The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides automated graders that execute shell commands to verify agent output through unit tests and static analysis.
Evidence: Graders/CodeBased/BinaryTests.ts uses bun $ to execute a test_command on test_files, and Graders/CodeBased/StaticAnalysis.ts executes arbitrary linter/compiler commands.
Context: These tools are essential for the skill's purpose of deterministic code evaluation but require the runner to use trusted configuration files.
[INDIRECT_PROMPT_INJECTION]: The skill features a system that ingests external data from failure logs and reflections to draft new evaluation cases, creating a potential surface for indirect injection.
Ingestion points: LIFEOS/MEMORY/LEARNING/REFLECTIONS/algorithm-reflections.jsonl and LIFEOS/MEMORY/OBSERVABILITY/ask-fidelity.jsonl are scanned by Tools/ProposeFromFailures.ts.
Boundary markers: The ingestion logic does not use explicit boundary markers when processing log entries.
Capability inventory: Drafted cases can utilize BinaryTestsGrader or StaticAnalysisGrader which perform shell command execution.
Sanitization: The risk is mitigated by the 'Discretionary' design described in Tools/GenerateCases.ts and Tools/ProposeFromFailures.ts, where all generated cases are placed in a _drafts/ directory for manual review and graduation by a human before they can be executed.
[DATA_EXPOSURE]: The evaluation runner incorporates local identity files into the system prompt of the agent being tested to ensure operational realism.
Evidence: Tools/EvalRunner.ts reads LIFEOS_SYSTEM_PROMPT.md and DA_IDENTITY.md to build the test context.
Context: While this exposes identity information to the inference models during evaluation, it is a necessary part of testing the agent's specific behavioral dispositions.

Evals