The Agent Skills Directory

[COMMAND_EXECUTION]: The orchestrator session executes multiple shell commands to manage the git lifecycle, including creating branches, worktrees, and squashing merges. It also interacts with the file system to create and update the 'overnight-plan.md' state file.
[REMOTE_CODE_EXECUTION]: The skill spawns sub-agent sessions using the 'claude -w' command with the '--permission-mode auto' flag. This configuration allows the worker sessions to execute commands, write files, and perform actions autonomously without requiring human approval for each tool call.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8). It reads project-level configuration files like 'CLAUDE.md' and the codebase to generate an execution plan. Malicious content in these files could influence the plan generation, leading the orchestrator to schedule and execute harmful sub-tasks.
Ingestion points: The orchestrator reads 'CLAUDE.md', 'README.md', and key source files to build context for decomposition (SKILL.md, Phase 1).
Boundary markers: No explicit delimiters or sanitization logic is used to isolate untrusted data when generating the task list.
Capability inventory: The system can spawn worker sessions with full tool permissions ('--permission-mode auto') and execute arbitrary shell commands defined as 'completion' checks.
Sanitization: No sanitization or validation of the generated task instructions or completion commands is performed before execution.
[DYNAMIC_EXECUTION]: The orchestrator dynamically assembles shell commands and prompts for the worker sessions based on the content of 'overnight-plan.md', which is generated at runtime.

overnight