toga-ai 1.0.994 → 1.0.996

package/knowledge/2.0/apps/api2/features/api-payload-interceptors.md

@@ -6,7 +6,7 @@ project: API
 client: shared
 type: feature
 status: active
-updated: 2026-10-08
+updated: 2026-10-09
 owners: ["mhammontree", "dfranks", "bala", "snaredla", "jcardinal", "rgirish"]
 files:
   - api2/Component/Api/V2/V2.php
@@ -25,6 +25,8 @@ files:
   - dbchanges2/Client_Compass/2026-08-06 - RemoveBrokenSalesOrderItemPostPostInterceptor.sql
   - dbchanges2/Client_Compass/2026-08-14a - AddSalesOrderInactiveStandaloneItemPreInterceptors.sql
   - dbchanges2/Client_CompassCanada/2026-08-14a - AddSalesOrderInactiveStandaloneItemPreInterceptors.sql
+  - _underscore/Model/Client/UserGlobalSetting.php
+  - dbchanges2/Core/2026-10-09a - UserGlobalSettingsInterceptors.sql
 related:
   - ../../worker2/features/netsuite-transferorder-outbound-push.md
   - ../../../clients/rate/features/aig-contract-creation.md
@@ -40,6 +42,8 @@ related:
   - ../../_underscore/features/email-template-sending.md
   - ../../../clients/elite/features/salesorder-netsuite-push.md
   - ../../../clients/compass-usa/features/sales-order-line-renumbering.md
+  - ./language-translation-layer.md
+  - ../../_underscore/features/persona-nav-visibility-gate.md
 ---
 
 How `_Model` prePost/postPost/prePut/postPut hooks are dispatched by the api2 /v2 engine only when a metadata row exists — the missing-row silent no-op, the wrong-row hard fatal, and the persistence/transaction traps; open before writing, activating, or debugging any interceptor.
@@ -73,7 +77,7 @@ The lookup predicate is **`apiId IS NULL OR apiId = <the caller's api>`**, so a
 ### The interceptor sees payload keys verbatim (no `c_` stripping, no rename)
 
 **Verified empirically** by a live HTTP probe against local dev (`test/@dave/Junk Drawer/probe_c_prefix_payload_mapping.php`):
-- `$payload` carries the keys **exactly as the client sent them**. Sending `mitsSalesOrder` yields `$payload->mitsSalesOrder`; sending `c_mitsSalesOrder` yields `$payload->c_mitsSalesOrder`. **Column mapping happens AFTER the interceptor runs.**
+- `$payload` carries the keys **exactly as the client sent them**. Sending `mitsSalesOrder` yields `$payload->mitsSalesOrder`; sending `c_mitsSalesOrder` yields `$payload->c_mitsSalesOrder`. **Column mapping happens AFTER the interceptor runs** — including alias keys: a PRE hook sees `user` / `setting`, and api2 maps them onto `userId` / `settingId` only afterwards (security impact: [own-row guards](#own-row-guards-in-a-pre-interceptor)).
 - The rename in **`Client_<X>.Apis_CustomRecordFields.overrideFieldName`** governs only which field names the API **accepts on input**. Without a rename row the renamed name is rejected with *"a field specified in your request does not exist"*; the **raw model field name is always accepted**. Rename rows are themselves `apiId`-scoped, so two APIs writing the same record can legitimately use different names for one field.
 
 > **⚠ Do not re-derive this from the source.** Static reading of `api2/Component/Api/V2/V2.php` around **L7384** — where `$modelVarsToSet` is cast to an object after a `$flippedRenamedFields` flip — *looks* like the interceptor receives MODEL-level field names. It does not. The probe result above is authoritative; the code path is misleading.
@@ -243,6 +247,23 @@ The reverse direction: not "queue a worker from an interceptor" but **"run the p
 - **Do not use the interceptor path just to send an email.** For a plain send there is a documented non-API entry point: `_Model_Client_EmailTemplate::send(string $clientIdentifier, ...)`, versus the `$api`-based `sendEmail()`. See [Email Template Sending](../../_underscore/features/email-template-sending.md). Faking `$api` is for reusing *business logic*, not for reaching the mailer.
 - **Notes will not be logged.** `_Model_Client_Logs_Record::addNote()` is flushed only by `Component/Api/V2/V2.php`, so notes staged by the interceptor you just ran are **dropped** in a worker. See [Record Change Audit Log](../../_underscore/features/record-change-audit-log.md).
 
