mixdog 0.9.3 → 0.9.5

package/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs

@@ -23,11 +23,8 @@
  * auth bundle; this module handles connection caching, delta framing, event
  * parsing, and tracing.
  */
-import WebSocket from 'ws';
-import { errText } from '../../../shared/err-text.mjs';
-import { createHash, randomBytes } from 'crypto';
+import { createHash } from 'crypto';
 import {
-    extractCachedTokens,
     traceAgentFetch,
     traceAgentSse,
     traceAgentUsage,
@@ -39,53 +36,47 @@ import {
     createStreamSafetyStamps,
     jitterDelayMs,
     MIDSTREAM_RETRY_POLICY,
-    populateHttpStatusFromMessage,
     sleepWithAbort,
 } from './retry-classifier.mjs';
-import { makeInvalidToolArgsMarker } from './openai-compat-stream.mjs';
-import { createLeakGuard, createToolCallDedupe, dedupeToolCallList } from './anthropic-leaked-toolcall.mjs';
 import {
     PROVIDER_RETRY_MAX_ATTEMPTS,
-    PROVIDER_WS_ACQUIRE_TIMEOUT_MS,
-    PROVIDER_WS_HANDSHAKE_TIMEOUT_MS,
-    PROVIDER_WS_INTER_CHUNK_TIMEOUT_MS,
-    PROVIDER_SEMANTIC_IDLE_TIMEOUT_MS,
-    PROVIDER_SSE_IDLE_WATCHDOG_ENABLED,
-    streamStalledError,
-    resolveTimeoutMs,
 } from '../stall-policy.mjs';
-import { customToolCallFromResponseItem } from './custom-tool-wire.mjs';
+import {
+    WS_IDLE_MS,
+    acquireWebSocket,
+    releaseWebSocket,
+    _sendFrame,
+    _closeAllPooledSockets,
+    drainOpenaiWsPool,
+} from './openai-ws-pool.mjs';
+import {
+    WS_PRE_RESPONSE_CREATED_MS,
+    WS_INTER_CHUNK_MS,
+    _sansInput,
+    _stableStringify,
+    _cloneJson,
+    _estimateFrameTokens,
+    _combineUsageWithWarmup,
+    _computeDelta,
+    _logicalResponseItemMatch,
+    parseToolSearchArgs,
+    _streamResponse,
+} from './openai-ws-stream.mjs';
+
+// Legacy import paths for scripts/tool-smoke.mjs, mixdog-session-runtime.mjs
+// (drainOpenaiWsPool), scripts/provider-toolcall-test.mjs (parseToolSearchArgs,
+// _logicalResponseItemMatch, _streamResponse) and other external callers.
+export {
+    _closeAllPooledSockets,
+    drainOpenaiWsPool,
+    _logicalResponseItemMatch,
+    parseToolSearchArgs,
+    _streamResponse,
+};
 
 globalThis.__mixdogOpenaiWsRuntimeLoaded = true;
 
-const CODEX_WS_URL = 'wss://chatgpt.com/backend-api/codex/responses';
-const CODEX_OAUTH_ORIGINATOR = 'codex_cli_rs';
-const OPENAI_WS_URL = 'wss://api.openai.com/v1/responses';
-const XAI_WS_URL = 'wss://api.x.ai/v1/responses';
-const WS_IDLE_MS = resolveTimeoutMs(
-    'MIXDOG_PROVIDER_WS_IDLE_MS',
-    20 * 60_000,
-    { minMs: 60_000, maxMs: 60 * 60_000 },
-);
-const WS_HANDSHAKE_TIMEOUT_MS = PROVIDER_WS_HANDSHAKE_TIMEOUT_MS;
-const WS_ACQUIRE_TIMEOUT_MS = PROVIDER_WS_ACQUIRE_TIMEOUT_MS;
-// Pre-`response.created` deadline. Once the socket is open and the
-// response.create frame is sent, a healthy server emits response.created
-// within seconds. If it stalls past this short bound the socket has wedged
-// post-upgrade with zero server events — treat it as a fast, retryable
-// first-byte timeout. This is the ONLY pre-stream watchdog; once any server
-// event arrives the single inter-chunk idle timer below takes over.
-// Only this short window is shortened; the post-`response.created`
-// inter-chunk / reasoning span keeps the longer deadlines below.
-const WS_PRE_RESPONSE_CREATED_MS = (() => {
-    const raw = process.env.MIXDOG_PROVIDER_WS_PRE_RESPONSE_CREATED_TIMEOUT_MS;
-    const n = Number(raw);
-    if (Number.isFinite(n) && n > 0) return Math.min(Math.max(n, 1_000), 120_000);
-    return 10_000;
-})();
-// Single inter-chunk idle timer. Resets on EVERY received frame — any frame,
-// including metadata/keepalive, proves the socket is live.
-const WS_INTER_CHUNK_MS = PROVIDER_WS_INTER_CHUNK_TIMEOUT_MS;
+// --- WS_PRE_RESPONSE_CREATED_MS / WS_INTER_CHUNK_MS: extracted to openai-ws-stream.mjs ---
 // Mid-stream retry budgets + backoff now live in the shared MIDSTREAM_RETRY_POLICY
 // table (retry-classifier.mjs). These aliases keep the local call sites readable
 // and ensure the numbers exist in exactly ONE place.
@@ -113,2019 +104,11 @@ const WS_MIDSTREAM_POLICY = {
 const HANDSHAKE_MAX_ATTEMPTS = PROVIDER_RETRY_MAX_ATTEMPTS;
 const HANDSHAKE_BACKOFF_BASE_MS = 500;
 const HANDSHAKE_BACKOFF_CAP_MS = 5000;
-// WS socket pool buckets are keyed by `poolKey` (the per-call sessionId)
-// to isolate parallel agent invocations — each gets its own socket so
-// a second caller cannot grab a sibling's mid-turn entry (openai-oauth would
-// otherwise reject the new response.create with "No tool output found
-// for function call ..."). The handshake `session_id` header/URL
-// uses `cacheKey` — a prefix-scoped cache key derived from the configured
-// provider namespace plus model/system/tools hash. Same-prefix sessions share
-// server-side prompt cache, while unrelated main/worker prefixes no longer
-// evict each other inside one static provider lane. The backend dedupes cache by
-// handshake session_id, not by body.prompt_cache_key alone (measured
-// 2026-04-19 after the v0.6.151 regression).
-const MAX_POOLED_SOCKETS_PER_KEY = 8;
-
-// poolKey -> Entry[]
-// Entry: { socket, busy, idleTimer, lastResponseId, lastRequestSansInput,
-//          lastRequestInput, lastResponseItems, lastInputLen, turnState,
-//          closing, ephemeral }
-const _wsPool = new Map();
-
-// Final prompt_cache_key/session_id lane guard for OpenAI OAuth transports.
-// The provider code may shard one logical prefix into N cache keys for 10+
-// total parallelism; inside each final key we still serialize requests because
-// Live probes show same-key concurrent WebSockets can randomly miss the
-// server prompt cache even after warm-up.
-const _openAiPromptCacheLanes = new Map();
-const _openAiPromptCacheLaneRates = new Map();
-const OPENAI_PROMPT_CACHE_LANE_RATE_WINDOW_MS = 60_000;
-const OPENAI_PROMPT_CACHE_LANE_RATE_MAX_KEYS = 4096;
-
-function _positiveInt(value, fallback = 0) {
-    const n = Number(value);
-    return Number.isFinite(n) && n > 0 ? Math.floor(n) : fallback;
-}
-
-function _nonNegativeInt(value, fallback = 0) {
-    const n = Number(value);
-    return Number.isFinite(n) && n >= 0 ? Math.floor(n) : fallback;
-}
-
-function _cacheLaneHash(value) {
-    return createHash('sha256').update(String(value || '')).digest('hex').slice(0, 12);
-}
-
-function _openAiPromptCacheLaneMaxInFlight(sendOpts = {}) {
-    return _positiveInt(
-        sendOpts?.openaiCacheLaneMaxInFlight
-            ?? sendOpts?.promptCacheLaneMaxInFlight
-            ?? process.env.MIXDOG_OPENAI_CACHE_LANE_MAX_INFLIGHT
-            ?? process.env.MIXDOG_OPENAI_CACHE_MAX_INFLIGHT,
-        1,
-    );
-}
-
-function _openAiPromptCacheLaneQueueTimeoutMs(sendOpts = {}) {
-    return _positiveInt(
-        sendOpts?.openaiCacheLaneQueueTimeoutMs
-            ?? sendOpts?.promptCacheLaneQueueTimeoutMs
-            ?? process.env.MIXDOG_OPENAI_CACHE_LANE_QUEUE_TIMEOUT_MS
-            ?? process.env.MIXDOG_OPENAI_CACHE_QUEUE_TIMEOUT_MS,
-        0,
-    );
-}
-
-function _openAiPromptCacheLaneRateLimitPerMin(sendOpts = {}) {
-    return _nonNegativeInt(
-        sendOpts?.openaiCacheLaneRateLimitPerMin
-            ?? sendOpts?.promptCacheLaneRateLimitPerMin
-            ?? process.env.MIXDOG_OPENAI_CACHE_LANE_RPM
-            ?? process.env.MIXDOG_OPENAI_CACHE_KEY_RPM,
-        12,
-    );
-}
-
-function _openAiPromptCacheLaneDeltaRateLimitPerMin(sendOpts = {}) {
-    return _nonNegativeInt(
-        sendOpts?.openaiCacheLaneDeltaRateLimitPerMin
-            ?? sendOpts?.promptCacheLaneDeltaRateLimitPerMin
-            ?? process.env.MIXDOG_OPENAI_CACHE_LANE_DELTA_RPM
-            ?? process.env.MIXDOG_OPENAI_CACHE_DELTA_RPM,
-        60,
-    );
-}
-
-function _openAiPromptCacheLaneDeltaMaxItems(sendOpts = {}) {
-    return _nonNegativeInt(
-        sendOpts?.openaiCacheLaneDeltaMaxItems
-            ?? sendOpts?.promptCacheLaneDeltaMaxItems
-            ?? process.env.MIXDOG_OPENAI_CACHE_LANE_DELTA_MAX_ITEMS
-            ?? process.env.MIXDOG_OPENAI_CACHE_DELTA_MAX_ITEMS,
-        8,
-    );
-}
-
-function _openAiPromptCacheLaneDeltaMaxTokens(sendOpts = {}) {
-    return _nonNegativeInt(
-        sendOpts?.openaiCacheLaneDeltaMaxTokens
-            ?? sendOpts?.promptCacheLaneDeltaMaxTokens
-            ?? process.env.MIXDOG_OPENAI_CACHE_LANE_DELTA_MAX_TOKENS
-            ?? process.env.MIXDOG_OPENAI_CACHE_DELTA_MAX_TOKENS,
-        20_000,
-    );
-}
-
-function _openAiPromptCacheLaneSlowTraceMs(sendOpts = {}) {
-    return _positiveInt(
-        sendOpts?.openaiCacheLaneSlowTraceMs
-            ?? sendOpts?.promptCacheLaneSlowTraceMs
-            ?? process.env.MIXDOG_OPENAI_CACHE_LANE_SLOW_MS,
-        3000,
-    );
-}
-
-function _isOpenAiPromptCacheLaneAuth(auth) {
-    return auth?.type !== 'xai';
-}
-
-function _getOpenAiPromptCacheLaneState(key, maxInFlight) {
-    let state = _openAiPromptCacheLanes.get(key);
-    if (!state) {
-        state = { key, active: 0, queue: [], maxInFlight, nextId: 0 };
-        _openAiPromptCacheLanes.set(key, state);
-    }
-    state.maxInFlight = maxInFlight;
-    return state;
-}
-
-function _cleanupOpenAiPromptCacheLane(state) {
-    if (state.active === 0 && state.queue.length === 0) {
-        _openAiPromptCacheLanes.delete(state.key);
-    }
-}
-
-function _getOpenAiPromptCacheLaneRateState(key) {
-    let state = _openAiPromptCacheLaneRates.get(key);
-    if (!state) {
-        state = { key, starts: [], lastUsedAt: Date.now() };
-        _openAiPromptCacheLaneRates.set(key, state);
-    }
-    state.lastUsedAt = Date.now();
-    if (_openAiPromptCacheLaneRates.size > OPENAI_PROMPT_CACHE_LANE_RATE_MAX_KEYS) {
-        let oldestKey = null;
-        let oldestAt = Infinity;
-        for (const [k, v] of _openAiPromptCacheLaneRates) {
-            if ((v?.lastUsedAt || 0) < oldestAt) {
-                oldestAt = v.lastUsedAt || 0;
-                oldestKey = k;
-            }
-        }
-        if (oldestKey) _openAiPromptCacheLaneRates.delete(oldestKey);
-    }
-    return state;
-}
-
-function _pruneOpenAiPromptCacheLaneRateState(state, now = Date.now()) {
-    const cutoff = now - OPENAI_PROMPT_CACHE_LANE_RATE_WINDOW_MS;
-    while (state.starts.length > 0 && state.starts[0] <= cutoff) state.starts.shift();
-    state.lastUsedAt = now;
-}
-
-function _sleepWithSignal(ms, signal) {
-    if (ms <= 0) return Promise.resolve();
-    return new Promise((resolve, reject) => {
-        let timer = null;
-        let abortListener = null;
-        const cleanup = () => {
-            if (timer) clearTimeout(timer);
-            if (signal && abortListener) signal.removeEventListener('abort', abortListener);
-        };
-        abortListener = () => {
-            cleanup();
-            const reason = signal?.reason;
-            reject(reason instanceof Error ? reason : new Error('OpenAI prompt cache lane rate wait aborted'));
-        };
-        if (signal?.aborted) {
-            abortListener();
-            return;
-        }
-        if (signal) signal.addEventListener('abort', abortListener, { once: true });
-        timer = setTimeout(() => {
-            cleanup();
-            resolve();
-        }, ms);
-        timer.unref?.();
-    });
-}
-
-async function _reserveOpenAiPromptCacheLaneRate({ key, limitPerMin, signal, beforeWait }) {
-    if (!limitPerMin || limitPerMin <= 0) {
-        return { rateLimitPerMin: 0, rateWaitMs: 0, rateWindowCount: 0 };
-    }
-    const state = _getOpenAiPromptCacheLaneRateState(key);
-    const startedAt = Date.now();
-    let beforeWaitCalled = false;
-    while (true) {
-        const now = Date.now();
-        _pruneOpenAiPromptCacheLaneRateState(state, now);
-        if (state.starts.length < limitPerMin) {
-            state.starts.push(now);
-            return {
-                rateLimitPerMin: limitPerMin,
-                rateWaitMs: now - startedAt,
-                rateWindowCount: state.starts.length,
-            };
-        }
-        const waitMs = Math.max(25, state.starts[0] + OPENAI_PROMPT_CACHE_LANE_RATE_WINDOW_MS - now);
-        if (!beforeWaitCalled && typeof beforeWait === 'function') {
-            beforeWaitCalled = true;
-            await beforeWait({
-                waitMs,
-                rateLimitPerMin: limitPerMin,
-                rateWindowCount: state.starts.length,
-            });
-        }
-        await _sleepWithSignal(waitMs, signal);
-    }
-}
-
-export function _resolveOpenAiPromptCacheRatePolicy(sendOpts = {}, info = {}) {
-    const mode = String(info?.mode || '').toLowerCase();
-    const frameInputItems = Number(info?.frameInputItems);
-    const deltaTokens = Number(info?.deltaTokens);
-    const hasPreviousResponseId = info?.hasPreviousResponseId === true;
-    const fullLimitPerMin = _openAiPromptCacheLaneRateLimitPerMin(sendOpts);
-    const deltaLimitPerMin = _openAiPromptCacheLaneDeltaRateLimitPerMin(sendOpts);
-    const deltaMaxItems = _openAiPromptCacheLaneDeltaMaxItems(sendOpts);
-    const deltaMaxTokens = _openAiPromptCacheLaneDeltaMaxTokens(sendOpts);
-    const itemCount = Number.isFinite(frameInputItems) ? frameInputItems : null;
-    const tokenCount = Number.isFinite(deltaTokens) ? deltaTokens : null;
-    const smallDeltaItems = itemCount == null || deltaMaxItems <= 0 || itemCount <= deltaMaxItems;
-    const smallDeltaTokens = tokenCount == null || deltaMaxTokens <= 0 || tokenCount <= deltaMaxTokens;
-
-    if (mode === 'delta' && hasPreviousResponseId && smallDeltaItems && smallDeltaTokens) {
-        return {
-            policy: deltaLimitPerMin > 0 ? 'delta_relaxed' : 'delta_unlimited',
