The Agent Skills Directory

[DATA_EXPOSURE]: The skill defines a protocol for managing sensitive data, specifically within references/authentication.md. It provides guidance for storing session material and access tokens in .tigerkit/secret-input/ using restrictive file permissions (0700/0600) and ensuring these paths are ignored by Git. This is documented as a security best practice for local credential management.
[COMMAND_EXECUTION]: The instructions involve the use of standard Git commands (git ls-files, git check-ignore) to verify the environment and ensure that the .tigerkit/ directory is properly excluded from version control. Additionally, it references the execution of a local utility script, scripts/durable_artifacts.py, for managing artifact status.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection as it processes user-provided data from input.json and task-related snapshots in .tigerkit/handoff.md. However, it includes significant mitigations, such as strict JSON schema validation, requiring explicit verification of facts, and instructing the agent to ignore instructions embedded within the data it processes.
[DYNAMIC_EXECUTION]: The skill describes the use of a background watcher to monitor changes to input files. While this involves a persistent process during the input phase, it is characterized as a finite, run-owned task intended to facilitate user interaction rather than an unauthorized persistence mechanism.

tk-handoff