Dashboard setup status, cost details and trend interaction

2026-10-04. Existing local-source, reference-pricing and history-retention rules apply.

Setup and collection status

Overview retains one status line with enable/details actions. Only missing configuration eligible for automatic setup counts as waiting to enable. Other items without verified valid data display “No data”, including those currently ineligible for automatic setup; valid incoming data changes status to “Verified”. Hide zero-count configuration-limit items. Show remaining configuration conflicts separately even with valid data. Details always distinguish configuration state, restriction reasons and data checks; “No data” does not permit otherwise restricted writes. Details inspect actual output targets automatically and read-only. File presence does not establish valid data: existing adapter field rules check parsable telemetry with read-size/range limits, without importing bodies or creating calls. Wait when data is absent; valid samples verify observed data, without clearing current configuration conflicts. Preserve external targets and do not claim success for shared output whose client cannot be identified. Fix duplicated JSONC insertion newlines while retaining comments, BOM, newline style and other keys.

Recheck every 30 seconds and merge concurrent checks. Inspect the last 2 MiB first; if no matching call appears, also inspect the first 8 MiB. Each sample allows at most 2,000 lines, 256 KiB per line and 200 ms reading time. Identify the actual client; another Agent’s records in shared files do not verify it. If length, mtime, creation time and first/last fingerprints are unchanged, cache checks for at most five minutes and 128 entries. New output invalidates cache. Valid samples establish observed export, without guaranteeing complete historical or current-day coverage.

Poll status every 500 ms during collection and refresh usage afterward. Idle polling remains ten seconds and checks status without redrawing. After a manual request is accepted but before running is visible, retain the collection indicator and disable repeat requests until running or completion time changes. A short task finishing between polls still refreshes; do not revert prematurely to idle waiting.

Costs and layout

Today’s hourly model, Agent and Agent-plus-model groups all plot token curves, including one-hour ranges with the same axis/curve logic. Do not switch to call counts or bars. Unknown totals remain gaps. Tooltip rows show only name/value (≥ for lower bounds, — for unknowns); explain lower bounds once. Full coverage explanations remain available on demand. Limit tooltip width and wrap long model names. Cost cards show compact amounts, with coverage counts in tooltips. The cost panel normally shows amounts, gap counts and curves, expanding explanations on demand so currencies/long text do not make it much taller than neighboring components.

Compact monetary rows, converted amounts and tier/substitute hints must fit their card at both four-column and eight-column breakpoints. Bound each child to its content width and truncate long single-line text with an ellipsis. Tooltips retain the full original and converted amount ranges. Never change rounding or amounts to make text fit.

Settings use the existing styled switches for on/off options, with stable accessible names, keyboard Space support and visible focus. Export user/host selections keep checkbox semantics with custom check marks and a system high-contrast fallback, following the WAI switch pattern.

Costs are off by default. The dashboard shows current-price “API usage-based price reference”; historical event-time estimates/source amounts remain in the database/API without repeated dashboard display. Unknown amounts remain gaps; currencies never merge. Full-width model details add reference amounts and per-currency totals, retaining unpriced/partially covered usage without filling unknown tokens with zero.

One cost request returns provider/model subtotals and current-price amounts grouped by local day/provider/model for panel, table and curve reuse. Switch curves by currency and compare models without mixing currencies. Cost curves are daily even when usage is hourly; never distribute a daily amount into one hour. Weekly/monthly curves sum relevant days. Unknown days remain gaps; verified zero amounts may display zero. Usage/call charts, costs and model table use full width; two share charts sit side by side. Activity heatmap and weekday distribution, unaffected by chart selection, move last; preserve other order/sizes/visibility. Heatmap covers the selected whole year and weekday distribution the full queried date range, with independent scope explanations. Pie charts reserve separate legend space to prevent clipping/overlap; stack vertically on narrow screens.

