toga-ai 1.0.522 → 1.0.524

package/knowledge/1.0/apps/library/features/error-capture-1-0.md

@@ -134,11 +134,37 @@ written for a reader and carries no file, line, or trace.
 ### `Logs.Event.context` shape
 
 Identical in both frameworks: **every top-level key holds an object**, and scalars are grouped
-under a `Reference` object. Top-level keys are the display names — `Reference`, `GET`, `POST`,
-`Runtime`, `AWS`, `SERVER`. `GET`/`POST` are cast to objects so an empty one serializes as `{}`,
-not `[]`. `$_POST` is captured and redacted on the same key-name basis as `$_GET` (note PHP only
-populates `$_POST` for form encodings, so JSON API bodies deliberately do not appear). The
-over-length shed list was renamed in lockstep and its marker is an object too.
+under a `REFERENCE` object. Top-level keys are **UPPERCASE** display names (changed 2026-08-04) —
+`REFERENCE`, `GET`, `POST`, `RUNTIME`, `AWS`, `SERVER`, `GLOBALS`. `GET`/`POST` are cast to objects
+so an empty one serializes as `{}`, not `[]`. `$_POST` is captured and redacted on the same
+key-name basis as `$_GET` (note PHP only populates `$_POST` for form encodings, so JSON API bodies
+deliberately do not appear). The over-length shed list was renamed in lockstep and its marker is an
+object too. Four shapes of this payload coexist for the 7-day retention window — readers keep the
+`REFERENCE` → `Reference` → `reference` fallback chain.
+
+### `$GLOBALS` IS captured (reversal, 2026-08-04) — and 1.0 benefits most
+
+`capture.php` mirrors 2.0's `buildGlobalsContext()` / `sanitizeForContext()` /
+`CONTEXT_GLOBALS_TO_EXCLUDE` / `CONTEXT_GLOBALS_MAX_DEPTH` / `CONTEXT_GLOBALS_MAX_STRING_LENGTH`
+exactly. Full rationale, guards, the hostile-fixture results and the **open PII security note**
+live in the
+[2.0 doc](../../../2.0/apps/_underscore/features/error-reporting-issue-event.md#globals-capture--the-earlier-refusal-is-reversed-2026-08-04)
+— read it before touching either copy.
+
+Two things are specifically 1.0:
+
+- **1.0 crons are the big winner.** They keep almost all working state in globals, so a 1.0 cron's
+  captured context is now genuinely rich, where a 2.0 request (state in objects and locals) stays
+  thin. The limitation is that `$GLOBALS` reaches **global scope only** — PHP exposes no API for
+  another frame's locals; the trace's arguments are the nearest substitute.
+- **`App_Error::redactSensitiveKeys()` is NOT the sanitizer used here.** It is still live on the
+  **1.0 Sentry path**, and it redacts only the **top level** of an array and never sees inside an
+  object. That weakness is precisely why globals capture was refused before. Do not reach for it
+  from capture code, and do not assume it protects anything nested.
+
+**PHP 7.2 constraints on the mirror:** no `mixed` type, no named arguments, and no
+`ReflectionProperty::isInitialized()` (7.4+) — `getValue()` is wrapped in its own `try`/`catch`
+instead.
 
 ### Elastic Beanstalk environment name on AL1
 
@@ -182,6 +208,20 @@ exists in `api2`/`worker2`. So 1.0 reads the AL1 container config at
 
 ## Change history
 
+- 2026-08-04 (latest, **uncommitted/undeployed** at time of writing) — **Mirrored 2.0's reversal:
+  `$GLOBALS` is now captured** into a top-level `GLOBALS` group in `capture.php`
+  (`buildGlobalsContext()`, `sanitizeForContext()`, `CONTEXT_GLOBALS_TO_EXCLUDE`,
+  `CONTEXT_GLOBALS_MAX_DEPTH`, `CONTEXT_GLOBALS_MAX_STRING_LENGTH`), byte-compatible with
+  `_underscore/Error.php`. 1.0 crons benefit most because they keep working state in globals — which
+  is why Sentry's Globals section was so much richer than ours. Recorded that
+  `App_Error::redactSensitiveKeys()` (still live on the 1.0 Sentry path) redacts only the top level
+  of an array and never enters an object, which is exactly why capture was refused before the
+  Reflection-based sanitizer existed. Added `dsn`/`connection`/`passphrase`/`salt` to
+  `CONTEXT_SENSITIVE_KEY_FRAGMENTS` (a DSN embeds `user:password`). Top-level context keys are now
+  **UPPERCASE**. PHP 7.2 sweep on this copy: no `mixed`, no named arguments, no
+  `ReflectionProperty::isInitialized()` — `getValue()` is wrapped in its own try/catch. The **open
+  PII exposure** this widens is documented on the 2.0 doc; `library` and `tools` must both deploy
+  before any of it is observable. (jcardinal)
 - 2026-08-04 (later, **uncommitted/undeployed** at time of writing) — **Fixed: 1.0 errors recorded
   `clientId` NULL on every cron but one** (2,093 NULL `Logs.Event` rows vs 107 with a client; issue
   `1Q` from `crons/toga2/quad/import_po_and_asn.php`). `App_Api_Toga2::authenticate()` — 1.0's real

