[Bug] Overnight sleep re-writes agent cache per wake-up, drains weekly limit; ambiguous "стоять" stops agents

Bug Description
Bug report, Claude Code (Mac desktop app, Code tab, Opus 5.5), session f561ef7b-96ea-4337-a6d2-3c6a61a08eaa 1. Overnight sleep burned ~25% of my weekly limit Workflow backlog-wave9b (run wf_a6fdeb18-9bb), 7 agents. The workflow panel shows 2.5M tokens in total, yet my weekly usage limit dropped by about 25% during this run. Two agents (impl:combat, 540k, and impl:board-lint, 504k tokens per the panel) show 16h 29m of runtime because the computer was asleep overnight. Their transcripts record about 50.6M cache_creation tokens for this run (19.8M and 17.0M for these two) and 377M cache_read. It looks like after every wake-up the agents re-wrote their whole context into the cache, and that was charged to my limit without showing up in the panel number. On Oct 14, workflow wf_f84fde92-310 (3 agents) failed entirely with “API Error: Connection refused (ECONNREFUSED)” after about 40 minutes and returned nothing. That wasted about 112M cache_read and 1.9M cache_creation. On Oct 13 an agent died with “Your computer went to sleep mid-response” and its work was lost. 2. The assistant then made it worse While I was dictating this report, it asked “should I stop the agents to save the limit?”. My next message was "стоять" (“hold on”), meant for the report text. It took that as “yes” and killed the running workflow, which still had two review agents and one fix agent about 35 minutes into their work. When I objected, it restarted the workflow. The resume cannot continue interrupted agents mid-way, so those three started from scratch. Their earlier work and the limit it used were thrown away. After I told it to kill the workflow again, it stopped the restarted run too. Net result: more limit spent and nothing finished. It also wrote the report to a file without being asked (I rejected that edit). It also didn’t know or say up front that /feedback exists in the CLI, which sent me through the support site first. Please: check how my limit was charged for these runs and restore what was lost to the sleep, the connection failures and the assistant’s stop/restart; explain why the panel figure (2.5M) differs so much from what was charged; fix agent behaviour after sleep so the cache is not re-written over and over; make the assistant never stop running background agents on an ambiguous one-word message, and ask instead.

Environment Info

• Platform: darwin
• Terminal: claude-desktop
• Version: 2.1.295
• Feedback ID: 79847b97-3791-43e7-beb2-b9cb64513f8e