WildDragonLLC/claudecodeui
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

feat: stamp per-model CONTEXT_WINDOW into subprocess env

Browse source 
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.
This commit is contained in:
Zac Gaetano 2026-05-31 17:20:55 -04:00
parent 69e478224c
commit 07b6f4eb87
1 changed files with 14 additions and 9 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

23
server/claude-sdk.js
View file

@ -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' }));
}