12 Commits

Author SHA1 Message Date
mAi
173d7ddbb2 feat(views): Phase 5j slice A — paliad-shape schema redesign
Hard-replaces the 5i projax.views table per m's Q10 pick (2026-05-29):
no real data to preserve after a few hours, and the shape changes are
big enough that a clean recreate beats a 6-step ALTER.

Schema (migration 0017_views_redesign.sql):
- id (uuid), slug (text, format-CHECK'd, UNIQUE), name, icon,
  filter_json (jsonb — INCLUDES view_type per m's Q2), sort_field,
  sort_dir, group_by, sort_order, show_count, last_used_at,
  created_at, updated_at.
- DROPPED: pinned, is_default_for, view_type column. m's Q9 picked
  MRU (last_used_at) over per-page-default; Q2 placed view_type
  inside filter_json so the JSON owns the canonical render spec.
- Constraints: slug regex, sort_dir enum. NO view_type CHECK — the
  JSON-shape validator owns it now.
- Indexes: slug UNIQUE, (sort_order, name), (last_used_at DESC).
- updated_at trigger reused; projax_admin ownership preserved.

Store (store/views.go rewrite):
- View struct: Slug as the user-facing key; uuid kept on ID for the
  legacy `?view=<uuid>` 302-redirect path that lands in slice C.
- ListViews ordered by sort_order, name (matches sidebar).
- GetView(slug) + GetViewByID(uuid). MostRecentView() drives the
  /views landing redirect (slice B).
- TouchView(slug) bumps last_used_at fire-and-forget.
- ReorderViews([]slugs) wires the column for slice G's drag UI.
- CreateView server-assigns sort_order = MAX+1 inside the tx.
- UpdateView replaces every writeable field; renames are supported.
- Validation: slug format regex + reserved-list rejection +
  filter_json JSON well-formed check before round-trip.
- ErrViewNotFound / ErrViewSlugTaken / ErrViewSlugReserved /
  ErrViewSlugFormat surface to handlers as the typed error set.

Cleanup of the 5i overlay (drops what the new shape obsoletes):
- web/views.go: gutted to a stub. applySavedView, applyDefaultView,
  overlayURLFields, filterQueryToJSON, filterJSONToQuery,
  filterFromJSONPayload, anySliceToStrings + every old handler
  (handleViewsIndex, handleViewCreate, handleViewWrite, handleViewEdit,
  handleViewRedirect, handleViewDelete) deleted.
- web/server.go: dropped the /views route registrations and the
  applySavedView + applyDefaultView calls in handleTree.
  DefaultBanner data-map field removed.
- web/tree_filter.go: TreeFilter.ViewID field removed; ParseTreeFilter
  and QueryString stop reading/emitting ?view=.
- web/templates/views.tmpl and view_edit.tmpl deleted.
- web/templates/tree_section.tmpl: default-banner block deleted.
- web/views_test.go: deleted (every test was against the 5i shape).

Between slice A and slice B, /views/* URLs return 404 by design.
Slice B reintroduces the route family in paliad-shape:
  GET /views          → MRU landing
  GET /views/{slug}   → render
  GET /views/new      → editor
  GET /views/{slug}/edit → editor
  POST /views, /views/{slug}, /views/{slug}/delete → CRUD

Tests (store/views_test.go, new):
- TestViewSlugCRUD — create / get-by-slug / get-by-id / rename /
  delete round-trip, including rename-leaves-old-slug-gone.
- TestViewSlugFormatRejected — uppercase, underscore, leading dash,
  length-cap, empty all surface ErrViewSlugFormat.
- TestViewReservedSlugRejected — tree/dashboard/calendar/timeline/graph
  and friends all reject with ErrViewSlugReserved.
- TestViewSlugCollision — duplicate slug surfaces ErrViewSlugTaken.
- TestViewMRU — TouchView + MostRecentView ordering against a
  controlled pair of slugs (resilient to other suites' touched views).
- TestViewReorder — ReorderViews rewrites sort_order ascending.

Web tests stay green (the 5i overlay tests are gone, the rest don't
touch the views shape).
2026-05-29 11:41:28 +02:00
mAi
2f47b28f39 feat(views): Phase 5i slice D — saved views table + CRUD + sidebar entry
Persists named bundles of (filter + view_type + sort + group_by). Per m's
Q2 pick (2026-05-26), views are page-agnostic — `is_default_for` lets a
view become the auto-applied default for a page, otherwise views render
on whichever page accepts their view_type.

