Skip to content

armgit / macro

Public Imported from GitHub
0

Macro is a unified workspace for teams: email, chat, docs, tasks, agents, calls, and CRM — @-linked together with shared AI memory.

feat(ai): add accelerated model speeds and pricing (#7802)

Wolf Mermelstein9c8d51f64b56
diff --git a/apps/web/src/components/ui/components/SpeedToggle.docs.tsx b/apps/web/src/components/ui/components/SpeedToggle.docs.tsx
new file mode 100644
index 0000000000..d866d454f6
--- /dev/null
+++ b/apps/web/src/components/ui/components/SpeedToggle.docs.tsx
@@ -0,0 +1,44 @@
+import { defineDoc } from '@app/features/ui-gallery/types';
+import { SpeedToggle } from '@core/component/AI/component/input/SpeedToggle';
+import { Model } from '@core/component/AI/constant/model';
+
+// #region demo:speeds
+function SpeedsDemo() {
+ return (
+ <div class="flex flex-col gap-4">
+ <div class="flex items-center gap-3">
+ <SpeedToggle model={Model.gpt6Astra} /> GPT-6 Astra · 6×
+ </div>
+ <div class="flex items-center gap-3">
+ <SpeedToggle model={Model.gpt61Sol} /> GPT-6.1 Sol · 6×
+ </div>
+ <div class="flex items-center gap-3">
+ <SpeedToggle model={Model.opus55} /> Opus 5.5 · 2×
+ </div>
+ <div class="flex items-center gap-3">
+ <SpeedToggle model={Model.sonnet55} /> Sonnet 5.5 · standard only
+ </div>
+ </div>
+ );
+}
+// #endregion
+
+export default defineDoc({
+ name: 'AI speed toggle',
+ category: 'Inputs',
+ guidelines: {
+ do: ['Show the usage multiplier when offering accelerated model speeds.'],
+ dont: [
+ 'Show the lightning control for models that do not support fast mode.',
+ ],
+ },
+ description:
+ 'A persistent, explicit paid-speed preference shared by chat composers. Click a bolt to preview the animation; the preference survives reloads.',
+ demos: [
+ {
+ id: 'speeds',
+ title: 'Model speed support',
+ render: SpeedsDemo,
+ },
+ ],
+});
diff --git a/apps/web/src/features/agents-view/views/NewChatPage.test.tsx b/apps/web/src/features/agents-view/views/NewChatPage.test.tsx
index 69adefff1e..144ea446d3 100644
--- a/apps/web/src/features/agents-view/views/NewChatPage.test.tsx
+++ b/apps/web/src/features/agents-view/views/NewChatPage.test.tsx
@@ -17,6 +17,7 @@ import {
type PersistedAgentLike,
type RuntimeLike,
} from '../core/roster';
+import { NEW_CONVERSATION_ATTACHMENTS_KEY } from '../primitives/composer-draft';
import { AgentPicker } from './AgentPicker';
import { NewChatPage } from './NewChatPage';
@@ -332,6 +333,7 @@ describe('agent-led new conversation', () => {
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenCalledWith(
expect.objectContaining({
+ onFailure: expect.any(Function),
prompt: 'Phone draft',
botId: CURSOR_BOT_ID,
modelOverride: 'gpt-5',
@@ -375,6 +377,7 @@ describe('agent-led new conversation', () => {
await waitFor(() => expect(screen.queryByRole('menu')).toBeNull());
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenCalledWith({
+ onFailure: expect.any(Function),
prompt: 'Prompt',
botId: undefined,
repoUrl: undefined,
@@ -386,6 +389,7 @@ describe('agent-led new conversation', () => {
const send = page();
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenCalledWith({
+ onFailure: expect.any(Function),
prompt: 'Prompt',
botId: undefined,
repoUrl: undefined,
@@ -417,6 +421,7 @@ describe('agent-led new conversation', () => {
).toBe('Shared draft');
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenLastCalledWith({
+ onFailure: expect.any(Function),
prompt: 'Shared draft',
botId: undefined,
repoUrl: undefined,
@@ -427,6 +432,7 @@ describe('agent-led new conversation', () => {
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenLastCalledWith({
botId: CURSOR_BOT_ID,
+ onFailure: expect.any(Function),
prompt: 'Shared draft',
repoUrl: 'https://github.com/macro-inc/macro',
repoBranch: 'feature/home',
@@ -447,6 +453,7 @@ describe('agent-led new conversation', () => {
await waitFor(() => expect(screen.queryByRole('menu')).toBeNull());
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenCalledWith({
+ onFailure: expect.any(Function),
prompt: 'Prompt',
botId: undefined,
repoUrl: undefined,
@@ -495,6 +502,7 @@ describe('agent-led new conversation', () => {
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenLastCalledWith({
botId: CURSOR_BOT_ID,
+ onFailure: expect.any(Function),
prompt: 'Prompt',
repoUrl: undefined,
});
@@ -529,6 +537,7 @@ describe('agent-led new conversation', () => {
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenLastCalledWith({
