The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill relies on external 'PLAN.md' files which contain instructions and executable commands that are parsed and run by the agent.
Ingestion points: The PLAN.md file serves as the primary input for the build, run, and next operations, where it is parsed for task metadata and logic.
Boundary markers: None. The skill extracts commands based on standard Markdown bold patterns (e.g., **Verify:**, **QC:**) without explicit delimiters to separate untrusted content from the execution engine.
Capability inventory: The skill executes arbitrary shell commands via the orca-graph.py script, manages a background daemon process, performs file system operations (writes to events.jsonl and project graphs), and executes Git commands (git checkout).
Sanitization: There is no explicit sanitization or validation of the command strings extracted from the plan file before they are passed to the shell.
[COMMAND_EXECUTION]: The skill provides a mechanism to execute arbitrary shell commands provided as arguments to the run command or defined within the PLAN.md file under **Verify:** and **QC:** headers. This is a core feature of the orchestrator but represents a direct command execution sink.
[PERSISTENCE]: The watch and run commands implement a persistence mechanism by spawning a background daemon process. This daemon tracks project states across the machine and maintains a global registry file at ~/.orca-graph/registry.json to coordinate task execution across different sessions.
[DYNAMIC_EXECUTION]: The skill dynamically generates and executes shell commands at runtime based on the task graph's state. This includes the automated execution of verification scripts and the optional **QC:** (Quality Control) review commands that run after a task is completed.

orca-graph