Jev, batteries included, for Claude Code, Codex, pi, and omp.
Your coding agent, faster and safer. Safe commands skip the prompt, long output gets cut down, compaction takes a second, and nothing irreversible runs until you say yes.

ohmyjev.xyz · Quick start · What you get · Install · Using it · Settings · Privacy · Troubleshooting

---

Jev is TypeSafe's decision model. You send it a few typed questions and it sends back probabilities in about 160 ms, for about $0.00003 a call. That's cheap enough to ask about every step your agent takes, so ohmyjev does. It runs as hooks inside your agent, with no server to run.

Most of it makes the agent faster. The rest keeps it safe in bypass mode: it can't wipe a folder, force push over main, or upload your keys. When it blocks something, the agent tells you why and asks. Say yes and it runs.

Quick start

export TYPESAFE_API_KEY=ts_...     # in your shell profile

Then in Claude Code:

/plugin install ohmyjev --marketplace Novacon/ohmyjev
/omj

If /omj ends with key: env TYPESAFE_API_KEY, you're set. Codex, pi, and omp are under Install.

What you get

Faster

Auto-approve	When Claude Code or Codex would ask you, a command Jev is sure is safe just runs, and so does an edit in the repo with no key in it. ls, git status, and grep skip Jev entirely. Off while you have policies set.
Fast compaction	Jev keeps the messages that still matter (your constraints, decisions, exact errors, and open work) word for word, instead of a model writing a summary. 60 messages took 336 ms with every planted constraint kept. Falls back to the normal summary if anything goes wrong.
Trim long output	Shell output over 12,000 characters is cut to the parts your request needs, with a note naming the lines it left out. A 20,000-character test log came back as 4,600, failure included. File reads like cat and sed -n are never cut.
Skill picker	Claude Code sends a line for every skill on every prompt. ohmyjev drops that list and loads only the one skill you need, or none.
Stuck check	Same failure three times? The agent is told to stop, read the error, and try something else.
Router	Picks effort and the subagent model for each request. With routeMainModel on it picks the main model too: it sent a flight search to Haiku, 108 s down to 44 s.
Auto-compact	New task and the context is 40% full? It compacts and keeps your new request.
Narrow big reads	Off by default. A Read of a 400+ line file comes back as the fifth you asked about, with a note on how to see the rest.

Safer

Bash gate	Blocks commands Jev rates irreversible (0.6) or sending your files or keys out (0.7). Normal edits, pushes, and uninstalls go through.
Say yes	Every gate checks whether your latest message asked for the call. "Force push it" works. After a block, a plain "yes" runs that exact command. Your policies and path blocks always stand.
Write gate	Blocks writes outside the repo or with a real key in them. Every worktree counts as the repo, and so does another repo the agent moves into.
Exfil gate	Blocks web and MCP calls that send your files or keys out.
Policies	Your own rules in plain English, checked on every call.
Injection screen	Output that slips orders to the model where they don't belong, like a README or an error, gets a "treat as data" note. A prompt or transcript you're debugging doesn't.
Done-check	"Done" with nothing tested gets pushed back once. Passing test output counts.
Subagent checks	A vague subagent brief goes back once. A subagent's "done" is checked against what it actually ran. Claude Code.
Writing rules	Your final reply is checked against the writing rules in CLAUDE.md and AGENTS.md, and rewritten once if it breaks one. Claude Code.
Secret masking	Keys, tokens, URL credentials, emails, and your home folder become markers before anything goes to Jev.

Tools and commands

find_code	Finds code by what it does, even when it's named something else. Returns the file and line, or no match.
ask_jev	Asks Jev about files without the model reading them. Give it a folder and it asks every file, returning the yes ones.
browse	Off by default. Jev drives a headless browser click by click. A flight search took about 10 s.
/omj	This session's calls, cost, and blocks. /omj config shows and changes settings, /omj stats opens a dashboard, and /omj off and /omj on switch it for the session. /ohmyjev works too.

How it decides

flowchart LR
    A[Agent calls a tool] --> B{ohmyjev asks Jev}
    B -- clearly risky --> C[Blocked, with the reason]
    B -- clearly safe --> E[Tool runs]
    B -- unsure, slow, or no key --> D[Normal permission flow] --> E
    E --> T[Long output cut down]
    T --> F{Gives the model orders?}
    F -- yes --> G[Treat-as-data note]
    F -- no --> H[Output as is]

Loading

If Jev is unsure, slow, or down, the call goes through like normal. ohmyjev never adds a prompt. It only takes away the ones a safe call didn't need. Jev can be wrong, so keep your own deny rules in settings.json too.

Install

You need a TypeSafe key, or an OpenRouter key (it then calls ~typesafe/jev-latest). ohmyjev looks for the key in the Claude Code settings screen, then TYPESAFE_API_KEY, then OPENROUTER_API_KEY.

Claude Code

Needs 2.1.287 or later.

