@terreno/mcp 57.15.0 → 57.16.1

package/dist/docs/versioned/57.15.0/implementationPlans/agent-ui-rich-replies.md

@@ -0,0 +1,359 @@
+# Agent UI Rich Replies: steppers, checklists, galleries, and copy
+
+**Status:** approved 2026-10-07
+**Task list:** [`docs/tasks/agent-ui-rich-replies.md`](../tasks/agent-ui-rich-replies.md)
+**Supporting skills:** `terreno-ui`, `terreno-backend-api`, `ai-prompt-governance`, `update-docs`, `verify-ui-changes`, `backend-test-env`
+**Source:** OpenAI, "GPT-6 and Intelligent UI for everyone", the "Planning a Sunday roast" GPT-5.6 vs GPT-6 Instant comparison (fetched 2026-10-07).
+
+Terreno agents can already reply with whole-reply YAML block documents: headings, markdown,
+metrics, cards, charts, dataset tables, images, callouts, and buttons that reply, open,
+select, or call a host callback. OpenAI's GPT-6 launch shows a meal-plan reply that goes
+further: a summary card at the top, a three-photo collage, a photo menu list, a guest-count
+stepper that recalculates the shopping list, a "Copy shopping list" button, and a timed
+cooking checklist with a "1 of 8" counter. The GPT-5.6 version beside it uses a markdown
+"How much lamb?" table. This plan adds the missing pieces, so a Terreno agent can
+return both replies. The new blocks are on by default (D6); a host whose shipped clients cannot render them yet sets `uiBlocks.richBlocks: false`.
+
+## The idea
+
+Six additions to the closed block catalog in `@terreno/blocks`, each rendered by
+`BlocksView` with `@terreno/ui` components:
+
+| Demo element | Terreno block | Interaction |
+| --- | --- | --- |
+| "Your dinner plan" summary with a small label above the title | `card.eyebrow` (new optional field) | none |
+| Three grouped hero photos | `gallery` (2–6 images, optional captions) | none |
+| Menu items: photo, bold title, description | `list` (items with optional image) | none |
+| Guest count − / + with recalculated quantities | `stepper` | **host callback round trip** |
+| "Copy shopping list" | `copy` action kind on a button | on the device (clipboard) |
+| Cooking checklist with times and "n of m" | `checklist` | **host callback per tick** (local when the host has no checklist action) |
+| "How much lamb?" guests table | `table` renders typed columns; markdown tables in `text` get theme styling | none |
+
+The stepper follows the decision made in chat: **− and + are callback buttons like the
+existing ones.** A tap sends `POST /gpt/actions` with the stepper's callback name and
+`{...callback.payload, value: <next>}`. The host action returns
+`{replace: "block", blocks: {v: 1, blocks: [<new stepper>]}}`, and `GPTChat` swaps the block by
+id through `overrides`, as it does today. `@terreno/ai` ships an opt-in host action,
+`scaleStepperHostAction`. It reads the stepper the agent wrote from the stored history, not
+from the tap request, and scales every item linearly from it. Scaling from the agent's
+original keeps rounding from drifting tap after tap. The owner of a history can still edit
+its stored prompts through `PATCH /gpt/histories/:id`, so an app whose numbers matter
+(prices, stock) must compute them from its own data, not from the stored block.
+
+```yaml
+- type: card
+  eyebrow: How much should you buy?
+  children:
+    - type: stepper
+      id: guests
+      label: Number of people
+      unit: People
+      value: 5
+      min: 1
+      max: 20
+      callback: {name: scaleStepper}
+      items:
+        - {label: Bone-in leg of lamb, amount: 2, unit: kg, decimals: 1}
+        - {label: Carrots, amount: 8, round: up}
+      note: Generous portions, with a little extra.
+    - type: actions
+      id: copy_list
+      elements:
+        - {type: button, id: copy, text: Copy shopping list, action: {kind: copy, target: guests}}
+```
+
+The checklist works the same way (D2). Each tick sends a callback with the full set of ticks,
+and `toggleChecklistHostAction` returns the checklist with those ticks. Photos (D3) come from
+a library of images generated once with the Vertex image model and stored through
+`FileStorageService`. The agent picks from it with a `findPhotos` tool and writes `file:` ids.
+
+Rejected: a shared document `state` with `{{formula}}` bindings computed on the device.
+It needs an expression language, and the human chose a server round trip.
