@k2b/cloud 0.6.0 → 0.8.0

package/src/ai/projects-routes.ts

@@ -2,7 +2,7 @@ import { Hono, type MiddlewareHandler } from "hono";
 import { z } from "zod";
 import { getCapability } from "../_internal/registry";
 import { CloudResourceRefSchema, cloudResourceRefAppId, resolveCapabilityResourceReader } from "../contracts/capabilities";
-import { PrincipalSchema } from "../contracts/shared";
+import { AuthenticatedPrincipalSchema } from "../contracts/shared";
 import { type AuthContext, auth, err, fail, ok, rateLimit, respond, v } from "../server";
 import { decodeAiFileContent } from "./files-store";
 import {
@@ -28,7 +28,7 @@ const ProjectFieldsSchema = z.object({
 });
 
 const UpdateProjectSchema = ProjectFieldsSchema.partial().refine((input) => Object.keys(input).length > 0, "No changes supplied.");
-const ProjectAccessSchema = z.object({ principal: PrincipalSchema, permission: z.enum(["read", "write", "admin"]) });
+const ProjectAccessSchema = z.object({ principal: AuthenticatedPrincipalSchema, permission: z.enum(["read", "write", "admin"]) });
 const ProjectAccessUpdateSchema = z.object({ permission: z.enum(["read", "write", "admin"]) });
 const KnowledgeSchema = z.object({
   title: z.string().trim().min(1).max(200),
@@ -70,17 +70,13 @@ type AiProjectsRouteDependencies = {
 const buildAiProjectsRoutes = (dependencies: AiProjectsRouteDependencies = {}) =>
   new Hono<AuthContext>()
     .use(dependencies.limit ?? rateLimit())
-    .use("*", dependencies.authenticate ?? auth.requireRole("*"))
-    .get("/", async (c) =>
-      respond(c, ok({ projects: (await aiProjects.list(c.get("accessSubject") ?? null)).map(publicProject) })),
-    )
+    .use("*", dependencies.authenticate ?? auth.requireRole("authenticated"))
+    .get("/", async (c) => respond(c, ok({ projects: (await aiProjects.list(c.get("accessSubject") ?? null)).map(publicProject) })))
     .post("/", dependencies.authenticate ?? auth.requireRole("authenticated"), v("json", ProjectFieldsSchema), async (c) =>
       respond(
         c,
         ok({
-          project: publicProject(
-            await aiProjects.create({ subject: c.get("accessSubject"), ...c.req.valid("json") }),
-          ),
+          project: publicProject(await aiProjects.create({ subject: c.get("accessSubject"), ...c.req.valid("json") })),
         }),
         201,
       ),
@@ -108,9 +104,7 @@ const buildAiProjectsRoutes = (dependencies: AiProjectsRouteDependencies = {}) =
     })
     .post("/:projectId/access", v("json", ProjectAccessSchema), async (c) => {
       const project = await aiProjects.getByShortId(c.req.param("projectId")!, c.get("accessSubject") ?? null, "admin");
-      const access = project
-        ? await aiProjects.grantAccess(project.id, c.get("accessSubject") ?? null, c.req.valid("json"))
-        : null;
+      const access = project ? await aiProjects.grantAccess(project.id, c.get("accessSubject") ?? null, c.req.valid("json")) : null;
       return access ? respond(c, ok({ access }), 201) : notFound(c);
     })
     .patch("/:projectId/access/:accessId", v("json", ProjectAccessUpdateSchema), async (c) => {
@@ -133,8 +127,7 @@ const buildAiProjectsRoutes = (dependencies: AiProjectsRouteDependencies = {}) =
     .delete("/:projectId/access/:accessId", async (c) => {
       const project = await aiProjects.getByShortId(c.req.param("projectId")!, c.get("accessSubject") ?? null, "admin");
       try {
-        return project &&
-          (await aiProjects.revokeAccess(project.id, c.req.param("accessId")!, c.get("accessSubject") ?? null))
+        return project && (await aiProjects.revokeAccess(project.id, c.req.param("accessId")!, c.get("accessSubject") ?? null))
           ? respond(c, ok({ deleted: true }))
           : notFound(c, "Access entry");
       } catch (error) {
@@ -148,35 +141,27 @@ const buildAiProjectsRoutes = (dependencies: AiProjectsRouteDependencies = {}) =
       return respond(
         c,
         ok({
-          knowledge: (
-            await aiProjects.listKnowledge(project.id, c.get("accessSubject") ?? null, c.req.valid("query").q)
-          ).map((item) => publicKnowledge(item, project.shortId)),
+          knowledge: (await aiProjects.listKnowledge(project.id, c.get("accessSubject") ?? null, c.req.valid("query").q)).map((item) =>
+            publicKnowledge(item, project.shortId),
+          ),
         }),
       );
     })
     .post("/:projectId/knowledge", v("json", KnowledgeSchema), async (c) => {
       const project = await aiProjects.getByShortId(c.req.param("projectId")!, c.get("accessSubject") ?? null, "write");
-      const knowledge = project