botId: CURSOR_BOT_ID,
+ onFailure: expect.any(Function),
prompt: 'Prompt',
repoUrl: 'https://github.com/macro-inc/macro',
repoBranch: 'feature/other',
@@ -546,6 +555,7 @@ describe('agent-led new conversation', () => {
await selectAgent(/Reviewer/);
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenCalledWith({
+ onFailure: expect.any(Function),
prompt: 'Prompt',
botId: 'saved-agent',
repoUrl: undefined,
@@ -575,6 +585,7 @@ describe('agent-led new conversation', () => {
expect(screen.getByRole('button', { name: 'Branch' })).toBeTruthy();
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenCalledWith({
+ onFailure: expect.any(Function),
prompt: 'Prompt',
botId: 'saved-cloud-agent',
repoUrl: undefined,
@@ -608,6 +619,7 @@ describe('agent-led new conversation', () => {
expect(branch.hasAttribute('disabled')).toBe(true);
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenCalledWith({
+ onFailure: expect.any(Function),
prompt: 'Prompt',
botId: 'local-agent',
repoUrl: 'https://github.com/macro-inc/macro',
@@ -689,6 +701,7 @@ describe('agent-led new conversation', () => {
await waitFor(() => expect(screen.queryByRole('menu')).toBeNull());
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenLastCalledWith({
+ onFailure: expect.any(Function),
prompt: 'Prompt',
botId: CURSOR_BOT_ID,
repoUrl: undefined,
@@ -772,6 +785,7 @@ describe('agent-led new conversation', () => {
expect(screen.getByTestId('drawer').hasAttribute('hidden')).toBe(true);
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(send).toHaveBeenLastCalledWith({
+ onFailure: expect.any(Function),
prompt: 'Prompt',
botId: undefined,
repoUrl: undefined,
@@ -929,6 +943,7 @@ it('starts a conversation with an uploaded image and no text', () => {
fireEvent.click(screen.getByRole('button', { name: 'Send' }));
expect(start).toHaveBeenCalledWith(
expect.objectContaining({
+ onFailure: expect.any(Function),
prompt: '',
attachments: [
{
@@ -942,6 +957,57 @@ it('starts a conversation with an uploaded image and no text', () => {
);
});
+it('restores a text-only draft when startup configuration fails', () => {
+ mocks.attachments = [];
+ const start = page();
+ fireEvent.input(screen.getByRole('textbox', { name: 'Draft' }), {
+ target: { value: 'Keep my unsent prompt' },
+ });
+ fireEvent.click(screen.getByRole('button', { name: 'Send' }));
+ fireEvent.input(screen.getByRole('textbox', { name: 'Draft' }), {
+ target: { value: '' },
+ });
+ start.mock.calls[0][0].onFailure();
+ expect(
+ (screen.getByRole('textbox', { name: 'Draft' }) as HTMLInputElement).value
+ ).toBe('Keep my unsent prompt');
+});
+
+it('retains uploaded attachments until first delivery is accepted', () => {
+ mocks.attachments = [
+ {
+ id: 'retained-file',
+ name: 'keep.txt',
+ kind: 'document',
+ mimeType: 'text/plain',
+ size: 12,
+ },
+ ];
+ localStorage.setItem(
+ `${NEW_CONVERSATION_ATTACHMENTS_KEY}:user`,
+ JSON.stringify(mocks.attachments)
+ );
+ const start = page();
+ fireEvent.click(screen.getByRole('button', { name: 'Send' }));
+ expect(
+ JSON.parse(
+ localStorage.getItem(`${NEW_CONVERSATION_ATTACHMENTS_KEY}:user`)!
+ )
+ ).toEqual(mocks.attachments);
+ start.mock.calls[0][0].onFailure();
+ expect(
+ JSON.parse(
+ localStorage.getItem(`${NEW_CONVERSATION_ATTACHMENTS_KEY}:user`)!
+ )
+ ).toEqual(mocks.attachments);
+ start.mock.calls[0][0].onDelivered();
+ expect(
+ JSON.parse(
+ localStorage.getItem(`${NEW_CONVERSATION_ATTACHMENTS_KEY}:user`)!
+ )
+ ).toEqual([]);
+});
+
it('isolates drafts and attachment projections as identity arrives and changes', () => {
const [identity, setIdentity] = createSignal<string>();
mocks.userId = identity;
@@ -968,3 +1034,34 @@ it('isolates drafts and attachment projections as identity arrives and changes',
'Alice private'
);
});
+
+it('does not restore a failed first prompt into another account', () => {
+ const [identity, setIdentity] = createSignal<string | undefined>('alice');
+ mocks.userId = identity;
+ mocks.attachments = [];
+ const start = page();
+ fireEvent.input(screen.getByLabelText('Draft'), {
+ target: { value: 'Alice prompt' },
+ });
+ fireEvent.click(screen.getByRole('button', { name: 'Send' }));
+ setIdentity('bob');
+ start.mock.calls[0][0].onFailure();
+ expect((screen.getByLabelText('Draft') as HTMLInputElement).value).toBe('');
+});
+
+it('clears delivered attachments after switching away and back to the submitting account', () => {
+ const [identity, setIdentity] = createSignal<string | undefined>('alice');
