#5 — SYSTEM_ARTICLES_IMPORT_WORKER hoisted into @mana/shared-ai
The worker built its actor inline, bypassing the SystemSource union
that's the blessed list for system-write principals. Now uses
makeSystemActor(SYSTEM_ARTICLES_IMPORT_WORKER) like every other
server-side system writer (mission-runner, projection, …).
#7 — sync-db helper hoisted out of mcp/ into lib/
Implementation moved to apps/api/src/lib/sync-db.ts; mcp/sync-db.ts
is a re-export shim so existing MCP imports keep working. Articles
bulk-import + future modules import from lib/ directly — no more
"articles depending on mcp" layering smell.
#11 — Prometheus metrics for the worker
New counters + histogram in lib/metrics.ts under
mana_api_articles_import_*:
- ticks_total{result=processed|skipped|error}
- items_total{result=extracted|error|consent_wall|cancelled}
- extract_duration_seconds (histogram, 0.25–30s buckets)
- jobs_completed_total{result=done}
- pickup_gc_rows_total
Worker tick + extractor instrumented at the right transition points.
Steady-state pickup_gc_rows_total > 0 over time signals a stuck
consumer somewhere — useful operator alert.
Plan: docs/plans/articles-bulk-import.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Final milestone of docs/plans/llm-fallback-aliases.md. Every backend
caller now requests models via the `mana/<class>` alias system instead
of hardcoded `ollama/...` strings. mana-llm resolves aliases through
`services/mana-llm/aliases.yaml` with health-aware fallback (M3) and
emits resolved-model + fallback metrics (M4).
SSOT moved to `packages/shared-ai/src/llm-aliases.ts` so apps/api,
apps/mana/apps/web, and services/mana-ai all import the same
`MANA_LLM` constant via the existing `@mana/shared-ai` workspace
dependency. Three additional sites (memoro-server, mana-events,
mana-research) inline the alias string with a SSOT comment because
they don't pull @mana/shared-ai today.
Migrated 14 sites across 10 files:
- apps/api: writing(LONG_FORM), comic(STRUCTURED), context(FAST_TEXT),
food(VISION), plants(VISION), research orchestrator (3 tiers
collapsed to STRUCTURED+FAST_TEXT/LONG_FORM)
- apps/mana/apps/web: voice/parse-task + parse-habit (STRUCTURED)
- services/mana-ai: planner llm-client + tick.ts (REASONING)
- services/mana-events: website-extractor (STRUCTURED, inlined)
- services/mana-research: mana-llm client (FAST_TEXT, inlined)
- apps/memoro/apps/server: ai.ts (FAST_TEXT, inlined)
Legacy env-vars removed: WRITING_MODEL, COMIC_STORYBOARD_MODEL,
VISION_MODEL, MANA_LLM_DEFAULT_MODEL. The chain in aliases.yaml is
now the single tuning surface; SIGHUP reloads it without redeploys.
New `scripts/validate-llm-strings.mjs` regex-scans 2538 files for
hardcoded `<provider>/<model>` strings and fails the build if any
land outside the SSOT or the explicitly-allowed paths (image-gen
modules, model-inspector code, this validator itself, the registry).
Wired into `validate:all` next to the i18n + theme validators.
Verified: `pnpm validate:llm-strings` clean, `pnpm --filter @mana/api
type-check` clean, `pnpm --filter @mana/ai-service type-check`
clean. Web type-check has 2 pre-existing errors in
SettingsSidebar.svelte (i18n MessageFormatter type drift, last
touched in 988c17a67 — unrelated to this work).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Server:
- New llmText() helper in apps/api/src/lib/llm.ts for plain-text
(non-streaming) completions with token-usage reporting.
- POST /api/v1/writing/generations (Hono + requireTier('beta'))
accepts system+user prompts, forwards to mana-llm (default model
ollama/gemma3:4b), returns raw output + model + tokenUsage. The
endpoint is stateless — draft/version bookkeeping is entirely
client-side so the same route serves refinement calls later.