-            limitPerMin: deltaLimitPerMin,
-            fullLimitPerMin,
-            deltaLimitPerMin,
-            deltaMaxItems,
-            deltaMaxTokens,
-            frameInputItems: itemCount,
-            deltaTokens: tokenCount,
-        };
-    }
-
-    return {
-        policy: mode === 'delta' ? 'delta_guarded' : 'full_guard',
-        limitPerMin: fullLimitPerMin,
-        fullLimitPerMin,
-        deltaLimitPerMin,
-        deltaMaxItems,
-        deltaMaxTokens,
-        frameInputItems: itemCount,
-        deltaTokens: tokenCount,
-    };
-}
-
-function _removeQueuedOpenAiPromptCacheLaneRequest(state, request) {
-    const index = state.queue.indexOf(request);
-    if (index >= 0) state.queue.splice(index, 1);
-    _cleanupOpenAiPromptCacheLane(state);
-}
-
-function _releaseOpenAiPromptCacheLane(state) {
-    state.active = Math.max(0, state.active - 1);
-    while (state.queue.length > 0 && state.active < state.maxInFlight) {
-        const next = state.queue.shift();
-        next.cleanup?.();
-        state.active += 1;
-        next.resolve(_makeOpenAiPromptCacheLaneHandle(state, next.requestId, next.enqueuedAt, true));
-    }
-    _cleanupOpenAiPromptCacheLane(state);
-}
-
-function _makeOpenAiPromptCacheLaneHandle(state, requestId, enqueuedAt, queued) {
-    let released = false;
-    return {
-        requestId,
-        queued,
-        waitedMs: Date.now() - enqueuedAt,
-        activeCount: state.active,
-        queueDepth: state.queue.length,
-        release() {
-            if (released) return;
-            released = true;
-            _releaseOpenAiPromptCacheLane(state);
-        },
-    };
-}
-
-function _acquireOpenAiPromptCacheLane({ key, maxInFlight, signal, timeoutMs }) {
-    const state = _getOpenAiPromptCacheLaneState(key, maxInFlight);
-    const requestId = ++state.nextId;
-    const enqueuedAt = Date.now();
-    if (state.active < state.maxInFlight) {
-        state.active += 1;
-        return Promise.resolve(_makeOpenAiPromptCacheLaneHandle(state, requestId, enqueuedAt, false));
-    }
-    return new Promise((resolve, reject) => {
-        const request = { requestId, enqueuedAt, resolve, reject, cleanup: null, timer: null, abortListener: null };
-        const cleanup = () => {
-            if (request.timer) clearTimeout(request.timer);
-            if (signal && request.abortListener) signal.removeEventListener('abort', request.abortListener);
-        };
-        request.cleanup = cleanup;
-        request.abortListener = () => {
-            cleanup();
-            _removeQueuedOpenAiPromptCacheLaneRequest(state, request);
-            const reason = signal?.reason;
-            reject(reason instanceof Error ? reason : new Error('OpenAI prompt cache lane wait aborted'));
-        };
-        if (signal?.aborted) {
-            request.abortListener();
-            return;
-        }
-        if (signal) signal.addEventListener('abort', request.abortListener, { once: true });
-        if (timeoutMs > 0) {
-            request.timer = setTimeout(() => {
-                cleanup();
-                _removeQueuedOpenAiPromptCacheLaneRequest(state, request);
-                reject(new Error(`OpenAI prompt cache lane wait timed out after ${timeoutMs}ms`));
-            }, timeoutMs);
-            request.timer.unref?.();
-        }
-        state.queue.push(request);
-    });
-}
-
-async function _withOpenAiPromptCacheLane({ auth, cacheKey, sendOpts, poolKey, iteration, traceProvider, useModel, externalSignal }, fn) {
-    if (!_isOpenAiPromptCacheLaneAuth(auth) || !cacheKey) {
-        return await fn({ enabled: false, maxInFlight: 0 });
-    }
-    const maxInFlight = _openAiPromptCacheLaneMaxInFlight(sendOpts);
-    if (maxInFlight <= 0) {
-        return await fn({ enabled: false, maxInFlight: 0 });
-    }
-    const laneKey = `openai-prompt:${traceProvider || 'openai'}:${useModel || 'default'}:${cacheKey}`;
-    const laneKeyHash = _cacheLaneHash(laneKey);
-    const state = _getOpenAiPromptCacheLaneState(laneKey, maxInFlight);
-    const queued = state.active >= state.maxInFlight;
-    if (queued) {
-        appendAgentTrace({
-            sessionId: poolKey,
-            iteration,
-            kind: 'cache_lane',
-            provider: traceProvider,
-            model: useModel,
-            event: 'queued',
-            lane_key_hash: laneKeyHash,
-            max_in_flight: maxInFlight,
-            active: state.active,
-            queue_depth: state.queue.length,
-        });
-    }
-    const timeoutMs = _openAiPromptCacheLaneQueueTimeoutMs(sendOpts);
-    let handle = await _acquireOpenAiPromptCacheLane({ key: laneKey, maxInFlight, signal: externalSignal, timeoutMs });
-    let handleActive = true;
-    const laneMeta = {
-        enabled: true,
-        laneKeyHash,
-        maxInFlight,
-        ratePolicy: 'pending',
-        rateLimitPerMin: 0,
-        rateWaitMs: 0,
-        rateWindowCount: 0,
-        rateReleasedForWait: false,
-        rateReacquireWaitMs: 0,
-        queued: queued || handle.queued === true,
-        waitMs: handle.waitedMs,
-        activeAfterAcquire: handle.activeCount,
-        queueDepthAfterAcquire: handle.queueDepth,
-        async reserveRate(info = {}) {
-            if (laneMeta.ratePolicy !== 'pending') return laneMeta;
-            const policy = _resolveOpenAiPromptCacheRatePolicy(sendOpts, info);
-            let releasedForRateWait = false;
-            const rateMeta = await _reserveOpenAiPromptCacheLaneRate({
-                key: laneKey,
-                limitPerMin: policy.limitPerMin,
-                signal: externalSignal,
-                beforeWait: () => {
-                    if (!handleActive) return;
-                    handle.release();
-                    handleActive = false;
-                    releasedForRateWait = true;
-                },
-            });
-            let reacquireWaitMs = 0;
-            if (releasedForRateWait) {
-                const reacquired = await _acquireOpenAiPromptCacheLane({
-                    key: laneKey,
-                    maxInFlight,
-                    signal: externalSignal,
-                    timeoutMs,
-                });
-                handle = reacquired;
-                handleActive = true;
-                reacquireWaitMs = reacquired.waitedMs;
-                laneMeta.queued = laneMeta.queued || reacquired.queued === true;
-                laneMeta.waitMs = (Number(laneMeta.waitMs) || 0) + reacquireWaitMs;
-                laneMeta.activeAfterAcquire = reacquired.activeCount;
-                laneMeta.queueDepthAfterAcquire = reacquired.queueDepth;
-            }
-            Object.assign(laneMeta, {
-                ratePolicy: policy.policy,
-                rateLimitPerMin: rateMeta.rateLimitPerMin,
-                rateWaitMs: rateMeta.rateWaitMs,
-                rateWindowCount: rateMeta.rateWindowCount,
-                rateReleasedForWait: releasedForRateWait,
-                rateReacquireWaitMs: reacquireWaitMs,
-                rateFullLimitPerMin: policy.fullLimitPerMin,
-                rateDeltaLimitPerMin: policy.deltaLimitPerMin,
-                rateDeltaMaxItems: policy.deltaMaxItems,
-                rateDeltaMaxTokens: policy.deltaMaxTokens,
-                ratePolicyFrameInputItems: policy.frameInputItems,
-                ratePolicyDeltaTokens: policy.deltaTokens,
-            });
-            if (rateMeta.rateWaitMs > 0) {
-                appendAgentTrace({
-                    sessionId: poolKey,
-                    iteration,
-                    kind: 'cache_lane',
-                    provider: traceProvider,
-                    model: useModel,
-                    event: 'rate_wait',
-                    lane_key_hash: laneKeyHash,
-                    rate_policy: laneMeta.ratePolicy,
-                    rate_limit_per_min: rateMeta.rateLimitPerMin,
-                    rate_wait_ms: rateMeta.rateWaitMs,
-                    rate_window_count: rateMeta.rateWindowCount,
-                    released_for_rate_wait: releasedForRateWait,
-                    reacquire_wait_ms: reacquireWaitMs,
-                    frame_input_items: policy.frameInputItems,
-                    delta_tokens: policy.deltaTokens,
-                });
-            }
-            return laneMeta;
-        },
-    };
-    try {
-        return await fn(laneMeta);
-    } finally {
-        const slowTraceMs = _openAiPromptCacheLaneSlowTraceMs(sendOpts);
-        const slowWaitMs = Math.max(Number(laneMeta.rateWaitMs) || 0, Number(laneMeta.waitMs) || 0);
-        if (slowTraceMs > 0 && slowWaitMs >= slowTraceMs) {
-            appendAgentTrace({
-                sessionId: poolKey,
-                iteration,
-                kind: 'cache_lane_slow',
-                provider: traceProvider,
-                model: useModel,
-                event: laneMeta.rateWaitMs > 0 && laneMeta.waitMs > 0
-                    ? 'rate_and_queue_wait'
-                    : laneMeta.rateWaitMs > 0
-                        ? 'rate_wait'
-                        : 'queue_wait',
-                lane_key_hash: laneKeyHash,
-                payload: {
-                    event: laneMeta.rateWaitMs > 0 && laneMeta.waitMs > 0
-                        ? 'rate_and_queue_wait'
-                        : laneMeta.rateWaitMs > 0
-                            ? 'rate_wait'
-                            : 'queue_wait',
-                    provider: traceProvider,
-                    model: useModel,
-                    lane_key_hash: laneKeyHash,
-                    threshold_ms: slowTraceMs,
-                    max_wait_ms: slowWaitMs,
-                    rate_policy: laneMeta.ratePolicy,
-                    rate_limit_per_min: laneMeta.rateLimitPerMin,
-                    rate_wait_ms: laneMeta.rateWaitMs,
-                    rate_window_count: laneMeta.rateWindowCount,
-                    released_for_rate_wait: laneMeta.rateReleasedForWait,
-                    reacquire_wait_ms: laneMeta.rateReacquireWaitMs,
-                    rate_full_limit_per_min: laneMeta.rateFullLimitPerMin,
-                    rate_delta_limit_per_min: laneMeta.rateDeltaLimitPerMin,
-                    rate_delta_max_items: laneMeta.rateDeltaMaxItems,
-                    rate_delta_max_tokens: laneMeta.rateDeltaMaxTokens,
-                    rate_policy_frame_input_items: laneMeta.ratePolicyFrameInputItems,
-                    rate_policy_delta_tokens: laneMeta.ratePolicyDeltaTokens,
-                    max_in_flight: laneMeta.maxInFlight,
-                    queued: laneMeta.queued,
-                    wait_ms: laneMeta.waitMs,
-                    active_after_acquire: laneMeta.activeAfterAcquire,
-                    queue_depth_after_acquire: laneMeta.queueDepthAfterAcquire,
-                },
-            });
-        }
-        if (handleActive) handle.release();
-    }
-}
-
-function _getPoolArr(poolKey) {
-    if (!poolKey) return null;
-    let arr = _wsPool.get(poolKey);
-    if (!arr) {
-        arr = [];
-        _wsPool.set(poolKey, arr);
-    }
-    return arr;
-}
-
-function _removeFromPool(poolKey, entry) {
-    if (!poolKey) return;
-    const arr = _wsPool.get(poolKey);
-    if (!arr) return;
-    const idx = arr.indexOf(entry);
-    if (idx >= 0) arr.splice(idx, 1);
-    if (arr.length === 0) _wsPool.delete(poolKey);
-}
-
-function _scheduleIdleClose(poolKey, entry) {
-    if (!entry) return;
-    if (entry.idleTimer) clearTimeout(entry.idleTimer);
-    entry.idleTimer = setTimeout(() => {
-        if (entry.busy) return;
-        try { entry.socket.close(1000, 'idle_timeout'); } catch {}
-        _removeFromPool(poolKey, entry);
-    }, WS_IDLE_MS);
-    try { entry.idleTimer.unref?.(); } catch {}
-}
-
-function _clearIdle(entry) {
-    if (entry?.idleTimer) {
-        clearTimeout(entry.idleTimer);
-        entry.idleTimer = null;
-    }
-}
-
-function _isOpen(entry) {
-    return entry?.socket?.readyState === WebSocket.OPEN;
-}
-
-// Awaited frame send. Asserts the socket is OPEN and resolves only after
-// the underlying transport reports the buffered write succeeded (or fails)
-// via the WebSocket send callback. Raw `socket.send(JSON.stringify(...))`
-// is fire-and-forget — a wedged or half-closed socket silently queues the
-// payload and the caller assumes it landed, then later times out waiting
-// for a server event that will never arrive. Tag any failure with
-// `wsSendFailed=true` so _classifyMidstreamError routes the next attempt
-// through a fresh socket.
-function _sendFrame(entry, frame) {
-    return new Promise((resolve, reject) => {
-        const socket = entry?.socket;
-        if (!socket || socket.readyState !== WebSocket.OPEN) {
-            const err = new Error(`WS send: socket not OPEN (readyState=${socket?.readyState ?? 'n/a'})`);
-            err.wsSendFailed = true;
-            reject(err);
-            return;
-        }
-        let payload;
-        try { payload = JSON.stringify(frame); }
-        catch (e) {
-            const err = e instanceof Error ? e : new Error(String(e));
-            err.wsSendFailed = true;
-            reject(err);
-            return;
-        }
-        try {
-            // Do NOT await the send callback: on a wedged-but-OPEN socket the
-            // ws write callback may never fire, which would hang this Promise
-            // before _streamResponse arms its first-byte watchdog. Fire and
-            // resolve immediately; transport failures surface via the socket
-            // 'error'/'close' handlers and the first-byte watchdog.