+ mocks.userId = identity;
+ mocks.attachments = [{ id: 'sent-file', name: 'sent.txt', kind: 'document' }];
+ localStorage.setItem(
+ `${NEW_CONVERSATION_ATTACHMENTS_KEY}:alice`,
+ JSON.stringify(mocks.attachments)
+ );
+ const start = page();
+ fireEvent.click(screen.getByRole('button', { name: 'Send' }));
+ setIdentity('bob');
+ setIdentity('alice');
+ expect(screen.getByTestId('attachments').textContent).toBe('sent.txt');
+ start.mock.calls[0][0].onDelivered();
+ expect(screen.getByTestId('attachments').textContent).toBe('');
+});
diff --git a/apps/web/src/features/agents-view/views/NewChatPage.tsx b/apps/web/src/features/agents-view/views/NewChatPage.tsx
index bc12adb2e1..3ec7f821a9 100644
--- a/apps/web/src/features/agents-view/views/NewChatPage.tsx
+++ b/apps/web/src/features/agents-view/views/NewChatPage.tsx
@@ -6,6 +6,12 @@ import {
uploadInputAttachments,
} from '@channel/Input';
import { FloatRegionOrInline } from '@components/app/mobile/float-regions/FloatRegion';
+import { SpeedToggle } from '@core/component/AI/component/input/SpeedToggle';
+import {
+ acceleratedSpeed,
+ speedForModel,
+} from '@core/component/AI/constant/speed';
+import { fastModeEnabled } from '@core/component/AI/signal/speed';
import { useSettingsState } from '@core/constant/SettingsState';
import { useUserId } from '@core/context/user';
import { uploadFile } from '@core/util/upload';
@@ -47,6 +53,9 @@ export type StartConversation = {
/** The catalog was still loading, so an in-memory agent runs its persona default. */
modelFallback?: boolean;
effortOverride?: { configId: string; value: string };
+ speedOverride?: { configId: string; value: string };
+ onDelivered?: () => void;
+ onFailure?: () => void;
};
/** One agent choice determines the session kind, default model, and repository context. */
@@ -227,10 +236,26 @@ export function NewChatPage(props: {
const model = composerModelOverride();
const inMemory =
persona.harness === 'macro-inmem' || persona.harness === 'in-memory';
+ const submittedTracker = attachmentTracker();
+ const submittedUserId = userId();
props.onStart({
prompt,
+ onFailure: () => {
+ if (userId() === submittedUserId && !draft()) setDraft(prompt);
+ },
...(attachments.length > 0
- ? { attachments: promptActionOf(prompt, attachments).attachments }
+ ? {
+ attachments: promptActionOf(prompt, attachments).attachments,
+ onDelivered: () => {
+ const tracker =
+ userId() === submittedUserId
+ ? attachmentTracker()
+ : submittedTracker;
+ for (const attachment of attachments) {
+ tracker.removeAttachment(attachment.id);
+ }
+ },
+ }
: {}),
botId: persona.botId,
repoUrl: repo,
@@ -240,8 +265,15 @@ export function NewChatPage(props: {
? { modelFallback: true }
: {}),
effortOverride: effortOverride(),
+ ...(model && inMemory && acceleratedSpeed(model)
+ ? {
+ speedOverride: {
+ configId: 'speed',
+ value: speedForModel(model, fastModeEnabled()),
+ },
+ }
+ : {}),
});
- attachmentTracker().clearAttachments();
// Macro's preferred model stays; coding-agent submenu picks are one-shot.
setModelOverride(undefined);
setEffortSelection(undefined);
@@ -292,7 +324,18 @@ export function NewChatPage(props: {
draft={draft()}
onDraftChange={setDraft}
blockedReason={blocked()}
- selector={agentSelector()}
+ selector={
+ <div class="flex min-w-0 items-center gap-1">
+ {agentSelector()}
+ <Show
+ when={['macro-inmem', 'in-memory'].includes(
+ selected()?.harness ?? ''
+ )}
+ >
+ <SpeedToggle model={composerModelOverride() ?? ''} />
+ </Show>
+ </div>
+ }
drawer={
<RepositoryPicker
onOpenChange={setRepositoryPickerOpen}
diff --git a/apps/web/src/features/block-agent/component/AgentComposer.test.tsx b/apps/web/src/features/block-agent/component/AgentComposer.test.tsx
index 3e645bbf46..ff87ad55af 100644
--- a/apps/web/src/features/block-agent/component/AgentComposer.test.tsx
+++ b/apps/web/src/features/block-agent/component/AgentComposer.test.tsx
@@ -19,6 +19,8 @@ const mocks = vi.hoisted(() => ({
steerQueued: vi.fn(),
upload: vi.fn(),
consumeNotes: vi.fn(),
+ metadata: () =>
+ undefined as { model: string; configOptions: { id: string }[] } | undefined,
input: undefined as AgentInputProps | undefined,
model: undefined as AgentModelSelectorProps | undefined,
queued: undefined as QueuedPromptsProps | undefined,
@@ -48,7 +50,7 @@ vi.mock('../context/AgentSessionContext', () => ({
selectModel: mocks.selectModel,
loadFailed: () => false,
messages: () => [],
- metadata: () => undefined,
+ metadata: () => mocks.metadata(),