The collapsible cost panel lists each used model’s actually matched current unit price per million tokens: provider, currency, input/output/cache-read/write tiers, context threshold, channel and price snapshot. Reuse each event’s matched price_id rather than averaging or selecting tiers again. Unmatched models show unavailable prices. Original model attribution and reference provider remain separate; unknown currency/rates stay unknown. The reference model table uses one amount column and per-currency totals. Today’s/current-range summaries include reference amounts over the same range when enabled. Selected hours filter retained individual calls by exact local hour, including both DST occurrences; never place whole-day costs in hourly cards. Seven metrics and costs use compact equal-height grids; input-coverage explanation remains in tooltips. Quota uses a separate compact row with expandable scope, snapshot time and explanation that account quota is independent of local tokens.

Local read-only checks found input/output on VS per-request spans for 2026-10-01/02, but the old adapter had not derived total. Derive only when both fields in the same span are known; missing fields, failed calls or overflow stay unknown. Rule correction replays consumed unchanged cursors; only complete valid replay marks completion. Keep the database/history. Old VS Code turn-input lower bounds plus whole-turn output yield only an observed lower bound, without replacing complete total. Keep Agent/model names in legends; tooltips use “name: ≥ value” or “name: —”. Explain lower bounds once, with full semantics available from the short note’s tooltip below the chart rather than repeated after every series.

Queries and selection

Measure real-database-copy timing, SQL query plans and call counts before optimization. Weekday distribution reuses daily summaries restricted by date/user filters, without reading raw messages. Archives unassignable to days retain unknown coverage. Avoid repeated per-model cost requests. Narrow price candidates by model while retaining priority/tier decisions. Add indexes based on query plans and preserve old-database compatibility; do not clear the database or rewrite history for an index.

“Last two calendar days” means yesterday/today in the selected timezone, using date boundaries rather than a rolling 24-hour window. This range and today default to hourly periods; a DST calendar day is not assumed to last 24 hours.

Trend usage, calls/sessions and cost curves, plus overview history/today’s hourly charts, support point/x-axis-label clicks and horizontal mouse-drag selection. On release query the continuous range by corresponding period boundaries, including reverse drags. Dragging selects rather than pans; wheel/slider zoom still chooses visible ranges. Show highlight, start/end and reset; reset clears every related highlight. Trend selection restricts summaries, model/Agent distribution, model-table usage and current-price references together. Overview history uses the same summary/distribution rules; today’s hourly selection independently restricts today’s summary/distribution/model/Agent table. History selection does not change today’s query. Daily cost points in hourly views select all hours of that day; hourly usage selections query exact call-hour costs, without allocating daily amounts. Bound selection to the queried range and keep cross-day timezone/ week-start consistent. Stale responses cannot overwrite new selections. Filter/period/user changes clear selection. Charts retain the complete query range; selection neither refetches curves nor creates query loops. During waiting/failure, show loading/gaps rather than the previous range’s distribution/model table.

summary accepts optional first_period/last_period label boundaries to restrict period rows after date/instance filters. Sessions/duration still use retained details from the same range, without adding period DISTINCT counts. Select hours by local labels; the chart’s single repeated-DST-hour label includes both actual occurrences. Weekly/monthly selections show actual ranges clipped by queried date boundaries. Pending summaries retain card layout and loading values to prevent jumps.

Keyboard selection

Time charts are focusable with localized accessible names. Arrows move, Home/End choose first/last period, Shift extends contiguous range, Enter applies the query. Mouse and keyboard use the same range/revision rules.

Validation and rollback

Test configured/waiting/valid/invalid/shared mixed data, current conflicts despite historical samples, newline preservation, model subtotals equaling currency totals, curves equaling current-price summaries, unit prices matching selected tiers, date/user/model filters, archive coverage, actual forward/reverse mouse drags, no query before release, rapid clicks/stale responses, consistent selected distribution/model usage/cost, calendar-day boundaries, full reset and wide/narrow layouts. Original sources/configuration stay read-only; scripts/copies belong under the task’s root build/ directory. Record final commands, environment, exit codes, performance measurements and gaps. Rollback removes this batch’s code/optional indexes without deleting sources, statistics or price snapshots.