@k2b/cloud 0.42.0 → 0.44.0

package/src/ai/store.ts

@@ -279,6 +279,8 @@ type TurnSteerRow = {
 };
 
 type ConversationResourceRefRow = {
+  context: string | null;
+  status: string | null;
   resource_type: string;
   resource_id: string;
   title: string | null;
@@ -298,6 +300,8 @@ type ConversationResourceOccurrenceRow = ConversationResourceRefRow & {
 };
 
 type ConversationSourceRow = {
+  context: string | null;
+  status: string | null;
   source_kind: AiConversationSource["kind"];
   source_key: string;
   title: string;
@@ -340,6 +344,7 @@ type InterChatMessageRow = {
 };
 
 type TimelineRow = {
+  model_replacement: { kind: "model_replaced"; previousLabel: string; nextLabel: string } | null;
   id: string;
   seq: number;
   loop_id: string | null;
@@ -356,6 +361,8 @@ const iso = (value: Date | string): string => (value instanceof Date ? value.toI
 
 const rowToConversationResourceRef = (row: ConversationResourceRefRow): AiConversationResourceRef => ({
   ref: { type: row.resource_type, id: row.resource_id },
+  context: row.context,
+  status: row.status,
   title: row.title,
   preview: row.preview,
   icon: row.icon,
@@ -397,6 +404,8 @@ const rowToConversationResourceOccurrence = (row: ConversationResourceOccurrence
 
