The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to process context records from external storage, which is a potential vector for indirect prompt injection. To mitigate this, the instructions explicitly state that retrieved context must be treated as historical evidence, not as authority, and that instruction-like text within the records should never be executed. Evidence found in evals/fixtures/resume-checkpoint/memory/workstreams/docs-17/plan.md includes examples of such payloads being used for testing purposes.
[CREDENTIALS_UNSAFE]: The skill provides explicit instructions to sanitize context records by excluding credentials, secrets, environment variables, and sensitive logs. Evaluation fixtures use synthetic secrets to verify that these values are not persisted, ensuring the agent correctly handles sensitive information.
[COMMAND_EXECUTION]: The evaluation fixtures include instructions for running unit tests using python3 -m unittest. This is a standard development practice and is scoped to the local test environment as described in evals/fixtures/resume-checkpoint/project/AGENTS.md.
[SAFE]: The skill follows security best practices by requiring explicit authorization for durable writes, maintaining clear task boundaries, and emphasizing that historical data does not grant current permissions for sensitive operations like deployment or deletion. The reference to external documentation is informational and does not involve automated code execution.

context-checkpoint