pending: () => false,
initialInput: 'Document context',
queue: {
@@ -87,6 +89,7 @@ beforeEach(() => {
localStorage.clear();
vi.resetAllMocks();
mocks.turn = () => 'idle';
+ mocks.metadata = () => undefined;
mocks.session = () => ({ canEdit: false });
mocks.issue.mockResolvedValue({ isErr: () => false });
});
@@ -194,3 +197,59 @@ it('explicitly sends the queue head and disables it while the turn transitions',
}
expect(mocks.sendNext).toHaveBeenCalledTimes(1);
});
+
+it('retains an unsent draft when speed configuration is rejected', async () => {
+ mocks.session = () => ({ canEdit: true });
+ mocks.metadata = () => ({
+ model: 'openai/gpt-6-astra',
+ configOptions: [{ id: 'speed' }],
+ });
+ mocks.selectModel.mockRejectedValue(new Error('Unavailable'));
+ render(() => <AgentComposer />);
+ mocks.input?.onDraftChange?.('Keep this draft');
+ mocks.input?.onDraftChange?.('');
+ mocks.input?.onSend('Keep this draft', []);
+ await vi.waitFor(() => expect(mocks.input?.draft).toBe('Keep this draft'));
+ expect(mocks.issue).not.toHaveBeenCalled();
+ expect(mocks.consumeNotes).not.toHaveBeenCalled();
+});
+
+it('leaves external agent speed settings alone', () => {
+ mocks.session = () => ({ canEdit: true });
+ mocks.metadata = () => ({
+ model: 'external/model',
+ configOptions: [{ id: 'speed' }],
+ });
+ render(() => <AgentComposer />);
+ mocks.input?.onSend('Use external settings', []);
+ expect(mocks.selectModel).not.toHaveBeenCalled();
+ expect(mocks.issue).toHaveBeenCalledWith({
+ type: 'prompt',
+ prompt: 'Use external settings',
+ });
+});
+
+it('retains the draft if editing access is revoked while applying speed', async () => {
+ const [canEdit, setCanEdit] = createSignal(true);
+ mocks.session = () => ({ canEdit: canEdit() });
+ mocks.metadata = () => ({
+ model: 'openai/gpt-6-astra',
+ configOptions: [{ id: 'speed' }],
+ });
+ let finish!: () => void;
+ mocks.selectModel.mockImplementation(
+ () =>
+ new Promise<void>((resolve) => {
+ finish = resolve;
+ })
+ );
+ render(() => <AgentComposer />);
+ mocks.input?.onSend('Keep after permission change', []);
+ setCanEdit(false);
+ finish();
+ await vi.waitFor(() =>
+ expect(mocks.input?.draft).toBe('Keep after permission change')
+ );
+ expect(mocks.issue).not.toHaveBeenCalled();
+ expect(mocks.consumeNotes).not.toHaveBeenCalled();
+});
diff --git a/apps/web/src/features/block-agent/component/AgentComposer.tsx b/apps/web/src/features/block-agent/component/AgentComposer.tsx
index 82b7e06f25..476fa2eba2 100644
--- a/apps/web/src/features/block-agent/component/AgentComposer.tsx
+++ b/apps/web/src/features/block-agent/component/AgentComposer.tsx
@@ -1,3 +1,9 @@
+import { SpeedToggle } from '@core/component/AI/component/input/SpeedToggle';
+import {
+ acceleratedSpeed,
+ speedForModel,
+} from '@core/component/AI/constant/speed';
+import { fastModeEnabled } from '@core/component/AI/signal/speed';
import ArrowUp from '@phosphor/arrow-up.svg';
import { Button } from '@ui';
/**
@@ -161,8 +167,32 @@ export function AgentComposer(props: {
// Attachments ride the prompt action itself, so they take the same path as
// the text: issued once, speculated by the fold, and queued server-side
// behind a running turn with the files still on them.
- const send = (markdown: string, attachments: InputAttachmentData[]) => {
- if (readOnly()) return;
+ const send = async (markdown: string, attachments: InputAttachmentData[]) => {
+ if (readOnly() || configuring()) return;
+ const meta = metadata();
+ if (
+ meta?.model &&
+ acceleratedSpeed(meta.model) &&
+ meta.configOptions.some((option) => option.id === 'speed')
+ ) {
+ setConfiguring(true);
+ try {
+ await selectModel(meta.model, {
+ configId: 'speed',
+ value: speedForModel(meta.model, fastModeEnabled()),
+ });
+ } catch {
+ persistedDraft.setDraft(markdown);
+ toast.failure('The speed setting could not be applied. Please retry.');
+ return;
+ } finally {
+ setConfiguring(false);
+ }
+ }
+ if (readOnly()) {
+ persistedDraft.setDraft(markdown);
+ return;
+ }
// Queued review notes ride this send: taking them here marks them sent
// before the prompt is issued, so a second Enter cannot post them again
// as their own queued prompt (which would then stop-and-flush).
@@ -304,9 +334,9 @@ export function AgentComposer(props: {
// Prompts go straight to the service, so sending needs a session to
// post to — a block whose create is still on the wire can be typed
// into, but not sent from, until the id lands.