 const rowToConversationSource = (row: ConversationSourceRow): AiConversationSource => ({
   kind: row.source_kind,
+  context: row.context,
+  status: row.status,
   key: row.source_key,
   title: row.title,
   preview: row.preview,
@@ -649,7 +658,7 @@ const rowToEnrichmentRun = (row: EnrichmentRunRow): AiEnrichmentRun => ({
   conversationId: row.conversation_id,
   status: row.status,
   trigger: row.trigger,
-  modelProfileId: row.model_profile_id,
+  modelId: row.model_profile_id,
   mode: row.mode,
   durationMs: row.duration_ms === null ? null : Number(row.duration_ms),
   titleUpdated: row.title_updated,
@@ -689,7 +698,7 @@ const rowToMessage = (row: MessageRow): AiStoredMessage => {
     kind: row.kind,
     message,
     loopId: row.loop_id,
-    modelProfileId: row.model_profile_id,
+    modelId: row.model_profile_id,
     providerModel: row.provider_model,
     usage: row.usage ? parseJsonValue<Usage>(row.usage) : null,
     stopReason: row.stop_reason,
@@ -719,7 +728,7 @@ const rowToTurn = (row: TurnRow): AiTurn => ({
   conversationId: row.conversation_id,
   status: row.status,
   attempt: Number(row.attempt ?? 0),
-  modelProfileId: row.model_profile_id,
+  modelId: row.model_profile_id,
   createdAt: iso(row.created_at),
   completedAt: row.completed_at ? iso(row.completed_at) : null,
   error: row.error,
@@ -1140,7 +1149,7 @@ const insertMessageLocked = async (
     kind?: "message" | "summary";
     seq?: number;
     loopId?: string | null;
-    modelProfileId?: string | null;
+    modelId?: string | null;
     meta?: AiStoredMessage["meta"];
   },
   db: typeof sql = sql,
@@ -1181,7 +1190,7 @@ const insertMessageLocked = async (
       ${JSON.stringify(input.message)}::text::jsonb,
       ${messageSearchText(input.message)},
       ${input.loopId ?? null},
-      ${input.modelProfileId ?? null},
+      ${input.modelId ?? null},
       ${providerModel},
       ${usage ? JSON.stringify(usage) : null}::text::jsonb,
       ${stopReason},
@@ -1217,7 +1226,7 @@ const appendTurnOwnedMessage = async (input: {
   kind?: "message" | "summary";
   seq?: number;
   loopId: string | null;
-  modelProfileId?: string | null;
+  modelId?: string | null;
   meta?: AiStoredMessage["meta"];
 }): Promise<AiStoredMessage | null> => {
   return sql.begin(async (tx) => {
@@ -1252,7 +1261,7 @@ const appendTurnOwnedMessage = async (input: {
       ${input.message.role},
       ${JSON.stringify(input.message)}::text::jsonb,
       ${input.loopId},
-      ${input.modelProfileId ?? null},
+      ${input.modelId ?? null},
       ${providerModel},
       ${usage ? JSON.stringify(usage) : null}::text::jsonb,
       ${stopReason},
@@ -1777,15 +1786,17 @@ export const aiConversations: AiConversationService = {
       for (const resource of resources) {
         await tx`
           INSERT INTO ai.conversation_resource_refs (
-            conversation_id, resource_type, resource_id, title, preview, icon, href, source_turn_id, source_call_id
+            conversation_id, resource_type, resource_id, title, preview, icon, href, context, status, source_turn_id, source_call_id
           ) VALUES (
             ${input.conversationId}, ${resource.ref.type}, ${resource.ref.id}, ${resource.title ?? null}, ${resource.preview ?? null},
-            ${resource.icon ?? null}, ${resource.href ?? null}, ${input.turnId ?? null}, ${input.callId ?? null}
+            ${resource.icon ?? null}, ${resource.href ?? null}, ${resource.context ?? null}, ${resource.status ?? null}, ${input.turnId ?? null}, ${input.callId ?? null}
           )
           ON CONFLICT (conversation_id, resource_type, resource_id)
           DO UPDATE SET
             title = COALESCE(EXCLUDED.title, ai.conversation_resource_refs.title),
             preview = COALESCE(EXCLUDED.preview, ai.conversation_resource_refs.preview),
+            context = COALESCE(EXCLUDED.context, ai.conversation_resource_refs.context),
+            status = COALESCE(EXCLUDED.status, ai.conversation_resource_refs.status),
             icon = COALESCE(EXCLUDED.icon, ai.conversation_resource_refs.icon),
             href = COALESCE(EXCLUDED.href, ai.conversation_resource_refs.href),
             -- Project context is indexed on every turn without a call. It must not move a resource a tool call
@@ -1811,7 +1822,7 @@ export const aiConversations: AiConversationService = {
     const pattern = searchPattern(input.search);
     const cursor = decodeResourceCursor(input.before);
     const rows = await sql<(ConversationResourceRefRow & { cursor_at: string })[]>`
-      SELECT indexed.resource_type, indexed.resource_id, indexed.title, indexed.preview, indexed.icon, indexed.href,
+      SELECT indexed.resource_type, indexed.resource_id, indexed.title, indexed.preview, indexed.icon, indexed.href, indexed.context, indexed.status,
              source_turn.short_id AS source_turn_id, indexed.source_call_id, indexed.first_seen_at, indexed.last_seen_at,
              to_char(indexed.last_seen_at AT TIME ZONE 'UTC', 'YYYY-MM-DD"T"HH24:MI:SS.US"Z"') AS cursor_at
       FROM ai.conversation_resource_refs indexed
@@ -1847,7 +1858,7 @@ export const aiConversations: AiConversationService = {
     const pattern = searchPattern(input.search);
     const cursor = decodeResourceOccurrenceCursor(input.before);
     const rows = await sql<(ConversationResourceOccurrenceRow & { cursor_at: string })[]>`
-      SELECT indexed.resource_type, indexed.resource_id, indexed.title, indexed.preview, indexed.icon, indexed.href,
+      SELECT indexed.resource_type, indexed.resource_id, indexed.title, indexed.preview, indexed.icon, indexed.href, indexed.context, indexed.status,
              source_turn.short_id AS source_turn_id, indexed.source_call_id, indexed.first_seen_at, indexed.last_seen_at,
              conversation.short_id AS conversation_short_id, conversation.title AS conversation_title,
              conversation.updated_at AS conversation_updated_at,
@@ -1908,6 +1919,7 @@ export const aiConversations: AiConversationService = {
         SELECT *
         FROM (
           SELECT source.kind AS source_kind, source.source_key, source.title, source.preview,
+                 NULL::text AS context, NULL::text AS status,
                  source.icon, source.href, file.path, file.media_type, file.size,
                  source.resource_type, source.resource_id, source.occurrences,
                  turn.short_id AS source_turn_id, source.source_turn_id AS turn_uuid, source.source_call_id,
@@ -1922,6 +1934,7 @@ export const aiConversations: AiConversationService = {
 
