generalcontext 0.2.0 → 0.3.0

package/dist/index.js

@@ -12,9 +12,11 @@
 // the background. Raw fetch capture is opt-in with instrumentFetch: true. Your agent never waits on General Context, and a failure on our side is
 // logged and dropped, never thrown. No runtime dependencies.
 import { AsyncLocalStorage } from 'node:async_hooks';
-import { createHmac, randomBytes, randomUUID } from 'node:crypto';
+import { createHash, createHmac, randomBytes, randomUUID } from 'node:crypto';
+import { request as httpRequest } from 'node:http';
+import { request as httpsRequest } from 'node:https';
 import { createRequire } from 'node:module';
-export const VERSION = '0.2.0';
+export const VERSION = '0.3.0';
 const BATCH = 20;
 const MAX_QUEUE = 5000;
 const MAX_CALL_BYTES = 4_000_000; // the server rejects larger calls
@@ -23,6 +25,8 @@ const MAX_BATCH_BYTES = 4_000_000;
 const SEND_TIMEOUT_MS = 15_000;
 const DIRECT_TIMEOUT_MS = 10_000; // reportOutcome and logEvent
 const EXIT_FLUSH_MS = 2_000;
+const FLUSH_TIMEOUT_MS = 10_000; // flush() default, like the Python SDK
+const RETRY_PAUSE_MS = 30_000; // after the collector fails to answer, wait this long before sending again
 const DEFAULT_BASE_URL = 'https://generalcontextlabs.com';
 const config = {
     apiKey: undefined,
@@ -41,6 +45,10 @@ let timer = null;
 let originalFetch = globalThis.fetch;
 let exitHooked = false;
 let stopped = false; // set when the exit flush runs out of time
+let pausedUntil = 0; // background sends wait until then after a failed send; flush() still tries
+// Prepared calls the collector refused for now (busy, failing, unreachable). Sent before the queue.
+const retry = [];
+let retryBytes = 0;
 const inflight = new Set();
 const apiKey = () => config.apiKey ?? process.env.GENERALCONTEXT_API_KEY;
 // Explicit option, then env var, then default. Trailing slashes are trimmed without a regex.
@@ -90,13 +98,58 @@ export function initialize(options = {}) {
     if (!secureUrl(baseUrl()))
         warn(`baseUrl ${baseUrl()} is not https; nothing will be sent`);
     const modules = options.instrumentModules ?? autoload();
+    // openai, @anthropic-ai/sdk and @google/genai ship separate CommonJS and ES module builds with their own classes.
+    // require() above found the CommonJS ones, which is what a CommonJS app uses. An ES module app (import OpenAI
+    // from 'openai') gets the other build, so patch that too. patchMethod skips a method already patched.
+    // The imports settle a few ticks after initialize() returns, before an app's own `await import(...)` of the
+    // same module resumes. A call made synchronously right after initialize(), before any await, can run
+    // unrecorded: await ready() first in that case.
+    // OpenAI Realtime (voice) connection classes live in their own entry points. openai/realtime/ws needs the
+    // optional `ws` package, so either may be missing.
+    if (!options.instrumentModules || options.instrumentModules.openAI) {
+        const require = createRequire(import.meta.url);
+        for (const name of REALTIME_MODULES) {
+            try {
+                patchRealtime(require(name));
+            }
+            catch { /* not installed */ }
+        }
+        const realtime = Promise.all(REALTIME_MODULES.map(name => import(name).then(patchRealtime, () => { }))).then(() => { });
+        pending = pending.then(() => realtime);
+    }
+    if (!options.instrumentModules) {
+        const esm = [['openai', patchOpenAI], ['@anthropic-ai/sdk', patchAnthropic]];
+        if (modules.googleGenAI)
+            esm.push(['@google/genai', patchGemini]);
+        // @aws-sdk/client-bedrock-runtime resolves to one class from both require() and import(), so it needs nothing here.
+        let found = Object.values(modules).some(Boolean);
+        const realtime = pending;
+        pending = Promise.all(esm.map(([name, patch]) => import(name).then(m => { found = true; patch(m); }, () => { })))
+            .then(() => { if (!found)
+            warn(NO_SDK); }).then(() => realtime);
+    }
     if (modules.openAI)
         patchOpenAI(modules.openAI);
     if (modules.anthropic)
         patchAnthropic(modules.anthropic);
+    if (modules.googleGenAI)
+        patchGemini(modules.googleGenAI);
+    if (modules.bedrockRuntime)
+        patchBedrock(modules.bedrockRuntime);
     if (options.instrumentFetch ?? false)
         patchFetch();
 }
+let pending = Promise.resolve();
+// Resolution is from this file's location, so an SDK installed in another project's node_modules is not found.
+const NO_SDK = 'found no model SDK to record (openai, @anthropic-ai/sdk, @google/genai, @aws-sdk/client-bedrock-runtime). '
+    + 'Install generalcontext in the same project as your model SDK, or pass instrumentModules.';
+/**
+ * Resolves once initialize() has patched every SDK build it found, including the ES module builds it loads
+ * asynchronously. Await it when a model call may run before the app's next await after initialize(). Never rejects.
+ */
+export function ready() {
+    return pending;
+}
 // Load installed SDKs when the caller did not pass them. Bundlers can hide these; pass instrumentModules there.
 function autoload() {
     const require = createRequire(import.meta.url);
@@ -106,7 +159,10 @@ function autoload() {
     catch {
         return undefined;
     } };
-    return { openAI: load('openai'), anthropic: load('@anthropic-ai/sdk') };
+    return {
+        openAI: load('openai'), anthropic: load('@anthropic-ai/sdk'),
+        googleGenAI: load('@google/genai'), bedrockRuntime: load('@aws-sdk/client-bedrock-runtime'),
+    };
 }
 // Accept a module namespace, its default export, or the client class itself.
 function clientClass(value, name) {
@@ -124,6 +180,45 @@ function patchAnthropic(value) {
     const Anthropic = clientClass(value, 'Anthropic');
     // messages.stream() calls messages.create({ stream: true }) internally, so it is covered too.
     patchMethod(Anthropic?.Messages?.prototype, 'create', 'anthropic');
+    // client.beta.messages: create, parse, stream, toolRunner and threads all post through beta.messages.create.
+    patchMethod(Anthropic?.Beta?.Messages?.prototype, 'create', 'anthropic');
+}
+const exported = (value, name) => {
+    const v = value;
+    return v?.[name] ?? v?.default?.[name] ?? (typeof v === 'function' && v.name === name ? v : undefined);
+};
+// @google/genai assigns generateContent and generateContentStream as instance properties in the Models
+// constructor. A setter on the prototype sees that assignment and stores the recording version instead.
