The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and process potentially untrusted external artifacts, including network traces, event logs, HAR files, and core dumps, to reproduce and diagnose issues (Phase 1, Phase 2).
Ingestion points: SKILL.md (Phase 1, Item 5 "Replay a captured trace"; "When you genuinely cannot build a loop").
Boundary markers: The instructions do not specify the use of delimiters or boundary markers to isolate untrusted debugging data during processing.
Capability inventory: The agent is authorized to execute shell commands, write to the filesystem (e.g., creating harnesses and tests), and interact with platform tools like gh and crucible.
Sanitization: No sanitization or validation of the ingested debugging artifacts is mentioned before they are replayed or analyzed.
[DYNAMIC_EXECUTION]: The skill's methodology revolves around the runtime generation and execution of code and scripts to build reproduction loops and instrumentation probes (Phase 1, Phase 4, Phase 5).
The agent is guided to create "Throwaway harnesses," "Bisection harnesses," and bash scripts (using scripts/hitl-loop.template.sh as a template) to reproduce bugs and verify fixes.
[COMMAND_EXECUTION]: The workflow depends on executing shell commands for tasks like bisection, reproduction, and network testing, including curl, git bisect, and custom CLI invocations (Phase 1).

diagnose