The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs the benchflow package using the uv tool. This package is the primary utility for the skill's trajectory management functions and is provided by the author of the skill.
[DATA_EXFILTRATION]: The skill accesses and uploads AI session logs from various application directories, specifically ~/.claude/, ~/.codex/, ~/.cursor/, and ~/.local/share/opencode/. To mitigate the risk of unauthorized data transfer, it utilizes a local viewer for human review, performs dry-runs to identify secrets for redaction, and requires an explicit confirmation signal from the user before initiating the upload to the BenchFlow service.
[COMMAND_EXECUTION]: The skill uses the Bash tool to execute uv for installation and several bench CLI commands to list, view, and upload trajectory data from the local machine.
[PROMPT_INJECTION]: The skill contains instructions to suppress the display of internal URLs in chat to protect user privacy. Additionally, it processes local session logs which creates an indirect prompt injection surface.
Ingestion points: Local session log files (.jsonl) and SQLite databases scanned during the discovery step in SKILL.md.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are used when the agent extracts and displays snippets from these logs.
Capability inventory: The skill possesses Bash tool access and the ability to execute arbitrary CLI commands via the shell.
Sanitization: No sanitization or filtering logic is defined for the text snippets extracted from the session logs before they are processed by the agent.

benchflow-traj-upload