The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a mechanism where current session context, task IDs, implementation decisions, and assumptions are written to disk and later consumed by a resume skill. This creates a feedback loop where data from one session could influence the behavior of the agent in a subsequent session.
Ingestion points: Data is collected from the active session environment and written to agents/memory/RESUME_POINTER.md, AUDIT_LOG.md, and SESSION_SNAPSHOT/*.md files.
Boundary markers: The skill uses Markdown headers and lists as structural delimiters, but lacks explicit instructions for escaping or quoting content to prevent it from being interpreted as instructions during the resume process.
Capability inventory: The skill requires file-write capabilities within the agents/memory/ and project/ directories to maintain state.
Sanitization: While the skill explicitly forbids the inclusion of secrets (API keys, PII) in the state files, it does not include measures to sanitize natural language instructions that might be embedded in the persisted task data.

state-preserve