Release: gpt-6-luna food-photo default

README.md

@@ -124,7 +124,7 @@ Useful optional environment variables:
 | `NEXT_CACHE_MAX_MEMORY_MB` | Optional Next.js in-memory cache cap in MB. Defaults to `0`, which disables the in-memory data cache entirely; set a non-zero value where cached fetches are expected to hit. |
 | `AI_GATEWAY_URL` | OpenAI-compatible chat-completions URL that powers food-photo estimates, for example `http://cliproxyapi.railway.internal:8317/v1/chat/completions`. Must be `https` unless the host is loopback or `*.railway.internal`. See `infra/cliproxyapi/`. Food-photo analysis is unavailable without it. |
 | `AI_GATEWAY_API_KEY` | Backend-only bearer key for the AI gateway. Required for food-photo estimates. |
-| `AI_GATEWAY_MODELS` | Optional comma-separated model list. Defaults to `gpt-5.6-luna(low),gpt-5.6-luna(medium)`; the effort suffix is translated by CLIProxyAPI into the reasoning-effort parameter. Set on the web service too (for the admin benchmark page) if customized. |
+| `AI_GATEWAY_MODELS` | Optional comma-separated model list. Defaults to `gpt-6-luna(low),gpt-6-luna(medium)`; the effort suffix is translated by CLIProxyAPI into the reasoning-effort parameter. Set on the web service too (for the admin benchmark page) if customized. |
 | `AI_GATEWAY_MODEL_TIMEOUT_MS` | Optional per-model attempt timeout for food-photo estimates. Defaults to `20000`, clamped to 3–30s. |
 | `ENABLE_TEST_ROUTES` | Enables controlled test-only routes. Leave off in production unless you are doing a controlled test run. |
 | `TEST_ROUTES_SECRET` | Required whenever `ENABLE_TEST_ROUTES=true`; send it in the `x-test-route-secret` header. |

apps/backend/src/legacy_api.rs

@@ -29,7 +29,7 @@ mod benchmark_fixtures;
 use benchmark_fixtures::{BENCHMARK_FIXTURES, CATEGORIES};
 
 /// The `(effort)` suffix is CLIProxyAPI's reasoning-effort selector; the retry raises it.
-const DEFAULT_FOOD_PHOTO_MODELS: &[&str] = &["gpt-5.6-luna(low)", "gpt-5.6-luna(medium)"];
+const DEFAULT_FOOD_PHOTO_MODELS: &[&str] = &["gpt-6-luna(low)", "gpt-6-luna(medium)"];
 /// Output budget shared by the reasoning tokens and the ~300-token JSON answer.
 const FOOD_PHOTO_MAX_TOKENS: u16 = 4_000;
 /// Reasoning models need headroom per attempt even at low effort.
@@ -484,7 +484,7 @@ async fn admin_benchmark(
             StatusCode::BAD_REQUEST,
             json!({
                 "ok": false,
-                "error": "Enter a model id, for example gpt-5.6-luna(low)."
+                "error": "Enter a model id, for example gpt-6-luna(low)."
             }),
         );
     };

apps/backend/src/legacy_api/tests/mod.rs

@@ -352,7 +352,7 @@ async fn food_photo_succeeds_through_the_gateway() {
         .await;
 
     let result = analyze_food_photo_url_with_limits(
-        &gateway_test_state(&endpoint, Some("gpt-5.6-luna(low)")),
+        &gateway_test_state(&endpoint, Some("gpt-6-luna(low)")),
         "data:image/png;base64,AA==",
         "",
         None,
@@ -404,17 +404,17 @@ fn gateway_models_come_from_config_with_luna_defaults() {
     assert_eq!(
         configured_food_photo_models(&config),
         vec![
-            "gpt-5.6-luna(low)".to_string(),
-            "gpt-5.6-luna(medium)".to_string()
+            "gpt-6-luna(low)".to_string(),
+            "gpt-6-luna(medium)".to_string()
         ]
     );
 
     config.ai_gateway_models =
-        Some("gpt-5.6-luna(medium), gpt-5.6-terra(low)\ngpt-5.6-luna(medium)".to_string());
+        Some("gpt-6-luna(medium), gpt-5.6-terra(low)\ngpt-6-luna(medium)".to_string());
     assert_eq!(
         configured_food_photo_models(&config),
         vec![
-            "gpt-5.6-luna(medium)".to_string(),
+            "gpt-6-luna(medium)".to_string(),
             "gpt-5.6-terra(low)".to_string()
         ]
     );
@@ -423,15 +423,15 @@ fn gateway_models_come_from_config_with_luna_defaults() {
 #[test]
 fn food_photo_request_body_omits_fields_reasoning_models_reject() {
     let body = serde_json::to_value(build_food_photo_request_body(
-        "gpt-5.6-luna(low)",
+        "gpt-6-luna(low)",
         "data:image/png;base64,AA==",
         "",
         "user-1",
         false,
     ))
     .expect("request body should serialize");
 
