fix(ai): remember host session model for opencodeModel inherit

.github/i18n/README.ar.md

@@ -450,7 +450,7 @@ memory({ mode: "import", inputPath: "./memories.json" });
 
 إذا عُيِّن `opencodeProvider` و`opencodeModel`، يتقدمان على إعدادات `memoryProvider` اليدوية أدناه.
 
-**اتبع نموذج الجلسة:** عيّن `"opencodeModel": "inherit"` لاستخدام نموذج OpenCode ملموس وقت الاستدعاء بدل معرّف مثبت. لـ **الالتقاط التلقائي**، تُسجَّل كل مطالبة عبر خطاف `chat.params` ويعيد طلب الالتقاط استخدام مزود/نموذج تلك المطالبة. لـ **تعلّم الملف** ومسارات المخرجات المنظمة الأخرى (غير المرتبطة برسالة مستخدم واحدة)، يرجع `inherit` إلى أحدث نموذج في قائمة `model.json` الحديثة في OpenCode (مفضّلًا `opencodeProvider` المُعَد). إرسال معرّف النموذج الحرفي `inherit` غير صالح أبدًا وتسبب سابقًا في `ProviderModelNotFoundError: Model not found: <provider>/inherit` على تلك المسارات. ما زال `opencodeProvider` مطلوبًا كبوابة الإعداد العادية.
+**اتبع نموذج الجلسة:** عيّن `"opencodeModel": "inherit"` لاستخدام نموذج OpenCode ملموس وقت الاستدعاء بدل معرّف مثبت. لـ **الالتقاط التلقائي**، تُسجَّل كل مطالبة عبر خطاف `chat.params` ويعيد طلب الالتقاط استخدام مزود/نموذج تلك المطالبة. لـ **تعلّم الملف** ومسارات المخرجات المنظمة الأخرى (غير المرتبطة برسالة مستخدم واحدة)، يعيد `inherit` أولاً استخدام آخر نموذج أبلغ به المضيف عبر `chat.params` / خطاف سياق V2، ثم يرجع إلى قائمة `model.json` الحديثة في OpenCode إن وُجدت (مفضّلًا `opencodeProvider` المُعَد). OpenCode v2 لم يعد يكتب `model.json`، لذا نموذج الجلسة داخل العملية هو الاحتياطي الأساسي هناك. إرسال معرّف النموذج الحرفي `inherit` غير صالح أبدًا وتسبب سابقًا في `ProviderModelNotFoundError: Model not found: <provider>/inherit` على تلك المسارات. ما زال `opencodeProvider` مطلوبًا كبوابة الإعداد العادية.
 
 **الرجوع:** إعداد API يدوي (إن لم تستخدم opencodeProvider):
 

.github/i18n/README.de.md

@@ -451,7 +451,7 @@ Unterstützte Provider: jeder von `opencode providers list` gelistete Provider (
 
 Wenn `opencodeProvider` und `opencodeModel` gesetzt sind, haben sie Vorrang vor den manuellen `memoryProvider`-Einstellungen unten.
 
-**Sitzungsmodell folgen:** `"opencodeModel": "inherit"` setzen, um zur Aufrufzeit ein konkretes OpenCode-Modell statt einer gepinnten ID zu nutzen. Für **Auto-Capture** wird jeder Prompt über den `chat.params`-Hook aufgezeichnet und der Capture-Request wiederverwendet Provider/Modell dieses Prompts. Für **Profil-Lernen** und andere Structured-Output-Pfade (die nicht an eine einzelne User-Nachricht gebunden sind) fällt `inherit` auf das neueste Modell in der Recent-Liste von OpenCodes `model.json` zurück (bevorzugt den konfigurierten `opencodeProvider`). Das wörtliche Modell-ID `inherit` zu senden ist nie gültig und verursachte zuvor `ProviderModelNotFoundError: Model not found: <provider>/inherit` auf diesen Pfaden. `opencodeProvider` bleibt als normale Config-Gate erforderlich.
+**Sitzungsmodell folgen:** `"opencodeModel": "inherit"` setzen, um zur Aufrufzeit ein konkretes OpenCode-Modell statt einer gepinnten ID zu nutzen. Für **Auto-Capture** wird jeder Prompt über den `chat.params`-Hook aufgezeichnet und der Capture-Request wiederverwendet Provider/Modell dieses Prompts. Für **Profil-Lernen** und andere Structured-Output-Pfade (die nicht an eine einzelne User-Nachricht gebunden sind) nutzt `inherit` zuerst das zuletzt vom Host über `chat.params` / den V2-Context-Hook gemeldete Modell und fällt danach auf die Recent-Liste in OpenCodes `model.json` zurück, falls vorhanden (bevorzugt den konfigurierten `opencodeProvider`). OpenCode v2 schreibt `model.json` nicht mehr; dort ist das In-Process-Sitzungsmodell der primäre Fallback. Das wörtliche Modell-ID `inherit` zu senden ist nie gültig und verursachte zuvor `ProviderModelNotFoundError: Model not found: <provider>/inherit` auf diesen Pfaden. `opencodeProvider` bleibt als normale Config-Gate erforderlich.
 
