Jev Router — mejiasd3v

Yes, the pin is deliberate. Model and effort stay fixed for the session, and effort stays fixed too unless you opt into adaptive effort for Codex Astra. Reloading after editing settings does not move an existing pin, and Pi's own thinking picker neither controls nor tracks the router's effort. /jev shows the pin, the selected effort, and fork suggestions; a fork chooses afresh.

Initial routing reads recent messages within a bounded request budget. Effort checks send the latest user-text excerpt plus up to eight recent user, assistant, and tool-result excerpts, each keeping up to 1,600 characters split between its beginning and end. Tool names and error flags are included; tool-call arguments, reasoning blocks, images, and system messages are excluded. Tool-result text can contain secrets and is not redacted, and these evaluations are billed separately.

Every routing path is bounded by timeoutMs (five seconds by default) with no retries. On adaptive-effort failures the current effort is retained and cancellation stops the request. A route with no eligible effort level is excluded entirely, including non-reasoning models when the floor is above off, and the fallback errors if it has no eligible choice. Skill loading and monitoring degrade to no-ops rather than partial results.

One global file. You merge jevRouter into ~/.pi/agent/settings.json, and project settings cannot override routing; PI_CODING_AGENT_DIR is respected for locating that global directory. Routes replace the default list rather than merging with it, so only models you list and have authenticated are eligible, and the configured fallback must itself be listed or startup fails.

Luna at max effort, Sol at auto, Astra at xhigh, Astra as fallback, a five-second timeout, and monitoring on. Configuration replaces the default list rather than merging, and only listed, authenticated models are eligible, with the fallback also required to be listed.

Jev picks the model by task fit first, then the lowest sufficient allowed effort within it. thinking can be auto, a forced level clamped to the model's capabilities, a custom map of choices, or omitted to inherit Pi's level. A global or per-model minThinking floor can rule out levels below it.

Both are opt-in. With skills enabled, Jev matches Pi's discovered skills against recent text and loads up to three with a probability of at least 0.8, injecting at most 50,000 bytes per turn. Adaptive effort applies only to Codex Astra and re-checks effort before each main-model request, bounded to one extra evaluation per request context.