Regression Ledger

| `/health:check --scope=stack` (and `agentic`, and both `--fix` rows) told the agent to "follow the steps" in `${CLAUDE_PLUGIN_ROOT}/skills/health-audit/SKILL.md`, whose steps run `bash ${CLAUDE_SKILL_DIR}/scripts/enabled-plugins.sh` and `claude plugin list --json --available`; read that way, `${CLAUDE_SKILL_DIR}` is not substituted for health-audit's directory and health-check grants neither command, so the user entry point to the stack audit ran a wrong path or prompted. | Claude Code substitutes `${CLAUDE_SKILL_DIR}` and applies `allowed-tools` only for the skill it loads; a SKILL.md read with the Read tool is plain text. health-check now invokes `health-plugin:health-audit` and `health-plugin:health-agentic-audit` with the Skill tool (passing `--fix`, `--dry-run`, `--verbose`), so each runs with its own directory and grants. | `health-plugin/scripts/tests/test-health-check-skill.sh` invariant 5: no `${CLAUDE_PLUGIN_ROOT}/skills/*/SKILL.md` path in the body, and both audits named beside "Skill tool" (restoring the follow-the-steps text in a copy fails all three assertions) | 0 |