Map model token limits to native settings

apps/daemon/internal/agent/claudesdk/local_test.go

@@ -40,7 +40,7 @@ func TestWorkspaceProviderCredentialsReplaceAmbientSelection(t *testing.T) {
 	config := workspaceFixture(t)
 	original := slices.Clone(config.Env)
 	req := workspaceRequest()
-	req.AgentOptions["model_provider"] = map[string]any{"protocol": "anthropic", "base_url": "https://provider.example/anthropic", "api_key": "selected-secret"}
+	req.AgentOptions["model_provider"] = map[string]any{"protocol": "anthropic", "base_url": "https://provider.example/anthropic", "api_key": "selected-secret", "context_window": 64000, "max_output_tokens": 4096}
 	start, env, err := prepare(config, req)
 	if err != nil {
 		t.Fatal(err)
@@ -52,6 +52,9 @@ func TestWorkspaceProviderCredentialsReplaceAmbientSelection(t *testing.T) {
 	if !slices.Contains(env, "ANTHROPIC_AUTH_TOKEN=selected-secret") || slices.Contains(env, "ANTHROPIC_AUTH_TOKEN=selected-provider-fixture") {
 		t.Fatal("provider selection was not exclusive")
 	}
+	if !slices.Contains(env, "CLAUDE_CODE_MAX_CONTEXT_TOKENS=64000") || !slices.Contains(env, "CLAUDE_CODE_MAX_OUTPUT_TOKENS=4096") {
+		t.Fatal("token limits were not native")
+	}
 	for _, value := range []any{nil, "secret", map[string]any{"protocol": "anthropic", "base_url": "http://provider.example", "api_key": "secret"}, map[string]any{"protocol": "anthropic", "base_url": "https://user:pass@provider.example", "api_key": "secret"}} {
 		req.AgentOptions["model_provider"] = value
 		if _, _, err := prepare(config, req); err == nil || strings.Contains(err.Error(), "secret") {

apps/daemon/internal/agent/claudesdk/provider.go

@@ -1,6 +1,7 @@
 package claudesdk
 
 import (
+	"strconv"
 	"strings"
 
 	harnessconfiguration "github.com/MiniMax-AI/OpenAgentCore/internal/harnessconfig/claudesdk"
@@ -12,7 +13,15 @@ func providerEnvironment(value any) ([]string, error) {
 	if err != nil {
 		return nil, err
 	}
-	return []string{"ANTHROPIC_BASE_URL=" + provider.BaseURL, "ANTHROPIC_AUTH_TOKEN=" + provider.APIKey}, nil
+	env := []string{"ANTHROPIC_BASE_URL=" + provider.BaseURL, "ANTHROPIC_AUTH_TOKEN=" + provider.APIKey}
+	// Native uses the context window only for a model it does not know.
+	if provider.ContextWindow > 0 {
+		env = append(env, "CLAUDE_CODE_MAX_CONTEXT_TOKENS="+strconv.Itoa(int(provider.ContextWindow)))
+	}
+	if provider.MaxOutputTokens > 0 {
+		env = append(env, "CLAUDE_CODE_MAX_OUTPUT_TOKENS="+strconv.Itoa(int(provider.MaxOutputTokens)))
+	}
+	return env, nil
 }
 
 func withProvider(env, provider []string) []string {
@@ -22,7 +31,7 @@ func withProvider(env, provider []string) []string {
 	out := make([]string, 0, len(env)+len(provider))
 	for _, entry := range env {
 		key, _, _ := strings.Cut(entry, "=")
-		if key != "ANTHROPIC_API_KEY" && key != "ANTHROPIC_AUTH_TOKEN" && key != "ANTHROPIC_BASE_URL" {
+		if key != "ANTHROPIC_API_KEY" && key != "ANTHROPIC_AUTH_TOKEN" && key != "ANTHROPIC_BASE_URL" && key != "CLAUDE_CODE_MAX_CONTEXT_TOKENS" && key != "CLAUDE_CODE_MAX_OUTPUT_TOKENS" {
 			out = append(out, entry)
 		}
 	}

apps/daemon/internal/agent/claudesdk/workspace.go

@@ -4,6 +4,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"slices"
 	"strings"
 