           SELECT 'file' AS source_kind, file.path AS source_key,
                  regexp_replace(file.path, '^.*/', '') AS title, 'Attached to the conversation' AS preview,
+                 NULL::text AS context, NULL::text AS status,
                  CASE WHEN file.media_type LIKE 'image/%' THEN 'ti ti-photo' ELSE 'ti ti-file' END AS icon,
                  NULL::text AS href, file.path, file.media_type, file.size,
                  NULL::text AS resource_type, NULL::text AS resource_id, 1 AS occurrences,
@@ -1934,7 +1947,7 @@ export const aiConversations: AiConversationService = {
 
           SELECT 'resource' AS source_kind, indexed.resource_type || ':' || indexed.resource_id AS source_key,
                  COALESCE(indexed.title, indexed.resource_type || ' ' || indexed.resource_id) AS title,
-                 indexed.preview, indexed.icon, indexed.href, NULL::text AS path, NULL::text AS media_type, NULL::bigint AS size,
+                 indexed.preview, indexed.context, indexed.status, indexed.icon, indexed.href, NULL::text AS path, NULL::text AS media_type, NULL::bigint AS size,
                  indexed.resource_type, indexed.resource_id, 1 AS occurrences,
                  turn.short_id AS source_turn_id, indexed.source_turn_id AS turn_uuid, indexed.source_call_id,
                  indexed.first_seen_at, indexed.last_seen_at
@@ -2135,7 +2148,7 @@ export const aiConversations: AiConversationService = {
         "idx_ai_turns_conversation_short_id",
         (attempt, shortId) => attempt<TurnRow[]>`
           INSERT INTO ai.turns (short_id, conversation_id, model_profile_id, status, run_config)
-          VALUES (${shortId}, ${message.target_conversation_id}, ${input.modelProfileId}, 'queued', (${JSON.stringify(input.runConfig)}::text)::jsonb)
+          VALUES (${shortId}, ${message.target_conversation_id}, ${input.modelId}, 'queued', (${JSON.stringify(input.runConfig)}::text)::jsonb)
           RETURNING *
         `,
       );
@@ -2434,7 +2447,7 @@ export const aiConversations: AiConversationService = {
         ${input.conversationId},
         ${input.status},
         ${input.trigger},
-        ${input.modelProfileId ?? null},
+        ${input.modelId ?? null},
         ${input.mode ?? null},
         ${input.durationMs ?? null},
         ${input.titleUpdated ?? false},
@@ -2691,6 +2704,7 @@ export const aiConversations: AiConversationService = {
       SELECT
         users.short_id AS id,
         users.seq,
+        users.meta_payload->'modelReplacement' AS model_replacement,
         COALESCE(timeline_turn.short_id, users.loop_id) AS loop_id,
         left(
           regexp_replace(
@@ -2722,6 +2736,7 @@ export const aiConversations: AiConversationService = {
       id: row.id,
       seq: Number(row.seq),
       loopId: row.loop_id,
+      ...(row.model_replacement ? { modelReplacement: row.model_replacement } : {}),
       userPreview: row.user_preview?.trim() || "Message",
       assistantPreview: row.assistant_preview?.trim() || "",
       isSteer: Boolean(row.is_steer),
@@ -2939,7 +2954,7 @@ export const aiConversations: AiConversationService = {
           kind: "summary",
           seq: checkpointSeq,
           loopId: null,
-          modelProfileId: input.modelProfileId ?? null,
+          modelId: input.modelId ?? null,
           meta: { compactedCount: archived[0]?.count ?? 0, ...(todoPlan ? { todoPlan } : {}) },
         },
         tx,
@@ -2958,7 +2973,7 @@ export const aiConversations: AiConversationService = {
         "idx_ai_turns_conversation_short_id",
         (attempt, shortId) => attempt<TurnRow[]>`
       INSERT INTO ai.turns (short_id, conversation_id, model_profile_id, status, run_config)
-      VALUES (${shortId}, ${input.conversationId}, ${input.modelProfileId}, 'queued',
+      VALUES (${shortId}, ${input.conversationId}, ${input.modelId}, 'queued',
         (${input.runConfig ? JSON.stringify(input.runConfig) : null}::text)::jsonb)
