The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute several local Python scripts and system tools as part of its workflow orchestration. Key examples include running .agents/skills/wtk-lean/scripts/workflow_route.py for routing features and .agents/skills/wtk-deep-review/scripts/token_metrics.py for usage analysis. It also relies on standard development commands such as git, make check, and docker-compose for environment management.
[DATA_EXPOSURE]: The skill is designed to read metadata and token usage events from sensitive local directories associated with AI agent harnesses, such as ~/.codex/sessions, ~/.claude/projects, and $CLAUDE_CONFIG_DIR. This access is specifically targeted at session metrics; the instructions explicitly forbid reading unrelated session content, prompts, or credentials.
[INDIRECT_PROMPT_INJECTION]: As an orchestrator, the skill defines a large attack surface by ingesting untrusted external data such as feature plans (plan.md), check definitions (checks.md), and UI mockups.
Ingestion points: Primary entry points include SKILL.md (route selection), references/evidence.md (claim verification), and references/ui-ux.md (visual requirements).
Boundary markers: The instructions place heavy emphasis on strict adherence to contract schemas and approved artifacts, which serves as a delimiter against arbitrary data.
Capability inventory: The skill possesses capabilities to execute Python scripts, Git commands, and build tools.
Sanitization: It relies on pre-installed, non-author session readers and specialized scripts to parse metrics and validate outputs.
[EXTERNAL_DOWNLOADS]: The skill references multiple official documentation sites for AI platforms and development tools, including code.claude.com, cursor.com, openai.com, and docker.com. These are well-known services and trusted organizations used solely for providing implementation guidance and pricing information.

wtk