 	"github.com/MiniMax-AI/OpenAgentCore/apps/daemon/internal/localworkspace"
@@ -102,7 +103,7 @@ func workspaceEnvironment(config Config) (*workspaceProfile, []string, error) {
 	env := []string{}
 	for _, entry := range os.Environ() {
 		name, _, _ := strings.Cut(entry, "=")
-		if name != "ANTHROPIC_API_KEY" && name != "ANTHROPIC_AUTH_TOKEN" && name != "CLAUDE_CODE_OAUTH_TOKEN" && name != "CLAUDE_CONFIG_DIR" && name != "HOME" && name != "TMPDIR" {
+		if !slices.Contains([]string{"ANTHROPIC_API_KEY", "ANTHROPIC_AUTH_TOKEN", "CLAUDE_CODE_OAUTH_TOKEN", "CLAUDE_CODE_MAX_CONTEXT_TOKENS", "CLAUDE_CODE_MAX_OUTPUT_TOKENS", "CLAUDE_CONFIG_DIR", "HOME", "TMPDIR"}, name) {
 			env = append(env, entry)
 		}
 	}
@@ -124,7 +125,8 @@ func workspaceEnvName(name string) bool {
 	switch name {
 	case "ANTHROPIC_API_KEY", "ANTHROPIC_AUTH_TOKEN", "ANTHROPIC_BASE_URL",
 		"ANTHROPIC_DEFAULT_SONNET_MODEL", "ANTHROPIC_DEFAULT_OPUS_MODEL", "ANTHROPIC_DEFAULT_HAIKU_MODEL",
-		"CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS", "HTTP_PROXY", "HTTPS_PROXY", "NO_PROXY":
+		"CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS", "CLAUDE_CODE_MAX_CONTEXT_TOKENS", "CLAUDE_CODE_MAX_OUTPUT_TOKENS",
+		"HTTP_PROXY", "HTTPS_PROXY", "NO_PROXY":
 		return true
 	default:
 		return false

apps/daemon/internal/agent/codex/model_route_test.go

@@ -2,6 +2,7 @@ package codex
 
 import (
 	"reflect"
+	"slices"
 	"strings"
 	"testing"
 )
@@ -35,3 +36,17 @@ func TestPlanRejectsIncompleteExplicitProvider(t *testing.T) {
 		}
 	}
 }
+
+func TestPlanMapsTheContextWindowNatively(t *testing.T) {
+	t.Setenv("HOME", t.TempDir())
+	provider := map[string]any{"protocol": "responses", "base_url": "https://model.example/v1", "api_key": "fixture", "context_window": 64000}
+	plan, err := BuildSessionPlan(testStateRoot(t), "limits", map[string]any{"model": "m", "model_provider": provider}, nil)
+	if err != nil || !slices.Contains(plan.ExtraConfig, [2]string{"model_context_window", "64000"}) {
+		t.Fatalf("context window not native: %v %v", plan.ExtraConfig, err)
+	}
+	plan.Cleanup()
+	provider["max_output_tokens"] = 4096
+	if _, err := BuildSessionPlan(testStateRoot(t), "limits", map[string]any{"model": "m", "model_provider": provider}, nil); err == nil {
+		t.Fatal("max_output_tokens without a native home accepted")
+	}
+}

