@bman654/clodex 2.11.6 → 2.12.0

package/README.md

@@ -106,6 +106,9 @@ flowchart LR
 > [!TIP]
 > Using Claude Code's agents view or background agents? Ask your Claude Code agent to read [docs/background-agents.md](docs/background-agents.md) and set it up for you — one global `clodex server --proxy` plus the `clodex-claude` wrapper bin bridges every claude process automatically.
 
+> [!TIP]
+> On Windows with the Claude Code VS Code extension? See [docs/windows-setup.md](docs/windows-setup.md) — proxy-mode env vars route the extension through clodex, and an optional process wrapper makes clodex models appear in its model picker.
+
 ## CLI reference
 
 ### `clodex claude [options] [claude-flags]`

package/dist/cli.js

@@ -382,7 +382,7 @@ import { join } from "path";
 // package.json
 var package_default = {
   name: "@bman654/clodex",
-  version: "2.11.6",
+  version: "2.12.0",
   publishConfig: {
     access: "public"
   },
@@ -8717,8 +8717,8 @@ var FAILURE_EVENT_TYPES = /* @__PURE__ */ new Set(["error", "response.failed", "
 var RESPONSES_WS_HARD_TTL_MS = 55 * 6e4;
 var RESPONSES_WS_IDLE_TTL_MS = 30 * 6e4;
 var RESPONSES_WS_NURSERY_IDLE_TTL_MS = 5 * 6e4;
-var RESPONSES_WS_MAX_CONNECTIONS = 32;
-var RESPONSES_WS_MAX_NURSERY_CONNECTIONS = 8;
+var RESPONSES_WS_MAX_CONNECTIONS = 64;
+var RESPONSES_WS_MAX_NURSERY_CONNECTIONS = 48;
 var diagnosticContext = new AsyncLocalStorage();
 function withResponsesWebSocketDiagnosticContext(context, fn) {
   return diagnosticContext.run(context, fn);
@@ -8847,7 +8847,7 @@ function instructionChangeSummary(previous, current) {
   const firstDiffLine = previous.slice(0, prefix).split("\n").length;
   return `instructions changed: previous_chars=${previous.length} current_chars=${current.length} common_prefix_chars=${prefix} common_suffix_chars=${suffix} first_diff_line=${firstDiffLine}`;
 }
-function responsesWebSocketPartitionKey(wsUrl, payload, options = {}, authorizationFingerprint = "") {
+function responsesWebSocketPartitionKey(wsUrl, payload, options = {}, authorizationFingerprint = "", claudeAgentId = "") {
   const promptCacheKey = payload.prompt_cache_key;
   const model = payload.model;
   if (typeof promptCacheKey !== "string" || !promptCacheKey || typeof model !== "string" || !model) return void 0;
@@ -8860,7 +8860,8 @@ function responsesWebSocketPartitionKey(wsUrl, payload, options = {}, authorizat
     model,
     effort,
     promptCacheKey,
-    authorizationFingerprint
+    authorizationFingerprint,
+    claudeAgentId
   ].join("");
   return createHash6("sha256").update(material).digest("hex");
 }
@@ -9100,13 +9101,16 @@ function mismatchDumpLine(items, index) {
 function canonicalItemStrings(items) {
   return items.map((item) => canonicalJson(normalizeToolCallJson([item])));
 }
-function isStrictPrefix(head, client) {
-  if (client.length <= head.length) return false;
+function isPrefixOrEqual(head, client) {
+  if (client.length < head.length) return false;
   for (let index = 0; index < head.length; index += 1) {
     if (head[index] !== client[index]) return false;
   }
   return true;
 }
+function isStrictPrefix(head, client) {
+  return client.length > head.length && isPrefixOrEqual(head, client);
+}
 function continuationMatch(entry, payload, clientItems) {
   if (!entry.responseId || !entry.requestInput || !entry.expectedAssistant) return void 0;
   const full = inputArray(payload);
@@ -9599,11 +9603,16 @@ function evictOldestIdleGeneration(generation, maxConnections, reason) {
   while (connectionCountByGeneration(generation) >= maxConnections && idle.length) {
     const oldest = idle.shift();
     if (oldest) {
+      const idleMs = Math.max(0, oldest.options.now() - oldest.lastUsedAt);
+      oldest.debug(
