distill-feedback

Session feedback can be turned into durable agent rules by processing JSONL feedback logs to extract standing preferences and converting them into atomic configuration rules. This requires semantic analysis to filter out one-off requests from permanent corrections.

Semantic feedback extraction uses LLM-based analysis to process session transcripts, identifying durable user corrections while filtering out one-off requests or chatter. It evaluates feedback queues to determine which corrections represent standing preferences suitable for rule generation.

Automatic application without manual review is not supported. The system requires human-gated approval for all rule updates to ensure accuracy and prevent configuration poisoning, meaning a user must validate proposed rules before they are applied.

The best way to update agent configuration is using delta-merge application, which safely updates rule sets through atomic edits. This prevents full-file rewrites and maintains project-specific memory when applying newly approved user corrections.

Processing session feedback queues requires JSONL feedback logs as input. There are no external dependencies needed, as the system uses LLM-based semantic analysis to extract durable corrections from the provided JSONL session transcripts.

Human-in-the-loop validation is necessary to ensure accuracy and prevent configuration poisoning. It requires user approval of all proposed rules before they are applied, guaranteeing that only correct and durable corrections modify the agent configuration.