📊 Usage HQ

Both land on the same Usage settings (shared pool). Shows progress bars: current 5-hour session (used + time left) and weekly limits (Opus-only reset + all-other-models reset). Paid plans only (Pro/Max/Team/Ent).

Codex in the CLI: type /cost for this session's spend. The platform link shows account-level usage. (Confirm exact in-app path next time you're in Codex.)

Every Claude surface shares ONE usage limit — claude.ai, Cowork, AND Claude Code all draw from the same pool. It's a rolling 5-hour session window + weekly caps (with a separate, tighter weekly cap just for Opus). So Opus on Cowork eats the same bucket as Sonnet in Code.

• Long conversations — every past message is resent every turn. A long thread pays for its whole history on each new message. (Biggest drain by far.)
• Model choice — Opus costs several× Sonnet, which costs more than Haiku.
• Effort / extended thinking — higher = more tokens.
• Tools & connectors — web search, Research, MCP connectors are token-heavy.
• Big attachments / pasted files / artifacts.

• Clear/compact between tasks — new task = fresh thread (in Code: /clear; mid-task: /compact). A wandering thread pays for every old topic.
• Match model to job — Sonnet for most builds, Opus only when genuinely stuck or wide refactor, Haiku for quick mechanical stuff. Plan with Opus, execute with Sonnet (/model opusplan in Code).
• Projects = caching — content in a Project is cached and doesn't recount when reused. Put core docs in the Project knowledge base. (This is why we build inside ONE project.)
• Point at file paths, don't paste — pasted text sits in context all session; a path lets Claude read just the part it needs.
• Keep CLAUDE.md / instructions lean — prepended every turn. "Two strikes" rule: only add a note after correcting the same thing twice. Under ~200 lines.
• Batch related asks into one message; be specific to cut back-and-forth.
• Turn off tools you're not using (web search, Research, MCP) in Search & tools settings.

Since there's no live API for usage, the Timekeeper agent opens the links above on a timer, screenshots them into a _USAGE/ folder, and logs the reading — then Mission Control shows the latest snapshot + a "run report now" button. A UI hotkey is fine here because the agent presses it, not you (buttons for you, hotkeys for the robot).

Needs from Tony, once: confirm the exact Codex usage click-path/hotkey + the working Claude usage URL while logged in.