v0.5.2 — /atelier:doctor fully prompt-free (F16b)

Follow-up to v0.5.1 (M7.1.F16). After v0.5.1 added allowed-tools to commands/doctor.md, the M7.1 dogfood-3 doctor run still surfaced two permission prompts — both for cat <path> || echo "MISSING" patterns. v0.5.2 closes the gap.

What's new

• M7.1.F16b (PR #78) — commands/doctor.md narrative gains explicit "use Read tool, not cat" guidance. Root cause: Claude Code routes cat <path> through a path-scoped read-approval check that bypasses the Bash(cat:*) allow-list (the prompt format "Yes, allow reading from <dir>/ from this project" is the file-read pipeline, NOT the Bash command-pattern pipeline). Other compounds in doctor — gh api ... || gh api ..., command -v X && echo ..., [ -d ... ] — all work cleanly because they aren't subject to the file-read interception.

  The fix is narrative-level: tell the LLM running doctor to use the Read tool (already in allow-list) for file contents, and Bash(test -f path) for existence checks. Bash(cat:*) stays in the frontmatter for non-path edge cases.

After upgrading

claude plugin update atelier@akalab-tech
# Restart any open Claude session
atelier /atelier:doctor
# Should complete with zero permission prompts.

Spawned follow-ups (still deferred, captured in ROADMAP)

• M7.1.F14 — doctor drift checks should read from local marketplace clone, not source repo API (was bypassed for dogfood-3 by making atelier repo public).
• M7.1.F15 — doctor parallel checks should fail independently, not cascade-cancel.

Links

• Implementation: PR #78
• Preceded by: v0.5.1 (F16 — incomplete frontmatter fix)