generateObject() in the AI SDK falls back to a tool-call mode when the
provider doesn't advertise structured-output support — and tool calling
through Ollama isn't reliable enough that the schema-validation step
passes. The response was failing with 'No object generated: response
did not match schema' even though the underlying mana-llm + Ollama
roundtrip works correctly when called with response_format directly
(verified via curl).
Set supportsStructuredOutputs:true on the createOpenAICompatible
factory so the AI SDK uses response_format json_schema mode. mana-llm
already routes that to Ollama's native format field thanks to the
companion fix in services/mana-llm/src/providers/ollama.py — verified
end-to-end with the MealAnalysisSchema and Gemma 3 4B.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
mana-llm on the live Mac Mini does not have GOOGLE_API_KEY configured —
only the Ollama provider is registered. The previous default
'google/gemini-2.0-flash' would error with 'Provider google not
available' on every photo analysis.
Switch to ollama/gemma3:4b which is locally available via the
gpu-proxy bridge to the Windows GPU box (192.168.178.11). Gemma 3 is
multimodal and verified end-to-end with the new mana-llm structured-
output passthrough — see the 5520f1385 fix landing the response_format
plumbing on the Pydantic side and the Ollama provider's native format
field translation.
VISION_MODEL env var still wins, so prod can flip to
google/gemini-2.0-flash later by adding GOOGLE_API_KEY to mana-llm's
docker-compose env block.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Found while smoke-testing the AI SDK refactor: both nutriphi and planta
were calling `${MANA_LLM_URL}/api/v1/chat/completions` and passing
`gemini-2.0-flash` as the model name. Both wrong:
1. mana-llm exposes routes under /v1/, not /api/v1/. The original
pre-refactor code had the same bug — it predates this commit and
was apparently never noticed because the photo workflow was never
wired into the unified app's UI until last week. /api/v1 returned
404 against the live mana-llm container; now we hit /v1.
2. mana-llm's router parses model strings as `provider/model`
(services/mana-llm/src/providers/router.py:_parse_model). Without
a prefix, `gemini-2.0-flash` was being routed as
`ollama/gemini-2.0-flash` and only worked via the auto-fallback
to Google when ollama failed. Be explicit: `google/gemini-2.0-flash`
hits the Google provider directly and skips the failed-ollama
round-trip.
VISION_MODEL env var still wins over the default, so prod overrides
remain possible.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds AI_SCHEMA_VERSION + AiResponseEnvelope<T> in @mana/shared-types so
every AI structured-output endpoint speaks { schemaVersion, data }.
Backend wraps via envelope() in each module routes.ts; frontend api.ts
unwraps via unwrapEnvelope<T>() which throws AiSchemaVersionMismatchError
on drift — actionable network-panel error instead of cascading
'field is undefined' bugs further down the stack.
Also adds providerOptions.anthropic.cacheControl on the system message
in nutriphi + planta routes via SYSTEM_CACHE_HINT. NO-OP today (Gemini
backend, ~50-token prompts under the 1024-token cache minimum) but
lights up automatically when mana-llm routes to Claude or prompts grow
past the threshold. ~5 lines per route, no risk.
System messages migrated from system: shorthand to a full messages[]
entry — the only way to attach providerOptions per-message in the AI SDK.
13 new tests in nutriphi/ai-schemas.test.ts cover the version constant,
the mismatch error shape, and Zod accept/reject for both schemas. Total
nutriphi + planta suite: 62/62.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replaces hand-rolled fetch + JSON.parse + cast-to-any with generateObject
from the AI SDK. The model is constrained to the shared Zod schemas in
@mana/shared-types, so the response is validated at the boundary instead
of trusting Gemini to emit the right shape.
Routes refactored:
- nutriphi/analysis/photo (image_url → multimodal `image:` content)
- nutriphi/analysis/text (free-text meal description)
- planta/analysis/identify (plant photo identification)
Why this is materially better than the old code:
- Runtime validation: if Gemini drifts, the AI SDK throws before the
response leaves the route. Frontend never sees malformed payloads.