+
+## The plan
+
+| # | Task | Depends on | Lands in | Proves it |
+| --- | --- | --- | --- | --- |
+| T1 | Stepper block round trips through a host callback (tracer) | none | blocks, ui, ai | fixture, renderer, handler, and route tests |
+| T2 | Prompt offers the new blocks only when the host opts in | T1 | blocks, ai | prompt + chat-turn tests |
+| T3 | Table block renders typed columns | T1 | ui | `DataTable` + `BlocksView` tests |
+| T4 | Markdown tables in `text` get theme styling | T3 | ui | `MarkdownView` test |
+| T5 | `checklist` contract | T2 | blocks | fixtures |
+| T6 | `gallery` contract | T5 | blocks | fixtures + image-host lint |
+| T7 | `list` contract | T6 | blocks | fixtures + image-host lint |
+| T8 | `card.eyebrow` contract | T7 | blocks | fixtures |
+| T9 | `copy` action contract + `blockPlainText` | T8 | blocks | fixtures + serializer tests |
+| T10 | Checklist renders and toggles | T3, T4, T5 | ui | `BlocksView` test |
+| T11 | Gallery renders | T6, T10 | ui | `BlocksView` test |
+| T12 | List renders | T7, T11 | ui | `BlocksView` test |
+| T13 | Card eyebrow renders | T8, T12 | ui | `BlocksView` test |
+| T14 | Copy action writes the clipboard | T9, T13 | ui | `BlocksView` test |
+| T15 | Sunday roast golden document + playground story | T14 | blocks, demo | fixture + story screenshot |
+| T16 | Example app: the roast reply against the real callback routes | T15, T20, T21 | example-backend, example-frontend | Playwright + recording |
+| T17 | Model smoke: a real model writes the roast reply | T15, T20 | example-backend | smoke script output |
+| T18 | Docs sweep, changelog, rules | T15, T20, T21 | docs, rules | `website:build`, `rules:check` |
+| T19 | Generated photo library: generate images once and store them | none | example-backend | model + script tests |
+| T20 | `findPhotos` tool and photo URLs in the example app | T19 | example-backend, example-frontend | tool + action tests |
+| T21 | Checklist ticks round trip through a host callback | T10 | ai | handler + route tests |
+
+Tracer: T1, from the agent's YAML through `validateBlocks`, `BlocksView`, the callback
+event, `POST /gpt/actions`, and `scaleStepperHostAction`, back to a replaced block.
+Out of scope: formulas or shared state, scripted html, persisting overrides, maps, tabs,
+interactive diagrams, generated tools (see Expansions).
+Open risks:
+- Each tap is one round trip. `pendingElementIds` disables both buttons until the reply
+  lands, so fast taps are dropped, not queued (Q4).
+- `scaleStepperHostAction` finds the agent's block by `messageId`, which is `msg-<index>`
+  when the host sets no message id. When that misses, it uses the only assistant prompt that
+  holds a stepper with that `blockId`. If more than one does, it returns 409 rather than
+  guess.
+- Stored history prompts are writable by their owner, so the stored stepper is not a trust
+  boundary (see the idea above; the how-to says so).
+
+## Contracts
+
+Field names are the contract. Limits go in `BLOCK_LIMITS` and are printed in the prompt
+from there. Lint paths use the existing `blocks[0].images[2].src` form.
+
+### `stepper` (T1)
+
+| Field | Required | Rule |
+| --- | --- | --- |
+| `id` | yes | block id pattern, at most 54 characters (room for the `_increase` suffix) |
+| `label` | yes | 1–80 |
+| `value`, `min`, `max` | yes | numbers; `min` < `max`; `min` ≤ `value` ≤ `max` (`OUT_OF_RANGE`) |
+| `step` | no | > 0, default 1 |
+| `unit` | no | 1–40, shown under the value ("People") |
+| `callback` | yes | `{name, payload?}`; `name` is checked by `UNKNOWN_HOST_ACTION` when `hostActions` is passed |
+| `itemsTitle` | no | 1–80, a small heading above the items ("Your shopping quantities") |
+| `items` | no | 0–12 of `{label (1–80), amount (number), unit? (1–20), decimals? (0–3, default 0), round?: nearest \| up (default nearest)}` |
+| `note` | no | 1–280, muted under the items |
+
+The renderer reserves the element ids `<id>_decrease` and `<id>_increase`. Lint reports
+`DUPLICATE_ID` when any other block or element uses them.
+
+Rendering: the label as an eyebrow, `IconButton` − and +, then the value with its unit,
+`itemsTitle`, the items as a two-column label / amount grid formatted with `decimals` ("2.0 kg"), then the
+note. − is disabled when one more step would go below `min`, + when it would go above `max`.
+Both are disabled while either element is pending,
+and when `callback.name` is not in `hostActions`. Accessible labels: "Decrease <label>",
+"Increase <label>".
+
+Callback event: `blockId: <id>`, `elementId: <id>_increase` or `<id>_decrease`, action
+`{kind: "callback", name, payload: {...payload, value: value ± step}}`.
+
+`scaleStepperHostAction` (in `@terreno/ai`) is `{handler, payload, handles: "stepper", logResponse: false}`. Its `payload` schema is `z.object({value: z.number()}).passthrough()`, and the handler reads only `value`.
+
+1. Loads the stepper the agent wrote with `findAgentBlock({history, messageId, blockId, type})`,
+   a shared helper (`ai/src/service/agentBlocks.ts`). It reads `history.prompts[n]` when
+   `messageId` is `msg-<n>` (stored prompts have no ids), parses it with `parseBlocks`, and
+   finds the block of that type whose id is `blockId`. When that misses, it uses the only
+   assistant prompt holding such a block. With more than one, it returns 409. With none, 404.
+2. Checks that `value` is a number within `[min, max]` on the `step` grid. Anything else
+   returns 400.
+3. Returns `{replace: "block", blocks: {v: 1, blocks: [stepper]}}`. The returned stepper has
+   the new `value`, and each item's `amount` is `original amount × value / original value`,
+   rounded by that item's `decimals` and `round`.
+
+`logResponse: false` is a new optional host-action field. With it, the `AIRequest` row for
+`POST /gpt/actions` keeps today's ids-only prompt, adds the numeric `value`, and leaves out
+the returned block. Every other host action keeps logging the full response, as today.
+
+### Default on, with an opt-out (T2)
+
+| Option | Effect |
+| --- | --- |
+| `uiBlocks.richBlocks` (default `true`) | The prompt offers `checklist`, `gallery`, `list`, `card.eyebrow`, and `copy`. `false` removes them. |
+| `richBlocks` on **and** a host action with `handles: "stepper"` | The prompt offers `stepper` and names those actions as its callbacks. Without a stepper action, the prompt never mentions `stepper`, because its buttons would be dead. |
+| `richBlocks` on **and** a host action with `handles: "checklist"` | The checklist prompt line tells the model to set `callback` to that action. Without one, it says to leave `callback` out (ticks stay local). |
+
+`blocksPromptSection` gains `richBlocks` (default `true`, matching the server), `stepperActions`, and `checklistActions`.