+### Own-row guards in a PRE interceptor
+
+A PRE hook that limits a user to their own rows must close four bypasses — each found by review on the `user-global-settings` guard (TRUE-82613):
+1. **Alias keys.** api2 maps `user` / `setting` (FK objects) onto `userId` / `settingId` **after** PRE runs, so a hook that only checks `$payload->userId` is skipped by sending `user: {uuid}`. **Allowlist the payload keys** and reject anything else.
+2. **Parent routes.** On `POST /v2/users/<uuid>/user-global-settings` api2 overwrites the FK the hook stamped with the parent's id **after** the hook. Refuse any POST with a parent segment.
+3. **PRE runs BEFORE the field-level write ACL check** — the hook cannot count on the ACL to have rejected a field yet.
+4. **PUT/PATCH carry no owner in the body.** Load the row by `$api->recordUuid`, refuse not-own rows, refuse changes to the owner/type columns, then `unset` them before the save.
+
+Refuse API-key callers when the rule is "the signed-in user" (no user on the token). Throw `_Exception_Validation` (400); use **one** message for "not yours" and "not allowed" so the response does not reveal whether a row exists.
+
+**Reference implementation:** `_Model_Client_UserGlobalSetting::prePost/prePut/prePatch` (Core.Records **279**), rows in `dbchanges2/Core/2026-10-09a - UserGlobalSettingsInterceptors.sql` (`clientId` + `apiId` NULL → **every tenant, every API**). Keys: POST `settingId, value`; PUT/PATCH `userId, settingId, value`. Only the language setting (Core.Settings id 2) is writable; `value` must be a code in the **client** `Languages` table (`_Model_Client_Language` — there is no Core Languages table and no active flag). POST stamps `userId` to the caller and refuses a second language row. No ACL grant ships with it yet.
+
+**Known own-row holes (open, not fixed — 2026-10-09):**
+- `Client_CompassCanada/2026-01-15 - UserGlobalSettingsAcl.sql` grants UserGlobalSettings writes with no own-row check.
+- `UserNavPreferences` checks the user on create, not on PUT/PATCH.
+- `TableViewPresets` has no PATCH check.
+
 ### Worked example — the EV-10 that was not a code bug
 
 `_Model_Client_ItemFulfillment::prePost()` defaults `itemFulfillmentStageId` to the shipped stage on a `POST /v2/item-fulfillments`. In **dev-sandbox** the `(recordId 28, PRE, POST)` row was missing, so `prePost` never ran, `itemFulfillmentStageId` stayed `NULL`, and the write failed with **`EV-10` "Column 'itemFulfillmentStageId' cannot be null"** — while the identical code worked in production. Fix was purely metadata: apply the existing `dbchanges2/Core/2026-06-30a - ItemFulfillmentStageDefaultInterceptor.sql`. **Diagnostic that works:** do a **full-table compare of `ApiPayloadInterceptors` between production and the failing environment**. It is a small table, and the drift is usually exactly one row.
@@ -262,6 +283,7 @@ The reverse direction: not "queue a worker from an interceptor" but **"run the p
 - **⚠ Validate that a payload ARRAY's elements are objects before touching them — `property_exists()` THROWS.** `property_exists()` raises a `TypeError` when argument 1 is an `array`, `int`, `null` or `bool` (a *string* merely returns `false`). So a `foreach` over e.g. `$payload->purchaseOrderItems` that calls `property_exists($item, …)` turns a malformed payload into an **HTTP 500** where the interceptor was supposed to answer **400**. Reject non-array containers and non-object elements **first**, with a plain `_Exception_Validation` (a sender coding fault is not a business condition — it should not raise an Issue or page anyone). Worked example: [Compass MITS PO zero-line-item rejection](../../../clients/compass-usa/features/mits-po-zero-line-item-rejection.md).
 - **⚠ A throw from a PRE interceptor leaves NO `Logs_<Client>.Api` row.** The transaction log is written near the END of `_Component_Api_V2::execute()` (~L2225-2271), *after* route processing (~L1742-2043), so an exception raised from `prePost` never reaches it and the rejected request looks like it never arrived. See [V2 request logging](./request-logging.md#-a-throw-from-a-pre-interceptor-leaves-no-api-log-row-at-all).
 - **⚠ Client behaviour never goes on `_Model_Client_X`.** Every tenant with a row for that record inherits it through `parent::` calls. Put it on `_Model_<Client>_X`.
+- **⚠ An own-row PRE guard that checks only `userId` is bypassable** — alias keys (`user`) map after PRE, and a parent-route POST overwrites the stamped FK. Allowlist keys + refuse parent routes. See [own-row guards](#own-row-guards-in-a-pre-interceptor).
 - **A client's rows may be in `Core` (with `clientId`), not in `Client_<X>`** — check both before concluding a client has no interceptors.
 - **⚠ The same hook in BOTH `Core` (with `clientId`) and `Client_<X>` fires TWICE** — both tables are read, nothing dedupes. Adyen (2026-10-08): Core ids 15–17 duplicated client rows 1, 2, 5 → double worker jobs and double NetSuite queueing; fixed by deactivating the Core rows. Keep each `(record, phase, method)` in one table only. See [Adyen reclaim flow](../../../clients/adyen/features/service-request-salesorder-netsuite.md#gotchas).
 - **The method name is derived, so a typo'd enum silently misses.** A row with `prePostProcessing = 'PRE'`, `httpMethod = 'PUT'` resolves to `prePut`, not `prePost`; the engine finds no method and moves on.
@@ -291,3 +313,4 @@ The reverse direction: not "queue a worker from an interceptor" but **"run the p
 