package/knowledge/1.0/apps/tools/features/errors-curation-console.md

@@ -107,6 +107,17 @@ statement about *response time*, not about the error, and nothing on the page le
 those apart. **Nothing renders while the clock is running but under 24h** — below that the clock
 contributes nothing to urgency, so saying "neglect" would overstate it.
 
+**The note must render on ONE line** (fixed 2026-08-04). `.neglect-note` was
+`display: inline-block`, so "neglect NORMAL · 26h" broke after "neglect" and the level dropped to a
+second line — the two halves then read as unrelated fragments. It is now `display: block` +
+`white-space: nowrap`.
+
+> **⚠ Widening the column is PART of that fix, not a nicety.** Under `table-layout: fixed`, a
+> `nowrap` element in a column too narrow to hold it **overflows the cell** rather than shrinking —
+> so `nowrap` alone would have traded a wrap for an overflow. `col-urgency` went **96px → 136px**
+> and `.issue-listing` `min-width` **1026px → 1066px** to match. Any future `nowrap` in this table
+> needs the same paired column-width change.
+
 **Deliberately scoped to the neglect contribution only.** The console never claims *which* axis is
 winning: computing the full driving axis in Tools would mean copying the volume, sustained and
 release-band thresholds too, and that much duplicated policy would rot.
@@ -209,15 +220,34 @@ They were renamed from plural on 2026-08-01, after the pipeline was already in p
 - **The distinct-EB-environment query is a full scan of `Event`.** Acceptable only because that
   table is bounded by the 7-day purge. At volume the proper fix is denormalising the EB environment
   onto its own `Event` column **at capture time**.
-- **Read both the old and new `Event.context` key positions.** The context shape was restructured
-  on 2026-08-04 (one object per top-level key, scalars under `Reference`); with 7-day retention,
-  both shapes coexist for a week after any such change.
+- **⚠ FOUR shapes of `Event.context` coexist — every reader must tolerate all of them.** The shape
+  was restructured on 2026-08-04 (one object per top-level key, scalars grouped) and the top-level
+  keys were then **uppercased** (`Reference` → `REFERENCE`, `Runtime` → `RUNTIME`). With 7-day
+  retention, older shapes stay in the table for a week after each change. `issue/get.php` therefore
+  reads **`'REFERENCE'` first, then `'Reference'`, then `'reference'`**, and the JSON-tree
+  open-by-default test is **case-insensitive** (`strtolower($key) === 'reference'`) so the reference
+  card stays expanded for both casings. The key humanizer needed no change — it already passes
+  all-caps keys through untouched. **This fallback-chain pattern is the rule for any context key
+  rename, not a one-off.** A new `GLOBALS` group also now appears on newer events (see the
+  [2.0 doc](../../../2.0/apps/_underscore/features/error-reporting-issue-event.md)) — it is the
+  largest group by far, so the largest-first card ordering will usually put it on top.
 - **Curated text is authoritative.** Once `isManaged` is set, later occurrences no longer update
   subject/description, so a stale curated subject stays stale until someone edits it again.
 