+        `evicting the oldest idle ${generation} connection to stay within its cap: connection=${oldest.debugId} idle_ms=${idleMs} cap=${maxConnections} reason=${reason}`
+      );
       evictions.push({
         connectionId: oldest.debugId,
         partitionKey: oldest.key,
         generation: oldest.generation,
-        reason
+        reason,
+        idleMs
       });
       deleteEntry(oldest);
     }
@@ -10011,28 +10020,42 @@ function createResponsesWebSocketFetch(wsUrl, log12, options = {}) {
     }
     if (hasResponsesLiteHeader(headers)) payload = applyResponsesLiteShape(payload);
     const authorizationFingerprint = authorizationHeaderFingerprint(headers);
+    const diagnosticCorrelation = diagnosticContext.getStore();
+    const claudeAgentId = diagnosticCorrelation?.claudeAgentId ?? "";
     const partitionKey = responsesWebSocketPartitionKey(
       wsUrl,
       payload,
       options,
-      authorizationFingerprint
+      authorizationFingerprint,
+      claudeAgentId
     );
     const promptFingerprint = responsesWebSocketPromptFingerprint(payload);
     const promptFieldHashes = responsesWebSocketPromptFieldHashes(payload);
     const instructionsSnapshot = instructionsFromPayload(payload);
-    const diagnosticCorrelation = diagnosticContext.getStore();
     let now = resolvedOptions.now();
     const evictions = cleanupExpiredConnections(now);
+    let canonicalClientItems;
+    const clientItems = () => canonicalClientItems ??= canonicalItemStrings(inputArray(payload));
     const scanForHeads = () => {
       const scanned = partitionKey ? connectionEntries(partitionKey) : [];
       const idle = scanned.filter((entry) => !entry.inFlight);
-      const clientItems = idle.length ? canonicalItemStrings(inputArray(payload)) : [];
+      const canonical = idle.length ? clientItems() : [];
       return {
         candidates: scanned,
         idleCandidates: idle,
-        matches: idle.map((entry) => ({ entry, match: continuationMatch(entry, payload, clientItems) })).filter((candidate) => candidate.match !== void 0).sort((left, right) => left.match.delta.length - right.match.delta.length || (left.match.mode === right.match.mode ? 0 : left.match.mode === "exact" ? -1 : 1))
+        matches: idle.map((entry) => ({ entry, match: continuationMatch(entry, payload, canonical) })).filter((candidate) => candidate.match !== void 0).sort((left, right) => left.match.delta.length - right.match.delta.length || (left.match.mode === right.match.mode ? 0 : left.match.mode === "exact" ? -1 : 1))
       };
     };
+    const couldPrecedeThisRequest = (entry) => {
+      if (entry.responseId && entry.requestInput && entry.expectedAssistant) {
+        return continuationMatch(entry, payload, clientItems()) !== void 0;
+      }
+      const streaming = entry.current;
+      if (!streaming) return true;
+      streaming.canonicalInput ??= canonicalItemStrings(inputArray(streaming.originalPayload));
+      return isPrefixOrEqual(streaming.canonicalInput, clientItems());
+    };
+    const blockingHead = (entries) => entries.find((entry) => entry.inFlight && couldPrecedeThisRequest(entry));
     let { candidates, idleCandidates, matches } = scanForHeads();
     let selected = matches[0]?.entry;
     let selectedMatch = matches[0]?.match;
@@ -10075,9 +10098,10 @@ function createResponsesWebSocketFetch(wsUrl, log12, options = {}) {
       );
       return "continuation";
     };
+    let arrivalBlockingHead = selected ? void 0 : blockingHead(candidates);
     if (selected && selectedDelta) {
       decision = continueOnHead(selected, selectedMatch);
-    } else if (candidates.some((entry) => entry.inFlight)) {
+    } else if (arrivalBlockingHead) {
       selected = void 0;
       persistent = false;
       decision = "parallel_isolated";