+function patchGemini(value) {
+    const proto = exported(value, 'Models')?.prototype;
+    if (!proto)
+        return;
+    const wrappers = {
+        generateContent: fn => recording(fn, 'gemini'),
+        generateContentStream: fn => geminiStreamRecording(fn),
+    };
+    for (const [name, wrapFn] of Object.entries(wrappers)) {
+        const existing = Object.getOwnPropertyDescriptor(proto, name);
+        if (existing?.set && existing.set[MARK])
+            continue;
+        if (typeof existing?.value === 'function') {
+            patchMethod(proto, name, 'gemini');
+            continue;
+        }
+        const set = function (fn) {
+            const v = typeof fn === 'function' && !fn[MARK] ? wrapFn(fn) : fn;
+            Object.defineProperty(this, name, { value: v, writable: true, configurable: true, enumerable: true });
+        };
+        set[MARK] = true;
+        Object.defineProperty(proto, name, { get: () => undefined, set, configurable: true });
+    }
+}
+function patchBedrock(value) {
+    const proto = exported(value, 'BedrockRuntimeClient')?.prototype;
+    const original = proto?.send;
+    if (typeof original !== 'function' || original[MARK])
+        return;
+    proto.send = bedrockRecording(original);
 }
 const MARK = Symbol.for('generalcontext.recorded');
 function patchMethod(owner, name, api) {
@@ -142,6 +237,9 @@ export function withSession(session, run) {
         tags.channel = s.channel;
     if (s.product)
         tags.product = s.product;
+    const account = s.account == null ? '' : String(s.account).trim().slice(0, 120);
+    if (account)
+        tags.account = account;
     if (s.metadata)
         tags.metadata = s.metadata;
     return scope.run({ ...scope.getStore(), tags }, run);
@@ -158,6 +256,54 @@ export function conversation(tags, run) {
 export function task(name, run) {
     return scope.run({ ...scope.getStore(), task: name }, run);
 }
+const CLIP_OVER = 200_000; // strings longer than this many characters are clipped
+const CLIP_HEAD = 100_000;
+const CLIP_TAIL = 20_000;
+const BINARY_MIN = 256; // shorter base64 payloads are left alone
+const BASE64 = /^[A-Za-z0-9+/_-]+={0,2}$/;
+const DATA_URL = /^data:([^,;]*)((?:;[^,;]*)*);base64,/;
+const omitted = (payload) => `[omitted ${payload.length} bytes sha256:${createHash('sha256').update(payload).digest('hex').slice(0, 12)}]`;
+/**
+ * The copy of a call that is recorded. Never changes `value`, and never touches what the caller sends to the provider.
+ * Agents resend the whole run on every call, so without this, later calls pass the per-call limit and are lost.
+ *
+ * - Inline binary becomes a placeholder that keeps the shape: a base64 data URL keeps its prefix
+ *   (`data:image/png;base64,[omitted N bytes sha256:ab12cd34ef56]`), and a base64 `data`, `result` or
+ *   `b64_json` field (Anthropic `source.data`, Gemini `inlineData.data`, OpenAI `input_audio.data`, Responses
+ *   `image_generation_call.result`) becomes `[omitted N bytes sha256:...]`. N is the length of the base64 text
+ *   and the hash is the first 12 hex digits of the sha256 of that text, so the same image gets the same placeholder
+ *   in every call. Media types stay. Binary is never recorded: it does not help review and it is a privacy risk.
+ * - Any string over 200,000 characters keeps its first 100,000 and last 20,000 characters around
+ *   `\n[... clipped N characters ...]\n`.
+ *
+ * Runs before beforeSend and redaction. A call still over the limit afterwards is dropped.
+ */
+export function compactForRecording(value, key, depth = 0) {
+    if (typeof value === 'string') {
+        if (value.length > BINARY_MIN) {
+            const url = DATA_URL.exec(value.slice(0, 512));
+            if (url)
+                return value.slice(0, url[0].length) + omitted(value.slice(url[0].length));
+            if ((key === 'data' || key === 'result' || key === 'b64_json') && BASE64.test(value))
+                return omitted(value);
+        }
+        if (value.length > CLIP_OVER)
+            return `${value.slice(0, CLIP_HEAD)}\n[... clipped ${value.length - CLIP_HEAD - CLIP_TAIL} characters ...]\n${value.slice(-CLIP_TAIL)}`;
+        return value;
+    }
+    if (value === null || typeof value !== 'object' || depth > 64)
+        return value;
+    if (typeof value.toJSON === 'function')
+        value = value.toJSON();
+    if (value === null || typeof value !== 'object')
+        return compactForRecording(value, key, depth + 1);
+    if (Array.isArray(value))
+        return value.map(item => compactForRecording(item, undefined, depth + 1));
+    const out = {};
+    for (const [k, v] of Object.entries(value))
+        out[k] = compactForRecording(v, k, depth + 1);
+    return out;
+}
 // Runs before the caller's await continues, so only take the snapshot here. Redaction and beforeSend run at send time.
 function record(call, at, scrub) {
     try {
@@ -165,7 +311,7 @@ function record(call, at, scrub) {
             warnLimited('no-key', 'GENERALCONTEXT_API_KEY is not set; a model call was not recorded');
             return;
         }
-        const json = JSON.stringify({ ...call, conversation: at?.tags ?? null, task: at?.task ?? null });
+        const json = JSON.stringify(compactForRecording({ ...call, request: withoutRequestOptions(call.request), conversation: at?.tags ?? null, task: at?.task ?? null }));
         const bytes = Buffer.byteLength(json);
         if (bytes > MAX_CALL_BYTES) {
             warnLimited('oversized', `a model call of ${bytes} bytes is over the ${MAX_CALL_BYTES} byte limit and was dropped`);
@@ -178,6 +324,8 @@ function record(call, at, scrub) {
         queue.push({ json, bytes, scrub });
         queuedBytes += bytes;
         hookExit();
+        if (Date.now() < pausedUntil)
+            return; // the collector just failed; a timer resumes sending
         if (config.disableBatch || queue.length >= BATCH)
             void send();
         else if (!timer) {
@@ -201,6 +349,43 @@ function hookExit() {
         void flush().finally(() => clearTimeout(cap));
     });
 }
+// Credentials are never sent, whatever `redact` says. They reach a recording when an app pastes a key into a
+// prompt, a tool returns one, or a request carries one in its options. Each pattern needs the provider's
+// fixed prefix, so ordinary text is never touched. The key name stays, so a reader knows one was there.
+const SECRETS = [
+    ['anthropic-key', /\bsk-ant-[A-Za-z0-9_-]{20,}/g],
+    ['openai-key', /\bsk-(?:proj-|svcacct-|admin-)?[A-Za-z0-9_-]{20,}/g],
+    ['stripe-key', /\b[rsp]k_(?:live|test)_[A-Za-z0-9]{16,}/g],
+    ['google-key', /\bAIza[A-Za-z0-9_-]{35}\b/g],
+    ['aws-key', /\b(?:AKIA|ASIA)[A-Z0-9]{16}\b/g],
+    ['github-token', /\b(?:gh[pousr]_[A-Za-z0-9]{36,}|github_pat_[A-Za-z0-9_]{40,})/g],
+    ['slack-token', /\bxox[abposr]-[A-Za-z0-9-]{10,}/g],
+    ['generalcontext-key', /\bgck_[a-f0-9]{48}\b/g],
+    ['jwt', /\beyJ[A-Za-z0-9_-]{8,}\.eyJ[A-Za-z0-9_-]{8,}\.[A-Za-z0-9_-]{8,}/g],
+    ['bearer-token', /(?<=\bBearer\s)[A-Za-z0-9._~+/-]{20,}=*/g],
+    // A password inside a connection string or URL: postgres://user:PASSWORD@host. Only the password goes.