-        ? await aiProjects.createKnowledge(project.id, c.get("accessSubject") ?? null, c.req.valid("json"))
-        : null;
+      const knowledge = project ? await aiProjects.createKnowledge(project.id, c.get("accessSubject") ?? null, c.req.valid("json")) : null;
       return knowledge ? respond(c, ok({ knowledge: publicKnowledge(knowledge, project!.shortId) }), 201) : notFound(c);
     })
     .patch("/:projectId/knowledge/:knowledgeId", v("json", UpdateKnowledgeSchema), async (c) => {
       const project = await aiProjects.getByShortId(c.req.param("projectId")!, c.get("accessSubject") ?? null, "write");
       const knowledge =
         project &&
-        (await aiProjects.updateKnowledge(
-          project.id,
-          c.req.param("knowledgeId")!,
-          c.get("accessSubject") ?? null,
-          c.req.valid("json"),
-        ));
+        (await aiProjects.updateKnowledge(project.id, c.req.param("knowledgeId")!, c.get("accessSubject") ?? null, c.req.valid("json")));
       return knowledge ? respond(c, ok({ knowledge: publicKnowledge(knowledge, project!.shortId) })) : notFound(c, "Knowledge entry");
     })
     .delete("/:projectId/knowledge/:knowledgeId", async (c) => {
       const project = await aiProjects.getByShortId(c.req.param("projectId")!, c.get("accessSubject") ?? null, "write");
-      return project &&
-        (await aiProjects.deleteKnowledge(project.id, c.req.param("knowledgeId")!, c.get("accessSubject") ?? null))
+      return project && (await aiProjects.deleteKnowledge(project.id, c.req.param("knowledgeId")!, c.get("accessSubject") ?? null))
         ? respond(c, ok({ deleted: true }))
         : notFound(c, "Knowledge entry");
     })
@@ -186,20 +171,14 @@ const buildAiProjectsRoutes = (dependencies: AiProjectsRouteDependencies = {}) =
       return respond(
         c,
         ok({
-          files: (await aiProjects.listFiles(project.id, c.get("accessSubject") ?? null)).map((file) =>
-            publicFile(file, project.shortId),
-          ),
+          files: (await aiProjects.listFiles(project.id, c.get("accessSubject") ?? null)).map((file) => publicFile(file, project.shortId)),
         }),
       );
     })
     .post("/:projectId/files", v("json", ProjectFileSchema), async (c) => {
       const body = c.req.valid("json");
       try {
-        const project = await aiProjects.getByShortId(
-          c.req.param("projectId")!,
-          c.get("accessSubject") ?? null,
-          "write",
-        );
+        const project = await aiProjects.getByShortId(c.req.param("projectId")!, c.get("accessSubject") ?? null, "write");
         const file = project
           ? await aiProjects.writeFile(project.id, c.get("accessSubject") ?? null, {
               path: body.path,
@@ -214,9 +193,7 @@ const buildAiProjectsRoutes = (dependencies: AiProjectsRouteDependencies = {}) =
     })
     .get("/:projectId/files/:fileId", async (c) => {
       const project = await aiProjects.getByShortId(c.req.param("projectId")!, c.get("accessSubject") ?? null);
-      const file = project
-        ? await aiProjects.readFile(project.id, c.req.param("fileId")!, c.get("accessSubject") ?? null)
-        : null;
+      const file = project ? await aiProjects.readFile(project.id, c.req.param("fileId")!, c.get("accessSubject") ?? null) : null;
       if (!file) return notFound(c, "Project file");
       return respond(
         c,
@@ -229,8 +206,7 @@ const buildAiProjectsRoutes = (dependencies: AiProjectsRouteDependencies = {}) =
     })
     .delete("/:projectId/files/:fileId", async (c) => {
       const project = await aiProjects.getByShortId(c.req.param("projectId")!, c.get("accessSubject") ?? null, "write");
-      return project &&
-        (await aiProjects.deleteFile(project.id, c.req.param("fileId")!, c.get("accessSubject") ?? null))
+      return project && (await aiProjects.deleteFile(project.id, c.req.param("fileId")!, c.get("accessSubject") ?? null))
         ? respond(c, ok({ deleted: true }))
         : notFound(c, "Project file");
     })
@@ -265,8 +241,7 @@ const buildAiProjectsRoutes = (dependencies: AiProjectsRouteDependencies = {}) =
     })
     .delete("/:projectId/references/:referenceId", async (c) => {
       const project = await aiProjects.getByShortId(c.req.param("projectId")!, c.get("accessSubject") ?? null, "write");
-      return project &&
-        (await aiProjects.deleteReference(project.id, c.req.param("referenceId")!, c.get("accessSubject") ?? null))
+      return project && (await aiProjects.deleteReference(project.id, c.req.param("referenceId")!, c.get("accessSubject") ?? null))
         ? respond(c, ok({ deleted: true }))
         : notFound(c, "Project reference");
     });

