iamwhatever · Oct 11, 2026 · Oct 5, 2026
@@ -201,6 +201,7 @@ and Forget. Legacy schedule rows state that the run uses Global Memory V1.
 | Embeddings | Enable the vector store, pick its model, and view localized setup errors, warnings, and deferred repair status | Memory browser → vector card | `pages/overview/VectorMemoryCard.tsx`, `pages/overview/EmbeddingModelCard.tsx`, `pages/overview/embeddingStatusText.ts` (shared status-text helper; no new entry point) | `handlers/memory.py` | `GET /api/memory/embedding-status`, `POST /api/memory/enable-embeddings`, `POST /api/memory/embedding-model` |
 | Memory graph | Entity/relation visualizer over the memory store | `/developer?tab=memory` | `pages/overview/MemoryGraphTab.tsx` | `handlers/memory.py` | `GET /api/memory/graph`, `GET /api/memory/observability` |
 | Usage | Token and turn usage over time, including estimated Kiro CLI tokens for this month, last month, and each day | `/settings/overview?view=usage` | `pages/overview/UsageTab.tsx` | `handlers/usage.py`, `handlers/telemetry.py` | `GET /api/usage`, `GET /api/usage/kiro`, `GET /api/usage/turns` |
+| Spend over time (credits) | Stacked area chart of the last 30 days' credit spend, layered by channel (where each session came from), agent, model, or the week a session was first seen in the window (the oldest layer, **Started by** a date, also holds sessions that began before it); cumulative by default or per day; hover or the arrow keys read out a day's total and each layer's value, the top layers are kept with the rest folded into **Other**, and a window the server had to shorten says how many of its oldest turns were left out and from which day every turn is counted; shown only when the default harness bills in credits (kiro-cli, KAS), with a notice in its place when that setting could not be read | `/settings/overview?view=usage` → **Spend over time (credits)** card, above Session activity (30 days) | `pages/overview/TokenStackedAreaChart.tsx` | `handlers/usage_series.py` | `GET /api/usage/series` |
 | Kiro credit balance | The top-bar credit pill and its account modal. The gateway refreshes the reading on a timer: the free `GetUsageLimits` API first, then the `kiro-cli /usage` scrape (a local slash command over the same free call) whenever the API returns no plan, parked for 6 h after three consecutive scrape attempts that failed (timed out, errored) or produced no parseable plan. The modal's **Refresh** button runs that same refresh now — under the state's notice when there is no reading ("Could not read your balance." for a failed read, "No balance reading is available for this account yet." for `none` — a dash the pill shows only while the selected harness is kiro-cli (`agent.acp_backend` names the kiro id, read positively by `api/acpBackend.ts`); on any other harness that payload hides the pill, as it always did), compact beside the meter when there is one — and writes the result into the pill's query. Single-flight (409 while a refresh runs, shown as an `ErrorNotice` with the hand-off like every other refresh outcome, Refresh held until the modal's ~3 s re-read lands and the notice clears); one notice at a time (a failed refresh replaces the standing failure notice with the same one sentence, the stale line yields to the stale-refresh notice); Refresh is held while the parked notice says when to try again; a parked scrape is reported after the API attempt (`skipped: scrape_parked`): no new reading is fetched, and the cache then holds a same-identity prior reading dimmed `stale` or an unavailable marker. A refresh that comes back `stale` is reported as such — an `ErrorNotice` with the hand-off — rather than stamped as checked now. Every credential read — the API call and the `/usage` scrape alike — is bracketed by a whoami immediately before and one immediately after, and its reading is published only when the pair PROVES one account (`_whoami_agreement` = `same`: at least one matching identity field, none disagreeing); `different` and `unproven` publish nothing (two identity-less whoamis around the API fall through to the scrape; around the scrape they end the refresh with no reading), and every failure path keeps a prior reading only for the same account (`_same_identity`), so an account switch never shows the previous account's balance. Under the meter, one muted line projects whether the plan lasts to the reset (`api/creditForecast.ts`): the pace is the account's own used credits over the days the cycle has run (cycle start = one calendar month before `resets`), shaped by weekday from the Usage tab's Daily History credits (the shared `providerUsageQuery` over `GET /api/usage/kiro`, last 28 days, flat until 14 days of history); the line is omitted when no reset date is known, the plan is used up, bonus credits still hold a balance, less than a day of the cycle has run, or the report is still loading or refreshing, a failed read shows an `ErrorNotice` with the hand-off, and the run-out case carries the warn tone. No notification yet (#17112). No config key, no Settings control, no cost copy | Top bar → credit pill → **Kiro Account** (desktop); on the phone the pill is not shown and the same modal opens from the nav drawer → bottom group → **Kiro Account** row, or on the chat route from the chat drawer → nav rail → **Kiro Account** tile | `shell/topbar/kiroUsageReadout.tsx`, `App.tsx` (mounts the modal), `components/KiroAccountModal.tsx`, `api/kiroUsage.ts`, `api/creditForecast.ts` | `handlers/sessions.py`, `handlers/usage.py` | `GET /api/sessions/usage`, `POST /api/sessions/usage/refresh`, `GET /api/usage/kiro` |
 | WakaTime coding activity and export | Coding stats for a named range plus project-grouped hours over a date range as a CSV or JSON download, for the productivity view and billable-hours invoicing | `/settings/overview?view=wakatime` | `pages/overview/WakaTimeTab.tsx` | `handlers/wakatime.py` | `GET /api/wakatime/stats`, `GET /api/wakatime/export` |
 | Portability | Export and import the whole memory/config bundle, optionally with persistent chat history (add-only on import); both ends warn about crew agent templates the bundle does not carry. A memory-only export (`?components=memory`, never chats) asks first with a dialog listing what the file holds; a memory-only import (`?components=memory`, or any archive whose manifest declares memory only) merges just the memory and refuses Replace | `/settings/imports` | `pages/overview/PortabilityTab.tsx` | `handlers/portability.py` | `GET /api/portability/export`, `POST /api/portability/import`, `POST /api/portability/preview` |

@@ -137,6 +137,74 @@ refresh can recover without clearing the report. A provider without usage
 support shows neutral status text, not an error. No report is persisted to disk,
 and the top-bar credit readout keeps its separate billing refresh policy.
 
