fix(opencode): honor catalog output limits in V1

packages/opencode/src/provider/transform.ts

@@ -15,7 +15,9 @@ function mimeToModality(mime: string): Modality | undefined {
   return undefined
 }
 
-export const OUTPUT_TOKEN_MAX = 32_000
+export const OUTPUT_TOKEN_MAX = 256_000
+const OUTPUT_TOKEN_FALLBACK = 32_000
+const REASONING_BUDGET_MAX = 32_000
 
 // OpenAI Responses `include` value that returns the encrypted reasoning state
 // needed for stateless multi-turn reasoning (store: false). Hoisted so every
@@ -1479,7 +1481,7 @@ export function providerOptions(model: Provider.Model, options: { [x: string]: a
 }
 
 export function maxOutputTokens(model: Provider.Model, outputTokenMax = OUTPUT_TOKEN_MAX): number {
-  return Math.min(model.limit.output, outputTokenMax) || outputTokenMax
+  return Math.min(model.limit.output > 0 ? model.limit.output : OUTPUT_TOKEN_FALLBACK, outputTokenMax)
 }
 
 type JsonRecord = Record<string, unknown>
@@ -1747,7 +1749,7 @@ function effortVariants(model: Provider.Model, values: readonly unknown[]) {
 }
 
 function budgetVariants(model: Provider.Model, min?: number, max?: number) {
-  const maximum = Math.min(max ?? OUTPUT_TOKEN_MAX - 1, model.limit.output - 1, OUTPUT_TOKEN_MAX - 1)
+  const maximum = Math.min(max ?? REASONING_BUDGET_MAX - 1, model.limit.output - 1, REASONING_BUDGET_MAX - 1)
   if (maximum <= 0) return {}
   const high = Math.min(Math.max(min ?? 0, Math.floor((maximum + 1) / 2)), maximum)
   return Object.fromEntries(

packages/opencode/src/session/llm/request.ts

@@ -16,6 +16,7 @@ import type { Plugin } from "@/plugin"
 import { mergeDeep } from "remeda"
 
 const USER_AGENT = `opencode/${InstallationVersion}`
+const SUMMARY_OUTPUT_MAX = 32_000
 
 type PrepareInput = {
   readonly user: SessionV1.User
@@ -111,6 +112,7 @@ export const prepare = Effect.fn("LLMRequestPrep.prepare")(function* (input: Pre
           ...input.messages,
         ]
 
+  const outputTokenLimit = ProviderTransform.maxOutputTokens(input.model, input.flags.outputTokenMax)
   const params = yield* input.plugin.trigger(
     "chat.params",
     {
@@ -126,7 +128,8 @@ export const prepare = Effect.fn("LLMRequestPrep.prepare")(function* (input: Pre
         : undefined,
       topP: input.agent.topP ?? ProviderTransform.topP(input.model),
       topK: ProviderTransform.topK(input.model),
-      maxOutputTokens: ProviderTransform.maxOutputTokens(input.model, input.flags.outputTokenMax),
+      maxOutputTokens:
+        input.agent.name === "compaction" ? Math.min(outputTokenLimit, SUMMARY_OUTPUT_MAX) : outputTokenLimit,
       options,
     },
   )

packages/opencode/test/provider/transform-output-limit.test.ts

@@ -0,0 +1,28 @@
+import { describe, expect, test } from "bun:test"
+import type { ConfigV1 } from "@opencode-ai/core/v1/config/config"
+import type { Provider } from "@/provider/provider"
+import { ProviderTransform } from "@/provider/transform"
+import { usable } from "@/session/overflow"
+
+const model = (output: number) => ({ limit: { context: 1_000_000, output } }) as Provider.Model
+
+describe("ProviderTransform.maxOutputTokens", () => {
+  test("uses the catalog output limit for LongCat instead of 32K", () => {
+    expect(ProviderTransform.maxOutputTokens(model(131_072))).toBe(131_072)
+    expect(usable({ cfg: {} as ConfigV1.Info, model: model(131_072) })).toBe(1_000_000 - 131_072)
+  })
+
+  test("keeps smaller model limits and the 32K fallback for unknown output limits", () => {
+    expect(ProviderTransform.maxOutputTokens(model(8_192))).toBe(8_192)
+    expect(ProviderTransform.maxOutputTokens(model(0))).toBe(32_000)
+  })
+
+  test("caps exceptionally large catalog limits at the V2 256K ceiling", () => {
+    expect(ProviderTransform.maxOutputTokens(model(384_000))).toBe(256_000)
+  })
+
+  test("preserves the explicit experimental output-token override", () => {
+    expect(ProviderTransform.maxOutputTokens(model(131_072), 16_000)).toBe(16_000)
+    expect(ProviderTransform.maxOutputTokens(model(384_000), 384_000)).toBe(384_000)
+  })
+})

packages/opencode/test/session/llm.test.ts

@@ -824,7 +824,7 @@ describe("session.llm.stream", () => {
   )
 
   it.instance(
-    "sends temperature, tokens, and reasoning options for openai-compatible models",
+    "sends catalog output limits and caps compaction requests for openai-compatible models",
     () =>
       Effect.gen(function* () {
         const fixture = loadFixture(vivgridFixture.providerID, vivgridFixture.modelID)
@@ -840,6 +840,7 @@ describe("session.llm.stream", () => {
           ProviderV2.ID.make(vivgridFixture.providerID),
           ModelV2.ID.make(fixture.model.id),
         )
+        const model = { ...resolved, limit: { ...resolved.limit, output: 131_072 } }
         const sessionID = SessionID.make("session-test-1")
         const agent = {
           name: "test",
@@ -862,7 +863,7 @@ describe("session.llm.stream", () => {
         yield* drain({
           user,
           sessionID,
-          model: resolved,
+          model,
           agent,
           system: ["You are a helpful assistant."],
           messages: [{ role: "user", content: "Hello" }],
@@ -884,11 +885,28 @@ describe("session.llm.stream", () => {
         expect(body.stream).toBe(true)
 
         const maxTokens = (body.max_tokens as number | undefined) ?? (body.max_output_tokens as number | undefined)
-        const expectedMaxTokens = ProviderTransform.maxOutputTokens(resolved)
-        expect(maxTokens).toBe(expectedMaxTokens)
+        expect(maxTokens).toBe(131_072)
 
         const reasoning = (body.reasoningEffort as string | undefined) ?? (body.reasoning_effort as string | undefined)
         expect(reasoning).toBe("high")
+
+        const compaction = waitRequest(
+          "/chat/completions",
+          new Response(createChatStream("Summary"), {
+            status: 200,
+            headers: { "Content-Type": "text/event-stream" },
+          }),
+        )
+        yield* drain({
+          user: { ...user, agent: "compaction" },
+          sessionID,
+          model,
+          agent: { ...agent, name: "compaction" },
+          system: ["Summarize the conversation."],
+          messages: [{ role: "user", content: "Summarize" }],
+          tools: {},
+        })
+        expect((yield* Effect.promise(() => compaction)).body.max_tokens).toBe(32_000)
       }),
     {
       config: () => ({