EPS
← All tasks·#2132Proposed

daily-held: perform the #681 worktree bind cutover

kind: infra#daily-held#needs-human
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Held item

The #681 design (heavy worktree footprint on the 512 GB data disk, bind-mounted onto .claude/worktrees) was never activated: probed 2026-08-06 — findmnt .claude/worktrees → NOT a mount point; df -h → / 945G at 97% (32G free) vs /mnt/eps-data 1007G at 23% (782G free after the overnight cleanup). ~212 GB of .claude/worktrees sits on the root disk, which spent 2026-08-05/06 oscillating 19.9 → 13.1 GiB free with the watcher warning that foreground Bash spawns fail silently (#552 class) — while the data disk has ~782 G free. Thomas hit this live ("what is this… the #681 bind was never activated", "how can we solve this") and the #2054 capture had to be re-routed to a RunPod for storage.

Which carve-out holds it: destructive/irreversible actions — the cutover moves live worktrees (the standing memory note says the migration was deliberately DEFERRED because it would break live sessions), so the timing/window is Thomas's call.

Related open tracking (named so the PM reconciles): #2099 (proposed) covers the CLAUDE.md § Disk hygiene DOC claim ("bind status runtime-checked, not asserted"); #2095/ #2097 cover janitor/headroom mechanization. None performs the cutover.

Suggested action

Pick a low-activity window (fleet idle, no live gate runs), then: quiesce sessions → rsync .claude/worktrees to /mnt/eps-data → bind-mount → verify new_worktree.sh's EPS_WORKTREE_REQUIRE_BIND=1 assertion passes → resume. The worktree-migration.LOCK machinery for exactly this cutover already exists (new_worktree.sh refuses while held). If preferred, file it as a scheduled kind: infra task with the window named and Thomas's go/no-go as the approval gate. With 782 G free on the data disk, the window is open now.

Provenance

  • origin: /daily 2026-08-05 problem sweep — miner 3 P5 (probed findmnt/df) + miner 4 P2; Thomas's live correction in session 7a1632b8 (~22:0xZ 08-05).

Folded in from #1038 (2026-08-09)

#1038 ("schedule #681 bind-migration cutover + disk sign-off", filed 2026-07-04) is archived in favor of this task — same cutover decision, older numbers. What survives from it:

  • Fresh numbers (probed 2026-08-09): .claude/worktrees = 263 GB on the root disk (212 GB at this task's filing on 08-06; ~97 GB in early July — steady growth), / at 87% used / 130 GB free, /mnt/eps-data at 27% used / 738 GB free. findmnt -T .claude/worktrees → still root ext4, no bind. The cutover window remains open.
  • Deletion sign-off remainder: /tmp/i779_mirror_dl (59 GB in July) no longer exists; ~/.cache/huggingface is 64 GB (above the disk guard's 50 GB HOME-HF-cache cap arm) and is the main surviving sign-off candidate from the 07-03 list.