 **Fallback:** Manuelle API-Konfiguration (wenn opencodeProvider nicht genutzt wird):
 

.github/i18n/README.nl.md

@@ -452,7 +452,7 @@ Ondersteunde providers: elke provider uit `opencode providers list` (bijv. `anth
 
 Als `opencodeProvider` en `opencodeModel` zijn ingesteld, hebben ze voorrang op de handmatige `memoryProvider`-instellingen hieronder.
 
-**Volg het sessiermodel:** stel `"opencodeModel": "inherit"` in om op call-tijd een concreet OpenCode-model te gebruiken in plaats van een vastgezette id. Voor **auto-capture** wordt elke prompt via de `chat.params`-hook vastgelegd en hergebruikt het capture-verzoek provider/model van die prompt. Voor **profiel-leren** en andere structured-output-paden (die niet aan één gebruikersbericht zijn gebonden) valt `inherit` terug op het meest recente model in OpenCode’s `model.json`-recentelijst (met voorkeur voor de geconfigureerde `opencodeProvider`). Het letterlijke model-id `inherit` versturen is nooit geldig en veroorzaakte eerder `ProviderModelNotFoundError: Model not found: <provider>/inherit` op die paden. `opencodeProvider` blijft vereist als normale configgate.
+**Volg het sessiermodel:** stel `"opencodeModel": "inherit"` in om op call-tijd een concreet OpenCode-model te gebruiken in plaats van een vastgezette id. Voor **auto-capture** wordt elke prompt via de `chat.params`-hook vastgelegd en hergebruikt het capture-verzoek provider/model van die prompt. Voor **profiel-leren** en andere structured-output-paden (die niet aan één gebruikersbericht zijn gebonden) hergebruikt `inherit` eerst het laatste model dat de host via `chat.params` / de V2-context-hook meldde, en valt daarna terug op OpenCode’s `model.json`-recentelijst indien aanwezig (met voorkeur voor de geconfigureerde `opencodeProvider`). OpenCode v2 schrijft `model.json` niet meer; daar is het in-process-sessiemodel de primaire fallback. Het letterlijke model-id `inherit` versturen is nooit geldig en veroorzaakte eerder `ProviderModelNotFoundError: Model not found: <provider>/inherit` op die paden. `opencodeProvider` blijft vereist als normale configgate.
 
 **Fallback:** Handmatige API-configuratie (als je geen opencodeProvider gebruikt):
 

.github/i18n/README.tr.md

@@ -452,7 +452,7 @@ Desteklenen sağlayıcılar: `opencode providers list` tarafından listelenen he
 
 `opencodeProvider` ve `opencodeModel` ayarlanmışsa, aşağıdaki manuel `memoryProvider` ayarlarına göre önceliklidirler.
 
