Windows app?

[BuggyButLearning]

This project looks amazing.

I see you have a mac build. Do you have a plan for a windows build anytime?

[iamtoruk]

Yes its almost build a bit difficult though. I would create a separate branch and let you know. It would be nice if you can test and give feedback as in my case currently I have been developing using VM

[iamtoruk]

The Windows menubar app is ready for testing on the feat/windows-menubar-tauri branch. It's a native system tray app built with Tauri (same React frontend as the macOS version).

What's included:

System tray icon with popup dashboard (spending, trends, activity, tips)
Dark/light theme toggle via right-click menu
Runs in background, refreshes automatically
No console window flash on data refresh
To try it:

git clone https://github.com/getagentseal/codeburn.git
cd codeburn
git checkout feat/windows-menubar-tauri
cd desktop
npm install
npm run tauri dev
Requires: Node.js, Rust toolchain, and the Tauri prerequisites for Windows.

Or if you just want to build a release binary:

npm run tauri build
The .msi installer will be in desktop/src-tauri/target/release/bundle/msi/.

Would appreciate feedback on any issues - especially around tray icon behavior, theme rendering, and data accuracy. Once it's solid we'll merge to main and add codeburn menubar support for Windows.

[soumyadebroy3]

Hey — I forked codeburn and built a Windows tray companion in Tauri 2.x (Rust shell + WebView frontend) to scratch exactly this itch. Sharing in case anyone else here is waiting for a native Windows option:

• Repo: https://github.com/soumyadebroy3/codeburn
• npm: @soumyadebroy3/codeburn (v2.1.1)
• Tray release: https://github.com/soumyadebroy3/codeburn/releases/tag/tray-v0.1.0

Install on Windows:

npm install -g @soumyadebroy3/codeburn
codeburn tray

That codeburn tray command downloads the signed .msi from the release page, verifies SHA-256, runs the installer, launches the tray.

Windows tray (the direct answer)

What works in v0.1, matching the macOS Swift menubar feature-by-feature:

• System tray icon + popover with Mica blur on Win11 (Acrylic fallback on Win10)
• Provider tabs, period switcher (Today / 7d / 30d / Month / All)
• Activity panel, Models panel, basic plan section, 30-second auto-refresh
• Single-instance, autostart-on-login, auto-updater (all native Tauri plugins)
• Same JSON contract as the macOS menubar — both consume codeburn report --format json, so feature parity is mostly porting React components

Not yet ported from the Swift app:

• Capacity-estimator confidence tiers (port from CapacityEstimator.swift pending)
• Quota-threshold notifications, compact mode, Claude OAuth keychain

While I was at it — other changes in the fork

Pricing accuracy:

• Anthropic 1-hour vs 5-minute cache rates read separately. Upstream was treating all cache_creation tokens at the 5m rate; that undercounted heavy-cache workloads by ~8.5% on my data. New behaviour reads usage.cache_creation.ephemeral_1h_input_tokens / ephemeral_5m_input_tokens from the JSONL and prices them at 2× and 1.25× the input rate respectively, with a defensive fallback when LiteLLM doesn't expose the 1h rate.
• Plan-aware leverage UX. When codeburn plan set is configured, the dashboard / status bar / menubar / HTML report all flip from raw API spend to Today value / Month value · 6.2× leverage on your $200 Max plan so flat-subscription users don't misread the metered total as a bill.
• Multi-provider plan presence detection. Cursor / Codex / Copilot / Kiro / Antigravity each get a one-line codeburn plan set snippet surfaced when the data shows you're using them but no plan is configured.

Bug fixes (highest blast radius):

• Cross-platform cwd security tightening. Hardened git invocation through a single safeRunGit helper that scrubs GIT_* env vars, forces core.fsmonitor= / core.sshCommand= / safe.directory=*, and rejects cwds outside $HOME / strict subpaths of $TMPDIR so a hostile .git/config at /tmp can't execute under our spawn.
• Calendar / daily-timeline rendering. Calendar heatmap cells used to merge into one fused rectangle in some browsers because adjacent empty <span>s collapsed the grid gap; rewritten as per-week flex rows with explicit flex: 0 0 22px cells. Daily timeline now fills quiet days with cost=0 so the x-axis is calendar-linear instead of data-day-indexed.
• Anomaly detector noise floor. EWMA spike detector was raising spurious flags on tiny noise (e.g. $0.02 vs $0.01 baseline → spike!). Added a MAD floor (max(rawMad, dataMedian * 0.1, 0.01)) and a minCost: 5 filter so spikes only fire when both relative and absolute thresholds are crossed.
• Atomic redact salt write. src/redact.ts cross-process race fixed via tempfile + atomic-rename, so two concurrent codeburn export --redact-paths runs can't corrupt the salt.
• Plan command race. preAction was racing with plan set doing read-modify-write on config.json; plan subcommand now skips auto-detect to avoid the lost-update.

Optimization (refactor work):

• Three highest cognitive-complexity functions extracted into helpers: codex.ts:parse() 75 → <15, cli.ts:export-action 39 → <15, cli.ts:buildJsonReport 32 → <15. Behaviour preserved, all tests green.
• Closed ~28 SonarQube findings (S1854 useless assignments, S5843 regex-complexity, S4624 nested template literals, S6582 optional chaining, S3776 worst-3 only). The long-tail S3776 / S4325 / S7778 smells are deliberately deferred — each was a real refactor with regression risk; gate stays green at 248 smells.

New features:

• HTML report. codeburn export --format html --all-projects emits a single self-contained HTML file with a calendar heatmap, daily timeline (anomaly spikes marked), per-provider donuts, optimize findings, and the yield panel. Handy for sharing / reviewing offline.
• Repo-scoped reports. When run inside a git repo (or with --project), every panel — calendar, daily timeline, donuts, per-project breakdown — scopes to that repo, not just the yield panel. Earlier the panels disagreed about scope, which was confusing.

Quality infra:

• Reproducible builds. scripts/litellm-pin.json pins the upstream LiteLLM model-prices file by commit SHA + SHA-256, so npm run build is deterministic from the tarball alone.
• SonarQube quality-gated CI with a custom MCP server (mcp-sonar/) that lets you drive scans + issue triage from inside Claude Code or any other MCP-compatible client. Quality Gate ✅, 0 bugs, 0 vulnerabilities, 0 hotspots, 81% coverage on new code.
• OIDC Trusted Publishing to npm (no NPM_TOKEN secret stored anywhere; sigstore provenance attestations attached to every published version).

Full diff vs upstream, with a per-feature breakdown: https://github.com/soumyadebroy3/codeburn/blob/main/README.md#fork-notes

Happy to upstream any of these as targeted PRs if they fit the project's direction — the tray is ~600 LOC of TS + Rust, the security ports are clean cherry-picks, and the cache-tier pricing fix is ~10 lines in src/parser.ts. Either way I'll keep the fork alive for my own use and anyone else who wants Windows support or any of the above today.