reach for Jev when you can describe the state and the outcomes, keep numbers and counting in code, and treat a security gate as one signal. Cite github.com/disler/ten-levels-of-jev README and the YouTube walkthrough at https://youtu.be/_U-O5lYhJ7Q.
Sourced performance claims
- Ladder: 10 levels, 3 options each, 30 concrete uses. Level 1 is one Noul (injection gate, urgency gate, ticket classifier). Level 10 is one ask_jev tool over state, paths, and a command (triage a failure, judge a diff, classify a request).Source: github.com/disler/ten-levels-of-jev README The ladder
- README price contrast: 1 million yes or no calls cost $3,500 on an expensive SOTA model and $16.80 on Jev. The same section says a call returns in about 300 ms.Source: github.com/disler/ten-levels-of-jev README Why this exists
- Level 1 injection gate, 5 input sets in the README: 0.99, 0.83, 0.61, 0.27, and 0.01 on the same question. Level 4 bash table: git push --force origin main is irreversible at 0.99 and blocked. ls -la src is read only at 1.00 and runs. rm -rf node_modules && npm install is reversible at 0.35 and asks a human. Thresholds live in level04/confidence.ts: below 0.5 a person decides, above 0.9 a destructive action skips confirmation.Source: github.com/disler/ten-levels-of-jev README Levels 1 and 4
- Level 8 live lab quote: ask_jev_file_bool on session.ts and seed.ts returned yes 0.98 and no 0.12 while agent context sat at 2k tokens. Jev judged 9,089 tokens for $0.00049. README prices one read of the same files at $0.091 on an expensive SOTA input price, 187x more. Level 9: 17 files, 2 yes answers at 0.93 and 0.96, pick_first_file chose src/domain/billing.ts at 0.93. The file cap is 255.Source: github.com/disler/ten-levels-of-jev README Levels 8 and 9
- Level 10 live run in the README: told the tests are red, the agent ran npm test through ask_jev, got bug_in_code at 0.99, fixed rounding, scored its own diff at risk 0.53, and confirmed tests pass at 0.99. Ledger: 3 Jev calls, 4 questions, $0.000084. Level 6 block example: rm -rf node_modules .sessions was irreversible 0.86 and destructive intent 0.99.Source: github.com/disler/ten-levels-of-jev README Levels 6 and 10
- Test commands: just test is 201 offline tests on the deterministic mock. just test-live is 10 live tests, 1 per level. README failure notes: the mock decides by word overlap. One live level 6 run routed around a write gate with a bash heredoc. Level 10 gates ask_jev's command, and the agent's own bash tool on that level is not gated.Source: github.com/disler/ten-levels-of-jev README Commands and Where it can still fail
- Public GitHub repo disler/ten-levels-of-jev had 147 stars and 45 forks on 2026-10-02. License MIT. Created 2026-09-27. Last push 2026-09-27. Language TypeScript. Homepage field is the YouTube walkthrough.Source: GitHub API 2026-10-02
Features and stack
Features
- 10 levels, 3 options each, questions and thresholds in one TypeScript file per option
- Vue lab on port 4399 with request bodies, probability bars, and a cost table
- pi extensions for guard hooks, compaction, cheap file reads, fan-out, and ask_jev
- 201 offline tests on the mock and 10 live tests, 1 per level
- hyper-jev skill with SKILL.md, cookbook, and a standalone starter
- README failure notes: mock word overlap, a write-gate bypass, and an ungated bash tool on level 10
Stack
- TypeScript
- Vue lab
- bun, just, and Node 24
- pi coding agent 0.85.1+
- OpenRouter or TypeSafe for live runs
- MIT
Pricing: MIT code. Live levels spend your OpenRouter or TypeSafe key. README contrast: 1 million yes or no calls at $16.80 on Jev versus $3,500 on an expensive SOTA model. Offline levels 1 to 5 run on the mock with no key.
Links
FAQ
- Is this the same as building-with-jev-skill?
- dbreunig/building-with-jev-skill is a SKILL.md that teaches Choice, Score, and Noul design for jev-1.13. ten-levels-of-jev is a 10-level lab: 30 TypeScript options, a Vue UI on port 4399, pi sessions for levels 6 to 10, and a separate hyper-jev skill that copies the client and the examples.
- Which levels run without an API key?
- README Install: with no key, levels 1 to 5 still run on the offline mock. Levels 6 to 10 need both pi 0.85.1+ and the key. just test runs 201 offline tests on that mock. just test-live runs 10 live tests, 1 per level.
- What changes at level 10?
- Levels 6 and 7 are hooks. The agent never asks, and Jev runs on the tool call, the tool result, and the turn end. Levels 8 and 9 are file tools the agent chooses. Level 10 is one ask_jev tool: the agent passes state, paths, a command, and its own question block. The README's example ledger is 3 Jev calls, 4 questions, $0.000084.
- What do the README dollar figures measure?
- They are the author's live-lab quotes, not a third-party invoice. The opening contrast is $3,500 versus $16.80 for 1 million yes or no calls. Level 8 prices 9,089 judged tokens at $0.00049 against a $0.091 file read. Level 10's ledger line is $0.000084. Your bill follows whichever key the lab is pointed at.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- Jev & TypeSafe
What Jev is, how System One fits your stack, and where this directory ends and TypeSafe docs begin.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.