Client:
- writing/api.ts — Bearer-authed fetch client (follows the food/
news-research pattern).
- writing/utils/prompt-builder.ts — pure builder turning a briefing
(+ optional style preset / extracted principles) into a system+user
pair. Forbids preamble / sign-off / meta commentary so the output is
ready to paste into a version.
- writing/stores/generations.svelte.ts — orchestrates the full flow:
queued → running → call → new LocalDraftVersion → pointer flip →
succeeded. On failure leaves the current version untouched with the
error on the generation record. Emits WritingDraftGenerationStarted /
WritingDraftVersionCreated / WritingDraftGenerationFailed events.
UI:
- Generate button in DetailView.svelte (label flips "Generate" / "Neu
generieren" based on whether the draft already has content).
- GenerationStatus.svelte strip surfaces queued / running / failed with
model + duration badges; succeeded generations auto-disappear because
the new version is already live via the currentVersionId pointer.
M3 is synchronous and non-streaming by design. M7 adds mission-based
long-form with streaming + outline stage + reference injection. M6 will
reuse the same /generations endpoint for selection-refinement prompts.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
gpt-image-1 answered the last Try-On attempt with
invalid_image_file: Invalid image file or mode for image 2
because one of the references (face/body/garment) was in a format or
color mode OpenAI's edits endpoint rejects — typical culprits are
HEIC from iPhones, CMYK JPEG, palette-mode PNG, APNG, or JPEG with an
ICC profile gpt-image-1 doesn't honour. mana-media stores originals
verbatim so whatever the user uploaded is what we were forwarding.
Route the references through mana-media's existing on-the-fly
/transform endpoint (format=png, w/h=1024, fit=inside) which pipes
the buffer through sharp server-side. One call per ref, all run in
parallel, same latency budget as before. Output is guaranteed
- PNG / RGB (or RGBA if the source had alpha, which gpt-image-1 accepts),
- no more than 1024 px on the longest side → well under OpenAI's
4 MB/image cap,
- aspect-ratio-preserving (fit=inside) so a portrait body photo
doesn't get squished into a square.
New helper `getMediaBufferAsPng(mediaId, longestSide)` in lib/media.ts
encapsulates the transform-URL build. The Try-On path in the picture
route now uses it instead of `getMediaBuffer`; all Blob filenames
pin to `.png` since the buffer is already normalized.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
M4 of docs/plans/wardrobe-module.md — the loop closes. A user with at
least a face-ref in the active space can click "Anprobieren" on an
outfit detail page; the client composes a reference call against the
existing M3 `/generate-with-reference` endpoint, persists the result
into the Picture gallery with a `wardrobeOutfitId` back-reference,
and pins a `lastTryOn` snapshot on the outfit so its card instantly
shows the AI preview next time.
Server side — picture/routes.ts:
- verifyMediaOwnership now accepts `apps: string | readonly string[]`.
Under the hood it runs one list() per app-tag and unions the owned
set before the missing-id check. Preserves the 500-row per-app
sanity cap. Single-tag callers unchanged — it's an additive widen.
- Picture /generate-with-reference passes `['me', 'wardrobe']` so
face/body portraits (me-images) and garment photos (wardrobe) can
ride in the same referenceMediaIds array. Anything outside those
two tags still 404s — no expansion of the trust surface.
Client side — wardrobe/api/try-on.ts:
- `runOutfitTryOn({ outfit, garments, faceRefMediaId, bodyRefMediaId?, ... })`
composes the ref list (face → body → up to 6 garments, respecting
the 8-slot server cap), picks portrait 1024x1536 by default (or
1024x1024 in accessory-only mode), and POSTs with
`model='openai/gpt-image-2'`, `quality='medium'`, `n=1`. One render
per click; multi-variant is a future Generator-style extension.
- Default prompts are composed in DE from the outfit meta (name +
occasion); callers can override via `prompt`. Accessory-only mode
uses a tighter studio-portrait phrasing since the fullbody ref is
dropped there.
