The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from the project workspace through the view_file tool and the anchor.ts script (e.g., when pinning code snippets or verifying task completion). This content is then interpolated into anchor files. If a project file contains malicious instructions in comments or strings, they could influence the agent's behavior during resumption.
Ingestion points: Project files read via view_file and parsed by scripts/anchor.ts (e.g., pinAttentionContext function).
Boundary markers: The skill uses markdown headers and blockquotes to delimit anchor content.
Capability inventory: File reads/writes via tools, and local git execution via the anchor.ts script.
Sanitization: The skill provides a 'Context Hierarchy & Trust' protocol instructing agents to treat external/browser content as untrusted and never obey it as an instruction.
[COMMAND_EXECUTION]: The scripts/anchor.ts engine executes local git commands (branch, status, log, diff, checkout) using spawnSync. This is necessary for the skill's stated purpose of verifying task ground truth and performing workspace rollbacks during failure loops. The command arguments are strictly defined within the script and do not allow arbitrary shell execution from untrusted inputs.

context-anchor