[Adambomb210]

A stopped box keeps billing for its disk until it's resumed or destroyed. Only /workspace is saved, and only for that same box. Members weren't told most of this:

• the job page and krater show gave a stopped box's full GPU price;
• the guide said its disk still bills "a little";
• the Slack notices said "its last saved workspace is kept", though no member can reach it once the box is destroyed: the workspace isn't a readable area, and it's stored per job.

What members now see

• Job page:
  • for a stopped box: "$0.0088/h while stopped (about $0.22 a day): only its disk bills", the full running price, and when it will be destroyed;
  • for every box: when /workspace was last saved, and what is and isn't saved;
  • Cancel for a box says what it loses. A running box loses work since its last save, with "stop the box instead". A stopped box's cancel ends the disk charge, and the box can't be brought back.
• API: adds machine.usd_per_hour_now and box.{destroy_at, destroy_after_hours, workspace_saved_at, workspace_save_minutes}. usd_per_hour is unchanged.
• CLI:
  • krater show prints the stopped cost, the destroy time and the last save. Against an older Krater it falls back to the old machine line;
  • krater cancel on a box says the box can't be brought back.
• Slack:
  • "Box stopped" gives the disk cost and the destroy time. Notices already queued at deploy time lack the new fields and render as before;
  • "destroyed soon" says to copy out anything needed;
  • "destroyed" no longer says the workspace is kept.
• Guide:
  • the concrete disk cost;
  • Cancel for boxes;
  • "the saved copy is for this box only";
  • two troubleshooting rows.

The rate comes from spend.hourly_rate_microusd, the same rule the accrual uses. There is no migration.

Not in this PR

A destroyed box's /workspace can't be downloaded, and a new box can't start from it. That's a separate decision.

Tests

• New tests for:
  • stopped and running box pages;
  • the API fields, including a job having no box field;
  • the CLI show and cancel lines, including against an older Krater;
  • the Slack notices with and without the new payload fields;
  • the scheduler's stopped-notice payload.
• The full server suite passes locally: 1940, plus the one updated assertion that pinned the old "workspace is kept" reason. The CLI's 195 tests pass. Ruff is clean.
• Not tried live: a stopped box on staging needs a paid box run.

🤖 Generated with Claude Code