Code, tests and
operational records

// Extracted from ClaudeCode Octo on 9 October 2026; implementation unchanged.
export interface AskGateInput {
  /** Session has a live turn RIGHT NOW per the engine trackers
   *  (activeStreams / activeRemote / activeCodex) — ground truth, not the gate's
   *  own bookkeeping, so a stale gate entry can never wedge a session. */
  isLive: boolean
  /** askId of the registered in-flight ask for this session, if any. */
  inFlightAskId?: string | null
  /** There IS a registered in-flight completion to join. */
  hasInFlight: boolean
  /** askId of the most recently COMPLETED ask for this session, if any. */
  recentAskId?: string | null
  /** When that recent ask completed (ms epoch). */
  recentAt?: number
  /** The incoming ask's client-generated id (absent on legacy clients and
   *  engine-internal dispatches). */
  askId?: string | null
  now: number
}

export type AskGateDecision = 'run' | 'join' | 'replay' | 'inject'

/** How long a completed askId still swallows a late transport retry. Chromium
 *  retries fire immediately on connection failure, so minutes of slack is ample;
 *  10 min also covers a sleep/wake gap. */
export const RECENT_ASK_TTL_MS = 10 * 60_000

export function decideAsk(i: AskGateInput): AskGateDecision {
  if (i.isLive) {
    // A turn is genuinely running. The ONLY ask that may attach to it is its own
    // transport-level retry (identical askId). Everything else — different
    // askId, or no askId at all — is a different message and must be steered in,
    // never run as a second concurrent query.
    if (i.askId && i.hasInFlight && i.inFlightAskId === i.askId) return 'join'
    return 'inject'
  }
  // No live turn. A matching recently-completed askId means this is a late
  // retry of a turn that already ran to completion — do NOT run it again.
  if (
    i.askId &&
    i.recentAskId === i.askId &&
    typeof i.recentAt === 'number' &&
    i.now - i.recentAt < RECENT_ASK_TTL_MS
  ) {
    return 'replay'
  }
  return 'run'
}

// Extracted from ClaudeCode Octo on 9 October 2026; implementation unchanged.
const EXTERNAL_PROTOCOLS = new Set(['http:', 'https:', 'mailto:'])

/**
 * Resolve an anchor/navigation target to a URL that may be handed to the
 * operating system. Relative links are resolved against the current document,
 * which matters for the served thin-client UI: `/docs` is still a web link and
 * must not replace the Electron app.
 */
export function resolveExternalUrl(href: string, baseUrl?: string): string | null {
  try {
    const url = baseUrl ? new URL(href, baseUrl) : new URL(href)
    return EXTERNAL_PROTOCOLS.has(url.protocol.toLowerCase()) ? url.href : null
  } catch {
    return null
  }
}

/**
 * Electron emits will-navigate for renderer-initiated top-level loads, but not
 * for BrowserWindow.loadURL/loadFile. The only renderer navigation we retain is
 * an exact reload of the current document (needed by dev/HMR and normal reload).
 */
export function isExactDocumentReload(targetUrl: string, currentUrl: string): boolean {
  try {
    return new URL(targetUrl).href === new URL(currentUrl).href
  } catch {
    return targetUrl === currentUrl && targetUrl.length > 0
  }
}

Window: 8 September–7 October 2026 UTC. Aggregate queries ran against live PostgreSQL on 9 October in a repeatable-read, read-only transaction. Only aggregate results were exported

Of 3,889 turns marked completed, 3,529 retained a nonempty final-answer record. The completed total includes 380 turns with a stop request. A further 58 turns failed and 7 were interrupted. These are distinct operational states, not customer acceptance labels

The provisional cohort excludes configured operator accounts and the earlier test-like email filter. It is not a verified paying-customer roster and can contain staff-assisted work. Retention and deletion affect coverage. Source semantics were reviewed at snapshot 3b91d462f; historical deployed versions across the window were not reconstructed

For claims specifically about agent accuracy, time saved or cost per completed task: a stable task ID linking all attempts, model/tool charges and final artifacts; an explicit accepted/rejected/abandoned outcome; recorded interventions and review time; and a predefined acceptance rubric. A prospective comparison should retain failed attempts and separate active human time from wall-clock time. This would extend the engineering evidence; it is not a prerequisite for presenting the existing customer-sales results