The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes local binaries located in the ~/.claude/skills/gstack/bin/ directory, including gstack-skill-start, gstack-slug, gstack-paths, gstack-decision-search, and gstack-skill-end. These commands are used for session management, path resolution, logging, and telemetry.
[DYNAMIC_EXECUTION]: The skill uses the eval command to execute shell output produced by other binaries (e.g., eval "$(~/.claude/skills/gstack/bin/gstack-slug ...)"). This allows the execution of dynamically generated code strings at runtime.
[DYNAMIC_EXECUTION]: The skill is designed to ingest and follow instructions dynamically provided by external tools via GSTACK_INSTRUCTION_BEGIN blocks. This mechanism allows the framework to inject new directives into the agent's context after the skill has already been loaded.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external files (SEED_PLAN) and dynamically injected instruction blocks, which serves as a potential vector for indirect prompt injection.
Ingestion points: Reads instruction blocks from the output of gstack-skill-start and test plans from file paths determined during the discovery step.
Boundary markers: Uses specific delimiters (GSTACK_INSTRUCTION_BEGIN / GSTACK_INSTRUCTION_END) and requires a matching SESSION_ID generated at the start of the session to validate instruction blocks.
Capability inventory: The skill has access to powerful tools including Bash, Write, Edit, Read, and Grep, enabling significant file system and shell operations.
Sanitization: Instructions mandate that the agent must only honor blocks that appear in direct tool results and carry a matching SESSION_ID, ignoring any similar patterns found in general file or page content.

test-audit