The Agent Skills Directory

[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface by ingesting untrusted data from external sources during the debugging workflow.
Ingestion points: Ingests symptom descriptions, error messages, stack traces, application logs, and git blame/log history (SKILL.md Step 1, Step 3; reference/debug-flow.md Method 6).
Boundary markers: No explicit delimiters or instructions are provided to the agent to treat ingested logs or error text as untrusted content.
Capability inventory: The skill maintains high-privilege capabilities including file modification (Edit/Write), shell execution (Bash), and agent dispatching.
Sanitization: The instructions lack evidence of sanitization or escaping for data interpolated from logs or user-provided error messages into the agent's context.
[COMMAND_EXECUTION]: The skill's primary function involves executing a variety of shell commands to reproduce and verify bugs.
Evidence: The workflow includes running test suites (vitest, playwright), build tools (tsc), linting (eslint), and custom reproduction scripts (Node.js, Python) as detailed in the feedback loop methods (reference/debug-flow.md Section 1).
Context: These executions are confined to the local development environment and are triggered as part of the agent's systematic debugging protocol.
[DYNAMIC_EXECUTION]: The skill dynamically generates and injects code during the investigation and remediation phases.
Evidence: Step 5 of SKILL.md involves "Instrumenting" the codebase by adding log statements, and Method 4 in the reference guide describes writing and executing throwaway reproduction scripts.
Mitigation: The skill includes a mandatory cleanup phase (Step 6) using random session tags to remove all instrumentation, reducing the risk of accidental persistence.

debug