delegating-with-context
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill transmits conversation transcript segments to external endpoints (api.cloudflare.com and api.typesafe.ai) to identify context relevant to a specific task.
- [COMMAND_EXECUTION]: The skill uses a hook to execute local Python scripts whenever an agent delegation or session creation tool is used. Evidence: Defined in hooks/hooks.json to execute scripts/context_hook.py.
- [INDIRECT_PROMPT_INJECTION]: The skill injects historical session data into subagent prompts, which may include untrusted content from previous tool outputs or user messages.
- Ingestion points: Reads transcript data from ~/.claude/projects/*.jsonl via scripts/chunking.py.
- Boundary markers: scripts/context_hook.py wraps the injected context in clear headers that distinguish reference data from primary instructions.
- Capability inventory: The skill does not execute code from the transcript; it only appends the text to a prompt for another agent to process.
- Sanitization: Includes proactive secret scrubbing to remove credentials from the ingested context.
Audit Metadata