v0.1.3 — cache breakdown + fresh-only totals option

• Panel menu now shows today's input / output / cache-read breakdown
• New Settings option 'Include cache-read in totals' (config: panel.count_cache_read): when off, label/session/total counters count fresh tokens only (input + output + cache-write). Applies live, no restart.

v0.1.2 — cost from fresh tokens only

Cost estimation now uses only fresh tokens (input, output, cache-write). cache-read — re-sent context that dominates token totals but is cheap or free on most plans — no longer inflates the estimate.

v0.1.1 — fix: kimi tokens counted twice

Fixed

• Token counts were exactly 2× too high for Kimi Code. Its wire.jsonl
  stores every model call twice — as usage.record and as an identical
  step.end loop event — and both were summed. The parser now only counts
  usage.record. Existing stats and pending flush state are migrated once
  (kimi rows halved, guarded by PRAGMA user_version).

Added

• agent-bar stats now breaks usage down into input / output / cache-read /
  cache-write per CLI, with a note explaining that cache reads dominate the
  token total but not the cost.

v0.1.0 — first public release

First public release.

Added

• Top-bar indicator (AppIndicator/StatusNotifierItem) for AI coding
  agents: Kimi Code, Claude Code and Codex CLI. Works on GNOME, KDE Plasma
  and any SNI-capable desktop, X11 and Wayland.
• "Needs you" alerts: pulsing orange icon (with per-session count badge)
  plus a desktop notification containing the exact question or permission an
  agent is waiting on. Kimi (PermissionRequest, Notification) and Claude
  (Notification, PermissionRequest) supported; Codex reports turn
  completion and tokens only (no waiting-for-input hook exists).
• Live token counter in the panel: today's spend parsed incrementally
  from each CLI's own session transcripts (Kimi wire.jsonl, Claude
  transcripts, Codex rollouts).
• Context-window fill (ctx NN%) per session, from the most recent model
  call — warns before a compaction.
• Cost estimation (~$) per day/session with a built-in, user-editable
  price table ([prices.<cli>] in the config).
• Multi-session menu: per-CLI colored dots, status, tokens, time in
  state, and the pending question inline.
• Persistent statistics (SQLite): today/week totals per CLI, top
  sessions, and the agent-bar stats command. Flushes are journaled so
  daemon restarts never double-count.
• Settings submenu in the panel (token counter, animation, notifications,
  sound, icon style, blink speed) with instant apply; everything also
  editable in ~/.config/agent-bar/config.toml, hot-reloaded within ~2 s.
• Icon themes: monochrome symbolic (GNOME-style) and color, plus an
  animated working spinner; generated by scripts/gen_icons.py.
• Hook management: agent-bar install-hooks / uninstall-hooks register
  and cleanly remove hooks in Kimi Code, Claude Code and Codex configs, with
  one-time backups (*.agent-bar.bak).
• agent-bar doctor: environment self-check for bug reports (GTK stack,
  notifications, sound, hooks, systemd unit).
• Packaging: scripts/install.sh / uninstall.sh (apt deps, pipx or
  PEP-668-safe venv install, hooks, systemd user autostart).
• CI workflow (pytest on Python 3.11/3.12, icon reproducibility check) and
  39 unit tests covering parsers, bridge, config, pricing and resource caps.

Security / robustness

• Transcript reading is capped (64 MB per pass, line-aligned) so a runaway
  log writer cannot exhaust daemon memory.
• Hook bridge is fail-open and exception-swallowing: agent-bar can never
  block or break a CLI session.
• UI refreshes are coalesced; menu widgets are destroyed on rebuild and the
  menu is not rebuilt while open.