@oar-sh/oar 0.9.17 → 0.9.18

package/server/services/claude-billing-switch-service.mjs

@@ -0,0 +1,343 @@
+'use strict';
+
+/**
+ * The Claude billing switch: a Claude chat moves between the host's Claude
+ * subscription (`claude`) and the API key saved in Settings → Providers →
+ * Claude API (`claude-api`), and keeps its Claude Code session: both run the
+ * same CLI on the same `~/.claude`.
+ *
+ * The chat's billing is `runtime_sessions.provider_type`. A switch, in one
+ * transaction:
+ * - checks the target: the provider usable (a key for the API, the Claude
+ *   provider on for the subscription), the monthly cap, the chat's model in
+ *   the target's list, and the history rule of decision 23;
+ * - writes the billing, when and by whom;
+ * - re-stamps the chat's rows that have not started (`queue.provider`), so
+ *   they run on the new billing; a row already running finishes on its own;
+ * - to the API key, releases a turn held at the subscription's usage limit;
+ * - leaves one note in the transcript (a second switch before the next turn
+ *   rewrites it);
+ * - and, for an agent's `send`, enqueues its message (`enqueue`).
+ *
+ * The worker swap itself is the delivery's (claude-billing-swap-service.mjs):
+ * the next row of the new billing is not handed to a worker of the old kind,
+ * which is replaced once it is idle.
+ */
+
+import { randomUUID } from 'node:crypto';
+
+import { claudeBaseModelId, isClaudeLongContextModelId } from '../../shared/model-id.mjs';
+
+export const CLAUDE_BILLING_PROVIDERS = Object.freeze(['claude', 'claude-api']);
+export const CLAUDE_BILLING_NOTE_KIND = 'billing-switch';
+export const CLAUDE_BILLING_SWITCHED_BY = Object.freeze(['user', 'agent', 'usage-limit']);
+
+// Decision 23: whether a chat with turns on the other billing may move to
+// this one. Set from the WP3 live check (a subscription session's history on
+// the key, and back); 'blocked' disables the switch for such chats.
+export const CLAUDE_BILLING_HISTORY_RULE = Object.freeze({
+  'claude-api': 'allowed',
+  claude: 'allowed',
+});
+
+export const CLAUDE_BILLING_LABELS = Object.freeze({
+  claude: 'Subscription',
+  'claude-api': 'API key',
+});
+
+export function normalizeClaudeBilling(value) {
+  const billing = String(value || '').trim().toLowerCase();
+  return CLAUDE_BILLING_PROVIDERS.includes(billing) ? billing : '';
+}
+
+export function otherClaudeBilling(billing) {
+  return normalizeClaudeBilling(billing) === 'claude' ? 'claude-api' : 'claude';
+}
+
+/** The transcript note for a switch. */
+export function buildClaudeBillingNoteText({ to, by = 'user' } = {}) {
+  const who = by === 'agent' ? ' (switched by an agent)' : (by === 'usage-limit' ? ' (switched at the usage limit)' : '');
+  if (normalizeClaudeBilling(to) === 'claude-api') {
+    return `💳 Billing: this chat runs on the Claude API key from the next turn${who}. `
+      + 'The first turn on the key has no prompt cache and costs more.';
+  }
+  return `💳 Billing: this chat runs on the Claude subscription from the next turn${who}.`;
+}
+
+function refusal(status, code, error, extra = {}) {
+  return { ok: false, status, code, error, ...extra };
+}
+
+function modelListFor(settings) {
+  return new Set([
+    String(settings?.model || '').trim(),
+    ...(Array.isArray(settings?.availableModels) ? settings.availableModels : []),
+    ...(Array.isArray(settings?.models) ? settings.models : []),
+  ].map((value) => String(value || '').trim().toLowerCase()).filter(Boolean));
+}
+
+export function createClaudeBillingSwitchService({
+  db,
+  stmts,
+  emit = () => {},
+  getClaudeProviderSettings = () => null,
+  getClaudeApiProviderSettings = () => null,
+  // The monthly cap: `{ reached, error? }`, from the runtime's
+  // getClaudeApiCapState seam (the claim check asks the same one; WP5 fills
+  // it in). Not reached by default.