 ## Change history
 
-- 2026-08-04 (latest, **uncommitted/undeployed** at time of writing) — **Built:** a **neglect
+- 2026-08-04 (latest, **uncommitted/undeployed** at time of writing) — **Fixed:** the neglect
+  indicator wrapped onto two lines in the listing — `.neglect-note` was `display: inline-block`, so
+  "neglect NORMAL · 26h" broke after "neglect" and read as two unrelated fragments. Now
+  `display: block` + `white-space: nowrap`, **with `col-urgency` 96px → 136px and `.issue-listing`
+  min-width 1026px → 1066px as part of the same fix**: under `table-layout: fixed` a `nowrap`
+  element in a too-narrow column overflows the cell instead of shrinking, so nowrap alone would have
+  traded a wrap for an overflow. **Changed:** capture now emits **UPPERCASE** top-level
+  `Event.context` keys, so `issue/get.php` reads `'REFERENCE'` → `'Reference'` → `'reference'` and
+  its JSON-tree open-by-default test became case-insensitive; **four shapes of the payload now
+  coexist** across the 7-day retention window and every reader must tolerate them. A new `GLOBALS`
+  context group also starts appearing on newer events. (jcardinal)
+- 2026-08-04 (earlier) — **Built:** a **neglect
   indicator** on both pages — listing gained an `openTaskAgeHours` correlated subquery and a
   `$describeNeglect` amber note (`.neglect-note`, `#8a5a00`) under the urgency badge; the detail
   page's *Acknowledged* row now reads "not yet — waiting 3 days" with a NORMAL/HIGH neglect chip.

package/knowledge/2.0/apps/_underscore/features/error-reporting-issue-event.md

@@ -177,18 +177,98 @@ read the ambient value.
 
 ### `Event.context` shape (uniform, both frameworks)
 
-**Every top-level key holds an object**; scalars are grouped under a `Reference` object. The
-top-level keys *are* the display names: `Reference`, `GET`, `POST`, `Runtime`, `AWS`, `SERVER`.
+**Every top-level key holds an object**; scalars are grouped under a `REFERENCE` object.
+**Top-level keys are UPPERCASE** (changed 2026-08-04): `REFERENCE`, `GET`, `POST`, `RUNTIME`,
+`AWS`, `SERVER`, `GLOBALS`. `GET`/`POST`/`AWS`/`SERVER` already were; `Reference` → `REFERENCE`
+and `Runtime` → `RUNTIME` bring the rest into line. These are **group headings, not data**, and
+the mixed casing read as though the difference meant something.
+
 `GET`/`POST` are cast to objects so an empty one serializes as `{}` rather than `[]`. `$_POST` is
-now captured (redacted on the same key-name basis as `$_GET`; PHP only populates `$_POST` for form
+captured (redacted on the same key-name basis as `$_GET`; PHP only populates `$_POST` for form
 encodings, so JSON API bodies deliberately do not appear). The duplicate top-level
 `peakMemoryMb` was dropped, the over-length shed list was renamed in lockstep with its marker now
 an object, and 2.0 gained `framework` / `isBusiness` / `issueKey` / `errorFile` / `errorLine` for
 1.0 parity.
 