 ## Change history
 - 2026-10-08 — `internalApiRequest` "[]" throw also fires on the OUTER request's warnings (e.g. WO-1); throw site now ~L2436; Core + client duplicate rows fire twice. (rgirish)
+- 2026-10-09 — Own-row guards in PRE hooks (alias keys + parent-route bypasses, PRE before field ACL); user-global-settings reference; known open holes. (jcardinal)

package/knowledge/2.0/apps/api2/features/language-translation-layer.md

@@ -6,7 +6,7 @@ project: API
 client: shared
 type: feature
 status: active
-updated: 2026-09-16
+updated: 2026-10-09
 owners: ["jcardinal", "bala"]
 files:
   - api2/Component/Api/V2/V2.php
@@ -18,9 +18,12 @@ files:
   - _underscore/Model/Client/FeatureTranslation.php
   - _underscore/Model/Client/ItemCategoryFeatureGroupTranslation.php
   - _underscore/Model/Client/ItemFeatureTranslation.php
+  - _underscore/Model/Client/UserGlobalSetting.php
+  - dbchanges2/Client/2026-10-08l - LanguageOverridable.sql
 related:
   - ./acl-permission-chain.md
   - ../../_underscore/features/model-magic-field-access.md
+  - ./api-payload-interceptors.md
 ---
 
 How api2 serves per-language item/feature text via `audience.language` + per-table sidecar tables resolved in the API layer (ORM untouched); open when adding a translatable field, debugging a missing translation, or a collation-mismatch SQL error.
@@ -38,20 +41,18 @@ Uniform across all clients (no client variations). The sidecar table + ACL ship
 - `api2/Component/Api/V2/V2.php`: `resolveSettingValue($slug, $context)` (the 8-layer Settings-cascade resolver, private); `resolveValidLanguageCode()` (validates against `Languages.code`, default `en`); `getTranslatedFieldValue()` / `loadTranslationSidecar()` / `shouldTranslate()` / `addTranslationFallbackWarning()` (read path); `extractTranslationWrites()` / `saveTranslationWrites()` (write path); `lookupTranslatableFieldByRecordFieldId` built in the RecordFields load (~line 200).
 - `_underscore/Model/Client/ItemTranslation.php` — the first sidecar model.
 
-**Language resolution (auth time, user auth only).** `resolveSettingValue('language', …)` walks the Settings Matrix in order — DefaultGlobal, DefaultApp, ClientGlobal, ClientApp, PersonaGlobal, PersonaApp, UserGlobal, UserApp — each overriding the previous when set. A Client/Persona layer with `isOverridable = 0` **locks** the value against all later layers. For the persona layers, the **first persona** (in the user's persona order) that sets it wins. The resolved code is validated against `Languages.code` (fallback `en`), embedded in the JWT `id.language` claim, and echoed on every response as `audience.language`. API-credential auth gets no language.
+**Language resolution (auth time, user auth only).** `resolveSettingValue('language', …)` walks the Settings Matrix in order — DefaultGlobal, DefaultApp, ClientGlobal, ClientApp, PersonaGlobal, PersonaApp, UserGlobal, UserApp — each overriding the previous when set. A Client/Persona layer with `isOverridable = 0` **locks** the value against all later layers. For the persona layers, the **first persona** (in the user's persona order) that sets it wins. The resolved code is validated against the client `Languages.code` (see *Invalid saved language* below), embedded in the JWT `id.language` claim, and echoed on every response as `audience.language`. API-credential auth gets no language.
 