+  getClaudeApiCapState = () => ({ reached: false }),
+  historyRule = CLAUDE_BILLING_HISTORY_RULE,
+  // Releases a turn held at the subscription's usage limit
+  // (usage-limit-pause-service.mjs); runs inside the switch's transaction.
+  releaseUsageLimitPause = null,
+  // The runtime row → what the worker swap waits for (claude-billing-swap-
+  // service.mjs, pendingFor), or null.
+  getPending = () => null,
+  onQueueChanged = () => {},
+  now = () => Date.now(),
+  logger = console,
+} = {}) {
+  if (!db || !stmts) throw new Error('createClaudeBillingSwitchService requires db and stmts');
+
+  const runtimeColumns = new Set(db.prepare(`PRAGMA table_info(runtime_sessions)`).all().map((column) => column.name));
+  const queueColumns = new Set(db.prepare(`PRAGMA table_info(queue)`).all().map((column) => column.name));
+  const messageColumns = new Set(db.prepare(`PRAGMA table_info(messages)`).all().map((column) => column.name));
+  const recordsSwitch = runtimeColumns.has('billing_switched_at') && runtimeColumns.has('billing_switched_by');
+
+  const getRuntime = db.prepare(`SELECT * FROM runtime_sessions WHERE conversation_id = ? LIMIT 1`);
+  const setBilling = recordsSwitch
+    ? db.prepare(`
+        UPDATE runtime_sessions
+        SET provider_type = ?, billing_switched_at = ?, billing_switched_by = ?, last_used_at = ?
+        WHERE id = ?
+      `)
+    : db.prepare(`UPDATE runtime_sessions SET provider_type = ?, last_used_at = ? WHERE id = ?`);
+  const setModel = db.prepare(`UPDATE runtime_sessions SET provider_model = ?, model = ? WHERE id = ?`);
+  const restampRows = queueColumns.has('provider')
+    ? db.prepare(`UPDATE queue SET provider = ? WHERE conversation_id = ? AND status IN ('pending', 'parked')`)
+    : null;
+  const countUnstartedRows = db.prepare(`
+    SELECT COUNT(*) AS n FROM queue WHERE conversation_id = ? AND status IN ('pending', 'parked')
+  `);
+  // The turns that ran on a billing other than `?`: replies with a recorded
+  // provider. A reply from before it was recorded (or a relay note) names
+  // none and does not count: an old chat is judged by the turns it has run
+  // since, and a history the API still refuses meets the turn's own refusal.
+  const countOtherBillingTurns = messageColumns.has('executed_provider')
+    ? db.prepare(`
+        SELECT COUNT(*) AS n FROM messages
+        WHERE conversation_id = ? AND role = 'assistant'
+          AND executed_provider IS NOT NULL AND LOWER(executed_provider) != ?
+      `)
+    : null;
+  const latestMessage = messageColumns.has('kind')
+    ? db.prepare(`SELECT id, role, kind, timestamp FROM messages WHERE conversation_id = ? ORDER BY timestamp DESC, rowid DESC LIMIT 1`)
+    : null;
+  const rewriteNote = db.prepare(`UPDATE messages SET text = ?, timestamp = ? WHERE id = ?`);
+
+  function readSettings(read) {
+    try {
+      const settings = read();
+      return settings && typeof settings === 'object' ? settings : {};
+    } catch {
+      return {};
+    }
+  }
+
+  function chatModel(runtime) {
+    return String(runtime?.provider_model || runtime?.model || '').trim();
+  }
+
+  /**
+   * Whether the chat may move to `to` now: `{ ok: true, runtime, from, to,
+   * model }`, or a refusal `{ ok: false, status, code, error }`. `model` is a
+   * model asked for with the switch, checked in place of the chat's own.