Schema (db/migrations/0016_views.sql):
- projax.views table with check constraints on view_type (5-value enum),
  sort_dir, is_default_for, and the kanban-needs-group rule.
- Case-insensitive unique name index (live rows only).
- One-default-per-page partial unique index.
- updated_at trigger; projax_admin ownership / grants.

Store (store/views.go):
- View struct + ViewInput; ListViews / GetView / CreateView / UpdateView
  / SoftDeleteView / DefaultViewFor.
- CreateView and UpdateView clear the prior default for a page in the
  same transaction when IsDefaultFor is set — defends against the
  partial unique index outside the SECURITY DEFINER path.
- Validation mirrors the DB check constraints so handlers can surface
  friendlier errors before round-tripping.

Handlers (web/views.go) + routes (web/server.go):
- GET  /views            list + create form (templates/views.tmpl).
- POST /views            create (filter_query form field is parsed into
                         canonical filter_json shape — design.md §2).
- GET  /views/<id>       redirect to the target page + ?view=<id>.
- POST /views/<id>       update.
- POST /views/<id>/delete soft delete.

Resolution path:
- handleTree now calls applySavedView when ?view=<uuid> is present;
  fields the saved filter_json + view_type back into the TreeFilter and
  the view-type slot. view_type then revalidates against the route
  catalog so a saved kanban-view URL on / lands on list with kanban
  shown locked until slice C ships it. Failures fall back gracefully
  (log + URL-derived filter), no 500.

UI:
- Sidebar gains a Views entry (4-square icon) next to Admin in
  layout.tmpl.
- /views renders a flat table + inline create form. The form accepts a
  URL-query filter string (e.g. `tag=work&mgmt=mai`) which is canonised
  into filter_json on save.

Tests:
- TestViewsCRUDRoundTrip — full create / list / open-redirect / soft-
  delete cycle via HTTP, plus filter_json shape assertion.
- TestSavedViewAppliedOnQueryParam — seed a card view scoped to dev,
  hit /?view=<id>, assert the page renders card grid + scoped chip-on.

Out of scope for slice D (per design.md §7):
- HTMX modal save UI from any page (the inline-create-on-/views/ form
  works; a modal lands in a polish pass).
- MCP read tools for views (deferred to a follow-up — m manages views
  via the UI).
2026-05-26 13:42:51 +02:00
mAi
0bea9c1ba4 feat(phase 4f): per-item timeline_exclude flag (hide noise from /timeline)
m's stated use case: home VTODOs (shopping list) shouldn't pollute the
chronological /timeline by default, but they should stay visible on the
home detail page itself. This adds an item-level switch with four kinds
and a URL override to peek at everything when wanted.

## Schema (migration 0015)
- timeline_exclude text[] NOT NULL DEFAULT '{}'
- items_timeline_exclude_idx GIN
- items_unified view rebuilt to surface the new column
- Behaviour-neutral: empty array = unchanged from today. m flips the
  toggle himself via /admin/bulk or the detail-page form.

## Aggregation
- web/timeline.go: pre-compute the per-kind keep-list via keepFor(kind)
  before fanning out — items with the kind in their exclude array are
  dropped entirely (no CalDAV call wasted on excluded sources). Doc and
  creation rows check the per-item flag inline. `?include_excluded=1`
  (URL) and `include_excluded:true` (MCP arg) override the filter.
- store.Item.ExcludesTimelineKind(kind) helper accepts either singular
  ("todo") or plural ("todos") to bridge the kind-constant / persisted-
  value naming choice — see comment for the why.

## UI
- /i/{path} grows a "Timeline behaviour" collapsible section with four
  checkboxes (todos / events / docs / creation) and helper text. Open by
  default when any kind is excluded, so m can see at a glance what's
  hidden for this item.
- /admin/bulk gains a "timeline todos" select with "Exclude from timeline"
  and "Re-include on timeline" — the other three kinds stay editable
  per-item only per the task brief (most common use case is just todos).

## MCP
- update_item accepts timeline_exclude as a partial-update field with an
  enum-restricted whitelist; unknown values dropped silently.
- itemView always emits timeline_exclude (defaults to []) so consumers
  can render the toggle state without a second round-trip.

