feat: allow customizing context window and max output tokens

src/components/ProviderManager.test.tsx

@@ -961,13 +961,13 @@ test('ProviderManager clears hidden Hicap auth fields when editing', async () =>
     mounted.stdin.write('\r')
     await waitForFrameOutput(mounted.getOutput, frame =>
       frame.includes('Edit provider profile') &&
-      frame.includes('Step 1 of 6'),
+      frame.includes('Step 1 of 8'),
     )
 
-    for (let step = 2; step <= 6; step++) {
+    for (let step = 2; step <= 8; step++) {
       mounted.stdin.write('\r')
       await waitForFrameOutput(mounted.getOutput, frame =>
-        frame.includes(`Step ${step} of 6`),
+        frame.includes(`Step ${step} of 8`),
       )
     }
     mounted.stdin.write('\r')
@@ -1037,25 +1037,25 @@ test('ProviderManager skips advanced fields for legacy Kimi Code profiles', asyn
     await waitForFrameOutput(mounted.getOutput, frame =>
       frame.includes('Edit provider profile') &&
       frame.includes('Provider name') &&
-      frame.includes('Step 1 of 4'),
+      frame.includes('Step 1 of 6'),
     )
 
     mounted.stdin.write('\r')
     await waitForFrameOutput(mounted.getOutput, frame =>
       frame.includes('Base URL') &&
-      frame.includes('Step 2 of 4'),
+      frame.includes('Step 2 of 6'),
     )
 
     mounted.stdin.write('\r')
     await waitForFrameOutput(mounted.getOutput, frame =>
       frame.includes('Default model') &&
-      frame.includes('Step 3 of 4'),
+      frame.includes('Step 3 of 6'),
     )
 
     mounted.stdin.write('\r')
     const output = await waitForFrameOutput(mounted.getOutput, frame =>
       frame.includes('API key') &&
-      frame.includes('Step 4 of 4'),
+      frame.includes('Step 4 of 6'),
     )
 
     expect(output).not.toContain('API mode')
@@ -1798,49 +1798,61 @@ test('ProviderManager editing an active multi-model provider keeps app state on
     mounted.getOutput,
     frame =>
       frame.includes('Edit provider profile') &&
-      frame.includes('Step 1 of 8'),
+      frame.includes('Step 1 of 10'),
+  )
+
+  mounted.stdin.write('\r')
+  await waitForFrameOutput(
+    mounted.getOutput,
+    frame => frame.includes('Step 2 of 10'),
+  )
+
+  mounted.stdin.write('\r')
+  await waitForFrameOutput(
+    mounted.getOutput,
+    frame => frame.includes('Step 3 of 10'),
   )
 
   mounted.stdin.write('\r')
   await waitForFrameOutput(
     mounted.getOutput,
-    frame => frame.includes('Step 2 of 8'),
+    frame => frame.includes('Step 4 of 10'),
   )
 
   mounted.stdin.write('\r')
   await waitForFrameOutput(
     mounted.getOutput,
-    frame => frame.includes('Step 3 of 8'),
+    frame => frame.includes('Step 5 of 10'),
   )
 
   mounted.stdin.write('\r')
   await waitForFrameOutput(
     mounted.getOutput,
-    frame => frame.includes('Step 4 of 8'),
+    frame => frame.includes('Step 6 of 10'),
   )
 
   mounted.stdin.write('\r')
   await waitForFrameOutput(
     mounted.getOutput,
-    frame => frame.includes('Step 5 of 8'),
+    frame => frame.includes('Step 7 of 10'),
   )
 
   mounted.stdin.write('\r')
   await waitForFrameOutput(
     mounted.getOutput,
-    frame => frame.includes('Step 6 of 8'),
+    frame => frame.includes('Step 8 of 10'),
   )
 
   mounted.stdin.write('\r')
   await waitForFrameOutput(
     mounted.getOutput,
-    frame => frame.includes('Step 7 of 8'),
+    frame => frame.includes('Step 9 of 10'),
   )
 
   mounted.stdin.write('\r')
   await waitForFrameOutput(
     mounted.getOutput,
-    frame => frame.includes('Step 8 of 8'),
+    frame => frame.includes('Step 10 of 10'),
   )
 
   mounted.stdin.write('\r')

src/components/ProviderManager.tsx

@@ -111,6 +111,8 @@ type DraftField =
   | 'authHeader'
   | 'authHeaderValue'
   | 'customHeaders'
+  | 'contextWindowSize'
+  | 'maxOutputTokens'
 