-            socket.send(payload, () => {});
-            resolve();
-        } catch (e) {
-            const err = e instanceof Error ? e : new Error(String(e));
-            err.wsSendFailed = true;
-            reject(err);
-        }
-    });
-}
-
-function _buildHandshakeHeaders({ auth, sessionToken, turnState, cacheKey: _cacheKey }) {
-    // xAI WS: do NOT pin x-grok-conv-id. Measured parallel runs show that
-    // forcing a routing shard via that header alternates cold caches across
-    // parallel workers; the automatic prompt-prefix cache holds up better
-    // when each handshake is unpinned. Reference: vercel/ai xai provider.
-    const headers = auth.type === 'xai'
-        ? {
-            'Authorization': `Bearer ${auth.apiKey}`,
-        }
-        : auth.type === 'openai-direct'
-        ? {
-            'Authorization': `Bearer ${auth.apiKey}`,
-            'OpenAI-Beta': 'responses_websockets=2026-02-06',
-        }
-        : {
-            'Authorization': `Bearer ${auth.access_token}`,
-            'chatgpt-account-id': auth.account_id || '',
-            'originator': CODEX_OAUTH_ORIGINATOR,
-            'OpenAI-Beta': 'responses_websockets=2026-02-06',
-        };
-    if (sessionToken) {
-        const sid = String(sessionToken);
-        headers['session_id'] = sid;
-    }
-    // x-client-request-id must be a per-request value so server-side request
-    // traces stay distinguishable across retries / reconnects sharing the same
-    // session_id. Reusing sessionToken (= cacheKey) collapsed every request
-    // for the same conversation onto one trace bucket.
-    headers['x-client-request-id'] = randomBytes(16).toString('hex');
-    if (turnState) headers['x-codex-turn-state'] = turnState;
-    return headers;
-}
-
-// handshake session_id is the conversation slot openai-oauth uses for in-memory
-// prefix state. All orchestrator-internal dispatches for this provider share
-// the same cacheKey (built in manager.mjs via providerCacheKey()), so they
-// share the server-side prefix-cache shard across roles/sources.
-function _mintSessionToken(cacheKey, auth) {
-    // xAI's public WebSocket endpoint uses the open connection plus
-    // response ids for continuation; unlike openai-oauth, it does not need the
-    // OAuth-specific session_id handshake shard.
-    if (auth?.type === 'xai') return null;
-    return cacheKey || 'mixdog-default';
-}
-
-function _openSocket({ auth, sessionToken, turnState, externalSignal, cacheKey }) {
-    const headers = _buildHandshakeHeaders({ auth, sessionToken, turnState, cacheKey });
-    const baseUrl = auth.type === 'xai'
-        ? XAI_WS_URL
-        : auth.type === 'openai-direct'
-            ? OPENAI_WS_URL
-            : CODEX_WS_URL;
-    const _wsOpenStart = Date.now();
-    if (process.env.MIXDOG_DEBUG_AGENT) {
-        process.stderr.write(`[agent-trace] ws-open-start url=${baseUrl} tokenHash=${createHash('sha256').update(String(sessionToken)).digest('hex').slice(0, 8)} ts=${_wsOpenStart}\n`);
-    }
-    const url = baseUrl + (sessionToken ? `?session_id=${encodeURIComponent(String(sessionToken))}` : '');
-    return new Promise((resolve, reject) => {
-        let settled = false;
-        let abortListener = null;
-        let acquireTimer = null;
-        const settle = (ok, val) => {
-            if (settled) return;
-            settled = true;
-            if (acquireTimer) {
-                clearTimeout(acquireTimer);
-                acquireTimer = null;
-            }
-            if (abortListener && externalSignal) {
-                try { externalSignal.removeEventListener('abort', abortListener); } catch {}
-            }
-            (ok ? resolve : reject)(val);
-        };
-        const socket = new WebSocket(url, { headers, handshakeTimeout: WS_HANDSHAKE_TIMEOUT_MS });
-        acquireTimer = setTimeout(() => {
-            if (settled) return;
-            if (process.env.MIXDOG_DEBUG_AGENT) {
-                process.stderr.write(`[agent-trace] ws-open-fail kind=acquire_timeout timeoutMs=${WS_ACQUIRE_TIMEOUT_MS} elapsed=${Date.now() - _wsOpenStart}ms\n`);
-            }
-            try { socket.terminate(); } catch {}
-            settle(false, Object.assign(
-                new Error(`${_wsErrLabel(auth?.type === 'xai' ? 'xai' : auth?.type === 'openai-direct' ? 'openai-direct' : 'openai-oauth')} acquire timed out before open (${WS_ACQUIRE_TIMEOUT_MS}ms)`),
-                { code: 'EWSACQUIRETIMEOUT', acquireTimeoutMs: WS_ACQUIRE_TIMEOUT_MS },
-            ));
-        }, WS_ACQUIRE_TIMEOUT_MS);
-        try { acquireTimer.unref?.(); } catch {}
-        const capturedHeaders = { turnState: null };
-        socket.once('upgrade', (res) => {
-            try {
-                const ts = res?.headers?.['x-codex-turn-state'];
-                if (typeof ts === 'string' && ts.length) capturedHeaders.turnState = ts;
-            } catch {}
-        });
-        socket.once('open', () => {
-            if (process.env.MIXDOG_DEBUG_AGENT) {
-                process.stderr.write(`[agent-trace] ws-open-ok elapsed=${Date.now() - _wsOpenStart}ms\n`);
-            }
-            settle(true, { socket, turnState: capturedHeaders.turnState });
-        });
-        socket.once('error', (err) => {
-            if (process.env.MIXDOG_DEBUG_AGENT) {
-                process.stderr.write(`[agent-trace] ws-open-fail kind=error msg=${String(err?.message || err).slice(0, 120)} elapsed=${Date.now() - _wsOpenStart}ms\n`);
-            }
-            try { socket.terminate(); } catch {}
-            settle(false, err instanceof Error ? err : Object.assign(new Error(errText(err) || 'openai-oauth WS error'), { wsErrorEvent: true, original: err }));
-        });
-        socket.once('close', (code, reason) => {
-            // Half-open handshake: the peer closed before 'open'/'error' fired
-            // (TCP RST / TLS edge). Without this the connect Promise never
-            // settles and only the 600s outer watchdog can break the stall
-            // (observed stage=requesting 601s hang). Open-path closes are
-            // no-ops here because settle() has already flipped `settled`.
-            if (settled) return;
-            try { socket.terminate(); } catch {}
-            settle(false, Object.assign(
-                new Error(`${_wsErrLabel(auth?.type === 'xai' ? 'xai' : auth?.type === 'openai-direct' ? 'openai-direct' : 'openai-oauth')} handshake closed before open (code=${code})`),
-                { wsCloseCode: code, wsCloseReason: (reason && reason.toString) ? reason.toString('utf-8') : '' }));
-        });
-        socket.once('unexpected-response', (_req, res) => {
-            if (settled) return;
-            const status = res?.statusCode || 0;
-            let body = '';
-            res.on('data', c => { if (body.length < 2048) body += c.toString('utf-8'); });
-            res.on('end', () => {
-                if (process.env.MIXDOG_DEBUG_AGENT) {
-                    process.stderr.write(`[agent-trace] ws-open-fail kind=http status=${status} body=${body.slice(0, 120)} elapsed=${Date.now() - _wsOpenStart}ms\n`);
-                }
-                try { socket.terminate(); } catch {}
-                settle(false, Object.assign(new Error(`${_wsErrLabel(auth?.type === 'xai' ? 'xai' : auth?.type === 'openai-direct' ? 'openai-direct' : 'openai-oauth')} handshake ${status}: ${body.slice(0, 200)}`), { httpStatus: status, httpBody: body }));
-            });
-        });
-        if (externalSignal) {
-            const onAbort = () => {
-                try { socket.terminate(); } catch {}
-                const reason = externalSignal.reason;
-                settle(false, reason instanceof Error ? reason : new Error(`${_wsErrLabel(auth?.type === 'xai' ? 'xai' : auth?.type === 'openai-direct' ? 'openai-direct' : 'openai-oauth')} handshake aborted`));
-            };
-            if (externalSignal.aborted) { onAbort(); return; }
-            abortListener = onAbort;
-            externalSignal.addEventListener('abort', onAbort, { once: true });
-        }
-    });
-}
-
-async function acquireWebSocket({ auth, poolKey, cacheKey, forceFresh, externalSignal }) {
-    const _acqStart = Date.now();
-    if (process.env.MIXDOG_DEBUG_AGENT) {
-        process.stderr.write(`[agent-trace] acquire-start poolKey=${poolKey} cacheKey=${cacheKey} forceFresh=${forceFresh} externalAborted=${!!externalSignal?.aborted} ts=${_acqStart}\n`);
-    }
-    if (externalSignal?.aborted) {
-        const reason = externalSignal.reason;
-        throw reason instanceof Error ? reason : new Error('OpenAI OAuth WS acquire aborted');
-    }
-    if (poolKey && !forceFresh) {
-        const arr = _wsPool.get(poolKey) || [];
-        // Prune dead entries first.
-        for (let i = arr.length - 1; i >= 0; i--) {
-            if (!_isOpen(arr[i]) || arr[i].closing) {
-                _clearIdle(arr[i]);
-                arr.splice(i, 1);
-            }
-        }
-        if (arr.length === 0) _wsPool.delete(poolKey);
-        // Reuse any idle open entry (cache-warm path).
-        const idle = arr.find(e => !e.busy);
-        if (idle) {
-            _clearIdle(idle);
-            idle.busy = true;
-            // Defensive: pre-existing pooled entries created before the
-            // prefix-hash field was introduced may not have it set. Normalize
-            // to null so the first delta check reads a deterministic value
-            // (and falls back to full-create instead of silently passing).
-            if (idle.lastInputPrefixHash === undefined) idle.lastInputPrefixHash = null;
-            if (idle.lastRequestInput === undefined) idle.lastRequestInput = null;
-            if (idle.lastResponseItems === undefined) idle.lastResponseItems = null;
-            if (process.env.MIXDOG_DEBUG_AGENT) {
-                process.stderr.write(`[agent-trace] acquire-reuse poolKey=${poolKey} openSockets=${arr.length} elapsed=${Date.now() - _acqStart}ms\n`);
-            }
-            return { entry: idle, reused: true };
-        }
-        // All entries busy and bucket at cap: fall through to ephemeral socket.
-        if (arr.length >= MAX_POOLED_SOCKETS_PER_KEY) {
-            if (process.env.MIXDOG_DEBUG_AGENT) {
-                process.stderr.write(`[agent-trace] acquire-ephemeral cacheKey=${cacheKey} reason=cap elapsed=${Date.now() - _acqStart}ms\n`);
-            }
-            const ephSessionToken = _mintSessionToken(cacheKey, auth);
-            const { socket, turnState } = await _openSocket({ auth, sessionToken: ephSessionToken, turnState: null, externalSignal, cacheKey });
-            // Drain-complete fence: same invariant as the normal acquire path —
-            // if drain fired during the await, do NOT push an ephemeral entry
-            // back into the pool.
-            if (_drainComplete) {
-                try { socket.close(1000, 'drain-complete'); } catch {}
-                throw new Error('WS pool drained — process exiting');
-            }
-            const entry = {
-                socket,
-                busy: true,
-                idleTimer: null,
-                lastResponseId: null,
-                lastRequestSansInput: null,
-                lastRequestInput: null,
-                lastResponseItems: null,
-                lastInputLen: 0,
-                lastInputPrefixHash: null,
-                turnState: turnState || null,
-                closing: false,
-                ephemeral: true,
-                sessionToken: ephSessionToken,
-            };
-            socket.on('close', () => { entry.closing = true; });
-            return { entry, reused: false };
-        }
-    }
-    // Parallel sockets must not inherit sibling turnState or the openai-oauth server
-    // treats the new request as a continuation of another in-flight turn and
-    // returns "No tool output found for function call …". turnState only
-    // propagates within a single entry across its own iterations.
-    const sessionToken = _mintSessionToken(cacheKey, auth);
-    if (process.env.MIXDOG_DEBUG_AGENT) {
-        process.stderr.write(`[agent-trace] acquire-new tokenHash=${createHash('sha256').update(String(sessionToken)).digest('hex').slice(0, 8)} elapsed=${Date.now() - _acqStart}ms\n`);
-    }
-    const { socket, turnState } = await _openSocket({ auth, sessionToken, turnState: null, externalSignal, cacheKey });
-    const entry = {
-        socket,
-        busy: true,
-        idleTimer: null,
-        lastResponseId: null,
-        lastRequestSansInput: null,
-        lastRequestInput: null,
-        lastResponseItems: null,
-        lastInputLen: 0,
-        lastInputPrefixHash: null,
-        turnState: turnState || null,
-        closing: false,
-        ephemeral: false,
-        sessionToken,
-    };
-    if (poolKey && !forceFresh) _getPoolArr(poolKey).push(entry);
-    socket.on('close', () => {
-        entry.closing = true;
-        _removeFromPool(poolKey, entry);
-    });
-    return { entry, reused: false };
-}
-
-function releaseWebSocket({ entry, poolKey, keep }) {
-    if (!entry) return;
-    entry.busy = false;
-    if (!keep || !_isOpen(entry) || !poolKey || entry.ephemeral) {
-        try { entry.socket.close(1000, keep ? 'no_session' : 'release_no_keep'); } catch {}
-        _removeFromPool(poolKey, entry);
-        return;
-    }
-    _scheduleIdleClose(poolKey, entry);
-}
-
-// If the cached request (sans input) matches the current one and the current
-// input starts with the cached input, return only the tail. Otherwise return
-// the full input (fresh turn).
-function _sansInput(body) {
-    const { input: _ignored, previous_response_id: _prevIgnored, ...rest } = body;
-    return rest;
-}
-
-function _stableStringify(obj) {
-    // Shallow stable-ish: JSON.stringify with sorted top-level keys. Nested
-    // arrays (tools, include) are order-sensitive and reflect intent, so we
-    // do not sort them.
-    if (obj === null || typeof obj !== 'object' || Array.isArray(obj)) return JSON.stringify(obj);
-    const keys = Object.keys(obj).sort();
-    const parts = [];
-    for (const k of keys) parts.push(JSON.stringify(k) + ':' + _stableStringify(obj[k]));
-    return '{' + parts.join(',') + '}';
-}
-
-function _cloneJson(value) {
-    if (value == null) return value;
-    try { return JSON.parse(JSON.stringify(value)); } catch { return value; }
-}
-
-function _responseItemKey(item, fallbackIndex = 0) {
-    if (!item || typeof item !== 'object') return `primitive:${fallbackIndex}`;
-    if (item.id) return `${item.type || 'item'}:id:${item.id}`;
-    if (item.call_id) return `${item.type || 'item'}:call:${item.call_id}`;
-    try { return `${item.type || 'item'}:json:${_stableStringify(item)}`; } catch {}
-    return `${item.type || 'item'}:${fallbackIndex}`;
-}
-
-function _normalizeArguments(value) {
-    if (value == null) return '';
-    if (typeof value === 'string') {
-        const trimmed = value.trim();
-        try { return _stableStringify(JSON.parse(trimmed || '{}')); } catch { return trimmed; }
-    }
-    return _stableStringify(value);
-}
-
-function _normalizeContentPart(part) {
-    if (!part || typeof part !== 'object') return part;
-    const type = part.type === 'input_text' ? 'output_text' : part.type;
-    if (type === 'output_text') return { type, text: part.text || '' };
-    return part;
-}
-
-function _contentPartsEqual(a, b) {
-    const aa = Array.isArray(a) ? a.map(_normalizeContentPart) : [];
-    const bb = Array.isArray(b) ? b.map(_normalizeContentPart) : [];
-    return _stableStringify(aa) === _stableStringify(bb);
-}
-
-export function _logicalResponseItemMatch(inputItem, responseItem) {
-    if (!inputItem || !responseItem) return false;
-    const inputType = inputItem.type || (inputItem.role === 'assistant' ? 'message' : '');
-    const responseType = responseItem.type || '';
-    if (responseType === 'function_call') {
-        if (inputType !== 'function_call') return false;
-        const inputCallId = String(inputItem.call_id || '');
-        const responseCallId = String(responseItem.call_id || '');
-        const inputName = String(inputItem.name || '');
-        const responseName = String(responseItem.name || '');
-        if (inputCallId && responseCallId) {
-            // call_id is the server-side anchor. The replayed history may carry
-            // locally compacted arguments, but previous_response_id already
-            // points at the canonical output item.
