The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes task definitions (.hamster/.../tasks/*.md) and specification files (brief.md, spec.md) which serve as the primary instructions for the subagents. This creates a surface where malicious content in these files could influence agent behavior.
Ingestion points: Found in references/brief-selection.md and references/agents/task-executor.md, where the skill reads local markdown files to determine execution logic.
Boundary markers: The skill provides clear hierarchy instructions, telling agents to treat brief.md and spec.md as authoritative over individual task files to mitigate contradictions.
Capability inventory: Includes file system modification, shell command execution (git, hamster CLI, gh CLI), and the ability to launch further subagents.
Sanitization: Relies on a "Plan Feedback Protocol" and manual user confirmation checkpoints before significant actions.
[COMMAND_EXECUTION]: The skill performs extensive shell operations to manage the development environment and reconstruct state.
System Inspection: Uses pgrep and lsof to identify running synchronization processes and inspect the /proc filesystem to determine working directories.
Build Tooling: Automatically detects and executes project-specific commands like npm run, cargo check, go build, and make during the validation phase of the execution loop.
Repository Management: Uses git for status checking and committing, and the GitHub CLI (gh) for pull request creation and modification.
[DYNAMIC_EXECUTION]: The skill utilizes dynamic script execution and subagent orchestration for its core logic.
Embedded Scripts: Executes a Python script via a heredoc in references/brief-selection.md to parse JSON state and resolve account information.
Subagent Orchestration: Dynamically launches specialized agents (task-executor and wave-reviewer) by injecting markdown instruction files from its reference directory into the agent's context.
[PRIVILEGE_ESCALATION]: The instructions in SKILL.md advise users to run readiness checks with elevated permissions on macOS if the default workspace sandbox cannot access the system Keychain, which involves running shell commands with higher privileges.

resume-hamster