The Agent Skills Directory

[COMMAND_EXECUTION]: The skill workflow requires the agent to execute various system commands to establish feedback loops. This includes using curl or HTTP scripts for local server testing, running CLI tools for diffing snapshots, and executing test runners like Playwright or Puppeteer.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external, potentially untrusted data to identify bug root causes.
Ingestion points: The agent processes network traces, event logs, HAR files, and user-provided error messages in SKILL.md (Phase 1 and 2).
Boundary markers: There are no specific delimiters used for external data ingestion beyond standard markdown blocks.
Capability inventory: The skill possesses capabilities for shell execution (bash), network access via curl, and file system interaction.
Sanitization: The skill contains a mandatory 'Desensitization' (脱敏) phase that explicitly instructs the agent to redact keys and credentials (replacing them with <REDACTED>) before displaying or utilizing captured data.
[DYNAMIC_EXECUTION]: The skill utilizes the generation and execution of local scripts to facilitate debugging.
Evidence: The agent is instructed to modify a provided template (scripts/hitl-loop.template.sh) to gather user input and to create 'one-off scaffolds' (minimal system subsets) to exercise code paths. These generated scripts are then executed via bash.

diagnosing-bugs