v0.12.1

Bug fixes

• Daily history is now a continuous calendar. The "Daily Tokens (14 days)"
  and cost charts skipped days with no Claude usage, packing active days
  together so the date axis wasn't continuous. GetHistory now returns the full
  window of N days ending today, zero-filling idle days so they render as empty
  bars with their date labels. This also corrects a latent off-by-one where the
  window spanned N+1 days.

• Stopped a self-inflicted rate-limit storm after a rejected token. When the
  usage API rejected the access token with a 401, the poller marked auth expired
  but the next credential reload reset it to OK on the still-future local
  ExpiresAt, so polling resumed, drew another 401, and looped every 30s until
  the API returned 429. The auth provider now remembers the rejected token and
  stays expired until a genuinely different token is written to disk (i.e.
  Claude Code refreshed it), so the poller waits quietly instead of hammering
  the API.

v0.12.0

Forecast model → v1.2

• The projection interval was over-spread - the complement of the v0.11.1
  fix, not a reversal. That release corrected an under-spread interval by
  widening the rate-uncertainty term. This one corrects a different
  component, the path-noise term: its §5 calibration fit the variance with
  an unweighted regression of squared errors, which is heteroskedastic, so the
  few long-horizon points dominated the fit and inflated it (and the
  over-subtraction silently pinned the rate prior). On real data the over-spread
  was path-dominated, so the two fixes touch orthogonal terms of the same
  interval. v1.2 weights that regression by 1/Δt², calibrating the spread and
  reviving the prior; net, out-of-sample coverage converges toward its 80%
  target. The v1.1 spec is preserved under
  internal/forecast/archive/v1.1/; the
  math is documented in
  internal/forecast/CHANGELOG.md.

• Removed the Low/Medium/High confidence badge. It scored how much recent
  data the forecast had, not how reliable the forecast actually was, so it
  could read "High" next to a wide interval and was uncorrelated with real
  accuracy. The 80% CI already conveys the uncertainty.

New: claumon bench

• An out-of-sample benchmark for the forecast model: leave-one-session-out and
  temporal-holdout protocols, CRPS/pinball proper scoring with coverage, MAE,
  and bias breakdowns, segmented by engagement and horizon. Datasets are
  reproducible: frozen fixtures exported from any device's store
  (claumon bench export --db ...) and seeded synthetic regimes. A development
  and validation tool; it does not affect the dashboard.

v0.11.1

Fixes

• Forecast 80% CI was systematically under-spread. Live monitoring
  showed only ~25% of session outcomes falling inside the displayed band.
  The §5 calibration regression was discarding its quadratic coefficient,
  which carries the historical-average rate variance; v1.1 retains it and
  uses it as a floor on the per-forecast rate uncertainty. CIs widen
  accordingly, most visibly at long horizons.

  Forecast model version → v1.1. The v1.0 spec is preserved under
  internal/forecast/archive/v1.0/;
  the math change is documented in
  internal/forecast/CHANGELOG.md.

Diagnostics

• claumon diagnostics now prints a "Spread sanity" block: mean squared
  error of F, mean predicted variance, an underspread ratio
  (1.0 = calibrated, > 1 = bands too narrow), and the components feeding
  the new EffectiveRateVar. Use it after a few days of v1.1 data to check
  the fix actually landed for your usage pattern.

Docs

• Model spec §3 picks up a paragraph documenting the Brownian-motion
  simplification (utilization is monotone, BM isn't) and the conditions
  under which the bias matters.

v0.11.0

Features

• Forecast trajectory modal. Click "Projected X% at reset" on a session or weekly gauge to open a modal showing the actual Monte Carlo trajectories the §6 simulator is producing: a fog of ~120 sampled paths, the empirical 10/90 percentile band traced from the same paths, the posterior mean line, observed snapshots so far, and a small first-passage histogram on the threshold line with the MC median ETA. New endpoint GET /api/forecast/sample?gauge=session|weekly re-runs the MC with trajectories collected; payload is subsampled in both the trajectory and time dimensions so weekly stays under ~200KB. The §6 MC core is shared with EstimateETA so the modal and the cached ETA summary always agree.

• Versioned forecast model. The forecast spec now carries a forecast.ModelVersion identifier (v1.0 today). It's surfaced on every /api/forecast and /api/forecast/sample response, in the modal subtitle, and in the diagnostics report. Retired specs will move to internal/forecast/archive/<version>/ when the model changes meaningfully; internal/forecast/CHANGELOG.md summarises each bump.

• claumon diagnostics subcommand. Replays the forecaster across past completed sessions and prints calibration metrics (80% CI coverage, MAE of F, ETA accuracy) stamped with the model version. Useful for deciding whether a model change actually improves accuracy before shipping it.

