The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill explicitly directs the agent to execute shell commands extracted from the body of the markdown document it is auditing.
Evidence: In Analysis 6 ("Premise Verification"), the agent is instructed to "Run every PROBEABLE probe now" and specifically to "run the cell as written in the executing shell" based on contents of a Verifications table.
Evidence: The structural-compliance-rules.md (Step 8b) mandates running "verification commands" found in the plan to generate receipts.
[INDIRECT_PROMPT_INJECTION]: The skill has a high surface area for indirect injection as it processes and executes technical instructions from user-provided or external plan files.
Ingestion points: .plan.md files, specifically the ### Verifications table and **Exit gate:** sections.
Capability inventory: Shell execution using git, rg, gh, npm, and Test-Path, plus arbitrary commands defined in the plan.
Boundary markers: None identified to separate plan content from the shell execution context.
Sanitization: No explicit validation or escaping of commands derived from the plan document is implemented beyond standard agent safety filters.

tl-agent-plan-audit