The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes commands provided in the corpus-manifest.json (e.g., setup_commands like cargo fetch and test_commands like cargo test). Additionally, in Step 5 of the pipeline, it executes 'reproducers' (code snippets or test cases) generated or identified by the auditor worker during the analysis of external code.
[EXTERNAL_DOWNLOADS]: The skill fetches code from remote repositories to establish a baseline for auditing. It uses git to check out specific commits from URLs provided in the manifest, which can point to untrusted external sources.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted source code and documentation from external repositories. This content could be crafted to include malicious instructions targeting the auditor worker LLM to bias findings or generate malicious reproducers.
Ingestion points: Source code and corpus manifests from external repositories defined in the input corpus manifest (SKILL.md, fixtures/synthetic/corpus-manifest.json).
Boundary markers: The skill task prompt instructs workers to "Derive findings only from the sanitized source" and follow evidence contracts (SKILL.md).
Capability inventory: The system executes preflight scripts, setup/test commands, and reproducer code (SKILL.md).
Sanitization: Step 2 of the pipeline involves a sanitization process to remove existing audit reports, security reviews, and remediation notes from the target source tree (SKILL.md).

qedgen-auditor-bench