 type ProviderDraft = Record<DraftField, string>
 
@@ -194,6 +196,20 @@ const FORM_STEPS: Array<{
     helpText: 'Optional. Extra non-auth request headers for providers that support them.',
     optional: true,
   },
+  {
+    key: 'contextWindowSize',
+    label: 'Context window size',
+    placeholder: 'e.g. 128000',
+    helpText: 'Optional. Max tokens for context (input + output + history).',
+    optional: true,
+  },
+  {
+    key: 'maxOutputTokens',
+    label: 'Max output tokens',
+    placeholder: 'e.g. 4096',
+    helpText: 'Optional. Max tokens the model can generate in a single turn.',
+    optional: true,
+  },
 ]
 
 const GITHUB_PROVIDER_ID = '__github_models__'
@@ -215,6 +231,8 @@ function toDraft(profile: ProviderProfile): ProviderDraft {
     authHeader: profile.authHeader ?? '',
     authHeaderValue: profile.authHeaderValue ?? '',
     customHeaders: serializeProfileCustomHeaders(profile.customHeaders) ?? '',
+    contextWindowSize: profile.contextWindowSize ? String(profile.contextWindowSize) : '',
+    maxOutputTokens: profile.maxOutputTokens ? String(profile.maxOutputTokens) : '',
   }
 }
 
@@ -251,6 +269,8 @@ function presetToDraft(preset: ProviderPreset): ProviderDraft {
     authHeader: '',
     authHeaderValue: '',
     customHeaders: '',
+    contextWindowSize: '',
+    maxOutputTokens: '',
   }
 }
 
@@ -282,7 +302,13 @@ function profileSummary(profile: ProviderProfile, isActive: boolean): string {
     routeSupportsAuthHeaders(routeId) && profile.authHeader
       ? ` · ${profile.authHeader} auth`
       : ''
-  return `${providerKind} · ${profile.baseUrl} · ${modelDisplay}${modeInfo}${authInfo} · ${keyInfo}${activeSuffix}`
+  const contextInfo = profile.contextWindowSize
+    ? ` · ${Math.round(profile.contextWindowSize / 1000)}k ctx`
+    : ''
+  const tokensInfo = profile.maxOutputTokens
+    ? ` · ${profile.maxOutputTokens} out`
+    : ''
+  return `${providerKind} · ${profile.baseUrl} · ${modelDisplay}${modeInfo}${authInfo}${contextInfo}${tokensInfo} · ${keyInfo}${activeSuffix}`
 }
 
 function getGithubCredentialSourceFromEnv(
@@ -1134,6 +1160,8 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
       authHeader: '',
       authHeaderValue: '',
       customHeaders: '',
+      contextWindowSize: '',
+      maxOutputTokens: '',
     }
     setEditingProfileId(null)
     setDraftProvider(provider)
@@ -1229,6 +1257,12 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
         Object.keys(parsedCustomHeaders.headers).length > 0
           ? parsedCustomHeaders.headers
           : undefined,
+      contextWindowSize: nextDraft.contextWindowSize
+        ? parseInt(nextDraft.contextWindowSize, 10)
+        : undefined,
+      maxOutputTokens: nextDraft.maxOutputTokens
+        ? parseInt(nextDraft.maxOutputTokens, 10)
+        : undefined,
     }
 
     const saved = profileId
@@ -1469,6 +1503,18 @@ export function ProviderManager({ mode, onDone }: Props): React.ReactNode {
       return
     }
 
+    if (
+      (currentStepKey === 'contextWindowSize' ||
+        currentStepKey === 'maxOutputTokens') &&
+      trimmed.length > 0
+    ) {
+      const parsed = parseInt(trimmed, 10)
+      if (isNaN(parsed) || parsed <= 0) {
+        setErrorMessage(`${currentStep.label} must be a positive number.`)
+        return
+      }
+    }
+
     const nextDraft = {
       ...draft,
       [currentStepKey]: trimmed,