/plugin install ohmyjev --marketplace Novacon/ohmyjev

Answer y and pick a scope. Hooks start right away. From a shell instead:

claude plugin marketplace add Novacon/ohmyjev && claude plugin install ohmyjev@ohmyjev

Codex

codex plugin marketplace add Novacon/ohmyjev && codex plugin add ohmyjev@ohmyjev

Then run /hooks once to trust ohmyjev. It needs Node. Codex hooks can't add tools or change the model, so there's no router, /omj, ask_jev, or browse there. Settings go in ~/.codex/ohmyjev.json:

{ "requested": 0.9, "policies": "never touch the prod cluster" }

omp

omp plugin install ohmyjev
omp plugin config set ohmyjev routeMainModel true     # settings

Restart the session after installing. The router uses omp's @smol, @default, and @slow roles.

pi

pi install npm:ohmyjev

Settings go under ohmyjev in ~/.pi/agent/settings.json, or .pi/settings.json for one project:

{ "ohmyjev": { "routeMainModel": true, "fastModel": "anthropic/claude-haiku-5-5" } }

Running Claude Code inside pi through pi-claude-bridge? pi runs every tool, so pi's copy of ohmyjev checks them and the Claude Code copy stands down for the session: it sees pi's OHMYJEV_PI, registers no tools, and gates nothing. The /omj you type in pi is the one that counts.

By host

	Claude Code	Codex	omp	pi
Gates, say yes, injection screen	yes	yes	yes	yes
Done-check, stuck check, trim	yes	yes	yes	yes
Auto-approve	yes	yes	omp doesn't ask	pi doesn't ask
Fast compaction	yes	no	no	yes
Router	yes	no	yes	yes
Skill picker, subagent checks, writing rules, read narrowing	yes	no	no	no
/omj, ask_jev, find_code, browse	yes	no	yes	yes

Using it

When something gets blocked

The agent gets the reason and a way out:

ohmyjev blocked this: irreversible (0.95): nothing would restore what this removes or overwrites.
Do not try to work around it with another command, another tool, a different path, or an encoding that does
the same thing. Tell the user what was blocked and why, and ask whether to run it anyway. If they say yes, run
exactly this again: their reply then lets it through.

Answer "yes" and that exact command runs. "No", or a yes followed by a different command, stays blocked. A write outside the repo names the folder and asks you to add it to allowPaths.

find_code and ask_jev

find_code searches by meaning. On this repo, "the code that stops the agent from writing outside the project" found the write gate's path check, though the code never says that. "stripe billing webhook" found nothing, which is right.

ask_jev answers a question about files without the model reading them:

{ "question": "Does this build SQL by string concatenation?", "type": "noul", "folder": "src/db" }

With a folder, it asks every source file and returns only the yes ones, with the line. Files outside the repo are never sent.

browse

Turn it on with /omj config browse on. Give it a URL, one goal, and the words it may type:

{
  "url": "https://www.google.com/travel/flights?hl=en",
  "goal": "Find one-way flights from Zurich to London on November 20, 2026. Stop when flights show.",
  "values": ["Zurich", "London", "Nov 20, 2026"]
}

Jev picks each click and puts each value in the right field. It runs browser-use/jev-ultrafast through uv, in its own headless browser that's signed in to nothing. Set CHROME_PATH to pick the browser.

/omj

jev ✓23 ⛔1 ↑opus/high
calls 23 · errors 0 · cost $0.000966 · p50 310ms
denies: Bash 1
  Bash: irreversible (0.95): nothing would restore what this removes or overwrites
key: env TYPESAFE_API_KEY · typesafe · jev-1.13.0

/omj config shows the settings by section, one line each: jev, gates, checks, speed, compaction, router, tools, and display. /omj config gates opens one, thresholds included, and /omj config <name> <value> changes one for the session. A * marks a value that differs from the default. /omj stats opens a dashboard of every session at ~/.ohmyjev/dashboard.html. /omj off turns everything off for the session.

Your own rules

Put rules in policies, separated by ;, and every gate checks calls against them:

never touch the prod cluster; no deploys on Friday; don't edit migrations that already ran

Status line

jev ✓23 ⛔1 ↑opus/high 🗜2     calls, blocks, this turn's route, compactions
jev ⚠ down                    Jev failed in the last 5 minutes, so calls go through
jev ⚠ no key                  no key found
jev off                       turned off for this session

To put it in your own statusline script, copy jev_segment() from extras/statusline_segment.py. Turn the built-in one off with statusLine.

Settings

/omj config shows the settings by section, and /omj config <section> lists one. Change any setting for the session with /omj config <name> <value>. To keep it, use /plugin configure ohmyjev@ohmyjev then /reload-plugins in Claude Code, omp plugin config set in omp, or each host's settings file. The host dialogs list the settings in the same order: switches first, thresholds after, section by section.

