planning-with-files

Pass

Audited by Gen Agent Trust Hub on May 7, 2026

Risk Level: SAFEDATA_EXFILTRATIONPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DATA_EXFILTRATION]: The scripts/session-catchup.py script accesses internal agent session history files located in ~/.claude/projects/ and ~/.codex/sessions/. This allows the agent to restore context from previous interactions by reading session interaction logs from the user's home directory. While intended for local context recovery, these logs contain a record of user prompts and agent responses.
  • [COMMAND_EXECUTION]: The skill requires the agent to execute several bundled local scripts (init-session.sh, check-complete.sh, session-catchup.py) to manage planning files and recover session context within the local environment.
  • [PROMPT_INJECTION]: The core workflow relies on the agent reading and following goals defined in project files (task_plan.md, findings.md). This creates a surface for indirect prompt injection.
  • Ingestion points: The agent is instructed to read from project planning files and the output of the session recovery script.
  • Boundary markers: The templates do not include explicit delimiters or instructions to ignore embedded commands within the planning files.
  • Capability inventory: The skill supports file modification and the execution of bundled shell and Python scripts.
  • Sanitization: Content from these files is directly interpolated into the agent's context without sanitization or validation.
Audit Metadata
Risk Level
SAFE
Analyzed
May 7, 2026, 04:33 AM