The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from various repository sources to perform its analysis and evolution tasks.
Ingestion points: Data is collected from git log output, file structure listings, pyproject.toml, package.json, and agent journals located in the .agents/ directory (reference/signal-collection.md).
Boundary markers: The skill uses instructional constraints such as "Propose, never force" and "Integrate, don't duplicate" in SKILL.md, and references a "review-before-merge default" for pattern synthesis in reference/subsystems.md.
Capability inventory: The skill has the ability to write to .agents/ECOSYSTEM.md and .agents/darwin.md, and it generates evolution proposals that influence the behavior and routing of other agents (e.g., Architect, Nexus).
Sanitization: There is no explicit evidence of sanitization, validation, or escaping of the ingested content (e.g., commit messages or journal entries) before it is processed by the agent.
[COMMAND_EXECUTION]: The skill's reference documentation defines specific shell commands that the agent is expected to execute to collect project signals.
Evidence: reference/signal-collection.md lists commands such as git log --since="30 days" --oneline | wc -l, git shortlog -sn --since="90 days", git log --stat, and find . -type f | wc -l. While standard, these represent a direct command execution interface with the host environment.
[METADATA_POISONING]: The skill incorporates a significant amount of deceptive or fictitious metadata regarding the current time and industry state.
Evidence: Throughout the reference files (e.g., reference/assessment-models.md, reference/evolution-actions.md), the skill cites a "2026-05 baseline" and references industry events, research papers, and product releases with future dates (e.g., May 2026). This creates a misleading context for the agent's operations and the validity of its internal benchmarks.

darwin