ccmeter v0.3.0

[0.3.0] — 2026-04-26

Seven new features that knock items off the README roadmap. All in-tree;
no new infrastructure required.

Added

• ccmeter prompts — per-prompt quality scoring. Ranks every assistant
  turn by output-tokens-per-dollar. The top of the list is high-yield
  prompts (concise input → useful output); the bottom is low-yield (lots
  of cache + input → little output). Use the bottom list for self-coaching.
• iCal export — ccmeter export --format ics (alias ical) emits a
  standards-compliant .ics file with one VEVENT per session. Drop into
  Apple Calendar / Google Calendar / Outlook for billable-time tracking.
• Git-branch auto-tagging — ccmeter --auto-tag-git ... (or env
  CCMETER_GIT_AUTOTAG=1) reads .git/HEAD from each session's working
  directory and applies a branch:<name> tag to non-main sessions. Manual
  tags from ccmeter tag always win. Worktrees and detached-HEAD handled.
• ccmeter notify — desktop budget notifications. One-shot or
  --watch polling daemon. macOS osascript, Linux notify-send,
  Windows toast (best-effort). Threshold and quiet-window are tunable.
  Reads your saved monthly budget from ccmeter budget.
• AI Coach in ccmeter live — rules-based real-time warnings inline
  in the live ticker. Catches: idle-bust imminent (4+ minutes since last
  turn with active cache), 3 consecutive >$0.30 turns (likely loop),
  same tool fired 5+ times in last 10 turns (stuck), >$1 burned in the
  last minute (rate alarm). Each rule is 4 lines; PRs to add more are
  trivial.
• ccmeter reconcile — diffs ccmeter's local total against
  Anthropic's authoritative /v1/organizations/usage_report. Requires
  ANTHROPIC_API_KEY in the environment with org-admin scope. Reports
  delta and percentage; flags >15% drift with diagnostic suggestions.
  Defensive about Anthropic's evolving API shape.
• python/pyccmeter.py — Python companion module. Wraps the CLI's
  JSON export and exposes the analysis as Python dataclasses. Useful
  for piping into pandas / matplotlib / per-team rollups. Single-file,
  no install — copy into your project, from pyccmeter import load_analysis.

Not built (multi-day projects)

The README roadmap also lists items that need infrastructure outside this
repo's scope: native macOS menu-bar app (Xcode + signing + notarization),
VS Code / JetBrains extensions (separate marketplaces), per-team
aggregation server (hosting + auth), anonymized community comparator
(SaaS backend), Go port (separate ecosystem). These remain on the
roadmap but live in companion repos when they ship, not in ccmeter core.

ccmeter v0.2.0

[0.2.0] — 2026-04-26

Fixed (critical, blocked launch)

• Opus 4.x pricing. Rate was hard-coded at $15/$75 per 1M tokens; the actual published Anthropic rate for every Opus 4.x model is $5/$25. ccmeter would have over-stated Opus spend by ~3× — fixed and verified against multiple secondary sources on 2026-04-26. Cache rates derive from the universal multipliers (5m=1.25×, 1h=2×, read=0.10×).
• Cache TTL framing. README, LAUNCH copy, recommendation rule, dashboard callout, and demo dataset all asserted "April 2, 2026" as the TTL change date with "30–200%" cost impact. Real story per anthropics/claude-code#46829 and The Register, Apr 13 2026: rollout was staggered through early March 2026; impact is 30–60% for typical heavy use; Anthropic's official position is the change should not increase costs. All copy now reflects the verified facts.

Added

• ccmeter selftest — diagnostic command that validates the parser & pricing against your real ~/.claude/projects data. Reports parse-success ratio, presence of usage objects, observed model ids, unknown schema keys, and a redacted sample of turn shapes. --json output is formatted to paste directly into a GitHub issue.
• ccmeter pricing — print the active pricing table (built-in plus any user overrides). Trust through transparency: when Console numbers don't match, this is the first thing to diff.
• ccmeter tools — per-tool cost breakdown answering "which tool ate the budget"; rolls up tool_use blocks across sessions and attributes per-turn cost.
• ccmeter tag <id> <label> — annotate a session for grouped reporting; persists to ~/.config/ccmeter/tags.json. Adds --tag filter on ccmeter sessions. Tags surface as a column.
• ccmeter share — Markdown stat-card or SVG suitable for Reddit/Twitter posts. Fully redacted by default; --fuzzy rounds to friendly bands like "~$200".
• ccmeter compare — week-over-week (or any two-period) deltas across totals, projects, cache hit-rate, busts.
• ccmeter live — full-screen ANSI ticker that animates each new turn as it lands. Plays nicely on a recording.
• Dashboard "Tools" page with attributed-cost bars per tool name.
• New API endpoints: /api/tools, /api/tags.
• Export schema versioning. The exported JSON now includes schemaVersion: 1; ccmeter merge rejects incompatible majors instead of silently corrupting reports.
• Parser tolerates usage at the top level of a turn (older Claude Code shape) and model at the top level (subagent traces). Skips known non-session log files: history.jsonl, summary-*.jsonl, summary.*.jsonl.
• Windows project-dir decode heuristic (C--Users-foo-bar → C:\Users\foo\bar).
• Switched primary CLI build to tsc for readable per-file source maps and zero-bundle distribution; tsup config retained for advanced single-bundle builds. Vite emptyOutDir is now opt-in to support read-only or sandboxed filesystems.
• Code-split the dashboard bundle. Recharts ships in its own ~340 KB chunk so the main page paints from a 250 KB bundle (gzip 77 KB).

Added (tests)

• pricing: explicit assertions for Opus 4.x ($5/$25), Claude 3 Opus legacy ($15/$75), and the top-level-usage fallback
• paths: POSIX + Windows decode heuristic
• tools: rollup correctness across sessions
• tags: round-trip read/write/delete + applyTags decoration