The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the repository being analyzed to build its model.
Ingestion points: The agent ingests file content from the target repository tree and architectural documentation during the reconstruction process.
Boundary markers: The skill instructions enforce a strict structure for claims, requiring the agent to classify statements as 'fact', 'inference', or 'unknown' and provide specific 'file:line' citations for every assertion. This helps maintain a clear boundary between the repository's data and the agent's synthesis.
Capability inventory: The skill's primary capability is codebase reconstruction. It is limited to reading files and writing reports to a specific artifact directory (.agents/recon/). It does not have permissions for network operations, privileged system access, or execution of the code it analyzes.
Sanitization: The output JSON is validated for structural correctness and path existence by a provided shell script, though the textual content of the report is generated by the LLM based on repository inputs.
[COMMAND_EXECUTION]: The skill includes a local validation script, scripts/validate-output.sh.
The script uses jq to perform schema validation on the agent's output artifacts.
It verifies that cited evidence exists on the filesystem using standard shell file-existence checks, correctly handling symlinks and line-number suffixes without executing the target files.
The script uses defensive coding practices, including set -euo pipefail and proper variable quoting.

codebase-recon