The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents a vulnerability surface for indirect prompt injection as it ingests and processes untrusted data from test failure output (stdout/stderr) and codebase source files. Maliciously crafted test outputs could attempt to manipulate the agent's classification logic or fix generation. Evidence Chain:
- Ingestion points: Phase 1 (parsing test results via regex) and Phase 2 (reading production and test source code).
- Boundary markers: The skill uses a structured 'Plan Artifact' (.agent/{branch}/test-auto-fix-plan.md) to manage context, but does not specify explicit delimiters or 'ignore' instructions for the raw test data processed at runtime.
- Capability inventory: The skill possesses extensive capabilities across its rules, including shell command execution (Bash), file writing/editing (Edit), and external tool invocation (Skill calls).
- Sanitization: There is no evidence of specific sanitization or filtering for the test output data before it is interpolated into the agent's reasoning process.
[COMMAND_EXECUTION]: The skill executes shell commands to detect test failures, run individual tests, and manage the environment. These commands are either detected during bootstrap (e.g., extracting scripts from package.json) or supplied by the user. It also performs git operations such as 'git remote get-url' to compute project keys and 'git restore'/'git revert' for regression recovery.
[DATA_EXFILTRATION]: The skill externalizes metadata about the project codebase, including failure patterns, file paths, and diagnostic verdicts, by sending them to the 'lorekit-memory' skill. While this is a functional requirement for the 'self-improvement loop' to track recurring issues, it involves transmitting internal project state to an external memory service.
[DYNAMIC_EXECUTION]: The skill implements a complete code-fix loop where it dynamically drafts edits to test files and immediately executes those modified files using the project's test runner to verify results. This involves executing code that was generated by the AI agent during the same session.

test-auto-fix