@zhixuan92/multi-model-agent-core 3.12.2 → 3.12.4

package/dist/runners/openai-usage-interceptor.js

@@ -0,0 +1,121 @@
+// HTTP-level usage capture for openai-compatible providers, independent of
+// @openai/agents' internal aggregation.
+//
+// Why this exists: the @openai/agents SDK's stream consumer
+// (node_modules/@openai/agents-openai/dist/openaiChatCompletionsStreaming.js
+// line ~38) overwrites its local `usage` variable on every SSE chunk:
+//
+//     usage = chunk.usage || undefined;
+//
+// For OpenAI proper this is benign — only the final `[DONE]`-adjacent chunk
+// carries usage, so the last write is the right one. For DeepSeek (and
+// likely other openai-compatible providers) with multi-turn tool-use, later
+// chunks can have `usage:undefined` AFTER an earlier chunk reported real
+// numbers, wiping the captured usage. The SDK then ends with
+// `state.usage.inputTokens=0`, costUSD=0, despite real tokens having
+// flowed. mma 3.12.2 telemetry showed every DeepSeek-as-reviewer call
+// logging 21+ turns and zero tokens, untraceable until the SDK source was
+// inspected.
+//
+// Fix: wrap the OpenAI client's `chat.completions.create` so we see the
+// raw HTTP response (or the raw SSE stream) BEFORE the SDK consumes it, and
+// accumulate usage into an out-of-band counter the runner can read at
+// result-time as a source-of-truth fallback. The wrapper does not modify
+// SDK-visible behavior — it observes only.
+function addUsage(acc, u) {
+    return {
+        promptTokens: acc.promptTokens + (u.prompt_tokens ?? 0),
+        completionTokens: acc.completionTokens + (u.completion_tokens ?? 0),
+        cachedReadTokens: acc.cachedReadTokens + (u.prompt_tokens_details?.cached_tokens ?? 0),
+        reasoningTokens: acc.reasoningTokens + (u.completion_tokens_details?.reasoning_tokens ?? 0),
+        responses: acc.responses + 1,
+    };
+}
+/** Wrap an OpenAI client so usage from every chat.completions.create response
+ *  is captured into the returned UsageAccumulator.
+ *
+ *  Mutates `client.chat.completions.create` in place. The wrapped client is
+ *  otherwise unchanged — same return shapes, same error behavior. The SDK
+ *  sees the same response stream byte-for-byte (we tee, we don't transform).
+ */
+export function wrapClientForUsageCapture(client) {
+    let acc = {
+        promptTokens: 0,
+        completionTokens: 0,
+        cachedReadTokens: 0,
+        reasoningTokens: 0,
+        responses: 0,
+    };
+    let observedNonZero = false;
+    const observe = (u) => {
+        if (!u)
+            return;
+        const before = acc;
+        acc = addUsage(acc, u);
+        // hasObservedUsage flips true the first time a response carries real
+        // input or output tokens. Counters can stay at 0 for empty responses
+        // (auth failures, immediate refusals) — those don't count as evidence
+        // that the provider IS reporting usage.
+        if ((u.prompt_tokens ?? 0) > 0 || (u.completion_tokens ?? 0) > 0) {
+            observedNonZero = true;
+        }
+        void before;
+    };
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    const completions = client.chat.completions;
+    const originalCreate = completions.create.bind(completions);
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    completions.create = async function wrappedCreate(...args) {
+        const result = await originalCreate(...args);
+        // Non-streaming branch: result is a ChatCompletion object with usage.
+        // Detect via the presence of `choices` (eagerly resolved) and `usage`.
+        if (result && typeof result === 'object' && Array.isArray(result.choices)) {
+            observe(result.usage);
+            return result;
+        }
+        // Streaming branch: result is an AsyncIterable of chunks. We wrap the
+        // iterator so each yielded chunk's usage feeds the accumulator without
+        // changing what the SDK sees.
