cctally 1.9.0 → 1.10.0

package/CHANGELOG.md

@@ -5,6 +5,30 @@ based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+## [1.10.0] - 2026-05-20
+
+### Added
+- Add Projects panel + modal to the dashboard — top-5 current-week leaderboard with horizontal bars and attributed `Used %`, a configurable `1w` / `4w` / `8w` / `12w` Projects modal with a stacked-area trend chart, 7-column per-project table (cost desc by default), and an in-place drill into model breakdown + recent sessions; two-way navigation with Sessions (clicking a project cell in Sessions opens the Projects modal pre-expanded on that project's drill, and clicking a session in the drill opens the SessionModal as a replace). New global `5` shortcut opens the Projects modal directly (the panel sits at index 4 of `DEFAULT_PANEL_ORDER`); `0` now opens whichever panel occupies position 10 (vim-style "10 wraps to 0"; `alerts` in the default order). A lazy `GET /api/project/<key>?weeks=N` endpoint mirrors the SessionModal stale-while-revalidate pattern — the modal renders from the envelope's top-N summary first, then hydrates the full drill on demand. Three new share templates (`projects-recap`, `projects-visual`, `projects-detail`) join the kernel and carry the active `windowWeeks` into the share flow.
+
+### Fixed
+- ProjectsModal mobile (≤640px) is no longer cramped: the 7-column table reflows to a stacked-card layout (project key + cost on row 1; sessions · used % · % week · first → last seen on row 2), a sort-cycle pill replaces the SortableHeader, and the per-project drill renders inline directly under the selected card. Closes #73.
+- dashboard Projects modal: `$/1%` column replaced with `% of week`. The original column was mathematically degenerate — `attributed_pct` is computed as `(cost[p] / total_cost) * weekly_used_pct`, so `cost[p] / attributed_pct` collapses to the constant `total_cost / weekly_used_pct` for every project (independent of `p`). Every row in the table showed the same value at every window pill, adding no per-row signal. Replaced with `% of week` = `windowCost / sum(windowCost)` over the active window — the one orthogonal axis the data already carries — rendered as a 0–100 percent and sorted desc by default (same shape as the column it replaces). `windowPct` (the "Used %" column) is unchanged. Column id `share_of_window`; field `ProjectsTableRow.shareOfWindow` (replacing `dollarsPerPct`). Spec §3.4 row updated; the F5 follow-up `$/1%` scatter chart deferred in §1 / §A.7 is unrelated and intact. Regression `<ProjectsModal /> > '% of week' column (#72)` in `dashboard/web/src/modals/ProjectsModal.test.tsx` asserts the new header, the absence of the legacy `$/1%` header, and the per-row share values on a 3-project / 1w fixture. Closes #72.
+- dashboard Projects modal: stacked-area trend chart now renders visibly when only one week of data is in scope (fresh installs or the `1w` pill). Previously `ProjectsTrendChart`'s `xFor` collapsed every point to `VW/2`, drawing each polygon as a zero-width vertical line; the chart appeared empty even though the table and totals showed real activity. `xFor` now synthesizes a `[VW*0.1, VW*0.9]` horizontal span when `weekCount === 1`, and the polygon walk emits both edges of that span at each accum-y so the series closes as a 4-corner rectangle (a wide stacked-bar segment) instead of a line. Multi-week renders (`4w` / `8w` / `12w`) are byte-identical — the new branch is gated on `weekCount === 1`. Regression `<ProjectsTrendChart /> 1-week render` in `dashboard/web/src/modals/ProjectsTrendChart.test.tsx` asserts non-trivial x-extent + closed-quad point counts on the 1-week path and pins multi-week geometry as a sanity check. Closes #68.
+- Mirror tool now propagates file deletions/additions when `.mirror-allowlist` edits change classification, with a release-time fingerprint precheck + `--reconcile` recovery mode. Closes #62.
+- dashboard Projects panel: `projects.current_week.rows` now populates correctly after an Anthropic mid-week reset. Previously `_build_projects_envelope` left the shifted `TuiCurrentWeek.week_start_at` (e.g. Friday 13:00 UTC post `_apply_midweek_reset_override`) in place as the bucket-lookup key while the entry aggregator anchored every row to its ISO-Monday — the lookup missed every Monday-keyed bucket and emitted `rows: []` / `total_cost_usd: 0.0` despite real activity. The envelope now canonicalizes `cw_start` through `_projects_week_start_monday_utc` so the lookup matches whichever bucket the aggregator wrote, matching the documented invariant: "`TuiCurrentWeek.week_start_at` is NOT a valid `week_start_date` lookup key after a mid-week reset." Regression `test_current_week_rows_populated_after_midweek_reset` in `tests/test_projects_envelope.py` drives a Friday-13:00-UTC stub and asserts populated rows; `tests/fixtures/dashboard/{reset-week,ok,warn,over,tz-override}/golden-data.json` re-rolled to capture the corrected output.