       RETURNING *
     `,
@@ -3015,7 +3030,7 @@ export const aiConversations: AiConversationService = {
         VALUES (
           ${shortId},
           ${input.conversationId},
-          ${input.modelProfileId},
+          ${input.modelId},
           'queued',
           (${JSON.stringify(input.runConfig)}::text)::jsonb
         )
@@ -3064,7 +3079,12 @@ export const aiConversations: AiConversationService = {
           conversationId: input.conversationId,
           message: input.userMessage,
           loopId: turn.id,
-          meta: input.expectedDraftRevision === undefined ? undefined : { submittedDraftRevision: input.expectedDraftRevision },
+          meta: {
+            ...(input.expectedDraftRevision === undefined ? {} : { submittedDraftRevision: input.expectedDraftRevision }),
+            modelRequest: input.runConfig.model,
+            modelResolved: input.runConfig.modelResolved,
+            ...(input.runConfig.modelReplacement ? { modelReplacement: input.runConfig.modelReplacement } : {}),
+          },
         },
         tx,
       );
@@ -3074,15 +3094,17 @@ export const aiConversations: AiConversationService = {
       for (const resource of resources) {
         await tx`
           INSERT INTO ai.conversation_resource_refs (
-            conversation_id, resource_type, resource_id, title, preview, icon, href, source_turn_id
+            conversation_id, resource_type, resource_id, title, preview, icon, href, context, status, source_turn_id
           ) VALUES (
             ${input.conversationId}, ${resource.ref.type}, ${resource.ref.id}, ${resource.title ?? null},
-            ${resource.preview ?? null}, ${resource.icon ?? null}, ${resource.href ?? null}, ${turn.id}
+            ${resource.preview ?? null}, ${resource.icon ?? null}, ${resource.href ?? null}, ${resource.context ?? null}, ${resource.status ?? null}, ${turn.id}
           )
           ON CONFLICT (conversation_id, resource_type, resource_id)
           DO UPDATE SET
             title = COALESCE(EXCLUDED.title, ai.conversation_resource_refs.title),
             preview = COALESCE(EXCLUDED.preview, ai.conversation_resource_refs.preview),
+            context = COALESCE(EXCLUDED.context, ai.conversation_resource_refs.context),
+            status = COALESCE(EXCLUDED.status, ai.conversation_resource_refs.status),
             icon = COALESCE(EXCLUDED.icon, ai.conversation_resource_refs.icon),
             href = COALESCE(EXCLUDED.href, ai.conversation_resource_refs.href),
             source_turn_id = EXCLUDED.source_turn_id,
@@ -3120,7 +3142,7 @@ export const aiConversations: AiConversationService = {
   getLatestTurn: async (input) => {
     const [row] = await sql<
       TurnRow[]
-    >`SELECT * FROM ai.turns WHERE conversation_id = ${input.conversationId}::uuid AND NOT COALESCE(run_config ? 'background', false)
+    >`SELECT * FROM ai.turns WHERE conversation_id = ${input.conversationId}::uuid AND NOT COALESCE(run_config ? 'background', false) AND (${input.kind ?? null}::text IS NULL OR run_config->>'kind'=${input.kind ?? null})
       ORDER BY created_at DESC, id DESC LIMIT 1`;
     return row ? rowToTurn(row) : null;
   },
@@ -3978,7 +4000,7 @@ export const aiConversations: AiConversationService = {
           kind: opts?.kind,
           seq: opts?.seq,
           loopId: opts?.kind === "summary" && !input.background ? null : input.turnId,
-          modelProfileId: input.modelProfileId,
+          modelId: input.modelId,
           meta,
         });
         if (!appended) {
@@ -3997,7 +4019,7 @@ export const aiConversations: AiConversationService = {
             kind: opts?.kind,
             seq: opts?.seq,
             loopId: input.turnId && opts?.kind !== "summary" ? input.turnId : null,
-            modelProfileId: input.modelProfileId,
+            modelId: input.modelId,
             meta,
           },
           tx,

