The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest untrusted data from 'technical discussion pastes' and evaluate them against a codebase, creating a surface for indirect prompt injection.\n
Ingestion points: The agent receives user-supplied pasted content from parallel technical sessions for analysis, as detailed in SKILL.md.\n
Boundary markers: The skill enforces a strict 'Iron Law' ('ANSWER FROM THE CODE, NOT FROM THE PASTE') and utilizes a set of separate instructional files (plain-card.md, grounding.md, session-map.md) to enforce clear separation between user claims and codebase reality. It also uses explicit 'Fence' sections to define decision limits.\n
Capability inventory: The agent has the capability to read any file in the local repository and execute git commands such as git log -p and git blame. It also has the ability to write project-specific metadata to the .skills/interpret-session/ directory.\n
Sanitization: The skill uses a sub-skill called speak-outer to review and sanitize 'carry-back' blocks before they are presented to the user. It explicitly differentiates between verified facts (from the repo) and derived commitments (to be checked by the user) to ensure design hallucinations or malicious injections are flagged as unverified.

interpret-session