The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The orchestrator ingests untrusted project goals from the .agent/goal.md file and interpolates them into continuation prompts during autonomous loops.
Ingestion points: The scripts/long_task.py helper script reads objective content from .agent/goal.md to provide context for turn-based continuation.
Boundary markers: The ingested content is enclosed in <objective> tags, and the agent is explicitly instructed to treat the block as context rather than higher-priority instructions to mitigate override attempts.
Capability inventory: The skill coordinates subagents via the Agent tool (with worktree isolation), modifies local project files, and executes shell commands.
Sanitization: The objective text is truncated to a 1500-character limit to prevent context window saturation.
[COMMAND_EXECUTION]: The skill relies on a bundled Python script (scripts/long_task.py) to handle lifecycle management and project state operations.
Evidence: The orchestrator invokes this script for tasks such as setting phases, checking status, and installing a 'Stop' hook in the agent platform's local configuration (settings.json). The script uses secure practices like shlex.quote to prevent command injection when building its internal execution strings.

long-task