feat: stamp per-model CONTEXT_WINDOW into subprocess env

Uses getModelContextWindow() to override CONTEXT_WINDOW in the env
passed to the Claude Code subprocess. Local/constrained models (claudecode,
cc/qwen72b, etc.) now self-limit their prompt budget to their actual VRAM
context ceiling instead of using the global 160k default.

server/claude-sdk.js

@@ -17,7 +17,7 @@ import crypto from 'crypto';
 import { promises as fs } from 'fs';
 import path from 'path';
 import os from 'os';
-import { CLAUDE_MODELS } from '../shared/modelConstants.js';
+import { CLAUDE_MODELS, getModelContextWindow } from '../shared/modelConstants.js';
 import { resolveClaudeCodeExecutablePath } from './shared/claude-cli-path.js';
 import {
   createNotificationEvent,
@@ -205,7 +205,14 @@ function mapCliOptionsToSDK(options = {}) {
   // Map model (default to sonnet)
   // Valid models: sonnet, opus, haiku, opusplan, sonnet[1m]
   sdkOptions.model = options.model || CLAUDE_MODELS.DEFAULT;
-  // Model logged at query start below
+
+  // Stamp per-model CONTEXT_WINDOW into the subprocess env so Claude Code
+  // self-limits its prompt budget to the model's actual capacity.
+  // Local/constrained models (e.g. claudecode, cc/qwen72b) declare a lower
+  // ceiling in MODEL_CONTEXT_OVERRIDES; cloud models use the global default.
+  const modelContextWindow = getModelContextWindow(sdkOptions.model);
+  sdkOptions.env.CONTEXT_WINDOW = String(modelContextWindow);
+  sdkOptions.env.VITE_CONTEXT_WINDOW = String(modelContextWindow);
 
   // Map system prompt configuration
   sdkOptions.systemPrompt = {
@@ -287,9 +294,10 @@ function transformMessage(sdkMessage) {
 /**
  * Extracts token usage from SDK result messages
  * @param {Object} resultMessage - SDK result message
+ * @param {string} modelValue - The model value used for this session
  * @returns {Object|null} Token budget object or null
  */
-function extractTokenBudget(resultMessage) {
+function extractTokenBudget(resultMessage, modelValue) {
   if (resultMessage.type !== 'result' || !resultMessage.modelUsage) {
     return null;
   }
@@ -312,11 +320,8 @@ function extractTokenBudget(resultMessage) {
   // Total used = input + output + cache tokens
   const totalUsed = inputTokens + outputTokens + cacheReadTokens + cacheCreationTokens;
 
-  // Use configured context window budget from environment (default 160000)
-  // This is the user's budget limit, not the model's context window
-  const contextWindow = parseInt(process.env.CONTEXT_WINDOW) || 160000;
-
-  // Token calc logged via token-budget WS event
+  // Use per-model context window so the budget meter reflects the model's actual ceiling.
+  const contextWindow = getModelContextWindow(modelValue);
 
   return {
     used: totalUsed,
@@ -682,7 +687,7 @@ async function queryClaudeSDK(command, options = {}, ws) {
         if (models.length > 0) {
           // Model info available in result message
         }
-        const tokenBudgetData = extractTokenBudget(message);
+        const tokenBudgetData = extractTokenBudget(message, sdkOptions.model);
         if (tokenBudgetData) {
           ws.send(createNormalizedMessage({ kind: 'status', text: 'token_budget', tokenBudget: tokenBudgetData, sessionId: capturedSessionId || sessionId || null, provider: 'claude' }));
         }

shared/modelConstants.js

@@ -48,6 +48,8 @@ export const CLAUDE_MODELS = {
     { value: "ollama/glm-4.7-flash", label: "GLM-4.7 Flash (ollama)" },
     { value: "ollama/qwen3.5", label: "Qwen 3.5 (ollama)" },
     { value: "ollama-local/r1-14b-32k", label: "R1 14B 32K (local)" },
+    { value: "claudecode", label: "claudecode (local)" },
+    { value: "cc/qwen72b", label: "qwen72b (local)" },
     { value: "gemini/gemini-3.1-pro-preview", label: "Gemini 3.1 Pro Preview" },
     { value: "gemini/gemini-3.1-flash-lite-preview", label: "Gemini 3.1 Flash Lite" },
     { value: "gemini/gemini-3-flash-preview", label: "Gemini 3 Flash Preview" },
@@ -104,6 +106,37 @@ export const CLAUDE_MODELS = {
   DEFAULT: "cc/claude-sonnet-4-6",
 };
 
+/**
+ * Per-model context window overrides (in tokens).
+ *
+ * Models listed here will have CONTEXT_WINDOW stamped into the Claude Code
+ * subprocess env so the agent self-limits its prompt budget accordingly.
+ * Models NOT listed here fall back to the global CONTEXT_WINDOW env var (default 160000).
+ *
+ * Add an entry here whenever you add a local/constrained model whose VRAM or
+ * architecture limits the usable context below the global default.
+ */
+export const MODEL_CONTEXT_OVERRIDES = {
+  'claudecode': 60000,
+  'cc/qwen72b': 60000,
+  'ollama-local/r1-14b-32k': 32000,
+};
+
+/**
+ * Returns the effective context window size for a given model value.
+ * Checks MODEL_CONTEXT_OVERRIDES first, then falls back to the
+ * global CONTEXT_WINDOW env var, then 160000.
+ *
+ * @param {string} modelValue - The model value string (e.g. "claudecode", "cc/claude-sonnet-4-6")
+ * @returns {number} Context window in tokens
+ */
+export function getModelContextWindow(modelValue) {
+  if (modelValue && MODEL_CONTEXT_OVERRIDES[modelValue] !== undefined) {
+    return MODEL_CONTEXT_OVERRIDES[modelValue];
+  }
+  return parseInt(process.env.CONTEXT_WINDOW, 10) || 160000;
+}
+
 /**
  * Cursor Models
  */