fix(dsh): export a known output limit as maxTokens

src/clients/config-export/dsh.ts

@@ -2,7 +2,7 @@
 import type { ExportModel, ExportContext, ManagedContribution } from "./contracts";
 import type { OcxConfig } from "../../types";
 import { providerCodexAccountMode } from "../../providers/registry";
-import { normalizeExportModels, authoritativeContextWindow, exportModelLabel, singleFragment } from "./model-metadata";
+import { normalizeExportModels, authoritativeContextWindow, exportModelLabel, knownOutputBudget, singleFragment } from "./model-metadata";
 import { OPENCODE_PROVIDER_ID } from "./constants";
 
 
@@ -26,6 +26,11 @@ export interface DshModelEntry {
   name: string;
   input: string[];
   contextWindow?: number;
+  /**
+   * Output ceiling DSH should treat as the model's real `max_tokens`. DSH falls back to its
+   * own 32768 default when this is absent, so a limit it does not know stays undeclared.
+   */
+  maxTokens?: number;
   reasoningEfforts?: Partial<Record<DshReasoningEffort, DshWireReasoningEffort>>;
 }
 
@@ -97,12 +102,14 @@ export function buildDshClientConfig(ctx: ExportContext): DshGeneratedConfig {
     const input = dshInputModalities(model.inputModalities);
     if (input === null) continue;
     const contextWindow = authoritativeContextWindow(model.contextWindow);
+    const maxTokens = contextWindow === undefined ? undefined : knownOutputBudget(contextWindow, model);
     const reasoningEfforts = dshReasoningEfforts(model);
     models.push({
       id: model.namespaced,
       name: exportModelLabel(model),
       input,
       ...(contextWindow !== undefined ? { contextWindow } : {}),
+      ...(maxTokens !== undefined ? { maxTokens } : {}),
       ...(reasoningEfforts ? { reasoningEfforts } : {}),
     });
   }

src/clients/config-export/model-metadata.ts

@@ -31,6 +31,25 @@ export function outputBudgetFor(context: number, model: OpencodeCatalogModel): n
   return Math.min(limit, context);
 }
 
+/**
+* Known model output limit without the schema stand-in, clamped to context.
+*
+* `outputBudgetFor` always returns a number because the clients that require the field
+* cannot express "unknown". A client that instead treats an absent `maxTokens` as a real
+* default must not be handed the 32000 stand-in: that would claim a limit no model declared
+* (#5828). Use this variant where "no claim" is representable.
+* @param context - the context window this model already exports.
+* @param model - the catalog row, which may carry an authoritative output limit.
+* @returns the known limit clamped to `context`, or undefined when nothing is known.
+*/
+export function knownOutputBudget(context: number, model: OpencodeCatalogModel): number | undefined {
+  const provider = model.provider ?? "";
+  const metadata = getModelMetadata(resolveMetadataProvider(provider) ?? provider, model.id ?? model.namespaced);
+  const limit = authoritativeContextWindow(model.maxTokens)
+    ?? authoritativeContextWindow(metadata?.maxTokens);
+  return limit === undefined ? undefined : Math.min(limit, context);
+}
+
 /** Input limits are optional: unlike output, a missing limit needs no schema fallback. */
 export function inputBudgetFor(context: number, model: OpencodeCatalogModel): number | undefined {
   const limit = authoritativeContextWindow(model.maxInputTokens);

tests/config/client-config-export.test.ts

@@ -544,7 +544,14 @@ describe("DSH rc.6 serializer", () => {
     expect(models.find(model => model.id === "a/context")?.contextWindow).toBe(128_000);
     expect(models.find(model => model.id === "a/fraction")).not.toHaveProperty("contextWindow");
     expect(models.find(model => model.id === "a/zero")).not.toHaveProperty("contextWindow");
+    // DSH reads maxTokens as the real output ceiling and falls back to 32768 without it,
+    // so a known limit must ride along with the context window (see #5828 for the same
+    // failure in the other exporters). No context window means no limit claim either.
     for (const model of models) expect(model).not.toHaveProperty("maxTokens");
+    const withLimit = dshConfig(ctx({
+      models: [{ namespaced: "a/limit", provider: "a", id: "limit", contextWindow: 200_000, maxTokens: 128_000 }],
+    }))["llm-pi-ai"].providers.opencodex!.models;
+    expect(withLimit[0]!.maxTokens).toBe(128_000);
   });
 
   test("maps only authoritative effort levels to DSH level-to-wire entries", () => {