src/screens/Doctor.tsx

@@ -4,7 +4,7 @@ import { join } from 'path';
 import React, { Suspense, use, useCallback, useEffect, useMemo, useState } from 'react';
 import { KeybindingWarnings } from 'src/components/KeybindingWarnings.js';
 import { McpParsingWarnings } from 'src/components/mcp/McpParsingWarnings.js';
-import { getModelMaxOutputTokens } from 'src/utils/context.js';
+import { getModelMaxOutputTokens, getContextWindowForModel } from 'src/utils/context.js';
 import { getClaudeConfigHomeDir } from 'src/utils/envUtils.js';
 import type { SettingSource } from 'src/utils/settings/constants.js';
 import { getOriginalCwd } from '../bootstrap/state.js';
@@ -152,6 +152,10 @@ export function Doctor(t0) {
     }, {
       name: "CLAUDE_CODE_MAX_OUTPUT_TOKENS",
       ...getModelMaxOutputTokens("claude-opus-4-6")
+    }, {
+      name: "CLAUDE_CODE_MAX_CONTEXT_TOKENS",
+      default: getContextWindowForModel("claude-opus-4-6"),
+      upperLimit: 1_000_000
     }];
     t4 = envVars.map(_temp8).filter(_temp9);
     $[5] = t4;

src/utils/config.ts

@@ -200,6 +200,8 @@ export type ProviderProfile = {
   authScheme?: OpenAICompatibleAuthScheme
   authHeaderValue?: string
   customHeaders?: Record<string, string>
+  contextWindowSize?: number
+  maxOutputTokens?: number
 }
 
 export type GlobalConfig = {

src/utils/context.ts

@@ -9,9 +9,13 @@ import {
 } from '../integrations/routeMetadata.js'
 import { getCanonicalName } from './model/model.js'
 import { getModelCapability } from './model/modelCapabilities.js'
+import { resolveAntModel } from './model/antModels.js'
+
+import { validateBoundedIntEnvVar } from './envValidation.js'
 
 // Model context window size (200k tokens for all models right now)
 export const MODEL_CONTEXT_WINDOW_DEFAULT = 200_000
+export const MODEL_CONTEXT_WINDOW_UPPER_LIMIT = 1_000_000
 
 // Fallback context window for unknown 3P models. Must be large enough that
 // the effective context (this minus output token reservation) stays positive,
@@ -80,20 +84,26 @@ export function getContextWindowForModel(
   model: string,
   betas?: string[],
 ): number {
-  // Allow override via environment variable (internal-only)
+  const contextWindowSize = getContextWindowForModelInternal(model, betas)
+
+  // Allow override via environment variable
   // This takes precedence over all other context window resolution, including 1M detection,
   // so users can cap the effective context window for local decisions (auto-compact, etc.)
   // while still using a 1M-capable endpoint.
-  if (
-    process.env.USER_TYPE === 'ant' &&
-    process.env.CLAUDE_CODE_MAX_CONTEXT_TOKENS
-  ) {
-    const override = parseInt(process.env.CLAUDE_CODE_MAX_CONTEXT_TOKENS, 10)
-    if (!isNaN(override) && override > 0) {
-      return override
-    }
-  }
+  // Note: CLAUDE_CODE_MAX_CONTEXT_TOKENS now applies to all users.
+  const result = validateBoundedIntEnvVar(
+    'CLAUDE_CODE_MAX_CONTEXT_TOKENS',
+    process.env.CLAUDE_CODE_MAX_CONTEXT_TOKENS,
+    contextWindowSize,
+    MODEL_CONTEXT_WINDOW_UPPER_LIMIT,
+  )
+  return result.effective
+}
 
+function getContextWindowForModelInternal(
+  model: string,
+  betas?: string[],
+): number {
   // [1m] suffix — explicit client-side opt-in, respected over all detection
   if (has1mContext(model)) {
     return 1_000_000

src/utils/providerProfile.test.ts

@@ -1521,3 +1521,24 @@ test('atomic-chat launch ignores mismatched persisted openai env', async () => {
   assert.equal(env.CODEX_API_KEY, undefined)
   assert.equal(env.CHATGPT_ACCOUNT_ID, undefined)
 })
+
+test('buildStartupEnvFromProfile preserves persisted token limits', async () => {
+  const env = await buildStartupEnvFromProfile({
+    persisted: {
+      profile: 'openai',
+      env: {
+        OPENAI_BASE_URL: 'https://api.openai.com/v1',
+        OPENAI_MODEL: 'gpt-4o',
+        CLAUDE_CODE_MAX_CONTEXT_TOKENS: '50000',
+        CLAUDE_CODE_MAX_OUTPUT_TOKENS: '4096',
+      },
+      createdAt: new Date().toISOString(),
+    },
+    goal: 'balanced',
+    processEnv: {},
+  })
+
+  assert.equal(env.OPENAI_MODEL, 'gpt-4o')
+  assert.equal(env.CLAUDE_CODE_MAX_CONTEXT_TOKENS, '50000')
+  assert.equal(env.CLAUDE_CODE_MAX_OUTPUT_TOKENS, '4096')
+})