 **Language re-resolution on token refresh (user tokens only).** `/auth/refresh` re-resolves the user's language and embeds it in the newly minted tokens, so a language switch takes effect on the next hourly refresh **without a full re-login** (previously the refresh path copied the decoded JWT payload verbatim and only reset `exp`/`iat`, so a language change was invisible until re-auth). Runs **only** for user tokens — reads `id.client.user.uuid` off the token and skips entirely for API-credential and public tokens (no language claim). Rebuilds the resolution context from the existing token — loads `_Model_Client_User` by `uuid` (→ `userId`), `_Model_Core_App` by `id.app.uuid` (→ `appId`), takes `personaIds` from `id.client.personas` — then calls the **same** private `resolveSettingValue(slug: _Model_Core_Setting::SLUG_LANGUAGE, context: {appId, personaIds, userId})` helper the login path uses (same 8-layer cascade; the client DB is already registered). The resolved code is validated against `_Model_Client_Language` (load by `code`):
 - **Valid** → sets `$this->jwtPayload->id->language` to the resolved code.
 - **Unset/empty** → defaults to `'en'`.
-- **Invalid** (non-empty but absent from `Languages`) → **keeps the token's current language** and adds a `WZ-4` warning. It deliberately does **not** fail the refresh — a bad setting must not log an active user out on their hourly refresh (contrast the login path, which hard-fails with EV-16).
+- **Invalid** (non-empty but absent from `Languages`) → **keeps the token's current language** and adds a `WZ-4` warning. Never fails the refresh.
 
 Placement is load-bearing: the `id.language` mutation happens **after** client-DB registration and **before** both the access and refresh tokens are signed, so both new tokens carry the updated language.
 
-**Invalid-language hard failure (auth time).** The resolved code is validated against the client's `Languages` table. An **unset/empty** setting defaults to `en` (no error). But a **non-empty** code that does **not** exist in `Languages` now **fails authentication** with HTTP 400 `EV-16 DEFINED_MESSAGE_ERROR_INVALID_LANGUAGE_SETTING` (validation group, defined in `Response.php`), with the offending `language` code in the message `identifiers` — instead of silently falling back to `en`.
+**Invalid saved language — warn, never block (login + refresh).** Since 2026-10-09 (TRUE-82613) `/v2/auth/login` no longer fails with EV-16 when the user's resolved code is missing from the client `Languages` table. It falls back via private `resolveClientFallbackLanguageCode($appId)` — the same cascade with **no persona/user layers** (default + client layers only), validated again — else `_Model_Client_Language::CODE_BASE` (`en`), and adds `WZ-4 DEFINED_MESSAGE_WARNING_INVALID_LANGUAGE_SETTING` (WARNING, `{language}` identifier, `Response.php`). Status stays `201` with tokens. OAuth sign-in unsets `messages`, so OAuth callers never see WZ-4. Refresh keeps the token's current language instead (above). `EV-16` is still defined in `Response.php` but nothing raises it. Unset/empty → `en`.
 
-**Invalid-language soft warning (refresh time).** On the refresh path an invalid code is handled with a distinct, non-fatal WARNING instead of EV-16: `WZ-4 DEFINED_MESSAGE_WARNING_INVALID_LANGUAGE_SETTING` (type WARNING, group VALIDATION, `{language}` identifier placeholder; defined in `Response.php`, added via `addDefinedMessage(...)` so the warning has a stable catalogued code). Because it is a WARNING it carries no `status` key and does not change the response status — the refresh succeeds and the token keeps its current language. Deliberately softer than EV-16 (status 400), which would fail the refresh and log the user out.
-
-**Auth-mint status finalization (subtle).** The auth success path sets `status = 201` and attaches the tokens. EV-16 is signalled by a local flag `$invalidLanguageSetting` set at the resolution site; finalization branches `if (empty($invalidLanguageSetting))` → `201` + tokens, else `addDefinedMessage(EV-16)` + `data = null`. See Gotchas — do **not** convert this to an `is_null($this->response->status)` guard.
+**Writing a user's language.** `user-global-settings` writes (Core.Records 279) are guarded by PRE interceptors on `_Model_Client_UserGlobalSetting` — own row, language setting only, value must exist in client `Languages`. Rules + the bypasses they close: [API Payload Interceptors → own-row guards](./api-payload-interceptors.md#own-row-guards-in-a-pre-interceptor). Clients get the new language on the next refresh as `audience.language`.
 