- Provider-portable: createOpenAICompatible({ baseURL: MANA_LLM_URL })
keeps mana-llm as the central routing/auth/observability point. The
AI SDK speaks the OpenAI dialect to mana-llm. If we ever swap the
backend (e.g. claude-sonnet-4-6 for plant ID), it's a one-line model
name change.
- System prompts moved from a multi-line example-laden string to a
short instruction. The schema itself (with .describe() field hints)
now carries the structural contract that the JSON-by-example
paragraph used to encode. Token cost goes down, accuracy goes up.
- Drops manual fetch error handling (status checks, JSON.parse, cast)
in favour of try/catch around generateObject. Errors are typed.
mana-llm itself is unchanged — it's still the OpenAI-compatible proxy
in front of Gemini Vision. The AI SDK just gives us a typed client and
a schema-aware decoder on top of it.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Mirror the planta two-step pattern: a FormData upload endpoint that
returns mediaId/publicUrl from mana-media, and a separate Gemini Vision
analysis endpoint that takes a photoUrl. Drops the base64 inline path
and the half-finished parallel-upload kludge in the old combined route.
Why: the old endpoint was wired neither in the frontend nor used
elsewhere, and the combined base64+upload+analyze design made it
impossible to show the photo to the user before AI ran or to re-analyze
without re-uploading.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Pre-launch theme system audit found multiple parallel layers in themes.css
(--theme-X full hsl strings, --X partial shadcn aliases, --color-X populated
by runtime store with raw channels) plus dead-code companion files. The
inconsistency caused light-mode regressions when scoped-CSS consumers
wrote `var(--color-X)` standalone — the variable holds raw HSL channels
which is invalid as a color value, browser fell back to inherited (white).
Rewrite to one consistent layer:
- Source of truth: --color-X defined as raw HSL channels (e.g.
`0 0% 17%`) in :root, .dark, and all variant [data-theme="..."]
blocks. Matches the format the runtime store
(@mana/shared-theme/src/utils.ts) writes, eliminating the
static-fallback-vs-runtime mismatch and the corresponding flash
of unstyled content on hydration.
- @theme inline uses self-reference + Tailwind v4 <alpha-value>
placeholder so utility classes generate correctly AND opacity
modifiers work: `text-foreground/50` → `hsl(var(--color-foreground) / 0.5)`.
- @layer components (.btn-primary, .card, .badge, etc.) wraps
var(--color-X) refs with hsl() — they were broken in light mode
too for the same reason.
Convention going forward (also documented in the file header):
1. Markup: use Tailwind utility classes (text-foreground, bg-card, …)
2. Scoped CSS: hsl(var(--color-X)) — always wrap with hsl()
3. NEVER raw var(--color-X) in CSS — that's the bug pattern
Net file: 692 → 580 LOC. Single source layer, no indirection.
Also delete dead companion files (zero imports anywhere):
- tailwind-v4.css (had broken self-reference, never imported)
- theme-variables.css (legacy hex-based palette)
- components.css (legacy component utilities)
- index.js / preset.js / colors.js (Tailwind v3 preset format,
irrelevant under Tailwind v4)
package.json exports map shrinks accordingly to just `./themes.css`.
Consumers using `hsl(var(--color-X))` (~379 files across mana-web,
manavoxel-web, arcade-web) keep working unchanged — the public API
name `--color-X` is preserved. Only the broken pattern `var(--color-X)`
(~61 files) needs a follow-up sweep, handled in a separate commit.
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>
Complete consolidation of all 15 app servers into one Hono/Bun process.
Modules added: chat, context, picture, storage, todo, planta, nutriphi,
guides, moodlit, news, traces, presi
Total: 15 modules, one server, one port (3050), ~2400 LOC.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