-            return inputCallId === responseCallId && inputName === responseName;
-        }
-        return inputName === responseName
-            && _normalizeArguments(inputItem.arguments) === _normalizeArguments(responseItem.arguments);
-    }
-    if (responseType === 'tool_search_call') {
-        if (inputType !== 'tool_search_call') return false;
-        const inputCallId = String(inputItem.call_id || '');
-        const responseCallId = String(responseItem.call_id || '');
-        if (inputCallId && responseCallId) return inputCallId === responseCallId;
-        return _normalizeArguments(inputItem.arguments) === _normalizeArguments(responseItem.arguments);
-    }
-    if (responseType === 'custom_tool_call') {
-        if (inputType !== 'custom_tool_call') return false;
-        const inputCallId = String(inputItem.call_id || '');
-        const responseCallId = String(responseItem.call_id || '');
-        const inputName = String(inputItem.name || '');
-        const responseName = String(responseItem.name || '');
-        if (inputCallId && responseCallId) return inputCallId === responseCallId && inputName === responseName;
-        return inputName === responseName && String(inputItem.input || '') === String(responseItem.input || '');
-    }
-    if (responseType === 'message') {
-        const inputRole = inputItem.role || (inputType === 'message' ? 'assistant' : '');
-        const responseRole = responseItem.role || 'assistant';
-        return inputType === 'message'
-            && inputRole === responseRole
-            && _contentPartsEqual(inputItem.content, responseItem.content);
-    }
-    if (responseType === 'reasoning') {
-        return inputType === 'reasoning'
-            && (!!responseItem.id ? inputItem.id === responseItem.id : true)
-            && (!!responseItem.encrypted_content
-                ? inputItem.encrypted_content === responseItem.encrypted_content
-                : true);
-    }
-    if (responseType === 'web_search_call') {
-        return inputType === 'web_search_call'
-            && (!!responseItem.id ? inputItem.id === responseItem.id : true)
-            && _stableStringify(inputItem.action || null) === _stableStringify(responseItem.action || null);
-    }
-    if (inputType !== responseType) return false;
-    const stripVolatile = (item) => {
-        if (!item || typeof item !== 'object') return item;
-        const { id: _id, status: _status, ...rest } = item;
-        return rest;
-    };
-    return _stableStringify(stripVolatile(inputItem)) === _stableStringify(stripVolatile(responseItem));
-}
 
-function _requestInputItemsMatch(a, b) {
-    return _stableStringify(a) === _stableStringify(b);
-}
-
-function _stripRequestPrefix(curInput, prevInput) {
-    const current = Array.isArray(curInput) ? curInput : [];
-    const previous = Array.isArray(prevInput) ? prevInput : [];
-    if (current.length < previous.length) return null;
-    for (let i = 0; i < previous.length; i += 1) {
-        if (!_requestInputItemsMatch(current[i], previous[i])) return null;
-    }
-    return current.slice(previous.length);
-}
-
-function _isReplayLikeHead(item, responseItem) {
-    if (!item || !responseItem) return false;
-    const inputType = item.type || (item.role === 'assistant' ? 'message' : '');
-    const responseType = responseItem.type || '';
-    if (responseType === 'message') return inputType === 'message';
-    if (responseType === 'function_call') return inputType === 'function_call';
-    if (responseType === 'tool_search_call') return inputType === 'tool_search_call';
-    return inputType === responseType;
-}
-
-function _stripResponseItemsFromHead(items, responseItems) {
-    const tail = Array.isArray(items) ? items : [];
-    const outputs = Array.isArray(responseItems) ? responseItems : [];
-    let cursor = 0;
-    let stripped = 0;
-    let skipped = 0;
-    for (const output of outputs) {
-        if (cursor >= tail.length) break;
-        if (_logicalResponseItemMatch(tail[cursor], output)) {
-            cursor += 1;
-            stripped += 1;
-            continue;
-        }
-        if (_isReplayLikeHead(tail[cursor], output)) {
-            return {
-                ok: false,
-                reason: `response_output_mismatch:${output?.type || 'unknown'}`,
-                tail,
-                stripped,
-                skipped,
-            };
-        }
-        skipped += 1;
-    }
-    return { ok: true, reason: null, tail: tail.slice(cursor), stripped, skipped };
-}
-
-function _computeDelta({ entry, body }) {
-    if (!entry || !entry.lastRequestSansInput || !entry.lastResponseId) {
-        return { mode: 'full', reason: 'no_anchor', frame: { type: 'response.create', ...body } };
-    }
-    if (!Array.isArray(entry.lastRequestInput)) {
-        return { mode: 'full', reason: 'no_input_snapshot', frame: { type: 'response.create', ...body } };
-    }
-    const curSans = _stableStringify(_sansInput(body));
-    if (curSans !== entry.lastRequestSansInput) {
-        return { mode: 'full', reason: 'request_properties_changed', frame: { type: 'response.create', ...body } };
-    }
-    const curInput = Array.isArray(body.input) ? body.input : [];
-    const afterPreviousInput = _stripRequestPrefix(curInput, entry.lastRequestInput);
-    if (!afterPreviousInput) {
-        return { mode: 'full', reason: 'input_prefix_mismatch', frame: { type: 'response.create', ...body } };
-    }
-    const stripped = _stripResponseItemsFromHead(afterPreviousInput, entry.lastResponseItems);
-    if (!stripped.ok) {
-        return { mode: 'full', reason: stripped.reason, frame: { type: 'response.create', ...body } };
-    }
-    return {
-        mode: 'delta',
-        reason: null,
-        strippedResponseItems: stripped.stripped,
-        skippedResponseItems: stripped.skipped,
-        frame: {
-            ...body,
-            type: 'response.create',
-            previous_response_id: entry.lastResponseId,
-            input: stripped.tail,
-        },
-    };
-}
-
-function _estimateFrameTokens(frame) {
-    try {
-        const s = JSON.stringify(frame);
-        return Math.ceil(s.length / 4);
-    } catch { return 0; }
-}
-
-function _usageNum(value) {
-    const n = Number(value || 0);
-    return Number.isFinite(n) ? n : 0;
-}
-
-function _combineUsageWithWarmup(actual, warmup) {
-    if (!warmup) return actual;
-    if (!actual) return warmup;
-    const actualRaw = actual.raw || {};
-    const warmupRaw = warmup.raw || {};
-    const actualTicks = _usageNum(actualRaw.cost_in_usd_ticks);
-    const warmupTicks = _usageNum(warmupRaw.cost_in_usd_ticks);
-    return {
-        ...actual,
-        inputTokens: _usageNum(actual.inputTokens) + _usageNum(warmup.inputTokens),
-        outputTokens: _usageNum(actual.outputTokens) + _usageNum(warmup.outputTokens),
-        cachedTokens: _usageNum(actual.cachedTokens) + _usageNum(warmup.cachedTokens),
-        promptTokens: _usageNum(actual.promptTokens) + _usageNum(warmup.promptTokens),
-        warmupInputTokens: _usageNum(warmup.inputTokens),
-        warmupCachedTokens: _usageNum(warmup.cachedTokens),
-        warmupOutputTokens: _usageNum(warmup.outputTokens),
-        raw: {
-            ...actualRaw,
-            warmup_usage: warmupRaw,
-            ...(actualTicks || warmupTicks ? { cost_in_usd_ticks: actualTicks + warmupTicks } : {}),
-        },
-    };
-}
-
-function _parseEvent(raw) {
-    try { return JSON.parse(raw); } catch { return null; }
-}
-
-function _incompleteReasonFromEvent(event) {
-    const reasonObj = event?.response?.incomplete_details
-        || event?.incomplete_details
-        || event?.response?.status_details
-        || null;
-    return String(reasonObj?.reason || event?.response?.status || 'incomplete');
-}
-
-function _isMaxOutputIncompleteReason(reason) {
-    return /^(?:max_output_tokens|max_tokens|length|output_token_limit)$/i.test(String(reason || '').trim());
-}
-
-function _httpStatusFromWsClose(code, reason) {
-    const n = Number(code || 0);
-    const r = String(reason || '').toLowerCase();
-    if (n === 4401
-        || /\b(?:unauthorized|unauthorised|authentication|auth(?:enticated?)?|not authenticated|token expired|access token)\b/.test(r)) {
-        return 401;
-    }
-    if (n === 4403 || /\b(?:forbidden|policy|permission denied)\b/.test(r)) return 403;
-    if (n === 4429 || /\b(?:rate limit|quota)\b/.test(r)) return 429;
-    return 0;
-}
-
-function _wsErrLabel(p) {
-    if (p === 'xai') return 'xAI WS';
-    if (p === 'openai-direct' || p === 'openai') return 'OpenAI WS';
-    return 'OpenAI OAuth WS';
-}
-// tool_search_call.arguments parse. Module-scope (exported) for direct test
-// coverage. Same policy as the function_call_arguments.done path and
-// openai-oauth _parseJsonObject —
-// object passes through; null/non-string/empty/whitespace → {} (no args); a
-// non-empty string that fails JSON.parse is deterministic bad JSON, surfaced
-// as an invalid-args MARKER (not silently swallowed to {}) so the dispatch
-// loop returns an is_error tool_result and the model self-corrects in the same
-// turn.
-export function parseToolSearchArgs(value) {
-    if (value && typeof value === 'object') {
-        // Reject arrays — the tool_search schema is an object
-        // ({query,select,limit}); an array must never pass through as args.
-        return Array.isArray(value) ? {} : value;
-    }
-    if (typeof value !== 'string' || !value.trim()) return {};
-    try {
-        const parsed = JSON.parse(value);
-        return parsed && typeof parsed === 'object' && !Array.isArray(parsed) ? parsed : {};
-    } catch (err) {
-        return makeInvalidToolArgsMarker(value, err instanceof Error ? err.message : String(err));
-    }
-}
-export async function _streamResponse({
-    entry,
-    externalSignal,
-    onStreamDelta,
-    onToolCall,
-    onTextDelta,
-    state,
-    logSuppressedReasoningDeltas = true,
-    traceProvider = 'openai-oauth',
-    _timeouts = null,
-    knownToolNames = null,
-}) {
-    const errLabel = _wsErrLabel(traceProvider);
-    const socket = entry.socket;
-    const preResponseCreatedMs = _positiveInt(_timeouts?.preResponseCreatedMs, WS_PRE_RESPONSE_CREATED_MS);
-    const interChunkMs = _positiveInt(_timeouts?.interChunkMs, WS_INTER_CHUNK_MS);
-    const _streamingStart = Date.now();
-    let _firstDeltaEmitted = false;
-    let content = '';
-    let model = '';
-    let responseId = '';
-    let responseServiceTier = '';
-    const toolCalls = [];
-    const webSearchCalls = [];
-    const webSearchCallKeys = new Set();
-    const responseItemsAdded = [];
-    const responseItemKeys = new Set();
-    const citations = [];
-    const citationKeys = new Set();
-    const pendingCalls = new Map();
-    // Tool-work-in-flight flag: set the moment a function/custom tool call's
-    // input starts streaming (before it lands in pendingCalls/toolCalls).
-    // Gates partial-final SUCCESS so a stall mid tool-input never looks text-only.
-    let _toolInFlight = false;
-    // Fix 2: cross-path name+args dedupe. A text-leaked synthetic and an
-    // identical native function_call must fire onToolCall exactly once. Every
-    // dispatch site routes through emitToolCallDedupe.
-    const _toolDedupe = createToolCallDedupe();
-    const emitToolCallDedupe = (call) => {
-        if (!_toolDedupe.shouldDispatch(call?.name, call?.arguments)) return;
-        midState.emittedToolCall = true;
-        try { onToolCall?.(call); } catch {}
-    };
-    // Reasoning items collected from response.output_item.done (or salvaged
-    // from response.completed.response.output). The request still includes
-    // `reasoning.encrypted_content` so the server keeps emitting the blobs,
-    // but explicit input-side replay is INTENTIONALLY OMITTED in
-    // convertMessagesToResponsesInput (openai-oauth.mjs:233-238) — openai-oauth
-    // rejects the same `rs_*` id twice in one handshake session_id with a
-    // "Duplicate item" error. Server-side conversation state already carries
-    // the prefix forward across the WS_IDLE_MS window. The collected
-    // reasoningItems below are surfaced for trace/debugging only; they do
-    // not feed back into the next request body.