@@ -10176,11 +10200,15 @@ function createResponsesWebSocketFetch(wsUrl, log12, options = {}) {
           pacingRescanOutcome = "no_change";
         }
       }
-      if (!selected && persistent && partitionKey && connectionEntries(partitionKey).some((entry) => entry.inFlight)) {
-        persistent = false;
-        decision = "parallel_isolated";
-        debug("parallel request using an isolated socket after pacing");
-        if (pacingRescanOutcome === "no_change") pacingRescanOutcome = "parallel_isolated";
+      if (!selected && persistent && partitionKey) {
+        const blocked = blockingHead(connectionEntries(partitionKey));
+        if (blocked) {
+          arrivalBlockingHead = blocked;
+          persistent = false;
+          decision = "parallel_isolated";
+          debug("parallel request using an isolated socket after pacing");
+          if (pacingRescanOutcome === "no_change") pacingRescanOutcome = "parallel_isolated";
+        }
       }
     }
     let suppressedMismatchWarnings;
@@ -10211,7 +10239,9 @@ function createResponsesWebSocketFetch(wsUrl, log12, options = {}) {
         accountIdHash: options.accountId ? createHash6("sha256").update(options.accountId).digest("hex").slice(0, 16) : "",
         model: typeof payload.model === "string" ? payload.model : void 0,
         effort: typeof payload.reasoning?.effort === "string" ? String(payload.reasoning.effort).trim().toLowerCase() : "",
-        promptCacheKey: typeof payload.prompt_cache_key === "string" ? payload.prompt_cache_key : void 0
+        promptCacheKey: typeof payload.prompt_cache_key === "string" ? payload.prompt_cache_key : void 0,
+        claudeAgentId: claudeAgentId || void 0,
+        claudeParentAgentId: diagnosticCorrelation?.claudeParentAgentId
       },
       promptFingerprint,
       promptFieldHashes,
@@ -10236,6 +10266,10 @@ function createResponsesWebSocketFetch(wsUrl, log12, options = {}) {
       createdConnectionId: selected ? void 0 : nextConnectionDebugId,
       ...pacingWaitedMs !== void 0 ? { pacingWaitedMs } : {},
       ...pacingRescanOutcome !== void 0 ? { pacingRescanOutcome } : {},
+      // Which busy head this request could not be told apart from. Without it an
+      // isolated turn reads as unexplained, and isolation is the single largest
+      // source of uncached prompt tokens on this transport.
+      ...decision === "parallel_isolated" && arrivalBlockingHead ? { isolatedByConnectionId: arrivalBlockingHead.debugId } : {},
       ...suppressedMismatchWarnings !== void 0 ? { suppressedMismatchWarnings } : {},
       createdGeneration: selected ? void 0 : persistent ? "nursery" : "isolated",
       incrementalInputItems: selectedDelta?.length,
@@ -12105,6 +12139,18 @@ function extractClaudeSessionId(body, headerFallback) {
   }
   return validClaudeSessionId(headerFallback);
 }
+var CLAUDE_AGENT_ID_RE = /^[A-Za-z0-9._-]{1,128}$/;
+function extractClaudeAgentIds(headers) {
+  const read = (name) => {
+    const raw = headers[name];
+    const value = (Array.isArray(raw) ? raw[0] : raw)?.trim();
+    return value && CLAUDE_AGENT_ID_RE.test(value) ? value : void 0;
+  };
+  return {
+    claudeAgentId: read("x-claude-code-agent-id"),
+    claudeParentAgentId: read("x-claude-code-parent-agent-id")
+  };
+}
 function claudeSessionPromptCacheKey(sessionId) {
   return "relay-session-" + createHash8("sha256").update(sessionId).digest("hex").slice(0, 32);
 }
@@ -13316,6 +13362,7 @@ async function startProxyCatalog(routes, defaultAliasId, debug = false, inferenc
         const openAiOAuth = isOpenAiOAuthRoute(route);
         const claudeSessionIdHeader = Array.isArray(req.headers["x-claude-code-session-id"]) ? req.headers["x-claude-code-session-id"][0] : req.headers["x-claude-code-session-id"];