- disabled={loadFailed() || pending() || readOnly()}
+ disabled={loadFailed() || pending() || readOnly() || configuring()}
commands={() => metadata()?.availableCommands ?? []}
- onSend={send}
+ onSend={(markdown, attachments) => void send(markdown, attachments)}
onStop={() =>
void act({ type: 'stop' }, 'The agent could not be stopped')
}
@@ -330,37 +360,46 @@ export function AgentComposer(props: {
}}
registerQuoteInsert={registerQuoteInsert}
modelControl={
- <ModelSelector
- model={metadata()?.model ?? null}
- changingTo={changingModel(messages(), metadata()?.model ?? null)}
- options={metadata()?.supportedModels ?? []}
- effortLabel={effortLabel(effort(), changingEffort())}
- disabled={
- loadFailed() ||
- readOnly() ||
- pending() ||
- configuring() ||
- changingEffort() !== undefined
- }
- onSelect={(model) => void chooseModel(model)}
- modelRow={(row) => (
- <AgentModelMenuItem
- {...row}
- harness={session()?.harness}
- effort={
- row.option.id === metadata()?.model ? effort() : undefined
- }
- effortValue={
- row.option.id === metadata()?.model
- ? changingEffort()
- : undefined
- }
- onSelectEffort={(selection) =>
- void chooseModel(row.option.id, selection)
- }
- />
- )}
- />
+ <div class="flex min-w-0 items-center gap-1">
+ <ModelSelector
+ model={metadata()?.model ?? null}
+ changingTo={changingModel(messages(), metadata()?.model ?? null)}
+ options={metadata()?.supportedModels ?? []}
+ effortLabel={effortLabel(effort(), changingEffort())}
+ disabled={
+ loadFailed() ||
+ readOnly() ||
+ pending() ||
+ configuring() ||
+ changingEffort() !== undefined
+ }
+ onSelect={(model) => void chooseModel(model)}
+ modelRow={(row) => (
+ <AgentModelMenuItem
+ {...row}
+ harness={session()?.harness}
+ effort={
+ row.option.id === metadata()?.model ? effort() : undefined
+ }
+ effortValue={
+ row.option.id === metadata()?.model
+ ? changingEffort()
+ : undefined
+ }
+ onSelectEffort={(selection) =>
+ void chooseModel(row.option.id, selection)
+ }
+ />
+ )}
+ />
+ <Show
+ when={metadata()?.configOptions.some(
+ (option) => option.id === 'speed'
+ )}
+ >
+ <SpeedToggle model={metadata()?.model ?? ''} />
+ </Show>
+ </div>
}
/>
</>
diff --git a/apps/web/src/features/block-agent/context/configure-session-model.test.ts b/apps/web/src/features/block-agent/context/configure-session-model.test.ts
index 8e8451fe10..9046d5409d 100644
--- a/apps/web/src/features/block-agent/context/configure-session-model.test.ts
+++ b/apps/web/src/features/block-agent/context/configure-session-model.test.ts
@@ -119,3 +119,28 @@ describe('combined model and effort selection', () => {
await result;
});
});
+
+it('confirms speed independently of reasoning effort before the caller sends a prompt', async () => {
+ const f = fixture();
+ f.snapshot.metadata.configOptions.push({
+ ...f.config,
+ id: 'speed',
+ currentValue: 'standard',
+ options: [{ value: 'standard' }, { value: 'ultrafast' }],
+ });
+ let finished = false;
+ const result = configureSessionModel(f.session, undefined, {
+ configId: 'speed',
+ value: 'ultrafast',
+ }).then(() => {
+ finished = true;
+ });
+ await flush();
+ expect(f.issue.mock.calls).toEqual([
+ [{ type: 'setConfigOption', configId: 'speed', value: 'ultrafast' }],
+ ]);
+ expect(finished).toBe(false);
+ f.confirm(1);
+ await result;
+ expect(finished).toBe(true);
+});
diff --git a/apps/web/src/features/block-agent/context/configure-session-model.ts b/apps/web/src/features/block-agent/context/configure-session-model.ts
index 971a93b43a..5629047bd5 100644
--- a/apps/web/src/features/block-agent/context/configure-session-model.ts
+++ b/apps/web/src/features/block-agent/context/configure-session-model.ts
@@ -54,14 +54,21 @@ export async function configureSessionModel(
await confirmSessionControl(session, { type: 'setModel', model });
}
if (!effort) return;
- const option = effortConfigOption(
- (await session.snapshot()).metadata.configOptions
- );
+ const options = (await session.snapshot()).metadata.configOptions;
+ const candidate = options.find((option) => option.id === 'speed');
+ const option =
+ effort.configId === 'speed'
+ ? candidate?.type === 'select'
+ ? candidate
+ : undefined
+ : effortConfigOption(options);
if (
option?.id !== effort.configId ||
!option.options.some((choice) => choice.value === effort.value)
) {
- throw new Error('The selected model no longer supports this effort.');
+ throw new Error(
+ `The selected model no longer supports this ${effort.configId === 'speed' ? 'speed' : 'effort'}.`
+ );
}
if (option.currentValue === effort.value) return;
await confirmSessionControl(session, {
