The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data from bug reports and external source code, creating a potential surface for indirect injection where malicious data could influence agent instructions. 1. Ingestion points: Bug description arguments in SKILL.md and project manifests/source files scanned during Phase 0 as described in references/preflight.md. 2. Boundary markers: The protocol does not explicitly define prompt delimiters or 'ignore' instructions for data processed during localization waves. 3. Capability inventory: High-privilege actions including shell command execution, filesystem modification in the fix phase, and network operations for agent coordination. 4. Sanitization: The skill focuses on verbatim evidence gathering and does not mention filtering or sanitizing external content.
[EXTERNAL_DOWNLOADS]: The 'Preflight' protocol in references/preflight.md enables the agent to automatically install missing dependencies to clear environmental blockers, which involves downloading code from external package registries.
[COMMAND_EXECUTION]: The skill performs extensive shell command execution based on the repository's stack (e.g., npm, pytest, go test) to baseline behavior and verify fixes as defined in references/stack-adapters.md.
[REMOTE_CODE_EXECUTION]: The skill dynamically generates and runs reproduction scripts and permanent regression tests to validate bug fixes, which involves the execution of code generated during the debugging process.

nuke-debug