apps/daemon/internal/agent/codex/options.go

@@ -123,6 +123,9 @@ func BuildSessionPlan(stateRoot, agentStateKey string, opts map[string]any, cont
 		plan.DisableFeatures = append(plan.DisableFeatures, "shell_snapshot")
 		env = append(env, providerAPIKeyEnv+"="+provider.BearerToken)
 		excludeFromShellEnvironment(&plan, providerAPIKeyEnv)
+		if provider.ContextWindow > 0 {
+			plan.ExtraConfig = append(plan.ExtraConfig, [2]string{"model_context_window", fmt.Sprint(provider.ContextWindow)})
+		}
 	}
 
 	plan.Env = env
@@ -258,7 +261,7 @@ func normaliseProviderConfig(raw any) (providerConfig, bool, error) {
 	if err != nil {
 		return providerConfig{}, false, err
 	}
-	return providerConfig{BaseURL: provider.BaseURL, BearerToken: provider.APIKey, WireAPI: "responses"}, true, nil
+	return providerConfig{BaseURL: provider.BaseURL, BearerToken: provider.APIKey, WireAPI: "responses", ContextWindow: provider.ContextWindow}, true, nil
 }
 
 func stringOpt(opts map[string]any, key string) string {

apps/daemon/internal/agent/codex/provider_config.go

@@ -42,6 +42,8 @@ type providerConfig struct {
 	// WireAPI is "responses" (codex Responses API) or "chat" (rejected
 	// upstream since codex-rs 0.140). Empty defaults to "responses".
 	WireAPI string
+	// ContextWindow is the native model_context_window when set.
+	ContextWindow int32
 }
 
 // writeCodexProviderConfig writes (or appends to) <codexHome>/config.toml

apps/web/e2e/fixture-console.mjs

@@ -540,7 +540,7 @@ function providerProblem(harness, input) {
   if (!httpsBase(input.base_url)) return "model provider requires an HTTPS base_url without credentials, query or fragment";
   if (!["anthropic", "responses", "chat_completions"].includes(input.protocol)) return "unsupported model provider protocol";
   if (!input.api_key.trim() || Buffer.byteLength(input.api_key) > 16384 || /[\0\r\n]/.test(input.api_key)) return "invalid model provider API key";
-  if ((input.max_output_tokens ?? 0) > (input.context_window ?? 0)) return "invalid model token limits";
+  if (input.context_window && (input.max_output_tokens ?? 0) > input.context_window) return "invalid model token limits";
   return null;
 }
 
@@ -553,7 +553,7 @@ function providerProblem(harness, input) {
  * DELETE is 204 and safe to repeat. A disabled harness may still be configured.
  */
 async function harnessRoute(request, response, path) {
-  const support = (id) => ({ protocols: id === "codex" ? ["responses"] : id === "claude_sdk" ? ["anthropic"] : ["anthropic", "responses", "chat_completions"], accepts_harness_config: id === "claude_sdk" });
+  const support = (id) => ({ protocols: id === "codex" ? ["responses"] : id === "claude_sdk" ? ["anthropic"] : ["anthropic", "responses", "chat_completions"], accepts_harness_config: id === "claude_sdk", accepts_max_output_tokens: id !== "codex" });
   const view = (id) => ({ model_configuration_support: support(id), object: "core.harness", id, enabled: state.harnesses[id].enabled, default: state.harnesses[id].default, model_configuration: state.harnesses[id].provider });
   if (path === "/harnesses" && request.method === "GET") return send(response, 200, { object: "list", data: Object.keys(state.harnesses).map(view) });
   const match = path.match(HARNESS_PROVIDER);