 A uniform shape is what lets a viewer render context **generically** (see the Tools console's
-card/JSON-tree renderer). Consumers must read **both** the old and new key positions during the
-7-day `Event` retention overlap.
+card/JSON-tree renderer).
+
+> **⚠ FOUR shapes of this payload now coexist in `Event`.** Retention is 7 days, so every reader
+> must tolerate the older ones until they purge. Tools reads `'REFERENCE'` first and falls back to
+> `'Reference'` then `'reference'`; its JSON-tree open-by-default test is **case-insensitive**
+> (`strtolower($key) === 'reference'`) so the reference card stays expanded for every shape. The
+> Tools key humanizer already passes all-caps keys through untouched. **Never rename a context key
+> without leaving the fallback chain in place.**
+
+### `GLOBALS` capture — the earlier refusal is REVERSED (2026-08-04)
+
+**`$GLOBALS` IS now captured**, into a new top-level `GLOBALS` group, in **both** frameworks
+(`_underscore/Error.php` `buildGlobalsContext()`; mirrored in `library/app/error/capture.php`).
+
+> This **reverses** the earlier decision on this page that `$GLOBALS` would deliberately never be
+> captured. That refusal was correct *for the sanitizer that existed at the time* — a key-name
+> blocklist that recursed **arrays only** and walked straight past framework singletons holding DB
+> credentials in **object properties**. That rationale is now the reason **the sanitizer had to be
+> rewritten**, not a reason to refuse capture.
+
+**Why reversed:** our context was compared against Sentry's "Globals" section for the same error
+and carried none of it. The variable that explains a failure is routinely the one nobody thought
+to allowlist in advance, and 1.0 crons in particular keep almost all working state in globals —
+which is exactly why the Sentry payload was so much richer.
+
+**What made it safe enough:** `sanitizeForContext()` **recurses into objects via Reflection,
+including PRIVATE and PROTECTED properties**, and tests every key at every depth.
+**Reflection, not `get_object_vars()`** — a credential is far likelier to be private than public,
+and *a sanitizer that cannot see a value cannot redact it*. (1.0's older
+`App_Error::redactSensitiveKeys()`, still live on the 1.0 Sentry path, redacts only the **top
+level** of an array and never looks inside an object. That is the old behaviour, not this one.)
+
+Guards:
+
+- `CONTEXT_GLOBALS_MAX_DEPTH = 6` — an object graph with parent pointers must not walk a whole
+  schema into a log row.
+- `CONTEXT_GLOBALS_MAX_STRING_LENGTH = 2000` per string, with the **true byte length appended**.
+- `CONTEXT_GLOBALS_TO_EXCLUDE` — the superglobals (already their own groups) plus
+  `GLOBALS`, `appRoot`, `newAppRoot`, `argv`, `argc`.
+- `Closure`s and resources render as placeholders.
+- **`GLOBALS` is shed FIRST by `encodeContext()`** when the payload exceeds
+  `CONTEXT_MAX_LENGTH` (60000), so an oversized dump costs itself and never the curated groups.
+
+Verified against a deliberately hostile fixture: a private `$dbPassword` nested inside an object
+property → `[REDACTED]`; `apiSecret` three levels deep in an array → `[REDACTED]`;
+`shippingAddress1`/`shippingPhone`/`shippingEmail` → `[REDACTED]`; a 5000-char string truncated; a
+resource and a `Closure` rendered safely.
+
+> **Known limitation — GLOBAL SCOPE ONLY.** A variable local to the function that threw is *not*
+> reachable from a handler; PHP exposes no API for another frame's locals. The stack trace's
+> arguments are the nearest substitute, and the trace already carries them. Practical effect: rich
+> capture for **1.0 crons**, comparatively thin for a **2.0 request** where state lives in objects
+> and locals.
+
+#### ⚠ SECURITY — residual PII that key-name redaction cannot catch (open, not solved)
+
+`Event.context` is a table **every client shares**, and its contents also flow into **ClickUp task
+bodies and escalation emails**. Key-name redaction cannot recognise PII **by value**, so from a
+real Compass MITS payload these are still stored in clear: `shippingName`, `shippingAttention`
+(person and company names), `shippingCity` / `shippingStateCode` / `shippingZip` (the rest of an
+address whose *line 1* is redacted), and `mitsClientId` (an `odp_pk_`-style publishable-looking
+key that no blocklist flags).
+
+This is the **same exposure that was deliberately closed earlier the same day** by dropping the
+`print_r($payload)` dump from
+`worker/crons/toga2/compass/workflow/1_transmit_compass_sales_orders_to_mits.php` — reopened
+through a wider door, knowingly, at the developer's direction. Sentry was an acceptable home for
+this data because it is access-controlled, single-tenant and not mirrored into tickets;
+`Logs.Event` is **none of those**.
+
+Three mitigations were offered and **not yet applied** (follow-ups):
+
+1. Extend the sensitive-fragment list (`city`, `zip`, `postal`, `attention`, `recipient`).
+2. Add **value-pattern** redaction — things that look like emails, phone numbers and card numbers
+   regardless of key name.
+3. Exclude named globals per-cron via `CONTEXT_GLOBALS_TO_EXCLUDE`.
+
+**Recommended: 2 + 1.**
 