-    const reasoningItems = [];
-    let reasoningTextDeltaCount = 0;
-    let reasoningSummaryTextDeltaCount = 0;
-    let reasoningOtherDeltaCount = 0;
-    let reasoningDeltaLogEmitted = false;
-    const pushReasoningItem = (item) => {
-        if (!item || item.type !== 'reasoning') return;
-        if (!item.encrypted_content) return;
-        reasoningItems.push({
-            id: item.id || '',
-            encrypted_content: item.encrypted_content,
-            summary: Array.isArray(item.summary) ? item.summary : [],
-        });
-    };
-    const pushResponseItem = (item) => {
-        if (!item || typeof item !== 'object') return;
-        const key = _responseItemKey(item, responseItemsAdded.length);
-        if (responseItemKeys.has(key)) {
-            const existing = responseItemsAdded.find((candidate, index) => _responseItemKey(candidate, index) === key);
-            if (existing?.type === 'function_call' && item.type === 'function_call') {
-                if (!existing.call_id && item.call_id) existing.call_id = item.call_id;
-                if (!existing.name && item.name) existing.name = item.name;
-                if ((existing.arguments == null || existing.arguments === '') && item.arguments != null) existing.arguments = item.arguments;
-            }
-            return;
-        }
-        responseItemKeys.add(key);
-        responseItemsAdded.push(_cloneJson(item));
-    };
-    const enrichFunctionCallResponseItem = ({ itemId = '', callId = '', name = '', argumentsText = '' } = {}) => {
-        for (const item of responseItemsAdded) {
-            if (item?.type !== 'function_call') continue;
-            if (itemId && item.id && item.id !== itemId) continue;
-            if (callId && item.call_id && item.call_id !== callId) continue;
-            if (!item.call_id && callId) item.call_id = callId;
-            if (!item.name && name) item.name = name;
-            if ((item.arguments == null || item.arguments === '') && argumentsText) item.arguments = argumentsText;
-        }
-    };
-    const pushCitation = (raw, fallbackTitle = '') => {
-        const url = raw?.url || raw?.uri || raw?.href || '';
-        if (!url || citationKeys.has(url)) return;
-        citationKeys.add(url);
-        citations.push({
-            title: raw?.title || fallbackTitle || '',
-            url,
-            snippet: raw?.snippet || raw?.text || raw?.description || '',
-            source: 'openai-oauth',
-        });
-    };
-    const pushOutputTextAnnotations = (contentPart) => {
-        const annotations = Array.isArray(contentPart?.annotations) ? contentPart.annotations : [];
-        for (const annotation of annotations) pushCitation(annotation);
-    };
-    const pushWebSearchCall = (item) => {
-        if (!item || item.type !== 'web_search_call') return;
-        let key = item.id || '';
-        if (!key) {
-            try { key = JSON.stringify(item.action || item); } catch { key = `${webSearchCalls.length}`; }
-        }
-        if (webSearchCallKeys.has(key)) return;
-        webSearchCallKeys.add(key);
-        webSearchCalls.push({
-            id: item.id || '',
-            status: item.status || '',
-            action: item.action || null,
-        });
-        const action = item.action || {};
-        if (action.url) pushCitation({ url: action.url, title: action.query || '' });
-        if (Array.isArray(action.urls)) {
-            for (const url of action.urls) pushCitation({ url, title: action.query || '' });
-        }
-    };
-    const pushCustomToolCall = (item) => {
-        const call = customToolCallFromResponseItem(item);
-        if (!call || toolCalls.some((existing) => existing.id === call.id)) return;
-        toolCalls.push(call);
-        emitToolCallDedupe(call);
-    };
-    const pushToolSearchCall = (item) => {
-        if (!item || item.type !== 'tool_search_call') return;
-        const callId = item.call_id || item.id || '';
-        if (!callId || toolCalls.some((call) => call.id === callId)) return;
-        const call = {
-            id: callId,
-            name: 'tool_search',
-            arguments: parseToolSearchArgs(item.arguments),
-            nativeType: 'tool_search_call',
-        };
-        toolCalls.push(call);
-        emitToolCallDedupe(call);
-    };
-    // Leaked tool-call guard. The model sometimes emits a tool call as plain
-    // text (XML `<invoke>`/`<function_calls>` or gpt-oss harmony
-    // `<|channel|>...to=functions.NAME...<|call|>`) inside
-    // `response.output_text.delta` instead of a native function_call. Route
-    // text through the guard so leaked calls are suppressed from the visible
-    // stream, synthesized (native `call_...` id shape), and dispatched like
-    // native ones. Additive: the native function_call path is untouched.
-    const leakGuard = createLeakGuard({ knownToolNames, harmony: true });
-    const dispatchLeakedCall = (recovered) => {
-        let args = recovered?.arguments;
-        if (args === null || typeof args !== 'object' || Array.isArray(args)) args = {};
-        const call = {
-            id: `call_leaked_${randomBytes(8).toString('hex')}`,
-            name: recovered.name,
-            arguments: args,
-        };
-        if (!_toolDedupe.shouldDispatch(call.name, call.arguments)) return;
-        toolCalls.push(call);
-        midState.emittedToolCall = true;
-        try { onToolCall?.(call); } catch {}
-    };
-    const relayLeakText = (delta) => {
-        if (!leakGuard.enabled) {
-            content += delta || '';
-            if (delta && onTextDelta) {
-                if (state) state.emittedText = true;
-                try { onTextDelta(delta); } catch {}
-            }
-            return;
-        }
-        const { text, calls } = leakGuard.push(delta);
-        if (text) {
-            content += text;
-            if (onTextDelta) {
-                if (state) state.emittedText = true;
-                try { onTextDelta(text); } catch {}
-            }
-        }
-        for (const c of calls) dispatchLeakedCall(c);
-    };
-    const flushLeak = () => {
-        if (!leakGuard.enabled) return;
-        const { text, calls } = leakGuard.flush();
-        if (text) {
-            content += text;
-            if (onTextDelta) {
-                if (state) state.emittedText = true;
-                try { onTextDelta(text); } catch {}
-            }
-        }
-        for (const c of calls) dispatchLeakedCall(c);
-    };
-    const logReasoningDeltaSuppression = () => {
-        if (!logSuppressedReasoningDeltas) return;
-        const total = reasoningTextDeltaCount + reasoningSummaryTextDeltaCount + reasoningOtherDeltaCount;
-        if (reasoningDeltaLogEmitted || total === 0) return;
-        reasoningDeltaLogEmitted = true;
-        process.stderr.write(`[openai-oauth-ws] suppressed reasoning text deltas from user content count=${total} text=${reasoningTextDeltaCount} summary=${reasoningSummaryTextDeltaCount} other=${reasoningOtherDeltaCount}\n`);
-    };
-    let usage;
-    let stopReason = null;
-    let incompleteReason = null;
-    let done = false;
-    let terminalError = null;
-    // Mid-stream retry classifier needs to distinguish "stream died before we
-    // even saw response.created" from "stream died after we had a partial
-    // response but before completion". Mutate the shared state object so the
-    // caller can inspect flags on the error path without us having to attach
-    // them manually at every reject site.
-    const midState = state || {};
-    midState.sawResponseCreated = midState.sawResponseCreated || false;
-    midState.sawCompleted = midState.sawCompleted || false;
-    midState.wsCloseCode = null;
-    midState.responseFailedPayload = null;
-    let idleTimer = null;
-    let keepaliveTimer = null;
-    let abortHandler = null;
-    let messageHandler = null;
-    let closeHandler = null;
-    let errorHandler = null;
-    // SEMANTIC idle timer: distinct from the inter-chunk
-    // timer, which resets on EVERY frame (rate_limits/metadata/keepalive keep
-    // the socket "alive"). This timer resets ONLY on meaningful output deltas
-    // (text/reasoning/tool args — the same events that call onStreamDelta) so a
-    // stream that emits some deltas then goes silent (server keepalive frames
-    // only) trips a short, named terminal StreamStalledError instead of coasting
-    // to the 300s inter-chunk cap / 30-min agent watchdog.
-    let semanticIdleTimer = null;
-    const semanticIdleMs = PROVIDER_SEMANTIC_IDLE_TIMEOUT_MS;
-    const semanticIdleEnabled = PROVIDER_SSE_IDLE_WATCHDOG_ENABLED && semanticIdleMs > 0;
-
-    return new Promise((resolve, reject) => {
-        // Pre-stream watchdog: the timer fires if the server never sends a
-        // first event (response.created) within preResponseCreatedMs
-        // after our last frame. The socket is open and the response.create
-        // frame was sent, but no server event has come back — a wedged
-        // post-upgrade socket. Healthy servers ack within seconds, so this
-        // window is intentionally short (WS_PRE_RESPONSE_CREATED_MS, ~10s).
-        // Once ANY server event arrives, resetIdle() cancels this watchdog and
-        // the single inter-chunk idle timer takes over — silent gaps
-        // mid-reasoning (openai-oauth spending 50s+ producing reasoning
-        // tokens) are normal and should not abort the turn.
-        const armPreStreamWatchdog = () => {
-            if (idleTimer) clearTimeout(idleTimer);
-            idleTimer = setTimeout(() => {
-                if (process.env.MIXDOG_DEBUG_AGENT) {
-                    process.stderr.write(`[agent-trace] ws-timeout kind=first-byte afterMs=${preResponseCreatedMs}\n`);
-                }
-                traceWsTimeout('first_byte_timeout', preResponseCreatedMs);
-                const err = new Error(`WS stream: no first server event within ${preResponseCreatedMs}ms`);
-                // Tag the close code so _classifyMidstreamError sees a 4000
-                // (our local pre-stream watchdog code) and routes through
-                // the post-upgrade-no-first-event retryable bucket.
-                err.wsCloseCode = 4000;
-                // Tag the error object itself (not just midState): the warmup
-                // path streams under a separate warmupState and rethrows on
-                // timeout BEFORE it can copy flags to the outer midState, so the
-                // outer catch's _classifyMidstreamError would otherwise see
-                // sawResponseCreated=false + close 4000 and hit the pre-created
-                // deny gate. err.firstByteTimeout makes both paths retryable.
-                err.firstByteTimeout = true;
-                midState.firstByteTimeout = true;
-                terminalError = err;
-                try { socket.close(4000, 'first_byte_timeout'); } catch {}
-                // socket.close() may not settle a half-open WS (closeHandler never
-                // fires) — reject directly so the turn retries instead of hanging
-                // until the 600s watchdog. finish() is idempotent (Promise settles
-                // once; cleanup is null-safe).
-                finish();
-            }, preResponseCreatedMs);
-        };
-        let interChunkTimer = null;
-        const traceWsTimeout = (event, timeoutMs) => {
-            try {
-                const iteration = Number(midState.iteration);
-                const attemptIndex = Number(midState.attemptIndex);
-                const payload = {
-                    provider: midState.traceProvider || traceProvider,
-                    transport: 'websocket',
-                    event,
-                    timeout_ms: timeoutMs,
-                    elapsed_ms: Date.now() - _streamingStart,
-                    model: midState.model || model || null,
-                    attempt_index: Number.isFinite(attemptIndex) ? attemptIndex : null,
-                    warmup: midState.warmup === true,
-                    saw_response_created: midState.sawResponseCreated === true,
-                };
-                appendAgentTrace({
-                    sessionId: midState.sessionId || null,
-                    iteration: Number.isFinite(iteration) ? iteration : null,
-                    kind: 'ws_timeout',
-                    ...payload,
-                    payload,
-                });
-            } catch {}
-        };
-        const clearPreStreamWatchdog = () => {
-            if (idleTimer) {
-                clearTimeout(idleTimer);
-                idleTimer = null;
-            }
-        };
-        const resetInterChunk = () => {
-            if (interChunkTimer) clearTimeout(interChunkTimer);
-            interChunkTimer = setTimeout(() => {
-                if (process.env.MIXDOG_DEBUG_AGENT) {
-                    process.stderr.write(`[agent-trace] ws-timeout kind=inter-chunk afterMs=${interChunkMs}\n`);
-                }
-                traceWsTimeout('inter_chunk_timeout', interChunkMs);
-                terminalError = new Error(`WS stream: inter-chunk inactivity for ${interChunkMs}ms`);
-                try { socket.close(4000, 'inter_chunk_timeout'); } catch {}
-                // Same half-open guard as the pre-stream watchdog: reject directly
-                // so a stuck socket.close() cannot leave the Promise pending.
-                finish();
-            }, interChunkMs);
-        };
-        // pi per-event idle: (re)armed only on meaningful output deltas via
-        // bumpSemanticIdle(). Keepalive/metadata frames DON'T touch it, so a
-        // deltas-then-silent wedge trips this short semantic window.
-        const resetSemanticIdle = () => {
-            if (!semanticIdleEnabled) return;
-            if (semanticIdleTimer) clearTimeout(semanticIdleTimer);
-            semanticIdleTimer = setTimeout(() => {
-                traceWsTimeout('semantic_idle_timeout', semanticIdleMs);
-                terminalError = streamStalledError('Responses WS', semanticIdleMs, { emittedToolCall: !!midState?.emittedToolCall });
-                // Partial-final recovery: attach streamed partial state so
-                // a wedged FINAL no-tool summary can be accepted as partial-final
-                // success by the loop. pendingToolUse gates out mid-flight tools.
-                try {
-                    terminalError.partialContent = content;
-                    terminalError.partialToolCalls = toolCalls.length ? toolCalls.slice() : undefined;
-                    terminalError.pendingToolUse = pendingCalls.size > 0
-                        || !!midState?.emittedToolCall
-                        || toolCalls.length > 0
-                        || _toolInFlight === true;
-                    terminalError.partialModel = model || undefined;
-                } catch { /* best-effort enrichment */ }
-                try { terminalError.wsCloseCode = 4000; } catch {}
-                try { socket.close(4000, 'semantic_idle_timeout'); } catch {}
-                finish();
-            }, semanticIdleMs);
-            try { semanticIdleTimer.unref?.(); } catch {}
-        };
-        // Single idle reset — called on EVERY parsed server event (matches
-        // codex, which resets one idle timer on every received WS frame). Any
-        // frame proves the socket is live; there is no separate "meaningful
-        // output" gate. Also clears the pre-stream watchdog defensively in case
-        // the first event is not response.created.
-        const resetIdle = () => {
-            clearPreStreamWatchdog();
-            resetInterChunk();
-        };
-        // Meaningful-output progress bump: called by the same delta cases that
-        // call onStreamDelta (text/reasoning/tool args). Arms the semantic idle.
-        const bumpSemanticIdle = () => { resetSemanticIdle(); };
-        const cleanup = () => {
-            if (idleTimer) clearTimeout(idleTimer);
-            if (interChunkTimer) { clearTimeout(interChunkTimer); interChunkTimer = null; }
-            if (semanticIdleTimer) { clearTimeout(semanticIdleTimer); semanticIdleTimer = null; }
-            if (keepaliveTimer) { clearInterval(keepaliveTimer); keepaliveTimer = null; }
-            if (messageHandler) socket.off('message', messageHandler);
-            if (closeHandler) socket.off('close', closeHandler);
-            if (errorHandler) socket.off('error', errorHandler);
-            if (abortHandler && externalSignal) externalSignal.removeEventListener('abort', abortHandler);
-        };
-        const finish = () => {
-            logReasoningDeltaSuppression();
-            // Flush any partial-sentinel tail held back mid-stream so
-            // legitimate trailing text is never lost (streamed-text path).
-            flushLeak();
-            cleanup();
-            if (terminalError) { reject(terminalError); return; }
-            resolve({
-                content,
-                model,
-                reasoningItems: reasoningItems.length ? reasoningItems : undefined,
-                responseItems: responseItemsAdded.length ? responseItemsAdded : undefined,
-                // Dedupe by name+args (Fix 2, array side) so an identical
-                // synthetic-leaked + native pair can't run the tool twice.
-                toolCalls: toolCalls.length ? dedupeToolCallList(toolCalls) : undefined,
-                citations: citations.length ? citations : undefined,
-                webSearchCalls: webSearchCalls.length ? webSearchCalls : undefined,
-                usage,
-                stopReason: stopReason || undefined,
-                // P1 audit fix: mirror the HTTP/SSE fallback's truncated flag
-                // for the WS path (sendViaWebSocket spreads this result
-                // through to the provider caller unchanged).
-                ...(stopReason === 'length' && content.length > 0 ? { truncated: true } : {}),
-                incompleteReason: incompleteReason || undefined,
-                responseId: responseId || undefined,
-                serviceTier: responseServiceTier || undefined,
-            });
-        };
-
-        messageHandler = (data) => {
-            resetIdle();
-            // resetIdle() above resets the SINGLE inter-chunk idle timer on
-            // EVERY received frame — response.created, metadata,
-            // rate_limits, and all deltas keep the socket alive. Separately, do
-            // NOT call onStreamDelta for every frame — metadata/keepalive frames
-            // must not reset the agent stall watchdog's lastStreamDeltaAt. Only
-            // meaningful output (text delta / tool call) updates that timestamp.