+    ['password', /(?<=\b[a-z][a-z0-9+.-]{1,20}:\/\/[^\s:/@]{1,64}:)[^\s@/]{1,128}(?=@)/g],
+    ['private-key', /-----BEGIN [A-Z ]{0,20}PRIVATE KEY-----(?:[A-Za-z0-9+/=\s:,]|-(?!-)){1,16384}?-----END [A-Z ]{0,20}PRIVATE KEY-----/g],
+];
+export function maskSecrets(text) {
+    if (text.length < 16)
+        return text;
+    return SECRETS.reduce((out, [kind, pattern]) => out.replace(pattern, `<${kind}>`), text);
+}
+// Request options that carry credentials or transport settings, not what the model was asked. Dropped from the
+// recorded request: Gemini's httpOptions (per-request headers) and anything spelled like a header map.
+const REQUEST_OPTIONS = new Set(['httpOptions', 'http_options', 'headers', 'extra_headers', 'extraHeaders', 'extra_query', 'extraQuery', 'abortSignal', 'signal']);
+function withoutRequestOptions(value, depth = 0) {
+    if (!value || typeof value !== 'object' || depth > 3)
+        return value;
+    if (Array.isArray(value))
+        return value;
+    const out = {};
+    for (const [k, v] of Object.entries(value))
+        if (!REQUEST_OPTIONS.has(k))
+            out[k] = k === 'config' ? withoutRequestOptions(v, depth + 1) : v;
+    return out;
+}
 // Apply beforeSend and redaction. Returns the JSON to send, or null to drop.
 function prepare(entry) {
     let call = JSON.parse(entry.json);
@@ -215,7 +400,59 @@ function prepare(entry) {
         if (call == null)
             return null;
     }
-    return JSON.stringify(scrubbed(call, entry.scrub === undefined ? config.scrub : entry.scrub));
+    const scrub = entry.scrub === undefined ? config.scrub : entry.scrub;
+    return JSON.stringify(scrubbed(call, scrub ? (text) => scrub(maskSecrets(text)) : maskSecrets));
+}
+// Answers worth retrying: the collector is busy (429) or failing (5xx). Anything else (a bad key, a body too
+// large) will fail again the same way, so that batch is dropped with a warning.
+const retryable = (status) => status === 429 || status === 408 || status >= 500;
+// Retry-After in seconds (or an HTTP date), kept between 1 s and 5 minutes.
+function retryAfterMs(value) {
+    const raw = Array.isArray(value) ? value[0] : value;
+    if (!raw)
+        return undefined;
+    const seconds = /^\d+$/.test(raw.trim()) ? Number(raw) * 1000 : Date.parse(raw) - Date.now();
+    return Number.isFinite(seconds) ? Math.min(300_000, Math.max(1000, seconds)) : undefined;
+}
+// While a caller awaits flush(), reportOutcome() or logEvent(), our sockets keep the process alive
+// so the await can finish. Otherwise they are unref'd: background sending never keeps a finished
+// process alive, even against a collector that accepts and never answers.
+let waiters = 0;
+const sockets = new Set();
+function holdOpen(work) {
+    if (waiters++ === 0)
+        for (const socket of sockets)
+            socket.ref();
+    return work.finally(() => { if (--waiters === 0)
+        for (const socket of sockets)
+            socket.unref(); });
+}
+/**
+ * POST to the collector with node:http(s), not fetch, so the socket can be unref'd (see above).
+ * Redirects are never followed (node:http does not follow them), so the key never reaches another host.
+ */
+function postRaw(url, body, signal) {
+    return new Promise((resolve, reject) => {
+        const target = new URL(url);
+        const send = target.protocol === 'https:' ? httpsRequest : httpRequest;
+        const req = send(target, {
+            method: 'POST', signal, agent: false,
+            headers: { Authorization: `Bearer ${apiKey()}`, 'Content-Type': 'application/json', 'Content-Length': Buffer.byteLength(body) },
+        }, res => {
+            res.resume(); // drain, so the socket is freed
+            res.on('end', () => resolve({ ok: (res.statusCode ?? 0) >= 200 && (res.statusCode ?? 0) < 300, status: res.statusCode ?? 0,
+                retryAfterMs: retryAfterMs(res.headers['retry-after']) }));
+            res.on('error', reject);
+        });
+        req.on('socket', socket => {
+            sockets.add(socket);
+            socket.once('close', () => sockets.delete(socket));
+            if (waiters === 0)
+                socket.unref();
+        });
+        req.on('error', reject);
+        req.end(body);
+    });
 }
 async function post(path, body, timeoutMs) {
     const url = `${baseUrl()}${path}`;
@@ -230,13 +467,7 @@ async function post(path, body, timeoutMs) {
     timeout.unref?.();
     inflight.add(controller);
     try {
-        return await originalFetch(url, {
-            method: 'POST',
-            headers: { Authorization: `Bearer ${apiKey()}`, 'Content-Type': 'application/json' },
-            body,
-            redirect: 'error', // never forward the API key to another host
-            signal: controller.signal,
-        });
+        return await postRaw(url, body, controller.signal);
     }
     catch (error) {
         warnLimited(`send:${path}`, `could not reach ${path}: ${error.message}`);
@@ -252,10 +483,11 @@ function send() {
         return sending;
     sending = (async () => {
         const envelope = JSON.stringify({ sdk: `node/${VERSION}`, app: config.appName });
-        while (queue.length && !stopped) {
-            // Up to BATCH calls and MAX_BATCH_BYTES per request.
-            const parts = [];
-            let size = envelope.length + 16;
+        while ((queue.length || retry.length) && !stopped) {
+            // Up to BATCH calls and MAX_BATCH_BYTES per request. Calls refused earlier go first, already prepared.
+            const parts = retry.splice(0, BATCH);
+            retryBytes = retry.reduce((sum, part) => sum + part.length, 0);
+            let size = envelope.length + 16 + parts.reduce((sum, part) => sum + part.length + 1, 0);
             while (queue.length && parts.length < BATCH) {
                 const entry = queue[0];
                 let json;
@@ -284,24 +516,57 @@ function send() {
                 continue;
             const body = `{"calls":[${parts.join(',')}],${envelope.slice(1)}`;
             const response = await post('/api/platform-collect/sdk', body, SEND_TIMEOUT_MS);
-            if (response && !response.ok)
+            if (response && !response.ok && !retryable(response.status))
                 warnLimited('rejected', `${parts.length} calls not accepted: ${response.status}`);
+            if ((!response || (!response.ok && retryable(response.status))) && !stopped) {
+                // Busy, failing, unreachable, hung or timed out: put this batch back at the front, already prepared, and
+                // try again after a pause. Nothing recorded is lost to a deploy or a brief outage.
+                retry.unshift(...parts);
+                retryBytes += parts.reduce((sum, part) => sum + part.length, 0);
+                if (retryBytes > MAX_QUEUE_BYTES) {
+                    const dropped = retry.splice(BATCH * 10);
+                    retryBytes = retry.reduce((sum, part) => sum + part.length, 0);
+                    warnLimited('queue-full', `${dropped.length} calls dropped while the collector was unavailable`);
+                }
+                const pause = response?.retryAfterMs ?? RETRY_PAUSE_MS;
+                // Unreachable, hung or timed out. Stop here instead of trying every queued batch in turn (15 s each),
+                // which kept the process alive for minutes after the app finished. The queue stays bounded and is
+                // retried after a pause, on flush(), or at exit.
+                pausedUntil = Date.now() + pause;
+                const resume = setTimeout(() => { if (queue.length || retry.length)
+                    void send(); }, pause);
+                resume.unref?.();
+                break;
+            }
         }
     })().catch(() => { }).finally(() => { sending = null; });
     return sending;
 }
-/** Wait until recorded calls are sent. Call before a serverless function or script exits. Never throws. */
-export async function flush() {
+/**
+ * Wait until recorded calls are sent, for at most `timeoutMs` (default 10 s). Call before a serverless
+ * function or script exits. Resolves true when the queue was sent, false at the deadline. Never throws.