- `isAccessoryOnlyOutfit()` helper — iff every garment is in
FACE_ONLY_CATEGORIES, skip body-ref and render square. Covers the
Brille-Try-On headline use case.
- On success: inserts a `picture.images` row with generationMode=
'reference', referenceImageIds, and wardrobeOutfitId set; then
calls wardrobeOutfitsStore.setLastTryOn() with imageId + imageUrl
so OutfitCard + DetailOutfitView immediately flip to the AI cover.
TryOnButton — wardrobe/components/TryOnButton.svelte:
- Three states: ready (click to render), missing-references (shows
UserCircle + link to /profile/me-images, with the right hint for
accessory-only vs. fullbody), loading (spinner).
- Credit estimate on the button (10c medium quality).
- Hints: accessory-only, too-many-garments (>6, over server cap),
and non-personal-space disclosure — the family-space case gets its
own sentence since "Try-On rendert dich, nicht dein Kind" is
non-obvious.
- Reads face-ref/body-ref via useImageByPrimary (space-scoped after
the v40 meImages migration — brand/club/family spaces need their
own references uploaded).
UI wiring:
- DetailOutfitView replaces the M3 stub button with <TryOnButton/>.
The existing "Try-On Verlauf"-Strip already reads
`useOutfitTryOns(outfit.id)` which filters `picture.images` by
wardrobeOutfitId — it lights up automatically on first render.
Not in M4 (punted to follow-ups):
- Solo-garment try-on on DetailGarmentView ("nur diese Brille auf
mein Gesicht"). Plan called it out as optional; the outfit flow
already covers it when the outfit contains only that one garment.
- Multi-variant rendering (n=2/4). Usable "show me 3 looks" needs a
picker UI on top, not just a param bump.
- Quality + prompt override in the button. A power-user panel can
come later; default medium + auto-prompt keeps M4's click-to-try-on
one-tap.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
M1 of docs/plans/wardrobe-module.md — pure data layer + backend plumbing,
zero UI (that's M2). A user can now hold a digital wardrobe per space:
brand merch, club Trikots, family Kleiderschrank, team Kostüme, practice
Dresscode, and personal closet all live as separate pools under the same
Dexie tables, space-scoped like tags/scenes/agents after Phase 2c.
Data model — two tables, no join:
- wardrobeGarments (Dexie v41): single clothing items / accessories.
Indexed on `category` + `createdAt` + `isArchived`. Encrypted:
name/brand/color/size/material/tags/notes. Plaintext: category,
mediaIds, counters, timestamps — all indexed or structural.
`mediaIds[0]` is the primary photo used for try-on; additional
ids are alternate views (back, detail) for M7.
- wardrobeOutfits (Dexie v41): named compositions referencing
garment ids. Encrypted: name/description/tags. Plaintext:
garmentIds (FK array), occasion (closed enum — useful for
undecrypted filtering), season, booleans, lastTryOn snapshot.
- picture.images gains `wardrobeOutfitId?: string | null` as a
plaintext back-reference. Try-on results land in the Picture
gallery like any other generation; the outfit detail view
queries them via this id rather than maintaining a third table.
Space scope:
- `wardrobe` added to all five explicit allowlists in shared-types/
spaces.ts (personal is wildcard, no edit needed). Each space type
gets a one-line comment explaining the real-world use case.
- App registry: `wardrobe` entry in shared-branding/mana-apps.ts
with a rose→fuchsia gradient icon (T-shirt on hanger silhouette),
color #e11d48, tier 'beta', status 'beta'.
- Module registry: wardrobeModuleConfig imported + appended to
MODULE_CONFIGS so SYNC_APP_MAP picks it up automatically.
Backend:
- MAX_REFERENCE_IMAGES bumped 4 → 8 in picture/generate-with-
reference (plus the client-side default in ReferenceImagePicker).
Justified with a comment: face + body + top + bottom + shoes +
outerwear + 2 accessories = 8. Cost doesn't scale with ref count
(OpenAI bills per output), so the bump is a pure capability
expansion with no credit-side risk.