+   */
+  function checkSwitch({ conversationId, to, model = '' } = {}) {
+    const convId = String(conversationId || '').trim();
+    const target = normalizeClaudeBilling(to);
+    if (!target) return refusal(400, 'CLAUDE_BILLING_INVALID', 'Billing must be "claude" (subscription) or "claude-api" (API key)');
+    const runtime = convId ? (getRuntime.get(convId) || null) : null;
+    if (!runtime) return refusal(404, 'CONVERSATION_NOT_FOUND', 'Conversation not found');
+    const from = normalizeClaudeBilling(runtime.provider_type);
+    if (!from) {
+      return refusal(409, 'CLAUDE_BILLING_NOT_CLAUDE', 'Only a Claude chat can switch between the subscription and the API key', {
+        provider: String(runtime.provider_type || '').trim().toLowerCase() || 'github',
+      });
+    }
+    if (from === target) return { ok: true, runtime, from, to: target, model: chatModel(runtime), changed: false };
+    const settings = target === 'claude-api'
+      ? readSettings(getClaudeApiProviderSettings)
+      : readSettings(getClaudeProviderSettings);
+    if (target === 'claude-api' && settings.enabled !== true) {
+      return refusal(409, 'CLAUDE_API_NOT_CONFIGURED', settings.configured === true
+        ? 'Claude API is switched off. Switch it on in Settings → Providers → Claude API.'
+        : 'Claude API has no key. Save one in Settings → Providers → Claude API.');
+    }
+    if (target === 'claude' && settings.enabled !== true) {
+      return refusal(409, 'CLAUDE_NOT_CONFIGURED', 'The Claude provider is switched off. Switch it on in Settings → Providers → Claude.');
+    }
+    if (target === 'claude-api') {
+      let cap = null;
+      // A cap that cannot be read is taken as reached: the key's bill is at stake.
+      try { cap = getClaudeApiCapState() || null; } catch { cap = { reached: true }; }
+      if (cap?.reached === true) {
+        return refusal(409, 'CLAUDE_API_CAP_REACHED', String(cap.error || '').trim()
+          || "This month's Claude API spend has reached the cap set in Settings → Providers → Claude API.");
+      }
+    }
+    const wanted = String(model || '').trim() || chatModel(runtime);
+    if (wanted) {
+      const list = modelListFor(settings);
+      const exact = list.has(wanted.toLowerCase());
+      // A `[1m]` model needs its long-context variant listed; a plain id is
+      // also taken by its base.
+      const listed = exact || (!isClaudeLongContextModelId(wanted) && list.has(claudeBaseModelId(wanted).toLowerCase()));
+      if (!listed) {
+        return refusal(409, 'CLAUDE_BILLING_MODEL_UNAVAILABLE',
+          `${wanted} is not in the ${target === 'claude-api' ? 'Claude API' : 'Claude'} model list. Pick a model it has, then switch.`,
+          { model: wanted });
+      }
+    }
+    if (String(historyRule?.[target] || 'allowed') === 'blocked' && countOtherBillingTurns) {
+      const turns = Number(countOtherBillingTurns.get(runtime.conversation_id, target)?.n || 0);
+      if (turns > 0) {
+        return refusal(409, 'CLAUDE_BILLING_HISTORY', target === 'claude-api'
+          ? 'This chat already has subscription turns, and the API refuses that history. Start a new Claude API chat.'
+          : 'This chat already has API key turns, and the subscription refuses that history. Start a new Claude chat.');
+      }
+    }
+    return { ok: true, runtime, from, to: target, model: wanted, changed: true };
+  }
+
+  /** What the composer's toggle and the usage-limit note read about a chat's billing. */
+  function billingState(conversationId) {
+    const runtime = getRuntime.get(String(conversationId || '').trim()) || null;
+    const provider = normalizeClaudeBilling(runtime?.provider_type);
+    if (!provider) return null;
+    const to = otherClaudeBilling(provider);
+    const check = checkSwitch({ conversationId, to });
+    let pending = null;
+    try { pending = getPending(runtime) || null; } catch { pending = null; }
+    return {
+      provider,
+      pending,
+      switchedAt: runtime?.billing_switched_at || null,
+      switchedBy: runtime?.billing_switched_by || null,
+      switchable: {
+        to,
+        allowed: check.ok === true,
+        code: check.ok ? null : check.code,
+        reason: check.ok ? null : check.error,
+      },
+    };
+  }
+
+  /**
+   * Switch the chat's billing. `enqueue({ runtime, billing })` runs inside
+   * the transaction after the switch (an agent's message), also when the
+   * chat already has that billing; a throw undoes the switch. A switch to
+   * the key releases a turn held at the subscription's usage limit.