+- dashboard share modal: `windowWeeks` selected via the Projects modal's `1w` / `4w` / `8w` / `12w` pill now reaches the server. Previously `shareModal.params.windowWeeks` was set by the modal but never read by the share path (`ShareModal`/`ActionBar`/`PreviewPane` ignored `slot.params`), so the server's `options.get("windowWeeks", 1)` fallback rendered every share artifact as 1w regardless of the pill. `windowWeeks` is now a typed field on `ShareOptions`; `ShareModalRoot` forwards `slot.params` as `initialParams` to `ShareModal`, which seeds it into the initial options state — the existing preview + export fetches naturally include it in the request body. Regression `I1 regression: windowWeeks param threads through /api/share/render body` in `dashboard/web/src/share/ShareModalRoot.test.tsx` asserts on the captured fetch body, not the store slot.
+- dashboard Projects modal: "Sessions" / "First seen" / "Last seen" columns are now window-scoped per spec §3.4 — they reflect the active `1w` / `4w` / `8w` / `12w` pill instead of fixed 12w / all-time values. `_build_projects_envelope` emits new per-week arrays `sessions_per_week` / `first_seen_per_week` / `last_seen_per_week` on each `trend.projects[]` entry (replacing the window-unaware `sessions_count_12w` / `first_seen_at` / `last_seen_at` scalars); `ProjectsModal.tsx` slices these to the active window pill and reduces them client-side. Column labels revert to bare "Sessions" / "First seen" / "Last seen" (the I2 stopgap's "(12w)" / "(all-time)" widening is gone). Reconcile R-PROJ6 in `bin/cctally-reconcile-test` locks in the per-week-array invariants (length == `len(weeks)`, count ≥ 0, presence parity between sessions count and first/last seen). Regression `flipping the window pill rescopes Sessions / First seen / Last seen cells` in `dashboard/web/src/modals/ProjectsModal.test.tsx`. The four dashboard goldens (`tests/fixtures/dashboard/{ok,warn,over,reset-week}/golden-data.json`) re-rolled. Closes #71.
+- ProjectsTrendChart legend on mobile (≤640px): replaced the 2-line silent clamp with a horizontal-scroll row (`scroll-snap-type: x proximity`), `(other)` pinned last, right-edge fade-mask affordance, and tap-to-drill on each project's legend item. Closes #74.
+- dashboard share Projects path: thin-history dashboards no longer mis-label exported artifacts. `_build_projects_share_panel_data` clamped its row aggregation via `take = min(weeks_back, n_weeks)` but historically returned the unclamped `weeks_back` for both `window_weeks` and the `period_start` bound, so a fresh install requesting the 12-week share template rendered as `Last 12 weeks` with a 12-week date range over only (say) 3 weeks of data. The builder now derives `effective_weeks = max(1, take)` and routes it through both the label and the period bounds so the share artifact's framing always matches the rows it carries. Regression: `tests/test_share_projects_window_clamp.py` (4 cases — clamp, no-clamp, empty-trend, panel-mode unaffected).
+- dashboard Projects modal sort: descending sorts on nullable columns (`Last seen`, `Used %`, `% of week`) now keep `—` rows pinned at the bottom. The previous `nullsLast` helper returned `1` / `-1` from each column's comparator, but `applyTableSort` multiplies the comparator's return by `sign` to flip direction — under desc the null-parking constants flipped sign too, pulling the `—` rows to the top. Null-last semantics moved off the comparator and onto a new `TableColumn.nullKey` extractor that `applyTableSort` checks BEFORE applying the direction sign, so null rows are parked at the end unconditionally. Regression: `dashboard/web/__tests__/tableSort.test.ts` (asc + desc both place null rows last on a mixed-null fixture).
