Fixes
- Worker model is now per instance: several Distill windows no longer overwrite each other's worker,
/worker-modelalways applies to the current window, and new/resumed sessions keep that window's worker. - The
<orchestration>prompt section follows the session's real worker and main model, including after switches and on resumed sessions.
Cheaper parallel delegation
- Batch delegation rules in the main prompt;
exploresearches before reading whole files. - Subagent reports capped (full text stored on disk);
repository_reviewlimited to the child's own files and shown in the TUI. - Visible warning when a child falls back to the main model;
model=in<subagent_meta>. - Background children finishing together wake the main model once; long children get one "report now" reminder.
- Sibling subagents share prompt cache; file edits lock per path across sessions.
- Jev B6 delegation decision recorded in shadow mode (hint behind
b6_delegation_hint). - Late background-child usage counted in the turn that spawned it; new
cohort-parallel-v1cost benchmark.