 **Which fields are translatable (metadata-driven).** `Core.RecordFields.translationRecordFieldId` (new column, positioned after `recordId`) points a source field's RecordField at the sidecar field's RecordField. `buildLookups`/RecordFields-load resolves this into `lookupTranslatableFieldByRecordFieldId[sourceRecordFieldId] => {sidecarRecordId, sidecarField}`.
 
@@ -88,9 +89,9 @@ Placement is load-bearing: the `id.language` mutation happens **after** client-D
 - The resolver/read/write deliberately use `_Model::load()` per row, not a SQL JOIN/COALESCE — less SQL-efficient but the team's chosen approach; the sidecar is cached per row to avoid N-per-field loads.
 - `audience.language` reflects the *requested* language even if a given field lacks a translation; the per-field fallback warning signals the English fallback.
 - The base language is `en`; when the resolved language is `en` (or API auth) all translation logic is skipped and responses are byte-identical to pre-feature.
-- **"Why isn't my language setting applying?"** A `ClientGlobalSettings` row with `isOverridable = 0` **locks** the value against all downstream persona/user layers — a `UserGlobalSettings` (or persona) override never wins while the client-global layer is locked. By design (the 8-layer cascade with `isOverridable` locking). To let user/persona overrides apply, the locking layer must be `isOverridable = 1` (or its value cleared). A data-config gotcha, not a code bug.
+- **"Why isn't my language setting applying?"** A `ClientGlobalSettings` row with `isOverridable = 0` **locks** the value against all persona/user layers — a `UserGlobalSettings` override never wins. `dbchanges2/Client/2026-10-08l - LanguageOverridable.sql` (fan-out) set the language row (`settingId 2`) to `isOverridable = 1` on every existing client. **New-client onboarding seeds still create it locked** (left as-is by decision) — a new tenant whose users cannot change language needs the same `UPDATE`.
 - **Namespace gotcha in `V2.php` (`namespace api;`).** The api-namespaced classes (`_Component_Api_V2_Response`, `_Component_Api_V2_Response_Message`) must be referenced **unqualified** (resolves to `api\…`) or with the full `\api\_Component_Api_V2_Response_Message`. A bare leading backslash (`\_Component_Api_V2_Response_Message`) forces the global namespace and throws `EO-1 "Could not find required file"` via the autoloader. The translation fallback-warning code hit this.
-- **Do NOT guard the auth-mint finalization with `is_null($this->response->status)`.** The auth response auto-adds client `name` + user `firstName`/`lastName` and runs `processRoutePairs()`, which sets `$this->response->status` to `200` **before** finalization. An `is_null` guard there suppresses the `201` + tokens and yields a broken `200`/`data:null` login. Branch on the explicit `$invalidLanguageSetting` flag instead.
+- **Do NOT guard the auth-mint finalization with `is_null($this->response->status)`.** The auth response runs `processRoutePairs()` (client `name`, user `firstName`/`lastName`), which sets `status = 200` **before** finalization, so an `is_null` guard yields a broken `200`/`data:null` login. Set `201` unconditionally.
 - **The all-fields (expand-everything) serializer branch must also translate.** `V2.php` has two nested-child serialization paths: when the caller requests **specific** fields it already ran each value through `getTranslatedFieldValue()`, but when the caller requested **no** specific fields (empty `$varOptions`, i.e. "return everything I have access to") the code copied raw child values straight into `$outData` and never translated — so any field returned via the all-fields expansion stayed in base English even for a non-base-language user. Both branches must route through `getTranslatedFieldValue($record, $field, $sourceModel, $defaultValue)` (preserving the `renamedFields` alias). Symptom: on the Compass Canada French storefront item-detail page (loaded via `/bundles`), feature-spec **values** were French but the feature-group **headers** (`itemCategoryFeatureGroup.name`, e.g. "Processeur"/"Mémoire") stayed English, because the group object serialized through the all-fields branch. **Not a query-depth problem** — a commerce-side `depth:6` theory was wrong and reverted; the fix returns French headers at the default `meta.depth 3`.
 - **Sidecar/source collation mismatch — and the collations DIFFER PER CLIENT DATABASE.** The sidecar tables use `utf8mb4_unicode_ci` (matching `ItemTranslations`) while the source field's collation varies by tenant: verified in prod **and** dev-sandbox, `Client_CompassCanada.Items.title` is `utf8mb4_0900_ai_ci` but `Client_Compass.Items.title` is `utf8mb4_unicode_ci`. So a hardcoded `COLLATE utf8mb4_bin` "fix" is brittle — works in one client DB, unnecessary or wrong in another. Consequences:
   - **Never build a mixed-collation fallback expression in SQL.** `COALESCE(NULLIF(ItemTranslations.title, ''), Items.title)` produces a result whose collation is forced to `utf8mb4_bin` (MySQL cannot reconcile the two inputs). Plain `SELECT`ing it works, so it looks fine — but **any** comparison, `ORDER BY`, `GROUP BY` or `DISTINCT` on that expression throws **MySQL error 1267 "Illegal mix of collations"**. Reproduced against prod. A latent landmine: the query ships green and fails the first time someone sorts or dedupes on it.
@@ -110,3 +111,6 @@ Placement is load-bearing: the `id.language` mutation happens **after** client-D
 - [api2 Architecture](../architecture.md)
 - [Persona Name Translation](../../_underscore/features/persona-name-translation.md) — metadata-only worked example (no api2 change).
 - [_Model magic-field access](../../_underscore/features/model-magic-field-access.md)
+
+## Change history
+- 2026-10-09 — Login no longer fails on an unknown saved language (EV-16 retired): client-language fallback + WZ-4; language row made overridable on all clients; user-global-settings write guard linked. (jcardinal)