+ * Sending continues in the background after the deadline.
+ */
+export async function flush(timeoutMs = FLUSH_TIMEOUT_MS) {
     if (timer) {
         clearTimeout(timer);
         timer = null;
     }
+    let deadline;
     try {
-        await send();
+        const sent = send().then(() => queue.length === 0 && retry.length === 0, () => false);
+        // The deadline timer stays referenced: an awaited flush() must be able to finish and report.
+        const late = new Promise(resolve => { deadline = setTimeout(() => resolve(false), Math.max(0, timeoutMs)); });
+        return await holdOpen(Promise.race([sent, late]));
+    }
+    catch {
+        return false;
+    }
+    finally {
+        clearTimeout(deadline);
     }
-    catch { /* send already logs */ }
 }
-const scrubText = (value) => (config.scrub ? config.scrub(value) : value);
+const scrubText = (value) => (config.scrub ? config.scrub(maskSecrets(value)) : maskSecrets(value));
 /** Mark how a session ended. Returns true when the platform accepted it. Never throws. */
 export async function reportOutcome(sessionId, outcome, note) {
     try {
@@ -311,7 +576,7 @@ export async function reportOutcome(sessionId, outcome, note) {
         }
         await flush();
         const body = { session_id: scrubText(String(sessionId)), outcome, note: note == null ? null : scrubText(note) };
-        const response = await post('/api/platform-collect/outcomes', JSON.stringify(body), DIRECT_TIMEOUT_MS);
+        const response = await holdOpen(post('/api/platform-collect/outcomes', JSON.stringify(body), DIRECT_TIMEOUT_MS));
         return Boolean(response?.ok);
     }
     catch {
@@ -327,12 +592,13 @@ export async function logEvent(event) {
         }
         await flush();
         const timestamp = event.timestamp instanceof Date ? event.timestamp.toISOString() : event.timestamp ?? new Date().toISOString();
+        const sessionId = scrubText(String(event.sessionId));
         const properties = Object.fromEntries(Object.entries(event.properties ?? {}).map(([k, v]) => [k, typeof v === 'string' ? scrubText(v) : v]));
         const body = { events: [{
                     event_id: event.eventId ?? randomUUID(), event_type: scrubText(event.eventType), timestamp,
-                    session_id: scrubText(String(event.sessionId)), properties,
+                    session_id: sessionId, properties,
                 }] };
-        const response = await post('/api/platform-collect/events', JSON.stringify(body), DIRECT_TIMEOUT_MS);
+        const response = await holdOpen(post('/api/platform-collect/events', JSON.stringify(body), DIRECT_TIMEOUT_MS));
         return Boolean(response?.ok);
     }
     catch {
@@ -349,9 +615,45 @@ function recording(create, api, scrub) {
         const reply = scope.run({ ...at, inSdkCall: true }, () => create.apply(this, args));
         if (!reply || typeof reply.then !== 'function')
             return reply;
+        // The SDKs' APIPromise reads the body inside parseResponse. Helpers such as parse() build a new promise that
+        // calls it again, so keep the first result: the body is read once and both sides get the same object.
+        const api_ = reply;
+        if (typeof api_.parseResponse === 'function') {
+            try {
+                const parseOnce = api_.parseResponse;
+                const seen = new WeakMap();
+                const parse = function (...a) {
+                    const key = a[1];
+                    if (!key || typeof key !== 'object')
+                        return parseOnce.apply(this, a);
+                    if (!seen.has(key))
+                        seen.set(key, parseOnce.apply(this, a));
+                    return seen.get(key);
+                };
+                api_.parseResponse = parse;
+                const thenUnwrap = api_._thenUnwrap;
+                if (typeof thenUnwrap === 'function') {
+                    api_._thenUnwrap = function (transform) {
+                        const next = thenUnwrap.call(this, transform);
+                        if (next && typeof next.parseResponse === 'function') {
+                            next.parseResponse = async (client, props) => {
+                                const parsed = await parse.call(reply, client, props);
+                                const out = transform(parsed, props);
+                                const id = parsed?._request_id;
+                                if (id && out && typeof out === 'object' && !('_request_id' in out))
+                                    Object.defineProperty(out, '_request_id', { value: id, enumerable: false });
+                                return out;
+                            };
+                        }
+                        return next;
+                    };
+                }
+            }
+            catch { /* leave the SDK promise as it was */ }
+        }
         reply.then(response => {
-            const done = (result) => record({ api, request, response: result, started_at: started, ended_at: new Date().toISOString(), error: null }, at, scrub);
-            if (request?.stream && isStream(response))
+            const done = (result, error = null) => record({ api, request, response: api === 'gemini' ? plainGemini(result) : result, started_at: started, ended_at: new Date().toISOString(), error }, at, scrub);
+            if (request?.stream && (isStream(response) || isLegacyStream(response)))
                 observeStream(response, api, done);
             else
                 done(response);
@@ -364,13 +666,18 @@ function recording(create, api, scrub) {
     return wrapped;
 }
 const isStream = (value) => typeof value?.iterator === 'function';
+// openai before 4.15 has no `iterator` field: the Stream class only defines [Symbol.asyncIterator].
+const isLegacyStream = (value) => !!value && typeof value === 'object' && typeof value[Symbol.asyncIterator] === 'function';
 // Both SDKs' Stream classes read chunks through `this.iterator`, including tee() and toReadableStream().
 // Replacing it keeps the object, its type and its methods, and lets us see each chunk once.
+// A stream the caller stops early (break, return or a throw inside for await, which aborts the request) or
+// that fails is still recorded, with what was assembled so far and an error. The caller's exception is untouched.
 function observeStream(stream, api, done) {
-    const source = stream.iterator.bind(stream);
-    stream.iterator = async function* () {
+    const legacy = !isStream(stream);
+    const source = legacy ? stream[Symbol.asyncIterator].bind(stream) : stream.iterator.bind(stream);
+    const observed = async function* () {
         const assemble = api === 'anthropic' ? anthropicAssembler() : openAIAssembler();
-        let finished = false;
+        let error = 'stream ended early: the caller stopped reading before the response completed';
         try {
             for await (const chunk of { [Symbol.asyncIterator]: source }) {
                 try {
@@ -379,23 +686,42 @@ function observeStream(stream, api, done) {
                 catch { /* never break the caller's stream */ }
                 yield chunk;
             }
-            finished = true;
+            error = null;
+        }
+        catch (failure) {
+            error = String(failure?.message ?? failure);
+            throw failure;
         }
         finally {
-            if (finished)
-                done(assemble.result());
+            let result = null;
+            try {
+                result = assemble.result();
+            }
+            catch { /* record the error alone */ }
+            done(result, error);
         }
     };
+    // An own property shadows the class method on this instance only.
+    if (legacy)
+        Object.defineProperty(stream, Symbol.asyncIterator, { value: observed, writable: true, configurable: true });
+    else
+        stream.iterator = observed;
 }
 function openAIAssembler() {
-    let model, id, usage = null, final = null;
+    let model, id, usage = null, final = null, partial = null;
+    const items = [];
     const choices = [];
     return {
         add(chunk) {
             // Responses API events: the final response arrives whole.
