The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads session history and configuration from local files to determine its state, creating a surface for instructions embedded in data to influence the agent.
Ingestion points: Reads the last 20 lines of .autoresearch/sessions/<session-id>/run.jsonl, .autoresearch/sessions/<session-id>/state.md, and research/learnings/<session-id>.md (documented in SKILL.md).
Boundary markers: None. Instructions do not specify delimiters or warnings to ignore data-embedded commands when reading history.
Capability inventory: Ability to execute arbitrary shell commands via bash for benchmarks and pre-flight checks.
Sanitization: Content is read and used directly for state management without validation or filtering.
[DYNAMIC_EXECUTION]: The skill generates shell scripts (benchmark.sh and checks.sh) from predefined templates and executes them using bash at runtime (found in references/experiment-protocol.md).
[COMMAND_EXECUTION]: The core loop executes shell commands through the session's benchmark wrapper (bash "$SESSION_DIR/benchmark.sh") to measure experiment results (found in SKILL.md).

autoresearch