feat: dashboard rework with eleven ranged pages and honest pricing

[KurutoDenzeru]

Description

Rebuilds the dashboard into a navigable page set and reworks the pricing ledger so every reported number traces to a recorded value.

The dashboard was one vertical scroll of components. It is now a sidebar shell with hash routing and eleven range-aware pages: Overview, Models, Providers, Costs, Requests, Errors, Traces, Tools, Frustration, Projects, Gain. Range picker in the header, 1–6 hotkeys, g + letter jumps.

The pricing ledger blended a default price into totals when it could not match a model id, so a single unknown model silently priced the whole account at Sonnet rates. A model id now resolves to its canonical vendor and provider, and three numbers are reported separately and never blended: measured (vendor-reported), API-equivalent (public rates), and cache savings. API-equivalent totals state coverage (N/M models priced) instead of hiding it. On this machine the API-equivalent total moved from $339.63 to $86.84 with 39/52 models priced and the 13 unpriced models named.

Related Issue

None. This branch started as a port of the omp-stats dashboard page set.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 📝 Documentation update
• 🎨 Style/UI update
• 💥 Breaking change (fix or feature that would cause existing functionality to change)
• ♻️ Refactor (no functional changes)
• 🧪 Test update
• 🔧 Configuration change

Changes Made

• Sidebar shell, hash routing, range picker, and the eleven pages above.
• Six ranges: 1h, 24h (default), 7d, 30d, 90d, all. A short range filters recent rows by exact timestamp and buckets them (5m inside the hour, 1h inside the day); a day range reads the payload's day tables. One payload serves every range.
• Session traces end to end: GET /sessions, /session/trace, /session/entry parse span timelines from the pi/omp transcripts (turns, model calls, tools, subagent tracks, markers, per-tool stats) with mtime and ETag caching. The Traces page renders a searchable session list, a per-track waterfall, and a span drawer with the raw transcript row. Every span comes from a recorded timestamp.
• Pricing: resolve a model to its real vendor instead of falling back to a default; report measured, API-equivalent, and cache savings as three numbers; state coverage and name the unpriced models.
• Usage rows now carry time to first token, provider, and working directory, so latency and project pages read recorded data rather than nothing.
• Requests gains a per-row TTFT column plus TTFT and median-elapsed strip stats. Overview gains Avg elapsed and TTFT, and its activity chart switches between tokens, requests, errors, and cost. Providers gets the sortable totals table, a stacked burn chart, and a peak-hour histogram. Errors groups failures by their recorded reason.
• Dashboard freeze fix: rangeView returned a fresh object each render, so every memo below it missed on each click and re-aggregated every day table on each expand, tab, and drawer. The view is now memoized on its inputs. The Models detail panel also mounted a recharts chart per expanded row; it now draws its day bars with plain elements.
• windowPartial counted any full recent list as a truncated window, so the default 24h range showed a false "lower bound" note on five pages. A window is truncated only when the list is full and every row in it falls inside the window.
• usage.db migrations guard against a truncated re-parse: a rebuild that keeps under half the rows restores the pre-migration backup rather than publishing the loss.

Screenshots/Recordings

Before	After
One vertical scroll of cards	Sidebar shell with 11 routed, range-aware pages

Visual verification during development was done against the omp-stats instance at 127.0.0.1:3847 on dark and light themes. Not captured here.

Testing

• I have tested this locally
• I have added/updated unit tests
• I have added/updated integration tests

bun run verify (frozen lockfile + build + test) is what CI runs: 62 files, 445 tests, all passing.

New coverage: test/shared/session-trace.test.ts (parser: omp rows, pi fallbacks, unterminated calls, child-track linking, pricing, directory shape) and dashboard/app/test/ for pages, routing, traces, theme, and settings. dashboard/app/test/interaction-cost.test.tsx spies on the aggregation and fails if a second fold runs inside a click, which is the regression guard for the freeze above.

Also verified: a served instance returning 200 with a /data.json carrying every field; the cap note appearing on 30d/90d/all and absent on 1h/24h.