-            const text = typeof data === 'string' ? data : data.toString('utf-8');
-            const event = _parseEvent(text);
-            if (!event) return;
-            if (event.error) {
-                const err = new Error(event.error.message || 'Responses WS error');
-                try {
-                    err.payload = event.error;
-                    populateHttpStatusFromMessage(err);
-                } catch {}
-                terminalError = err;
-                finish();
-                return;
-            }
-            if (typeof event.type !== 'string') return;
-            switch (event.type) {
-                case 'response.created':
-                    midState.sawResponseCreated = true;
-                    if (event.response?.model) model = event.response.model;
-                    if (event.response?.id) responseId = event.response.id;
-                    // Server ack (first event). resetIdle() at the top of
-                    // messageHandler already cleared the pre-stream watchdog and
-                    // armed the single idle timer — no extra bookkeeping here.
-                    break;
-                case 'response.output_text.delta':
-                    try {
-                        if (!_firstDeltaEmitted) {
-                            _firstDeltaEmitted = true;
-                            if (process.env.MIXDOG_DEBUG_AGENT) {
-                                process.stderr.write(`[agent-trace] ws-first-delta sinceStreaming=${Date.now() - _streamingStart}ms\n`);
-                            }
-                        }
-                        onStreamDelta?.();
-                    } catch {}
-                    bumpSemanticIdle();
-                    // Live text relay (gateway): forward the raw text chunk so
-                    // the client renders first tokens before the final replay.
-                    // Tool-call/argument deltas intentionally stay off this path.
-                    // Invariant: once a non-empty chunk has been relayed live it
-                    // cannot be withdrawn, so flag the attempt so a later
-                    // mid-stream/truncated failure is NOT retried (retry would
-                    // concatenate a second attempt onto rendered text).
-                    // Routed through the leaked-tool-call guard: appends to
-                    // `content`, forwards visible text via onTextDelta, and
-                    // recovers/dispatches any leaked known-tool call.
-                    relayLeakText(event.delta || '');
-                    break;
-                case 'response.reasoning_text.delta':
-                case 'response.reasoning_summary_text.delta':
-                    if (event.type === 'response.reasoning_text.delta') reasoningTextDeltaCount += 1;
-                    else reasoningSummaryTextDeltaCount += 1;
-                    // Reasoning text is live model progress — refresh
-                    // lastStreamDeltaAt so stream-watchdog does not flag a
-                    // long reasoning span as a stall. The local WS idle timer
-                    // was already reset by resetIdle() at the top of
-                    // messageHandler. Reasoning is still suppressed from user
-                    // content (no `content +=` here).
-                    try { onStreamDelta?.(); } catch {}
-                    bumpSemanticIdle();
-                    break;
-                case 'response.output_item.added':
-                    if (event.item?.type === 'function_call') {
-                        pendingCalls.set(event.item.id || '', {
-                            name: event.item.name || '',
-                            callId: event.item.call_id || '',
-                        });
-                        _toolInFlight = true;
-                    } else if (event.item?.type === 'custom_tool_call') {
-                        _toolInFlight = true;
-                    } else if (event.item?.type === 'tool_search_call') {
-                        // Mark tool_search in-flight at item-added time, same
-                        // as function_call/custom_tool_call above, so the
-                        // semantic-idle stall gate's pendingToolUse never
-                        // drops a mid-flight tool_search before
-                        // response.output_item.done.
-                        _toolInFlight = true;
-                    }
-                    break;
-                case 'response.function_call_arguments.delta':
-                    _toolInFlight = true;
-                    try { onStreamDelta?.(); } catch {}
-                    bumpSemanticIdle();
-                    break;
-                case 'response.custom_tool_call_input.delta':
-                    _toolInFlight = true;
-                    try { onStreamDelta?.(); } catch {}
-                    bumpSemanticIdle();
-                    break;
-                case 'response.function_call_arguments.done': {
-                    const itemId = event.item_id || '';
-                    const pending = pendingCalls.get(itemId);
-                    // function_call_arguments.done is a completion signal:
-                    // empty/whitespace → no args ({}); a non-empty string that
-                    // fails JSON.parse is deterministic bad JSON. Surface an
-                    // invalid-args MARKER (not silent {}) so the dispatch loop
-                    // returns an is_error tool_result and the model re-issues
-                    // valid JSON in the same turn.
-                    let args = {};
-                    {
-                        const _argText = typeof event.arguments === 'string' ? event.arguments : '';
-                        if (_argText.trim() !== '') {
-                            try {
-                                args = JSON.parse(_argText);
-                            } catch (err) {
-                                args = makeInvalidToolArgsMarker(_argText, err instanceof Error ? err.message : String(err));
-                            }
-                        }
-                    }
-                    enrichFunctionCallResponseItem({
-                        itemId,
-                        callId: pending?.callId || event.call_id || '',
-                        name: pending?.name || event.name || '',
-                        argumentsText: event.arguments || JSON.stringify(args),
-                    });
-                    if (pending?.callId && pending?.name) {
-                        const call = { id: pending.callId, name: pending.name, arguments: args };
-                        toolCalls.push(call);
-                        emitToolCallDedupe(call);
-                    } else {
-                        // Synthesizing a `tc_${Date.now()}` callId here would
-                        // make the next turn fail to match the model's
-                        // function_call_output reference. Defer instead and
-                        // salvage call_id/name from the final
-                        // response.completed.output bundle below. If salvage
-                        // also fails we fail the stream explicitly — masking
-                        // the gap with a synthetic id just shifts the failure
-                        // one turn later under a confusing "No tool output
-                        // found for function call" error.
-                        toolCalls.push({
-                            id: pending?.callId || '',
-                            name: pending?.name || '',
-                            arguments: args,
-                            _pendingItemId: itemId,
-                            _deferred: true,
-                        });
-                    }
-                    try { onStreamDelta?.(); } catch {}
-                    bumpSemanticIdle();
-                    break;
-                }
-                case 'response.output_item.done':
-                    pushResponseItem(event.item);
-                    // function_call / output_text already captured via their
-                    // dedicated streaming events. The one shape we still need
-                    // here is `reasoning` — carries encrypted_content that
-                    // must be replayed on the next input to keep the openai-oauth
-                    // server-side prompt cache prefix warm.
-                    if (event.item?.type === 'reasoning') pushReasoningItem(event.item);
-                    if (event.item?.type === 'web_search_call') pushWebSearchCall(event.item);
-                    if (event.item?.type === 'tool_search_call') {
-                        pushToolSearchCall(event.item);
-                    }
-                    if (event.item?.type === 'custom_tool_call') {
-                        pushCustomToolCall(event.item);
-                    }
-                    break;
-                case 'response.completed': {
-                    const completedServiceTier = event.response?.service_tier || event.response?.serviceTier || '';
-                    if (completedServiceTier) responseServiceTier = String(completedServiceTier);
-                    if (event.response?.usage) {
-                        const u = event.response.usage;
-                        const rawUsage = responseServiceTier
-                            ? { ...u, service_tier: responseServiceTier }
-                            : u;
-                        usage = {
-                            inputTokens: u.input_tokens || 0,
-                            outputTokens: u.output_tokens || 0,
-                            cachedTokens: extractCachedTokens(u),
-                            // openai-oauth reports input_tokens as the total
-                            // prompt volume (cached portion is a subset, not
-                            // additive). Alias into the cross-provider
-                            // `promptTokens` field so downstream loggers have
-                            // uniform semantics.
-                            promptTokens: u.input_tokens || 0,
-                            raw: rawUsage,
-                        };
-                    }
-                    if (!model && event.response?.model) model = event.response.model;
-                    if (!responseId && event.response?.id) responseId = event.response.id;
-                    if (event.response?.output) {
-                        for (const item of event.response.output) {
-                            pushResponseItem(item);
-                            if (!content && item.type === 'message') {
-                                for (const c of item.content || []) {
-                                    if (c.type === 'output_text') {
-                                        // Completed-output fallback (no streamed
-                                        // text). Route through the leak guard so
-                                        // a tool call leaked only in the final
-                                        // bundle is recovered, not surfaced as
-                                        // visible content. final=true → full flush.
-                                        if (leakGuard.enabled) {
-                                            const { text, calls } = leakGuard.push(c.text || '', true);
-                                            content += text;
-                                            for (const lc of calls) dispatchLeakedCall(lc);
-                                        } else {
-                                            content += c.text || '';
-                                        }
-                                        pushOutputTextAnnotations(c);
-                                    }
-                                }
-                            }
-                            if (item.type === 'message') {
-                                for (const c of item.content || []) {
-                                    if (c.type === 'output_text') pushOutputTextAnnotations(c);
-                                }
-                            }
-                            if (item.type === 'web_search_call') pushWebSearchCall(item);
-                            if (item.type === 'tool_search_call') pushToolSearchCall(item);
-                            if (item.type === 'custom_tool_call') pushCustomToolCall(item);
-                            // Salvage path: some streams emit reasoning only
-                            // inside the final response.completed.output
-                            // bundle (no per-item .done event). Dedup by id.
-                            if (item.type === 'reasoning'
-                                && !reasoningItems.some(r => r.id === item.id)) {
-                                pushReasoningItem(item);
-                            }
-                            // Salvage path for function_call: when
-                            // arguments.done fired before (or without) a
-                            // matching output_item.added, the deferred tool
-                            // call placeholder has empty id/name. The
-                            // completed.output bundle carries the canonical
-                            // call_id/name; fill them in and emit onToolCall.
-                            if (item.type === 'function_call') {
-                                const tc = toolCalls.find(
-                                    (t) => t._deferred && t._pendingItemId === (item.id || ''),
-                                );
-                                if (tc) {
-                                    if (!tc.id && item.call_id) tc.id = item.call_id;
-                                    if (!tc.name && item.name) tc.name = item.name;
-                                    if (tc.id && tc.name) {
-                                        delete tc._deferred;
-                                        delete tc._pendingItemId;
-                                        emitToolCallDedupe(tc);
-                                    }
-                                }
-                            }
-                        }
-                    }
-                    // Salvage validation. Any deferred call still missing
-                    // id/name would propagate to the next turn as a
-                    // function_call_output the server can't anchor. Fail the
-                    // stream now so the caller sees a deterministic error
-                    // instead of a cryptic mismatch one turn later.
-                    const unresolved = toolCalls.find((t) => t._deferred);
-                    if (unresolved) {
-                        terminalError = new Error(
-                            `${errLabel} function_call salvage failed: missing call_id/name for item_id=${unresolved._pendingItemId || '?'}`,
-                        );
-                        finish();
-                        break;
-                    }
-                    midState.sawCompleted = true;
-                    done = true;
-                    finish();
-                    break;
-                }
-                case 'response.done': {
-                    // response.done is the terminal frame for some openai-oauth
-                    // streams that never emit a separate response.completed.
-                    // Route through the same completed/failed/incomplete
-                    // normalization based on event.response.status so a
-                    // server-side abort (incomplete / failed) does not slip
-                    // through as success. status === 'completed' falls
-                    // through to the success path with sawCompleted set;
-                    // anything else is converted into a terminal error.
-                    const status = event.response?.status || '';
-                    if (status === 'failed') {
-                        midState.responseFailedPayload = event;
-                        const msg = event.response?.error?.message
-                            || event.error?.message
-                            || event.message
-                            || 'response.done failed';
-                        terminalError = Object.assign(
-                            new Error(`${errLabel} response.done failed: ${msg}`),
-                            { responseFailed: event },
-                        );
-                        populateHttpStatusFromMessage(terminalError, msg);
-                        done = true;
-                        finish();
-                        break;
-                    }
-                    if (status === 'incomplete') {
-                        const reasonStr = _incompleteReasonFromEvent(event);
-                        if (_isMaxOutputIncompleteReason(reasonStr)) {
-                            incompleteReason = reasonStr;
-                            stopReason = 'length';
-                            midState.sawCompleted = true;
-                            done = true;
-                            finish();
-                            break;
-                        }
-                        terminalError = Object.assign(
-                            new Error(`${errLabel} response.done incomplete: ${reasonStr}`),
-                            { responseIncomplete: event, incompleteReason: reasonStr },
-                        );
-                        done = true;
-                        finish();
-                        break;
-                    }
-                    if (status && status !== 'completed') {
-                        terminalError = Object.assign(
-                            new Error(`${errLabel} response.done unexpected status: ${status}`),
-                            { responseDoneStatus: status },
-                        );
-                        done = true;
-                        finish();
-                        break;
-                    }
-                    midState.sawCompleted = true;
-                    done = true;
-                    finish();
-                    break;
-                }
-                case 'response.incomplete': {
-                    // Most incomplete reasons are real failures. max_output_tokens
-                    // maps cleanly to Anthropic's stop_reason=max_tokens; treating
-                    // it as an error makes Claude retry the same over-budget turn.
-                    const reasonStr = _incompleteReasonFromEvent(event);
-                    if (_isMaxOutputIncompleteReason(reasonStr)) {
-                        incompleteReason = reasonStr;
-                        stopReason = 'length';
-                        midState.sawCompleted = true;
-                        done = true;
-                        finish();
-                        break;
-                    }
-                    terminalError = Object.assign(
-                        new Error(`${errLabel} response.incomplete: ${reasonStr}`),
-                        { responseIncomplete: event, incompleteReason: reasonStr },
-                    );
-                    finish();
-                    break;
-                }
-                case 'response.failed': {
-                    // Stash the payload so the mid-stream classifier can sniff
-                    // network_error / stream_disconnected without re-parsing.
-                    midState.responseFailedPayload = event;
-                    const msg = event.response?.error?.message
-                        || event.error?.message
-                        || event.message
-                        || 'response.failed';
-                    terminalError = Object.assign(new Error(`${errLabel} response.failed: ${msg}`), {
-                        responseFailed: event,
-                    });
-                    // Sniff the server message for transient/auth/permanent
-                    // hints so the handshake / mid-stream retry classifiers
-                    // can route by httpStatus. Without this, server-side
-                    // events like "Our servers are currently overloaded"
-                    // surfaced as unclassified errors and skipped the
-                    // 5xx retry bucket entirely.
-                    populateHttpStatusFromMessage(terminalError, msg);
-                    finish();
-                    break;
-                }
-                case 'error': {
-                    const errMsg = String(event.message || event.error?.message || 'unknown');
-                    terminalError = new Error(`${errLabel} error: ${errMsg}`);
-                    populateHttpStatusFromMessage(terminalError, errMsg);
-                    finish();
-                    break;
-                }
-                default:
-                    // Catch any other reasoning-delta variants (e.g.
-                    // `response.reasoning.<sub>.delta`) so they are counted
-                    // and suppressed, never reaching the user content buffer.
-                    if (typeof event.type === 'string'
-                        && event.type.startsWith('response.reasoning')
-                        && event.type.endsWith('.delta')) {
-                        reasoningOtherDeltaCount += 1;
-                        // These ARE live model progress (reviewer Medium): a
-                        // provider that emits only these reasoning variants for a
-                        // long span would otherwise trip the SEMANTIC idle abort.
-                        // Refresh both the watchdog and the semantic idle timer,
-                        // matching the named reasoning_text.delta case above.
-                        try { onStreamDelta?.(); } catch {}
-                        bumpSemanticIdle();
-                    }
-                    // Trace-only events (response.in_progress, etc.)
