The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands via Node.js child_process.execSync to facilitate its diagnostic workflow.
The commands utilized include gcloud storage ls, gcloud storage rsync, and several gh CLI subcommands (issue create, project item-add, api graphql).
Execution is used exclusively for synchronizing evaluation artifacts and generating diagnostic reports.
Inputs for these commands are primarily derived from system state or restricted through regular expression filtering, limiting the risk of arbitrary command injection.
[EXTERNAL_DOWNLOADS]: The skill synchronizes diagnostic data from a remote Google Cloud Storage bucket.
Evaluation runs are retrieved from gs://guidance-evals/ using gcloud storage rsync to the local harness/results/ directory.
The downloaded content consists of JSON evaluation metadata used for local analysis scripts.
[PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it processes untrusted markdown content from the guides/ directory.
Ingestion points: The extractTaskPrompt and extractGuideDescription functions in investigate.ts read raw markdown content from the filesystem.
Boundary markers: The SKILL.md instructions define a CRITICAL CONSTRAINT and explicit workflow rules that prevent the agent from modifying the source evaluation files, enforcing a read-only diagnostic posture.
Capability inventory: The skill has access to the filesystem and CLI tools (gcloud, gh) for report generation and synchronization.
Sanitization: Content is extracted and embedded into markdown reports for human review; the processed data is not executed as shell commands or code.

nightly-eval-investigation