Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS
JavaScript ai-agentsdashboarddeveloper-toolsdigital-twinquantified-selfsecond-brainself-hostedtailscale
import { useCallback, useState } from 'react';
import { Scale } from 'lucide-react';
import Pill from '../ui/Pill';
import BrailleSpinner from '../BrailleSpinner';
import { scopeAdherenceReasonLabel } from '../../lib/scopeAdherenceReasons';
import { useInstanceFeatures } from '../../hooks/useInstanceFeatures';
import { scoreAppScopeAdherence } from '../../services/api';
/**
* An on-demand, ADVISORY answer to "does this change advance what the product
* says it is for?", rendered under one issue or pull-request row.
*
* Three deliberate constraints:
*
* 1. **Nothing is gated on it.** No button is disabled by a verdict, no row is
* hidden, no claim or merge is blocked. The output is a clause the reader
* can go and argue with.
* 2. **It runs only when clicked.** The scorer is a 9 GB local model that
* loads on first use; scoring every visible row on mount would be exactly
* the cold-bootstrap work the AI Provider Usage Policy forbids.
* 3. **It disappears when the feature is off.** An install that has not opted
* into the jev scorer never sees a button whose only outcome would be
* "not installed".
*/
const TONE = { aligned: 'success', contradicts: 'warning', unrelated: 'muted' };
const LABEL = { aligned: 'Advances', contradicts: 'Works against', unrelated: 'Unrelated' };
export default function ScopeAdherenceCheck({ appId, kind, title, body = '', diffSummary = '' }) {
const { isFeatureEnabled } = useInstanceFeatures();
const [result, setResult] = useState(null);
const [scoring, setScoring] = useState(false);
const check = useCallback(() => {
setScoring(true);
setResult(null);
scoreAppScopeAdherence(appId, { kind, title, body, diffSummary })
.then((response) => setResult(response || { ok: false }))
// The tab owns its error UI and an uninstalled scorer is an ordinary
// answer here, so a failed request reads as "no advisory", never a toast.
.catch(() => setResult({ ok: false }))
.finally(() => setScoring(false));
}, [appId, kind, title, body, diffSummary]);
if (!isFeatureEnabled('jev')) return null;
const verdict = result?.ok ? result.verdict : null;
return (
<div className="mt-2 flex flex-wrap items-center gap-2 text-xs">
<button
type="button"
onClick={check}
disabled={scoring}
className="inline-flex items-center gap-1 rounded border border-port-border px-2 py-0.5 text-gray-400 hover:text-port-accent hover:border-port-accent/40 disabled:opacity-50"
title="Advisory only — scores this change against the repository's PRD.md and GOALS.md locally"
>
{scoring ? <BrailleSpinner /> : <Scale size={12} />}
{scoring ? 'Scoring…' : 'Check scope'}
</button>
{verdict && verdict !== 'abstained' && (
<>
<Pill tone={TONE[verdict]} size="xs">{LABEL[verdict]}</Pill>
{/* The clause, not just the score. A bare margin is noise. The
citation is composed server-side so its spelling has one home. */}
<span className="text-gray-400">{result.clause?.citation}</span>
<span className="text-gray-600 italic">advisory only</span>
</>
)}
{verdict === 'abstained' && (
<span className="text-gray-500">No advisory — the scorer could not separate the options.</span>
)}
{result && !result.ok && <span className="text-gray-500">{scopeAdherenceReasonLabel(result.code)}</span>}
</div>
);
}