AI Task Manager 0.1.2

Install

macOS, Intel and Apple Silicon (one universal build): download AI.Task.Manager_0.1.2_universal.dmg, open it and drag AI Task Manager to Applications. The app is not notarized yet, so macOS blocks the first launch: open it once, then go to System Settings, Privacy & Security, scroll to Security and click Open Anyway (on macOS 14 or earlier, right-click the app and choose Open). Or run xattr -dr com.apple.quarantine "/Applications/AI Task Manager.app".

Windows 10 and 11: download AI.Task.Manager_0.1.2_x64-setup.exe (or AI.Task.Manager_0.1.2_x64_en-US.msi). The installer is not signed yet, so SmartScreen will warn: More info, then Run anyway.

Installs with update checks on are offered this release at their next launch or popover open, or within four hours. Update checks stay off by default; with them off, download and install over the old copy.

Both installers were built by this repository's own GitHub Actions workflow from the tagged commit. Nothing leaves your computer: docs/privacy.md lists every network call the app can make and the commands that prove it.

What is in 0.1.2

Added

• Agent folders resolve on Windows too. Running now reads a live agent's
  working folder through its own PEB, one process at a time, the way a
  debugger does, since Windows has no batch call like lsof. A process that
  refuses the read, and any 32-bit target, still reads "folder unknown"
  rather than a guess.
• Agent spend names the client it mostly went to. Inventory's thirty-day
  agent rows roll each agent's cost up to a client the same way the Clients
  tab does, and say so only when one client is a true majority -- never a
  mere plurality among several smaller ones.
• An opt-in /v1/agents feed. Thirty days of agent spend alongside who
  is running right now, behind the same loopback-only switch as the other
  local feeds. A running row carries a tool's pace, work area and client;
  the working folder, the process id and the session id never reach it.
• Team-wide agent spend on the collector. Built-in agent names (Explore,
  Plan and the rest) come through as themselves; anything a seat named
  itself is folded into one custom row before it ever leaves that seat, so
  a name you chose yourself never leaves the machine that reported it. The
  collector re-applies the same fold to whatever a client actually posts.
• Two new usage findings, and a third promoted to the audit.
  cache-read-share fires when at least 90% of the last 30 days' tokens
  were cache re-reads and at least $50 of Claude spend sits behind it -- the
  share alone is just how Claude Code works, so the dollar floor is what
  decides whether it is worth reading. subagent-share fires when
  subagents account for at least 10% of the last 30 days' spend, and at
  least $5. perm-deny-only, until now shown only in Inventory, also
  reaches the audit. All three are educational, not scored: unscored
  "consider" entries that stay quiet on a machine with nothing to say.

Changed

• The local scan cache re-scans once. The new re-read finding needed a
  day-by-day cache-read count the persisted cache did not keep before, so
  the first launch after updating re-reads session logs once to backfill
  it; every other total is unaffected.
• Live token pace now covers Codex and Gemini CLI, not only Claude Code.
  Running now's pace reads each tool's own newest session file for that
  folder, cached by tool and folder together, so two different tools
  sharing one folder can never show each other's numbers.