package/src/ai/projects.ts

@@ -1,5 +1,6 @@
 import { type SQL, type SQLQuery, sql } from "bun";
 import type { CloudResourceRef } from "../contracts/capabilities";
+import { AuthenticatedPrincipalSchema } from "../contracts/shared";
 import type { AccessSubject } from "../server";
 import {
   buildAccessPrincipalCondition,
@@ -10,6 +11,7 @@ import {
   type Principal,
 } from "../server/services/access";
 import { toPgUuidArray } from "../services/postgres";
+import { AiFileVersionConflict, AiFileWriteError, aiFileContentVersion } from "./file-content-version";
 import { mountAiProjectFilePath } from "./file-mount";
 import { withAiShortIdForDb } from "./short-id";
 import type { AiProjectPromptSnapshot } from "./types";
@@ -168,15 +170,17 @@ const principalForSubject = (subject: AccessSubject): Principal =>
     : { type: "service_account", serviceAccountId: subject.serviceAccountId };
 
 const accessMatch = (subject: AccessSubject | null): SQLQuery =>
-  buildAccessPrincipalCondition({
-    subject,
-    columns: {
-      userId: sql`access.user_id`,
-      groupId: sql`access.group_id`,
-      serviceAccountId: sql`access.service_account_id`,
-      authenticatedOnly: sql`access.authenticated_only`,
-    },
-  });
+  subject === null
+    ? sql`FALSE`
+    : buildAccessPrincipalCondition({
+        subject,
+        columns: {
+          userId: sql`access.user_id`,
+          groupId: sql`access.group_id`,
+          serviceAccountId: sql`access.service_account_id`,
+          authenticatedOnly: sql`access.authenticated_only`,
+        },
+      });
 