+`validateBlocks` gains the same two lists: a stepper or checklist whose `callback.name` is not
+in its list fails `UNKNOWN_HOST_ACTION`. `chatTurn` derives both lists from the host actions'
+`handles` field. The server always accepts the new blocks when it validates. The
+opt-out only stops the model from writing them for clients that cannot render them yet.
+
+Upgrading changes behaviour: every host with `uiBlocks` on gets the new blocks in its prompt
+after the bump. The changelog entry says so under a "Behaviour change" heading and names
+`richBlocks: false` as the opt-out.
+
+### `checklist` (T5)
+
+`{type: checklist, id (at most 31 characters), title? (1–120), callback?: {name, payload?}, items: 1–30 of {id (at most 32 characters), text (1–120), detail? (1–280), meta? (1–40), checked?: boolean}}`.
+Item ids are unique. The element id of a tick is `<id>_<item id>`, reserved like the stepper's
+(`DUPLICATE_ID`). Rendering: the title on the left and "n of m" on the right, then one
+`CheckBox` row per item: the bold `text` first, then `meta` (the time, small and muted), then a muted
+`detail` (D7).
+
+With a `callback`, an `onAction` handler, and either no `hostActions` list or one that names the
+callback (matching the stepper), a tick sends
+`{kind: "callback", name, payload: {...payload, itemId, checked, state: {<item id>: boolean, ...}}}`
+with `blockId: <id>` and `elementId: <id>_<item id>`. `state` is the full set of ticks after
+this one, because overrides live only in the client (D4). The whole checklist is disabled until
+the reply lands, so a second tick cannot send a `state` that misses the first, and the checklist
+then shows the returned block. Ticks are not applied ahead of
+the reply. Without a `callback`, without `onAction`, or when a `hostActions` list leaves the name
+out, ticks are local.
+
+`toggleChecklistHostAction` (in `@terreno/ai`, T21) is `{handler, payload, handles: "checklist", logResponse: false}`
+with payload `z.object({itemId: z.string(), checked: z.boolean(), state: z.record(z.string(), z.boolean())}).passthrough()`.
+It loads the agent's checklist with `findAgentBlock`, rejects item ids the checklist does not
+have (400), and returns `{replace: "block", blocks: {v: 1, blocks: [checklist]}}` with
+`checked` set from `state`. Apps that record progress register their own `handles: "checklist"`
+action.
+
+### `gallery` (T6)
+
+`{type: gallery, id?, images: 2–6 of {src, alt (1–200), caption? (1–120)}}`. Each `src` follows
+the `image` rules, including `IMAGE_HOST_NOT_ALLOWED`. Rendering: up to 3 images sit in one
+row of equal 4:3 tiles; more than 3 wrap into a 3-column grid. Narrow screens scroll the row
+sideways (Q5).
+
+### `list` (T7)
+
+`{type: list, id?, items: 1–12 of {title (1–120), text? (1–500, plain text), meta? (1–40), image?: {src, alt}}}`.
+Image sources follow the `image` rules. Rendering: a 3:4 portrait thumbnail on the left,
+then the bold title and muted text.
+
+### `card.eyebrow` (T8)
+
+Optional `eyebrow` (1–60) on `card`, rendered small and muted above `title`.
+
+### Generated photo library (T19, T20)
+
+The agent cannot make photos on the fly, and `https` images need an allowlist. So the example
+app generates a library once and stores it.
+
+| Piece | Contract |
+| --- | --- |
+| `PhotoLibraryEntry` model (example-backend) | `{fileAttachmentId, gcsKey, alt (1–200), tags: string[] (1–12), prompt}`, every field with a `description`; `createdUpdatedPlugin`, `isDeletedPlugin`; `prompt` is unique. A soft-deleted entry still holds its prompt, and `--force` leaves the old stored object and its `FileAttachment` behind |
+| `bun run photos:generate` (example-backend script) | Reads `src/scripts/photoPrompts.ts` (about 12 entries: roast lamb, crispy potatoes, honey carrots and parsnips, lemony greens, apple crumble, a laid table, and so on). It runs each through AI SDK `generateImage` with the Vertex image model (`imagen-4.0-fast-generate-001` unless `PHOTO_IMAGE_MODEL` is set), uploads it with `FileStorageService.upload` as a system user, and upserts a `PhotoLibraryEntry` by `prompt`. Re-runs skip prompts that already exist unless `--force` is passed. Requires `GOOGLE_VERTEX_PROJECT`, Application Default Credentials for Vertex, and `GCS_BUCKET`, validated up front; `GOOGLE_VERTEX_LOCATION` (default `us-central1`) picks a regional endpoint for Imagen. Uploads are owned by a find-or-create system user `photo-library@system.invalid`. A failed prompt is logged, the run continues, and the exit code is 1 if any failed |
+| `findPhotos({query, count})` tool | Matches `query` words against `tags` and `alt`, and returns up to `count` (1–6) `{src: "file:<entry id>", alt}` |
+| `photoLibrary` `modelRouter` with an instance action `url` | `GET /photoLibrary/:id/url` returns `{url}` (a `FileStorageService.getSignedUrl`) to any authenticated user. The library is shared, so the per-user `GET /files/*` route does not fit. List and read are `IsAuthenticated`; create, update, and delete are disabled |
+| example-frontend `resolveImage` | Resolves `file:<id>` through the generated SDK hook for `url` |
+
+The library holds only generated food photos, no user data.