+        if (result && typeof result[Symbol.asyncIterator] === 'function') {
+            const inner = result;
+            // Build a thin wrapper that preserves the original async-iterable
+            // identity (so the SDK can still .controller-abort if it does), but
+            // observes usage on every chunk that has it. Note: we accumulate the
+            // FIRST non-null usage we see from each request (chunks-with-usage are
+            // typically only 1 per stream, sent on the [DONE] frame). If a
+            // provider sends multiple usage frames, we add each one — that matches
+            // the cumulative semantics.
+            const wrapped = {
+                async *[Symbol.asyncIterator]() {
+                    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+                    let lastUsage;
+                    for await (const chunk of inner) {
+                        if (chunk && chunk.usage) {
+                            lastUsage = chunk.usage;
+                        }
+                        yield chunk;
+                    }
+                    // Accumulate at end-of-stream so multiple intra-stream `usage`
+                    // chunks (DeepSeek behavior) don't double-count. The last seen
+                    // usage is the cumulative-for-this-request value the provider
+                    // wanted us to see.
+                    observe(lastUsage);
+                },
+            };
+            // Some SDK call sites poke at `.controller` on the stream object for
+            // abort handling. Forward it if present.
+            const ctrl = result.controller;
+            if (ctrl !== undefined)
+                wrapped.controller = ctrl;
+            return wrapped;
+        }
+        // Unknown shape (shouldn't happen for chat.completions.create) — pass
+        // through unmodified.
+        return result;
+    };
+    return {
+        snapshot: () => ({ ...acc }),
+        hasObservedUsage: () => observedNonZero,
+    };
+}
+//# sourceMappingURL=openai-usage-interceptor.js.map

package/dist/runners/openai-usage-interceptor.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"openai-usage-interceptor.js","sourceRoot":"","sources":["../../src/runners/openai-usage-interceptor.ts"],"names":[],"mappings":"AAAA,2EAA2E;AAC3E,wCAAwC;AACxC,EAAE;AACF,4DAA4D;AAC5D,6EAA6E;AAC7E,sEAAsE;AACtE,EAAE;AACF,wCAAwC;AACxC,EAAE;AACF,4EAA4E;AAC5E,uEAAuE;AACvE,4EAA4E;AAC5E,yEAAyE;AACzE,6DAA6D;AAC7D,qEAAqE;AACrE,sEAAsE;AACtE,0EAA0E;AAC1E,aAAa;AACb,EAAE;AACF,wEAAwE;AACxE,4EAA4E;AAC5E,sEAAsE;AACtE,yEAAyE;AACzE,2CAA2C;AA4B3C,SAAS,QAAQ,CAAC,GAAkB,EAAE,CAAY;IAChD,OAAO;QACL,YAAY,EAAE,GAAG,CAAC,YAAY,GAAG,CAAC,CAAC,CAAC,aAAa,IAAI,CAAC,CAAC;QACvD,gBAAgB,EAAE,GAAG,CAAC,gBAAgB,GAAG,CAAC,CAAC,CAAC,iBAAiB