apps/web/e2e/system.spec.ts

@@ -41,21 +41,15 @@ test("sets, replaces and clears a harness's default model configuration, and kee
   // Codex accepts no native Harness configuration.
   await set.getByText("Advanced settings", { exact: true }).click();
   await expect(set.getByLabel("Harness configuration (JSON)")).toHaveCount(0);
-  // Limits are Core's 32-bit whole numbers, and max output fits a set context window.
-  await set.getByLabel("Max output tokens").fill("32000");
-  await expect(set.getByLabel("Context window", { exact: true })).not.toHaveAttribute("aria-invalid", "true");
-  await set.getByLabel("Context window", { exact: true }).fill("1000");
-  await expect(set.getByText("Set a context window at least this large.")).toBeVisible();
-  await expect(set.getByLabel("Context window", { exact: true })).toHaveAttribute("aria-invalid", "true");
-  await expect(set.getByLabel("Context window", { exact: true })).toHaveAccessibleDescription(/Set a context window at least this large/);
-  await expect(set.getByLabel("Max output tokens", { exact: true })).not.toHaveAttribute("aria-invalid", "true");
+  // Codex has no native output limit; limits are Core's 32-bit whole numbers.
+  await expect(set.getByLabel("Max output tokens")).toHaveCount(0);
   await set.getByLabel("Context window").fill("2147483648");
   await expect(set.getByText("Enter a whole number up to 2147483647.")).toBeVisible();
   await expect(set.getByRole("button", { name: "Save" })).toBeDisabled();
   await set.getByLabel("Context window").fill("200000");
   await set.getByRole("button", { name: "Save" }).click();
   await expect(set).toBeHidden();
-  for (const text of ["Ready", "OpenAI Responses", "https://model.example/v1", "Configured", "200,000", "32,000"]) await expect(codex).toContainText(text);
+  for (const text of ["Ready", "OpenAI Responses", "https://model.example/v1", "Configured", "200,000"]) await expect(codex).toContainText(text);
   expect(await page.content()).not.toContain(KEY);
 
   // Replacing starts from the saved fields but never from the key; closing the form forgets a typed key.
@@ -131,6 +125,15 @@ test("sets, replaces and clears a harness's default model configuration, and kee
   await limits.getByLabel("Base URL").fill("https://model.example/anthropic");
   await limits.getByLabel("API key").fill(KEY);
   await limits.getByLabel("Default model ID").fill("fixture-model");
+  // Max output alone is valid and must fit a set context window.
+  await limits.getByText("Advanced settings", { exact: true }).click();
+  await limits.getByLabel("Max output tokens").fill("32000");
+  await expect(limits.getByLabel("Context window", { exact: true })).not.toHaveAttribute("aria-invalid", "true");
+  await limits.getByLabel("Context window", { exact: true }).fill("1000");
+  await expect(limits.getByText("Set a context window at least this large.")).toBeVisible();
+  await expect(limits.getByLabel("Context window", { exact: true })).toHaveAccessibleDescription(/Set a context window at least this large/);
+  await expect(limits.getByLabel("Max output tokens", { exact: true })).not.toHaveAttribute("aria-invalid", "true");
+  await limits.getByLabel("Context window", { exact: true }).fill("");
   await expect(limits.getByRole("button", { name: "Save" })).toBeEnabled();
   await limits.getByRole("button", { name: "Cancel" }).click();
   await expect(mcode).toContainText("Not set");

apps/web/src/features/overview/getting-started.test.ts

@@ -12,7 +12,7 @@ const deployment = (overrides: Partial<SandboxDeployment> = {}): SandboxDeployme
 const fleet = (value: SandboxDeployment, nodes = [node("n1")]): FleetState => ({ status: "ready", snapshot: { deployment: value, nodes, allocations: [], loadedAt: 0 }, targetGeneration: value.generation, refreshing: false, error: null });
 const sandboxes = (state: FleetState) => gettingStartedSteps({ sandboxReset: false, fleet: state, localOnly: false, projects: [], sessions: 0, harnesses: [] }).sandboxes;
 const provider = { object: "core.model_configuration", model: "fixture-model", harness_config: {}, model_provider: { protocol: "responses", base_url: "https://model.example/v1", api_key_configured: true }, last_used_at: null, last_error_code: null, last_error_at: null, updated_at: "2026-09-25T00:00:00Z" } as const;
-const harness = (id: CoreHarness["id"], fields: Partial<CoreHarness> = {}): CoreHarness => ({ object: "core.harness", model_configuration_support: { protocols: ["responses"], accepts_harness_config: true }, id, enabled: true, default: false, model_configuration: null, ...fields });
+const harness = (id: CoreHarness["id"], fields: Partial<CoreHarness> = {}): CoreHarness => ({ object: "core.harness", model_configuration_support: { protocols: ["responses"], accepts_harness_config: true, accepts_max_output_tokens: false }, id, enabled: true, default: false, model_configuration: null, ...fields });
 