+
+### `copy` action (T9)
+
+A new `action.kind`: `{kind: copy, text?, target?}`. Lint requires exactly one of the two,
+reporting `MISSING_REQUIRED` as it does for `open`. `target` must name a `stepper`,
+`checklist`, `list`, `table`, or `text` block (`COPY_TARGET_INVALID`). `text` holds at most
+4,000 characters (`copyTextMaxLength`); `target` is a block id.
+`blockPlainText(block, {datasets?, checked?})` in `@terreno/blocks` turns a block into the
+copied text:
+
+| Block | Text |
+| --- | --- |
+| `stepper` | `<label>: <value> <unit>`, then one `<item label>: <amount> <unit>` line per item |
+| `checklist` | one `[x] text` or `[ ] text` line per item, using the current toggles |
+| `list` | one `- title: text` line per item |
+| `table` | a header row, then the rows, tab-separated |
+| `text` | the markdown as written |
+
+`BlocksView` copies the text from the **current** block, with the override and toggles applied,
+using `expo-clipboard`. It then shows "Copied" in a polite live region next to the button for
+2 s. It does not call `onAction` or the server.
+
+### Tables (T3, T4)
+
+T3: `DataTable` gets real `number` and `date` cells. `number` is right-aligned. `date` is
+formatted with Luxon (`DateTime.fromISO(...).toLocaleString(DateTime.DATE_MED)`). The
+`table` block renderer stops forcing every column to `text` and `width: 120`. It maps the
+dataset's column types to these cells. It measures the container with `onLayout` and splits
+the width evenly, at least 96 per column, scrolling sideways past that. Number headers
+right-align to match their cells. Headers show column names with underscores as spaces, and the
+copied header row does the same. The table sits in a frame sized to its header plus rows (54
+each, up to 10 rows), so it never spills over the next block; longer tables scroll inside it.
+
+T4: `react-native-markdown-display` already lays GFM tables out as views. T4 adds theme
+styling: `border.default` cell borders, a bold header row on `surface.secondaryLight`, and
+horizontal scroll when the table is wider than its container. Its test fails on `master`
+first.
+
+## Acceptance criteria
+
+| ID | Criterion | Verification |
+| --- | --- | --- |
+| AC1 | Each new block and the `copy` action has at least one valid and one invalid fixture, and `jsonSchema.test.ts` agrees with Zod on all of them | `bun test blocks/` |
+| AC2 | Every new error code (`OUT_OF_RANGE`, `COPY_TARGET_INVALID`) appears in `docs/reference/blocks.md` | `blocksDocParity.test.ts` |
+| AC3 | With `richBlocks: false`, `blocksPromptSection` output is byte-identical to `master`, even when a stepper action is registered. With `richBlocks` unset, `addGptRoutes` and `blocksPromptSection` both behave as `true`. On, the prompt has one line per new block (added as each contract lands; the full snapshot is taken in T9), limits come from `BLOCK_LIMITS`, and the stepper line names the stepper actions | prompt snapshot tests |
+| AC4 | Tapping + on a 5-person stepper emits a callback with `payload.value: 6`. An item `{amount: 2, unit: kg, decimals: 1}` renders "2.0 kg". Rendering an override with `value: 6` shows 6 and the new amounts | `ui/src/blocks/BlocksView.test.tsx` |
+| AC5 | Through `POST /gpt/actions`, with a stored stepper (5 people; lamb `2` kg / 1 dp; parsnips `7` / up; broccoli `625` g / nearest) and `value: 6`: the result is `{replace: "block", blocks: {v: 1, blocks: [stepper]}}` with lamb 2.4, parsnips 9 (nearest would give 8), and broccoli 750. An extra agent `callback.payload` key is accepted. `value: 21` (max 20) returns 400. An unknown `blockId` returns 404. Two stored prompts with the same stepper id and a missed `msg-<n>` return 409. The `AIRequest` row has the ids and `value`, and no `response` | `ai/src/service/scaleStepper.test.ts`, `ai/src/routes/gptActions.test.ts` |
+| AC6 | Without a callback, toggling an item changes "1 of 8" to "2 of 8" on the device. With a callback in `hostActions`, the tick emits `payload: {itemId, checked: true, state}` with all 8 items, disables the whole checklist while the tick is pending, and the counter shows "2 of 8" only after the override arrives | `BlocksView.test.tsx` |
+| AC6b | Through `POST /gpt/actions`, `toggleChecklistHostAction` returns the stored checklist with `checked` from `state`. An unknown item id returns 400. The `AIRequest` row has no `response` | `ai/src/service/toggleChecklist.test.ts`, `ai/src/routes/gptActions.test.ts` |
+| AC7 | Gallery, list, and eyebrow render with `@terreno/ui` components only, with `alt` as the accessible label | `BlocksView.test.tsx` |
