The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data such as execution traces and datasets for failure analysis and grading.
Ingestion points: references/error-analysis.md instructs the agent to gather representative traces from production or synthetic sources. references/trajectory-grading.md involves evaluating tool-call sequences from various runs.
Boundary markers: The instructions for using 'LLM judges' in references/eval-techniques.md do not specify the use of delimiters or instructions for the judge to ignore malicious content embedded within the traces.
Capability inventory: The skill possesses the capability to execute local scripts (scripts/*.mjs) and perform autonomous file mutations during 'keep/discard' loops as defined in references/agent-loop.md.
Sanitization: No sanitization or escaping mechanisms are prescribed for the external trace content before it is processed by the agent or submitted to LLM-based grading rubrics.
[PROMPT_INJECTION]: Instructions within the skill promote a high degree of autonomy that explicitly avoids human oversight during execution loops.
Evidence: references/agent-loop.md contains the instruction: "Do not pause mid-loop for permission unless the user interrupted." Similarly, references/karpathy-patterns.md cites the pattern: "NEVER STOP until interrupted."
These instructions increase the risk that an agent redirected by a prompt injection (direct or indirect) would continue executing malicious actions without a verification gate.
[COMMAND_EXECUTION]: The skill includes and relies upon the execution of several JavaScript files to function correctly.
Evidence: scripts/loop-report.mjs, scripts/eval-eval.mjs, and scripts/check-description.mjs are provided within the skill folder. The agent is instructed to run these using Node.js for self-testing and reporting tasks. These scripts use the standard node:fs module to read and validate local files.

octocode-graph-eval