Release Notes

Changelog

Corrected August 30, 2026. An audit found entries below claiming features that were never built, and per-user figures written as if they were product facts. Those entries have been rewritten to match what the code actually does. Nothing was deleted to hide it.

September 8, 2026 — Spend by reasoning effort

The Effort tab splits the window's spend by the reasoning effort each agent recorded — one stacked bar and one matrix of models by effort level, with the levels' totals beneath. Connector 0.8.3 reads the level from the same local logs it already scans (Claude Code records it per message from 2.1.237, Codex per turn), aggregated per day, agent and model, and serves it over loopback only; it is not part of the hub sync payload. An older connector gets an upgrade notice instead of a chart.
Nothing on the tab is re-priced or guessed. Spend is the same estimate as the dashboard total, apportioned within each model-day by token mix at catalog rates, so the levels add up to the total exactly. A message whose agent recorded no effort is shown as "Not recorded", never assigned to a neighbouring level; Antigravity records none. A record the catalog cannot price keeps its tokens under its level and its spend is disclosed as unsplit rather than shown as $0.00.

September 6, 2026 — Disclosed coverage boundary

Codex threads hosted in ChatGPT's cloud are now disclosed, not silently omitted. Codex CLI and Codex-in-VS-Code threads that run locally are read from their rollout logs, as before. Codex threads that run in ChatGPT's cloud (the ChatGPT desktop app's hosted Codex) execute on OpenAI's servers and write no token log to the machine. The connector now detects those threads and counts them per day; their tokens and cost remain unobservable locally, and the dashboard says so rather than presenting a total that quietly excludes them. This is a coverage boundary being stated in the open, not a new capability — no token count is estimated or invented for those threads.
Privacy of the disclosure: Nothing about those threads except a per-day count is exposed, and that count is served over loopback only — it is not part of the hub sync payload.
Connector 0.8.2 carries the detection. An older connector never looked, so the dashboard says nothing about cloud-hosted threads rather than claiming there were none.

v1.3.0 — August 18, 2026

One-step connect: No folder picking and no API keys. The connect page probes the local connector over loopback and opens the dashboard once it answers.
Insights engine: Detection of prompt-cache waste, model-tier routing, reasoning-budget and agent-loop opportunities. Each carries a monthly figure computed from your own records — an estimate, not a quote.
Observed model efficiency: Per-model comparison of tokens per dollar, cost per million observed tokens, prompt-cache rate and reasoning share, all derived from your records. There is no latency or tokens-per-second measurement — the logs do not carry timing.
Coding streak & velocity: Current and longest active-day streaks, a 30-day activity heatmap, and peak usage day — computed per user, so the numbers are yours.
Reasoning tokens: Reasoning-token totals are tracked and broken out per model wherever the agent log reports them. Reasoning content itself is never collected.

v1.2.1 — August 18, 2026

Rebuilt Ingestion Interface: Direct browser telemetry integration and loopback daemon synchronization on port 41414.
Loopback origin allowlist: The connector answers the Usage Radar origin plus loopback origins on your own machine, and rejects a mismatched Host, which is what blocks DNS rebinding. No provider credential is involved — the product never asks for one.

v1.2.0 — August 18, 2026

One recommended start: /scan leads with the local connector. Direct provider-billing ingestion was described here but never shipped; there is no provider API path in the product.
Port migration: The supported connector endpoint is 127.0.0.1:41414; legacy 4317 probing remains only for migrations.

v1.1.1 — July 16, 2026

Honest session pseudonyms: Session identifiers are HMAC-SHA256 keyed with a secret generated per install, so they cannot be correlated across machines.
Snapshot truthfulness: Cached data warning banner when connector is unreachable.

v1.1.0 — July 16, 2026

One name everywhere: Normalized all directories, install paths, and binaries to Usage Radar (~/.usageradar).
Loopback server: Switched defaults to port 41414.

v1.0.0 — June 2026

Initial release: local collector, totals dashboard, spend velocity, and JSON export of the normalized counters.

September 8, 2026 — Spend by reasoning effort

The Effort tab splits the window's spend by the reasoning effort each agent recorded — one stacked bar and one matrix of models by effort level, with the levels' totals beneath. Connector 0.8.3 reads the level from the same local logs it already scans (Claude Code records it per message from 2.1.237, Codex per turn), aggregated per day, agent and model, and serves it over loopback only; it is not part of the hub sync payload. An older connector gets an upgrade notice instead of a chart.

Nothing on the tab is re-priced or guessed. Spend is the same estimate as the dashboard total, apportioned within each model-day by token mix at catalog rates, so the levels add up to the total exactly. A message whose agent recorded no effort is shown as "Not recorded", never assigned to a neighbouring level; Antigravity records none. A record the catalog cannot price keeps its tokens under its level and its spend is disclosed as unsplit rather than shown as $0.00.

September 6, 2026 — Disclosed coverage boundary

Codex threads hosted in ChatGPT's cloud are now disclosed, not silently omitted. Codex CLI and Codex-in-VS-Code threads that run locally are read from their rollout logs, as before. Codex threads that run in ChatGPT's cloud (the ChatGPT desktop app's hosted Codex) execute on OpenAI's servers and write no token log to the machine. The connector now detects those threads and counts them per day; their tokens and cost remain unobservable locally, and the dashboard says so rather than presenting a total that quietly excludes them. This is a coverage boundary being stated in the open, not a new capability — no token count is estimated or invented for those threads.

Privacy of the disclosure: Nothing about those threads except a per-day count is exposed, and that count is served over loopback only — it is not part of the hub sync payload.

Connector 0.8.2 carries the detection. An older connector never looked, so the dashboard says nothing about cloud-hosted threads rather than claiming there were none.

v1.3.0 — August 18, 2026

One-step connect: No folder picking and no API keys. The connect page probes the local connector over loopback and opens the dashboard once it answers.

Insights engine: Detection of prompt-cache waste, model-tier routing, reasoning-budget and agent-loop opportunities. Each carries a monthly figure computed from your own records — an estimate, not a quote.

Observed model efficiency: Per-model comparison of tokens per dollar, cost per million observed tokens, prompt-cache rate and reasoning share, all derived from your records. There is no latency or tokens-per-second measurement — the logs do not carry timing.

Coding streak & velocity: Current and longest active-day streaks, a 30-day activity heatmap, and peak usage day — computed per user, so the numbers are yours.

Reasoning tokens: Reasoning-token totals are tracked and broken out per model wherever the agent log reports them. Reasoning content itself is never collected.

v1.2.1 — August 18, 2026

Rebuilt Ingestion Interface: Direct browser telemetry integration and loopback daemon synchronization on port 41414.

Loopback origin allowlist: The connector answers the Usage Radar origin plus loopback origins on your own machine, and rejects a mismatched Host, which is what blocks DNS rebinding. No provider credential is involved — the product never asks for one.