-**Oturum modelini izle:** çağrı anında sabitlenmiş bir kimlik yerine somut bir OpenCode modeli kullanmak için `"opencodeModel": "inherit"` ayarlayın. **Otomatik yakalama** için her prompt `chat.params` kancası üzerinden kaydedilir ve yakalama isteği o prompt’un sağlayıcı/modelini yeniden kullanır. **Profil öğrenme** ve diğer yapılandırılmış çıktı yolları için (tek bir kullanıcı mesajına bağlı değildir), `inherit` OpenCode’un `model.json` son liste içindeki en son modele düşer (yapılandırılmış `opencodeProvider` tercih edilir). Literâl model kimliği `inherit` göndermek asla geçerli değildir ve daha önce bu yollarda `ProviderModelNotFoundError: Model not found: <provider>/inherit`’e neden olmuştur. `opencodeProvider` yine de normal yapılandırma kapısı olarak gereklidir.
+**Oturum modelini izle:** çağrı anında sabitlenmiş bir kimlik yerine somut bir OpenCode modeli kullanmak için `"opencodeModel": "inherit"` ayarlayın. **Otomatik yakalama** için her prompt `chat.params` kancası üzerinden kaydedilir ve yakalama isteği o prompt’un sağlayıcı/modelini yeniden kullanır. **Profil öğrenme** ve diğer yapılandırılmış çıktı yolları için (tek bir kullanıcı mesajına bağlı değildir), `inherit` önce host’un `chat.params` / V2 context kancasında bildirdiği son modeli yeniden kullanır; yoksa OpenCode’un `model.json` son listesine düşer (yapılandırılmış `opencodeProvider` tercih edilir). OpenCode v2 artık `model.json` yazmaz; orada birincil geri dönüş süreç içi oturum modelidir. Literâl model kimliği `inherit` göndermek asla geçerli değildir ve daha önce bu yollarda `ProviderModelNotFoundError: Model not found: <provider>/inherit`’e neden olmuştur. `opencodeProvider` yine de normal yapılandırma kapısı olarak gereklidir.
 
 **Yedek:** Manuel API yapılandırması (opencodeProvider kullanılmıyorsa):
 

.github/i18n/README.zh.md

@@ -446,7 +446,7 @@ memory({ mode: "import", inputPath: "./memories.json" });
 
 若设置了 `opencodeProvider` 与 `opencodeModel`，它们优先于下方的手动 `memoryProvider` 设置。
 
-**跟随会话模型：** 设置 `"opencodeModel": "inherit"` 以在调用时使用具体的 OpenCode 模型，而非固定 id。对于**自动捕获**，每个提示通过 `chat.params` 钩子记录，捕获请求重用该提示的提供商/模型。对于**画像学习**及其他结构化输出路径（不绑定到单条用户消息），`inherit` 回退到 OpenCode `model.json` 最近列表中的最新模型（优先使用已配置的 `opencodeProvider`）。发送字面模型 id `inherit` 永远无效，此前会在这些路径上导致 `ProviderModelNotFoundError: Model not found: <provider>/inherit`。`opencodeProvider` 仍作为正常配置门控是必需的。
+**跟随会话模型：** 设置 `"opencodeModel": "inherit"` 以在调用时使用具体的 OpenCode 模型，而非固定 id。对于**自动捕获**，每个提示通过 `chat.params` 钩子记录，捕获请求重用该提示的提供商/模型。对于**画像学习**及其他结构化输出路径（不绑定到单条用户消息），`inherit` 先复用主机在 `chat.params` / V2 context 钩子上报的最近模型，再回退到 OpenCode `model.json` 最近列表（若存在；优先使用已配置的 `opencodeProvider`）。OpenCode v2 不再写入 `model.json`，因此进程内会话模型是主要回退。发送字面模型 id `inherit` 永远无效，此前会在这些路径上导致 `ProviderModelNotFoundError: Model not found: <provider>/inherit`。`opencodeProvider` 仍作为正常配置门控是必需的。
 
 **回退：** 手动 API 配置（若不使用 opencodeProvider）：
 

README.md

@@ -494,7 +494,7 @@ If `autoCaptureEnabled` is `true` (default) but neither `opencodeProvider`/`open
 
 If `opencodeProvider` and `opencodeModel` are set, they take precedence over the manual `memoryProvider` settings below by default (`"autoCaptureProviderOrder": "opencode-first"`). Set `"autoCaptureProviderOrder": "external-first"` to prefer the manual API when both are configured (useful when the host session API is busy/slow); the other channel is used as fallback. Lower `"opencodeTimeoutMs"` to fail over from the opencode channel sooner.
 