package/src/ai/stream.ts

@@ -134,7 +134,7 @@ const turnSnapshotFromActive = (active: NonNullable<Awaited<ReturnType<typeof ai
   status: active.turn.status,
   seq: active.liveSeq,
   blocks: [...active.liveBlocks],
-  modelProfileId: active.turn.modelProfileId,
+  modelId: active.turn.modelId,
   createdAt: active.turn.createdAt,
   actionWaitMs: active.actionWaitMs,
   waitingSince: active.waitingSince,

package/src/ai/structured-runs.ts

@@ -15,7 +15,7 @@ export type AiStructuredRunRecord = {
   traceId?: string;
   task: string;
   appId?: string;
-  modelProfileId?: string;
+  modelId?: string;
   providerModel?: string;
   status: "ok" | "failed";
   durationMs: number;
@@ -40,7 +40,7 @@ export const recordAiStructuredRun = async (record: AiStructuredRunRecord): Prom
       input_tokens, output_tokens, total_tokens,
       mode, repaired, attempts, error_code, error, user_id, conversation_id, turn_id, workflow_run_id, trace_id
     ) VALUES (
-      ${record.task}, ${record.appId ?? null}, ${record.modelProfileId ?? null}, ${record.providerModel ?? null},
+      ${record.task}, ${record.appId ?? null}, ${record.modelId ?? null}, ${record.providerModel ?? null},
       ${record.status}, ${Math.max(0, Math.round(record.durationMs))},
       ${usageNumber(record.usage, "input")}, ${usageNumber(record.usage, "output")},
       ${usageNumber(record.usage, "total")},

package/src/ai/structured.ts

@@ -2,49 +2,20 @@ import type { Input, LoopAggregate, StructuredMeta, Usage } from "@k2b/nessi";
 import { nessi, StructuredOutputError } from "@k2b/nessi";
 import type { z } from "zod";
 import type { AccessSubject } from "../server";
-import { coreSettings } from "../services";
 import type { TraceContext } from "../services/logging";
 import { trace } from "../services/logging";
+import type { AiModelRequest, AiResolvedModelRef } from "../shared/ai-models";
 import { taskReasoningEffort } from "./provider";
+import { AiProviderError } from "./provider-errors";
 import { inferenceProvider } from "./quota-provider";
 import { resolveAiModel } from "./settings";
 import { type AiUsageAttribution, safelyRecordStructuredRun } from "./structured-runs";
 import type { AiResolvedModel } from "./types";
 