+The Usage tab's "Spend over time (credits)" card is a stacked area chart with one layer
+per bucket of the chosen dimension, backed by `GET /api/usage/series`
+(`dashboard/handlers/usage_series.py`). The route reads the same per-turn
+shards as the Daily History credits, admits rows by the same guard (a `tokens`
+row with a parseable local day), and answers a dense local-day axis -- every
+day of the shard retention window, zero-filled -- plus one series per bucket
+of the requested dimension: `by=channel` (the default), `agent`, `model`, or
+`cohort`, the ISO week in which the row's session was first seen inside the
+window, keyed by its Monday and assigned in a second pass once every session is
+known. The channel is derived from the row's session key with
+`telemetry_channel_of`, never read from the row's own `surface` field, for the
+reason `cost_breakdown` documents: historical rows can carry a wrong but
+non-empty surface and the row schema has no writer marker, so the key is the
+authority and this chart buckets spend by the same closed channel vocabulary as
+the Spend panel. The model is canonicalised across provider id migrations like
+the daily chart does. The value stacked is the row's `credits`, the unit the
+kiro-cli backend and its KAS relay bill in; a harness that bills in tokens or
+dollars writes rows with no credits, so the Usage tab mounts the card only once
+a loaded config names a credits-billing default harness (`agent.acp_backend`
+via `reportsCredits` in `api/acpBackend.ts`, stated positively like
+`isKiroBackend`), never on a guess and never as "not claude"; a config read
+that failed with nothing cached (`failedWithNoData`) is said in an
+`ErrorNotice` where the card would stand, carrying the failed request's report
+for the agent hand-off and a Try again button that re-reads, since on a
+kiro-cli install a silent absence would read as "this harness has no credits"; turns a crew
+member ran on another harness (`agent.member_acp_backend`) land in the same
+store and count whatever credits they recorded. The top seven buckets by
+window total (the session palette's hue count) are kept and the rest fold into
+one `other` series carrying its member count; a row whose dimension value is
+empty or absent lands in an explicit `unattributed` series rather than being
+guessed at or dropped, so the stack's top edge always equals the day's total
+spend. Series arrive in stack order, bottom first: value dimensions
+largest-first, cohorts oldest-first, then `other`, then `unattributed`. An
+unknown `by` is a 400 (`invalid_dimension`) rather than a silent substitution,
+and an app token is answered 404 with a
+SEL `app_isolation` audit record -- the payload is the whole install's spend
+with no slot filter, unlike the row-isolated `/api/usage/turns`. The parsed rows
+are cached on the shard fingerprint the other readers use, so switching
+dimension re-aggregates in memory; the cache is bounded on both axes
+(`MAX_ROWS` rows, `MAX_FIELD_CHARS` per retained string), newest shards are read
+first so a window past the cap keeps its newest rows, and the payload discloses
+the shortening (`truncated`, `dropped_rows`, and `complete_from`, the oldest
+day every row of which was kept, never reaching past an unreadable shard --
+`null` when the cap tripped inside the newest shard) instead of serving a
+partial window as the whole one. The chart stacks
+and, in its default cumulative view,
+prefix-sums the series client-side (`pages/overview/TokenStackedAreaChart.tsx`,
+`d3`'s `stack`/`area`, axis top rounded to a nice number); its view choices
+persist in browser storage; the top-N buckets take the session palette's hues
+and `other` / `unattributed` the muted tokens; channel ids are shown capitalised
+with the unrecognised-key channel labelled apart from the `other` fold; the plot
+is a keyboard-operable slider over the day axis whose value text reads the
+focused day out; an `unattributed` layer earns a one-line caption under the
+controls and a truncated window a warning-toned one that leads with the
+consequence (the totals undercount) and, when the server could name it, from
+which day every turn is counted. Cohort layers are labelled "Week of" their
+Monday except the oldest one drawn, labelled "Started by" its Sunday (or the
+axis's last day while that week is still running): a cohort is the week a
+session was first seen inside the window, so the oldest layer also holds
+sessions that began before the window, and the cohort caption says so. A failed first load shows a titled `ErrorNotice`, the
+server's text on its own line under the title, with a Try again button that
+refetches; a failed refresh over a drawn chart shows the same shape above the
+stack it keeps showing. Changing the dimension
+keeps the previous stack on screen, dimmed and `aria-busy` with a "Loading…"
+cue beside the select, until the new answer arrives (`keepPreviousData`); each
+answer carries the dimension it was split by, so the stand-in keeps its own
+labels and caption rather than the new selection's.
+
 ## Gateway restart
 
 The dashboard restart endpoint and successful update applies share

@@ -650,6 +650,7 @@ async def api_tasks_summary(request):
     api_kiro_usage,
     api_usage,
 )
+from kiro_crew.dashboard.handlers.usage_series import api_usage_series  # noqa: E402, F401
 from kiro_crew.dashboard.handlers.wakatime import (  # noqa: E402, F401
     api_wakatime_export,
     api_wakatime_stats,