-    assert_eq!(body["model"], json!("gpt-5.6-luna(low)"));
+    assert_eq!(body["model"], json!("gpt-6-luna(low)"));
     assert_eq!(body["max_tokens"], json!(4000));
     // Reasoning models reject temperature pins; provider-specific routing fields would be meaningless too.
     for absent in [

apps/web/README.md

@@ -25,7 +25,7 @@ a ChatGPT/Codex subscription). Configure it on the backend environment:
 ```bash
 AI_GATEWAY_URL=http://cliproxyapi.railway.internal:8317/v1/chat/completions
 AI_GATEWAY_API_KEY=...
-AI_GATEWAY_MODELS=gpt-5.6-luna(low),gpt-5.6-luna(medium)
+AI_GATEWAY_MODELS=gpt-6-luna(low),gpt-6-luna(medium)
 AI_GATEWAY_MODEL_TIMEOUT_MS=20000
 ```
 

apps/web/components/admin-ai-benchmark-client.tsx

@@ -437,7 +437,7 @@ export function AdminAiBenchmarkClient({
             type="text"
             value={model}
             onChange={(event) => setModel(event.target.value)}
-            placeholder="gpt-5.6-luna(low)"
+            placeholder="gpt-6-luna(low)"
             className="w-full rounded-2xl border border-[var(--color-border)] bg-[var(--color-app-bg)] px-4 py-3 font-mono text-sm text-[var(--color-ink)] outline-none transition focus:border-[var(--color-accent)]"
           />
         </label>

apps/web/lib/ai-food-photo.ts

@@ -45,8 +45,8 @@ export type AnalyzeFoodPhotoResult =
 
 // Mirrors the backend's defaults for the AI gateway; the effort suffix is translated into the reasoning parameter.
 export const DEFAULT_FOOD_PHOTO_MODELS = [
-  "gpt-5.6-luna(low)",
-  "gpt-5.6-luna(medium)",
+  "gpt-6-luna(low)",
+  "gpt-6-luna(medium)",
 ] as const;
 
 function parseModelList(value: string | undefined) {

apps/web/tests/unit/ai-food-photo.test.ts

@@ -23,18 +23,18 @@ describe("food photo model config", () => {
     expect(getConfiguredFoodPhotoModels()).toEqual([
       ...DEFAULT_FOOD_PHOTO_MODELS,
     ]);
-    expect(getConfiguredFoodPhotoModel()).toBe("gpt-5.6-luna(low)");
+    expect(getConfiguredFoodPhotoModel()).toBe("gpt-6-luna(low)");
   });
 
   it("parses and dedupes the configured model list", () => {
     process.env.AI_GATEWAY_MODELS =
-      "gpt-5.6-luna(medium), gpt-5.6-terra(low)\ngpt-5.6-luna(medium)";
+      "gpt-6-luna(medium), gpt-5.6-terra(low)\ngpt-6-luna(medium)";
 
     expect(getConfiguredFoodPhotoModels()).toEqual([
-      "gpt-5.6-luna(medium)",
+      "gpt-6-luna(medium)",
       "gpt-5.6-terra(low)",
     ]);
-    expect(getConfiguredFoodPhotoModel()).toBe("gpt-5.6-luna(medium)");
+    expect(getConfiguredFoodPhotoModel()).toBe("gpt-6-luna(medium)");
   });
 
   it("falls back to the defaults when the list is only separators", () => {

docs/architecture/godfiles-audit.md

@@ -1,8 +1,8 @@
 # Godfiles audit and refactoring plan
 
-Status: analysis and plan only; no production code has been refactored.
+Status: partially implemented. The September 16 update below records completed work and revises the remaining extraction plan; it does not implement further production changes.
 
-Audit snapshot: `main`, `dev`, and `staging` at
+Historical audit snapshot: `main`, `dev`, and `staging` at
 `5bbfe766df371e17fa9bcd00013ccd18276dac34` on 2026-08-31. This refresh supersedes the original
 `dev` snapshot at `b40658ef5343a49ce952a595489f934b59c1aaab`.
 
@@ -33,6 +33,66 @@ the code harder to follow through indirection, introduce Rust visibility/cycle p
 Next.js server-action boundaries. The recommended approach therefore keeps existing public facades,
 moves one cohesive responsibility at a time, and proves behavior after every move.
 