-export const AI_BACKGROUND_MODEL_SETTING_KEY = "ai.background_model_id";
-export const AI_WORKFLOW_MODEL_SETTING_KEY = "ai.workflow_model_id";
-
-export const selectAiWorkflowModelId = (input: {
-  requestedModelId?: string;
-  workflowModelId?: string;
-  backgroundModelId?: string;
-}): string | undefined => input.requestedModelId?.trim() || input.workflowModelId?.trim() || input.backgroundModelId?.trim() || undefined;
-
-/**
- * Resolve the model for background inference: explicit request →
- * `ai.background_model_id` setting → platform default. Throws when AI is
- * disabled or the model is unavailable — callers skip their work then.
- */
-export const resolveAiBackgroundModel = async (requestedModelId?: string): Promise<AiResolvedModel> => {
-  const backgroundModelId = String((await coreSettings.get<string>(AI_BACKGROUND_MODEL_SETTING_KEY)) ?? "").trim();
-  return resolveAiModel({ kind: "selectable" }, requestedModelId?.trim() || backgroundModelId || undefined);
-};
-
-/** Resolve a workflow model: action override → workflow setting → background setting → platform default. */
-export const resolveAiWorkflowModel = async (requestedModelId?: string): Promise<AiResolvedModel> => {
-  const [workflowModelId, backgroundModelId] = await Promise.all([
-    coreSettings.get<string>(AI_WORKFLOW_MODEL_SETTING_KEY),
-    coreSettings.get<string>(AI_BACKGROUND_MODEL_SETTING_KEY),
-  ]);
-  return resolveAiModel(
-    { kind: "selectable" },
-    selectAiWorkflowModelId({
-      requestedModelId,
-      workflowModelId: String(workflowModelId ?? ""),
-      backgroundModelId: String(backgroundModelId ?? ""),
-    }),
-  );
-};
+export const resolveAiBackgroundModel = (model: AiModelRequest = { class: "fast" }): Promise<AiResolvedModel> =>
+  resolveAiModel({ kind: "selectable" }, model);
+export const resolveAiWorkflowModel = (model: AiModelRequest = { class: "balanced" }): Promise<AiResolvedModel> =>
+  resolveAiModel({ kind: "selectable" }, model);
 
 export type RunAiStructuredInput<TOutput extends z.ZodType> = {
   /** Short machine name for tracing, e.g. "chat-enrich". */
@@ -53,7 +24,7 @@ export type RunAiStructuredInput<TOutput extends z.ZodType> = {
   output: TOutput;
   outputName?: string;
   systemPrompt?: string;
-  requestedModelId?: string;
+  model?: AiModelRequest;
   temperature?: number;
   maxOutputTokens?: number;
   signal?: AbortSignal;
@@ -65,17 +36,17 @@ export type RunAiStructuredInput<TOutput extends z.ZodType> = {
   /** Metadata-only attribution, supplied by the authorized caller. */
   attribution?: AiUsageAttribution;
   /** Model resolution seam — tests inject a fake so they never touch shared settings. */
-  resolveModel?: (requestedModelId?: string) => Promise<AiResolvedModel>;
+  resolveModel?: (model?: AiModelRequest) => Promise<AiResolvedModel>;
 };
 
 export type RunAiStructuredResult<TOutput extends z.ZodType> = {
   output: z.infer<TOutput>;
-  modelProfileId: string;
+  model: AiResolvedModelRef;
   usage?: Usage;
   structuredMeta: StructuredMeta;
 };
 
-/** nessi wraps provider failures in direct structured mode; callers rely on the provider's original error. */
+/** nessi wraps classified provider failures in direct structured mode; preserve their stable code for callers. */
 export const structuredProviderFailure = (error: unknown): unknown => {
   if (
     error instanceof StructuredOutputError &&
@@ -109,11 +80,17 @@ export const runAiStructured = async <TOutput extends z.ZodType>(
       const startedAt = Date.now();
       let resolved: AiResolvedModel | undefined;
       try {
-        resolved = await (input.resolveModel ?? resolveAiBackgroundModel)(input.requestedModelId);
+        resolved = await (input.resolveModel ?? resolveAiBackgroundModel)(input.model ?? { class: "balanced" });
         await trace.record({
           context: span,
           event: "model.resolved",
-          attributes: { model: resolved.profile.id, providerModel: resolved.profile.model, provider: resolved.profile.provider },
+          attributes: {
+            model: resolved.profile.id,
+            providerModel: resolved.profile.model,
+            provider: resolved.profile.provider,
+            requestedClass: resolved.model.requestedClass,
+            resolvedClass: resolved.model.class,
+          },
         });
 
         const result = await nessi.structured({
@@ -122,6 +99,7 @@ export const runAiStructured = async <TOutput extends z.ZodType>(
             kind: input.usageSubject ? "chat" : "background",
             subject: input.usageSubject,
             task: input.task,
+            modelRef: resolved.model,
             appId: input.appId,
             traceId: span.traceId,
             ...input.attribution,
@@ -155,7 +133,7 @@ export const runAiStructured = async <TOutput extends z.ZodType>(
           attribution: input.attribution,
           traceId: span.traceId,
           appId: input.appId,
-          modelProfileId: resolved.profile.id,
+          modelId: resolved.profile.id,
           providerModel: resolved.profile.model,
           status: "ok",
           durationMs,
@@ -167,7 +145,7 @@ export const runAiStructured = async <TOutput extends z.ZodType>(
 
