CI is red repo-wide: lancedb 0.5.x vanished from PyPI

[vicenteliu]

Every CI run since ~2026-08-22 (last green: the #212 merge push) now fails at the Install step, before any lint/test runs:

ERROR: Could not find a version that satisfies the requirement lancedb<0.6,>=0.5
       (from versions: 0.14.0, 0.15.0, ..., 0.38.0)

PyPI no longer serves any 0.5.x release — the oldest available is 0.14.0. Nothing in this repo changed; the constraint lancedb >= 0.5, < 0.6 (pyproject.toml:33, PR-4 era) is simply no longer satisfiable.

Blast radius

• CI: both pip install -e ".[dev]" jobs (Python, Rust ext) fail in seconds; observed on PR fix: the TUI tour gif stops racing its own workers #213, which is blocked through no fault of its own.
• Docker: the builder stage installs deps from PyPI before the --no-index --find-links=./wheels step, so image builds should fail the same way.
• Fresh clones / new contributors: uv sync / pip can't fetch the locked 0.5.7 either.
• Existing local venvs: unaffected (0.5.7 already cached/installed) — which is why this went unnoticed until a PR triggered CI.

Options

1. Migrate to modern lancedb (0.38.x today): real work — the 0.5 → 0.14+ API surface changed (connect/table/search APIs); touches src/opspilot/kb/lance_store.py and the pyarrow/pandas pins. Probably the right long-term answer.
2. Vendor the 0.5.7 wheel into the repo/CI cache as a stopgap: unblocks CI fast, but freezes on a dead version.
3. Reconsider the store: ADR-0017+ era code paths already abstract KB storage to some degree — worth checking how thick the lancedb coupling actually is before choosing 1.

Until one of these lands, treat every red CI run's Install-step failure as this issue, not the PR's content.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KXrF6HMB12V1hC9xFtooDE