,IAAI,CAAC,CAAC;QACnE,gBAAgB,EAAE,GAAG,CAAC,gBAAgB,GAAG,CAAC,CAAC,CAAC,qBAAqB,EAAE,aAAa,IAAI,CAAC,CAAC;QACtF,eAAe,EAAE,GAAG,CAAC,eAAe,GAAG,CAAC,CAAC,CAAC,yBAAyB,EAAE,gBAAgB,IAAI,CAAC,CAAC;QAC3F,SAAS,EAAE,GAAG,CAAC,SAAS,GAAG,CAAC;KAC7B,CAAC;AACJ,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,yBAAyB,CAAC,MAAc;IACtD,IAAI,GAAG,GAAkB;QACvB,YAAY,EAAE,CAAC;QACf,gBAAgB,EAAE,CAAC;QACnB,gBAAgB,EAAE,CAAC;QACnB,eAAe,EAAE,CAAC;QAClB,SAAS,EAAE,CAAC;KACb,CAAC;IACF,IAAI,eAAe,GAAG,KAAK,CAAC;IAE5B,MAAM,OAAO,GAAG,CAAC,CAA+B,EAAQ,EAAE;QACxD,IAAI,CAAC,CAAC;YAAE,OAAO;QACf,MAAM,MAAM,GAAG,GAAG,CAAC;QACnB,GAAG,GAAG,QAAQ,CAAC,GAAG,EAAE,CAAC,CAAC,CAAC;QACvB,qEAAqE;QACrE,qEAAqE;QACrE,sEAAsE;QACtE,wCAAwC;QACxC,IAAI,CAAC,CAAC,CAAC,aAAa,IAAI,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,iBAAiB,IAAI,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC;YACjE,eAAe,GAAG,IAAI,CAAC;QACzB,CAAC;QACD,KAAK,MAAM,CAAC;IACd,CAAC,CAAC;IAEF,8DAA8D;IAC9D,MAAM,WAAW,GAAG,MAAM,CAAC,IAAI,CAAC,WAAkB,CAAC;IACnD,MAAM,cAAc,GAAG,WAAW,CAAC,MAAM,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC;IAE5D,8DAA8D;IAC9D,WAAW,CAAC,MAAM,GAAG,KAAK,UAAU,aAAa,CAAC,GAAG,IAAW;QAC9D,MAAM,MAAM,GAAG,MAAM,cAAc,CAAC,GAAG,IAAI,CAAC,CAAC;QAE7C,sEAAsE;QACtE,uEAAuE;QACvE,IAAI,MAAM,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,OAAO,CAAC,EAAE,CAAC;YAC1E,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;YACtB,OAAO,MAAM,CAAC;QAChB,CAAC;QAED,sEAAsE;QACtE,uEAAuE;QACvE,8BAA8B;QAC9B,IAAI,MAAM,IAAI,OAAQ,MAAiC,CAAC,MAAM,CAAC,aAAa,CAAC,KAAK,UAAU,EAAE,CAAC;YAC7F,MAAM,KAAK,GAAG,MAAqD,CAAC;YACpE,kEAAkE;YAClE,oEAAoE;YACpE,qEAAqE;YACrE,uEAAuE;YACvE,+DAA+D;YAC/D,uEAAuE;YACvE,4BAA4B;YAC5B,MAAM,OAAO,GAAsD;gBACjE,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,aAAa,CAAC;oBAC3B,8DAA8D;oBAC9D,IAAI,SAAuC,CAAC;oBAC5C,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,KAAK,EAAE,CAAC;wBAChC,IAAI,KAAK,IAAI,KAAK,CAAC,KAAK,EAAE,CAAC;4BACzB,SAAS,GAAG,KAAK,CAAC,KAAK,CAAC;wBAC1B,CAAC;wBACD,MAAM,KAAK,CAAC;oBACd,CAAC;oBACD,+DAA+D;oBAC/D,+DAA+D;oBAC/D,8DAA8D;oBAC9D,oBAAoB;oBACpB,OAAO,CAAC,SAAS,CAAC,CAAC;gBACrB,CAAC;aACF,CAAC;YACF,qEAAqE;YACrE,yCAAyC;YACzC,MAAM,IAAI,GAAI,MAAmC,CAAC,UAAU,CAAC;YAC7D,IAAI,IAAI,KAAK,SAAS;gBAAE,OAAO,CAAC,UAAU,GAAG,IAAI,CAAC;YAClD,OAAO,OAAO,CAAC;QACjB,CAAC;QAED,sEAAsE;QACtE,sBAAsB;QACtB,OAAO,MAAM,CAAC;IAChB,CAAC,CAAC;IAEF,OAAO;QACL,QAAQ,EAAE,GAAG,EAAE,CAAC,CAAC,EAAE,GAAG,GAAG,EAAE,CAAC;QAC5B,gBAAgB,EAAE,GAAG,EAAE,CAAC,eAAe;KACxC,CAAC;AACJ,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zhixuan92/multi-model-agent-core",
-  "version": "3.12.2",
+  "version": "3.12.4",
   "type": "module",
   "license": "MIT",
   "description": "Core library for multi-model-agent: provider runners (Claude, Codex, OpenAI-compatible), routing logic, config schema, and tool/sandbox primitives.",