The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill configures all dynamically created sub-agents with the mode: "bypassPermissions" parameter in both SKILL.md and references/agents.md. This setting explicitly instructs the platform to skip user confirmation prompts for all tools executed by these agents, effectively removing human-in-the-loop oversight for filesystem modifications and command execution.
[INDIRECT_PROMPT_INJECTION]: The skill scouts the codebase and interpolates file content directly into the prompts of execution agents without sanitization or robust boundary markers, making it vulnerable to malicious instructions embedded in the project files.
Ingestion points: The scout agents in Phase 2 and the context injection in Phase 4 utilize Glob, Grep, and Read tools to ingest arbitrary content from the project directory into the agent's context.
Boundary markers: The prompt templates in references/prompt-templates.md use standard Markdown headers (e.g., ## Context, ## Goal) but lack specific instructions to ignore or isolate instructions found within the ingested codebase data.
Capability inventory: Sub-agents are granted broad capabilities including Bash, Write, and Edit tools, and they are instructed to run with suppressed permissions, which would cause them to follow any injected instructions autonomously.
Sanitization: There is no evidence of sanitization, escaping, or validation of external content before it is interpolated into agent prompts.
[COMMAND_EXECUTION]: The workflow relies heavily on the Bash tool for both codebase exploration and the execution of tasks across the project. These commands are executed without individual user approval when the bypassPermissions mode is active.
[DATA_EXFILTRATION]: While not directly malicious, the skill provides instructions in references/enable-agent-teams.md to modify sensitive configuration files like ~/.claude/settings.json and shell profile files (.bashrc, .zshrc). This guides the user to expose internal environment settings to the experimental agent features.

team-assemble