## Tests
- Migration + GIN index landed
- Item with timeline_exclude=['todos'] hides the VTODO from /timeline
- ?include_excluded=1 brings it back
- Bulk action toggles the array idempotently in both directions
- Detail page renders all 4 checkbox affordances

## docs/design.md
§12 gains a "Per-item exclusion" subsection documenting semantics, the
URL override, the bulk action, and the "detail page still shows everything"
invariant.

## Out of scope (per task brief)
- Per-tag exclusion (per-item is clearer)
- Per-day exclusion (overkill)
- Dashboard exclusion (m only flagged timeline; dashboard's "today" view
  should still show shopping today if it's due today)
- Auto-seeding home with timeline_exclude=['todos'] (m runs once himself
  via /admin/bulk after the deploy — schema change stays behaviour-neutral)
2026-05-17 19:28:49 +02:00
mAi
f6cf050c3f feat(phase 4d): public-listing fields so projax becomes the portfolio source of truth
Adds five additive columns on projax.items and propagates them through
every read/write path. flexsiebels.de (and any future portfolio renderer)
can now pull the public set via the MCP `list_items(public=true)` filter
and stop hard-coding project lists.

## Schema (migration 0014)
- public               boolean       default false (partial index when true)
- public_description   text          default ''
- public_live_url      text          default ''
- public_source_url    text          default ''
- public_screenshots   text[]        default '{}'
- items_unified view rebuilt to include the five new columns
- items_public_idx     PARTIAL INDEX WHERE public = true (5% of rows)

## Store
- Item struct + scan/scanItems extended (5 cols)
- UpdateInput accepts the new fields with full-replace semantics
- new SetPublic(ids, bool) for bulk write
- SearchFilters gains Public *bool — nil = no filter

## MCP
- list_items: new `public` boolean filter (input schema + handler)
- update_item: 5 new partial-update fields (nil pointer = leave alone)
- itemView always emits the 5 fields (even when public=false) so consumers
  can preview "what would publish" without a second round-trip
- 2 new integration tests against the DB

## Web
- /i/{path} grows a "Public listing" fieldset: toggle + textarea + 2 URL
  inputs + screenshot list editor with add/remove rows + inline JS for
  the editor. Values persist when public is off so toggling never
  destroys typed-in content.
- /admin/bulk action bar gains "Make public" / "Make private" via a new
  select; SQL update is a single statement per action.
- /?public=1 and /?public=0 chip parameters narrow the tree page.
  Active() + QueryString() + TogglePublic() round-trip the state.
- parseScreenshotList helper trims + drops empties + preserves order
- 5 integration tests: migration landed, form round-trip, bulk action
  round-trip, detail-page affordances, tree-filter narrowing

## docs/design.md §15
Documents the schema, MCP contract, UI surfaces, flexsiebels consumption
pattern, and what's NOT in scope (flexsiebels-side render, asset hosting,
approval workflows).

## Out of scope (per task brief)
- Flexsiebels rendering — separate task in m/flexsiebels.de after this ships
- Asset hosting (projax stores URLs, never bytes — same PER discipline)
- Multi-stage publish workflow (boolean is enough)
2026-05-17 19:11:26 +02:00
mAi
0e490bb600 feat(phase 3d auto-tag): backfill area tags, bulk-edit UI, soft-delete cleanup
- migration 0012: one-shot populate empty tags from each item's area-roots
  (so chips on /?tag=work etc. actually filter the 40+ mai-backfilled rows)
- migration 0013: cleanup 12 orphan item_links + BEFORE-UPDATE trigger that
  cascades soft-delete to item_links going forward — closes the data drift
  that made TestItemsUnifiedSurfacesMaiPointer fail since 3c
- /admin/bulk page: flat filter+checkbox list with one-tx Apply for add/
  remove tag, set management, set status. Per-row inline chip add/remove
  via /admin/bulk/chip. Reuses tree_filter URL params 1:1.
- design.md §3.2 + §4.1 updated; tag+management section notes 0012
- bulk + tag-backfill + soft-delete-cascade tests cover the new surface
2026-05-15 18:49:58 +02:00
mAi
e055e4607e feat(phase 3c per-events): event_date on item_links, Documents UI, PER URL resolver, MCP date-aware add_link
migration 0011_item_links_event_date.sql: ADD event_date date + partial
index (idempotent). Day granularity by design per the PER spec; the
column lands NULL on every existing row, no backfill.