 ### The quotable reference — base-32 over an unambiguous alphabet (reworked 2026-08-04)
 
@@ -320,12 +400,22 @@ in `_underscore` and `library` in the **same** release, and the `/errors` consol
   form. **Evaluation order matters:** in `IssueAreaOwners`, `consecutiveCount` is assigned
   **before** `clickupUserId` so the `IF()` compares against the *old* owner. Reversing them
   makes the condition always true and the counter never resets.
-- **`Events.context` is an allowlist, never a dump.** It replaced `print_r($GLOBALS)` filtered
-  by a key-name blocklist — that blocklist recursed **arrays only** and walked straight past
-  objects whose *properties* hold DB credentials, into a table shared by every client.
-  `requestUri` and `referer` have query strings stripped (a `?token=abc` was `[REDACTED]` under
-  `get` and stored in clear in `requestUri`). The sensitive-fragment list covers **PII**, not
-  just credentials. Persisted traces are rebuilt **argument-free**.
+- **`Event.context` is a curated allowlist PLUS a separately-sanitized `GLOBALS` group** (as of
+  2026-08-04 — see the GLOBALS section above; the earlier "never capture `$GLOBALS`" position is
+  **superseded**). The original `print_r($GLOBALS)` dump it replaced was filtered by a key-name
+  blocklist that recursed **arrays only** and walked straight past objects whose *properties* hold
+  DB credentials — that is why capture was refused until `sanitizeForContext()` gained Reflection
+  descent into private/protected properties. `requestUri` and `referer` have query strings stripped
+  (a `?token=abc` was `[REDACTED]` under `get` and stored in clear in `requestUri`). The
+  sensitive-fragment list covers **PII**, not just credentials — but only **by key name**, so
+  value-recognisable PII still gets through (see the security note above). Persisted traces are
+  rebuilt **argument-free**.
+- **`CONTEXT_SENSITIVE_KEY_FRAGMENTS` must include `dsn` and `connection`.** Added 2026-08-04
+  alongside `passphrase` and `salt`, in **both** frameworks. A property named `dsn` holding
+  `mysql:host=…` came through in clear on the hostile fixture — **a DSN embeds `user:password`**.
+  It had never mattered before because under the allowlist-only context no global ever reached the
+  redactor at all. Any new fragment must be added to `_underscore/Error.php` **and**
+  `library/app/error/capture.php` together.
 - **Keep the debug `print_r($GLOBALS)` dump inside the `isDebugMode()` branch.** It was being
   built on every exception — including the OOM path the memory reserve exists to protect.
 - **Strict-mysqli varchar overflow silently drops the row.** `_Error::initialize()` sets
@@ -430,6 +520,31 @@ clientUserId). **Neither was built.** As built instead:
 
 ## Change history
 
+- 2026-08-04 (latest, **uncommitted/undeployed** at time of writing) — **DECISION REVERSED:
+  `$GLOBALS` IS now captured**, as a new top-level `GLOBALS` group, in both frameworks. The earlier
+  entry recording that `$GLOBALS` would deliberately never be captured is **superseded and has been
+  corrected in place** — its rationale (a key-name blocklist recurses arrays only and steps past
+  framework singletons holding DB credentials in object properties) is now recorded as *the reason
+  the sanitizer had to be rewritten*. Reversed because our context was compared against Sentry's
+  "Globals" section and carried none of it; the variable that explains a failure is routinely one
+  nobody thought to allowlist, and 1.0 crons keep almost all state in globals. Made safe by a new
+  `sanitizeForContext()` that **recurses into objects via Reflection including private/protected
+  properties** (not `get_object_vars()` — a credential is likelier private, and a sanitizer that
+  cannot see a value cannot redact it), with depth cap 6, a 2000-char string cap that appends the
+  true byte length, superglobals + `GLOBALS`/`appRoot`/`newAppRoot`/`argv`/`argc` excluded,
+  Closure/resource placeholders, and `GLOBALS` **shed first** by `encodeContext()` past
+  `CONTEXT_MAX_LENGTH`. Verified against a hostile fixture. **Known limitation: global scope only** —
+  PHP exposes no API for another frame's locals, so 2.0 requests capture comparatively little.
+  **Fixed:** `dsn`, `connection`, `passphrase`, `salt` added to `CONTEXT_SENSITIVE_KEY_FRAGMENTS` in
+  both frameworks — a `dsn` property came through in clear on the fixture and a DSN embeds
+  `user:password`. **Changed:** every top-level `Event.context` key is now **UPPERCASE**
+  (`Reference`→`REFERENCE`, `Runtime`→`RUNTIME`); readers must tolerate **four** coexisting shapes
+  for the 7-day retention window. **Recorded as an OPEN security note (not solved):** key-name
+  redaction cannot recognise PII by value, so `shippingName`/`shippingAttention`/`shippingCity`/
+  `shippingStateCode`/`shippingZip`/`mitsClientId` are still stored into a table every client shares
+  and mirrored into ClickUp bodies and escalation emails — the same exposure closed earlier the same
+  day in the Compass MITS transmit cron, knowingly reopened at the developer's direction. Three
+  mitigations offered, none applied. (jcardinal)
 - 2026-08-04 — **Correction/diagnostic:** recorded that the `dbchanges2/Logs/` migration files lag
   the deployed schema — `2026-07-30a` describes a `code` identity column, but live `Logs.Issue` is
   keyed on **`issueKey` (varchar 128)** — and listed the live column set. Query