 const permissionFor = async (row: ProjectRow, subject: AccessSubject | null, db: SQL = sql): Promise<AiProjectPermission | "none"> => {
   const match = accessMatch(subject);
@@ -299,6 +303,7 @@ const createProjectAccess = async (
   input: { principal: Principal; permission: AiProjectPermission },
   db: SQL,
 ): Promise<AiProjectAccess> => {
+  AuthenticatedPrincipalSchema.parse(input.principal);
   const created = await createAccess(input, db);
   if (!created.ok) throw new Error(created.error.message);
   let shortId = "";
@@ -439,7 +444,6 @@ export const aiProjects = {
       JOIN auth.access access ON access.id = project_access.access_id
       WHERE access.permission <> 'none' AND ${match}
       ORDER BY lower(project.name), project.id
-      LIMIT 200
     `;
     return (await Promise.all(rows.map((row) => toProject(row, subject)))).filter((project): project is AiProject => Boolean(project));
   },
@@ -787,24 +791,36 @@ export const aiProjects = {
     });
   },
 
-  async listFiles(projectId: string, subject: AccessSubject | null): Promise<AiProjectFile[]> {
+  async listFiles(projectId: string, subject: AccessSubject | null, page?: { after?: string; limit: number }): Promise<AiProjectFile[]> {
     if (!(await requireProject(projectId, subject, "read"))) return [];
+    if (page && (!Number.isInteger(page.limit) || page.limit < 1 || page.limit > 1000)) throw new Error("Invalid file page limit");
     return (
       await sql<
         FileRow[]
-      >`SELECT id, short_id, project_id, path, media_type, size, updated_at FROM ai.project_files WHERE project_id = ${projectId}::uuid ORDER BY path LIMIT 500`
+      >`SELECT id, short_id, project_id, path, media_type, size, updated_at FROM ai.project_files WHERE project_id = ${projectId}::uuid AND path > ${page?.after ?? ""} ORDER BY path LIMIT ${page?.limit ?? 500}`
     ).map(toFile);
   },
 
   async writeFile(
     projectId: string,
     subject: AccessSubject | null,
-    input: { path: string; mediaType: string; bytes: Uint8Array },
+    input: { path: string; mediaType: string; bytes: Uint8Array; expectedVersion?: string | null },
   ): Promise<AiProjectFile | null> {
     if (!(await requireProject(projectId, subject, "write"))) return null;
-    if (input.bytes.byteLength > AI_PROJECT_FILE_MAX_BYTES) throw new Error("Project file exceeds the size limit.");
+    if (input.bytes.byteLength > AI_PROJECT_FILE_MAX_BYTES)
+      throw new AiFileWriteError("STORAGE_FULL", "Project file exceeds the size limit; nothing was written.");
     const path = normalizeProjectPath(input.path);
     return sql.begin(async (tx) => {
+      const [row] = await tx<ProjectRow[]>`SELECT * FROM ai.projects WHERE id=${projectId}::uuid FOR UPDATE`;
+      const project = row ? await toProject(row, subject, tx) : null;
+      if (!project || !hasPermission(project.permission, "write")) return null;
+      if (input.expectedVersion !== undefined) {
+        const [existing] = await tx<
+          (FileRow & { bytes: Uint8Array })[]
+        >`SELECT id,short_id,project_id,path,media_type,size,updated_at,bytes FROM ai.project_files WHERE project_id=${projectId}::uuid AND path=${path}`;
+        const version = existing ? aiFileContentVersion({ ...toFile(existing), bytes: existing.bytes }) : null;
+        if (version !== input.expectedVersion) throw new AiFileVersionConflict();
+      }
       const rows = await withAiShortIdForDb(
         tx,
         "idx_ai_project_files_short_id",
@@ -851,6 +867,9 @@ export const aiProjects = {
   async deleteFile(projectId: string, fileId: string, subject: AccessSubject | null): Promise<boolean> {
     if (!(await requireProject(projectId, subject, "write"))) return false;
     return sql.begin(async (tx) => {
+      const [row] = await tx<ProjectRow[]>`SELECT * FROM ai.projects WHERE id=${projectId}::uuid FOR UPDATE`;
+      const project = row ? await toProject(row, subject, tx) : null;
+      if (!project || !hasPermission(project.permission, "write")) return false;
       const rows = await tx<{ id: string }[]>`
         DELETE FROM ai.project_files WHERE short_id = ${fileId} AND project_id = ${projectId}::uuid RETURNING id
       `;