store:
- ItemLink gains an EventDate *time.Time (every read path scans it).
- AddLinkDated(ctx, item, refType, refID, rel, note, date, metadata)
  upserts with COALESCE(new, old) for note + event_date so partial
  callers don't clobber prior state.
- DatedLinks(item) returns event_date IS NOT NULL ordered DESC.

web:
- per.go: parsePER strips a trailing .YYMMDD (rejects invalid dates like
  Feb 30); collisionTag yields a/b/.../z/aa/ab/...; computePERs walks
  DatedLinks output and assigns render-time collision tags inside each
  date group. Tags are never stored.
- handleDetail: 404 retry with PER stripped — /i/mfin.house1.260515
  resolves to the house1 item with HighlightDate=2026-05-15.
- documents_section.tmpl: add-form (ref_type/date/ref_id/note),
  date-sorted rows with computed PER, ref-type badge, remove × with
  anti-forgery item-id check, highlight row when HighlightDate matches.
- POST /i/{path}/links/add and /links/remove handlers; HTMX swap on the
  fragment, redirect for non-HTMX callers.

mcp:
- add_link accepts event_date: "YYYY-MM-DD" (parsed strict, hands back
  fmt.Errorf on bad form). linkView.event_date surfaces it on responses.
- Existing add_link callers without event_date keep working unchanged.

docs:
- docs/standards/per.md gains an Implementation section pointing at
  item_links.event_date + ref_types + render-time collision policy.
- docs/design.md adds a Documents/dated artifacts section with the
  schema delta, conflict policy, and URL routing rules.

tests:
- per_test.go: parsePER (valid/invalid dates, non-numeric, wrong
  length); collisionTag (1..53); computePERs (bare-then-.a, skips
  undated, multi-date grouping).
2026-05-15 18:35:21 +02:00
mAi
1fcf6356f8 fix(db): track applied migrations in projax.schema_migrations
The destructive deltas in 0010 (drop parent_id, drop path) broke idempotent
re-runs of 0001/0002 on every boot — those expect the legacy columns to
exist. Stop relying on every migration being safe-to-rerun and track
applied versions in projax.schema_migrations instead.

ApplyMigrations now:
- ensures projax.schema_migrations exists,
- reads the set of applied filenames,
- applies only the missing ones, in lexicographic order,
- records each apply on success.

Existing fleet was bootstrapped via MCP-as-supabase-admin (where each
0001..0010 was applied directly to the live DB before this commit
existed). To match the new runner's expectations, the tracker was
seeded with all ten names before this push (matching DB state).

migrate_test.go: TestMigrationsAreIdempotent now asserts on parent_ids
shape instead of the old parent_id column.

This is the production-down fix — the previous deploy was crashlooping
on 'apply 0001_init.sql: ERROR: column "path" does not exist'.
2026-05-15 16:36:43 +02:00
mAi
41c1eaadaa feat(phase 1.5): tags + management + DAG + mai.projects sync
Big task. Five migrations, full store + web rewrite, and a model upgrade
that turns the parent_id tree into a parent_ids[] DAG.

Schema (db/migrations)
- 0006_tags_management_unify: adds tags + management text[] (GIN-indexed),
  collapses the area/project distinction (kind keeps the slot but 'area'
  is no longer a special value), drops the structural rules from the
  path trigger so root projects + non-root projects are both legal.
- 0007_backfill_mai_projects: one-shot, idempotent — for every row in
  mai.projects without a 'mai-project' item_link, create a projax.items
  row under a heuristic-chosen area (mhealth→health, msports/manjin→
  sports, kanzlai/hlckm/work/mworkrepo/paliad or HL/* repo→work,
  mhome→home, default→dev), insert the item_link, and tag the row
  management=['mai']. Also flips management='mai' on any already-linked
  pre-Phase-1.5 promotions.
- 0008_mai_projects_sync: bidirectional triggers. sync_to_mai runs as
  projax_admin and writes mai.projects directly (after the operator-run
  grant + RLS policy widening — documented in the migration header).
  sync_from_mai is SECURITY DEFINER so writes by the mai role fan out
  into projax.items. pg_trigger_depth() + projax.in_sync GUC keep the
  cycle suppressed. Slug stays the join key for new rows; the
  item_link pointer survives renames.
- 0009_items_unified_simplify: view collapses to a thin projection over
  projax.items now that mai.projects is a derived projection.
- 0010_multi_parent: parent_id → parent_ids uuid[], path → paths text[].
  compute_item_paths walks via parents' precomputed paths (no recursive
  CTE in the hot path; cycle detection uses one). New triggers:
  items_check_slug_collision (multi-parent uniqueness),
  items_after_delete (manual cascade since arrays don't carry FK).
  Trigger refresh_item_paths_recursive does parent-first DFS over
  descendants, guarded by projax.refreshing_paths GUC.

