@modelprofile.com/flexharness 10.15.0 → 10.16.0

package/ts_document/index.ts

@@ -489,10 +489,10 @@ export async function analyzeDocumentsWithUsage(options: IDocumentOptions): Prom
     ? plugins.wrapLanguageModel({
         model: options.model,
         middleware: {
-          wrapGenerate: async ({ doGenerate, params }) => {
+          wrapStream: async ({ doStream, params }) => {
             modelCalls.start();
             try {
-              return await doGenerate();
+              return await doStream();
             } catch (error) {
               modelCalls.settle(params.abortSignal?.aborted ? 'aborted' : 'failed');
               throw error;
@@ -515,14 +515,20 @@ export async function analyzeDocumentsWithUsage(options: IDocumentOptions): Prom
       ],
     },
   ];
-  let result: Awaited<ReturnType<typeof plugins.generateText>>;
+  // The call is streamed: some providers, such as the ChatGPT Codex backend, refuse a call that is not.
+  let streamError: unknown;
+  let text = '';
+  let usage: Awaited<ReturnType<typeof plugins.streamText>['totalUsage']>;
   try {
-    result = await plugins.generateText({
+    const result = plugins.streamText({
       model,
       instructions: options.systemMessage,
       messages,
       providerOptions: options.providerOptions,
       abortSignal,
+      onError: ({ error }) => {
+        streamError ??= error;
+      },
       ...(modelCalls
         ? {
             onLanguageModelCallEnd: ({ modelId, usage: callUsage }) => {
@@ -531,16 +537,21 @@ export async function analyzeDocumentsWithUsage(options: IDocumentOptions): Prom
           }
         : {}),
     });
+    for await (const delta of result.textStream) text += delta;
+    // A stream that fails delivers its error to onError and ends, and an aborted stream just ends;
+    // either way the text is incomplete.
+    if (streamError !== undefined) throw streamError;
+    abortSignal?.throwIfAborted();
+    usage = await result.totalUsage;
     modelCalls?.settle('missing');
   } catch (error) {
     modelCalls?.settle(abortSignal?.aborted ? 'aborted' : 'failed');
-    throw error;
+    throw streamError ?? error;
   }
-  const usage = result.totalUsage;
   const inputTokens = usage.inputTokens ?? 0;
   const outputTokens = usage.outputTokens ?? 0;
   return {
-    text: result.text,
+    text,
     usage: {
       inputTokens,
       outputTokens,

package/ts_document/plugins.ts

@@ -1,7 +1,7 @@
-import { generateText, wrapLanguageModel } from 'ai';
+import { streamText, wrapLanguageModel } from 'ai';
 import * as smartpdf from '@push.rocks/smartpdf';
 import { AgentModelCallUsageRecorder } from '@modelprofile.com/flexharness-agent';
-export { generateText, wrapLanguageModel, smartpdf, AgentModelCallUsageRecorder };
+export { streamText, wrapLanguageModel, smartpdf, AgentModelCallUsageRecorder };
 export type { ModelMessage } from 'ai';
 export type { TSmartAiLanguageModel, TSmartAiProviderOptions } from '@modelprofile.com/flexharness-models';
 export type { TAgentModelCallUsageReporter } from '@modelprofile.com/flexharness-agent';

package/ts_document/readme.md

@@ -67,6 +67,8 @@ Both functions accept these request options:
 - `maxImageEdge` is the longest edge, in pixels, of an image document as it is sent; see above.
 - `onUsage` receives the usage of each model call of the request, also of a request that fails or is aborted; see [Usage of a request that fails or is aborted](#usage-of-a-request-that-fails-or-is-aborted).
 