package/src/ai/protocol.ts

@@ -33,6 +33,8 @@ export type AiTurnBlock =
       name: string;
       args?: unknown;
       status: AiToolBlockStatus;
+      /** Localized transient host status, without tool payload data. */
+      progress?: string;
       result?: unknown;
       isError?: boolean;
       /** Present while status is awaiting_approval. */
@@ -62,6 +64,7 @@ export type AiWireEvent =
       /** Authoritative ordered baseline for this attempt. Optional for older senders. */
       blocks?: AiTurnBlock[];
     })
+  | (AiWireEventBase & { type: "message_saved"; message: AiStoredMessage })
   | (AiWireEventBase & { type: "block_set"; block: AiTurnBlock })
   | (AiWireEventBase & { type: "block_delta"; blockId: string; blockKind: "text" | "thinking"; delta: string })
   | (AiWireEventBase & {

package/src/ai/provider.ts

@@ -10,12 +10,11 @@ const commonOptions = (profile: AiModelProfile, apiKey?: string) => ({
   baseURL: profile.baseURL,
   contextWindow: profile.contextWindow,
   temperature: profile.temperature,
-  creditsPerInputToken: profile.creditsPerInputToken,
-  creditsPerOutputToken: profile.creditsPerOutputToken,
   timeouts: PROVIDER_TIMEOUTS,
 });
 
 export const createAiProvider = (profile: AiModelProfile, apiKey?: string): Provider => {
+  if (profile.capabilities.includes("transcription")) throw new Error("Transcription profiles cannot be used for chat generation.");
   switch (profile.provider) {
     case "openai":
       return openai(profile.model, commonOptions(profile, apiKey));
@@ -32,8 +31,6 @@ export const createAiProvider = (profile: AiModelProfile, apiKey?: string): Prov
         baseURL: profile.baseURL,
         contextWindow: profile.contextWindow,
         temperature: profile.temperature,
-        creditsPerInputToken: profile.creditsPerInputToken,
-        creditsPerOutputToken: profile.creditsPerOutputToken,
         timeouts: PROVIDER_TIMEOUTS,
       });
     case "vllm":
@@ -44,8 +41,6 @@ export const createAiProvider = (profile: AiModelProfile, apiKey?: string): Prov
         apiKey,
         contextWindow: profile.contextWindow,
         temperature: profile.temperature,
-        creditsPerInputToken: profile.creditsPerInputToken,
-        creditsPerOutputToken: profile.creditsPerOutputToken,
         timeouts: PROVIDER_TIMEOUTS,
         compat: {
           toolCallIdPolicy: "passthrough",
@@ -65,8 +60,6 @@ export const createAiProvider = (profile: AiModelProfile, apiKey?: string): Prov
         apiKey,
         contextWindow: profile.contextWindow,
         temperature: profile.temperature,
-        creditsPerInputToken: profile.creditsPerInputToken,
-        creditsPerOutputToken: profile.creditsPerOutputToken,
         timeouts: PROVIDER_TIMEOUTS,
       });
   }