package/knowledge/clients/compass-canada/profile.md

@@ -19,7 +19,7 @@ project: _Underscore
 client: compass-canada
 type: profile
 status: active
-updated: 2026-10-07
+updated: 2026-10-09
 owners: [jcardinal, bala, tcox, apeterson, ajean, mhammontree, rgirish]
 files: []
 related:
@@ -175,6 +175,9 @@ to but distinct from Compass USA. Like Compass USA it spans the **2.0** commerce
   block, and entity-escaped because the send is iso-8859-1. See
   [French Order Email Localization](features/french-order-email-localization.md). French
   `EmailTemplates` bodies must be pure ASCII.
+- **⚠ Open security hole (2026-10-09, not fixed):** `Client_CompassCanada/2026-01-15 - UserGlobalSettingsAcl.sql`
+  grants UserGlobalSettings writes with **no own-row check** — a user could write another user's row. See
+  [own-row guards](../../2.0/apps/api2/features/api-payload-interceptors.md#own-row-guards-in-a-pre-interceptor).
 - **Approval decisions use the shared Compass parent.** `_Model_Compass_Canada_ApprovalDecision`
   is an empty subclass of `_Model_Compass_ApprovalDecision`; the entire approval/notification/
   manager-reassignment/VIP-auto-approve flow lives in the parent and branches at runtime on

package/knowledge/standalone/apps/forward/architecture.md

@@ -86,5 +86,5 @@ Treat this as obfuscation only. Anything genuinely sensitive must not rely on it
 - New **encrypted** domains require a code change (`$encryptedDomains` in `index.php`), not just a `forward.ini` edit.
 - `index.php` only runs on a map miss, so a bad/over-broad `forward.ini` host-only line can shadow paths you expected PHP to handle.
 - **EB rejects bundles over 500 MB** ("The artifact is empty or exceeds the maximum size supported by AWS Elastic Beanstalk"), and the CodePipeline EB deploy ignores `.ebignore`. Never commit large static assets here — that broke every deploy from 2026-10-08 until demos moved to S3.
-- **Git history still holds ~500 MB of old demo blobs.** Removing them needs a history rewrite + force-push — deferred; developer's call.
+- **Git history was rewritten on 2026-10-09** (`git filter-repo`) to drop ~500 MB of old demo blobs; a clone is now ~39 MB. Any clone made before that date must be deleted and re-cloned — pulling into it brings the old history back. GitHub's `refs/pull/1/head` and `refs/pull/2/head` still point at old commits (GitHub blocks pushes to them).
 - **Removing a top-level folder:** first check no Route 53 name pointing at forward EB has a first label matching it (local directory handler). Checked for the 2026-10-09 removal; `togadesk.beta.agilantsolutions.com` is redirected by `forward.ini` first.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "toga-ai",
-  "version": "1.0.994",
+  "version": "1.0.996",
   "description": "TOGA Technology Team Claude Knowledge System — shared AI coding harness with skills, knowledge base CLI, and project installer for Claude Code.",
   "keywords": [
     "claude",