The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from external sources and uses it to generate prompts for sub-agents.
Ingestion points: The skill reads task content from GitHub issue bodies via the gh issue view command and from local markdown files (SKILL.md, lines 14-25).
Boundary markers: There are no explicit boundary markers or instructions to sub-agents to ignore embedded commands within the ingested task data.
Capability inventory: The sub-agents (referenced in references/harness-*.md) utilize powerful tools such as Agent, Workflow, and actor which possess filesystem read/write capabilities.
Sanitization: The skill lacks sanitization or filtering of the external content before interpolating it into sub-agent prompts.
[COMMAND_EXECUTION]: The skill uses the GitHub CLI (gh issue view) to fetch issue data from remote repositories. While GitHub is a trusted service, the output is treated as instructional data for other agents.
[REMOTE_CODE_EXECUTION]: The skill employs dynamic execution patterns to orchestrate sub-agents on specific platforms.
Script Generation: In references/harness-claude-code.md and references/harness-mimo-code.md, the skill generates inline JavaScript orchestration scripts at runtime to manage task concurrency and dependencies.
Execution Method: These scripts are executed via platform-native Workflow or workflow tools. While the scripts follow fixed templates, they incorporate external task data via string interpolation.

dispatch