The Agent Skills Directory

[DYNAMIC_EXECUTION]: The instructions in references/diagnosis/diagnosing-bugs.md lead the agent to dynamically create and execute reproduction harnesses, such as shell scripts and browser automation scripts (Playwright/Puppeteer), at runtime.
[EXTERNAL_DOWNLOADS]: The skill manifest (SOURCE-MANIFEST.json) identifies an external repository as its upstream source. Additionally, references/explaining-bugs.md requires the installation of the pele package via NPM for visual diagram generation.
[COMMAND_EXECUTION]: The skill utilizes extensive shell command functionality for diagnostic tasks, including curl operations, human-in-the-loop bash scripts, and automated test runners.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes untrusted external artifacts like network traces, logs, and HAR files. Combined with the agent's ability to execute commands and install packages, this introduces a surface where malicious content in analyzed files could influence agent execution.
Ingestion points: Files like HAR traces, event logs, and core dumps processed during Phase 1 and Phase 2 of the diagnosis loop.
Boundary markers: No specific prompt delimiters are defined to isolate untrusted content from instructions, though secret redaction is mandated.
Capability inventory: Includes shell access, NPM package installation, and browser automation.
Sanitization: Instructions specify redacting credentials using <REDACTED> but do not provide for the sanitization of embedded commands or prompt instructions in data files.

atlas-debug