The Agent Skills Directory

[DATA_EXPOSURE]: The investigator subagent prompts, specifically for the Change Analysis and Dependency/Environment roles, instruct the agent to check for changes in .env files and other configuration sources. Accessing .env files is considered a high-severity finding as they often contain sensitive application secrets and credentials, although this behavior is dropped to medium severity here as it is contextually appropriate for a debugging tool.- [COMMAND_EXECUTION]: The orchestrator and its subagents perform various shell operations, including git commands (diff, commit, revert) and executing test suites (e.g., via npm test or pytest). These commands are executed against the local codebase, which constitutes untrusted input, creating a risk for unintended code execution during the debugging process.- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process codebase contents, error logs, and user reports, providing a surface for indirect prompt injection. This is mitigated by a structured multi-agent architecture where the orchestrator does not read raw codebase data directly, instead relying on synthesized reports from specialized subagents that act as data filters.
Ingestion points: Codebase files, error logs, and user-provided bug descriptions.
Boundary markers: Use of Phase Handoff Manifests and Synthesis Reports to consolidate data and strip raw context between phases.
Capability inventory: Shell command execution (git and test runners), file system writes for session state, and subagent dispatching.
Sanitization: The Synthesis agent is tasked with cross-referencing findings and identifying contradictions, providing a layer of validation for data ingested by investigators.

debugging