+            // Keep the latest snapshot and finished output items too, so a stream stopped early still records what arrived.
             if (chunk.type) {
                 if (chunk.type === 'response.completed')
                     final = chunk.response;
+                else if (chunk.response)
+                    partial = chunk.response;
+                else if (chunk.type === 'response.output_item.done' && chunk.item)
+                    items[chunk.output_index ?? items.length] = chunk.item;
                 return;
             }
             model ??= chunk.model;
@@ -423,7 +749,7 @@ function openAIAssembler() {
                     choice.finish_reason = c.finish_reason;
             }
         },
-        result: () => final ?? ({ id, object: 'chat.completion', model, choices: choices.filter(Boolean), usage }),
+        result: () => final ?? (partial ? { ...partial, output: items.filter(Boolean) } : { id, object: 'chat.completion', model, choices: choices.filter(Boolean), usage }),
     };
 }
 function anthropicAssembler() {
@@ -530,6 +856,635 @@ function assembleSSE(text) {
     }
     return assemble.result();
 }
+// generateContentStream resolves to an async generator. Hand the caller a generator that yields the same chunks.
+function geminiStreamRecording(create, scrub) {
+    const wrapped = function (...args) {
+        const started = new Date().toISOString();
+        const at = scope.getStore();
+        const request = args[0];
+        const reply = scope.run({ ...at, inSdkCall: true }, () => create.apply(this, args));
+        if (!reply || typeof reply.then !== 'function')
+            return reply;
+        const fail = (error) => record({ api: 'gemini', request, response: null, started_at: started, ended_at: new Date().toISOString(), error: String(error?.message ?? error) }, at, scrub);
+        return reply.then(source => {
+            if (!source || typeof source[Symbol.asyncIterator] !== 'function')
+                return source;
+            return (async function* () {
+                const assemble = geminiAssembler();
+                let finished = false;
+                try {
+                    for await (const chunk of source) {
+                        try {
+                            assemble.add(chunk);
+                        }
+                        catch { /* never break the caller's stream */ }
+                        yield chunk;
+                    }
+                    finished = true;
+                }
+                catch (error) {
+                    fail(error);
+                    throw error;
+                }
+                finally {
+                    if (finished)
+                        record({ api: 'gemini', request, response: assemble.result(), started_at: started, ended_at: new Date().toISOString(), error: null }, at, scrub);
+                }
+            })();
+        }, error => { fail(error); throw error; });
+    };
+    wrapped[MARK] = true;
+    return wrapped;
+}
+// The response as plain JSON, without the raw HTTP headers the SDK attaches.
+function plainGemini(response) {
+    if (!response || typeof response !== 'object')
+        return response;
+    const { sdkHttpResponse, ...rest } = response;
+    return rest;
+}
+function geminiAssembler() {
+    const out = { candidates: [] };
+    const parts = [];
+    let candidate = null;
+    return {
+        add(chunk) {
+            const c = plainGemini(chunk);
+            if (c.usageMetadata)
+                out.usageMetadata = c.usageMetadata;
+            if (c.modelVersion)
+                out.modelVersion = c.modelVersion;
+            if (c.responseId)
+                out.responseId = c.responseId;
+            const first = c.candidates?.[0];
+            if (!first)
+                return;
+            candidate ??= { content: { role: 'model', parts } };
+            if (first.content?.role)
+                candidate.content.role = first.content.role;
+            if (first.finishReason)
+                candidate.finishReason = first.finishReason;
+            for (const part of first.content?.parts ?? []) {
+                const last = parts[parts.length - 1];
+                if (typeof part.text === 'string' && last && typeof last.text === 'string' && Boolean(last.thought) === Boolean(part.thought) && Object.keys(part).every(k => k === 'text' || k === 'thought'))
+                    last.text += part.text;
+                else
+                    parts.push({ ...part });
+            }
+        },
+        result: () => ({ ...out, candidates: candidate ? [candidate] : [] }),
+    };
+}
+// Images and documents arrive as bytes. Send a marker in their place, never the bytes.
+function omitBytes(value, depth = 0) {
+    if (value instanceof Uint8Array || value instanceof ArrayBuffer || ArrayBuffer.isView(value))
+        return { omitted: 'bytes' };
+    if (depth > 32 || value === null || typeof value !== 'object')
+        return value;
+    if (Array.isArray(value))
+        return value.map(item => omitBytes(item, depth + 1));
+    return Object.fromEntries(Object.entries(value).map(([k, v]) => [k, omitBytes(v, depth + 1)]));
+}
+const CONVERSE = new Set(['ConverseCommand', 'ConverseStreamCommand']);
+function bedrockRecording(send, scrub) {
+    const wrapped = function (...args) {
+        const command = args[0];
+        const name = command?.constructor?.name ?? '';
+        const input = command?.input;
+        // By class name, or by input shape when a bundler renamed the classes. InvokeModel has a body, not messages.
+        const converse = CONVERSE.has(name) || (typeof input?.modelId === 'string' && Array.isArray(input?.messages) && !('body' in input));
+        if (!converse)
+            return send.apply(this, args);
+        const started = new Date().toISOString();
+        const at = scope.getStore();
+        let request = null;
+        try {
+            request = omitBytes(command?.input);
+        }
+        catch { /* record without the request */ }
+        const reply = scope.run({ ...at, inSdkCall: true }, () => send.apply(this, args));
+        if (!reply || typeof reply.then !== 'function')
+            return reply;
+        const done = (response, error) => record({ api: 'bedrock-converse', request, response, started_at: started, ended_at: new Date().toISOString(), error }, at, scrub);
+        reply.then(output => {
+            try {
+                const { $metadata, ...rest } = output ?? {};
+                const stream = rest.stream;
+                if (stream && typeof stream[Symbol.asyncIterator] === 'function') {
+                    output.stream = (async function* () {
+                        const assemble = converseAssembler();
+                        let finished = false;
+                        try {
+                            for await (const event of stream) {
+                                try {
+                                    assemble.add(event);
+                                }
+                                catch { /* never break the caller's stream */ }
+                                yield event;
+                            }
+                            finished = true;
+                        }
+                        catch (error) {
+                            done(null, String(error?.message ?? error));
+                            throw error;
+                        }
+                        finally {
+                            if (finished)
+                                done(assemble.result(), null);
+                        }
+                    })();
+                }
+                else
+                    done(omitBytes(rest), null);
+            }
+            catch (error) {
+                warnLimited('record', `could not record a model call: ${error.message}`);
+            }
+        }, error => done(null, String(error?.message ?? error)));
+        return reply;
+    };
+    wrapped[MARK] = true;
+    return wrapped;
+}
+function converseAssembler() {
+    const content = [];
+    const json = [];
+    const out = { output: { message: { role: 'assistant', content } }, stopReason: null };
+    return {
+        add(event) {
+            if (event.messageStart?.role)
+                out.output.message.role = event.messageStart.role;
+            const start = event.contentBlockStart;
+            if (start) {
+                const tool = start.start?.toolUse;
+                content[start.contentBlockIndex ?? content.length] = tool ? { toolUse: { toolUseId: tool.toolUseId, name: tool.name, input: {} } } : {};
+            }
+            const delta = event.contentBlockDelta;
+            if (delta) {
+                const i = delta.contentBlockIndex ?? 0, d = delta.delta ?? {};
+                const block = content[i] ??= {};
+                if (typeof d.text === 'string')
+                    block.text = (block.text ?? '') + d.text;
+                if (d.toolUse) {
+                    block.toolUse ??= { input: {} };
+                    json[i] = (json[i] ?? '') + (d.toolUse.input ?? '');
+                }
+                if (d.reasoningContent) {
+                    const r = block.reasoningContent ??= { reasoningText: { text: '' } };
+                    if (d.reasoningContent.text)
+                        r.reasoningText.text += d.reasoningContent.text;
+                    if (d.reasoningContent.signature)
+                        r.reasoningText.signature = d.reasoningContent.signature;
+                }
+            }
+            const stop = event.contentBlockStop;
+            if (stop && json[stop.contentBlockIndex] && content[stop.contentBlockIndex]?.toolUse) {
+                try {
+                    content[stop.contentBlockIndex].toolUse.input = JSON.parse(json[stop.contentBlockIndex]);
+                }
+                catch {
+                    content[stop.contentBlockIndex].toolUse.input = json[stop.contentBlockIndex];
+                }
+            }
+            if (event.messageStop)
+                out.stopReason = event.messageStop.stopReason ?? null;
+            if (event.metadata?.usage)
+                out.usage = event.metadata.usage;
+            if (event.metadata?.metrics)
+                out.metrics = event.metadata.metrics;
+        },
+        result: () => ({ ...out, output: { message: { ...out.output.message, content: content.filter(Boolean) } } }),
+    };
+}
+// ---- OpenAI Realtime (voice) ----------------------------------------------------------------
+//
+// A realtime session is one WebSocket carrying many responses. Each response.done becomes one 'openai-realtime'
+// call shaped like a Responses call: request { model, instructions, tools, input } and response { id, output, usage },
+// so the platform folds a whole phone call into one conversation. Only text is kept: transcripts, text, tool calls
+// and tool results. Audio bytes are never read into this state. Events are observed, never changed or delayed.
