Skip to content

fix token totals #306

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
rajbos merged 4 commits into from
Feb 25, 2026
+160 −30
Merged

fix token totals #306

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
1d78123
Fix left over merge conflict markers
rajbos Feb 25, 2026
39ab2bb
Fix token calculation
rajbos Feb 25, 2026
9bf5995
Fix display when there is nothing to show
rajbos Feb 25, 2026
7c1bd34
Merge branch 'main' into fix-token-totals
rajbos Feb 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
149 changes: 139 additions & 10 deletions src/extension.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -348,7 +348,7 @@ interface WorkspaceCustomizationSummary {

class CopilotTokenTracker implements vscode.Disposable {
// Cache version - increment this when making changes that require cache invalidation
private static readonly CACHE_VERSION = 23; // Cache key format changed: per-edition file lock instead of per-session keys
private static readonly CACHE_VERSION = 27; // Support result.metadata.promptTokens/outputTokens (Insiders format)
// Maximum length for displaying workspace IDs in diagnostics/customization matrix
private static readonly WORKSPACE_ID_DISPLAY_LENGTH = 8;

Expand Down Expand Up @@ -1169,9 +1169,18 @@ class CopilotTokenTracker implements vscode.Disposable {
* Generate a cache identifier based on VS Code extension mode.
* VS Code editions (stable vs insiders) already have separate globalState storage,
* so we only need to distinguish between production and development (debug) mode.
* In development mode, each VS Code window gets a unique cache identifier using
* the session ID, preventing the Extension Development Host from sharing/fighting
* with the main dev window's cache.
*/
private getCacheIdentifier(): string {
return this.context.extensionMode === vscode.ExtensionMode.Development ? 'dev' : 'prod';
if (this.context.extensionMode === vscode.ExtensionMode.Development) {
// Use a short hash of the session ID to keep the key short but unique per window
const sessionId = vscode.env.sessionId;
const hash = sessionId.substring(0, 8);
return `dev-${hash}`;
}
return 'prod';
}

/**
Expand Down Expand Up @@ -2736,9 +2745,14 @@ class CopilotTokenTracker implements vscode.Disposable {

// Use actual usage if available, otherwise estimate from text
if (request.result?.usage) {
// OLD FORMAT (pre-Feb 2026)
const u = request.result.usage;
modelUsage[requestModel].inputTokens += typeof u.promptTokens === 'number' ? u.promptTokens : 0;
modelUsage[requestModel].outputTokens += typeof u.completionTokens === 'number' ? u.completionTokens : 0;
} else if (typeof request.result?.promptTokens === 'number' && typeof request.result?.outputTokens === 'number') {
// NEW FORMAT (Feb 2026+)
modelUsage[requestModel].inputTokens += request.result.promptTokens;
modelUsage[requestModel].outputTokens += request.result.outputTokens;
} else {
// Fallback to text-based estimation
if (request.message?.text) {
Expand All @@ -2755,6 +2769,20 @@ class CopilotTokenTracker implements vscode.Disposable {
}
}

// FALLBACK: If reconstruction missed result data, use regex extraction from raw lines
const rawModelUsage = this.extractPerRequestUsageFromRawLines(lines);
for (const [reqIdx, extracted] of rawModelUsage) {
const request = sessionState.requests?.[reqIdx];
if (!request) { continue; }
// Only use regex fallback if reconstruction didn't already provide usage
if (request.result?.usage || (typeof request.result?.promptTokens === 'number') || (request.result?.metadata && typeof request.result.metadata.promptTokens === 'number')) { continue; }
let requestModel = defaultModel;
if (request.modelId) { requestModel = request.modelId.replace(/^copilot\//, ''); }
if (!modelUsage[requestModel]) { modelUsage[requestModel] = { inputTokens: 0, outputTokens: 0 }; }
modelUsage[requestModel].inputTokens += extracted.promptTokens;
modelUsage[requestModel].outputTokens += extracted.outputTokens;
}

return modelUsage;
}

Expand All @@ -2773,9 +2801,18 @@ class CopilotTokenTracker implements vscode.Disposable {

// Use actual usage if available, otherwise estimate from text
if (request.result?.usage) {
// OLD FORMAT (pre-Feb 2026)
const u = request.result.usage;
modelUsage[model].inputTokens += typeof u.promptTokens === 'number' ? u.promptTokens : 0;
modelUsage[model].outputTokens += typeof u.completionTokens === 'number' ? u.completionTokens : 0;
} else if (typeof request.result?.promptTokens === 'number' && typeof request.result?.outputTokens === 'number') {
// NEW FORMAT (Feb 2026+)
modelUsage[model].inputTokens += request.result.promptTokens;
modelUsage[model].outputTokens += request.result.outputTokens;
} else if (request.result?.metadata && typeof request.result.metadata.promptTokens === 'number' && typeof request.result.metadata.outputTokens === 'number') {
// INSIDERS FORMAT (Feb 2026+): Tokens nested under result.metadata
modelUsage[model].inputTokens += request.result.metadata.promptTokens;
modelUsage[model].outputTokens += request.result.metadata.outputTokens;
} else {
// Fallback to text-based estimation
// Estimate tokens from user message (input)
Expand Down Expand Up @@ -4911,6 +4948,8 @@ class CopilotTokenTracker implements vscode.Disposable {