         return {
           output: result.output,
-          modelProfileId: resolved.profile.id,
+          model: resolved.model,
           usage: result.usage,
           structuredMeta: result.structuredMeta,
         };
@@ -182,20 +160,25 @@ export const runAiStructured = async <TOutput extends z.ZodType>(
           attribution: input.attribution,
           traceId: span.traceId,
           appId: input.appId,
-          modelProfileId: resolved?.profile.id,
+          modelId: resolved?.profile.id,
           providerModel: resolved?.profile.model,
           status: "failed",
           durationMs: Date.now() - startedAt,
           usage: details?.aggregate?.usage,
           attempts: details?.attempts,
-          errorCode: failure instanceof StructuredOutputError ? failure.code : failure instanceof Error ? failure.name : "unknown",
+          errorCode:
+            failure instanceof StructuredOutputError || failure instanceof AiProviderError
+              ? failure.code
+              : failure instanceof Error
+                ? failure.name
+                : "unknown",
           error: failure instanceof Error ? failure.message : String(failure),
         });
         throw failure;
       }
     },
     {
-      summarize: (result) => ({ model: result.modelProfileId, mode: result.structuredMeta.mode, repaired: result.structuredMeta.repaired }),
+      summarize: (result) => ({ model: result.model.id, mode: result.structuredMeta.mode, repaired: result.structuredMeta.repaired }),
       onError: (error) => (error instanceof StructuredOutputError ? structuredFailureSummary(error) : undefined),
     },
   );

package/src/ai/task-contracts.ts

@@ -1,7 +1,8 @@
 import { z } from "zod";
+import { AiModelRequestSchema } from "../shared/ai-models";
 
 const prompt = z.string().trim().min(1).max(20_000);
-const modelProfileId = z.string().trim().min(1).max(120).optional();
+const model = AiModelRequestSchema.optional();
 const choice = z.string().trim().min(1).max(200);
 const choices = z
   .array(choice)
@@ -20,6 +21,7 @@ const structuredField = z
     required: z.boolean().default(true),
     choices: z.array(choice).min(1).max(50).optional(),
     maxLength: z.number().int().min(1).max(20_000).optional(),
+    minLength: z.number().int().min(1).max(20_000).optional(),
   })
   .strict()
   .superRefine((field, ctx) => {
@@ -27,6 +29,8 @@ const structuredField = z
       ctx.addIssue({ code: "custom", path: ["choices"], message: "Enum fields require choices." });
     if (field.type !== "enum" && field.choices)
       ctx.addIssue({ code: "custom", path: ["choices"], message: "Choices are only valid for enum fields." });
+    if (field.minLength !== undefined && (field.type !== "text" || field.minLength > (field.maxLength ?? 20_000)))
+      ctx.addIssue({ code: "custom", path: ["minLength"], message: "minLength requires a text field and cannot exceed maxLength." });
     if (field.type !== "text" && field.maxLength)
       ctx.addIssue({ code: "custom", path: ["maxLength"], message: "maxLength is only valid for text fields." });
   });
@@ -42,7 +46,7 @@ export const AiTaskRequestSchema = z.discriminatedUnion("kind", [
     kind: z.literal("generate_text"),
     prompt,
     input: z.unknown().optional(),
-    modelProfileId,
+    model,
     maxOutputChars: z.number().int().min(1).max(20_000).default(4_000),
   }),
   z.object({
@@ -50,7 +54,7 @@ export const AiTaskRequestSchema = z.discriminatedUnion("kind", [
     prompt,
     input: z.unknown(),
     choices,
-    modelProfileId,
+    model,
   }),
   z
     .object({
@@ -60,7 +64,7 @@ export const AiTaskRequestSchema = z.discriminatedUnion("kind", [
       choices,
       minChoices: z.number().int().min(0).max(50).default(0),
       maxChoices: z.number().int().min(0).max(50).optional(),
-      modelProfileId,
+      model,
     })
     .superRefine((value, ctx) => {
       const maximum = value.maxChoices ?? value.choices.length;
@@ -76,9 +80,25 @@ export const AiTaskRequestSchema = z.discriminatedUnion("kind", [
     prompt,
     input: z.unknown(),
     fields: structuredFields,
-    modelProfileId,
+    model,
   }),
 ]);
 
