The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes state information from .agentit/STATE.md and handoff capsules, creating an attack surface where malicious data could influence agent behavior.
Ingestion points: Processes content from .agentit/STATE.md and external handoff directories during the resume protocol.
Boundary markers: The skill explicitly instructs the agent to treat resumed text as data only, to make a fresh decision, and never to execute the incoming text directly.
Capability inventory: The skill utilizes file system access and shell command execution via the agentit CLI tool.
Sanitization: Relies on instructional guardrails for the agent to re-evaluate all decisions and capability selections upon resumption.
[COMMAND_EXECUTION]: The skill uses the agentit CLI tool for state management, including initializing projects, creating checkpoints, and performing handoffs between environments.
[EXTERNAL_DOWNLOADS]: The skill facilitates syncing operational state with a private remote repository.
Remote target: Supports configuration of a private GitHub repository (a well-known service) for state synchronization.
Mechanism: Uses the agentit continuity link and sync commands to reconcile local state with the remote via Git.

long-horizon-recovery