// Extract turns from reconstructed requests array
const requests = sessionState.requests || [];
// Pre-compute regex-based token extraction for lines that failed JSON.parse
const rawUsageFallback = this.extractPerRequestUsageFromRawLines(lines);
for (let i = 0; i < requests.length; i++) {
const request = requests[i];
if (!request || !request.requestId) { continue; }
Expand All @@ -4933,13 +4972,39 @@ class CopilotTokenTracker implements vscode.Disposable {
// Extract actual usage data from request.result if available
let actualUsage: ActualUsage | undefined;
if (request.result?.usage) {
// OLD FORMAT (pre-Feb 2026): Tokens nested under request.result.usage
const u = request.result.usage;
actualUsage = {
completionTokens: typeof u.completionTokens === 'number' ? u.completionTokens : 0,
promptTokens: typeof u.promptTokens === 'number' ? u.promptTokens : 0,
promptTokenDetails: Array.isArray(u.promptTokenDetails) ? u.promptTokenDetails : undefined,
details: typeof request.result.details === 'string' ? request.result.details : undefined
};
} else if (typeof request.result?.promptTokens === 'number' && typeof request.result?.outputTokens === 'number') {
// NEW FORMAT (Feb 2026+): Tokens directly at request.result level
actualUsage = {
completionTokens: request.result.outputTokens,
promptTokens: request.result.promptTokens,
details: typeof request.result.details === 'string' ? request.result.details : undefined
};
} else if (request.result?.metadata && typeof request.result.metadata.promptTokens === 'number' && typeof request.result.metadata.outputTokens === 'number') {
// INSIDERS FORMAT (Feb 2026+): Tokens nested under result.metadata
actualUsage = {
completionTokens: request.result.metadata.outputTokens,
promptTokens: request.result.metadata.promptTokens,
details: typeof request.result.details === 'string' ? request.result.details : undefined
};
}

// FALLBACK: If reconstruction missed result data (bad escape chars), use regex extraction
if (!actualUsage) {
const extracted = rawUsageFallback.get(i);
if (extracted) {
actualUsage = {
completionTokens: extracted.outputTokens,
promptTokens: extracted.promptTokens
};
}
}

const turn: ChatTurn = {
Expand Down Expand Up @@ -5704,10 +5769,17 @@ class CopilotTokenTracker implements vscode.Disposable {

// Extract actual token counts from LLM API usage data
if (request.result?.usage) {
// OLD FORMAT (pre-Feb 2026)
const u = request.result.usage;
const prompt = typeof u.promptTokens === 'number' ? u.promptTokens : 0;
const completion = typeof u.completionTokens === 'number' ? u.completionTokens : 0;
totalActualTokens += prompt + completion;
} else if (typeof request.result?.promptTokens === 'number' && typeof request.result?.outputTokens === 'number') {
// NEW FORMAT (Feb 2026+)
totalActualTokens += request.result.promptTokens + request.result.outputTokens;
} else if (request.result?.metadata && typeof request.result.metadata.promptTokens === 'number' && typeof request.result.metadata.outputTokens === 'number') {
// INSIDERS FORMAT (Feb 2026+): Tokens nested under result.metadata
totalActualTokens += request.result.metadata.promptTokens + request.result.metadata.outputTokens;
}
}
}
Expand All @@ -5733,6 +5805,7 @@ class CopilotTokenTracker implements vscode.Disposable {
// is fragile. Reconstructing the state (like the logviewer does) is the reliable approach.
let sessionState: any = {};
let isDeltaBased = false;
let parseFailedLines = 0;

for (const line of lines) {
if (!line.trim()) { continue; }
Expand Down Expand Up @@ -5782,26 +5855,82 @@ class CopilotTokenTracker implements vscode.Disposable {
}
}
} catch (e) {
// Skip malformed lines
// Track parse failures for regex fallback
parseFailedLines++;
}
}