diff --git a/apps/web/src/features/block-agent/context/pending-session.test.tsx b/apps/web/src/features/block-agent/context/pending-session.test.tsx
index 1778bdd75c..f9376fa1e4 100644
--- a/apps/web/src/features/block-agent/context/pending-session.test.tsx
+++ b/apps/web/src/features/block-agent/context/pending-session.test.tsx
@@ -306,3 +306,49 @@ describe('pending navigation ownership', () => {
expect(fold.closeSession).not.toHaveBeenCalled();
});
});
+
+it('retains submitting context until the first prompt is accepted', async () => {
+ const delivery = deferred<ReturnType<typeof accepted>>();
+ harness.control.mockReturnValue(delivery.promise);
+ const onDelivered = vi.fn();
+ const onFailure = vi.fn();
+ startPendingSession({ prompt: 'Hello', onDelivered, onFailure });
+ await settle();
+ expect(onDelivered).not.toHaveBeenCalled();
+ delivery.resolve(accepted(harness.control.mock.calls[0][1]));
+ await settle();
+ expect(onDelivered).toHaveBeenCalledOnce();
+ expect(onFailure).not.toHaveBeenCalled();
+});
+
+it('restores submitting context when speed configuration fails', async () => {
+ fold.readSession.mockResolvedValue({
+ messages: [],
+ metadata: {
+ turn: 'idle',
+ model: 'openai/gpt-6-astra',
+ configOptions: [
+ {
+ id: 'speed',
+ type: 'select',
+ currentValue: 'standard',
+ options: [{ value: 'standard' }],
+ },
+ ],
+ },
+ });
+ const onDelivered = vi.fn();
+ const onFailure = vi.fn();
+ const id = startPendingSession({
+ prompt: 'Keep this draft',
+ modelOverride: 'openai/gpt-6-astra',
+ speedOverride: { configId: 'speed', value: 'ultrafast' },
+ onDelivered,
+ onFailure,
+ });
+ await settle();
+ expect(pendingSession(id)?.failed()).toBe(true);
+ expect(onDelivered).not.toHaveBeenCalled();
+ expect(onFailure).toHaveBeenCalledOnce();
+ expect(harness.control).not.toHaveBeenCalled();
+});
diff --git a/apps/web/src/features/block-agent/context/pending-session.ts b/apps/web/src/features/block-agent/context/pending-session.ts
index f2ce7b1641..ff3ad4fdbe 100644
--- a/apps/web/src/features/block-agent/context/pending-session.ts
+++ b/apps/web/src/features/block-agent/context/pending-session.ts
@@ -99,6 +99,12 @@ export type StartPendingSessionOptions = {
instructions?: string;
/** Opaque harness setting confirmed before the first prompt. */
effortOverride?: { configId: string; value: string };
+ /** Inference speed confirmed before the first prompt. */
+ speedOverride?: { configId: string; value: string };
+ /** Release the submitting composer's attachments after the first prompt is accepted. */
+ onDelivered?: () => void;
+ /** Restore the submitting draft when creation, configuration, or first delivery fails. */
+ onFailure?: () => void;
/**
* Explicit GitHub repository for the managed Cursor session.
*/
@@ -147,6 +153,7 @@ export function startPendingSession(
trace?.end('failed', cause ?? new Error(message));
releaseNavigation();
setError(message);
+ options.onFailure?.();
};
const expiry = setTimeout(
() => forgetPendingSession(id),
@@ -211,10 +218,15 @@ export function startPendingSession(
// an effort holds the block in preflight. Then adopt the session before
// issuing the first prompt so that prompt is folded speculatively while
// its POST is in flight.
- if (options.effortOverride || prompt || options.attachments?.length) {
+ if (
+ options.effortOverride ||
+ options.speedOverride ||
+ prompt ||
+ options.attachments?.length
+ ) {
const session = AgentSession.acquire(created);
try {
- if (options.effortOverride) {
+ if (options.effortOverride || options.speedOverride) {
await session.load();
trace?.stage('loaded');
await sessionConfigReported(session);
@@ -223,6 +235,13 @@ export function startPendingSession(
options.modelOverride,
options.effortOverride
);
+ if (options.speedOverride) {
+ await configureSessionModel(
+ session,
+ options.modelOverride,
+ options.speedOverride
+ );
+ }
trace?.stage('configured');
}
// The prompt's reference ends when its POST answers. Navigation
@@ -246,11 +265,12 @@ export function startPendingSession(
}
);
if (delivered.isErr()) {
- releaseNavigation();
- setError(
+ fail(
delivered.error.map((error) => error.message).join(' ') ||
'The first message could not be sent.'
);
+ } else {
+ options.onDelivered?.();
}
}
} catch (error) {
diff --git a/apps/web/src/features/block-chat/component/Chat.tsx b/apps/web/src/features/block-chat/component/Chat.tsx
index 268332bdda..838fef689f 100644
--- a/apps/web/src/features/block-chat/component/Chat.tsx
+++ b/apps/web/src/features/block-chat/component/Chat.tsx
@@ -284,6 +284,7 @@ function ChatInner(props: {
onSend({
content: pendingSend.content,