+| AC8 | Pressing a copy button whose target is a stepper at value 6 writes the scaled list to the clipboard (mocked `expo-clipboard`) and shows "Copied" | `BlocksView.test.tsx` |
+| AC9 | A `number` column renders right-aligned and a `date` column renders Luxon `DATE_MED`. A GFM table in `text` has themed borders, a bold header, and scrolls sideways when wide. Both tests fail on `master` | `DataTable.test.tsx`, `BlocksView.test.tsx`, `MarkdownView.test.tsx` |
+| AC10 | `blocks/src/fixtures/golden/sunday-roast.yaml` reproduces the GPT-6 demo reply (title, gallery, summary card, menu list, stepper + copy card with `itemsTitle`, checklist, note, tips, follow-up buttons) plus the GPT-5.6 "How much lamb?" table. It validates with `hostActions: ["scaleStepper", "toggleChecklist"]`, `stepperActions: ["scaleStepper"]`, `checklistActions: ["toggleChecklist"]`, and `imageHosts: ["images.example.com"]` in its own test, outside the option-less `valid/` loop | `bun test blocks/` |
+| AC11 | In example-frontend, the roast reply is stored in a real `GptHistory` and streamed by the mocked model. Pressing + changes 5 to 6 through the real `POST /gpt/actions` on example-backend. Copy writes the list. A checklist tick goes through `POST /gpt/actions` and updates the counter. Gallery and list images load through `GET /photoLibrary/:id/url`, with only the storage download stubbed (`page.route` on the signed-URL host serves a fixture image) | `bun run frontend:e2e` (`e2e/ai-chat.spec.ts`), with screenshots and a recording attached to the PR (`/opt/cursor/artifacts/` is read-only in this sandbox) |
+| AC12 | With a model key set, a generated photo library in the database, and the `findPhotos` tool registered, `bun run blocks:smoke` sends "Give me a plan for a sunday lamb roast, I'm having friends over still figuring out numbers tbh" through `AIService` with the opted-in prompt. It asserts that the reply validates and contains `gallery`, `list`, `stepper`, `checklist`, and a `copy` action, and that every image `src` is a `file:` id that `findPhotos` returned. Without a key it exits 0 and says it skipped | script exit code; output saved to `--out`. **Partly met:** the checks are unit-tested with a mock model, but no real model has run it (no key in the build sandbox) |
+| AC14 | `photos:generate` with a fake image model and a fake storage service creates one entry per prompt, and a second run creates none. `findPhotos({query: "roast lamb", count: 3})` returns 3 `file:` ids. `GET /photoLibrary/:id/url` returns a URL to a non-admin user, and create returns 405 | `example-backend` tests |
+| AC13 | Docs: reference rows, a how-to "Add a stepper callback", the explanation catalog fixed (it lacks `callout`, `image`, `details`, and `html` today), and a changelog entry | `bun run website:build`, `bun run rules:check` |
+
+## Assumptions
+
+- "Serving layout" means the stepper card: the label, − value +, the unit, the quantities grid,
+  and the note (Q6).
+- "Summary in an outline at the top" means the summary card: an eyebrow, a title, and a
+  bold-label bullet list written in markdown inside `text`. No new outline block.
+- Generating the library is an operator step run once per environment (`bun run photos:generate`). Tests use a fake image model and fake storage, so CI never calls Vertex.
+- "The tables" means the GPT-5.6 "How much lamb?" guests table, plus tables in general (Q7).
+- Copy feedback is inline "Copied" text in a live region, as in the demo, not a toast.
+- iOS and Android render through the same `BlocksView`. Builds already shipped do not know
+  the new blocks and would show an error banner, with the raw YAML collapsed, for the whole reply
+  that uses them. With the default on
+  (D6), a host with such builds must set `richBlocks: false` until they update. X10 (a
+  client-declared catalog version) would remove that manual step.
+- The `copy` action never reaches the server, so `hostActions` lint does not apply to it.
+  Local checklist ticks (no callback) don't either.
+- Image tiles reuse `useResolvedImages` for `file:` ids.
+- Storage: no new collection. Each stepper tap adds one `AIRequest` row. Like every host
+  action today, its prompt holds only ids; this adds the numeric `value`. Unlike other host
+  actions, the returned block is left out (`logResponse: false`), because an agent can put
+  anything in a stepper label.
+- Fixtures under `fixtures/valid/` use `data:image` or `file:` sources, because that loop
+  validates without `imageHosts`. Invalid cases that need options go in `lint.test.ts`.
