context-transfer

Pass

Audited by Gen Agent Trust Hub on Nov 8, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill aggregates project metadata, file paths, and task history into a checkpoint file which includes a natural language 'Resume prompt' intended for the agent in future sessions. Malicious instructions placed in repository files or session logs could be included in these checkpoints and influence the agent's logic during resumption.
  • Ingestion points: scripts/context_transfer.py reads Git status, diffs, and untracked file information.
  • Boundary markers: The checkpoint format utilizes JSON frontmatter and clear Markdown headers (e.g., ## Resume prompt). SKILL.md instructs the agent to treat these checkpoints as data rather than authoritative instructions.
  • Capability inventory: The skill can execute local Git commands and read/write checkpoint files in the user's home directory.
  • Sanitization: The skill implements SECRET_PATTERNS regex checks in scripts/context_transfer.py to identify and reject common secret formats (e.g., GitHub tokens, AWS keys, private keys) before they are written to a checkpoint.
  • [COMMAND_EXECUTION]: The skill executes local shell commands to interact with the Git version control system and to query status from sibling agent components.
  • Evidence: The _git function in scripts/context_transfer.py uses subprocess.run to call the git binary with arguments like rev-parse, status, and worktree. The capture_delegated function executes a sibling Python script (cross_agent.py) using sys.executable.
  • [DATA_EXFILTRATION]: While the skill does not perform network operations, it collects extensive telemetry about the local development environment, including file digests, Git history, and environment-specific session IDs.
  • Evidence: scripts/context_transfer.py collects comprehensive repository facts in the _project and _related_git_facts functions, including branch names, HEAD commits, and lists of untracked or modified files.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 8, 2026, 05:53 PM