Garry's Opinionated OpenClaw/Hermes Agent Brain
TypeScript
/**
* gbrain feedback — inspect and reset use-attributed retrieval feedback.
*
* Usage:
* gbrain feedback status [--json]
* gbrain feedback reset [--source ID] [--page SLUG]
*
* Zero LLM calls. Ratings come from `gbrain rate` / the rate_answer op.
*/
import type { BrainEngine } from '../core/engine.ts';
import { setCliExitVerdict } from '../core/cli-force-exit.ts';
import { loadFeedbackSettings } from '../core/feedback/settings.ts';
import { droppedFeedbackEvents } from '../core/feedback/record.ts';
import { feedbackStatus, resetRetrievalWeights } from '../core/feedback/store.ts';
import { feedbackStageErrors } from '../core/search/feedback-boost.ts';
const HELP = `Usage: gbrain feedback <status|reset> [options]
Inspect and reset the brain's use-attributed retrieval feedback (ratings from
\`gbrain rate\` / rate_answer and think/synthesize citations).
status [--json] counts, weight spread, lowest-weighted pages with a next step
reset [--source ID] [--page SLUG] return learned weights to neutral (all, one source, or one page)
Guide: docs/guides/retrieval-feedback.md
`;
function flag(args: string[], name: string): string | undefined {
const i = args.indexOf(name);
return i >= 0 ? args[i + 1] : undefined;
}
export async function runFeedback(engine: BrainEngine, args: string[]): Promise<void> {
const sub = args[0];
if (!sub || sub === '--help' || sub === '-h' || args.includes('--help')) {
process.stdout.write(HELP);
return;
}
if (sub === 'reset') {
const sourceId = flag(args, '--source');
const slug = flag(args, '--page');
const removed = await resetRetrievalWeights(engine, { sourceId, slug });
const scope = slug ? `page ${slug}` : sourceId ? `source ${sourceId}` : 'the whole brain';
process.stdout.write(`Reset ${removed} learned weight${removed === 1 ? '' : 's'} to neutral for ${scope}.\n`);
return;
}
if (sub !== 'status') {
process.stderr.write(`Unknown subcommand: ${sub}\n\n${HELP}`);
setCliExitVerdict(2);
return;
}
const [settings, status] = await Promise.all([loadFeedbackSettings(engine), feedbackStatus(engine)]);
const report = {
settings,
...status,
dropped_events_this_process: droppedFeedbackEvents(),
stage_errors_this_process: feedbackStageErrors(),
};
if (args.includes('--json')) {
process.stdout.write(JSON.stringify(report, null, 2) + '\n');
return;
}
const lines = [
`Retrieval feedback: ${settings.enabled ? 'on' : 'off'}${settings.enabled && !settings.learn ? ' (not learning)' : ''}, influence λ=${settings.influence}, implicit=${settings.implicit ? 'on' : 'off'}`,
`Answers recorded (last ${settings.eventRetentionDays} days): ${status.events}`,
`Ratings: ${status.ratings_explicit} explicit, ${status.ratings_cited} from citations${status.last_rating_at ? ` (last ${status.last_rating_at.toISOString()})` : ''}`,
`Learned weights: ${status.weights_total} (${status.weights_off_neutral} away from neutral)`,
];
if (status.top_client_share_7d !== null && status.top_client_share_7d > 0.8 && status.ratings_explicit > 10) {
lines.push(`Warning: one client made ${Math.round(status.top_client_share_7d * 100)}% of ratings in the last 7 days. If that is not expected, run \`gbrain feedback reset\` and review that client's write grant.`);
}
if (status.highest.length > 0) {
lines.push('', 'Most trusted pages:');
for (const h of status.highest) lines.push(` ${h.weight.toFixed(3)} ${h.source_id}:${h.slug}`);
}
if (status.lowest.length > 0) {
lines.push('', 'Least trusted pages (open each and check whether it is wrong, stale or simply off-topic for those questions):');
for (const l of status.lowest) {
lines.push(` ${l.weight.toFixed(3)} ${l.source_id}:${l.slug} rated low on ${l.low_ratings} answer${l.low_ratings === 1 ? '' : 's'}`);
}
}
if (status.events === 0) {
lines.push('', 'No answers recorded yet. Answers from query/search/think/synthesize/recall carry an answer_id; rate one with `gbrain rate <answer_id> 1-5`.');
}
process.stdout.write(lines.join('\n') + '\n');
}