fix: include agent/subagent sessions, fix Codex cache hit and cost calculation

- Remove agent-*.jsonl exclusion filter that was dropping ~46% of API calls
- Scan subagents/ directories for subagent session files
- Normalize Codex token semantics: OpenAI includes cached tokens inside
  input_tokens, subtract them to match Anthropic's separate reporting
- Fixes cost double-counting and 100% cache hit display for Codex users

src/parser.ts

@@ -1,4 +1,4 @@
-import { readdir, readFile, stat } from 'fs/promises'
+import { readdir, readFile } from 'fs/promises'
 import { basename, join } from 'path'
 import { calculateCost, getShortModelName } from './models.js'
 import { discoverAllSessions, getProvider } from './providers/index.js'
@@ -298,15 +298,30 @@ async function parseSessionFile(
   return buildSessionSummary(sessionId, project, classified)
 }
 
+async function collectJsonlFiles(dirPath: string): Promise<string[]> {
+  const files = await readdir(dirPath).catch(() => [])
+  const jsonlFiles = files.filter(f => f.endsWith('.jsonl')).map(f => join(dirPath, f))
+
+  for (const entry of files) {
+    if (entry.endsWith('.jsonl')) continue
+    const subagentsPath = join(dirPath, entry, 'subagents')
+    const subFiles = await readdir(subagentsPath).catch(() => [])
+    for (const sf of subFiles) {
+      if (sf.endsWith('.jsonl')) jsonlFiles.push(join(subagentsPath, sf))
+    }
+  }
+
+  return jsonlFiles
+}
+
 async function scanProjectDirs(dirs: Array<{ path: string; name: string }>, seenMsgIds: Set<string>, dateRange?: DateRange): Promise<ProjectSummary[]> {
   const projectMap = new Map<string, SessionSummary[]>()
 
   for (const { path: dirPath, name: dirName } of dirs) {
-    const files = await readdir(dirPath).catch(() => [])
-    const jsonlFiles = files.filter(f => f.endsWith('.jsonl') && !f.startsWith('agent-'))
+    const jsonlFiles = await collectJsonlFiles(dirPath)
 
-    for (const file of jsonlFiles) {
-      const session = await parseSessionFile(join(dirPath, file), dirName, seenMsgIds, dateRange)
+    for (const filePath of jsonlFiles) {
+      const session = await parseSessionFile(filePath, dirName, seenMsgIds, dateRange)
       if (session && session.apiCalls > 0) {
         const existing = projectMap.get(dirName) ?? []
         existing.push(session)

src/providers/codex.ts

@@ -227,6 +227,10 @@ function createParser(source: SessionSource, seenKeys: Set<string>): SessionPars
           const totalTokens = inputTokens + cachedInputTokens + outputTokens + reasoningTokens
           if (totalTokens === 0) continue
 
+          // OpenAI includes cached tokens inside input_tokens; Anthropic does not.
+          // Normalize to Anthropic semantics: inputTokens = non-cached only.
+          const uncachedInputTokens = Math.max(0, inputTokens - cachedInputTokens)
+
           const model = resolveModel(entry.payload, sessionModel)
           const timestamp = entry.timestamp ?? ''
           const dedupKey = `codex:${source.path}:${timestamp}:${cumulativeTotal}`
@@ -236,7 +240,7 @@ function createParser(source: SessionSource, seenKeys: Set<string>): SessionPars
 
           const costUSD = calculateCost(
             model,
-            inputTokens,
+            uncachedInputTokens,
             outputTokens + reasoningTokens,
             0,
             cachedInputTokens,
@@ -246,7 +250,7 @@ function createParser(source: SessionSource, seenKeys: Set<string>): SessionPars
           yield {
             provider: 'codex',
             model,
-            inputTokens,
+            inputTokens: uncachedInputTokens,
             outputTokens,
             cacheCreationInputTokens: 0,
             cacheReadInputTokens: cachedInputTokens,

tests/providers/codex.test.ts

@@ -153,7 +153,7 @@ describe('codex provider - JSONL parsing', () => {
     const call = calls[0]!
     expect(call.provider).toBe('codex')
     expect(call.model).toBe('gpt-5.3-codex')
-    expect(call.inputTokens).toBe(500)
+    expect(call.inputTokens).toBe(400)
     expect(call.cachedInputTokens).toBe(100)
     expect(call.cacheReadInputTokens).toBe(100)
     expect(call.outputTokens).toBe(200)