agent-implement-a2ui

name	agent-implement-a2ui
description	Add A2UI (Agent-to-UI) surfaces to a Prisme.ai MCP workspace so that LLM agents can render interactive UI through MCP tool calls. Scaffolds an MCP endpoint (or extends an existing one) with tools whose outputs include a __surface payload (components, data_model, actions) that the host UI renders using the prisme://blocks/v1 catalog. Use when the user says "ajoute des surfaces A2UI", "expose une UI via MCP", "/agent-implement-a2ui <workspace> <surfaces>", or wants the agent to draw cards/forms/tables/buttons inside a chat.
argument-hint	[workspace-id-or-folder] [surfaces: card, form, table, action-card, feedback, confirmation, loader, custom...]
allowed-tools	Read, Write, Edit, Bash, Glob, Grep, AskUserQuestion, Agent, mcp__prisme-ai-builder__pull_workspace, mcp__prisme-ai-builder__push_workspace, mcp__prisme-ai-builder__validate_automation, mcp__prisme-ai-builder__get_prisme_documentation, mcp__prisme-ai-builder__search_workspaces, mcp__prisme-ai-builder__list_automations

A2UI — Agent-to-UI surfaces over MCP

You are wiring A2UI into a Prisme.ai workspace. A2UI lets an LLM agent display interactive UI (cards, forms, tables, approval flows…) by calling MCP tools whose responses carry a __surface block. The host UI (AI Store, embedded chat, custom client) reads __surface and renders components from the prisme://blocks/v1 catalog.

Reference implementation: workspace rTFdCox (slug perfect-wolverine-5, name a2ui-test-mcp) in sandbox. Pull it locally if you need to inspect:

mcp__prisme-ai-builder__pull_workspace(workspaceId: "rTFdCox", environment: "sandbox", path: "./rTFdCox-workspace")

---

Mental model

LLM agent ──tools/call──► MCP endpoint ──► tool automation
                              ▲                  │
                              │                  ▼
                              │            output = {
                              │              content: [{type:text, text:"…"}],   ← LLM reads this
                              │              __surface: {                         ← UI renders this
                              │                catalog_id: prisme://blocks/v1,
                              │                components: [...],
                              │                data_model: {...},
                              │                wait_for_action: bool,
                              │                send_data_model: bool
                              │              }
                              │            }
                              │
        user clicks Button ───┘  (returns { action.name, data_model } back to agent)

Key idea: one tool = one surface. The content.text is the LLM's narration ("Form displayed, waiting…"); the __surface is what the human sees. Both ship in the same MCP response.

What the agent actually sees (agent-factory _route-tool.yml): when a tool result contains __surface, the host streams it as an a2ui.create part, then replaces the tool result given to the LLM with __surface.text (not content[0].text). If wait_for_action: true, the run pauses with a surface_action pending approval; when the user acts, the agent resumes with a user message of the form:

[Form submitted: <surfaceId>]
Action: <action name>
Data: <JSON of the filled data_model>     ← only if send_data_model: true

The A2UI protocol (v0.9) also defines a2ui.update / a2ui.data / a2ui.delete lifecycle parts in the SSE stream, but only a2ui.create is bridged from MCP tool outputs today — an MCP tool cannot update or delete an existing surface; it re-creates one (reuse the same surface_id for client-side reconciliation).

---

Arguments

Parse $ARGUMENTS:

• First token — target workspace: a workspace ID (e.g. rTFdCox), a local folder name (e.g. my-workspace), or new <slug> to scaffold a fresh A2UI-only workspace.
• Rest — comma- or space-separated list of surface types to generate. Built-in keywords map to ready templates; anything else is treated as a custom surface (you'll design it).