Go store + handlers
- Item gains ParentIDs []string + Paths []string. PrimaryPath /
  OtherPaths helpers feed the detail breadcrumb. Source always
  'projax' now; SourceRefDeref still surfaces the mai-id pointer.
- Update / Reparent / Create take ParentIDs []string. AddParent helper
  for the multi-parent UI's "also list under" action.
- GetByPath uses '$1 = any(paths)' so /i/work.paliad and /i/dev.paliad
  resolve to the same row.
- buildForest renders a multi-parent item under each of its parents
  (duplicated nodes in distinct branches). Tag-filter prune is
  branch-preserving.

Templates
- detail.tmpl: multi-select parents, tags + management chip inputs,
  "Also at: …" breadcrumb for multi-parent items.
- new.tmpl: same multi-select + chip inputs.
- tree.tmpl: tag-filter chip bar, "×N" badge on multi-parent rows,
  management chips visible on every row.
- classify.tmpl: re-parent workflow (no more promote-to-projax — the
  bidirectional sync removed the dichotomy).

Tests (DB + HTTP, all skip without env)
- TestMultiParentResolvesBothPaths   inserts an item with two parents,
                                     asserts both inherited paths.
- TestSlugCollisionUnderCommonParent  refuses a sibling clash.
- TestMultiParentBothPathsRouteToSameRow  HTTP-level: /i/dev.X and
                                          /i/work.X both 200, same row.
- TestReparentRoundTrip rewritten for parent_ids[] semantics.
- TestPathTriggerNestAndRename / Reparent rewritten to query paths[].

Docs (docs/design.md)
- §2 rewritten: items in a DAG, no area/project distinction.
- §3 schema: parent_ids + paths + tags + management + indices.
- §3.1 path-trigger overhaul incl. cycle detection via recursive CTE
  and slug-collision-under-common-parent guard.
- §3.2 view simplified.
- §3.4 NEW: mai.projects bidirectional sync, including the manual
  prereq.
- §4.1 + §4.2: classify becomes re-parent, tags+management UI section.

mai head start / mai hire / mai status / mai instruct keep working
because mai.projects retains its FK-target shape; the projax sync just
mirrors the row in lock-step.
2026-05-15 16:33:52 +02:00
mAi
840c1760c9 feat(auth): federate with mgmt.msbls.de via Supabase cookies
projax was deployed publicly through Dokploy/Traefik with a Let's
Encrypt cert; the earlier "Tailscale-only" claim was never true. Gate
every request at the application layer using the same Supabase JWT
cookie pair that mgmt.msbls.de issues, so projax inherits SSO without
running its own login.

Middleware (web/auth.go):
- GET <SUPABASE_URL>/auth/v1/user with the access_token cookie or a
  Bearer header. On 2xx → pass through.
- On expiry, swap the refresh_token via /auth/v1/token?grant_type=
  refresh_token and rotate both cookies (Domain=msbls.de, HttpOnly,
  Secure, SameSite=Lax, Path=/, Max-Age=1y). Cookie attributes match
  mgmt/auth.ts verbatim — refreshed sessions stay drop-in compatible
  with the rest of the .msbls.de fleet.
- Anything still invalid → 302 to <PROJAX_LOGIN_URL>?redirectTo=
  <original-absolute-url>. mgmt's safeRedirect() rejects absolute URLs
  and falls back to /, so after login the user lands on mgmt; manual
  click back to projax then succeeds with the fresh cookie. UX is
  rough but functional; broadening mgmt's safeRedirect is parked for a
  separate PR.
- /healthz remains ungated so Dokploy/Traefik probes don't hit the
  redirect.

main.go: enable the middleware only when SUPABASE_URL is set; require
SUPABASE_ANON_KEY when it is (refuse to start otherwise). New env
overrides: PROJAX_LOGIN_URL (default https://mgmt.msbls.de/login),
PROJAX_COOKIE_DOMAIN (default msbls.de). Local dev with no env stays
fully anonymous.

