The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze application execution traces, logs, and datasets which may contain untrusted external data.
Ingestion points: The skill reads dataset/goldens.jsonl, execution traces via normalize_traces.py, and human annotations to diagnose failures as described in SKILL.md and references/workflow.md.
Boundary markers: The instructions emphasize the use of reviewed failure taxonomies and specific development examples to scope analysis, which helps reduce the impact of potentially malicious data embedded in logs.
Capability inventory: The skill has the capability to generate and apply patches to the application's source code files during the experiment rounds described in SKILL.md and references/experiments.md.
Sanitization: references/workflow.md notes that output renderers escape content to prevent execution, and the skill protocol mandates that code changes are only delivered as patches for human review rather than being merged or deployed automatically.
[DYNAMIC_EXECUTION]: The skill performs automated setup and code generation tasks as part of its optimization loop.
Evidence: The scripts/setup_workspace.py script dynamically modifies template files (Python, TypeScript, JavaScript) by replacing hardcoded paths with the workspace root before writing them to the filesystem.
Evidence: The core functionality of the skill involves the AI agent generating and applying patches to the project's source code to test hypotheses for improvement, which is a form of dynamic code generation.

eval-descent