Built-in surface keywords:

Keyword	Template	Interactive
card	templates/show_card.yml	no
table	templates/show_table.yml	no
form	templates/show_form.yml	yes (submit/cancel)
action-card	templates/show_action_card.yml	yes (approve/reject)
feedback	templates/show_feedback.yml	yes (submit/skip)
confirmation	templates/show_confirmation.yml	yes (yes/no)
loader	templates/show_loader.yml	auto (fires its action on timeout or poll completion)
demo	templates/show_component_demo.yml	yes (showcase)

If $ARGUMENTS is empty, ask the user with AskUserQuestion:

1. Target — Existing workspace ID / Existing local folder / New workspace
2. Surfaces — multiSelect: card, table, form, action-card, feedback, confirmation, loader, demo, custom

---

Phase 1 — Resolve target

1. Pull or detect the workspace.

   • If user gave a workspace ID, run pull_workspace to ./<id>-workspace.
   • If user gave a local folder, ls it; verify it has index.yml.
   • If new <slug>, scaffold the minimal layout described in Phase 5 — Fresh workspace scaffold.

2. Detect the existing MCP endpoint if any:

   • Look for automations/mcp.yml with when.endpoint: */mcp.
   • If present, you will extend it (add tool branches to its dispatch).
   • If absent, you will create one (see Phase 4 — MCP endpoint).

3. Detect mcpTools list location:

   • Most workspaces declare tools in index.yml > config.value.mcpTools (canonical: same place the host's tools/list reads from).
   • imports/MCP Core.yml > config.mcpTools if the workspace uses the MCP Core app — in that case BOTH places matter for some tenants. See feedback [[feedback-mcp-core-config-is-source]] in memory.

---

Phase 2 — Plan surfaces

For each surface keyword:

1. Pick a slug — show_<name> (e.g. show_card, show_invoice). Snake_case, no -.
2. Decide arguments — what the LLM will pass in tools/call.arguments. Keep it minimal: title, description, rows, items, defaults. Heavy structure (component tree) lives in the automation, not in arguments.
3. Decide interactivity:
   • wait_for_action: true + send_data_model: true → agent pauses, gets the action + filled data_model back when user clicks.
   • wait_for_action: false → read-only display; agent continues immediately.
   • Loader surfaces: wait_for_action: true + send_data_model: false — the action fires automatically on timeout/poll completion, no user click.
4. Plan the data_model shape — what the UI binds to via { path: /xxx }.

For custom surfaces, sketch the component tree before coding. The mental tool is: list of components, each with an id, referenced by parent's children: []. No imbricated tree.

---

Phase 3 — Generate tool automations

For each built-in surface: copy templates/show_<keyword>.yml into automations/ of the target workspace, then customize (defaults, fields, etc.) to fit the requested use case.

For custom surfaces: hand-write the automation following the rules in § A2UI cheat sheet below. The shape is always:

slug: show_xxx
name: show_xxx
description: <what this surface displays — written for the LLM>
private: true                         # NEVER public; only the MCP endpoint calls it
arguments:
  arguments:                          # nested by convention — matches body.params.arguments
    type: object
    description: Tool arguments from MCP call
do:
  - set: { name: _args, value: '{{arguments}}' }
  # …unpack, defaults, build _components & _data_model…
output:
  content:
    - type: text
      text: "<one-line narration for the LLM>"
  isError: false
  __surface:
    surface_id: <stable-id-or-slug>
    catalog_id: prisme://blocks/v1
    wait_for_action: <bool>
    send_data_model: <bool>           # omit if wait_for_action is false
    text: "<same one-line narration>"
    components: [ ... ]               # flat list, each with id + component + props
    data_model: { ... }               # initial values keyed by /paths used in components

Rules of thumb

• Flat components list. No nesting. Parents reference children by id in children: [...].
• Bindings: any prop that should be data-driven becomes prop: { path: /a/b/c } instead of a literal. path is a JSON pointer into data_model.
• Actions (Button.action): { name: <event>, context?: { ... } }. The client returns name + filled data_model + context to the agent.
• surface_id is for client-side reconciliation — keep it stable per tool name; vary the suffix if multiple instances coexist.
• Don't pre-fill data_model with computed strings inside content. If the value comes from arguments, bind via path so the UI re-renders on data changes.