+
+## Open questions (recommendation assumed)
+
+| ID | Question | Recommendation | Why | If answered differently | Tasks affected |
+| --- | --- | --- | --- | --- | --- |
+| Q3 | Should `@terreno/ai` ship `scaleStepperHostAction`, or should each app write its own handler? | Ship it, opt-in, exported from `@terreno/ai` | Any app gets the roast reply in one line; apps with real logic (prices, stock) register their own `handles: "stepper"` action | Drop the export; example-backend owns the handler | T1, T16 |
+| Q4 | Should + / − send one callback per tap, or batch taps (300 ms debounce, then one call)? | One per tap, with the buttons disabled while pending | Matches the existing callback buttons; simplest to prove | Add a debounce and send the final value | T1 |
+| Q5 | Should `gallery` take a `layout` (`collage`, `grid`, `carousel`)? | No: one adaptive layout | Fewer choices for the model; the demo uses one row | Add the enum and two more renderer branches | T6, T11 |
+| Q6 | Does "serving layout" mean the stepper card (− 5 +, People, quantities grid)? | Yes | It sits between the stepper and the copy button in the demo | Name what it is; likely one more block or field | T1 |
+| Q7 | Does "the tables" mean the GPT-5.6 "How much lamb?" table (guests → lamb, potatoes)? | Yes: as a `table` block with typed columns, plus themed markdown tables | It is the only table in the post; the GPT-6 reply shows quantities as a grid | Name the table and what it should do (sort, scale with the stepper) | T3, T4, T15 |
+
+## Expansions (follow up later)
+
+| ID | Idea | Why it came up | Rough size | Depends on |
+| --- | --- | --- | --- | --- |
+| X1 | Generated tools in chat (savings calculator, bill splitter, game) | The post's "Create interactive experiences" section | L | a state + formulas design, or more host callbacks |
+| X2 | Interactive diagrams ("change an input to see what happens") | The post's central limit theorem example | L | X1 |
+| X3 | `map` block with stops and notes | The post's road-trip example | M | none |
+| X4 | Full eval that the model picks these blocks for planning questions (beyond the T17 smoke) | The post trains the model's design judgment | M | T17 |
+| X5 | Copy button on markdown code fences | A common chat copy target; not in the demo | S | T14 |
+| X6 | Sortable table columns | `DataTable` supports `sortable` already | S | T3 |
+| X7 | Write `blocks.json` to `blocks/schemas/` for native clients | `blocksJsonSchema` exists but is not published there | S | T9 |
+| X8 | Streaming inside a block (partial lists) | The post's progressive compiler | M | none |
+| X9 | Persist overrides (if Q2 stays no) | A reload resets the stepper | M | Q2 |
+| X11 | Generate images per reply (a `generatePhoto` tool) | D3 picked a stored library instead | M | T19 |
+| X10 | A client-declared catalog version, so the server picks blocks per client | Shipped native builds lag behind the server; more pressing now that D6 turns the blocks on by default | M | T2 |
+
+## Decisions
+
+| ID | Question asked | Answer | What it changes |
+| --- | --- | --- | --- |
+| D1 | How should the agent express the +/− servings interaction: fixed block types computed on the device, shared state with formulas, or scripted html? | "When they press the button it should round trip to the server for it to process the function callback. That's a callback button like the others we have." | `stepper` − / + are host callbacks through `POST /gpt/actions`; the server returns the replaced block. No formulas, no scripted html. |
+| D2 | Q1: Should checklist ticks also round-trip to the server, like the stepper? | "Callback per tick" | `checklist.callback`, `toggleChecklistHostAction`, and T21. Ticks wait for the server; without a registered checklist action they stay local. |
+| D3 | Q8: Where should a real agent get photos for the gallery and menu list? | First "Generated images", then: "change q8 to generate some images and store them" | A library generated once with the Vertex image model and stored through `FileStorageService` (T19). The agent picks with `findPhotos` and writes `file:` ids, resolved by `GET /photoLibrary/:id/url` (T20). No per-reply generation and no `imageHosts`. |
+| D4 | Q2: Should stepper and checklist changes survive a page reload? | "No (Recommended)" | Overrides stay in client memory. X9 stays parked. |
+| D5 | Approve the plan? | "Approve (Recommended)", given with D2–D4 | Status set to approved with these answers applied. |
+| D6 | (unprompted) | "enable rich blocks by default" | `uiBlocks.richBlocks` and `blocksPromptSection` default to `true`; `false` opts out. AC3 now pins the `false` case to `master` and checks the default. The changelog flags the behaviour change. The stepper still needs a `handles: "stepper"` action. |
+| D7 | (review of the built PR) | "in blocks playground, ensure every kind of block is available. also the serving callback doesn't work. if it cant work, wire up toasts or something. flip the checkboxes to be bold title, then time, then description. none of the images render." | The playground offers a preset per block type and an "All blocks" preset. The stepper and checklist callbacks run locally there through pure `scaleStepperBlock` / `applyChecklistState` helpers moved into `@terreno/blocks` (the host actions call the same helpers); other actions show a toast. Checklist rows read bold title, time, description. Demo stories and presets use bundled public-domain food photos through `file:` ids instead of colour swatches. Tasks T22–T25. |
+
+## Sign-off
+
+Status: approved 2026-10-07 (chat; answers in D2–D6). Built 2026-10-08; AC12 partly met (see AC table).
+Cut: 2 rounds, 23 findings: 21 fixed, 2 moved to questions or expansions (round 1 F4 → Q7, round 2 F2 → Q8, since answered as D3), 0 rebutted. Sign-off answers changed T5, T10, T16–T21; no cut re-run, since less than half the tasks changed and the tracer did not move

package/dist/docs/versioned/57.15.0/implementationPlans/agentic-sdlc-plugin.md

@@ -0,0 +1,236 @@
+# Implementation Plan: Loop-Engineering Lifecycle Plugin
+
+**Status:** Complete — shipped as `plugins/terreno-planning`
+**Approved:** 2026-08-22
+**Roadmap issue:** https://github.com/TerrenoLabs/terreno/issues/1006
+**Priority:** High
+**Owner:** agent
+**Created:** 2026-07-27
+**Program:** [OSS launch](oss-launch-program.md)
+
+## Goal
+
+Evolve Terreno's existing planning plugin into five reusable, bounded lifecycle
+transitions optimized for repeated fresh agent invocations:
+
+```text
+Grow (shape) → Pick (build) → Roast (prove) → Brew (submit) → Taste (react)
+```
+
+The outer loop owns invocation, persistence, Taste `PENDING` reinvocation, retry, stop,
+and escalation. Brew waits in-process for async review bots. Taste waits in-process for
+review bots and for product CI (GitHub CLI or CircleCI CLI watch loop). Before any
+push it pulls latest `master`, runs root `prepush` when present (otherwise affected
+package lint/typecheck/tests) in a no-context subagent, then pushes and watches CI.
+Lifecycle skills own how one stage is performed. Repository skills own how this codebase
+works. Roast owns independent acceptance proof. State/evidence bridge fresh invocations.
+
+This is a refactor of the existing strong workflow, not a parallel implementation.
