Release notes

FEATURE

The report charts when you reach for AI most across your day, without peeking at content

Want to know which hours of the day you actually lean on AI? The report now charts it directly. The analysis report gained a new "Usage Habits" section that plots a full-width, 24-hour activity histogram (a chart of high and low bars showing how much activity falls in each slot), highlights your peak hour, and prints a plain-language summary beside it like "You most often work with AI around HH:00 and HH:00". The data comes entirely from the message timestamps (strings marking when each message was sent, like 2026-07-14 14:30) in your local Claude Code logs, looking only at the timing of user and assistant messages — **never the conversation content**; what you actually discussed is never read. The whole feature runs on your own machine, sending nothing to the cloud and calling no API. It's like a clerk who watches only the clock when you badge in and never glances at your screen, summing up which hours you work most.

FEATURE

When Codex's number goes stale, the panel honestly flags it "N minutes ago" so you don't wait in vain

When Codex's number goes stale, the panel now tells you straight up "this is about N minutes old" instead of leaving you staring at an outdated figure. On the classic panel's Codex card, once the local usage snapshot (a frozen data snapshot taken at one point in time) is older than 15 minutes, an "about N minutes ago" tag appears along with an ⓘ info tooltip (a small explanation that pops up when you hover). Why does it lag? Unlike Claude Code, Codex has no live status-line hook (a hookup point that lets a program push its latest state into the menu bar in real time), so its usage figures can only come from the session data it writes into its logs intermittently, naturally trailing your real account. The tooltip also explains that staying deliberately offline — not querying online — is a design choice so it never burns your token (the unit AI usage is metered in; longer answers cost more). And this entire hint is computed from the existing rate_limits.updated_at field — no network, no API. It's like a fuel gauge whose reading was taken at the morning fill-up now wearing a sticky note saying "about 3 hours ago" so you don't trust it blindly, with an attached note explaining: this car deliberately doesn't report in remotely, so it won't rack up your data charges.

Original changelog

Original changelogThe report charts when you reach for AI most across your day, without peeking at content
**"Usage Habits" section in the HTML report**: fully local, zero API. The analysis report now shows a full-width 24-hour activity histogram of when you work, highlighting your peak hour with a plain-language summary ("You most often work with AI around HH:00 and HH:00"). Data comes from the message timestamps in your local Claude Code logs (user / assistant messages only) — **never the conversation content**. Parsing lives in a standalone `persona_loader.py` with a 300s TTL cache.
Original changelogWhen Codex's number goes stale, the panel honestly flags it "N minutes ago" so you don't wait in vain
**"Stale data" hint on the Codex card**: when the local Codex usage snapshot is older than 15 minutes, the classic panel's Codex card shows an "about N minutes ago" tag plus an info (ⓘ) tooltip. Unlike Claude Code, Codex has no live status-line hook, so its usage numbers come from session logs it writes only intermittently and can lag your real account; the tooltip also explains that staying offline is a deliberate choice so it never burns your tokens. Built from the existing `rate_limits.updated_at` — **no network, no API**.