- New POST /api/v1/wardrobe/garments/upload wraps uploadImageToMedia
with app='wardrobe'. Registered under /api/v1/wardrobe in index.ts.
Pattern 1:1 with the profile/me-images/upload endpoint; tier-gating
falls out of wardrobe NOT being in RESOURCE_MODULES (tier='guest'
works — consistent with picture's plain CRUD).
Stores emit domain events (WardrobeGarmentAdded, WardrobeOutfitCreated,
WardrobeOutfitTryOn, etc.) so later mana-ai missions can observe
activity without polling.
No UI in this commit. M2 (Garments-Grundlayer) wires the route + grid
+ upload-zone; M3 the Outfit composer; M4 the Try-On integration.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes the M3 loop from docs/plans/mana-mcp-and-personas.md. The
runner now picks up due personas, drives them through Claude + MCP
for one simulated turn, collects actions + ratings, and persists
them through service-key internal endpoints in mana-auth.
Internal endpoints (mana-auth, service-key-gated)
- GET /api/v1/internal/personas/due
Returns personas whose tickCadence + lastActiveAt say they're
due. Rules: hourly > 1h, daily > 24h, weekdays > 24h mon-fri.
NULLS FIRST so never-run personas go ahead of stale ones.
- POST /api/v1/internal/personas/:id/actions
Batch ≤ 500. Row ids are deterministic
(`${tickId}-${i}-${toolName}`) + ON CONFLICT DO NOTHING so the
runner can retry a tick without doubling audit rows. Also
bumps personas.last_active_at so the next /due call sees it.
- POST /api/v1/internal/personas/:id/feedback
Batch ≤ 100. Row id is `${tickId}-${module}` — natural key is
one rating per module per tick.
Runner tick pipeline (services/mana-persona-runner/src/runner/)
- claude-session.ts
Two phases per tick. runMainTurn feeds the persona's system
prompt + a German "simulate a day" user prompt to Claude Agent
SDK's query(), with mana-mcp wired in as a streamable-HTTP MCP
server. We iterate the returned AsyncGenerator and extract
tool_use blocks into ActionRows; tool_result with is_error=true
flips the most recent action. runRatingTurn is a fresh query()
with tools:[] asking Claude in character to rate each used
module 1-5 as strict JSON, which we parse with tolerance for
surrounding whitespace / fences. Unparseable output becomes a
synthetic '__parse' feedback row so operators see the failure.
- tick.ts
Orchestrator. Skips if config.paused. Fetches /due, processes
in batches of config.concurrency (Promise.allSettled so one
failure doesn't kill the batch), returns {due, ranSuccessfully,
failed[], durationMs}.
- types.ts
ActionRow and FeedbackRow shapes shared between claude-session
and the internal client; mirrors the mana-auth schema but in
narrow plain TS for the wire.
Runner bootstrap (src/index.ts)
- setInterval(config.tickIntervalMs) starts the tick loop on boot.
tickInFlight guards against overlap when Claude latency > interval.
If MANA_SERVICE_KEY or ANTHROPIC_API_KEY is missing, loop is
disabled with a warn line — /health still works, /diag/login
still works.
- New dev-only POST /diag/tick fires a single tick on demand and
returns the result, so you can verify without waiting 60 s.
- Graceful SIGTERM/SIGINT shutdown clears the interval.
Client
- clients/mana-auth-internal.ts
X-Service-Key client for the three endpoints above. Constructor
throws if serviceKey is empty — fail loud, not silent.
Boot smoke: /health + /diag/tick both return descriptive 500s when
keys are absent, 200/JSON when present. Warning lines show up on
boot for missing keys. Type-check green across mana-auth, tool-
registry, mcp, persona-runner.
End-to-end smoke recipe (docker up → db:push → seed:personas →
diag/tick → psql) documented in
services/mana-persona-runner/CLAUDE.md. That's the M3 exit gate.