         const claudeSessionId = extractClaudeSessionId(anthropicBody, claudeSessionIdHeader);
+        const claudeAgentIds = extractClaudeAgentIds(req.headers);
         const translationLifecycle = createTranslationLifecycle(
           inferenceLogPath,
           relayRequestId,
@@ -13389,7 +13436,7 @@ async function startProxyCatalog(routes, defaultAliasId, debug = false, inferenc
             keepAlive.unref();
             try {
               await withResponsesWebSocketDiagnosticContext(
-                { requestId: relayRequestId, claudeSessionId },
+                { requestId: relayRequestId, claudeSessionId, ...claudeAgentIds },
                 () => streamAnthropicResponse(
                   model,
                   params,
@@ -13420,7 +13467,7 @@ async function startProxyCatalog(routes, defaultAliasId, debug = false, inferenc
             res.end();
           } else {
             const anthropicResponse = await withResponsesWebSocketDiagnosticContext(
-              { requestId: relayRequestId, claudeSessionId },
+              { requestId: relayRequestId, claudeSessionId, ...claudeAgentIds },
               () => generateAnthropicResponse(
                 model,
                 params,
@@ -18189,6 +18236,7 @@ async function handleAnthropicMessages(req, res, options, modelCache, plog) {
   const requestId = randomUUID5();
   const claudeSessionIdHeader = Array.isArray(req.headers["x-claude-code-session-id"]) ? req.headers["x-claude-code-session-id"][0] : req.headers["x-claude-code-session-id"];
   const claudeSessionId = extractClaudeSessionId(body, claudeSessionIdHeader);
+  const claudeAgentIds = extractClaudeAgentIds(req.headers);
   if (options.webSocketDiagnosticsLogPath) {
     writeWebSocketDiagnosticRequestLog(options.webSocketDiagnosticsLogPath, {
       requestId,
@@ -18353,7 +18401,7 @@ async function handleAnthropicMessages(req, res, options, modelCache, plog) {
             res.write(chunk);
           };
           await withResponsesWebSocketDiagnosticContext(
-            { requestId, claudeSessionId },
+            { requestId, claudeSessionId, ...claudeAgentIds },
             () => streamAnthropicResponse(languageModel, params, responseModelId, writeStreamChunk, void 0, {
               abortSignal: clientAbort.signal,
               initialInputTokens: estimateAnthropicInputTokens(body),
@@ -18369,7 +18417,7 @@ async function handleAnthropicMessages(req, res, options, modelCache, plog) {
           res.end();
         } else {
           const anthropicResponse = await withResponsesWebSocketDiagnosticContext(
-            { requestId, claudeSessionId },
+            { requestId, claudeSessionId, ...claudeAgentIds },
             () => generateAnthropicResponse(languageModel, params, responseModelId, {
               forceStream: openAiOAuth,
               abortSignal: clientAbort.signal,
@@ -19462,6 +19510,9 @@ function forwardToAdapter(req, res, rawBody, adapter, adapterRequest = http2.req
         "Content-Length": String(rawBody.length),
         "x-api-key": adapter.token,
         ...typeof req.headers["x-claude-code-session-id"] === "string" ? { "x-claude-code-session-id": req.headers["x-claude-code-session-id"] } : {},
+        // Subagent identity: the relay partitions ChatGPT WebSocket heads by it.
+        ...typeof req.headers["x-claude-code-agent-id"] === "string" ? { "x-claude-code-agent-id": req.headers["x-claude-code-agent-id"] } : {},
+        ...typeof req.headers["x-claude-code-parent-agent-id"] === "string" ? { "x-claude-code-parent-agent-id": req.headers["x-claude-code-parent-agent-id"] } : {},
         ...lifecycle ? { "x-relay-request-id": lifecycle.requestId } : {}
       }
     }, (upstreamRes) => {