Record and organize evaluation trajectories across iterations with best-candidate tracking.
npx skills add https://github.com/2389-research/simmer --skill simmer-reflect
Please help me install this Agent Skill. Skill: simmer-reflect Source: https://github.com/2389-research/simmer/tree/main/skills/simmer-reflect Command: npx skills add https://github.com/2389-research/simmer --skill simmer-reflect
Reflect records the full iteration history, trajectory table, and best candidate, enabling regression rollback and ASI forwarding in simmer rounds.
After each judge round, instruct Reflect to update the trajectory and propagate the ASI to the next round.
High-intent search queries and answers about installing and using this skill.
Tracking iteration scores and identifying the best candidate involves recording the full evaluation trajectory history, preserving score changes across rounds to lock in the highest-performing output for subsequent generation.
ASI forwarding passes the judge's evaluation scores from the current round to the next generator, using the preserved trajectory history to ensure subsequent iterations build upon the best-so-far candidate.
Handling score regression involves signaling when to revert to a prior iteration by using the recorded trajectory table, allowing you to roll back to a previous best candidate and pass its context forward instead of continuing with degraded results.
Trajectory tracking supports both single-file and workspace modes, enabling you to organize evaluation histories and manage best-so-far candidates whether your iteration process operates on individual files or across a full workspace environment.
You should update the iteration trajectory table after each judge round completes, recording the latest scores and changes so the best candidate is identified, regression is detected, and the ASI is propagated to guide the next generator.
2389-researchScore candidate artifacts against user-defined criteria and generate an ASI.
2389-researchGenerate improved artifact versions from ASI feedback in the simmer loop.
fockusCoordinate three judge agents in iterative debates to produce consensus evaluations.
zebbernEnumerate AWS IAM misconfigurations and privilege escalation paths.
