thumbgate-feedback

Agent governance prevention rules block specific tool calls based on prior negative outcomes by applying YAML-defined triggers with negative constraints. They load at session start so the agent starts safer, preventing repeated mistakes without manual intervention.

Yes, explicit thumbs up/down user feedback is captured with context, failure or success details, and tags to create a structured audit trail. This feedback-to-memory capture workflow converts quality signals into lesson-like entries for future reference.

Define YAML triggers with negative trigger constraints to specify which actions to block based on captured failure context. The prevention rule auto-generation updates these rules from repeated negative outcomes and their associated context to enforce guardrails.

Yes, it applies to agent sessions in tools like Claude Code by pulling prior feedback summaries and prevention rules into the session start context. This ensures the agent starts safer by loading established guardrails before executing new tool calls.

Prevention rules require explicit thumbs down or worked/failed signals to trigger capture, meaning implicit dissatisfaction is ignored. Additionally, rules depend on YAML-defined triggers with negative constraints, so improperly configured constraints may fail to block erroneous tool calls.