The Silent Failure That Made Slots Run Blind
The slot system has a workspace clone for each slot — a wksp/ symlink that connects each repo clone back to its workspace. When it works, ctx.py detects the workspace, handoffs go to the right place, and session state flows naturally. When it breaks, the slot runs as SINGLE_REPO=yes — no workspace, no handoffs, no design artifacts. No error. No warning. Just a session that silently drops half its capabilities.
Five slots were running like this. I found them by accident.
The root cause was resolve_workspace_source(), which returns "work" as the default workspace clone name. The collision check only looked at the repos in the current slot — not all repos in the family. If the family had a repo called work (which casehub does), the workspace clone landed on top of it. After that, add_repo("work") would fail with repo_already_in_slot because the directory existed — but it was the workspace, not the repo.
The second problem was downstream. When git clone --shared creates the workspace clone, gitignored directories don’t exist in the clone. The workspace’s .gitignore lists child repo directories (/engine, /soredium) because those are separate git repos checked out as siblings. The code already calls mkdir and _unignore_subdir to create them — but with no validation afterward, a failure in that sequence was invisible. The wksp/ symlink would point to a directory that should exist but doesn’t.
The fix has three layers. First, _get_family_repo_names() scans all top-level git directories in the family root for collision detection — not just the current slot’s repo list. Second, validate_slot_wksp() runs after every create_slot and add_repo call, checking that each wksp/ symlink resolves. If any dangle, it exits 1 with the specific failure. Third, list_slots now surfaces WKSP=broken in its output so broken symlinks don’t hide until someone notices a session misbehaving.
Along the way, I hit a Python gotcha worth remembering: Path("").is_dir() returns True. An empty string becomes the current working directory. resolve_original_repo was calling git remote get-url local through a mock that returned an empty string — Path("") resolved to the CWD, which happened to have a wksp/ symlink, and validation failed against the wrong repo entirely. The fix is a one-liner (and url.strip()), but the failure mode was genuinely confusing to trace.
The broader pattern here is worth naming: silent degradation is worse than a crash. A slot that runs as SINGLE_REPO=yes still works — you can edit code, run tests, make commits. Everything looks fine until you try to find the handoff or the design spec and they’re not where they should be. Five slots ran like this for weeks. Post-creation validation catches the failure at the moment it happens, when the fix is obvious, instead of leaving it to surface as a confusing symptom three sessions later.