The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted data from CI logs, APM trackers, and alert payloads to identify and diagnose errors. This creates a surface for indirect prompt injection if an attacker can control the content of an error log or CI output. * Ingestion points: The skill reads from CI status (gh run list), error-log queries (e.g., logcli), APM trackers, and alert payloads as described in references/watch-and-heal.md. * Capability inventory: The skill has access to high-privilege tools including Bash for command execution and Edit for modifying the codebase. * Boundary markers: A strict HITL (Human-In-The-Loop) boundary is enforced for all irreversible actions such as merging PRs, deploying code, or mutating production data. * Sanitization: The workflow requires a deterministic reproduction in a sandbox environment before a fix is attempted, which provides a layer of validation against malicious signal noise.
[COMMAND_EXECUTION]: The skill uses the Bash tool to execute CI commands (gh run list, gh run view) and potentially other log-querying tools. These operations are essential for the skill's primary function but involve executing commands based on external configurations and data.

self-healing-loop