+- dashboard Projects panel: credited weeks no longer overstate `Used %`. `_build_projects_envelope` historically loaded the per-week `weekly_percent` via `MAX(weekly_percent) GROUP BY week_start_date`, which relies on the "weekly_percent is monotonic within a week" invariant that **breaks** on weeks where Anthropic ships an in-place goodwill credit (the same v1.7.2 surface that `week_reset_events` tracks): pre-credit `60.0%` and post-credit `5.0%` coexist for the same `week_start_date`, MAX returns the stale pre-credit high-water mark, and every per-project `attributed_pct` plus the trend's `total_pct` get pinned at the HWM forever. Switched to the canonical "latest snapshot per week" pattern (read rows ordered by `captured_at_utc ASC, id ASC` and let later non-NULL values overwrite — same shape as `_select_last_known_snapshot` and the doctor credited-week check). The `reset-week` dashboard fixture's golden flipped from `total_pct: 60.0` to `total_pct: 5.0` (and the two contributing projects' `attributed_pct` rescaled proportionally from `37.9% / 22.1%` to `3.16% / 1.84%`), confirming the post-credit values now drive the Projects panel/modal.
+- dashboard Projects modal: keyboard shortcuts (`1` / `4` / `8` / `0` / `s` / `↑` / `↓` / `Enter`) no longer leak through the share / composer overlay when it's layered above the modal. The share overlay only registers `Escape` at `overlay` scope, so single-char and named keys previously fell through to the `modal`-scope bindings and silently mutated `projectsWindowWeeks` / `projectsTrendYMode` / row selection on the hidden Projects modal underneath — dismissing the share dialog then dropped the user into a different view than the one they were sharing. Each binding now gates on `shareModal === null && composerModal === null` via a shared `isProjectsTopmost()` predicate.
+- dashboard `GET /api/project/<key>?weeks=N` drill endpoint: `window_attributed_pct` now scopes to the requested window. The HTTP path reuses `snap.projects_envelope` (built by the sync thread with `weeks_back=12`), so `matching_trend["weekly_pct"]` is always a 12-element array — summing it without slicing made `1w` / `4w` / `8w` drills report the 12-week attribution total, and the answer depended on whether the helper rebuilt the envelope or reused the snapshot. The drill now trims `weekly_pct` to the trailing `weeks_back` entries before summing (same `[-take:]` slice the share-envelope adapter at `_cctally_dashboard.py:1629` already uses).
+- dashboard Projects drill (`ProjectsDrillPanel`): stale-on-switch guard now also treats a window change for the *same* project as stale. `useProjectDetail` keeps the previously-rendered payload mounted across `(projectKey, windowWeeks)` changes (SWR pattern), but the guard only compared `data.key !== projectKey`. Switching the modal's window pill (e.g. 12w → 4w) on a heavy project therefore left the prior window's cost / models / sessions rendered under the new `{windowWeeks}w` heading for seconds while `/api/project` resolved. The guard now also checks `data.window_weeks !== windowWeeks`, so the drill always agrees with the visible window header during a re-fetch.
+- dashboard Projects modal: persisted selection is re-bound to the leader (or cleared) when the previously-selected project drops out of the current window's trend rows. Without the re-bind, narrowing the window after selecting an older project — or cross-navigating from an older Sessions row while the modal is on a 1w/4w window — left the drill pointed at a row that no longer exists in the data (endpoint 404 on desktop, missing inline anchor on mobile). A new effect re-evaluates `selectedKey` against `env.projects.trend.projects[]` whenever the trend rows change and falls back to the current-week leader when the prior selection is absent. Cross-nav (`store.openProjectKey`) still takes priority above this branch so explicit deep-links into the modal are never overridden.
+- dashboard Projects share path: `_build_projects_share_panel_data` now clips `period_end` to `min(cw_start + 7d, _share_now_utc())`. The rows in this panel_data are week-to-date (current_week rows aggregate through "now", and the multi-week branch's trailing slice is also week-to-date), so a mid-week export advertised `period_end = cw_start + 7d` — a future reset that the rows don't include — in the rendered period and frontmatter. The frontmatter now agrees with the live dashboard's "spent this week" KPI, which `_build_current_week_share_panel_data` symmetrically clips to `now`. Regression: `test_period_end_clipped_to_now_when_mid_week` + `test_period_end_uses_week_end_when_now_past_reset` in `tests/test_share_projects_window_clamp.py`; existing clamp tests pin `CCTALLY_AS_OF` past `cw_start + 7d` so their period_end assertions remain valid.
+- dashboard Projects modal mobile sort cycle: `first_seen` step now defaults to `asc` (matching `PROJECTS_COLUMNS`'s `defaultDirection`). The mobile pill previously hard-coded `first_seen: desc`, which (a) produced the reverse order from the desktop column header on the `first` step and (b) left a persisted desktop `first_seen` sort unrepresentable in the mobile cycle's `findIndex` lookup — the cycle would silently fall back to cost-desc on the first tap on mobile. The cycle and the column registry now agree on direction.
+
 ## [1.9.0] - 2026-05-19
 
 ### Removed