+const REALTIME_MODULES = ['openai/realtime/ws', 'openai/realtime/websocket'];
+const REALTIME_MAX_ITEMS = 200;
+// A finished response is recorded once the next response starts, the socket closes, or this long after
+// response.done, so an interruption that arrives while its audio is still playing is on the call.
+const REALTIME_HOLD_MS = 60_000;
+const RT_STATE = Symbol.for('generalcontext.realtime');
+const rtLive = new Set();
+let rtExitHooked = false;
+function patchRealtime(value) {
+    const v = value;
+    for (const cls of [v?.OpenAIRealtimeWS, v?.OpenAIRealtimeWebSocket, v?.default?.OpenAIRealtimeWS, v?.default?.OpenAIRealtimeWebSocket]) {
+        const proto = cls?.prototype;
+        if (!proto || typeof proto.send !== 'function' || typeof proto._emit !== 'function' || Object.prototype.hasOwnProperty.call(proto, MARK))
+            continue;
+        Object.defineProperty(proto, MARK, { value: true });
+        patchRealtimeMethods(proto, undefined);
+        // The constructor assigns this.socket in the app's async scope, so that is where the session id is read.
+        // Automatic reconnects assign it again; the state (and its session) stays the same.
+        const SOCKET = Symbol('socket');
+        Object.defineProperty(proto, 'socket', {
+            configurable: true,
+            get() { return this[SOCKET]; },
+            set(socket) {
+                Object.defineProperty(this, SOCKET, { value: socket, writable: true, configurable: true });
+                try {
+                    watchSocketClose(this, rtState(this, undefined), socket);
+                }
+                catch { /* never break the app */ }
+            },
+        });
+    }
+}
+function patchRealtimeMethods(owner, scrub) {
+    const emit = owner._emit, send = owner.send, sendRef = owner.sendResponseCreate;
+    if (typeof emit === 'function' && !emit[MARK]) {
+        owner._emit = function (name, ...args) {
+            try {
+                if (name === 'event')
+                    rtServerEvent(rtState(this, scrub), args[0]);
+                else if (name === 'close')
+                    rtClose(rtState(this, scrub)); // connections with automatic reconnects
+            }
+            catch { /* never break the app */ }
+            return emit.call(this, name, ...args);
+        };
+        owner._emit[MARK] = true;
+    }
+    for (const [key, original] of [['send', send], ['sendResponseCreate', sendRef]]) {
+        if (typeof original !== 'function' || original[MARK])
+            continue;
+        owner[key] = function (event, ...rest) {
+            try {
+                rtClientEvent(rtState(this, scrub), event);
+            }
+            catch { /* never break the app */ }
+            return original.call(this, event, ...rest);
+        };
+        owner[key][MARK] = true;
+    }
+}
+function rtState(conn, scrub) {
+    let state = conn[RT_STATE];
+    if (!state) {
+        state = { at: scope.getStore(), scrub, entries: [], seq: 0, speechStoppedAt: null, current: null, held: null, closed: false, bytesPerMs: 48, sentSeq: 0 };
+        try {
+            state.model = new URL(String(conn.url)).searchParams.get('model') ?? undefined;
+        }
+        catch { /* no url yet */ }
+        Object.defineProperty(conn, RT_STATE, { value: state, configurable: true });
+    }
+    else if (scrub !== undefined && state.scrub === undefined)
+        state.scrub = scrub;
+    return state;
+}
+const rtWatched = new WeakSet();
+function watchSocketClose(conn, state, socket) {
+    if (!socket || typeof socket !== 'object' || rtWatched.has(socket))
+        return;
+    rtWatched.add(socket);
+    const onClose = () => { if (conn.socket === socket || conn.socket === undefined)
+        rtClose(state); };
+    if (typeof socket.on === 'function')
+        socket.on('close', onClose);
+    else if (typeof socket.addEventListener === 'function')
+        socket.addEventListener('close', onClose);
+}
+const rtText = (item) => {
+    const parts = [];
+    for (const part of Array.isArray(item?.content) ? item.content : []) {
+        const text = typeof part?.text === 'string' ? part.text : typeof part?.transcript === 'string' ? part.transcript : null;
+        if (text)
+            parts.push(text);
+    }
+    return parts.length ? parts.join('') : null;
+};
+const rtHasAudio = (item) => (Array.isArray(item?.content) ? item.content : []).some((p) => /audio/.test(String(p?.type)));
+// A realtime item as a Responses item, text only. Returns null when there is no text yet (audio awaiting its transcript).
+function rtConvert(item) {
+    if (!item || typeof item !== 'object')
+        return null;
+    if (item.type === 'function_call')
+        return { type: 'function_call', call_id: item.call_id, name: item.name, arguments: item.arguments ?? '' };
+    if (item.type === 'function_call_output')
+        return { type: 'function_call_output', call_id: item.call_id, output: item.output };
+    if (item.type !== 'message')
+        return null;
+    const audio = rtHasAudio(item);
+    // Caller audio with no transcript (input transcription is off, or it has not arrived yet) is kept as an
+    // empty spoken turn, so the call still shows that the caller spoke. A transcript fills it in later.
