feat: Add Crush provider support and message-count model insights

[aaditagrawal]

Summary

Adds support for Charm’s Crush to slopmeter.

This wires Crush through the CLI, provider registry, heatmap rendering, JSON export, docs, and tests. Usage is read from local crush.db files, daily totals are aggregated from top-level sessions, and model insights are derived from assistant message counts when token-by-model data is not available.

What changed

• Added a new crush provider across CLI/provider/export surfaces
• Added --crush and included Crush in default auto-detection and --all
• Implemented a Crush loader that:
  • reads session totals from local SQLite databases
  • snapshots WAL-backed databases before inspection
  • resolves explicit Crush locations first
  • falls back to discovering project-local .crush/crush.db files under HOME
• Added a Crush heatmap theme in both CLI and registry renderers
• Added graceful placeholder handling for providers without token-based model breakdowns
• Added Crush model insights using assistant message count:
  • mostUsedModel
  • recentMostUsedModel
  • explicitly labeled as msgs
• Updated docs and JSON export version
• Added CLI/integration coverage for Crush detection, missing data, WAL snapshots, home-scan fallback, and message-count model insights

Notes

Crush does not persist reliable token usage by model in the local DB, so per-model token totals are not fabricated. Daily input/output totals still come from persisted session token totals, while model insights come from assistant message counts in messages.

Since there's no global db for Crush, results have to be parsed from subdirectories, which can be slow, so I would recommend having this as an optional flag; I only developed this since I had usage in Crush once upon a time and wanted to have that visualised.

Additional fix

Also hardens Cursor loading by switching the auth DB read path to lazy node:sqlite usage so default/all-provider scans are less brittle when better-sqlite3 is unavailable or ABI-mismatched.

Verification

• bun run --cwd packages/cli test

• bun run --cwd packages/registry registry:build

• manual sample run:

  • node packages/cli/dist/cli.js --crush --format json --output /tmp/slopmeter-crush-sample.json

  Thanks for building this project out, it's super useful :)

---

Summary by cubic

Adds Crush and Google Antigravity providers to slopmeter, with message-count model insights and more resilient data loading. Also improves help/output behavior and fixes model/token dedup across providers.

• New Features

  • New crush and antigravity providers across CLI (auto-detect, --crush, --antigravity, --all), registry, renderers, and JSON export.
  • Crush SQLite loader: reads session totals, snapshots WAL DBs, resolves explicit/global/project data dirs, and scans HOME for project .crush/crush.db (supports $CRUSH_GLOBAL_DATA).
  • Antigravity loader: parses local request logs and state DB (treats non-lock DB errors as warnings), normalizes Gemini model names, and aggregates daily totals.
  • UI/registry: themes for Crush and Antigravity, “msgs” unit for model insights with “Not tracked” placeholders when tokens aren’t available.
  • Combined graph title shortened to “All Providers”; “All” view and output suffixes now include Pi, Crush, and Antigravity. --help and READMEs updated.

• Bug Fixes

  • Cursor: lazy node:sqlite auth loading with fail-soft availability probe to avoid better-sqlite3/ABI issues.
  • Codex: trim $CODEX_HOME and fix token usage deduplication.
  • Gemini: deduplicate on normalized model name and token totals.
  • Amp: log a warning when thread JSON read fails instead of silently skipping.
  • Crush: availability check uses full DB-path resolution; add .crush/ to .gitignore.

^{Written for commit 68e8a30. Summary will update on new commits.}

[aaditagrawal]

Unrelated, but the Kilo CLI adapter can be supported the same way OpenCode is, since it's just a fork of OpenCode.