 export type AiTaskRequest = z.infer<typeof AiTaskRequestSchema>;
 export type AiTaskRequestInput = z.input<typeof AiTaskRequestSchema>;
+
+/** Existing durable task storage contract; measured in UTF-8 bytes, not characters. */
+export const AI_TASK_REQUEST_MAX_BYTES = 100_000;
+export class AiTaskInputTooLargeError extends Error {
+  readonly code = "WORKFLOW_AI_INPUT_TOO_LARGE";
+  readonly retryable = false;
+  constructor() {
+    super(
+      "AI input exceeds the 100,000-byte task budget. Narrow the supplied content or inspect the original source manually; nothing was truncated.",
+    );
+  }
+}
+
+export const assertAiTaskRequestSize = (request: unknown): void => {
+  if (Buffer.byteLength(JSON.stringify(request), "utf8") > AI_TASK_REQUEST_MAX_BYTES) throw new AiTaskInputTooLargeError();
+};

package/src/ai/task-execution.ts

@@ -11,7 +11,13 @@ const structuredOutputSchema = (fields: Extract<AiTaskRequest, { kind: "extract_
   for (const field of fields) {
     let schema: z.ZodType =
       field.type === "text"
-        ? z.string().max(field.maxLength ?? 20_000)
+        ? field.minLength === undefined
+          ? z.string().max(field.maxLength ?? 20_000)
+          : z
+              .string()
+              .trim()
+              .min(field.minLength)
+              .max(field.maxLength ?? 20_000)
         : field.type === "number"
           ? z.number()
           : field.type === "boolean"
@@ -28,13 +34,16 @@ const structuredOutputSchema = (fields: Extract<AiTaskRequest, { kind: "extract_
 /** Schema-validated AI calculations shared by workflows and interactive code. */
 export const executeAiTask = async (
   request: AiTaskRequest,
-  options: Pick<
-    RunAiStructuredInput<z.ZodType>,
-    "appId" | "attribution" | "signal" | "requestedModelId" | "resolveModel" | "usageSubject"
-  > & { taskPrefix: string },
+  options: Pick<RunAiStructuredInput<z.ZodType>, "appId" | "attribution" | "signal" | "model" | "resolveModel" | "usageSubject"> & {
+    taskPrefix: string;
+  },
   runStructured: typeof runAiStructured = runAiStructured,
 ) => {
-  const { taskPrefix, ...common } = options;
+  const { taskPrefix, ...provided } = options;
+  const common = {
+    ...provided,
+    model: provided.model ?? request.model ?? ({ class: "balanced" } satisfies import("../shared/ai-models").AiModelRequest),
+  };
   if (request.kind === "generate_text") {
     const result = await runStructured({
       ...common,

package/src/ai/task-model.ts

@@ -0,0 +1,9 @@
+import type { AccessSubject } from "../server/services/access";
+import { selectAssistantAiModel } from "./assistant-models";
+import { personalAiModelPolicy } from "./personal-agent";
+import { resolveAiModel } from "./settings";
+import { resolveAiWorkflowModel } from "./structured";
+
+/** Interactive application tasks retain Assistant grants and data-boundary policy. */
+export const resolveApplicationAiTaskModel = async (subject?: AccessSubject, model?: import("../shared/ai-models").AiModelRequest) =>
+  subject ? resolveAiModel(personalAiModelPolicy, await selectAssistantAiModel(subject, model)) : resolveAiWorkflowModel(model);