-**Follow the session model:** set `"opencodeModel": "inherit"` to use a concrete OpenCode model at call time instead of a pinned id. For **auto-capture**, each prompt is recorded via the `chat.params` hook and the capture request reuses that prompt's provider/model. For **profile learning** and other structured-output paths (which are not tied to a single user message), `inherit` falls back to the most recent model in OpenCode's `model.json` recent list (preferring the configured `opencodeProvider`). Sending the literal model id `inherit` is never valid and previously caused `ProviderModelNotFoundError: Model not found: <provider>/inherit` on those paths. `opencodeProvider` is still required as the normal config gate.
+**Follow the session model:** set `"opencodeModel": "inherit"` to use a concrete OpenCode model at call time instead of a pinned id. For **auto-capture**, each prompt is recorded via the `chat.params` hook and the capture request reuses that prompt's provider/model. For **profile learning** and other structured-output paths (which are not tied to a single user message), `inherit` first reuses the last model the host reported on `chat.params` / the V2 context hook, then falls back to OpenCode's `model.json` recent list when present (preferring the configured `opencodeProvider`). OpenCode v2 no longer writes `model.json`, so the in-process session model is the primary fallback there. Sending the literal model id `inherit` is never valid and previously caused `ProviderModelNotFoundError: Model not found: <provider>/inherit` on those paths. `opencodeProvider` is still required as the normal config gate.
 
 **Fallback:** Manual API configuration (if not using opencodeProvider):
 

src/index.ts

@@ -27,6 +27,7 @@ import { loadOpencodeProvider } from "./services/ai/opencode-provider-loader.js"
 import {
   isInternalStructuredSession,
   noteStructuredOutputStep,
+  rememberSessionModel,
   STRUCTURED_OUTPUT_AGENT,
   STRUCTURED_OUTPUT_MAX_STEPS,
   STRUCTURED_OUTPUT_TOOLS,
@@ -832,6 +833,13 @@ export const OpenCodeMemPlugin: Plugin = async (ctx: PluginInput) => {
     "chat.params": async (input, output) => {
       applyStructuredOutputChatParams(input, output);
 
+      const sessionModel = input?.model;
+      if (typeof sessionModel?.providerID === "string" && typeof sessionModel?.id === "string") {
+        // Keep the host-resolved model for inherit on idle paths (profile
+        // learning, dedup, …). OpenCode v2 no longer writes model.json (#453).
+        rememberSessionModel(sessionModel.providerID, sessionModel.id);
+      }
+
       if (!isConfigured() || CONFIG.opencodeModel !== "inherit") return;
 
       try {

src/services/ai/opencode-provider.ts

@@ -377,12 +377,32 @@ export interface StructuredOutputOptions<T> {
   retryCount?: number;
 }
 
+/** Last host-reported session model from `chat.params` / the V2 context hook. */
+let rememberedSessionModel: { providerID: string; modelID: string } | undefined;
+
+/**
+ * Remember the concrete model the host is using for the current session.
+ * OpenCode v2 no longer writes `model.json`, so profile-learning and other
+ * non-prompt structured-output paths need this in-process fallback for
+ * `opencodeModel: "inherit"` (issue #453).
+ */
+export function rememberSessionModel(providerID: string, modelID: string): void {
+  if (!providerID || !modelID || modelID === "inherit") return;
+  rememberedSessionModel = { providerID, modelID };
+}
+
+/** Test helper: drop the in-process session model between cases. */
+export function clearRememberedSessionModelForTests(): void {
+  rememberedSessionModel = undefined;
+}
+
 /**
  * Resolve `opencodeModel: "inherit"` to a concrete provider/model.
  *
- * Prefer an explicit prompt-recorded model (auto-capture path). Otherwise fall
- * back to OpenCode's recent model list so profile-learning / conflict / dedup
- * paths don't send the literal model id "inherit" (ProviderModelNotFoundError).
+ * Prefer an explicit prompt-recorded model (auto-capture path). Otherwise use
+ * the last host-reported session model, then OpenCode's recent `model.json`
+ * list (v1 hosts), so profile-learning / conflict / dedup paths don't send the
+ * literal model id "inherit" (ProviderModelNotFoundError).
  */
 export function resolveOpencodeModelRef(opts: {
   providerID: string;
@@ -397,6 +417,10 @@ export function resolveOpencodeModelRef(opts: {
     return { providerID: opts.prompt.providerId, modelID: opts.prompt.modelId };
   }
 