package/knowledge/2.0/apps/toga-blox/INDEX.md

@@ -9,4 +9,4 @@
 | [Primary Table templates (server/client, sizing, virtualization)](features/primary-table-templates.md) | `src/templates/PrimaryTable/` is the **production, wired-up table** built on the [Table component](table.md). | toga-blox/src/templates/PrimaryTable/PrimaryTable.tsx, toga-blox/src/templates/PrimaryTable/PrimaryTableServerTemplate.tsx, toga-blox/src/templates/PrimaryTable/PrimaryTableClientTemplate.tsx, toga-blox/src/templates/PrimaryTable/PrimaryTableHeaderCell.tsx, toga-blox/src/templates/PrimaryTable/PrimaryTableBodyCell.tsx, toga-blox/src/templates/PrimaryTable/PrimaryTableRow.tsx, toga-blox/src/templates/PrimaryTable/PrimaryTableExpandableRow.tsx, toga-blox/src/templates/PrimaryTable/types.ts |
 | [TableRecordModal (record-detail modal shell)](features/table-record-modal.md) | `TableRecordModal` is a **generic, presentational modal shell** for showing a single table record (row) in detail — typically opened from a table row click. | toga-blox/src/components/TableRecordModal/TableRecordModal.tsx, toga-blox/src/components/TableRecordModal/index.ts, toga-blox/src/components/TableRecordModal/tableRecordModal.module.css |
 | [Table component (cells, action cells, filters & sorts, hooks, theming)](features/table.md) | The `Table` component (`src/components/Table/`) is the **TanStack Table v8** building block behind the [Primary Table templates](primary-table-templates.md). | toga-blox/src/components/Table/index.ts, toga-blox/src/components/Table/types.ts, toga-blox/src/components/Table/utils/buildTanstackColumns.tsx, toga-blox/src/components/Table/utils/resolveCellType.tsx, toga-blox/src/components/Table/components/cellTypes, toga-blox/src/components/Table/components/actionCells, toga-blox/src/components/Table/components/columnFiltersAndSorts, toga-blox/src/components/Table/hooks, toga-blox/src/components/Table/themeConfig |
-| [Dynamic npm Publish Pipeline (branch → channel)](workflows/dynamic-publish-pipeline.md) | How `@agilant/toga-blox` (checkout folder `toga-blox-npm`, registry repo key `toga-blox`) publishes a per-environment npm **channel** (dist-tag) from a `_<mode> | toga-blox/.github/workflows/publish.yml, toga-blox/src/utils/getFontAwesomeIcon.tsx |
+| [Dynamic npm Publish Pipeline (branch → channel)](workflows/dynamic-publish-pipeline.md) | How `@agilant/toga-blox` (checkout folder `toga-blox-npm`, registry repo key `toga-blox`) publishes a per-environment npm **channel** (dist-tag) from a `_<mode> | toga-blox/.github/workflows/publish.yml, toga-blox/package.json, toga-blox/src/utils/getFontAwesomeIcon.tsx |

