@vincemakes/kiso-provider-openai 0.36.0 → 0.37.0

package/dist/index.js

@@ -99,12 +99,27 @@ export function createOpenAICompatAdapter(client, adapterOpts = {}) {
             // usage chunk is still accepted after the finish: some compat
             // providers send it late.
             let finishSeen = false;
+            // TRACE-F1: the model the SERVER says it served. Every chunk of an
+            // OpenAI-shaped stream carries `model`; we never read it, so an id
+            // the vendor silently aliases (a retired name, a migration id)
+            // looked identical to the one we asked for. ONE variable, read by
+            // every usage exit below — this adapter has three of them, and two
+            // places computing the same thing is what W22-R1 was.
+            let served = null;
             try {
                 for await (const chunk of stream) {
+                    // before the finishSeen branch below, which continues early.
+                    // FIRST statement wins: a provider that changes the id
+                    // mid-stream is telling us something is wrong, and the first
+                    // answer is the one the request was routed by.
+                    if (served === null && typeof chunk.model === "string" && chunk.model !== "")
+                        served = chunk.model;
                     if (finishSeen) {
                         if (chunk.usage) {
                             usageSent = true;
-                            const details = chunk.usage.prompt_tokens_details;
+                            const u = chunk.usage;
+                            const details = u.prompt_tokens_details;
+                            const reasoning = u.completion_tokens_details?.reasoning_tokens;
                             yield {
                                 seq: 0,
                                 type: "usage",
@@ -113,6 +128,8 @@ export function createOpenAICompatAdapter(client, adapterOpts = {}) {
                                 cacheRead: details?.cached_tokens ?? null,
                                 cacheWrite: null,
                                 known: true,
+                                ...(typeof reasoning === "number" ? { reasoningTokens: reasoning } : {}),
+                                ...(served !== null ? { servedModel: served } : {}),
                             };
                         }
                         continue; // content and finish reasons after the first finish: ignored
@@ -209,7 +226,14 @@ export function createOpenAICompatAdapter(client, adapterOpts = {}) {
                         // prompt_tokens_details — an absent value is null, NEVER
                         // faked as a zero-cache turn. OpenAI does not report a
                         // cache write; null is the honest answer.
-                        const details = chunk.usage.prompt_tokens_details;
+                        const u = chunk.usage;
+                        const details = u.prompt_tokens_details;
+                        // RSN-1: the completion side of the SAME object. We have
+                        // always read the prompt side for cache and never looked
+                        // here, so `canonical.reasoning` sat hardcoded null under a
+                        // comment saying no provider reports a split — true when
+                        // written, false since this vendor shipped one.
+                        const reasoning = u.completion_tokens_details?.reasoning_tokens;
                         yield {
                             seq: 0,
                             type: "usage",
@@ -218,6 +242,8 @@ export function createOpenAICompatAdapter(client, adapterOpts = {}) {
                             cacheRead: details?.cached_tokens ?? null,
                             cacheWrite: null,
                             known: true,
+                            ...(typeof reasoning === "number" ? { reasoningTokens: reasoning } : {}),
+                            ...(served !== null ? { servedModel: served } : {}),
                         };
                     }
                     const fr = chunk.choices?.[0]?.finish_reason;
@@ -255,7 +281,9 @@ export function createOpenAICompatAdapter(client, adapterOpts = {}) {
             if (!usageSent) {
                 // Area 6: no usage reported is expressed as UNKNOWN — nulls
                 // and known:false — never faked as a zero-cost turn.
-                yield { seq: 0, type: "usage", inputTokens: null, outputTokens: null, cacheRead: null, cacheWrite: null, known: false };
+                // known:false and a served model are not in tension: the server
+                // stated what it ran, it just never reported what it cost.
+                yield { seq: 0, type: "usage", inputTokens: null, outputTokens: null, cacheRead: null, cacheWrite: null, known: false, ...(served !== null ? { servedModel: served } : {}) };
             }
             // Area 6 hardening (review finding 4): a stream that ended with
             // NO finish_reason is a TRUNCATED turn — the stop is an explicit

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@vincemakes/kiso-provider-openai",
-	"version": "0.36.0",
+	"version": "0.37.0",
 	"description": "kiso OpenAI-compatible adapter \u2014 OpenAI + the compat family (GLM, Kimi, DeepSeek, OpenRouter) via base_url swap, reasoning dialects digested.",
 	"type": "module",
 	"license": "MIT",
@@ -21,7 +21,7 @@
 		"test": "vitest run"
 	},
 	"dependencies": {
-		"@vincemakes/kiso-core": "0.36.0",
+		"@vincemakes/kiso-core": "0.37.0",
 		"openai": "^7.3.0"
 	},
 	"devDependencies": {