+  if (rememberedSessionModel) {
+    return rememberedSessionModel;
+  }
+
   const recent = readRecentOpencodeModel(opts.providerID);
   if (recent) return recent;
 

tests/opencode-provider.test.ts

@@ -25,6 +25,9 @@ import {
   STRUCTURED_OUTPUT_PERMISSIONS,
   STRUCTURED_OUTPUT_TOOLS,
   takeStructuredOutputStepAbortError,
+  rememberSessionModel,
+  clearRememberedSessionModelForTests,
+  resolveOpencodeModelRef,
 } from "../src/services/ai/opencode-provider.js";
 
 const schema = z.object({
@@ -1054,9 +1057,11 @@ describe("resolveOpencodeModelRef / inherit", () => {
     const { join } = await import("node:path");
     tempRoot = mkdtempSync(join(tmpdir(), "opencode-mem-inherit-"));
     process.env.XDG_STATE_HOME = tempRoot;
+    clearRememberedSessionModelForTests();
   });
 
   afterEach(async () => {
+    clearRememberedSessionModelForTests();
     if (originalStateHome === undefined) {
       delete process.env.XDG_STATE_HOME;
     } else {
@@ -1069,16 +1074,15 @@ describe("resolveOpencodeModelRef / inherit", () => {
     }
   });
 
-  it("passes concrete model ids through unchanged", async () => {
-    const { resolveOpencodeModelRef } = await import("../src/services/ai/opencode-provider.js");
+  it("passes concrete model ids through unchanged", () => {
     expect(resolveOpencodeModelRef({ providerID: "local", modelID: "h100/qwen:latest" })).toEqual({
       providerID: "local",
       modelID: "h100/qwen:latest",
     });
   });
 
-  it("prefers the recorded prompt model when modelID is inherit", async () => {
-    const { resolveOpencodeModelRef } = await import("../src/services/ai/opencode-provider.js");
+  it("prefers the recorded prompt model when modelID is inherit", () => {
+    rememberSessionModel("opencode-go", "should-not-win");
     expect(
       resolveOpencodeModelRef({
         providerID: "local",
@@ -1091,6 +1095,32 @@ describe("resolveOpencodeModelRef / inherit", () => {
     });
   });
 
+  it("uses the remembered host session model before model.json", async () => {
+    const { mkdirSync, writeFileSync } = await import("node:fs");
+    const { join } = await import("node:path");
+    const stateDir = join(tempRoot!, "opencode");
+    mkdirSync(stateDir, { recursive: true });
+    writeFileSync(
+      join(stateDir, "model.json"),
+      JSON.stringify({
+        recent: [{ providerID: "local", modelID: "from-model-json" }],
+      })
+    );
+    rememberSessionModel("opencode-go", "host-session-model");
+    expect(resolveOpencodeModelRef({ providerID: "local", modelID: "inherit" })).toEqual({
+      providerID: "opencode-go",
+      modelID: "host-session-model",
+    });
+  });
+
+  it("ignores blank or literal inherit values when remembering the session model", () => {
+    rememberSessionModel("opencode-go", "inherit");
+    rememberSessionModel("", "ignored");
+    expect(() => resolveOpencodeModelRef({ providerID: "local", modelID: "inherit" })).toThrow(
+      /no session model was recorded and no recent OpenCode model is available/
+    );
+  });
+
   it("falls back to OpenCode recent model.json when inherit has no prompt model", async () => {
     const { mkdirSync, writeFileSync } = await import("node:fs");
     const { join } = await import("node:path");
@@ -1106,15 +1136,13 @@ describe("resolveOpencodeModelRef / inherit", () => {
       })
     );
 
-    const { resolveOpencodeModelRef } = await import("../src/services/ai/opencode-provider.js");
     expect(resolveOpencodeModelRef({ providerID: "local", modelID: "inherit" })).toEqual({
       providerID: "local",
       modelID: "h100/qwen:latest",
     });
   });
 
-  it("throws a clear error when inherit has no prompt model and no recent list", async () => {
-    const { resolveOpencodeModelRef } = await import("../src/services/ai/opencode-provider.js");
+  it("throws a clear error when inherit has no prompt model and no recent list", () => {
     expect(() => resolveOpencodeModelRef({ providerID: "local", modelID: "inherit" })).toThrow(
       /no session model was recorded and no recent OpenCode model is available/
     );