-                    break;
-            }
-        };
-        closeHandler = (code, reason) => {
-            if (done) return;
-            midState.wsCloseCode = code;
-            if (!terminalError) {
-                const r = reason?.toString?.('utf-8') || '';
-                const httpStatus = _httpStatusFromWsClose(code, r);
-                terminalError = Object.assign(
-                    new Error(`OpenAI OAuth WS closed before response.completed (code=${code}${r ? `, reason=${r}` : ''})`),
-                    { wsCloseCode: code, wsCloseReason: r, ...(httpStatus ? { httpStatus } : {}) },
-                );
-            } else if (terminalError && !terminalError.wsCloseCode) {
-                try { terminalError.wsCloseCode = code; } catch {}
-                try { terminalError.httpStatus = terminalError.httpStatus || _httpStatusFromWsClose(code, reason?.toString?.('utf-8') || ''); } catch {}
-            }
-            finish();
-        };
-        errorHandler = (err) => {
-            if (done) return;
-            const wrapped = err instanceof Error ? err : new Error(String(err));
-            if (terminalError) {
-                // Preserve the first terminalError; chain the later socket
-                // error in via `cause` (or `suppressed` if cause already set)
-                // so diagnostics keep the original failure visible.
-                try {
-                    if (!terminalError.cause) terminalError.cause = wrapped;
-                    else {
-                        const list = Array.isArray(terminalError.suppressed)
-                            ? terminalError.suppressed
-                            : [];
-                        list.push(wrapped);
-                        terminalError.suppressed = list;
-                    }
-                } catch {}
-            } else {
-                terminalError = wrapped;
-            }
-            try { socket.close(4001, 'stream_error'); } catch {}
-            finish();
-        };
-        if (externalSignal) {
-            abortHandler = () => {
-                if (done) return;
-                const reason = externalSignal.reason;
-                terminalError = reason instanceof Error ? reason : new Error('OpenAI OAuth WS aborted by session close');
-                // Tag: was this a user/caller abort, or a watchdog abort?
-                // Mid-stream retry must skip user aborts but may retry watchdog
-                // aborts. The caller-owned AbortController surfaces through
-                // externalSignal; the agent stall watchdog signals via a reason
-                // object whose name === 'AgentStallAbortError'. stream-watchdog
-                // uses StreamStalledAbortError. Anything else → treat as user.
-                const reasonName = reason?.name || '';
-                if (reasonName === 'AgentStallAbortError'
-                    || reasonName === 'StreamStalledAbortError') {
-                    midState.watchdogAbort = reasonName;
-                } else {
-                    midState.userAbort = true;
-                }
-                try { socket.close(4002, 'aborted'); } catch {}
-                finish();
-            };
-            if (externalSignal.aborted) { abortHandler(); return; }
-            externalSignal.addEventListener('abort', abortHandler, { once: true });
-        }
-        socket.on('message', messageHandler);
-        socket.on('close', closeHandler);
-        socket.on('error', errorHandler);
-        armPreStreamWatchdog();
-        // Periodic client-side WS ping while the stream is active. The server's
-        // server closes with 1011 "keepalive ping timeout" when it thinks the
-        // peer is silent during long reasoning windows where no data frames
-        // flow. Sending a ping every 10s from our side keeps the socket warm.
-        // The interval is unref'd so it never holds the event loop open, and
-        // cleanup() clears it on every terminal path (completed / close /
-        // error / abort / mid-stream retry teardown).
-        keepaliveTimer = setInterval(() => {
-            try {
-                if (socket.readyState !== WebSocket.OPEN) return;
-                socket.ping();
-            } catch {}
-        }, 10_000);
-        try { keepaliveTimer.unref?.(); } catch {}
-    });
-}
+// --- WS pool/handshake/acquire/release: extracted to openai-ws-pool.mjs ---
 
-/**
- * Classify a handshake error for retry eligibility.
- *
- * Default-deny: anything we don't recognize as transient returns null (treat
- * as permanent). Permanent buckets (401/403/404/429) also return null — the
- * server has made a deterministic decision that a retry can't change.
- *
- * Returns one of:
- *   'timeout' — `ws` handshakeTimeout fired
- *   'reset'   — ECONNRESET / socket hang up
- *   'dns'     — EAI_AGAIN / ENOTFOUND / EAI_NODATA
- *   'refused' — ECONNREFUSED
- *   'network' — ENETUNREACH / EHOSTUNREACH / EPIPE
- *   'acquire_timeout' — hard client-side open/acquire deadline fired
- *   'http_5xx' (with specific status e.g. 'http_503') — server overload
- *   null      — not retryable
- */
-// Thin re-export wrapper: handshake classification now lives in the shared
-// retry-classifier (classifyHandshakeError). Kept here as a named export so
-// internal call sites (_acquireWithRetry) and any external importer keep
-// resolving the same symbol.
+// --- Delta/matching helpers + parseToolSearchArgs: extracted to openai-ws-stream.mjs ---
+// --- Stream consumer (_streamResponse): extracted to openai-ws-stream.mjs ---
 export function _classifyHandshakeError(err) {
     return classifyHandshakeError(err);
 }
@@ -2204,6 +187,182 @@ function _sleepWithAbort(ms, externalSignal, sleepFn = _defaultSleep) {
     return sleepWithAbort(ms, externalSignal, sleepFn, 'OpenAI OAuth WS retry backoff aborted');
 }
 
+function _num(value, fallback = 0) {
+    const n = Number(value);
+    return Number.isFinite(n) ? n : fallback;
+}
+
+function _envPositiveInt(name, fallback) {
+    const n = Number(process.env[name]);
+    return Number.isFinite(n) && n > 0 ? Math.floor(n) : fallback;
+}
+
+function _envRatio(name, fallback) {
+    const n = Number(process.env[name]);
+    return Number.isFinite(n) && n >= 0 && n <= 1 ? n : fallback;
+}
+
+function _cleanMetaString(value) {
+    return typeof value === 'string' ? value.trim() : '';
+}
+
+function _hashText(value, chars = 24) {
+    return createHash('sha256').update(String(value || '')).digest('hex').slice(0, chars);
+}
+
+function _sessionStartedAtUnixMs(sessionId) {
+    const parts = String(sessionId || '').split('_');
+    for (const part of parts) {
+        if (/^\d{12,}$/.test(part)) {
+            const n = Number(part);
+            if (Number.isFinite(n) && n > 0) return Math.floor(n);
+        }
+    }
+    return Date.now();
+}
+
+function _codexRequestKind(sendOpts, sessionId) {
+    const explicit = _cleanMetaString(sendOpts?.requestKind || sendOpts?.codexRequestKind);
+    if (explicit) return explicit;
+    return String(sessionId || '').includes(':compact') ? 'compaction' : 'turn';
+}
+
+function _codexInstallationId(sendOpts) {
+    return _cleanMetaString(sendOpts?.installationId || sendOpts?.codexInstallationId || process.env.MIXDOG_CODEX_INSTALLATION_ID)
+        || `mixdog-${_hashText(`${process.env.USERPROFILE || process.env.HOME || ''}:${process.cwd()}`, 32)}`;
+}
+
+function _codexMetadataBase(entry, { poolKey, cacheKey, sendOpts } = {}) {
+    const sessionId = _cleanMetaString(sendOpts?.codexSessionId || sendOpts?.session?.codexSessionId || poolKey || cacheKey)
+        || 'mixdog-session';
+    const threadId = _cleanMetaString(sendOpts?.threadId || sendOpts?.codexThreadId || sendOpts?.session?.threadId || cacheKey || sessionId)
+        || sessionId;
+    const turnId = _cleanMetaString(sendOpts?.turnId || sendOpts?.codexTurnId || sendOpts?.session?.turnId || sessionId)
+        || sessionId;
+    const windowId = _cleanMetaString(sendOpts?.windowId || sendOpts?.codexWindowId || sendOpts?.session?.windowId)
+        || `${threadId}:1`;
+    const installationId = _codexInstallationId(sendOpts);
+    const startedAt = Number.isFinite(Number(sendOpts?.turnStartedAtUnixMs))
+        ? Math.floor(Number(sendOpts.turnStartedAtUnixMs))
+        : _sessionStartedAtUnixMs(sessionId);
+    const requestKind = _codexRequestKind(sendOpts, sessionId);
+    const turnMetadata = {
+        installation_id: installationId,
+        session_id: sessionId,
+        thread_id: threadId,
+        turn_id: turnId,
+        window_id: windowId,
+        request_kind: requestKind,
+        turn_started_at_unix_ms: startedAt,
+    };
+    const metadata = {
+        'x-codex-installation-id': installationId,
+        session_id: sessionId,
+        thread_id: threadId,
+        turn_id: turnId,
+        'x-codex-window-id': windowId,
+        'x-codex-turn-metadata': JSON.stringify(turnMetadata),
+    };
+    // NOTE: no entry-level caching — pooled sockets outlive turns, and codex
+    // rebuilds metadata per request (responses_metadata.rs client_metadata()).
+    // Caching the first turn's turn_id on the socket would replay stale turn
+    // identity for every later turn on that connection.
+    return metadata;
+}
+
+function _metadataTrace(metadata) {
+    if (!metadata || typeof metadata !== 'object') {
+        return { count: 0, hash: null, hasTurnMetadata: false, hasThreadId: false };
+    }
+    const keys = Object.keys(metadata).sort();
+    return {
+        count: keys.length,
+        hash: _hashText(keys.map((key) => `${key}=${metadata[key]}`).join('\n'), 12),
+        hasTurnMetadata: typeof metadata['x-codex-turn-metadata'] === 'string' && metadata['x-codex-turn-metadata'].length > 0,
+        hasThreadId: typeof metadata.thread_id === 'string' && metadata.thread_id.length > 0,
+    };
+}
+
+function _codexWsCompatibilityHeaders(context = {}) {
+    const metadata = _codexMetadataBase(null, context);
+    const headers = {};
+    if (metadata['x-codex-window-id']) headers['x-codex-window-id'] = metadata['x-codex-window-id'];
+    if (metadata['x-codex-turn-metadata']) headers['x-codex-turn-metadata'] = metadata['x-codex-turn-metadata'];
+    // codex sends x-codex-installation-id on EVERY request incl. the WS
+    // handshake (client.rs:582-584). We previously carried it only in the
+    // body client_metadata; the server may gate x-codex-turn-state issuance
+    // on handshake-level client identity, so mirror codex here.
+    if (metadata['x-codex-installation-id']) headers['x-codex-installation-id'] = metadata['x-codex-installation-id'];
+    return headers;
+}
+
+function _withCodexWsClientMetadata(frame, entry, enabled, context = {}) {
+    if (!enabled || !frame || typeof frame !== 'object') return frame;
+    const base = _codexMetadataBase(entry, context);
+    const metadata = {
+        ...base,
+        ...(frame.client_metadata && typeof frame.client_metadata === 'object' ? frame.client_metadata : {}),
+        'x-codex-ws-stream-request-start-ms': String(Date.now()),
+    };
+    if (entry && typeof entry === 'object') {
+        // codex scopes x-codex-turn-state to ONE turn (client.rs:263-279 —
+        // "must not send it between different turns"). Pooled sockets span
+        // turns, so key the stored token by the turn that captured it and
+        // drop it when the turn changes.
+        if (entry.turnState && entry.turnStateTurnId && entry.turnStateTurnId !== base.turn_id) {
+            entry.turnState = null;
+            entry.turnStateTurnId = null;
+        }
+        entry.currentTurnId = base.turn_id;
+    }
+    if (entry?.turnState) {
+        metadata['x-codex-turn-state'] = String(entry.turnState);
+    }
+    return {
+        ...frame,
+        client_metadata: metadata,
+    };
+}
+
+function _cacheObservation({ entry, result }) {
+    const inputTokens = _num(result?.usage?.inputTokens, 0);
+    const promptTokens = _num(result?.usage?.promptTokens, 0) || inputTokens;
+    const cachedTokens = _num(result?.usage?.cachedTokens, 0);
+    const previousMaxCached = _num(entry?.promptCacheMaxCachedTokens, 0);
+    const warmThreshold = _envPositiveInt('MIXDOG_OAI_CACHE_MISS_WARM_TOKENS', 2048);
+    const promptThreshold = _envPositiveInt('MIXDOG_OAI_CACHE_MISS_PROMPT_TOKENS', 4096);
+    const dropRatio = _envRatio('MIXDOG_OAI_CACHE_MISS_DROP_RATIO', 0.6);
+    const dropThreshold = Math.floor(previousMaxCached * dropRatio);
+    const wasWarm = previousMaxCached >= warmThreshold;
+    const cacheRatio = promptTokens > 0 ? cachedTokens / promptTokens : null;
+    const zeroMiss = wasWarm && promptTokens >= promptThreshold && cachedTokens === 0;
+    const partialDrop = wasWarm
+        && promptTokens >= promptThreshold
+        && cachedTokens > 0
+        && previousMaxCached > 0
+        && cachedTokens < dropThreshold;
+    const actualMiss = zeroMiss || partialDrop;
+    return {
+        inputTokens,
+        promptTokens,
+        cachedTokens,
+        uncachedTokens: Math.max(0, promptTokens - cachedTokens),
+        previousMaxCached,
+        wasWarm,
+        warmThreshold,
+        promptThreshold,
+        dropRatio,
+        dropThreshold,
+        cacheRatio,
+        actualMiss,
+        missReason: zeroMiss
+            ? 'warm_session_zero_cached_tokens'
+            : partialDrop
+                ? 'warm_session_cached_tokens_dropped'
+                : null,
+    };
+}
+
 /**
  * Run `_acquire({auth, poolKey, cacheKey})` with bounded exponential-backoff
  * retry on transient handshake failures. The injection seams (`_acquire`,
@@ -2218,6 +377,7 @@ export async function _acquireWithRetry({
     auth,
     poolKey,
     cacheKey,
+    codexHeaders,
     forceFresh,
     onRetry,
     externalSignal,
@@ -2237,7 +397,7 @@ export async function _acquireWithRetry({
                     process.stderr.write(`[agent-trace] ws-handshake-attempt n=${attempt}\n`);
                 }
             }
-            return await _acquire({ auth, poolKey, cacheKey, forceFresh, externalSignal });
+            return await _acquire({ auth, poolKey, cacheKey, codexHeaders, forceFresh, externalSignal });
         } catch (err) {
             lastErr = err;
             const classifier = _classifyHandshakeError(err);
@@ -2388,17 +548,12 @@ export async function sendViaWebSocket({
     // this carry-forward would not help and is therefore gated to xAI.
     let carryForwardCache = null;
     const emittedProgress = [];
+    const useCodexWsClientMetadata = traceProvider === 'openai-oauth';
+    const codexMetadataContext = { poolKey, cacheKey, sendOpts };
+    const codexHandshakeHeaders = useCodexWsClientMetadata
+        ? _codexWsCompatibilityHeaders(codexMetadataContext)
+        : null;
 
-    return await _withOpenAiPromptCacheLane({
-        auth,
-        cacheKey,
-        sendOpts,
-        poolKey,
-        iteration,
-        traceProvider,
-        useModel,
-        externalSignal,
-    }, async (promptCacheLane) => {
     for (let attemptIndex = 0; attemptIndex <= MAX_MIDSTREAM_RETRIES; attemptIndex++) {
         const handshakeStart = Date.now();
         let acquired;
@@ -2410,6 +565,7 @@ export async function sendViaWebSocket({
                 auth,
                 poolKey,
                 cacheKey,
+                codexHeaders: codexHandshakeHeaders,
                 // Retry attempt must not reuse a pooled socket — the prior
                 // one is either torn down or in an unknown state.
                 forceFresh: forceFresh || attemptIndex > 0,
@@ -2502,18 +658,23 @@ export async function sendViaWebSocket({
         let deltaReason = null;
         let strippedResponseItems = 0;
         let skippedResponseItems = 0;
+        let wireFrameHadTurnState = false;
+        let wireFrameMetadataTrace = _metadataTrace(null);
+        let framePrefixHash = null;
+        let framePrefixHeadHash = null;
+        let framePrefixPrevMatch = null;
         let result;
         const streamTimeouts = null;
         try {
-            if (warmupBody && typeof warmupBody === 'object' && attemptIndex === 0) {
-                await promptCacheLane?.reserveRate?.({
-                    mode: 'full',
-                    frameInputItems: Array.isArray(warmupBody.input) ? warmupBody.input.length : null,
-                    deltaTokens: _estimateFrameTokens({ type: 'response.create', ...warmupBody }),
-                    hasPreviousResponseId: false,
-                });
+            // codex prewarm gate (client.rs:1686-1688): only when the session
+            // has no prior request state. A reused pooled socket with a live
+            // chain must go straight to the real request.
