git clone https://github.com/AI45Lab/SentrySkills.git ~/SentrySkills
cd ~/SentrySkills
python install/install.pyRestart Claude Code after installation.
Claude Code is the best fit for the new version because hooks can enforce the rule-first frontend:
base_rule -> extra_rule -> rule_gate
Then the framework performs risk assessment and may choose:
sync model_stageasync/subagent model_stage
Key rules:
rule_stage_action == blockstops immediately- subagents are only allowed after rule gating
- only low-risk turns may use async/subagent execution
- knowledge writeback only happens after completed
model_stage - the main framework agent should run one proposal sweep at task end
- proposal sweep only affects subsequent turns
base_rule_actionextra_rule_actionrule_stage_actionframework_risk_levelmodel_dispatch_modemodel_stage_statusmodel_stage_actionmodel_executormodel_stage_result_availableproposal_sweep_effectknowledge_writeback_statusfinal_action
Claude Code runs against the current workspace, so runtime state stays local:
.sentryskills/base/.sentryskills/extra/