---

A2UI cheat sheet

__surface shape

Field	Type	Required	Notes
surface_id	string	yes	stable per tool; suffix when multiple instances
catalog_id	string	yes	always prisme://blocks/v1 for now
wait_for_action	bool	yes	pauses the agent when true (surface_action pending approval)
send_data_model	bool	no	when true + wait_for_action, returns filled data_model
text	string	yes	this is the tool result the LLM sees (agent-factory replaces content[0].text with it when __surface is present) — mirror content[0].text
components	array	yes	flat list of {id, component, ...props, children?}
data_model	object	yes-ish	required when any prop uses {path} bindings

Component catalog prisme://blocks/v1

(Authoritative source: services/platform/src/lib/a2ui/catalog.ts in the prisme.ai repo — check it when in doubt, the catalog evolves.)

Layout: Column, Row, Card, Divider

• Column / Row: gap (Tailwind scale, default 2), justify (start/center/end/between/around), align (start/center/end/stretch), children
• Card: title?, variant? (default/outline/ghost/elevated), className? (Tailwind-ish: p-4, flex-1…), children
• Divider: orientation? (horizontal/vertical)
• ⚠️ Tabs and Modal exist in the catalog but render as a plain div (children pass through, no tab/modal behavior) — avoid them, use Card + Column instead.

Display: Text, Badge, Alert, Progress, Avatar, Image, List, Loader

• Text: content or text (string or {path}), variant (heading/body/caption/code), weight? (normal/medium/semibold/bold)
• Badge: label or text (string or {path}), variant (default/secondary/outline/destructive/warning)
• Alert: title?, message or content (string or {path} — NOT description, it is ignored), variant (default/destructive/success/warning/info/error)
• Progress: value (number or {path}), max (default 100)
• Avatar: src?, alt, fallback?
• Image: src, alt, width?, height?
• List: variant (bullet/number/plain), items: [{id, label}] — objects, NOT plain strings (strings render as empty bullets)
• Loader: see Loader below

Inputs: TextField, TextArea, Select, CheckBox, Switch, Slider — all accept an optional label rendered above/beside the input

• TextField: label?, placeholder, type? (text/email/password), binding via value: {path} or text: {path}
• TextArea: label?, placeholder, rows, binding via value: {path} or text: {path}
• Select: label?, placeholder, options ([{value, label}] or plain string[]), value: {path}
• CheckBox / Switch: label, binding via value: {path} or checked: {path}
• Slider: min, max, step, value: {path} (or literal for read-only demo)
• Inputs are auto-disabled when the surface is no longer interactive (action already consumed).

Actions: Button

• label, variant (primary/outline/ghost/destructive — default is outline), size?, disabled?, action
• action is a top-level component field accepted as a string (action: submit) or {name} object. ⚠️ context is no longer returned to the agent — encode any context in the action name or in data_model.

Data: DataTable

• columns: [{key, header}], data: {path} (binding to array), title?
• Renders nothing when data or columns is empty — pair with a Text fallback if emptiness is possible.

URL safety: Image.src, Avatar.src are sanitized — only http(s) and relative URLs render; data:, javascript: and protocol-relative //host URLs are silently dropped.

Loader (auto-completing waits)

Loader is an interactive component that fires its action automatically — no user click:

• Timeout mode: timeout (seconds) → dispatches action when elapsed.
• Poll mode (wins if both set): poll (URL) + interval? (seconds, min 1, default 30) → GETs the URL until the JSON body returns {done: true}. Body fields progress (0-100) and message live-update the display. After 3 consecutive failures it dispatches <action>_error instead.
• style: spinner (default), progress (shows a progress bar once a progress value arrives — poll mode only), or image (image-shaped placeholder: mosaic of tiles glowing in a diagonal wave + centered picture icon — use when the awaited result is an image). message? initial label (shown with a small spinner below the placeholder in image style).
• aspect (with style: image): square (default) | landscape | portrait — pass it straight from your job metadata; unknown values fall back to square.
• ⚠️ poll MUST be a same-origin relative path (^/(?!/) — single leading /). Absolute URLs and //host are rejected by the renderer. To poll a workspace webhook, the path must be reachable on the host UI's origin.
• The final poll response is merged into the returned data_model under __loader_<componentId> (errors under __loader_<componentId>_error).
• Use wait_for_action: true on the surface so the agent pauses until the loader completes.