Tests (7 cases, no DB needed): stub Supabase via httptest covers
healthz-open, anonymous-redirect, bad-cookie-redirect, good-cookie
pass-through, Bearer-pass-through, stale-but-refreshable rotation
(verifies cookie Domain/HttpOnly/Secure/SameSite), final fail
redirect.

DB-backed integration tests now honour PROJAX_SKIP_MIGRATE=1 so they
don't deadlock against the live container's auto-migrate during a
deploy window.

README + dokploy.yaml: kill the Tailscale-only claim, document the
federated-auth trust model and the new SUPABASE_* env contract.
2026-05-15 14:58:43 +02:00
mAi
092a56cf24 feat(db): pivot to dedicated projax_admin role
Mirrors the mbrian_admin pattern: the binary connects as a role bounded
to the projax schema, so even a compromised projax process cannot read
mai.workers, otto.*, vault.*, etc.

- 0001: switch grants block from postgres → projax_admin (conditional
  on the role existing — bootstrap still works as superuser before the
  role is created). Wrap `create schema` in a guard so the migration
  is idempotent when re-run as a non-superuser app role that lacks
  database-level CREATE.
- 0005_reown_to_projax_admin.sql: enumerate every projax-namespaced
  object via pg_namespace + pg_class / pg_proc and ALTER OWNER to
  projax_admin. Explicitly scoped — no global REASSIGN OWNED that
  would yank ownership from other projects sharing the postgres role.
  Strips residual postgres grants. No-ops with a NOTICE when the role
  is missing.
- README: new "Manual prerequisite" deploy section. Documents the
  CREATE ROLE statement, the cross-schema USAGE + SELECT grants, AND
  the RLS policy `projax_read ON mai.projects` that's required because
  mai.projects has row-level security with policies scoped to `mai`
  and `anon` only. Without the policy, items_unified silently returns
  zero mai-source rows.
- deploy/dokploy.yaml: DSN comment now reflects projax_admin and
  points at the README prereq.

Verified locally against msupabase with a throwaway projax_admin role:
- 13/13 tests green
- mai.workers SELECT → permission denied
- mai.sessions SELECT → permission denied
- mai.projects SELECT → 59 rows (RLS policy in effect)
- projax.items_unified SELECT → 66 rows (7 projax + 59 mai)
2026-05-15 13:32:56 +02:00
mAi
c0466ade36 feat(db): items_unified adapter view + promotion hiding
projax.items_unified joins projax.items (deleted_at IS NULL) with
mai.projects so a single query feeds the tree UI. mai.projects.id is a
text key, so a deterministic placeholder UUID is derived from md5(p.id);
projax-native rows keep their gen_random_uuid().

When a projax item is created with an item_links row pointing back to a
mai.projects id (ref_type='mai-project'), the corresponding mai.projects
row drops out of the view — that's how the "Promote to projax" flow
makes the duplicate disappear without ever touching mai.projects.

Test coverage:
- both sources appear in the view
- promotion link hides the mai source row and surfaces the projax row
2026-05-15 13:17:51 +02:00
mAi
b8d3418876 feat(db): projax schema, path trigger, seed areas
- 0001_init.sql: projax.items + projax.item_links tables with indices,
  partial-unique root slug, updated_at trigger, schema grants to the
  application role.
- 0002_path_trigger.sql: BEFORE-write trigger maintains items.path via
  recursive parent walk; rejects cycles and structural-rule violations
  (areas at root, projects not at root). AFTER trigger rewrites
  descendant paths on slug rename or re-parent.
- 0003_seed_areas.sql: dev, sports, home, work, health, finances, social.
- db/migrate.go: embed.FS-backed sequential runner.
- db/migrate_test.go: integration suite covering idempotency, nest,
  rename propagation, re-parent propagation, cycle rejection, and
  structural rules. Skips when no DB env var is set.

Also ignores .m/events.log and .m/locks (per-worker scratch).
2026-05-15 13:16:24 +02:00