The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The README.md documentation includes instructions to install the uv tool by piping a script from astral.sh to a shell (curl -LsSf https://astral.sh/uv/install.sh | sh). astral.sh is the official domain for Astral, the developers of the uv and ruff tools.
[DATA_EXFILTRATION]: The skill accesses local directories containing sensitive conversation histories, such as ~/.claude/projects/, ~/.codex/sessions/, and the Cursor editor's storage directory. This access is the primary intended function of the skill, which acts as a unified reader for these histories.
[COMMAND_EXECUTION]: The agent-conv script executes external binaries including claude and codex to support thread forking and porting functionality. It uses os.execvp and subprocess.run to transition execution to these tools.
[INDIRECT_PROMPT_INJECTION]: The skill ingests historical conversation data that is later used to seed new agent sessions. This presents a potential attack surface where instructions found within the processed history could be interpreted by the target agent.
Ingestion points: Conversation logs are read from local filesystem paths associated with agent installations (e.g., SKILL.md, bin/agent-conv).
Boundary markers: When porting history between providers, the content is delimited using triple-dash --- markers to separate context from instructions.
Capability inventory: The tool has the ability to spawn new shells and execute agent binaries using history as context.
Sanitization: The script implements logic to strip internal system tags, thinking blocks, and synthetic turns from the conversation stream during display.

agent-conv-cli