The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local command-line tools using subprocess.run to provide interactive features. It utilizes fzf for a terminal-based session picker and calls the claude CLI (claude -p) to distill conversation transcripts into concise handover briefs. The command arguments are passed as a list, preventing shell injection.
[DATA_EXPOSURE]: As part of its core functionality, the skill accesses and modifies Claude Desktop's local session data. It reads session metadata JSON files from the user's AppData directory and transcript JSONL files from the ~/.claude/projects directory. While this involves sensitive conversational data, the operations are performed locally and facilitate cross-account session management.
[INDIRECT_PROMPT_INJECTION]: The session recovery feature processes historical transcript data through an LLM, which presents a surface for indirect prompt injection. Ingestion points: Local *.jsonl transcript files in scripts/summon.py. Boundary markers: The distillation process uses the _DISTILL_INSTRUCTION template to guide the LLM's summarization task. Capability inventory: File read/write access and the ability to execute fzf and claude CLI tools. Sanitization: The _text_only function in scripts/summon.py sanitizes the input by extracting only conversational text and deliberately skipping tool_use and tool_result blocks.
[DYNAMIC_EXECUTION]: The widget command assembles an HTML interactive picker by injecting session metadata into the assets/picker-widget.html template. The session data is injected as a JSON block via string replacement and does not involve unsafe evaluation or execution of arbitrary code.

summon