{"content":"fix(78d913e3, branch fix/assistant-mode-bubble-display): cerebellum effort-verdict rate — three-part repair.\n\nForensics from local session 42b6a16d (2026-11-09): a plain 'let's make a November version' message took ~28 minutes, and cowork.log showed the auto effort selector returned NO verdict on 72% of production turns (93/130 across 2026-11-04..09), every failure an abort at the old 1s budget, while successful verdicts showed a 701ms median. Assistant mode then silently rode the fallback effort, and on the GLM responses wire 'off' drops the param entirely — glm-5.3-flash ran server-default thinking for minutes.\n\nChanges: (1) CEREBELLUM_EFFORT_TIMEOUT_MS 1000->2500 covers the observed production tail; (2) CEREBELLUM_EFFORT_RACE_BUDGET_MS 400->1200 so the runner actually waits for verdicts that now land (worst-case turn-start stall ~1.2s on a hung backend, median unchanged); (3) resolveAutoEffortFallback spells 'low' instead of 'off' for GLM-family assistant sessions — GLM-5.x always thinks and cannot express 'off' on any wire, so 'low' is the honest floor.\n\nVerified: compile:electron clean; cerebellumEffortSelector 15/15; coworkComposerEffortDefault 7/7 incl. new GLM fallback cases. Next up on this branch: the assistant-mode activity status display (event-derived specific status lines + zero-marginal-cost bundle-authored goal phrase).","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}