The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on executing shell commands for its core functionality of verifying code changes and running tests. This is handled by scripts/evidence_runner.py, which utilizes subprocess.Popen (supporting a --shell flag) to execute provided commands and capture their output for evidence records.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, such as PR/MR feedback, issue descriptions, and reviewer comments, creating an attack surface for indirect prompt injection.
Ingestion points: External comments, review threads, and issue bodies are fetched from GitHub and GitLab APIs (documented in references/review-collection.md).
Boundary markers: The skill uses a specific protocol marker coding-workflow:v1 to identify and verify its own prior actions and communications.
Capability inventory: The agent can execute arbitrary shell commands via scripts/evidence_runner.py, perform file writes for task storage, and execute Git operations including commits and pushes.
Sanitization: The skill implements best practices for log redaction in references/redaction.md and scripts/evidence_runner.py, which filters sensitive information from captured command output.
[SAFE]: The skill includes several features designed to enhance security and prevent accidental data leakage:
Credential Redaction: scripts/evidence_runner.py contains built-in regular expressions to detect and redact API keys, bearer tokens, and private keys from stdout/stderr before storing evidence.
Task Storage Isolation: scripts/task_storage.py ensures that all workflow-related metadata, indices, and evidence files are stored in a dedicated directory outside of the target code repository, preventing local environment pollution.
Mandatory Independent Review: The workflow (references/precommit-review.md) enforces an "adversarial review" loop where a fresh, independent subagent must verify code changes before they can be committed.

coding-workflow