The Agent Skills Directory

[COMMAND_EXECUTION]: The skill spawns background processes using the claude -p command via a local helper script (plugins/development-harness/skills/kage-bunshin/scripts/spawn.py). It also executes several local shell scripts to manage the git environment.
[DYNAMIC_EXECUTION]: It implements a recursive AI pattern where the primary orchestrator spawns and monitors independent sub-sessions. These sub-sessions are launched with the --permission-mode auto flag, enabling them to execute tools autonomously without manual user approval.
[INDIRECT_PROMPT_INJECTION]: A "Discovery Relay" mechanism (Step 6d) facilitates the flow of data between waves of agents. Data generated by one agent is directly injected into the prompt of the next. Ingestion points: The orchestrator reads outcome data from JSON files located in /tmp/kb-work-{issue}.json (Step 6b). Boundary markers: The relay document uses standard markdown headers to structure the injected data but does not appear to use robust delimiters or instructions to prevent the sub-agents from following malicious instructions potentially embedded in task notes. Capability inventory: Each spawned sub-session is a full orchestrator with access to the Agent tool, shell execution, and all configured MCP servers. Sanitization: The skill lacks evidence of sanitization for the NOTES or COMMITS fields before they are interpolated into subsequent agent prompts.
[PRIVILEGE_ESCALATION]: The workflow automates sensitive git operations, including merging branches and pushing changes to the main repository branch, potentially bypassing manual oversight via the autonomous sub-agents.

work-milestone