+    const text = rtText(item) ?? (audio && item.role !== 'assistant' ? '' : null);
+    if (text == null)
+        return null;
+    if (item.role === 'assistant')
+        return { type: 'message', role: 'assistant', content: [{ type: 'output_text', text }], ...(audio ? { audio: true } : {}) };
+    return { role: item.role ?? 'user', content: [{ type: 'input_text', text }], ...(audio ? { audio: true } : {}) };
+}
+function rtEntry(state, id, match) {
+    let entry = id ? state.entries.find(e => e.id === id) : undefined;
+    if (!entry && match)
+        entry = state.entries.find(e => !e.id && match(e));
+    if (entry) {
+        if (id)
+            entry.id = id;
+        return entry;
+    }
+    entry = { id: id ?? null, item: null, seq: ++state.seq };
+    state.entries.push(entry);
+    if (state.entries.length > REALTIME_MAX_ITEMS)
+        state.entries.splice(0, state.entries.length - REALTIME_MAX_ITEMS);
+    return entry;
+}
+function rtUpsert(state, raw, responseId) {
+    const id = typeof raw?.id === 'string' ? raw.id : null;
+    const converted = rtConvert(raw);
+    // The app's own conversation.item.create usually has no id. Match the server's echo of it by content.
+    const same = converted && JSON.stringify(converted);
+    const entry = rtEntry(state, id, e => !!same && !!e.item && JSON.stringify(e.item) === same);
+    // An untranscribed echo of an item never erases a transcript that already arrived.
+    const blank = converted && 'role' in converted && converted.role !== 'assistant' && !converted.content?.[0]?.text;
+    if (converted && !(blank && entry.item))
+        entry.item = { ...entry.item, ...converted };
+    if (responseId)
+        entry.response = responseId;
+}
+function rtServerEvent(state, event) {
+    const type = typeof event?.type === 'string' ? event.type : '';
+    const now = new Date().toISOString();
+    switch (type) {
+        case 'session.created':
+        case 'session.updated': {
+            const s = event.session ?? {};
+            if (typeof s.model === 'string')
+                state.model = s.model;
+            if ('instructions' in s)
+                state.instructions = s.instructions;
+            if ('tools' in s)
+                state.tools = s.tools;
+            rtFormat(state, s);
+            break;
+        }
+        case 'conversation.item.added':
+        case 'conversation.item.created':
+        case 'conversation.item.done': {
+            const r = state.current;
+            const own = r && typeof event.item?.id === 'string' && r.output_ids.has(event.item.id) ? r.id : undefined;
+            rtUpsert(state, event.item, own);
+            break;
+        }
+        case 'conversation.item.input_audio_transcription.completed': {
+            const entry = rtEntry(state, event.item_id);
+            entry.item = { role: 'user', content: [{ type: 'input_text', text: String(event.transcript ?? '') }], audio: true, at: now };
+            break;
+        }
+        // Every user turn starts with speech_started. It interrupts the response being generated, or the last one
+        // while its audio is still playing (audio length estimated from the delta sizes, the bytes are not kept).
+        case 'input_audio_buffer.speech_started': {
+            const r = state.current ?? state.held;
+            if (r && (r === state.current || Date.now() < rtPlaybackEnd(state, r)))
+                r.voice.interrupted = true;
+            break;
+        }
+        case 'input_audio_buffer.speech_stopped':
+            state.speechStoppedAt = now;
+            break;
+        case 'response.created': {
+            rtRelease(state);
+            if (state.current)
+                rtFinish(state, 'a new response started before response.done');
+            state.current = rtStart(state, event.response?.id, now);
+            break;
+        }
+        case 'response.output_item.added':
+        case 'response.output_item.done':
+            rtOwn(state, event.response_id, event.item?.id);
+            rtUpsert(state, event.item, rtResponseFor(state, event.response_id)?.id);
+            break;
+        case 'response.output_audio.delta':
+        case 'response.audio.delta': {
+            const r = rtResponseFor(state, event.response_id);
+            if (r && !r.voice.first_audio_at) {
+                r.voice.first_audio_at = now;
+                r.firstAudioMs = Date.now();
+                // The caller has been waiting since they stopped speaking, through any silent response before
+                // this one (a tool call): the wait ends at the first audio they hear.
+                if (state.speechStoppedAt)
+                    r.voice.speech_stopped_at = state.speechStoppedAt;
+                state.speechStoppedAt = null;
+            }
+            if (r && typeof event.delta === 'string')
+                r.audioBytes += Math.floor(event.delta.length * 3 / 4);
+            break;
+        }
+        case 'response.output_audio_transcript.done':
+        case 'response.audio_transcript.done':
+        case 'response.output_text.done':
+        case 'response.text.done': {
+            const r = rtResponseFor(state, event.response_id);
+            rtOwn(state, event.response_id, event.item_id);
+            const entry = rtEntry(state, event.item_id);
+            const text = String(event.transcript ?? event.text ?? '');
+            const audio = type.includes('audio');
+            entry.item = { type: 'message', role: 'assistant', content: [{ type: 'output_text', text }], ...(audio ? { audio: true } : {}) };
+            if (r)
+                entry.response = r.id;
+            break;
+        }
+        case 'response.function_call_arguments.done': {
+            rtOwn(state, event.response_id, event.item_id);
+            const entry = rtEntry(state, event.item_id);
+            entry.item = { type: 'function_call', call_id: event.call_id, name: event.name, arguments: event.arguments ?? '' };
+            const r = rtResponseFor(state, event.response_id);
+            if (r)
+                entry.response = r.id;
+            break;
+        }
+        case 'conversation.item.truncated':
+            rtTruncate(state, event.item_id, event.audio_end_ms);
+            break;
+        case 'error': {
+            const r = state.current;
+            if (r)
+                r.error = `realtime error: ${String(event.error?.message ?? 'unknown error')}${event.error?.code ? ` (${event.error.code})` : ''}`;
+            break;
+        }
+        case 'response.done': {
+            const res = event.response ?? {};
+            let r = rtResponseFor(state, res.id);
+            if (!r)
+                r = state.current = rtStart(state, res.id, now);
+            for (const item of Array.isArray(res.output) ? res.output : []) {
+                rtOwn(state, r.id, item?.id);
+                rtUpsert(state, item, r.id);
+            }
+            r.ended_at = now;
+            r.usage = res.usage ?? null;
+            if (res.status === 'failed' && !r.error)
+                r.error = `response failed: ${JSON.stringify(res.status_details ?? null)}`;
+            if (res.status === 'cancelled')
+                r.voice.interrupted = true;
+            if (state.current === r)
+                state.current = null;
+            rtHold(state, r);
+            break;
+        }
+    }
+}
+function rtClientEvent(state, event) {
+    switch (event?.type) {
+        case 'session.update': {
+            const s = event.session ?? {};
+            if (typeof s.model === 'string')
+                state.model = s.model;
+            if ('instructions' in s)
+                state.instructions = s.instructions;
+            if ('tools' in s)
+                state.tools = s.tools;
+            rtFormat(state, s);
+            break;
+        }
+        case 'conversation.item.create':
+            rtUpsert(state, event.item);
+            break;
+        case 'response.create': {
+            const r = event.response ?? {};
+            if ('instructions' in r || 'tools' in r)
+                state.next = { instructions: r.instructions, tools: r.tools };
+            break;
+        }
+        case 'response.cancel':
+            rtInterrupt(state);
+            break;
+        case 'conversation.item.truncate':
+            rtTruncate(state, event.item_id, event.audio_end_ms);
+            break;
+    }
+}
+function rtStart(state, id, now) {
+    const r = {
+        id: typeof id === 'string' ? id : `resp_${randomUUID()}`, seq: ++state.seq, started_at: now, error: null, output_ids: new Set(), audioBytes: 0,
+        voice: { speech_stopped_at: state.speechStoppedAt, first_audio_at: null, interrupted: false },
+    };
+    if (state.next) {
+        r.instructions = state.next.instructions;
+        r.tools = state.next.tools;
+        state.next = undefined;
+    }
+    rtLive.add(state);
+    return r;
+}
+// Output audio is pcm16 at 24 kHz (48 bytes per ms) unless the session picked G.711 (8 bytes per ms).