Convention (reference workspace): one show_loader automation exposed as two MCP tools — show_loader_with_timeout and show_loader_with_polling — so the LLM picks the mode from the tool name instead of juggling mutually-exclusive args.

Bindings

Whenever a prop reads from the data model:

value: { path: /form/email }    # → reads/writes data_model.form.email
content: { path: /project/name }
content: { path: /project/name, default: "Untitled" }   # optional fallback

Initial values live in data_model:

data_model:
  form: { email: "", priority: "" }
  project: { name: "Alpha", progress: 73 }

JSON pointer rules: /foo/bar for nested, /list/0/key for array index (rarely needed since DataTable.data takes the array root).

Action result contract

When the user triggers a component whose action is set (or a Loader completes), the client emits:

{
  "name": "submit",
  "surfaceId": "interactive-form",
  "sourceComponentId": "submit-btn",
  "timestamp": "…",
  "dataModel": { "…filled values…" }     // {} when send_data_model: false
}

Agent-factory then resumes the LLM with a user message:

[Form submitted: <surfaceId>]
Action: <name>
Data: <JSON dataModel>        ← only when send_data_model: true

⚠️ There is no context field anymore — Button.action.context is dropped by the renderer. Disambiguate via distinct action names (approve / reject) or values seeded in data_model.

The agent decides next steps from this message. Design tool descriptions accordingly ("Pauses until user submits — returns the filled form").

---

Phase 4 — MCP endpoint

If the workspace has no automations/mcp.yml, create one. Choose an endpoint slug that fits the workspace (e.g. a2ui-test/mcp, <service>/mcp):

slug: mcp
name: MCP/endpoint
description: MCP JSON-RPC 2.0 endpoint for A2UI tool discovery and execution
when:
  endpoint: <endpoint-slug>/mcp
do:
  - conditions:
      '{{body.method}} == "initialize"':
        - set:
            name: output
            value:
              jsonrpc: '2.0'
              id: '{{body.id}}'
              result:
                protocolVersion: '2025-06-18'
                capabilities: { tools: {} }
                serverInfo: { name: <workspace-name>, version: 1.0.0 }
      '{{body.method}} == "notifications/initialized"':
        - set: { name: $http, value: { status: 200 } }
        - set: { name: output, value: '' }
      '{{body.method}} == "tools/list"':
        - set:
            name: output
            value:
              jsonrpc: '2.0'
              id: '{{body.id}}'
              result:
                tools: '{{config.mcpTools}}'
      '{{body.method}} == "tools/call"':
        - conditions:
            '{{body.params.name}} == "show_<surface1>"':
              - show_<surface1>: { arguments: '{{body.params.arguments}}', output: toolResult }
            # …one branch per surface…
            default:
              - set:
                  name: output
                  value:
                    jsonrpc: '2.0'
                    id: '{{body.id}}'
                    error:
                      code: -32601
                      message: 'Tool not found: {{body.params.name}}'
              - break: {}
        - set:
            name: output
            value:
              jsonrpc: '2.0'
              id: '{{body.id}}'
              result: '{{toolResult}}'
      default:
        - set:
            name: output
            value:
              jsonrpc: '2.0'
              id: '{{body.id}}'
              error:
                code: -32601
                message: 'Method not found: {{body.method}}'
output: '{{output}}'

If the endpoint exists, ADD one branch per new surface under the tools/call conditions block. Do not duplicate the surrounding scaffolding.

