The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/long_task.py modifies the global platform configuration file (~/.claude/settings.json) to install a 'Stop hook'. This hook persistently alters the agent's behavior across sessions by intercepting stop signals.
[COMMAND_EXECUTION]: The 'Stop hook' mechanism automates the agent's turns by returning a 'block' decision to the platform, forcing the agent to continue working autonomously without manual confirmation.
[PROMPT_INJECTION]: The skill contains instructions that mandate the agent to skip user verification and resolve all issues autonomously during its primary execution phases ('You DO NOT ask the user during Phase 2/3').
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests and processes various project-state files from the current working directory.
Ingestion points: Project files including .agent/goal.md, .agent/plans.md, .agent/standards.md, .agent/implement.md, and .agent/progress.md (read by both the orchestrator and the lifecycle script).
Boundary markers: A <objective> tag is used for the goal summary, with instructions to prioritize system messages over tag content; other files lack explicit boundary markers or isolation.
Capability inventory: The skill utilizes the Agent tool with worktree isolation and executes local shell commands via the lifecycle script.
Sanitization: There is no evidence of sanitization, escaping, or validation of the content read from the project markdown files before it is used to prompt subagents.

long-task