+
+## Scope
+
+- Canonical five plugin skills and migration from former names
+- Explicit transition contracts for every stage
+- Shared machine-readable result and execution-state schemas
+- Supporting-skill discovery/composition
+- Repository-specific detail removed from portable stages
+- Brew exits after PR setup and the review-bot wait
+- Taste performs one reactive iteration after waiting for review bots and product CI, then exits
+- Outer-loop reference including small-feature and CI/review scenarios
+- Static validation for names, status values, transitions, portability, and loop bounds
+- Documentation, metadata, and generated integrations
+
+## Non-scope
+
+- A daemon/hosted scheduler implementation
+- Automatic merging
+- Copying Terreno's repository skills into the plugin
+- Persisting chain-of-thought or conversational transcripts
+- Hundreds of generic micro-skills
+
+## Decisions
+
+| ID | Decision |
+| --- | --- |
+| AP1 | Public reusable plugin; repository-specific knowledge remains local |
+| AP2 | Sensitive-data rules cover credentials, customer data, PII/PHI, and evidence media |
+| AP3 | Canonical stages are Grow, Pick, Roast, Brew, Taste |
+| AP4 | Stages discover supporting skills by description; exact skill names are never universal dependencies |
+| AP5 | Taste is one observe/act/emit iteration after in-process waits for async review bots and product CI. Before any push: always pull latest `master`, record last-run failed tests and re-verify them locally, then run the repository root's `prepush` package script when present (otherwise affected-package lint, typecheck, and tests) in a fresh subagent with no parent conversation, then push and watch CI (`gh` / `circleci`). The outer loop reinvokes on Taste `PENDING` |
+| AP6 | Shared results use compact `v: 2` YAML; required keys are `v`, `stage`, `status`, `next`, `action`; empty keys are omitted; YAML is collapsed for humans |
+| AP7 | Existing repository state convention wins; fallback reuses loop-owned `.terreno/pipeline/<slug>.json`, not committed by default |
+| AP8 | Brew emits PR/head state and exits; direct Taste invocation is standalone compatibility only |
+| AP9 | No deprecated command aliases: old implementation-Roast conflicts with new verification-Roast and no maintained alias mechanism exists |
+| AP10 | Plugin major version is `2.0.0` because lifecycle semantics and command names are breaking |
+| AP11 | Grow ends with a standalone approval brief — orientation paragraph, optional background, idea, plan — then an unbounded Decisions table pairing every grilled decision with the question that prompted it, or omits that table when there were none; grilling stays on a question until the answer is executable |
+| AP12 | Brew and Taste wait until Bugbot, CodeQL, and similar review bots on the current head have reported, preferring hooks targeted to the matched bot or harness subscriptions over timer polling; unfiltered PR-check watches are product-CI waits owned by Taste |
+| AP13 | Product CI is every discovered host (GitHub Actions, CircleCI, Buildkite, GitLab CI, and similar). Taste observes native jobs when GitHub checks are incomplete; Brew confirms each host triggered or documented a not-applicable skip. An unexplained untriggered host prevents Brew `PASS`; a documented skip is terminal for Taste |
+| AP14 | Taste waits in-process with the provider's bounded native watch command (`gh pr checks --watch`, `gh run watch`, `circleci run watch`, `bk build watch`) in a loop until jobs are terminal or the wait times out. Outer loops honor Taste `PENDING` with the same hooks. Watch exit codes trigger a fresh classification rather than becoming stage verdicts directly |
+
+## Architecture
+
+```text
+                 OUTER LOOP
+       invoke / persist / wait / retry
+           stop / route / escalate
+                     |
+                     v
+ GROW → PICK → ROAST → BREW → TASTE
+ shape  build   prove   submit  react once
+                     |
+     repository skills + IP/task/state + evidence
+```
+
+### Ownership
+
+| Layer | Owns |
+| --- | --- |
+| Outer loop | when, who, next stage, persistence, Taste `PENDING` reinvocation, retry, stop, escalation |
+| Lifecycle plugin | portable stage procedure, invariants, evidence, transitions |
+| Repository skills | commands, frameworks, package conventions, safety rules, gotchas |
+| Roast | requirement → method → evidence → verdict |
+| State/evidence | durable facts required by a fresh invocation |
+
+### State transition table
+
+| Stage | PASS | FAIL | BLOCKED/PENDING |
+| --- | --- | --- | --- |
+| Grow | Pick | Grow with artifact defects | Human/external gate |
+| Pick | Roast | Pick with new evidence-based hypothesis | Human/external gate |
+| Roast | Brew | Pick with exact failed criteria | Classify environment/external/human |
+| Brew | Taste | Pick/Roast/Brew according to evidence | Human/access gate |
+| Taste | Merge-ready | Focused Taste retry only with new evidence | `PENDING`: loop waits then Taste; `BLOCKED`: named gate |
+
+## Shared result contract
+
+Every stage emits compact `v: 2` YAML: required `v`, `stage`, `status`, `next`, `action`;
+optional `ip`, `task`, `attempt`, `branch`, `sha`, `pr`, `skills`, `done`, `checks`,
+`artifacts`, `fail`, `block`, `ask`, `wait`. Omit nulls and empty arrays. Humans see
+`status` / `next` / `action`; the YAML is collapsed in chat and in the PR Details toggle.
+
+Schemas:
+
+- `plugins/terreno-planning/references/stage-result.schema.json`
+- `plugins/terreno-planning/references/execution-state.schema.json`
+
+No reasoning transcript is persisted.
