ai-lens 0.8.103 → 0.8.105

package/.commithash

@@ -1 +1 @@
-0aee413
+0a107d0

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 History of changes to the `ai-lens` CLI package on npm. New entries go on top. Format: `## X.Y.Z — YYYY-MM-DD`, followed by user-facing bullets.
 
+## 0.8.105 — 2026-06-18
+- fix: Cyrillic (and other non-ASCII) in Claude Code hook payloads is no longer mangled on Windows. The windowless hook launcher read stdin through PowerShell's console codepage (OEM, e.g. CP866 on Russian Windows), corrupting prompts/paths on the way in; it now reads the raw stdin bytes and decodes UTF-8 directly. The earlier 0.8.102 fix only covered the write side, so machines that adopted the windowless launcher saw the corruption reappear.
+
+## 0.8.104 — 2026-06-18
+- fix: subagent token usage is now captured for Claude Code. Tokens spent by Task / general-purpose subagents were silently dropped, so subagent-heavy sessions were badly undercounted — only the main thread's tokens showed up. Capture now reads each subagent's own transcript on `SubagentStop` instead of re-reading the parent session's.
+
 ## 0.8.103 — 2026-06-18
 - fix: `ai-lens status` no longer falsely reports the client as "outdated" on repo-path installs (where hooks run capture.js straight from a checkout that auto-updates on `git pull`). It now reads the client version from the repo the hooks actually run, instead of an unused leftover copy in `~/.ai-lens/client/` — which could be months stale and made the check show ✗ + "run init" even though the live client was current. The stale copy is now noted as ignored. Copy-mode installs are unchanged.
 

package/client/ai-lens-hook.ps1

@@ -13,8 +13,16 @@
 #
 # Payload source differs by caller, so read BOTH:
 #   - Cursor pipes it through the PowerShell pipeline → $input.
-#   - Claude Code pipes it to the process's OS stdin → [Console]::In.
-# Read $input first; if empty (Claude Code invocation), fall back to [Console]::In.
+#   - Claude Code pipes it to the process's OS stdin → read it as raw bytes.
+# Read $input first; if empty (Claude Code invocation), read the OS stdin stream.
+#
+# Why raw bytes (not [Console]::In.ReadToEnd()): [Console]::In decodes stdin via
+# [Console]::InputEncoding, which on Windows PowerShell 5.1 defaults to the OEM
+# codepage (e.g. CP866 on Russian Windows) — so Cyrillic/non-ASCII in the JSON
+# payload is mangled at READ time, before we ever re-encode to UTF-8 on the way
+# to node. Reading OpenStandardInput() as bytes and decoding UTF-8 ourselves
+# bypasses the console codepage entirely. (The 0.8.102 fix only covered the WRITE
+# side; the read side still mangled — observed live as mojibake'd prompts.)
 #
 # Args: $args[0] = node path, $args[1] = capture.js path.
 
@@ -25,7 +33,10 @@ try {
 
   $payload = @($input) -join "`n"
   if ([string]::IsNullOrEmpty($payload)) {
-    $payload = [Console]::In.ReadToEnd()
+    $stdin = [Console]::OpenStandardInput()
+    $ms = New-Object System.IO.MemoryStream
+    $stdin.CopyTo($ms)
+    $payload = [System.Text.Encoding]::UTF8.GetString($ms.ToArray())
   }
 
   $psi = New-Object System.Diagnostics.ProcessStartInfo

package/client/capture.js

@@ -892,12 +892,20 @@ function normalizeClaudeCode(event) {
   // TokenUsage event per real assistant API call. This is what makes Claude
   // Code's row-per-call granularity match Cursor and Codex.
   //
-  // Claude Code's Stop hook passes the path as `transcript_path`, but
-  // SubagentStop uses `agent_transcript_path` (see Claude Code hook docs and
-  // test/fixtures/claude-code-events/subagent-stop.json). Check both so
-  // subagent token usage is not silently dropped in production.
+  // Claude Code's Stop hook passes the path as `transcript_path`. SubagentStop
+  // carries BOTH `transcript_path` (the parent session's transcript) AND
+  // `agent_transcript_path` (the subagent's own transcript, a separate file in
+  // a `subagents/` subdir — the subagent's API calls live ONLY there, not in
+  // the parent transcript). So for SubagentStop we MUST prefer
+  // `agent_transcript_path`: a `transcript_path || agent_transcript_path`
+  // precedence always picks the parent transcript (whose delta is already
+  // consumed) and silently drops every subagent call — i.e. all the tokens
+  // burned by Task/general-purpose subagents. See
+  // test/fixtures/claude-code-events/subagent-stop.json for a real payload.
   if (hookType === 'Stop' || hookType === 'SubagentStop') {
-    const transcriptPath = event.transcript_path || event.agent_transcript_path;
+    const transcriptPath = hookType === 'SubagentStop'
+      ? (event.agent_transcript_path || event.transcript_path)
+      : (event.transcript_path || event.agent_transcript_path);
     const { calls, commitOffset } = extractNewClaudeApiCallsFromTranscript(transcriptPath);
     const tokenEvents = calls.map(call => ({
       event_id: null, // assigned in main() from a stable hash of call.uuid

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ai-lens",
-  "version": "0.8.103",
+  "version": "0.8.105",
   "type": "module",
   "description": "Centralized session analytics for AI coding tools",
   "bin": {