planning-with-files-de
Pass
Audited by Gen Agent Trust Hub on Nov 11, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses a set of shell, PowerShell, and Python scripts to automate context injection during agent lifecycle events. These scripts manage local state, calculate file hashes, and update plan status based on tool usage. This includes the
session-catchup.pyscript which can access local project metadata from the~/.claude/projects/directory when explicitly invoked by the user to restore session context. - [INDIRECT_PROMPT_INJECTION]: The skill reads user-modifiable files (
task_plan.md,progress.md) and injects them into the model's prompt. To mitigate injection risks, it implements high-fidelity framing with cryptographically derived nonces and provides an attestation tool (attest-plan.sh) to lock plans against unauthorized modifications. - Ingestion points: Planning files (
task_plan.md,progress.md,findings.md) read byscripts/inject-plan.shandscripts/inject-plan.py. - Boundary markers: Present. Context is injected within nonce-framed blocks (e.g.,
===BEGIN-PWF-DATA kind=plan nonce=...===). - Capability inventory: The skill utilizes file reading/writing and command execution tools to maintain its persistent state. Script-based automation is used across various lifecycle hooks.
- Sanitization: Implemented through SHA-256 attestation verification and secure data framing.
- [DYNAMIC_EXECUTION]: The skill employs dynamic techniques for efficiency and safety. The PowerShell scripts (e.g.,
attest-plan.ps1) useAdd-Typeto compile C# code at runtime for secure file operations that avoid symlink attacks by using Win32 handles. Additionally, shell scripts use Python heredocs to run complex logic in a single process, reducing overhead and enabling robust JSON handling while ensuring the project directory is excluded from the Python search path for security.
Audit Metadata