From 2f8ce576dccdc1e1b28ef22da936200652daa8e4 Mon Sep 17 00:00:00 2001
From: YeonGyu-Kim <code.yeon.gyu@gmail.com>
Date: Mon, 11 May 2026 08:47:49 +0900
Subject: [PATCH] fix(chat-params): use conservative token fallback for invalid
 limits

Avoid inflating non-positive maxOutputTokens values to model capability maxima by always falling back to a safe fixed budget.

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-openagent)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
---
 src/plugin/chat-params.test.ts | 46 +++++++++++++++++++++++++++++++---
 src/plugin/chat-params.ts      | 16 +++++++++---
 2 files changed, 54 insertions(+), 8 deletions(-)

diff --git a/src/plugin/chat-params.test.ts b/src/plugin/chat-params.test.ts
index de7acb5df..736e36d21 100644
--- a/src/plugin/chat-params.test.ts
+++ b/src/plugin/chat-params.test.ts
@@ -5,7 +5,7 @@ import { join } from "node:path"
 
 import { createChatParamsHandler, type ChatParamsOutput } from "./chat-params"
 import * as dataPathModule from "../shared/data-path"
-import { writeProviderModelsCache } from "../shared"
+import * as sharedModule from "../shared"
 import {
   clearSessionPromptParams,
   getSessionPromptParams,
@@ -21,13 +21,13 @@ describe("createChatParamsHandler", () => {
     getCacheDirSpy = spyOn(dataPathModule, "getOmoOpenCodeCacheDir").mockReturnValue(
       join(tempCacheRoot, "oh-my-opencode"),
     )
-    writeProviderModelsCache({ connected: [], models: {} })
+    sharedModule.writeProviderModelsCache({ connected: [], models: {} })
   })
 
   afterEach(() => {
     clearSessionPromptParams("ses_chat_params")
     clearSessionPromptParams("ses_chat_params_temperature")
-    writeProviderModelsCache({ connected: [], models: {} })
+    sharedModule.writeProviderModelsCache({ connected: [], models: {} })
     getCacheDirSpy?.mockRestore()
     if (tempCacheRoot) {
       rmSync(tempCacheRoot, { recursive: true, force: true })
@@ -101,7 +101,7 @@ describe("createChatParamsHandler", () => {
 
   test("applies stored prompt params for the session", async () => {
     //#given
-    writeProviderModelsCache({
+    sharedModule.writeProviderModelsCache({
       connected: ["openai"],
       models: {
         openai: [
@@ -256,6 +256,7 @@ describe("createChatParamsHandler", () => {
 
   test("falls back to default maxOutputTokens when stored and compatibility tokens are non-positive", async () => {
     //#given
+    const logSpy = spyOn(sharedModule, "log").mockImplementation(() => undefined)
     setSessionPromptParams("ses_chat_params", {
       maxOutputTokens: 0,
     })
@@ -282,6 +283,43 @@ describe("createChatParamsHandler", () => {
     //#when
     await handler(input, output)
 
+    //#then
+    expect(output.maxOutputTokens).toBe(4096)
+    expect(logSpy).toHaveBeenCalledWith(
+      "[plugin] maxOutputTokens=0 is non-positive; using safe fallback 4096",
+    )
+
+    logSpy.mockRestore()
+  })
+
+  test("uses safe fallback instead of model max when stored maxOutputTokens is non-positive", async () => {
+    //#given
+    setSessionPromptParams("ses_chat_params", {
+      maxOutputTokens: -1,
+    })
+
+    const handler = createChatParamsHandler({
+      anthropicEffort: null,
+    })
+
+    const input = {
+      sessionID: "ses_chat_params",
+      agent: { name: "oracle" },
+      model: { providerID: "openai", modelID: "gpt-5.4" },
+      provider: { id: "openai" },
+      message: {},
+    }
+
+    const output: ChatParamsOutput = {
+      topP: 1,
+      topK: 1,
+      maxOutputTokens: -1,
+      options: {},
+    }
+
+    //#when
+    await handler(input, output)
+
     //#then
     expect(output.maxOutputTokens).toBe(4096)
   })
diff --git a/src/plugin/chat-params.ts b/src/plugin/chat-params.ts
index ac665b62b..26f35d03d 100644
--- a/src/plugin/chat-params.ts
+++ b/src/plugin/chat-params.ts
@@ -1,5 +1,7 @@
 import { getSessionPromptParams } from "../shared/session-prompt-params-state"
-import { getModelCapabilities, resolveCompatibleModelSettings } from "../shared"
+import { getModelCapabilities, log, resolveCompatibleModelSettings } from "../shared"
+
+const SAFE_MAX_OUTPUT_TOKENS_FALLBACK = 4096
 
 export type ChatParamsInput = {
   sessionID: string
@@ -168,9 +170,15 @@ export function createChatParamsHandler(args: {
       if (compatibility.maxTokens !== undefined && compatibility.maxTokens > 0) {
         output.maxOutputTokens = compatibility.maxTokens
       } else {
-        const capabilitiesLimit = capabilities?.maxOutputTokens
-        output.maxOutputTokens =
-          typeof capabilitiesLimit === "number" && capabilitiesLimit > 0 ? capabilitiesLimit : 4096
+        const originalMaxOutputTokens = typeof output.maxOutputTokens === "number"
+          ? output.maxOutputTokens
+          : compatibility.maxTokens
+        output.maxOutputTokens = SAFE_MAX_OUTPUT_TOKENS_FALLBACK
+        if (typeof originalMaxOutputTokens === "number" && originalMaxOutputTokens <= 0) {
+          log(
+            `[plugin] maxOutputTokens=${originalMaxOutputTokens} is non-positive; using safe fallback ${SAFE_MAX_OUTPUT_TOKENS_FALLBACK}`,
+          )
+        }
       }
     }