The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core functionality relies on reading state and instructions from 'handoff' files written by previous sessions, creating a vector for instruction persistence.
Ingestion points: Instructions in SKILL.md (Resuming section) direct the agent to read the newest file in _agent/handoff/<GOAL>/.
Boundary markers: The skill does not define clear boundaries or 'ignore' instructions for the content of the handoff files, instructing the agent to follow the ## Next section directly.
Capability inventory: The skill utilizes shell execution for Git operations, file system management, and process spawning via the orca CLI.
Sanitization: There is no sanitization or validation of the content within the handoff files before the agent processes them.
[COMMAND_EXECUTION]: The skill heavily utilizes shell commands to manage the handoff environment.
It executes git commands (e.g., git branch, git rev-parse, git status) to derive goal names and repository paths.
It performs file system operations like mkdir -p and echo >> to manage the _agent/ directory and .git/info/exclude file.
[DYNAMIC_EXECUTION]: The skill dynamically assembles shell commands to launch subsequent agent sessions.
It constructs launch commands for various agents (claude, codex, grok, etc.) by interpolating prompt strings into shell templates.
It uses the orca CLI (e.g., orca terminal create --command "<launch command>") to execute these assembled strings, which involves nesting natural language prompts within shell quotes.
[DATA_EXPOSURE]: The skill accesses and references sensitive local environment data.
It instructs the agent to search for and link to local session logs in ~/.claude/projects/ to maintain a transcript history across sessions.

handoff