+
+## Execution state
+
+The IP owns design/scope; tasks own executable slices. State stores only current
+stage/attempt/head/PR, prior structured result (`last`), artifacts, attempted approaches
+(`tried`), and next transition. The outer loop preserves/transports it between fresh
+invocations.
+
+State files are excluded from Brew commits unless a repository explicitly tracks them.
+After a PR exists, each Taste invocation also reconstructs current truth from the PR and
+current head rather than trusting stale state.
+
+## Stage contracts
+
+### Grow
+
+Researches repository facts, distinguishes human decisions from discoverable facts and
+low-risk conventional details, grills until each answer is executable, then writes
+approved, implementation-ready IP/tasks. Every acceptance criterion maps to verification.
+Approval shows a standalone brief — orientation paragraph, optional background, idea,
+plan — plus a full Decisions table with prompting questions when any grilled decisions
+exist.
+
+### Pick
+
+Implements one slice with Specify → Encode failing test → Fulfill → Clean the Kitchen.
+It consumes prior Roast/Taste evidence, preserves verified behavior, records failed
+approaches, runs independent implementation and test-quality reviews, and emits evidence
+for Roast.
+
+### Roast
+
+Runs independently, preferably in fresh context. It builds a requirement/evidence matrix
+and executes objective checks. It does not normally fix implementation code. `FAIL`
+returns exact expected/actual evidence to Pick.
+
+### Brew
+
+Requires Roast proof, runs repository-defined final checks/review, commits/pushes, applies
+the PR template, attaches evidence, confirms product CI triggered on every discovered
+host, and waits for async review bots (Bugbot, CodeQL, and similar) on the current head.
+Waits prefer provider CLI watch hooks or harness subscriptions and use bounded polling
+only as fallback. Brew records PR/current head and exits; it does not wait for ordinary
+product CI or execute Taste.
+
+### Taste
+
+Waits in-process if Bugbot, CodeQL, or similar review bots are still running, preferring
+provider CLI watch hooks or harness subscriptions, then waits in a loop for product CI
+using GitHub CLI (`gh pr checks --watch`, `gh run watch`) or CircleCI CLI
+(`circleci run watch`) until jobs on every discovered host are terminal or the wait
+times out. It then classifies mergeability and reviews, performs one bounded set of
+actionable fixes. Before any push it always pulls latest `master`, records last-run
+failed tests and re-verifies them locally, then spawns a fresh
+subagent with no parent conversation to run root `prepush` when present, falling back to
+lint, typecheck, and locally affected tests in each affected package. It then pushes and
+watches review bots and product CI. It may act
+once more after that watch. It emits
+`PASS`/`PENDING`/`BLOCKED`/`FAIL` and exits. The outer loop reinvokes after timeout or a
+second post-fix push.
+
+## Supporting skills
+
+Every stage inventories skills exposed by the harness/repository and loads those whose
+descriptions match the affected domains. Useful project skills may describe APIs, UI,
+data, schemas, prompts, test environments, deployment, documentation, or verification.
+No exact name is hard-required globally.
+
+If repository policy requires a capability and it is missing, the stage returns
+`BLOCKED`. If no skill applies, the stage infers conventions from repository instructions,
+code, tests, scripts, and analogous changes.
+
+## Retry semantics and human gates
+
+Engineering retries are bounded and hypothesis-driven. Each attempt retains exact failure
+evidence and attempted approaches. Product CI may be observed indefinitely by repeated
+fresh Taste invocations. Outer loops use provider-native watch hooks or harness
+subscriptions during each bounded wait and timers only as fallback. Brew and Taste wait
+internally only for async review bots.
+
+Human gates include product semantics, architecture/security/data ownership, destructive
+or irreversible operations, permissions, public compatibility, major scope expansion,
+and policy-required approval. `BLOCKED` names the decision, evidence, options/tradeoffs,
+and recommended default.
+
+## Compatibility migration
+
+| Retired | Canonical |
+| --- | --- |
+| `terreno-1-blend` | `terreno-1-grow` |
+| `terreno-2-roast` (implementation) | `terreno-2-pick` |
+| `terreno-3-cupping` | `terreno-3-roast` (verification) |
+| `terreno-4-pour` | `terreno-4-brew` |
+| `terreno-5-dialin` | `terreno-5-taste` |
+
+Former Grind orchestration becomes an outer-loop feature profile. Former plugin code
+review becomes a shared reference. Neither remains a sixth plugin skill.
+
+## Acceptance criteria
+
+- Exactly five canonical lifecycle skill directories exist.
+- Every stage has explicit preconditions, inputs, procedure, supporting skills, evidence,
+  success/failure/blocked conditions, and next transition.
+- A fresh Pick/Roast/Taste invocation can reconstruct its work from durable artifacts.
+- Result/status/transition values validate against the common schema.
+- Taste has no internal indefinite wait/observation loop.
+- Brew does not require same-context Taste execution.
+- Lifecycle skills contain no Terreno package names/commands that belong to project skills.
+- Repository skills remain local and discoverable/composable.
+- Backend/API, UI, and multi-invocation CI/review scenarios are documented.
+- Canonical sources and rulesync-generated copies validate cleanly.
+
+## Remaining assumptions / weaknesses
+
+- The plugin defines state/results but does not ship a scheduler or durable remote state
+  service; the outer loop must preserve the fallback artifact between machines.
+- Some generated skill formats do not encode Cursor's `disable-model-invocation`
+  frontmatter. Enforcement of explicit invocation depends on that agent runtime.
+- Repository skill quality controls how much exact domain knowledge a stage can compose.
+  When no skill exists, agents still infer from instructions/code/tests.