---

Phase 5 — mcpTools declaration

In index.yml, under config.value.mcpTools, append an entry per surface:

- name: show_<surface>
  description: <LLM-facing description — say WHAT it displays + whether it pauses>
  inputSchema:
    type: object
    properties:
      <arg1>:
        type: string
        description: <…>
    required: []

Description rules (the LLM reads these to decide when to call):

• Lead with the visible outcome ("Display a project status card…").
• For interactive surfaces, mention pause: "The conversation pauses until user submits."
• For arrays, give shape hint: 'Custom field definitions. Each: { id, component, placeholder, type? }. Components: TextField, TextArea, Select, CheckBox.'
• Per memory [[feedback-mcp-tool-array-items]], any type: array MUST have items: { type: object | string } — OpenAI rejects otherwise.
• One automation may back several tools: when args are mutually exclusive modes (e.g. timeout vs poll), declare one tool per mode (show_loader_with_timeout, show_loader_with_polling) routed to the same automation in mcp.yml — the LLM picks by name, no conditional-args confusion.

---

Phase 6 — Fresh workspace scaffold (only if new <slug>)

Minimal file layout:

<slug>-workspace/
├── .import.yml
├── index.yml          # config.value.mcpTools list
├── security.yml       # standard (copy from rTFdCox)
└── automations/
    ├── mcp.yml        # endpoint
    └── show_<…>.yml   # one per surface

• id in index.yml is assigned by Prisme.ai on push — leave it as a placeholder name or let push_workspace resolve it.
• Add labels: [app-builder, one-product, mcp].
• security.yml: copy verbatim from rTFdCox-workspace/security.yml.

---

Phase 7 — Validate and push

For each generated/edited automation, run:

mcp__prisme-ai-builder__validate_automation(path: "<workspace>/automations/show_xxx.yml")

Trust this over example patterns — see project CLAUDE.md.

Then push:

mcp__prisme-ai-builder__push_workspace(path: "<workspace>", workspaceId: "<id>", environment: "sandbox")

After push, give the user:

• The MCP endpoint URL: https://api.<env>.prisme.ai/v2/workspaces/<id>/webhooks/<endpoint-slug>/mcp
• A list of tool names exposed
• Suggested test prompt for the LLM (e.g. "Show me a project status card for Project Beta at 45%")

---

Phase 8 — Review

Spawn a sub-task with Agent (subagent_type: code-review or general-purpose) to review the generated automations. Format issues as 🔴 MAJOR | 🟠 NEED_HUMAN bullets, per project CLAUDE.md. Fix MAJORs immediately; ask for the rest.

---

Common pitfalls

• Forgetting private: true on show_* automations → they become callable as public endpoints, bypassing the MCP envelope. Per [[feedback-private-does-not-block-webhook]], private: true does NOT block webhooks but it hides them from App instructions — still set it.
• Putting JS in {{}} — DSUL refuses. Use {% %} for math; otherwise unpack into intermediate set: steps. See memory.
• Naming a var output inside a do: step — collides with the call-site capture variable; rename to _output or similar. Memory: [[feedback-dsul-output-variable]].
• type: array without items: in inputSchema → silent break at LLM call time.
• Editing imports/MCP Core.yml > config.mcpTools only — the canonical source for THIS pattern is index.yml > config.value.mcpTools (the endpoint reads {{config.mcpTools}}). Memory: [[feedback-mcp-core-config-is-source]].
• Using {path} without initial value in data_model → the host may render empty/null. Always seed bound props in data_model, even with "" or false (or use {path, default}).
• Forgetting the text field inside __surface — it is the tool result the LLM reads (agent-factory swaps it in) AND the text-only fallback. Mirror content[0].text.
• List.items as plain strings → empty bullets. Items must be [{id, label}] objects.
• Alert.description is ignored by the renderer — use message (or content).
• Relying on Button.action.context — not returned anymore; use distinct action names or data_model.
• Tabs / Modal render as a bare div (no behavior) — don't use them.
• Loader.poll with an absolute URL → rejected client-side; only same-origin relative paths (/...) are allowed.
• data: URIs in Image.src/Avatar.src → silently dropped by URL sanitization; only http(s)/relative.
• Unknown component types are skipped with a console warning (typos render nothing — check casing: TextField, CheckBox, DataTable).

---

After running

End-of-turn summary: list surfaces added, endpoint URL, and the next test step. Do not write planning docs unless the user asks.