+## Update: 2026-09-16 follow-up audit (branch `dev` at `7a354d9d`)
+
+Status: analysis only. No production code was changed for this update; it records progress since
+the previous snapshot and revises the remaining plan.
+
+Implemented since the last snapshot (verified in the tree at `7a354d9d`):
+
+- Inline Rust test modules moved to child files: `api/tests/mod.rs`, `auth/tests/mod.rs`,
+  `config/tests/mod.rs`, `legacy_api/tests/mod.rs`, `db/tests/mod.rs`, and root `tests/mod.rs`.
+- `db/api_tokens.rs`, `db/weight.rs`, `db/gym.rs`, and `db/healthkit.rs` extracted from `db.rs`.
+- `db/sql.rs` owns the shared `jsonb_build_object` projections and insert fragments that were the
+  duplicate-scan clones (plan item B1), and the pool-only `_with_executor` wrappers are gone (B2).
+  The complexity plan closed with zero duplicate clones and clean knip/clippy
+  (`complexity-reduction-results.md`).
+- Web: `use-action-runner.ts`, `use-lazy-collection.ts`, and `use-copied-flash.ts` hooks exist;
+  `gym-shell.tsx` was split into `gym-buddies-panel.tsx` and `gym-slot-form-modal.tsx`.
+- DB package: the Drizzle-era `..._ignored` varargs are removed from `backend-queries.ts` (D1);
+  `queries/api-tokens.test.ts` and `queries/weight.test.ts` exist.
+
+Current line counts (physical lines, production files only, at `7a354d9d`):
+
+| File | Lines | Change vs. previous snapshot |
+| --- | ---: | --- |
+| `apps/backend/src/db.rs` | 6,275 | was 11,240 |
+| `apps/backend/src/legacy_api.rs` | 2,094 | was 3,499; tests moved out |
+| `apps/backend/src/api.rs` | 1,362 | was 2,521; tests moved out |
+| `apps/web/components/dashboard-shell.tsx` | 1,469 | roughly unchanged; still the main UI godfile |
+| `apps/backend/src/db/tests/mod.rs` | 3,124 | grew into a new test monolith |
+| `apps/backend/src/db/gym.rs` | 1,030 | new module; near-godfile |
+
+### Revised guidance
+
+The domain-only split remains the right direction. Three additions make it work:
+
+1. Separate cross-domain workflows from plain domain persistence. `apply_template_json` (db.rs
+   around line 3093) loads a template, bulk-checks product access, resolves meal-group labels,
+   normalizes every item, then inserts transactionally; it is not template CRUD. Give such
+   workflows their own owner (`db/template_workflows.rs`, `db/onboarding.rs`, `db/dashboard.rs`)
+   so domain modules do not become mutually dependent. Dispatcher arms that still contain inline
+   SQL and transactions, such as the meal-group operations at db.rs lines 964-1117 and dashboard
+   composition at lines 1124-1135, should move their implementations into the owning domain
+   before the dispatch table itself is extracted into `db/rpc.rs`.
+2. Keep one explicit string-dispatched `rpc_json` table. Do not introduce a registry, macro
+   routing, or per-domain dispatchers that each recognize the same operation names. The contract
+   tests that pin operation strings and argument objects stay authoritative.
+3. Prevent `db/sql.rs` and `db/tests/mod.rs` from becoming the next godfiles. Move domain-specific
+   SQL projections and tests beside their owning modules as each extraction lands; keep only
+   genuinely shared projections in `sql.rs`, and keep the `SCHEMA_SQL` fixture and parity
+   assertion in one place.
+
+Updated remaining sequence: extract schema readiness and shared input parsing first; then foods
+and meal groups (their SQL currently sits inside dispatcher arms); then users, meals, templates,
+and recipes; then admin (split into `admin/{users,barcodes,audit}.rs`) and stats; then a thin
+`db/rpc.rs`. `api.rs` should split into `contract.rs`, `auth.rs`, `response.rs`, `input.rs`, and
+`resources/*` while keeping the single request lifecycle (authentication, scopes, method checks,
+timeout, envelope, CORS) centralized as it is today. `legacy_api.rs` should split by capability
+with a shared `gateway.rs` used by both food-photo and benchmark paths, rather than nesting the
+gateway under food-photo. The dashboard split stays valid; extract pure draft helpers and
+render-only sections first, and keep the coupled draft/mutation state in one owner.
+
 ## Audit method and decision rule
 
 The audit counted physical lines in every tracked source-like file, then inspected all production

infra/cliproxyapi/README.md

@@ -56,7 +56,7 @@ strangers' traffic with it.
    address bar and paste it back into the panel's callback field. The
    panel exchanges it and writes the credential to `/data/auths`.
 4. Verify: `curl -H "Authorization: Bearer $AI_GATEWAY_API_KEY"
-   https://<public-domain>/v1/models` should list `gpt-5.6-*` models.
+   https://<public-domain>/v1/models` should list `gpt-6-*` models.
 
 Fallback if the panel flow fails: run the proxy locally with Docker
 (`docker run --rm -it -p 1455:1455 -v ${PWD}/auths:/root/.cli-proxy-api
@@ -78,7 +78,7 @@ if customized) so the admin AI-benchmark page reflects the gateway models;
 it does not need the API key.
 
 `AI_GATEWAY_MODELS` is optional and defaults to
-`gpt-5.6-luna(low),gpt-5.6-luna(medium)` — low effort first (~2s to first
+`gpt-6-luna(low),gpt-6-luna(medium)` — low effort first (~2s to first
 token), medium as the retry fallback. `AI_GATEWAY_MODEL_TIMEOUT_MS` bounds
 each model attempt (defaults to 20s, clamped to 3–30s).