The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to perform repository maintenance, file system scans, and invoke external CLI tools. Scripts in SKILL_detail.md utilize find, grep, git log, and sort to automate decision-making processes.
[DATA_EXFILTRATION]: The skill is configured to send project data, synthesized patterns, and skill proposals to external AI services for validation (Step 3.5-X).
Evidence: SKILL_detail.md contains bash commands that pipe data to gemini and claude --print to perform adversarial reviews.
Context: The destinations are well-known and trusted services, but the skill facilitates the transmission of internal project context to these external endpoints.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant surface area for indirect prompt injection as it ingests and processes untrusted data from git logs and session outputs to generate new logic and rules.
Ingestion points: Step 0-a (git diff), Step 0-b (git log), and Step 1 (field git diff and session outputs) serve as entry points for external data.
Boundary markers: Absent. The prompts defined in SKILL_detail.md (e.g., §Step3-75) interpolate ingested data without specific delimiters or warnings to ignore embedded instructions.
Capability inventory: The skill has Write and Bash access, allowing it to modify project instructions or execute shell commands based on the results of the automated analysis.
Sanitization: No sanitization or validation of the ingested git/output data is documented before it is passed to the LLM or used in script generation.
[AUTONOMY_CONCERNS]: The skill defines a 'Forced mode' intended for local development environments where the pipeline executes automatically at the end of a session without requiring user approval for the execution steps.

harvest-loop