+function rtFormat(state, session) {
+    const format = session?.audio?.output?.format?.type ?? session?.output_audio_format;
+    if (typeof format === 'string')
+        state.bytesPerMs = /pcmu|pcma|g711/.test(format) ? 8 : 48;
+}
+const rtPlaybackEnd = (state, r) => r.firstAudioMs === undefined ? 0 : r.firstAudioMs + r.audioBytes / state.bytesPerMs;
+const rtResponseFor = (state, id) => [state.current, state.held].find(r => r && (!id || r.id === id)) ?? null;
+function rtOwn(state, responseId, itemId) {
+    const r = rtResponseFor(state, responseId);
+    if (r && typeof itemId === 'string')
+        r.output_ids.add(itemId);
+}
+// The response the caller is hearing: the one in flight, else the last finished one still playing.
+function rtInterrupt(state) {
+    const r = state.current ?? state.held;
+    if (r)
+        r.voice.interrupted = true;
+}
+function rtTruncate(state, itemId, ms) {
+    const r = [state.current, state.held].find(x => x && typeof itemId === 'string' && x.output_ids.has(itemId)) ?? state.current ?? state.held;
+    if (!r)
+        return;
+    r.voice.interrupted = true;
+    if (typeof ms === 'number')
+        r.voice.truncated_at_ms = ms;
+}
+function rtHold(state, r) {
+    rtRelease(state);
+    state.held = r;
+    r.timer = setTimeout(() => { if (state.held === r)
+        rtRelease(state); }, REALTIME_HOLD_MS);
+    r.timer.unref?.();
+    if (!rtExitHooked) {
+        rtExitHooked = true;
+        process.once('beforeExit', () => {
+            for (const s of rtLive) {
+                rtRelease(s);
+            }
+            if (queue.length)
+                void flush(EXIT_FLUSH_MS);
+        });
+    }
+}
+function rtRelease(state) {
+    const r = state.held;
+    if (!r)
+        return;
+    state.held = null;
+    if (r.timer)
+        clearTimeout(r.timer);
+    rtRecord(state, r);
+    if (!state.current)
+        rtLive.delete(state);
+}
+function rtFinish(state, error) {
+    const r = state.current;
+    if (!r)
+        return;
+    state.current = null;
+    r.ended_at = new Date().toISOString();
+    r.error ??= error;
+    rtRecord(state, r);
+}
+function rtClose(state) {
+    if (state.closed)
+        return;
+    state.closed = true;
+    rtRelease(state);
+    rtFinish(state, 'connection closed before response.done');
+    rtLive.delete(state);
+}
+function rtRecord(state, r) {
+    try {
+        const input = [], output = [];
+        // Timing goes on the stored assistant item too, so every later call's input still carries it.
+        const stopped = r.voice.speech_stopped_at ? Date.parse(r.voice.speech_stopped_at) : NaN;
+        const audio = r.voice.first_audio_at ? Date.parse(r.voice.first_audio_at) : NaN;
+        const timing = {
+            latency_ms: Number.isFinite(stopped) && Number.isFinite(audio) ? audio - stopped : null, interrupted: r.voice.interrupted,
+        };
+        if (r.voice.truncated_at_ms !== undefined)
+            timing.truncated_at_ms = r.voice.truncated_at_ms;
+        for (const e of state.entries) {
+            if (!e.item)
+                continue;
+            const mine = e.response === r.id || (e.id && r.output_ids.has(e.id));
+            if (mine && e.item.role === 'assistant')
+                e.item = { ...e.item, voice: timing };
+            if (mine)
+                output.push(e.item);
+            else if (e.seq < r.seq && e.seq > state.sentSeq)
+                input.push(e.item);
+        }
+        const request = {
+            model: state.model ?? null,
+            instructions: r.instructions ?? state.instructions ?? null,
+            tools: r.tools ?? state.tools ?? [],
+            input,
+        };
+        // After the first reply, only what is new since the last one, chained to it, as the Responses API does.
+        if (state.lastId)
+            request.previous_response_id = state.lastId;
+        state.lastId = r.id;
+        state.sentSeq = Math.max(state.sentSeq, ...state.entries.filter(e => e.item && (e.seq < r.seq || e.response === r.id || (e.id && r.output_ids.has(e.id)))).map(e => e.seq));
+        record({
+            api: 'openai-realtime', request, response: { id: r.id, output, usage: r.usage ?? null },
+            started_at: r.started_at, ended_at: r.ended_at ?? new Date().toISOString(), error: r.error, voice: { ...r.voice },
+        }, state.at?.tags && !state.at.tags.channel ? { ...state.at, tags: { ...state.at.tags, channel: 'voice' } } : state.at, state.scrub);
+    }
+    catch (error) {
+        warnLimited('record', `could not record a realtime response: ${error.message}`);
+    }
+}
 const PATTERNS = [
     // Bounded repeats and a token-start lookbehind keep this linear on adversarial input.
     ['email', /(?<![A-Za-z0-9._%+-])[A-Za-z0-9._%+-]{1,64}@[A-Za-z0-9-]{1,63}(?:\.[A-Za-z0-9-]{1,63}){0,8}\.[A-Za-z]{2,24}/g],
@@ -581,11 +1536,32 @@ export function wrap(client, options = {}) {
         owners.push([target.responses, 'openai-responses']);
     if (!target.chat && target.messages?.create)
         owners.push([target.messages, 'anthropic']);
-    if (!owners.length)
-        throw new TypeError('wrap expects an OpenAI or Anthropic client');
+    if (!target.chat && target.beta?.messages?.create)
+        owners.push([target.beta.messages, 'anthropic']);
+    const models = target.models;
+    const gemini = typeof models?.generateContent === 'function';
+    const bedrock = typeof target.send === 'function' && target.config && !gemini && !owners.length;
+    // An OpenAIRealtimeWS or OpenAIRealtimeWebSocket connection. Its session is the one wrap() runs in.
+    const realtime = typeof target.send === 'function' && typeof target._emit === 'function' && 'socket' in target;
+    if (realtime) {
+        patchRealtimeMethods(target, scrub);
+        try {
+            watchSocketClose(target, rtState(target, scrub), target.socket);
+        }
+        catch { /* never break the app */ }
+        return client;
+    }
+    if (!owners.length && !gemini && !bedrock)
+        throw new TypeError('wrap expects an OpenAI, Anthropic, Google Gen AI or Bedrock Runtime client, or an OpenAI Realtime connection');
     for (const [owner, api] of owners) {
         if (!owner.create[MARK])
             owner.create = recording(owner.create.bind(owner), api, scrub);
     }
+    if (gemini && !models.generateContent[MARK])
+        models.generateContent = recording(models.generateContent.bind(models), 'gemini', scrub);
+    if (gemini && typeof models.generateContentStream === 'function' && !models.generateContentStream[MARK])
+        models.generateContentStream = geminiStreamRecording(models.generateContentStream.bind(models), scrub);
+    if (bedrock && !target.send[MARK])
+        target.send = bedrockRecording(target.send.bind(target), scrub);
     return client;
 }