package/knowledge/2.0/apps/toga-blox/workflows/dynamic-publish-pipeline.md

@@ -6,10 +6,11 @@ project: TOGa Blox
 client: shared
 type: workflow
 status: active
-updated: 2026-07-23
+updated: 2026-08-04
 owners: [jcardinal, apeterson]
 files:
   - toga-blox/.github/workflows/publish.yml
+  - toga-blox/package.json
   - toga-blox/src/utils/getFontAwesomeIcon.tsx
 related:
   - ../architecture.md
@@ -99,11 +100,36 @@ environment needs its own maintained `_<mode>` branch in `toga-blox-npm`.
   `React.CSSProperties & Record<\`--${string}\`, string>` — the value type must be `string`
   (**not** `string | number`; `CSSVariables` requires `string`). Runtime and function signature
   are unchanged, so it is safe for the ~40 consumers.
+- **The `_production`/`_beta` build copies only `.scss` — a plain `.css` component
+  stylesheet is silently dropped from the tarball.** On `_production` and `_beta` the
+  `build` script is `tsc && copyfiles -u 1 "src/**/*.scss" dist`: `tsc` compiles a
+  component's `import "./Foo.css"` into the emitted JS, but `copyfiles` never ships the
+  `.css`, so the published package contains JS importing a file that isn't in the tarball.
+  Consumers then fail their Vite/Rollup build with `Could not resolve "./Foo.css"`
+  (seen with `EnvironmentBadge` in `@agilant/toga-blox@beta` 1.0.316-beta.108, breaking
+  `toga2-commerce`). **Convention on the prod/beta build: author component stylesheets as
+  `.scss`, never `.css`** (a flat `.css` file is valid SCSS as-is, so renaming is the
+  zero-risk fix — no build-script change). Note the build description in `architecture.md`
+  (copies `*.scss/*.css/*.module.css`) reflects the in-progress `feature-new-table` branch,
+  **not** the current `_production`/`_beta` build; that feature branch is not to be merged
+  to prod/beta just to pick up the wider copy glob.
+- **Diagnose what a channel actually shipped with `npm pack`.** To see exactly which files a
+  published dist-tag contains, run `npm pack @agilant/toga-blox@<tag>` then
+  `tar tzf <tarball>` — this is how the missing `.css` (JS importing an unshipped stylesheet)
+  was confirmed against the tarball rather than the source tree.
 - **Secret hygiene done right.** `publish.yml` writes a transient `.npmrc` at build time from
   GitHub secrets (`${NODE_AUTH_TOKEN}` / FontAwesome token) — it never commits tokens. This is
   the pattern the consuming apps should follow (see the commerce workflow's secret-hygiene note).
 
 ## Change history
+- 2026-08-04 — Recorded the prod/beta `.scss`-only build gotcha: the `_production`/`_beta`
+  `build` copies only `src/**/*.scss`, so a component shipped with a plain `.css` + matching
+  `import "./X.css"` publishes JS referencing a stylesheet absent from the tarball, breaking
+  consumers' Vite build (`Could not resolve "./EnvironmentBadge.css"`, `@agilant/toga-blox@beta`
+  1.0.316-beta.108 → `toga2-commerce`). Fix was renaming `EnvironmentBadge.css`→`.scss` (no
+  build-script change), republished on `_beta` as 1.0.316-beta.115 and a `fix/env-badge-scss-prod`
+  PR into `_production`; `toga2-commerce` `_beta` bumped to the fixed version. Added the
+  `npm pack`/`tar tzf` tarball-inspection diagnostic. (apeterson)
 - 2026-07-23 — Documented CI version derivation (strip suffix → patch+1 → append `-<mode>.<run>`;
   dev owns only the base version, never the suffix) and the `_sandbox-client` deploy procedure now
   driven by the `deploy-sandbox-client` project skill. Recorded that `publish.yml` has diverged

package/knowledge/2.0/apps/worker2/features/error-escalation-cron.md

@@ -348,6 +348,14 @@ this controller needs the same treatment.**
 
 ## Gotchas / known issues
 
+- **⚠ This cron is an EXFILTRATION PATH for `Event.context` — treat anything captured as
+  ticket-visible and email-visible.** As of 2026-08-04 capture writes a `GLOBALS` group into
+  `Event.context`, and redaction is **by key name only**, so value-recognisable PII (names, city,
+  state, ZIP, publishable-looking ids) survives into ClickUp task bodies and escalation emails —
+  neither of which is single-tenant or access-controlled the way Sentry was. Before widening what
+  this cron copies out of `context`, read the security note in the
+  [2.0 error-reporting doc](../../_underscore/features/error-reporting-issue-event.md). Mitigations
+  are proposed but **not applied**.
 - **An issue held at HIGH purely by neglect never de-escalates on its own.**
   `applyDeEscalationDwell()` needs the computed target to sit **below** the stored urgency for
   `REQUIRED_QUIET_RUNS` (3) consecutive runs, but the neglect axis keeps recomputing the target

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "toga-ai",
-  "version": "1.0.522",
+  "version": "1.0.524",
   "description": "TOGA Technology Team Claude Knowledge System — shared AI coding harness with skills, knowledge base CLI, and project installer for Claude Code.",
   "keywords": [
     "claude",