Utility model
- The utility model (
[jev.local] model) runs on any catalog transport, including the ChatGPT subscription. Before, it only accepted Chat Completions with bearer auth, so it silently never ran. - Large tool results are compressed by unit selection. The utility picks source lines by id, and Distill copies the original text and always keeps required lines. The full output is stored, and the footer names its path. This covers:
- terminal and task output, including multi-task results;
- grep listings;
- web fetch;
- MCP results;
search_tool;- whole-file reads in read-only sessions.
- The utility now writes initial titles, shell autocomplete, prompt suggestions and recaps first.
- Memory capture trims large tool results with the utility before the main-model extraction.
- The end-of-turn report shows
Utility - Nx. Memory capture, Dream, the laziness and permission classifiers and image description now write usage rows.
Jev
- New decisions:
- early compaction timing (
d4_compaction_timing); - a memory-capture gate (
d5_memory_capture_gate); - plan and code-reviewer subagent model choice (
b7_subagent_model).
- early compaction timing (
- The post-edit review (C4) and the utility post-review act only at confidence 0.70 or higher.
- The footer shows
jev:no-keywhen Jev is enabled but its key is missing. - When
api_key_envis unset, the key variable follows the provider:OPENROUTER_API_KEYfor the OpenRouter providers. - Six levers without a call site were retired. Their config keys are ignored.
Subagents
- Each session keeps the worker model it started with.
/worker-modelalso updates the open sessions in the window. - Subagents inherit the parent's Jev effort policy, and their Jev decisions count in the parent's turn report.
Full changes: #25