package/src/ai/quota-provider.ts

@@ -0,0 +1,192 @@
+import { setTimeout as delay } from "node:timers/promises";
+import { estimateTokens, type Provider } from "@k2b/nessi";
+import { sql } from "bun";
+import type { AccessSubject } from "../server/services/access";
+import { logger } from "../services/logging";
+import { isAssistantChatTurn } from "./assistant-models";
+import { AiBackgroundAdmissionError, type AiCallContext, beginAiCall, finishAiCall } from "./inference-calls";
+import type { AiModelProfile } from "./types";
+
+const log = logger("ai:quotas");
+
+/** One wrapper owns call accounting. complete() can have a separate compaction purpose. */
+const callLifecycle = {
+  begin: beginAiCall,
+  finish: finishAiCall,
+  heartbeat: async (id: string) => {
+    await sql`UPDATE ai.inference_calls SET lease_expires_at=now()+interval '2 minutes' WHERE id=${id}::uuid AND finished_at IS NULL`;
+  },
+};
+export function inferenceProvider(
+  provider: Provider,
+  profile: AiModelProfile,
+  context: AiCallContext,
+  completeContext = context,
+  lifecycle = callLifecycle,
+): Provider {
+  const begin = async (request: Parameters<Provider["complete"]>[0], ctx: AiCallContext) => {
+    const messages = request.messages.map((message) =>
+      message.role === "user"
+        ? {
+            ...message,
+            content: message.content.filter((part) => typeof part === "string" || part.type !== "file"),
+          }
+        : message,
+    );
+    const inputTokens =
+      estimateTokens(messages) +
+      Math.ceil(
+        ((request.systemPrompt?.length ?? 0) +
+          JSON.stringify(request.tools ?? []).length +
+          JSON.stringify(request.responseFormat ?? {}).length) /
+          4,
+      );
+    // Wait up to one reservation lease before delegating to the caller's retry policy.
+    // No provider request or new reservation exists while waiting; cancellation stays active.
+    const deadline = Date.now() + 120_000;
+    const admit = async () => {
+      for (;;) {
+        request.signal?.throwIfAborted();
+        try {
+          return await lifecycle.begin(
+            profile,
+            ctx,
+            inputTokens,
+            request.maxOutputTokens ?? profile.maxOutputTokens,
+            // Nessi's Anthropic adapter defaults to 1024; other adapters use the model default.
+            provider.family === "anthropic" ? 1024 : provider.contextWindow,
+          );
+        } catch (error) {
+          if (!(error instanceof AiBackgroundAdmissionError) || !error.retryable || Date.now() >= deadline) throw error;
+          await delay(1_000, undefined, { signal: request.signal });
+        }
+      }
+    };
+    const call = await admit();
+    const heartbeat = setInterval(() => {
+      void lifecycle.heartbeat(call.id).catch(() => log.warn("AI accounting heartbeat failed", { callId: call.id }));
+    }, 30_000);
+    return { ...call, inputTokens, stop: () => clearInterval(heartbeat) };
+  };
+  const finish = async (id: string, usage: Parameters<typeof finishAiCall>[1], status: "ok" | "failed") => {
+    try {
+      await lifecycle.finish(id, usage, status);
+    } catch {
+      log.error("AI cost booking failed", { callId: id, code: "ai_cost_booking_failed" });
+    }
+  };
+  return {
+    name: provider.name,
+    family: provider.family,
+    model: provider.model,
+    contextWindow: provider.contextWindow,
+    capabilities: provider.capabilities,
+    complete: async (request) => {
+      const call = await begin(request, completeContext);
+      let usage: Parameters<typeof finishAiCall>[1];
+      let status: "ok" | "failed" = "failed";
+      try {
+        const result = await provider.complete({ ...request, maxOutputTokens: call.maxOutputTokens });
+        if (
+          result.usage &&
+          [result.usage.input, result.usage.output].every((n) => Number.isSafeInteger(n) && n >= 0) &&
+          result.usage.input + result.usage.output > 0
+        )
+          usage = { input: result.usage.input, output: result.usage.output };
+        status = ["error", "aborted", "interrupted"].includes(result.finishReason) ? "failed" : "ok";
+        return result;
+      } finally {
+        call.stop();
+        if (!usage && status === "failed") usage = { input: call.inputTokens, output: 0, estimated: true };
