[NaNIsaac]

merge_queue.py (after #33) made one search, one repository listing and one heavy gh pr list (files, checks, comments, body) per repository on every run, and a single failed gh call ended the run with a bare traceback and no gh message.

• gh failures: gh() raises GhError with the command, exit code and the last 400 characters of stderr, and prints it to stderr, so the sentinel's merge-queue.log shows it. A failed search or repository list ends the run with that message instead of a traceback.
• One failure does not stop the run: a repository or PR gh cannot read goes into a "조회 실패" section of the digest (name and reason) and the rest is still generated; the exit code is 0.
• Heavy fields only where needed: with a fresh prs.py snapshot (at most 30 minutes old, HERDR_BOARD_DIR overrides the directory) the repositories come from it, and a PR whose (head sha, updated, checks) match the last run is taken from merge-queue-cache.json next to the snapshot. Only changed PRs are fetched, one gh pr view each. Mergeable always comes from the snapshot. Without a usable snapshot it asks gh as before, one gh pr list per repository.

Measured here on the real account, same output both times: first run with an empty cache 75 s (one view per PR), next run 20 s. The remaining time is the per-base commit lookups for the stale check, unchanged.

local-ci: 2d0e974 | merge_queue.py selftest ok (new cases: no snapshot, cache hit fetches nothing, new head sha fetches one PR, mergeable from snapshot, old and corrupt snapshot, a PR gh cannot read is reported and the others stay, gh stderr in the error); ruff format clean | not run: ruff check tools/conventions.py has one existing import-order finding, left alone

🤖 Generated with Claude Code