 describe("Getting started steps", () => {
   it("uses live provider readiness independently of target state or a durable pin", () => {

apps/web/src/features/system/ModelProviderDialog.tsx

@@ -83,7 +83,7 @@ export function ModelProviderDialog({ harness, onClose, onSaved, onReread }: {
   const url = baseUrl.trim();
   const urlProblem = url && !isProviderUrl(url) ? t("models.form.baseUrlInvalid") : null;
   const context = tokenLimit(contextWindow);
-  const output = tokenLimit(maxOutputTokens);
+  const output = tokenLimit(support?.accepts_max_output_tokens ? maxOutputTokens : "");
   const limitProblem = (limit: ReturnType<typeof tokenLimit>) => (limit.problem === "whole" ? t("models.form.wholeNumber") : limit.problem === "large" ? t("models.form.tooLarge") : null);
   const outputProblem = limitProblem(output);
   // An undersized window is the field the administrator must fix.
@@ -220,7 +220,7 @@ export function ModelProviderDialog({ harness, onClose, onSaved, onReread }: {
             </div> : null}
             <div className="system-model-limits">
               {limitField("context", contextWindow, setContextWindow, contextProblem)}
-              {limitField("output", maxOutputTokens, setMaxOutputTokens, outputProblem)}
+              {support?.accepts_max_output_tokens ? limitField("output", maxOutputTokens, setMaxOutputTokens, outputProblem) : null}
             </div>
           </div>
         </details>

contracts/agents-api/core-errors.md

@@ -61,7 +61,7 @@ Each code returns HTTP 400 with `type: "invalid_request_error"`. A missing, malf
 | `model_provider_base_url_invalid` | `base_url` | omitted | Requires HTTPS without credentials, query or fragment |
 | `model_provider_protocol_unsupported` | `protocol` | `harness` and `allowed_protocols`, from the build's adapter catalog | The protocol is unknown or unsupported by the selected Harness |
 | `model_provider_api_key_invalid` | `api_key` | `max_length`: 16384 | The key is empty, too long or contains a prohibited character |
-| `model_provider_token_limits_invalid` | `context_window` or `max_output_tokens` | omitted | A limit is negative, or `max_output_tokens` exceeds a set `context_window` |
+| `model_provider_token_limits_invalid` | `context_window` or `max_output_tokens` | omitted | A limit is negative, `max_output_tokens` exceeds a set `context_window`, or the Harness has no native setting for a set limit |
 | `model_configuration_model_invalid` | `model` | omitted | The deployment default's model is not a nonempty model identifier |
 | `harness_config_invalid` | `harness_config` | omitted | The deployment default's native parameters are unsupported or invalid |
 | `invalid_sandbox_configuration` | `resources.cpus` | `min`: 1, `max`: 255 | The CPU count is outside the supported bounds |

contracts/agents-api/core.openapi.yaml

@@ -2537,12 +2537,16 @@ definitions:
     properties:
       accepts_harness_config:
         type: boolean
+      accepts_max_output_tokens:
+        description: AcceptsMaxOutputTokens reports that the Harness maps max_output_tokens natively.
+        type: boolean
       protocols:
         items:
           type: string
         type: array
     required:
       - accepts_harness_config
+      - accepts_max_output_tokens
       - protocols
     type: object
   v1.ModelProviderInput:

contracts/agents-api/model-execution.md

@@ -15,13 +15,15 @@ A Harness receives one frozen model provider bundle, defined by [`modelprovider.
 | `api_key` | Nonempty, at most 16 KiB, without NUL, CR or LF |
 | `context_window`, `max_output_tokens` | Optional nonnegative integers, with output no larger than context. Use the real model's limits |
 
