The Agent Skills Directory

[COMMAND_EXECUTION]: The skill automates a complex multi-stage review process by executing various CLI tools including git, opencode, ocr, claude, uv, jq, and serena. These are invoked via Python's subprocess module using command lists to manage repository state and launch review arms.
[EXTERNAL_DOWNLOADS]: The skill orchestrates remote model calls through opencode and ocr scanners. It relies on these external tools to communicate with AI model providers for the code review tasks. All model IDs are validated against a local whitelist before execution.
[DATA_EXPOSURE]: The skill reads the operator's local configuration files (e.g., ~/.config/opencode/opencode.json and ~/.serena-reviewer/serena_config.yml) to resolve API keys and tool settings. It carefully creates isolated configuration snapshots in a temporary state directory (.llmtmp/review-deep) to manage subagent permissions without exposing the primary environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes entire codebases as input for auditing, creating a surface for indirect prompt injection.
Ingestion points: Repository files are read by multiple review agents via opencode run and ocr scan (SKILL.md).
Boundary markers: Reviewers are scoped via specific ledgers (ledger.txt) and directory-aware partitions that define their context.
Capability inventory: The review subagents are highly restricted via configuration, with all shell, task, and skill tools disabled or denied. The orchestrator maintains necessary capabilities for repository management and reporting.
Sanitization: The skill performs a multi-model verification pass (verify.py) where findings are checked against a clean git archive export in an isolated directory, providing a layer of validation against hallucinations or injections that might rely on uncommitted artifacts.

review-deep