+   */
+  function switchBilling({ conversationId, to, by = 'user', model = '', enqueue = null, deferAnnounce = false } = {}) {
+    const who = CLAUDE_BILLING_SWITCHED_BY.includes(by) ? by : 'user';
+    let outcome = null;
+    let note = null;
+    let released = null;
+    let enqueued = null;
+    const run = db.transaction(() => {
+      const check = checkSwitch({ conversationId, to, model });
+      if (!check.ok) {
+        outcome = check;
+        return;
+      }
+      const at = new Date(now()).toISOString();
+      const { runtime } = check;
+      if (check.changed) {
+        if (recordsSwitch) setBilling.run(check.to, at, who, at, runtime.id);
+        else setBilling.run(check.to, at, runtime.id);
+        if (String(model || '').trim() && check.model !== chatModel(runtime)) setModel.run(check.model, check.model, runtime.id);
+        restampRows?.run(check.to, runtime.conversation_id);
+        note = writeNote(runtime.conversation_id, check.to, who, at);
+        if (check.to === 'claude-api' && typeof releaseUsageLimitPause === 'function') {
+          released = releaseUsageLimitPause({ conversationId: runtime.conversation_id, by: 'billing-switch' }) || null;
+        }
+      }
+      if (typeof enqueue === 'function') enqueued = enqueue({ runtime, billing: check.to }) ?? null;
+      outcome = {
+        ok: true,
+        changed: check.changed,
+        conversationId: runtime.conversation_id,
+        provider: check.to,
+        from: check.from,
+        at,
+        queuedRows: Number(countUnstartedRows.get(runtime.conversation_id)?.n || 0),
+      };
+    });
+    run();
+    if (!outcome?.ok) return outcome;
+    const result = {
+      ...outcome,
+      noteMessageId: note?.messageId || null,
+      released: released ? { messageId: released.messageId || null } : null,
+      enqueued,
+    };
+    // Called inside a caller's transaction (the send route's persist step),
+    // the switch is not committed yet: the caller announces it once it is.
+    if (deferAnnounce) return { ...result, announce: () => announce(outcome, note, released, who) };
+    announce(outcome, note, released, who);
+    return result;
+  }
+
+  // What the clients and the queue hear of a committed switch.
+  function announce(outcome, note, released, who) {
+    try { released?.announce?.(); } catch {}
+    if (note) {
+      try {
+        emit(note.created ? 'assistant_message' : 'message_updated', {
+          conversationId: outcome.conversationId,
+          messageId: note.messageId,
+          message: {
+            id: note.messageId,
+            role: 'assistant',
+            text: note.text,
+            kind: CLAUDE_BILLING_NOTE_KIND,
+            model: null,
+            mode: null,
+            attachments: [],
+            timestamp: note.timestamp,
+          },
+        });
+      } catch {}
+    }
+    if (outcome.changed) {
+      try {
+        logger.log?.(`[claude-billing] conv=${outcome.conversationId.slice(0, 8)} ${outcome.from} → ${outcome.provider} by ${who}; ${outcome.queuedRows} queued row(s) re-stamped`);
+      } catch {}
+      try { emit('conversation_billing', { conversationId: outcome.conversationId, billing: billingState(outcome.conversationId) }); } catch {}
+      try { onQueueChanged('claude-billing-switch'); } catch {}
+    }
+  }
+
+  // One note per run of switches: one written since the chat's last message
+  // is rewritten in place, so a switch there and back leaves one line.
+  function writeNote(conversationId, to, by, at) {
+    const text = buildClaudeBillingNoteText({ to, by });
+    const latest = latestMessage?.get(conversationId) || null;
+    if (latest && latest.role === 'assistant' && latest.kind === CLAUDE_BILLING_NOTE_KIND) {
+      rewriteNote.run(text, at, latest.id);
+      return { messageId: latest.id, text, timestamp: at, created: false };
+    }
+    const messageId = randomUUID();
+    stmts.insertMsg.run(messageId, conversationId, 'assistant', text, null, null, null, at, null, null, null);
+    stmts.setMessageKind?.run?.(CLAUDE_BILLING_NOTE_KIND, messageId);
+    stmts.updateConvTime?.run?.(at, conversationId);
+    return { messageId, text, timestamp: at, created: true };
+  }
+
+  return {
+    checkSwitch,
+    billingState,
+    switchBilling,
+  };
+}