The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions explicitly authorize the agent to utilize the local environment as an active reasoning system. This includes writing and executing scripts and running shell commands. It also instructs the agent to maintain its state within a hidden directory structure ('.theory-runs/').
Evidence in 'references/repo-lab-protocol.md': 'Write scripts, generate comparison tables, inspect local files, run shell commands...'
[EXTERNAL_DOWNLOADS]: The skill is designed to use web access and external sources to sharpen its reasoning and discipline conjectures.
Evidence found in 'SKILL.md' and 'references/repo-lab-protocol.md'.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from the web and external sources.
Ingestion points: Web search results and external source materials as noted in 'SKILL.md' and 'references/repo-lab-protocol.md'.
Boundary markers: None identified; the skill does not specify delimiters or instructions to ignore embedded commands in external content.
Capability inventory: High impact capabilities including local file system read/write access, script generation, and shell command execution across the workbench files.
Sanitization: No sanitization or escaping protocols are mentioned for content fetched from external domains before it is processed in the conjecture-criticism loop.

theorize-metaphysics