model: pendingSend.model ?? input.model(),
+ speed: pendingSend.speed,
attachments: pendingSend.attachments ?? [],
toolset: { type: 'all' },
});
diff --git a/apps/web/src/features/chat/SoupChatInput.test.tsx b/apps/web/src/features/chat/SoupChatInput.test.tsx
index 89f7fb5c36..39f76346d7 100644
--- a/apps/web/src/features/chat/SoupChatInput.test.tsx
+++ b/apps/web/src/features/chat/SoupChatInput.test.tsx
@@ -16,6 +16,9 @@ const mocks = vi.hoisted(() => ({
rename: vi.fn(),
storeModel: vi.fn(),
sendBackground: vi.fn(),
+ setMarkdown: vi.fn(),
+ setAttached: vi.fn(),
+ failureToast: vi.fn(),
background: false,
agentsEnabled: false,
agentsLoading: (): boolean => false,
@@ -38,7 +41,7 @@ vi.mock('@core/component/AI/component/input/buildChatEditor', () => ({
buildChatEditor: () => {
const builder = {
withAppLinkResolver: () => builder,
- withMentions: () => ({}),
+ withMentions: () => ({ controls: { setMarkdown: mocks.setMarkdown } }),
};
return builder;
},
@@ -69,7 +72,7 @@ vi.mock('@core/component/AI/context', () => ({
ChatInputProvider: (props: { children: unknown }) => props.children,
useChatInputContext: () => ({
model: () => Model.gpt56,
- attachments: {},
+ attachments: { setAttached: mocks.setAttached },
}),
}));
vi.mock('@core/component/AI/signal/attachment', () => ({
@@ -108,10 +111,15 @@ vi.mock('@service-cognition/client', () => ({
},
}));
+vi.mock('@core/component/Toast/Toast', () => ({
+ toast: { failure: mocks.failureToast },
+}));
+
import { SoupChatInput } from './SoupChatInput';
beforeEach(() => {
vi.clearAllMocks();
+ mocks.sendBackground.mockResolvedValue({ isErr: () => false });
mocks.background = false;
mocks.agentsEnabled = false;
mocks.agentsLoading = () => false;
@@ -205,3 +213,22 @@ describe('SoupChatInput', () => {
expect(mocks.storeModel).not.toHaveBeenCalled();
});
});
+
+it('restores a rejected background draft and reports the failure', async () => {
+ mocks.background = true;
+ mocks.createChat.mockResolvedValue({
+ isErr: () => false,
+ value: { id: 'new-chat' },
+ });
+ mocks.sendBackground.mockResolvedValue({ isErr: () => true });
+ render(() => <SoupChatInput />);
+ fireEvent.click(screen.getByRole('button', { name: 'Send' }));
+ await waitFor(() =>
+ expect(mocks.setMarkdown).toHaveBeenCalledWith('Summarize this document')
+ );
+ expect(mocks.setAttached).toHaveBeenCalledWith([
+ { entity_id: 'document-id', entity_type: 'document' },
+ ]);
+ expect(mocks.failureToast).toHaveBeenCalledOnce();
+ expect(mocks.replace).not.toHaveBeenCalled();
+});
diff --git a/apps/web/src/features/chat/SoupChatInput.tsx b/apps/web/src/features/chat/SoupChatInput.tsx
index a6950f8ab1..f3ad2e17af 100644
--- a/apps/web/src/features/chat/SoupChatInput.tsx
+++ b/apps/web/src/features/chat/SoupChatInput.tsx
@@ -17,6 +17,7 @@ import {
storeChatStateImmediate,
storeSoupInputModel,
} from '@core/component/AI/util/storage';
+import { toast } from '@core/component/Toast/Toast';
import { enableChatV3Agents } from '@core/constant/featureFlags';
import { PaywallKey, usePaywallState } from '@core/constant/PaywallState';
import { registerHotkey, useHotkeyDOMScope } from '@core/hotkey/hotkeys';
@@ -97,14 +98,23 @@ function SoupChatInputInner() {
if (backgroundSend) {
// Send the message in the background without navigating
- cognitionApiServiceClient.sendStreamChatMessage({
+ const sent = await cognitionApiServiceClient.sendStreamChatMessage({
content: request.content,
model: request.model,
+ speed: request.speed,
chat_id: chatId,
attachments:
request.attachments.length > 0 ? request.attachments : undefined,
toolset: { type: 'all' },
});
+ if (sent.isErr()) {
+ input.attachments.setAttached(request.attachments);
+ editor.controls.setMarkdown(request.content);
+ toast.failure(
+ 'Could not send the message. Your draft has been restored.'
+ );
+ return;
+ }
invalidateAllSoup();
} else {
// Store the pending send data for the chat to pick up
@@ -112,6 +122,7 @@ function SoupChatInputInner() {
content: request.content,
attachments: request.attachments,
model: request.model,
+ speed: request.speed,
});
// Replace the soup split with the chat split
diff --git a/apps/web/src/features/home/home-chat-input.test.tsx b/apps/web/src/features/home/home-chat-input.test.tsx
index 0d7c7b16c2..8a7bf0c2c1 100644
--- a/apps/web/src/features/home/home-chat-input.test.tsx
+++ b/apps/web/src/features/home/home-chat-input.test.tsx
@@ -91,6 +91,7 @@ const request: ChatSendInput = {
beforeEach(() => {
vi.resetAllMocks();
+ mocks.sendStreamChatMessage.mockResolvedValue(ok({}));
mocks.agentsEnabled = false;
// ChatInput clears the editor and attachments before invoking onSend.
