v0.5.1 — /atelier:doctor runs without permission prompts (F16)

Patch release fixing a frontmatter oversight discovered during M7.1 dogfood-3 first /atelier:doctor run. Plugin scope (commands/doctor.md). Pulled in by claude plugin update atelier@akalab-tech.

What's new

• M7.1.F16 (PR #77) — commands/doctor.md was the only slash command in atelier without an allowed-tools frontmatter entry. Every gh api, cat, jq, claude plugin list, docker compose version doctor ran triggered an interactive permission prompt — pure friction for a strictly read-only command. Fixed by adding a comprehensive read-only allow-list (Read, Glob, Grep + 15 Bash patterns covering plugin queries, gh API, file reads, JSON/text parsing, binary presence checks, docker probes). No write-capable tools included.

Spawned follow-ups (captured in ROADMAP, deferred)

The same /atelier:doctor run that surfaced F16 also surfaced two architectural gaps in doctor's design. Captured as separate deferred ROADMAP entries:

• M7.1.F14 — doctor's drift checks query gh api repos/AkaLab-Tech/<repo>/releases/latest, which fails 404 when the source repo is private and the gh identity lacks org access. Worked around for dogfood-3 by flipping AkaLab-Tech/atelier to public; the architectural fragility (depending on source-repo API access for drift detection) remains. Real fix: read from the local marketplace clone instead.
• M7.1.F15 — doctor's parallel checks cascade-cancel when any one fails (Claude Code's parallel-tool-call default behavior). Real fix: sequential checks, or wrap each in || true so non-zero exits don't bubble up.

Both will land as separate small PRs.

Upgrade

claude plugin update atelier@akalab-tech
# Then re-run doctor from a fresh atelier session:
atelier /atelier:doctor
# Should complete without permission prompts.

Links

• Implementation: PR #77
• Predecessor: v0.5.0 (M7.1.F7b orchestrator-side identity adoption)
• Series source: PR #69 dogfood-2 ROADMAP batch