feat: move mfb provider from azure to pair foundry

.env.example

@@ -117,17 +117,17 @@ FORMSG_SDK_MODE=
 # POSTMAN_INTERNAL_CAMPAIGN_API_KEY=
 # POSTMAN_BASE_URL=https://test.postman.gov.sg/api/v2
 
-## Azure OpenAI
-# AZURE_OPENAI_API_KEY=
-# AZURE_OPENAI_ENDPOINT=
-# AZURE_OPENAI_DEPLOYMENT_NAME=
-# AZURE_OPENAI_API_VERSION=
-
 ## Kill email mode configs, provide a valid storage form id
 # KILL_EMAIL_MODE_FEEDBACK_FORMID=
 
-# For WOG AD Login 
+# For WOG AD Login
 # WOGAD_AUTHORITY=
 # WOGAD_CLIENT_ID=
 # WOGAD_CLIENT_SECRET=
-# WOGAD_REDIRECT_URI=
+# WOGAD_REDIRECT_URI=
+
+## AI SDK
+# AI_SDK_PROVIDER_NAME=
+# AI_SDK_BASE_URL=
+# AI_SDK_API_KEY=
+# AI_SDK_MODEL_NAME=

.scratch/migrate-pair-foundry/01-ai-sdk-wrapper-migration.md

@@ -0,0 +1,44 @@
+# Migrate `ai-model` wrapper + both MFB flows to Pair Foundry via Vercel AI SDK
+
+**Type:** AFK
+**Triage:** ready-for-agent
+
+## Parent
+
+PRD: `.scratch/migrate-pair-foundry/PRD.md`
+ADR: `docs/adr/0001-pair-foundry-llm-provider.md`
+
+## What to build
+
+Replace the Azure-OpenAI-backed Magic Form Builder (MFB) LLM client with a Vercel AI SDK client pointed at Pair Foundry's PX Engine (`engine.pair.gov.sg`). The single application-facing seam — `sendPromptToModel({ messages, options, formId })` — keeps its function signature, but its internals are fully rewritten on top of ai-sdk's `generateText` (using `@ai-sdk/openai`'s `createOpenAI` provider). The `Message` and options types exposed to callers become ai-sdk native (`CoreMessage` / `ModelMessage`).
+
+Both MFB flows (text-prompt and vision-prompt) are switched over together because changing the exported `Message` type forces both callers to update in the same change. The vision flow rewrites its content parts from the OpenAI shape (`{ type: 'image_url', image_url: { url } }`) to ai-sdk's (`{ type: 'image', image: <dataURL> }`). JSON-mode is preserved per-call via `providerOptions.openai.responseFormat`, so downstream `JSON.parse` + zod validation in the assistance service is unchanged.
+
+Connection settings (`providerName`, `apiKey`, `baseUrl`, `modelName`) are read from the existing `aisdkConfig` (env-var-driven, SSM-supplied in prod). The default model is `claude-x` — verified to support both image content parts and JSON mode on PX Engine — and serves both text and vision flows.
+
+Errors from ai-sdk are mapped onto the existing `ModelGetClientFailureError` (provider/client construction failure) and `ModelResponseFailureError` (request failure or empty response) so that error-handling code paths in MFB controllers and the assistance service remain untouched. A null return for missing message content stays the contract for "model responded but produced nothing usable."
+
+New unit tests for the wrapper cover message/options forwarding, JSON-mode plumbing, error mapping, and null-response handling. The wrapper mocks ai-sdk at the `generateText` boundary, not deeper. Existing `admin-form.assistance.service` tests continue to mock `sendPromptToModel` directly and remain valid through the swap.
+
+The PR is structured as conventional commits (one logical step per commit: dependency add, wrapper rewrite, prompt rewrite, wrapper tests) so reviewers can step through it commit-by-commit, per the ADR's implementation considerations.
+
+## Acceptance criteria
+
+- [ ] `ai` package added to `apps/backend/package.json` and pnpm lockfile; `@ai-sdk/openai` already present.
+- [ ] `ai-model.ts` no longer imports anything from the `openai` npm package; uses `createOpenAI` from `@ai-sdk/openai` plus ai-sdk's `generateText`.
+- [ ] `ai-model.ts` reads `providerName`, `apiKey`, `baseUrl`, `modelName` from `aisdkConfig`; passes `providerName` as the provider `name`, `baseUrl` as `baseURL`, `apiKey` as `apiKey`, and `.chat(modelName)` selects the model.
+- [ ] `sendPromptToModel({ messages, options, formId })` retains its three named parameters and its `ResultAsync<string | null, ModelGetClientFailureError | ModelResponseFailureError>` return type.
+- [ ] `Message` re-exported by `ai-model.ts` is ai-sdk's `CoreMessage` (or `ModelMessage`); the `Role` enum is removed; callers in `admin-form.assistance.service.ts` are updated accordingly.
+- [ ] JSON-mode (`response_format: { type: 'json_object' }`) is plumbed via `providerOptions: { openai: { responseFormat: { type: 'json_object' } } }` on each call from the text and vision prompt flows.
+- [ ] Vision-flow content parts use ai-sdk's `{ type: 'image', image: <dataURL> }` shape; the `image_url` shape no longer appears in the prompt builders.
+- [ ] ai-sdk client-construction failures surface as `ModelGetClientFailureError`; ai-sdk request failures surface as `ModelResponseFailureError`; empty/missing model responses return `null`.
+- [ ] `temperature` is not pinned inside the wrapper; callers can still pass it via `options`.
+- [ ] New unit tests added for the `ai-model` wrapper covering: happy-path forwarding of messages and `providerOptions.openai.responseFormat`; null/empty response → `null`; provider construction failure → `ModelGetClientFailureError`; ai-sdk request failure → `ModelResponseFailureError` with `formId` in the log meta; options pass-through (caller options layered correctly with the wrapper's own).
+- [ ] Existing `admin-form.assistance.service.spec.ts` continues to pass without modification to its mocks.
+- [ ] Backend type-checks, lints, and tests pass.
+- [ ] Manual smoke: text-prompt MFB generates expected JSON form fields end-to-end against PX Engine; vision-prompt MFB does the same with an uploaded image.
+- [ ] Commits follow conventional commits and are reviewable in sequence.
+
+## Blocked by
+
+None — can start immediately.

