The Agent Skills Directory

[PERSISTENCE]: The skill installs a persistent background agent on macOS using launchctl. Specifically, it creates a launchd plist file (com.kntnt.model-selector.refresh.plist) in the user's LaunchAgents directory to run a daily catalogue update job.
Evidence: scripts/capture.py contains logic to install and remove the refresh schedule using launchd constants and labels.
[DYNAMIC_EXECUTION]: The skill employs multiple dynamic execution techniques to load logic and manage tool invocation.
It uses importlib.util to load internal "Library" and sibling scripts from file paths computed at runtime.
scripts/invoke.py implements a search algorithm that iterates through dozens of hidden tool-specific directories (e.g., ~/.claude/skills, ~/.codex/skills) to find and execute a management engine script (kntnt.py) via uv run.
[INDIRECT_PROMPT_INJECTION]: The skill processes recorded unit-of-work data, which includes excerpts of previous instructions and results, to generate prompts for a "judge" model that evaluates performance.
Ingestion points: Transcripts are read from paths like ~/.claude/projects/ and ~/.codex/sessions/ in scripts/capture.py and scripts/quota.py.
Boundary markers: Present. In scripts/grade.py, the prompt_for function uses Markdown-style headers ("Instruction:", "Result:") to delimit processed content.
Capability inventory: The skill uses subprocess.run and subprocess.Popen in scripts/grade.py and scripts/selection.py to execute commands and dispatch tasks.
Sanitization: Partial. Excerpts of instructions and results are capped at 4,000 and 12,000 characters respectively in scripts/capture.py.
[COMMAND_EXECUTION]: The skill frequently uses the subprocess module to interact with local CLI tools and manage its own lifecycle.
scripts/catalogue.py uses subprocess.Popen with start_new_session=True to communicate with the claude and codex CLIs for model discovery.
scripts/grade.py uses subprocess.run to call the selection.py engine and subprocess.Popen to dispatch grading tasks in a new session.
[EXTERNAL_DOWNLOADS]: The skill communicates with external services to keep its model catalogue current.
scripts/catalogue.py performs unauthenticated HTTPS GET requests to the OpenRouter API (openrouter.ai/api/v1/models) to retrieve pricing and metadata updates.

model-selector