The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external commands to perform workspace diagnostics.
Evidence: The script scripts/ecosystem_scans.py invokes the loom CLI tool using subprocess.run([loom_binary, "workspace", "doctor", "--json"], ...) to analyze projection drift and workspace health.
[INDIRECT_PROMPT_INJECTION]: The skill audits and processes SKILL.md files from multiple sources, creating a potential surface for instructions embedded in other skills to influence the audit logic.
Ingestion points: The tool reads frontmatter and content from SKILL.md files across various registry and projection roots in ecosystem_scans.py.
Boundary markers: The skill implements a hardening mechanism in ecosystem_reconcile.py that appends trigger boundary clauses to skill descriptions to prevent accidental agent activation.
Capability inventory: The skill can modify file contents (_atomic_write), create or replace symlinks (path.symlink_to), and execute shell commands (subprocess.run).
Sanitization: Data is processed using yaml.safe_load and json.loads to prevent unsafe deserialization of skill configuration.
[SAFE]: The skill implements secret detection with automated redaction to prevent data exfiltration.
Evidence: The scan_secrets function in scripts/ecosystem_scans.py uses predefined regex patterns to detect high-confidence secrets (API keys, private keys). The logic specifically suppresses the actual values in reports, only noting the rule, line number, and file path.
[SAFE]: The skill follows a plan-first authorization model for filesystem mutations.
Evidence: As defined in SKILL.md and scripts/ecosystem_reconcile.py, the tool separates the 'audit' and 'plan' modes from 'repair'. It generates a detailed repair plan and requires explicit user authorization before applying any changes, including backups for configuration files.

skill-ecosystem-doctor