// Extract actual tokens from the reconstructed state (handles all delta path patterns)
// Use per-request regex fallback (like the logviewer) so that requests whose result
// lines failed JSON.parse still contribute actual tokens instead of being silently lost.
let totalActualTokens = 0;
if (isDeltaBased && sessionState.requests && Array.isArray(sessionState.requests)) {
for (const request of sessionState.requests) {
if (request?.result?.usage) {
const u = request.result.usage;
const prompt = typeof u.promptTokens === 'number' ? u.promptTokens : 0;
const completion = typeof u.completionTokens === 'number' ? u.completionTokens : 0;
totalActualTokens += prompt + completion;
if (isDeltaBased) {
const rawUsageFallback = parseFailedLines > 0 ? this.extractPerRequestUsageFromRawLines(lines) : new Map<number, { promptTokens: number; outputTokens: number }>();
const requests = (sessionState.requests && Array.isArray(sessionState.requests)) ? sessionState.requests : [];
// Determine highest request index: max of reconstructed array length and regex-extracted keys
let maxIndex = requests.length;
for (const idx of rawUsageFallback.keys()) {
if (idx + 1 > maxIndex) { maxIndex = idx + 1; }
}
for (let i = 0; i < maxIndex; i++) {
const request = requests[i];
let found = false;
// Try reconstructed state first
if (request?.result) {
const result = request.result;
if (typeof result.promptTokens === 'number' && typeof result.outputTokens === 'number') {
totalActualTokens += result.promptTokens + result.outputTokens;
found = true;
} else if (result.metadata && typeof result.metadata.promptTokens === 'number' && typeof result.metadata.outputTokens === 'number') {
// INSIDERS FORMAT (Feb 2026+): Tokens nested under result.metadata
totalActualTokens += result.metadata.promptTokens + result.metadata.outputTokens;
found = true;
} else if (result.usage) {
const u = result.usage;
const prompt = typeof u.promptTokens === 'number' ? u.promptTokens : 0;
const completion = typeof u.completionTokens === 'number' ? u.completionTokens : 0;
totalActualTokens += prompt + completion;
found = true;
}
}
// Per-request fallback: if reconstruction missed this request's result, use regex
if (!found) {
const extracted = rawUsageFallback.get(i);
if (extracted) {
totalActualTokens += extracted.promptTokens + extracted.outputTokens;
}
}
}
}

return { tokens: totalTokens + totalThinkingTokens, thinkingTokens: totalThinkingTokens, actualTokens: totalActualTokens };
}

/**
* Extract per-request actual token usage from raw JSONL lines using regex.
* Handles cases where lines with result data fail JSON.parse due to bad escape characters.
* Supports both old format (usage.promptTokens/completionTokens) and new format (promptTokens/outputTokens).
*/
private extractPerRequestUsageFromRawLines(lines: string[]): Map<number, { promptTokens: number; outputTokens: number }> {
const usage = new Map<number, { promptTokens: number; outputTokens: number }>();
for (const line of lines) {
if (!line.includes('"result"')) { continue; }
const resultMatch = line.match(/"k":\s*\["requests",\s*(\d+),\s*"result"\]/);
if (!resultMatch) { continue; }
const requestIndex = parseInt(resultMatch[1], 10);
const promptMatch = line.match(/"promptTokens":(\d+)/);
const outputMatch = line.match(/"outputTokens":(\d+)/);
const completionMatch = line.match(/"completionTokens":(\d+)/);
if (promptMatch && (outputMatch || completionMatch)) {
usage.set(requestIndex, {
promptTokens: parseInt(promptMatch[1], 10),
outputTokens: parseInt(outputMatch?.[1] || completionMatch![1], 10)
});
}
}
return usage;
}

/**
* Get OpenCode messages for a session, trying DB first then JSON files.
*/
Expand Down
2 changes: 1 addition & 1 deletion src/webview/details/main.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -194,7 +194,7 @@ function buildMetricsSection(

const tbody = document.createElement('tbody');
const rows: Array<{ label: string; icon: string; color?: string; today: string; last30Days: string; lastMonth: string; projected: string }> = [
{ label: 'Tokens (total)', icon: '🟣', color: '#c37bff', today: (stats.today.actualTokens || 0) > 0 ? formatNumber(stats.today.tokens) : '—', last30Days: (stats.last30Days.actualTokens || 0) > 0 ? formatNumber(stats.last30Days.tokens) : '—', lastMonth: (stats.lastMonth.actualTokens || 0) > 0 ? formatNumber(stats.lastMonth.tokens) : '—', projected: formatNumber(projections.projectedTokens) },
{ label: 'Tokens (total)', icon: '🟣', color: '#c37bff', today: formatNumber(stats.today.tokens), last30Days: formatNumber(stats.last30Days.tokens), lastMonth: formatNumber(stats.lastMonth.tokens), projected: formatNumber(projections.projectedTokens) },
{ label: 'Tokens (user estimated)', icon: '📝', color: '#b39ddb', today: formatNumber(stats.today.estimatedTokens), last30Days: formatNumber(stats.last30Days.estimatedTokens), lastMonth: formatNumber(stats.lastMonth.estimatedTokens), projected: '—' },
{ label: 'Service overhead %', icon: '☁️', color: '#90a4ae', today: (stats.today.actualTokens || 0) > 0 ? formatPercent(((stats.today.tokens - stats.today.estimatedTokens) / stats.today.tokens) * 100) : '—', last30Days: (stats.last30Days.actualTokens || 0) > 0 ? formatPercent(((stats.last30Days.tokens - stats.last30Days.estimatedTokens) / stats.last30Days.tokens) * 100) : '—', lastMonth: (stats.lastMonth.actualTokens || 0) > 0 ? formatPercent(((stats.lastMonth.tokens - stats.lastMonth.estimatedTokens) / stats.lastMonth.tokens) * 100) : '—', projected: '—' },
{ label: 'Thinking tokens', icon: '🧠', color: '#a78bfa', today: formatNumber(stats.today.thinkingTokens || 0), last30Days: formatNumber(stats.last30Days.thinkingTokens || 0), lastMonth: formatNumber(stats.lastMonth.thinkingTokens || 0), projected: '—' },
Expand Down
Loading