The Agent Skills Directory

[SAFE]: The skill is designed for technical planning and grounding of software tasks. It explicitly instructs the agent to perform read-only investigations ('Default to read-only investigation and safe non-mutating checks') and prohibits writing code or specifications directly. The only permitted file write is to a dedicated .workflow/ directory to store the planning brief.
[SAFE]: Analysis of the evaluation files, such as evals/files/browser-boundary/input/src/TransportConfig.php, reveals the presence of tokens ('client_example', 'svc_live_example'). These are explicitly identified as 'fake values' and 'neutral names' in the accompanying documentation (evals/sources/browser-boundary-observed.txt) and are used solely for demonstrating security boundary logic during agent evaluation.
[COMMAND_EXECUTION]: The skill instructions mention 'reference search plus diff, value, and behavior checks'. These are standard development tool operations (like grep or git diff) used for legitimate repository investigation within the scoped task.
[DATA_EXPOSURE]: The skill provides instructions for identifying sensitive data boundaries (e.g., ensuring server credentials do not reach the browser). This is a defensive security planning activity, not an exfiltration attempt. The process is confined to the local repository context provided to the agent.

task-groundwork