M2.d (cross-space family/team memberships) still deferred.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
End-to-end testing surfaced a 404 from the synth path. mana-llm
(services/mana-llm/src/main.py) mounts the OpenAI-compatible API at
/v1/* — there's no /api prefix.
The first quick-depth e2e run only worked because the planner is
skipped on quick (it just uses the question itself), so llmJson never
fired; only llmStream did, and the streaming path also used the wrong
prefix but the test happened to land before this was caught.
The other apps/api modules (chat, guides, context, traces) all use the
wrong /api/v1/ path too — that's a separate, pre-existing bug to be
addressed in their own commits.
Verified by re-running a standard-depth research run end-to-end against
mana-llm pointed at the GPU server's ollama with gemma3:4b/12b: plan +
retrieve + extract + synth all succeed.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
End-to-end deep-research feature for the questions module: a fire-and-
forget orchestrator in apps/api that plans sub-queries with mana-llm,
retrieves sources via mana-search (with optional Readability extraction),
and streams a structured synthesis back to the web app over SSE.
Backend (apps/api/src/modules/research):
- schema.ts: pgSchema('research') with research_results + sources
- orchestrator.ts: three-phase pipeline (plan / retrieve / synthesise)
with depth-aware config (quick=1×, standard=3×, deep=6× sub-queries)
- pubsub.ts: in-process event bus, single-node, swappable for Redis
- routes.ts: POST /start (202, fire-and-forget), GET /:id/stream (SSE),
POST /start-sync (test only), GET /:id, GET /:id/sources
- Credit gating via @mana/shared-hono/credits — validate up-front,
consume best-effort on `done`. Failed runs cost nothing.
Helpers (apps/api/src/lib):
- llm.ts: llmJson() + llmStream() over mana-llm OpenAI-compat API
- search.ts: webSearch() + bulkExtract() over mana-search Go service
- responses.ts: shared errorResponse / listResponse / validationError
Schema deployment:
- drizzle.config.ts (research-scoped) + drizzle/research/0000_init.sql
hand-authored migration, deployable via psql -f or drizzle-kit push.
- drizzle-kit added as devDep with db:generate / db:push scripts.
Web client (apps/mana/apps/web/src/lib/api/research.ts):
- Typed start() / get() / listSources() / streamProgress(). The stream
uses fetch + ReadableStream (not EventSource) so we can attach the
JWT via Authorization header. Special-cases 402 for friendly toast.
- New PUBLIC_MANA_API_URL plumbing in hooks.server.ts + config.ts.
Module store (modules/questions/stores/answers.svelte.ts):
- New write-side store with createManual / startResearch / accept /
softDelete. startResearch creates an optimistic empty answer, opens
the SSE stream, debounces token deltas in 100ms batches into the
encrypted local row, and on `done` replaces the streamed text with
the parsed { summary, keyPoints, followUps } payload + citations
resolved against research.sources.id.
Citation rendering (modules/questions/components/AnswerCitations.svelte):
- Tokenises [n] markers in the answer body into clickable pills with
hover popovers showing title / host / snippet / external link.
- Lazy-loaded via a session-scoped source cache (stores/sources.svelte.ts)
that deduplicates concurrent fetches.
UI (routes/(app)/questions/[id]/+page.svelte):
- Recherche card with three-state button (start / cancel / re-run),
animated phase indicator, source counter.
- Confirmation dialog warning about web/LLM transmission since the
question itself is locally encrypted.
- Toasts for success / error / cancel via @mana/shared-ui/toast.
- Re-run flow soft-deletes prior research-driven answers but keeps
manual ones intact.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Picture, Contacts, Planta, Storage, and NutriPhi image uploads now go
through mana-media instead of directly to S3. This enables SHA-256
deduplication, automatic thumbnail generation, EXIF extraction, and
makes all images visible in the Photos gallery. Non-image files (PDFs,
audio, docs) continue to use shared-storage directly. SVG avatars in
Contacts also stay on shared-storage since Sharp can't process SVGs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
