Question 1

How does jev-opus change effort without breaking the prompt cache?

Accepted Answer

It only changes effort through Opus 5.5's per-message effort statement, inserted at the turn it governs via a local gateway or an Agent SDK hook. Every earlier insertion is replayed byte-identically on later requests, so the cached prefix and preserved-thinking blocks stay valid. The README records cache reads growing across one prompt while effort moved medium to low to high to low.

Question 2

Does jev-opus work without a TypeSafe API key?

Accepted Answer

Yes. Without a Jev key the CLI still runs using local heuristics. To have Jev pick the effort, provide either a TypeSafe key or an OpenRouter key, which reaches the same Jev model through TypeSafe. The doctor command checks Claude Code, the credential, Jev and a real Opus 5.5 call.

Question 3

How do I install it and what does it require?

Accepted Answer

npm install -g jev-opus, then jev-opus init and jev-opus claude, or npx jev-opus for anything without installing. You need Node 22.18 or newer and Claude Code 2.1.280 or newer already logged in with claude auth login. jev-opus init asks for the Jev key and writes ~/.config/jev-opus/.env.

Question 4

Which effort levels can it reach?

Accepted Answer

Only the ones you allow. The default range is low to high, so xhigh and max are never used unless you raise the ceiling, and the levels weakest to strongest are low, medium, high, xhigh and max. Set JEV_OPUS_MIN_EFFORT and JEV_OPUS_MAX_EFFORT in the env file, or use --min, --max, /bounds, /pin and /auto in driver mode. Running /effort yourself overrides Jev until your next prompt.

Question 5

Can I see when and why the effort changed?

Accepted Answer

The status line shows the current prompt's whole path as it happens and resets at the next prompt, and badges appear above Claude's text on the first response of each prompt and wherever the level changed, with the reason. JEV_OPUS_DISPLAY switches between changes, every-response and off, and JEV_OPUS_NO_INLINE_EFFORT or JEV_OPUS_NO_STATUSLINE turn pieces off. The badges show the effort selected, not how much the model reasoned.

Question 6

Which Claude Code surfaces does it cover?

Accepted Answer

The terminal claude is the tested path and works with a claude.ai subscription; the VS Code extension uses the same environment-variable mechanism but is not yet tested. jev-opus gateway serves http://127.0.0.1:47821 and prints ANTHROPIC_BASE_URL plus ANTHROPIC_CUSTOM_HEADERS, a random token to treat as a temporary secret. claude.ai web, mobile and cloud sessions cannot be routed at all.

Question 7

What does jev-opus leave unrouted on purpose?

Accepted Answer

Side requests never reach Jev: token counting, short requests without tools such as titles and summaries, and Claude Code's next-prompt suggestion. They still carry their own branch's inserted statements so they share the cache, but create no decision, no journal entry and no status change. Subagents are routed as their own threads, and if journal recovery fails the gateway errors out rather than forwarding a transcript with missing statements.

Jev Opus

About this project

Highlights

Questions about this entry

Reactions & coverage

Related terms

More Jev resources like this

From the Jev guides

From the Jev community