.scratch/migrate-pair-foundry/02-azure-cleanup.md

@@ -0,0 +1,32 @@
+# Remove orphaned Azure AI Foundry artifacts
+
+**Type:** AFK
+**Triage:** ready-for-agent
+
+## Parent
+
+PRD: `.scratch/migrate-pair-foundry/PRD.md`
+ADR: `docs/adr/0001-pair-foundry-llm-provider.md`
+
+## What to build
+
+Once the Magic Form Builder LLM wrapper is fully migrated to Pair Foundry (slice 1), the Azure AI Foundry remnants are dead weight: the Azure-specific convict config, the `openai` npm package, and `AZURE_OPENAI_*` env-var references in the repo. This slice deletes them so SSM / IaC stops carrying ghost parameters and the lockfile stops carrying an unused dependency.
+
+The scope is in-repo only. If FormSG's deployment-side IaC lives in a separate repository, those `AZURE_OPENAI_*` parameter definitions are out of scope for this slice and handled in a follow-up PR against that repo.
+
+After this slice the only LLM config surface in the codebase is `aisdk.config.ts` and the `AI_SDK_*` env vars.
+
+## Acceptance criteria
+
+- [ ] `apps/backend/src/app/config/features/azureopenai.config.ts` is deleted.
+- [ ] `openai` is removed from `apps/backend/package.json` dependencies; pnpm lockfile is regenerated and no longer contains the `openai` package.
+- [ ] Repo-wide grep confirms no remaining imports from `'openai'`, `'openai/error'`, or `'openai/resources/...'` in `apps/backend/src`.
+- [ ] Repo-wide grep confirms no remaining references to `AZURE_OPENAI_API_KEY`, `AZURE_OPENAI_ENDPOINT`, `AZURE_OPENAI_DEPLOYMENT_NAME`, `AZURE_OPENAI_API_VERSION`, or `AZURE_OPENAI_MODEL` in `.env` examples, docs, or in-repo IaC.
+- [ ] Repo-wide grep confirms no remaining references to `azureOpenAIConfig` or `azureopenai.config`.
+- [ ] Backend type-checks, lints, and tests pass.
+- [ ] Manual smoke: text-prompt MFB and vision-prompt MFB continue to work end-to-end against PX Engine (regression check after dep removal).
+- [ ] Commits follow conventional commits.
+
+## Blocked by
+
+- Slice 1: `.scratch/migrate-pair-foundry/01-ai-sdk-wrapper-migration.md` — the Azure config and `openai` package can only be safely removed once nothing imports them.