samuelfaj/distill v2.0.43 on GitHub

Jev 1.14 via OpenRouter, and agents can call Jev directly

OpenRouter only. Jev now runs only through OpenRouter Decisions (https://openrouter.ai/api/alpha/decisions, model ~typesafe/jev-latest, key OPENROUTER_API_KEY). Any other provider, host, key or model is rejected, and Jev turns off instead of sending decisions elsewhere. Redirects are not followed.
Long context. Requests over 32 KiB use Jev 1.14's long-context mode (up to 768 KiB of state). It was verified live at 200k input tokens.
New jev_decide tool. The main agent and its general-purpose, explore, plan and code-reviewer subagents can ask Jev typed questions (noul, choice, score) about a small state or stored tool outputs.
- Source lines are numbered verbatim, and each answer reports how much of each source was sent.
- Secrets are refused before anything is sent, and usage is recorded in the session.
- Images are refused for now, because OpenRouter currently ignores them.
More automatic decisions.
- Failed commands now get a failure category, and edits get a change type, both as advisory hints.
- Before a main turn ends, Jev checks whether the requested work or proof is missing and adds at most one reminder per prompt.
- The delegation hint and the model-tier choice are now on by default.

Jev 1.14 was compared with gpt-6.1-sol (main) and gpt-6-luna (worker). Each classification benchmark has 8 labelled synthetic cases.

Failure triage, change type and end-of-turn check: all three systems scored 100%. Jev answered in about 1.1 s versus about 1.6 s and cost 12–25× less than the main model. The end-of-turn check caught 4 of 4 unfinished turns.
Finding one error in a ~600 KB build log: the long-context request scored 100% for $0.054, versus $2.79 on the main model. The old 32 KiB limit scored 0%.
Two questions batched into one request were about 2× faster and 33% cheaper than two separate requests.

Full details are in #34.