-Supported protocols by Harness:
+Supported protocols and native token limits by Harness:
 
-| Harness | Supported protocols, default first |
-| --- | --- |
-| Codex | `responses` |
-| Claude Code | `anthropic` |
-| MiniMax Code | `anthropic`, `responses`, `chat_completions` |
+| Harness | Supported protocols, default first | Native token limits |
+| --- | --- | --- |
+| Codex | `responses` | `model_context_window`; no output limit |
+| Claude Code | `anthropic` | `CLAUDE_CODE_MAX_CONTEXT_TOKENS`, which native uses only for a model it does not know, and `CLAUDE_CODE_MAX_OUTPUT_TOKENS`, which native caps at the model's limit |
+| MiniMax Code | `anthropic`, `responses`, `chat_completions` | The custom model's `limit.context` and `limit.output` |
+
+A set limit with no native setting returns `model_provider_token_limits_invalid`.
 
 Core validates the resolved combination before writing the Session. Core and Runtime read the same ordered `protocols` declaration in `internal/harnessconfig`. There is no model API proxy, passthrough gateway or cross-protocol conversion, including inside a Harness. Unsupported saved configurations and Session snapshots fail when used; they are never rewritten, aliased or migrated.
 
@@ -132,7 +134,7 @@ PUT requires `model` and a complete `model_provider`; `harness_config` defaults
 
 Session creation decrypts the default of the resolved Harness and freezes it in the Session's encrypted snapshot like any other bundle, so changing or removing the default never reaches existing Sessions. The execution-configuration read shows the frozen safe view with source `deployment`. A missing or invalid provider snapshot fails closed, with no fallback to another model or provider.
 
-`model_configuration_support` derives from the adapter declaration that Core and Runtime share: `protocols` lists the selectable native protocols with the default first, and `accepts_harness_config` says whether native parameters are accepted. It describes the build, not a live Runtime or a remote model.
+`model_configuration_support` derives from the adapter declaration that Core and Runtime share: `protocols` lists the selectable native protocols with the default first, `accepts_harness_config` says whether native parameters are accepted, and `accepts_max_output_tokens` whether `max_output_tokens` has a native setting. It describes the build, not a live Runtime or a remote model.
 
 ### Deployment default observations
 

contracts/agents-api/v1/model_configuration.go

@@ -59,6 +59,8 @@ func ValidateNativeModelConfiguration(harness, model string, raw json.RawMessage
 type ModelConfigurationSupport struct {
 	Protocols            []string `json:"protocols" binding:"required"`
 	AcceptsHarnessConfig bool     `json:"accepts_harness_config" binding:"required"`
+	// AcceptsMaxOutputTokens reports that the Harness maps max_output_tokens natively.
+	AcceptsMaxOutputTokens bool `json:"accepts_max_output_tokens" binding:"required"`
 }
 
 func (p *ModelProviderInput) ValidateConfiguration(harness string, native json.RawMessage) error {

contracts/agents-api/v1/model_execution.go

@@ -75,5 +75,8 @@ func (p *ModelProviderInput) ValidateHarnessWithRegistry(harness string, registr
 	if err := configuration.ValidateProtocol(string(p.Protocol)); err != nil {
 		return &ModelProviderError{Code: "model_provider_protocol_unsupported", Param: "protocol", message: err.Error()}
 	}
+	if err := configuration.ValidateMaxOutputTokens(p.MaxOutputTokens); err != nil {
+		return &ModelProviderError{Code: "model_provider_token_limits_invalid", Param: "max_output_tokens", message: err.Error()}
+	}
 	return nil
 }