d8dfun/claudecodeui
1
0
Fork 0
mirror of https://github.com/siteboon/claudecodeui.git synced 2026-05-30 08:15:31 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix: refine token usage reporting

Browse Source 
The old token UI mixed context pressure, cache counters, and dollar estimates.

That made the percentage look precise even when provider data was incomplete or different.

The composer and /cost view now show concrete counts instead of a pie percentage.

Token payloads now share a smaller shape: used, inputTokens, outputTokens, and breakdown.

Claude uses per-step usage where available and Codex reads total_token_usage events.

Gemini reads its tokens object without inventing a context window.

OpenCode reads opencode.db session totals and includes all token columns in used.

The /cost backend no longer returns cache display fields or input/output dollar estimates.

This avoids derived values that look reliable but are not comparable across providers.

Verification: npm run typecheck; targeted eslint; OpenCode session provider test.
This commit is contained in:
Haileyesus
2026-05-29 15:06:55 +03:00
parent 661b8bd137
commit d0cc85e76b
15 changed files with 515 additions and 263 deletions
Download Patch File Download Diff File Expand all files Collapse all files

83
server/claude-sdk.js
View File

@@ -285,43 +285,68 @@ function transformMessage(sdkMessage) {
return sdkMessage;
}
function readNumber(value) {
const parsed = Number(value);
return Number.isFinite(parsed) ? parsed : 0;
}
/**
* Extracts token usage from SDK result messages
* @param {Object} resultMessage - SDK result message
* Extracts token usage from SDK messages.
* Prefers per-step `message.usage` (Claude message payload), then falls back
* to result-level usage/modelUsage for compatibility across SDK versions.
* @param {Object} sdkMessage - SDK stream message
* @returns {Object|null} Token budget object or null
*/
function extractTokenBudget(resultMessage) {
if (resultMessage.type !== 'result' || !resultMessage.modelUsage) {
function extractTokenBudget(sdkMessage) {
if (!sdkMessage || typeof sdkMessage !== 'object') {
return null;
}
// Get the first model's usage data
const modelKey = Object.keys(resultMessage.modelUsage)[0];
const modelData = resultMessage.modelUsage[modelKey];
const messageUsage = sdkMessage.message?.usage || sdkMessage.usage;
if (messageUsage && typeof messageUsage === 'object') {
const inputTokens = readNumber(messageUsage.input_tokens ?? messageUsage.inputTokens);
const outputTokens = readNumber(messageUsage.output_tokens ?? messageUsage.outputTokens);
const totalUsed = inputTokens + outputTokens;
const contextWindow = parseInt(process.env.CONTEXT_WINDOW, 10) || 160000;
if (!modelData) {
return {
used: totalUsed,
total: contextWindow,
inputTokens,
outputTokens,
breakdown: {
input: inputTokens,
output: outputTokens,
},
};
}
if (!sdkMessage.modelUsage || typeof sdkMessage.modelUsage !== 'object') {
return null;
}
// Use cumulative tokens if available (tracks total for the session)
// Otherwise fall back to per-request tokens
const inputTokens = modelData.cumulativeInputTokens || modelData.inputTokens || 0;
const outputTokens = modelData.cumulativeOutputTokens || modelData.outputTokens || 0;
const cacheReadTokens = modelData.cumulativeCacheReadInputTokens || modelData.cacheReadInputTokens || 0;
const cacheCreationTokens = modelData.cumulativeCacheCreationInputTokens || modelData.cacheCreationInputTokens || 0;
// Fallback for older SDK messages with only modelUsage
const modelKey = Object.keys(sdkMessage.modelUsage)[0];
const modelData = sdkMessage.modelUsage[modelKey];
// Total used = input + output + cache tokens
const totalUsed = inputTokens + outputTokens + cacheReadTokens + cacheCreationTokens;
if (!modelData || typeof modelData !== 'object') {
return null;
}
// 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
const inputTokens = readNumber(modelData.cumulativeInputTokens ?? modelData.inputTokens);
const outputTokens = readNumber(modelData.cumulativeOutputTokens ?? modelData.outputTokens);
const totalUsed = inputTokens + outputTokens;
const contextWindow = parseInt(process.env.CONTEXT_WINDOW, 10) || 160000;
return {
used: totalUsed,
total: contextWindow
total: contextWindow,
inputTokens,
outputTokens,
breakdown: {
input: inputTokens,
output: outputTokens,
},
};
}
@@ -684,16 +709,10 @@ async function queryClaudeSDK(command, options = {}, ws) {
ws.send(msg);
}
// Extract and send token budget updates from result messages
if (message.type === 'result') {
const models = Object.keys(message.modelUsage || {});
if (models.length > 0) {
// Model info available in result message
}
const tokenBudgetData = extractTokenBudget(message);
if (tokenBudgetData) {
ws.send(createNormalizedMessage({ kind: 'status', text: 'token_budget', tokenBudget: tokenBudgetData, sessionId: capturedSessionId || sessionId || null, provider: 'claude' }));
}
// Extract and send token budget updates from assistant/result usage payloads
const tokenBudgetData = extractTokenBudget(message);
if (tokenBudgetData) {
ws.send(createNormalizedMessage({ kind: 'status', text: 'token_budget', tokenBudget: tokenBudgetData, sessionId: capturedSessionId || sessionId || null, provider: 'claude' }));
}
}