+        await finish(call.id, usage, status);
+      }
+    },
+    stream: async function* (request) {
+      const call = await begin(request, context);
+      const id = call.id;
+      let usage: { input: number; output: number; estimated?: boolean } | undefined;
+      let completed = false;
+      let failed = false;
+      const outputBlocks = new Map<string, number>();
+      let generated = false;
+      try {
+        for await (const event of provider.stream({ ...request, maxOutputTokens: call.maxOutputTokens })) {
+          if (event.type === "issue" && event.issue.kind === "provider_error") failed = true;
+          if (event.type === "block_delta") outputBlocks.set(event.blockId, (outputBlocks.get(event.blockId) ?? 0) + event.delta.length);
+          if (event.type === "block_end") {
+            const block = event.block;
+            const size =
+              block.type === "text"
+                ? block.text.length
+                : block.type === "thinking"
+                  ? block.thinking.length
+                  : block.name.length + JSON.stringify(block.args).length;
+            outputBlocks.set(event.blockId, size);
+          }
+          if (event.type === "block_start" || event.type === "block_delta" || event.type === "block_end") generated = true;
+          if (event.type === "issue" && event.issue.kind === "provider_error" && event.issue.contextOverflow && !generated && !usage)
+            usage = { input: 0, output: 0 };
+          if (event.type === "usage") {
+            if (event.finishReason === "aborted" || event.finishReason === "interrupted" || event.finishReason === "error") failed = true;
+            const { input, output } = event.usage;
+            // Some adapters synthesize all-zero usage when the endpoint reports none.
+            if (Number.isSafeInteger(input) && Number.isSafeInteger(output) && input >= 0 && output >= 0 && input + output > 0)
+              usage = { input, output };
+          }
+          yield event;
+        }
+        completed = true;
+      } finally {
+        call.stop();
+        if (id) {
+          try {
+            // Interrupted adapters commonly omit their final usage event. Charge an
+            // explicitly labelled estimate, never invent a measured zero.
+            const interrupted = !completed || failed || request.signal?.aborted;
+            if (!usage && interrupted) {
+              // Encoded files are not text tokens. Their provider-specific costs
+              // cannot be reconstructed after interruption.
+              const messages = request.messages.map((message) =>
+                message.role === "user"
+                  ? {
+                      ...message,
+                      content: message.content.filter((part) => typeof part === "string" || part.type !== "file"),
+                    }
+                  : message,
+              );
+              const instructionSize =
+                (request.systemPrompt?.length ?? 0) +
+                JSON.stringify(request.tools ?? []).length +
+                JSON.stringify(request.responseFormat ?? {}).length;
+              usage = {
+                input: estimateTokens(messages) + Math.ceil(instructionSize / 4),
+                output: Math.ceil(Array.from(outputBlocks.values()).reduce((sum, size) => sum + size, 0) / 4),
+                estimated: true,
+              };
+            }
+            await finish(id, usage, completed && !failed && !request.signal?.aborted ? "ok" : "failed");
+          } catch (error) {
+            log.warn("Chat usage booking failed", { code: "quota_booking_failed", turnId: context.turnId, callId: id });
+          }
+        }
+      }
+    },
+  };
+}
+
+export function assistantQuotaProvider(
+  provider: Provider,
+  config: Parameters<typeof isAssistantChatTurn>[0],
+  subject: AccessSubject | null,
+  profile: AiModelProfile,
+  turnId: string,
+  conversationId: string,
+) {
+  const common = { subject, turnId, conversationId };
+  return inferenceProvider(
+    provider,
+    profile,
+    { ...common, kind: isAssistantChatTurn(config) ? "chat" : "background", task: isAssistantChatTurn(config) ? "chat" : "agent-turn" },
+    { ...common, kind: "background", task: "chat-compaction" },
+  );
+}