| name | iteration-recording |
| description | How and when to persist a rule attempt to the experiment record (metrics.csv and experiment.json). |
Iteration Recording
Use this skill after every attack execution within the current variant cycle. The orchestrator owns the experiment ID and the variant label; you only pass the per-attempt details.
Tool
record_iteration(
iteration: int,
attack_id: str,
arguments: list[str],
fired: bool,
evasion_rationale: str,
rule: str,
container_exit_code: int | None = None,
container_stderr: str = "",
) -> "recorded"
When to Call
- Once after every call to
trigger_attacker + check_alert_fired, regardless of outcome.
- Pass
iteration as the attempt number within the current variant cycle (start at 1, increment for each rule regeneration attempt).
fired must reflect the result of check_alert_fired.rule must be the rule that was deployed for this attempt (post-assign_sid).- Copy
container_exit_code and container_stderr from the trigger_attacker return value.
Why It Matters
The orchestrator can crash, the SSH session can drop, the LLM can run out of budget. Every recorded iteration is a row in metrics.csv and an entry in experiment.json — partial records are still useful for analysis. Always record before returning.