Internals

• CI workflows bumped to Node-24-compatible action versions (actions/checkout v4→v6, actions/setup-go v5→v6, goreleaser/goreleaser-action v6→v7); workflow Go version aligned to go.mod (1.26).

v0.10.1

Fixes

• Forecast lower bound no longer dips below current utilization. The 80% CI was being clipped to [0%, 100%] for display, but utilization within a reset window is monotone non-decreasing, so the Gaussian left tail going below u_now produced unphysical readouts like "current 6%, 80% CI: 0%-16%". The displayed bounds are now floored at u_now. The unclipped point forecast F is preserved for downstream ETA computation, and the model spec in internal/forecast/MODEL.pdf is updated to match.

v0.10.0

Forecasts on the rate-limit gauges

Each gauge now shows a projection of where utilization will land at reset, with an 80% credible band, an ETA to threshold (default 100%), and a LOW/MED/HIGH confidence pill. This is an empirical Bayes forecast: the prior on the rate and the path-noise variance are both estimated from past data and plugged in to a posterior update, rather than being marginalized under a hyperprior. The full spec lives in internal/forecast/MODEL.pdf; the gist:

• Rate posterior. Inside the open window, utilization is modeled as u(t) = u_now + r·t + W(t), with r an unknown per-hour rate and W Brownian path noise with per-hour variance σ². The current rate is fit by OLS on the last 30 minutes of snapshots; that OLS slope and its standard error are treated as a Gaussian likelihood r̂_OLS | r ~ N(r, SE²_OLS), then fused with a Gaussian empirical prior on r (mean μ₀, variance τ₀²) via the standard conjugate normal-normal update. The prior is refit daily from up to 200 completed past windows.

• Closed-form projection. Both the rate-uncertainty and path-noise pieces are Gaussian, so projected utilization at reset is F ~ N(u_now + r̂·ΔT, ΔT²·τ_post² + ΔT·σ²). The 80% credible interval is F ± z₀.₉·σ_F, clipped to [0%, 100%] for display. (Surfaced to users as "80% CI" for brevity.)

• Path-noise calibration. σ² is recovered by replaying the forecaster across past windows: at each replay point, the squared forecast error e² against the actual u_final is regressed against [ΔT, ΔT²] with no intercept. The linear coefficient is path noise; the quadratic coefficient absorbs rate-uncertainty contamination and is discarded. This is a method-of-moments estimator, then plugged into the posterior; the prior is refit once more with the noise correction applied to its sample variance.

• Monte Carlo ETA. For each threshold, 500 trajectories of the SDE are simulated at 5-minute steps, drawing one rate sample per trajectory from the posterior. The reported ETA is the median first-passage time with the 80% CI from the 10th/90th percentiles. If at least half the trajectories never cross before reset, the threshold is reported as unreachable (p_inf is exposed on the API payload).

• Confidence tag. Derived from effective sample size n_eff = min(n_recent, τ₀²/SE_OLS² + N_sessions). n_eff ≥ 50 → HIGH, ≥ 15 → MEDIUM, else LOW. Falls back to the prior alone when fewer than three recent snapshots exist.

The forecast is computed once per poll and shipped inside the existing usage SSE event; the same payload is also exposed at GET /api/forecast for pull-style clients. It is suppressed when fewer than two completed past windows exist, when the window has just reset, or when a drop in the recent snapshot series indicates a missed reset between polls.

Fixes

• Canonicalize reset_at timestamps at write time. The Claude API returns reset_at recomputed as now + remaining on each poll, so the same nominal window drifts by hundreds of milliseconds across polls and occasionally straddles a minute boundary. Snapshots were being written with these drifting strings, so GROUP BY session_reset_at shattered every window into singletons and downstream aggregation lost track of session boundaries. Reset times are now rounded to the nearest minute on write, and a one-time idempotent migration canonicalizes existing rows on startup.

v0.9.3

Performance

• Stop re-parsing every JSONL on every session change. The watcher callback called DiscoverDailyAggregates, which re-read and re-parsed every session file under ~/.claude/projects on each event. With many concurrent live sessions, this burned hundreds of MB/s of I/O and pegged a CPU core. Parsed per-day buckets are now memoized in-process and invalidated by (path, mtime, size), so only changed files are re-parsed.

v0.9.2

Fixed

• Poll loop no longer burns API requests after auth expires. When the API rejected the cached token with 401 and the keychain still held a stale token, authStatus could remain ok, so each wake-up hit the API four times before a 429 forced a 1h backoff. The provider is now force-marked expired on any 401 from the usage API, so subsequent polls skip the API entirely until credentials are refreshed externally.

v0.9.1

Fixed

• Heatmap and today's tokens no longer include prior days — sessions resumed across midnight credited their full lifetime tokens to today's aggregates and the hourly heatmap. Both are now bucketed by each message's own timestamp.
• Session detail timestamps now show dates across day boundaries.
• Process liveness detection on Windows.
• Time-sensitive store tests that assumed March 2026 dates.

Changed

• Expanded tool call detail in the session view.

v0.9.0

Added

• Claude Design weekly usage gauge — dashboard now surfaces the weekly Claude Design utilization (from the seven_day_omelette field in the OAuth usage API) alongside the existing Opus and Sonnet gauges. The card is hidden when usage is 0%.