+            if (warmupBody && typeof warmupBody === 'object' && attemptIndex === 0 && !entry.lastResponseId) {
                 const warmupFrame = { type: 'response.create', ...warmupBody };
-                await _sendFrameFn(entry, warmupFrame);
+                const wireWarmupFrame = _withCodexWsClientMetadata(warmupFrame, entry, useCodexWsClientMetadata, codexMetadataContext);
+                wireFrameHadTurnState = !!wireWarmupFrame?.client_metadata?.['x-codex-turn-state'];
+                wireFrameMetadataTrace = _metadataTrace(wireWarmupFrame?.client_metadata);
+                await _sendFrameFn(entry, wireWarmupFrame);
                 const warmupStart = Date.now();
                 const warmupState = {
                     attemptIndex,
@@ -2573,13 +734,15 @@ export async function sendViaWebSocket({
                         payload: warmupPayload,
                     });
                 } catch {}
-                requestBody = { ...body, previous_response_id: warmupResult.responseId };
-                delete requestBody.instructions;
-                delete requestBody.generate;
-                entry.lastRequestSansInput = _stableStringify(_sansInput({
-                    ...requestBody,
-                    input: warmupInputArr,
-                }));
+                // Do NOT rewrite the main request after warmup (R23 finding).
+                // The old prev_id+no-instructions rewrite made it=1's frame a
+                // different shape from it=2+ full frames, so the first real
+                // full-frame cache write only landed at it=2 and it=3 raced
+                // its propagation (cached=0 early misses). Keeping it=1 as a
+                // normal full frame makes it byte-identical to the warmup's
+                // prefix (instant full hit on the cache warmup just wrote)
+                // and keeps every subsequent frame one consistent shape.
+                // Delta opt-in still chains via entry.lastResponseId above.
             }
 
             const delta = _computeDelta({ entry, body: requestBody });
@@ -2587,13 +750,34 @@ export async function sendViaWebSocket({
             deltaReason = delta.reason || null;
             strippedResponseItems = delta.strippedResponseItems || 0;
             skippedResponseItems = delta.skippedResponseItems || 0;
-            deltaTokens = _estimateFrameTokens(frame);
-            await promptCacheLane?.reserveRate?.({
-                mode,
-                frameInputItems: Array.isArray(frame.input) ? frame.input.length : null,
-                deltaTokens,
-                hasPreviousResponseId: typeof frame.previous_response_id === 'string' && frame.previous_response_id.length > 0,
-            });
+            const wireFrame = _withCodexWsClientMetadata(frame, entry, useCodexWsClientMetadata, codexMetadataContext);
+            wireFrameHadTurnState = !!wireFrame?.client_metadata?.['x-codex-turn-state'];
+            wireFrameMetadataTrace = _metadataTrace(wireFrame?.client_metadata);
+            deltaTokens = _estimateFrameTokens(wireFrame);
+            // Prefix-consistency probe (item-level). Serialized-JSON byte
+            // prefixes can never match across appends (the shorter frame ends
+            // in "]}" where the longer has ","), so compare what the server's
+            // prefix cache actually sees: the non-input request header and the
+            // per-item content of the input array. prev_match=true means the
+            // current call's header is identical and its first N input items
+            // equal the previous call's N items (append-only history).
+            try {
+                const { client_metadata: _cm, input: frameInput, ...frameHeader } = frame;
+                const headerHash = _hashText(JSON.stringify(frameHeader), 16);
+                const itemHashes = (Array.isArray(frameInput) ? frameInput : [])
+                    .map((item) => _hashText(JSON.stringify(item), 12));
+                framePrefixHash = headerHash;
+                framePrefixHeadHash = _hashText(itemHashes.join(','), 16);
+                const prevHeader = entry.lastFrameHeaderHash;
+                const prevItems = entry.lastFrameItemHashes;
+                if (prevHeader && Array.isArray(prevItems)) {
+                    framePrefixPrevMatch = headerHash === prevHeader
+                        && itemHashes.length >= prevItems.length
+                        && prevItems.every((h, i) => itemHashes[i] === h);
+                }
+                entry.lastFrameHeaderHash = headerHash;
+                entry.lastFrameItemHashes = itemHashes;
+            } catch {}
 
             // Re-check abort after acquire/warmup — narrow window where
             // externalSignal could fire between successful acquire and
@@ -2606,7 +790,7 @@ export async function sendViaWebSocket({
                 const reason = externalSignal.reason;
                 throw reason instanceof Error ? reason : new Error('Aborted');
             }
-            await _sendFrameFn(entry, frame);
+            await _sendFrameFn(entry, wireFrame);
 
             if (process.env.MIXDOG_DEBUG_AGENT) {
                 process.stderr.write(`[agent-trace] ws-streaming-start sinceAcquire=${Date.now() - handshakeStart}ms\n`);
@@ -2748,6 +932,11 @@ export async function sendViaWebSocket({
             entry.lastInputPrefixHash = null;
         }
 
+        // Cache observation must see the MAIN request's usage only. Folding
+        // warmup usage in first (R18) made prompt_tokens spike on it=1 and
+        // then "shrink" on it=2, faking prefix-rewrite/cache-drop signals in
+        // every warmup session (debugger 2026-07-03).
+        const cacheObservation = _cacheObservation({ entry, result });
         if (warmupResult?.usage) {
             result.usage = _combineUsageWithWarmup(result.usage, warmupResult.usage);
         }
@@ -2768,11 +957,51 @@ export async function sendViaWebSocket({
             provider: traceProvider,
             serviceTier: responseServiceTier,
         });
+        const requestHasPreviousResponseId = typeof frame.previous_response_id === 'string' && frame.previous_response_id.length > 0;
+        const transportCacheKeyHash = cacheKey
+            ? createHash('sha256').update(String(cacheKey)).digest('hex').slice(0, 12)
+            : null;
+        if (cacheObservation.actualMiss) {
+            try {
+                appendAgentTrace({
+                    sessionId: poolKey,
+                    iteration,
+                    kind: 'cache_miss',
+                    provider: traceProvider,
+                    model: liveModel,
+                    transport: 'websocket',
+                    payload: {
+                        provider: traceProvider,
+                        model: liveModel,
+                        transport: 'websocket',
+                        ws_mode: mode,
+                        reason: cacheObservation.missReason || 'warm_session_cache_miss',
+                        cached_tokens: cacheObservation.cachedTokens,
+                        prompt_tokens: cacheObservation.promptTokens,
+                        input_tokens: cacheObservation.inputTokens,
+                        uncached_tokens: cacheObservation.uncachedTokens,
+                        cache_ratio: cacheObservation.cacheRatio,
+                        previous_max_cached_tokens: cacheObservation.previousMaxCached,
+                        cache_key_hash: transportCacheKeyHash,
+                        warm_threshold_tokens: cacheObservation.warmThreshold,
+                        prompt_threshold_tokens: cacheObservation.promptThreshold,
+                        drop_ratio: cacheObservation.dropRatio,
+                        drop_threshold_tokens: cacheObservation.dropThreshold,
+                        request_has_previous_response_id: requestHasPreviousResponseId,
+                        chain_delta_reason: mode === 'delta' ? null : deltaReason,
+                        body_input_items: Array.isArray(requestBody.input) ? requestBody.input.length : null,
+                        frame_input_items: Array.isArray(frame.input) ? frame.input.length : null,
+                        response_id: result.responseId || null,
+                    },
+                });
+            } catch {}
+        }
+        entry.promptCacheMaxCachedTokens = Math.max(
+            _num(entry.promptCacheMaxCachedTokens, 0),
+            cacheObservation.cachedTokens,
+        );
         // Extra WS-specific observability: transport + per-iteration delta bytes.
         try {
-            const transportCacheKeyHash = cacheKey
-                ? createHash('sha256').update(String(cacheKey)).digest('hex').slice(0, 12)
-                : null;
             const transportPayload = {
                 provider: traceProvider,
                 transport: 'websocket',
@@ -2789,22 +1018,26 @@ export async function sendViaWebSocket({
                 midstream_retries: attemptIndex,
                 response_id: result.responseId || null,
                 cache_key_hash: transportCacheKeyHash,
-                cache_lane_enabled: promptCacheLane?.enabled === true,
-                cache_lane_key_hash: promptCacheLane?.laneKeyHash || null,
-                cache_lane_rate_policy: promptCacheLane?.ratePolicy || null,
-                cache_lane_max_in_flight: Number.isFinite(Number(promptCacheLane?.maxInFlight)) ? Number(promptCacheLane.maxInFlight) : null,
-                cache_lane_rate_limit_per_min: Number.isFinite(Number(promptCacheLane?.rateLimitPerMin)) ? Number(promptCacheLane.rateLimitPerMin) : null,
-                cache_lane_rate_full_limit_per_min: Number.isFinite(Number(promptCacheLane?.rateFullLimitPerMin)) ? Number(promptCacheLane.rateFullLimitPerMin) : null,
-                cache_lane_rate_delta_limit_per_min: Number.isFinite(Number(promptCacheLane?.rateDeltaLimitPerMin)) ? Number(promptCacheLane.rateDeltaLimitPerMin) : null,
-                cache_lane_rate_wait_ms: Number.isFinite(Number(promptCacheLane?.rateWaitMs)) ? Number(promptCacheLane.rateWaitMs) : null,
-                cache_lane_rate_window_count: Number.isFinite(Number(promptCacheLane?.rateWindowCount)) ? Number(promptCacheLane.rateWindowCount) : null,
-                cache_lane_rate_released_for_wait: promptCacheLane?.rateReleasedForWait === true,
-                cache_lane_rate_reacquire_wait_ms: Number.isFinite(Number(promptCacheLane?.rateReacquireWaitMs)) ? Number(promptCacheLane.rateReacquireWaitMs) : null,
-                cache_lane_wait_ms: Number.isFinite(Number(promptCacheLane?.waitMs)) ? Number(promptCacheLane.waitMs) : null,
-                cache_lane_queued: promptCacheLane?.queued === true,
-                cache_lane_active: Number.isFinite(Number(promptCacheLane?.activeAfterAcquire)) ? Number(promptCacheLane.activeAfterAcquire) : null,
-                cache_lane_queue_depth: Number.isFinite(Number(promptCacheLane?.queueDepthAfterAcquire)) ? Number(promptCacheLane.queueDepthAfterAcquire) : null,
-                request_has_previous_response_id: typeof frame.previous_response_id === 'string' && frame.previous_response_id.length > 0,
+                request_has_previous_response_id: requestHasPreviousResponseId,
+                cached_tokens: cacheObservation.cachedTokens,
+                prompt_tokens: cacheObservation.promptTokens,
+                input_tokens: cacheObservation.inputTokens,
+                uncached_tokens: cacheObservation.uncachedTokens,
+                cache_ratio: cacheObservation.cacheRatio,
+                actual_cache_miss: cacheObservation.actualMiss,
+                actual_cache_miss_reason: cacheObservation.missReason,
+                previous_max_cached_tokens: cacheObservation.previousMaxCached,
+                cache_drop_threshold_tokens: cacheObservation.dropThreshold,
+                frame_prefix_hash: framePrefixHash,
+                frame_prefix_head_hash: framePrefixHeadHash,
+                frame_prefix_prev_match: framePrefixPrevMatch,
+                ws_client_metadata: useCodexWsClientMetadata,
+                ws_client_metadata_key_count: wireFrameMetadataTrace.count,
+                ws_client_metadata_hash: wireFrameMetadataTrace.hash,
+                ws_client_metadata_has_turn_metadata: wireFrameMetadataTrace.hasTurnMetadata,
+                ws_client_metadata_has_thread_id: wireFrameMetadataTrace.hasThreadId,
+                ws_client_metadata_has_turn_state: wireFrameHadTurnState,
+                ws_entry_turn_state_available: useCodexWsClientMetadata && !!entry.turnState,
                 chain_delta_reason: mode === 'delta' ? null : deltaReason,
                 chain_stripped_response_items: strippedResponseItems,
                 chain_skipped_response_items: skippedResponseItems,
@@ -2825,7 +1058,10 @@ export async function sendViaWebSocket({
                 ...transportPayload,
                 payload: transportPayload,
             });
-            if (mode !== 'delta' || deltaReason) {
+            const chainFallback = mode !== 'delta'
+                && deltaReason
+                && !['no_anchor', 'full_forced'].includes(deltaReason);
+            if (chainFallback || (mode === 'delta' && deltaReason)) {
                 appendAgentTrace({
                     sessionId: poolKey,
                     iteration,
@@ -2839,7 +1075,11 @@ export async function sendViaWebSocket({
                         ws_mode: mode,
                         reason: mode === 'delta' ? deltaReason : (deltaReason || 'full_frame'),
                         cache_key_hash: transportCacheKeyHash,
-                        cache_lane_key_hash: promptCacheLane?.laneKeyHash || null,
+                        cached_tokens: cacheObservation.cachedTokens,
+                        prompt_tokens: cacheObservation.promptTokens,
+                        uncached_tokens: cacheObservation.uncachedTokens,
+                        cache_ratio: cacheObservation.cacheRatio,
+                        actual_cache_miss: cacheObservation.actualMiss,
                         request_has_previous_response_id: transportPayload.request_has_previous_response_id,
                         chain_stripped_response_items: strippedResponseItems,
                         chain_skipped_response_items: skippedResponseItems,
@@ -2876,26 +1116,4 @@ export async function sendViaWebSocket({
     }
     // Unreachable — the loop either returns or throws above.
     throw _stampTool(_stampLiveText(firstAttemptError || new Error('sendViaWebSocket: unreachable')));
-    });
-}
-
-// Drain-complete fence — set true once _closeAllPooledSockets runs so any
-// in-flight acquire that resumes after drain throws instead of pushing a
-// fresh socket into the cleared pool. Single-set, process-lifetime invariant.
-let _drainComplete = false;
-
-// Drain hook — self-registered exit drain.
-// Force-closes pooled sockets and fences subsequent acquires.
-// `drainOpenaiWsPool` alias matches the registry's `drain*` naming convention;
-// `_closeAllPooledSockets` kept for backward compat with existing call sites.
-export function _closeAllPooledSockets(reason = 'shutdown') {
-    _drainComplete = true;
-    for (const arr of _wsPool.values()) {
-        for (const entry of arr) {
-            try { entry.socket.close(1000, reason); } catch {}
-        }
-    }
-    _wsPool.clear();
 }
-export const drainOpenaiWsPool = _closeAllPooledSockets;
-process.on('exit', drainOpenaiWsPool);