Checklist

• My code follows the project's coding standards
• I have performed a self-review of my code
• I have commented my code where necessary
• I have updated the documentation accordingly
• My changes generate no new warnings or errors
• I have checked for accessibility compliance
• I have verified responsive design (if applicable)

Additional Notes

Lint reports warnings that predate this branch (react-hooks/exhaustive-deps on models, unicorn/no-array-sort, nested cell components); no new classes.

Labels applied: bug, enhancement, documentation, tests. There is no style label on this repo, so the UI part is covered by enhancement — say the word and I will create a style/UI label.

Two limits worth stating. Providers' "Subscription windows" and "Window utilization" sections are honest notes rather than empty tables: each needs per-account plan, quota, reset, and rate-limit data that the transcripts do not carry. And tokens-per-second is not reported, because dividing tokens by elapsed wall clock would conflate tool time with model speed.

[KurutoDenzeru]

Dashboard rework: payload, tokens-per-second, and the last freeze

Four fixes on top of 47fb48a, each measured before and after on the same machine, same data, same browser.

• /data.json is now gzipped and carries an ETag the client replays. 797 KB → 141 KB on the wire, and an unchanged poll returns 304 with 0 bytes instead of re-downloading the whole report every 5 seconds.
• The client no longer builds a second 800 KB JSON string per poll to detect change; it compares a four-field stamp.
• Tokens/s was inflated by roughly the request count. The Models column divided a model's total output by the mean duration of a single request: 20 tokens over a 1050 ms mean printed 19.0 instead of 9.5. The divisor is now total generating time (mean duration × run count). Against OMP's live values: step-5-preview 16.4 (OMP 15.1), deepseek-v4.1-flash 61.1 (OMP 61.3), TTFT 6.2s (OMP 6.0s).
• Two gateways of one model merged into one row. The fold keyed on the display label alone, so cline/cline-free/mimo-v2.6-flash and opencode-zen/mimo-v2.6-flash collapsed into a single row and cost landed on the wrong provider. The key now carries the serving segment.
• The last freeze. renderExpanded passed a fresh arrow on every render, so TanStack re-derived its expanded-row model on each parent render, including the one an expand itself triggers. Under a 6× CPU throttle, measured with Profiler and a longtask observer: expanding a Models row went from 3076 ms across seven long tasks to 43 ms with none. Every range switch now reports zero long tasks.

Verification: npx tsc -b clean, bun run test 446/446, and the expanded panel confirmed open in a real browser with the detail content rendering (screenshot in this thread). The tokens-per-second test fails when the fix is reverted, which I confirmed by reverting it and re-running.

Note on the residual tokens/s gap versus OMP: OMP counts cache-read and cache-write in throughput; tersio counts output only. TTFT matches to the decimal.

[KurutoDenzeru]

Superseded by #53, which is the same work carried to its conclusion. Closed rather than merged.

feat/dashboard-rework is not an ancestor of feat/omp-stats-dashboard, so 64 files are unique to this branch — the dash/, pages/, settings/, and app/ component trees, the @tanstack/react-table dependency, and /data.json. All of it is the architecture #53 replaced: components/agents/ and components/charts/ now hold the pages and the page kit, charts/table.tsx is a native sorting and paging table that needs no dependency, and api/agent serves per-view payloads where /data.json carried one whole report.

The behaviour fixes behind the architecture all survive, each reimplemented against omp's own database rather than Tersio's ledger:

• The reparse guard that restores a pre-migration backup when a rebuild keeps under half the rows: extensions/shared/usage-store.ts.
• Tokens per second over total generating time, not the mean duration of one request: extensions/shared/agent-stats.ts sums output_tokens * 1000 / duration and divides by the count of timed rows.
• One row per model per provider, so two gateways of one model no longer merge: the byModel group carries the provider in its key.
• Time to first token, provider, and working directory on request rows: read from omp's messages, which carries all three.
• The interaction-cost guard against re-aggregating on a fold: the client polls api/agent every 30 seconds while visible, and the server memoizes the aggregate for 5 seconds keyed on both database file stamps.

Two things here were deliberately dropped rather than replaced: the Usage page and its ledger-only surface, and the gain and frustration pages that read a store the reference writes from. Tersio's own ledger stays a product through tersio usage, tersio reset, and tersio dashboard --export.