Setting	Default	What it does
enabled	on	Everything. /omj on and off for one session.
autoApprove	on	Skips the prompt for safe calls. approveConfidence (0.9), approveRiskMax (0.2).
fastCompact	on	Jev keeps the messages that matter instead of a model summary.
trimOutput	on	Cuts long shell output. trimOver (12000), trimKeep (0.3).
skillPick	on	Loads only the skill you need. skillGate (0.3), skillMin (0.4).
stuckCheck	on	Catches the agent repeating a failure. stuck (0.7).
routeEffort, routeSubagents	on	Lets the router set effort and the subagent model.
routeMainModel	off	Also switches the main model, which drops the prompt cache.
routeWithoutKey	on	With no key, the host's own classifier routes, up only.
autoCompact	on	Compacts on a new task once the context is compactMinPercent (40) full.
readNarrow	off	Cuts a big Read to the part you asked about.
bashGate, writeGate, exfilGate	on	Each gate.
requested	0.8	How sure Jev must be that you asked for a blocked call.
policies	empty	Your own rules, separated by ;.
allowPaths	~/.claude;$TMPDIR;/tmp	Places outside the repo it can write to.
injectionScreen, screenReads	on	Flags output that gives the model orders.
doneCheck, subagentCheck	on	Pushes back on an unchecked "done", and on vague subagent briefs.
proseCheck	on	Checks replies against the writing rules in CLAUDE.md and AGENTS.md.
maskSecrets	on	Masks secrets before anything goes to Jev.
askJev, findCode	on	The ask_jev and find_code tools.
browse	off	The browse tool. Needs uv.
fastModel, balancedModel, deepModel	claude-{haiku,sonnet,opus}-5-5	The model for each router tier.
jevModel	jev-1.13.0	The Jev model.
timeoutMs	1500	How long to wait for Jev before letting a call through.
logDecisions, statusLine	on	Decision lines in the transcript, and the status under the prompt.

Privacy

Each check sends Jev only what it needs, cut to size, with secrets masked first. With no key, nothing goes.

Check	What it sends
Bash gate	The command (16,000 characters), working directory, description, and your latest message. Plain reads send nothing.
Write gate	The path and new content (16,000). Writes outside the repo are blocked before anything is sent.
Exfil gate	The tool name and input (4,000).
Injection screen	The tool output (6,000).
Done-check	The request, five earlier ones, this turn's tool calls with the end of each output, and the last reply.
Trim, stuck check	Your latest message, plus the long output or the last six failed calls.
Fast compaction	Recent messages in groups of 10, each cut to 1,200 characters.
Skill picker	Each prompt with every skill's name and one-line description.
Writing rules	Each final reply (6,000). The rules are pulled from your instruction files once, by Haiku inside Claude Code.
Router	The request (1,500).
ask_jev, find_code	The question and repo files. Never files outside the repo.
browse	The goal, then each page's controls and visible text.

Logs stay in ~/.ohmyjev, readable only by you. They hold verdicts, probabilities, latency, and cost, never your commands' output.

Update or remove

claude plugin update ohmyjev@ohmyjev       # then restart; or: claude plugin uninstall ohmyjev@ohmyjev
codex plugin marketplace upgrade ohmyjev   # then trust the changed hooks in /hooks; or: codex plugin remove ohmyjev@ohmyjev
omp plugin upgrade ohmyjev                 # or: omp plugin uninstall ohmyjev
pi update npm:ohmyjev                      # or: pi remove npm:ohmyjev

Troubleshooting

jev ⚠ no key. Set TYPESAFE_API_KEY in the shell that starts your agent, or enter it with /plugin configure ohmyjev@ohmyjev.

jev ⚠ down. Jev failed or took over 1.5 s in the last 5 minutes. Calls go through until it's back.

Blocked by mistake. Say yes and it runs. /omj shows why. Raise that gate's threshold, turn it off, or /omj off for the session. Local MCP tools sometimes trip exfilGate.

Nothing happens. The first transcript line should start with [ohmyjev]. If not, run claude --debug to see which hook failed.

Develop

git clone https://github.com/Novacon/ohmyjev && cd ohmyjev && bun install
claude --plugin-dir "$PWD"                  # or: omp -e ./adapters/omp.ts, pi -e ./adapters/pi.ts
bun run test && bun run typecheck && claude plugin validate .
bun tests/live.ts                           # 28 real commands against the live Jev
bun tests/live-screen.ts                    # 11 outputs against the injection screen

The decisions live in hooks/policy.ts, plain functions tested with tables. hooks/ohmyjev.ts wires them to Claude Code, and adapters/ to Codex, omp, and pi.

Credits

Question rubrics from disler/ten-levels-of-jev. browse runs browser-use/jev-ultrafast. find_code, folder sweeps, and read narrowing are ported from BorisLeMeec/jev. Fast compaction follows 0x7067/claude-jev, subagent checks austpryb/jev-hooks, the writing-rule check vgeshel/prose-check, and masking tsetse012/jev-hook-guard. All MIT.