+A request makes one streamed model call, retries aside, so it also reads documents on providers that answer streamed calls only, such as the Codex backend of a ChatGPT account.
+
 `extractTextFromPdfWithUsage` and `analyzeDocumentsWithUsage` take the same arguments and return `{ text, usage, pageCount }`. `usage` has the shape of an agent run's usage (`inputTokens`, `outputTokens`, `totalTokens`, `cacheReadTokens`, `cacheWriteTokens`), summed over the request's model steps, with usage the provider leaves out counted as zero; `pageCount` is the number of rendered pages sent to the model.
 
 ```typescript

package/ts_media/index.ts

@@ -8,7 +8,9 @@ export interface IVisionOptions {
 }
 
 export async function analyzeImage(options: IVisionOptions): Promise<string> {
-  const result = await plugins.generateText({
+  // The call is streamed: some providers, such as the ChatGPT Codex backend, refuse a call that is not.
+  let streamError: unknown;
+  const result = plugins.streamText({
     model: options.model,
     messages: [
       {
@@ -23,6 +25,13 @@ export async function analyzeImage(options: IVisionOptions): Promise<string> {
         ],
       },
     ],
+    onError: ({ error }) => {
+      streamError ??= error;
+    },
   });
-  return result.text;
+  let text = '';
+  for await (const delta of result.textStream) text += delta;
+  // A stream that fails delivers its error to onError and ends; its text is incomplete.
+  if (streamError !== undefined) throw streamError;
+  return text;
 }

package/ts_media/plugins.ts

@@ -1,3 +1,3 @@
-import { generateText } from 'ai';
-export { generateText };
+import { streamText } from 'ai';
+export { streamText };
 export type { TSmartAiLanguageModel } from '@modelprofile.com/flexharness-models';

package/ts_stores_nosqldb/classes.nosqldbstores.ts

@@ -11,16 +11,19 @@ import type {
   IFlexProjectManagementWriteContext,
   IFlexProjectionSnapshotV3,
   IFlexProjectionStore,
+  IFlexResultStore,
   IFlexScopeSnapshot,
   IFlexScopeStore,
   IFlexToolJobStoreProvider,
   TFlexProjectionSnapshot,
   TFlexProjectManagementRecord,
+  TJsonValue,
 } from '../ts/interfaces.js';
 import {
   assertFlexPermissionSnapshot,
   assertFlexProjectionSnapshot,
   assertFlexScopeSnapshot,
+  assertJsonSerializable,
 } from '../ts/utils.json.js';
 import {
   assertFlexProjectManagementRecord,
@@ -297,7 +300,7 @@ class NoSqlAgentEventStore implements plugins.IAgentEventStoreV2 {
       domain: 'agentEventArchive' as const,
       storageKey: this.storageKey,
       sessionId: this.sessionId,
-      archiveId: archive.archiveId,
+      entryId: archive.archiveId,
     };
     await this.records.mutate(address, {
       validate: (value) => validateAgentEventArchive(value, this.sessionId, archive.archiveId),
@@ -318,7 +321,7 @@ class NoSqlAgentEventStore implements plugins.IAgentEventStoreV2 {
     this.assertSession(sessionId);
     validateIdentifier(archiveId, 'archiveId');
     return this.records.read(
-      { domain: 'agentEventArchive', storageKey: this.storageKey, sessionId, archiveId },
+      { domain: 'agentEventArchive', storageKey: this.storageKey, sessionId, entryId: archiveId },
       (value) => validateAgentEventArchive(value, this.sessionId, archiveId),
     );
   }
@@ -428,6 +431,38 @@ export interface INoSqlFlexHarnessStoresOptions {
   operationTimeoutMs?: number;
 }
 
+class NoSqlResultStore implements IFlexResultStore {
+  constructor(private readonly records: FlexRecordStore) {}
+
+  public async save(storageKey: string, sessionId: string, handle: string, value: TJsonValue): Promise<void> {
+    validateIdentifier(handle, 'handle');
+    assertJsonSerializable(value);
+    const conflict = () => new FlexHarnessStoreConflictError(`${providerKey(storageKey, sessionId)}:${handle}`, 0, 1);
+    await this.records.mutate({ domain: 'result', storageKey, sessionId, entryId: handle }, {
+      validate: validateResult,
+      decide: (current) => {
+        if (current !== undefined) throw conflict();
+        return { revision: 1, value };
+      },
+      contention: conflict,
+    });
+  }
+
+  public load(storageKey: string, sessionId: string, handle: string): Promise<TJsonValue | undefined> {
+    validateIdentifier(handle, 'handle');
+    return this.records.read({ domain: 'result', storageKey, sessionId, entryId: handle }, validateResult);
+  }
+
+  public deleteSession(storageKey: string, sessionId: string): Promise<void> {
+    return this.records.deleteOwner(storageKey, sessionId, ['result']);
+  }
+}
+
+function validateResult(value: unknown): TJsonValue {
+  assertJsonSerializable(value);
+  return value as TJsonValue;
+}
+
 /**
  * Every FlexHarness persistence domain on `@lossless.org/client/nosqldb`, with compare-and-swap
  * that holds across processes. Records live in `flexharness_records`; payloads longer than one
@@ -440,6 +475,7 @@ export class NoSqlFlexHarnessStores implements IFlexHarnessStores {
   public readonly projectManagement: IFlexProjectManagementStore;
   public readonly agentEvents: IFlexAgentEventStoreProvider;
   public readonly jobs: IFlexToolJobStoreProvider;
+  public readonly results: IFlexResultStore;
   private readonly records: FlexRecordStore;
 
   constructor(options: INoSqlFlexHarnessStoresOptions) {
@@ -453,6 +489,7 @@ export class NoSqlFlexHarnessStores implements IFlexHarnessStores {
     this.projectManagement = new NoSqlProjectManagementStore(this.records);
     this.agentEvents = new NoSqlAgentEventStoreProvider(this.records);
     this.jobs = new NoSqlToolJobStoreProvider(this.records);
+    this.results = new NoSqlResultStore(this.records);
   }
 
   /**

package/ts_stores_nosqldb/classes.recordstore.ts

@@ -12,14 +12,16 @@ export type TFlexRecordDomain =
   | 'projectManagement'
   | 'agentEvents'
   | 'agentEventArchive'
-  | 'jobs';
+  | 'jobs'
+  | 'result';
 
-/** Names one record: a domain snapshot of a storage key, a session or an archive. */
+/** Names one record: a domain snapshot of a storage key or a session, or one entry of a session. */
 export interface IFlexRecordAddress {
   domain: TFlexRecordDomain;
   storageKey: string;
   sessionId?: string;
-  archiveId?: string;
+  /** The archive or result the record holds. */
+  entryId?: string;
 }
 
 /** The next stored value a mutation decided on. */
@@ -256,7 +258,7 @@ export class FlexRecordStore {
       address.domain,
       address.storageKey,
       address.sessionId ?? null,
-      address.archiveId ?? null,
+      address.entryId ?? null,
     ]));
     return {
       recordId,