mocks.draft = '';
@@ -207,3 +208,14 @@ describe('Home Agent V3 mounting', () => {
expect(document.querySelector('[data-testid="agent-composer"]')).toBeNull();
});
});
+
+it('restores a rejected background draft with its attachments', async () => {
+ mocks.createChat.mockResolvedValue(ok({ id: 'new-chat' }));
+ mocks.sendStreamChatMessage.mockResolvedValue(
+ err([{ code: 'HTTP_ERROR', message: 'Speed unavailable' }])
+ );
+ await mocks.onSend?.({ ...request, metaKey: true, speed: 'fast' });
+ expect(mocks.draft).toBe(request.content);
+ expect(mocks.attached).toEqual(request.attachments);
+ expect(mocks.invalidateSoup).not.toHaveBeenCalled();
+});
diff --git a/apps/web/src/features/home/home-chat-input.tsx b/apps/web/src/features/home/home-chat-input.tsx
index 23a7880695..b12b67ffd2 100644
--- a/apps/web/src/features/home/home-chat-input.tsx
+++ b/apps/web/src/features/home/home-chat-input.tsx
@@ -139,14 +139,19 @@ export const LegacyHomeChatInput = (props: HomeChatInputProps) => {
if (backgroundSend) {
// Send the message in the background without navigating
- cognitionApiServiceClient.sendStreamChatMessage({
+ const sent = await cognitionApiServiceClient.sendStreamChatMessage({
content: request.content,
model: request.model,
+ speed: request.speed,
chat_id: chatId,
attachments:
request.attachments.length > 0 ? request.attachments : undefined,
toolset: request.toolset,
});
+ if (sent.isErr()) {
+ restoreDraft(request);
+ return;
+ }
invalidateAllSoup();
} else {
// Store the pending send data for the chat to pick up
@@ -154,6 +159,7 @@ export const LegacyHomeChatInput = (props: HomeChatInputProps) => {
content: request.content,
attachments: request.attachments,
model: request.model,
+ speed: request.speed,
});
if (props.openChat) {
diff --git a/apps/web/src/lib/core/component/AI/component/input/ChatInput.tsx b/apps/web/src/lib/core/component/AI/component/input/ChatInput.tsx
index d4358d22d6..b8a8ea22e3 100644
--- a/apps/web/src/lib/core/component/AI/component/input/ChatInput.tsx
+++ b/apps/web/src/lib/core/component/AI/component/input/ChatInput.tsx
@@ -36,7 +36,10 @@ import { createElementSize } from '@solid-primitives/resize-observer';
import { createCallback } from '@solid-primitives/rootless';
import { Button, ComposerSurface, cn, SendButton as UiSendButton } from '@ui';
import { createEffect, createMemo, createSignal, Show } from 'solid-js';
+import { speedForModel } from '../../constant/speed';
+import { fastModeEnabled } from '../../signal/speed';
import { AttachmentList } from './Attachment';
+import { SpeedToggle } from './SpeedToggle';
import { useAiDataConsentGate } from './useAiDataConsent';
/**
@@ -200,6 +203,7 @@ export function ChatInput(props: ChatInputComponentProps) {
const sendInput: ChatSendInput = {
content: markdownText(),
model: opts?.modelOverride ?? model(),
+ speed: speedForModel(opts?.modelOverride ?? model(), fastModeEnabled()),
attachments: attachments.attached(),
toolset: toolsetSignal[0](),
metaKey: opts?.metaKey,
@@ -318,6 +322,7 @@ export function ChatInput(props: ChatInputComponentProps) {
onLocked={() => showPaywall(PaywallKey.O1_LIMIT)}
compact={compactSelector()}
/>
+ <SpeedToggle model={model()} />
<DictationButton dictation={dictation} />
<Show when={generating() && props.onStop} fallback={<SendButton />}>
<StopButton />
@@ -433,6 +438,7 @@ export function ChatInput(props: ChatInputComponentProps) {
/>
<div class="size-7 not-touch:size-[33.75px] shrink-0" />
<div class="size-7 not-touch:size-[33.75px] shrink-0" />
+ <div class="size-7 not-touch:size-[33.75px] shrink-0" />
</div>
</Show>
<div
diff --git a/apps/web/src/lib/core/component/AI/component/input/ModelSelector.tsx b/apps/web/src/lib/core/component/AI/component/input/ModelSelector.tsx
index 83e1b435c3..8707a6f56a 100644
--- a/apps/web/src/lib/core/component/AI/component/input/ModelSelector.tsx
+++ b/apps/web/src/lib/core/component/AI/component/input/ModelSelector.tsx
@@ -41,6 +41,7 @@ const MODEL_DESCRIPTION: Record<TModel, string> = {
[Model.opus55]: 'Complex tasks and deeper analysis',
[Model.haiku45]: 'Quick answers and lighter tasks',
[Model.gpt6Astra]: 'Frontier reasoning for the hardest problems',
+ [Model.gpt61Sol]: 'Balanced reasoning, coding, and speed',
[Model.gpt56]: 'Reasoning, writing, and problem solving',
[Model.gpt56Mini]: 'Fast help with everyday tasks',
[Model.gemini38Flash]: 'Fast answers over data and documents',
diff --git a/apps/web/src/lib/core/component/AI/component/input/SpeedToggle.test.tsx b/apps/web/src/lib/core/component/AI/component/input/SpeedToggle.test.tsx
new file mode 100644
index 0000000000..832c91cdcb
--- /dev/null
+++ b/apps/web/src/lib/core/component/AI/component/input/SpeedToggle.test.tsx
@@ -0,0 +1,66 @@
+import { cleanup, fireEvent, render, screen } from '@solidjs/testing-library';
+import type { JSX } from 'solid-js';
+import { afterEach, beforeEach, expect, it, vi } from 'vitest';
+import { Model } from '../../constant/model';
+import { speedForModel } from '../../constant/speed';
+import { setFastModeEnabled } from '../../signal/speed';
+import { SpeedToggle } from './SpeedToggle';
+
+vi.mock('@ui', () => ({
+ cn: (...classes: unknown[]) => classes.filter(Boolean).join(' '),
+ Button: (props: JSX.ButtonHTMLAttributes<HTMLButtonElement>) => (
+ <button {...props} />
+ ),
+}));
+const animate = vi.fn(() => ({ cancel: vi.fn() }));
+const originalAnimate = Object.getOwnPropertyDescriptor(
+ HTMLElement.prototype,
+ 'animate'
+);
+beforeEach(() => {
+ setFastModeEnabled(false);
+ vi.stubGlobal(
+ 'matchMedia',
+ vi.fn(() => ({ matches: false }))
+ );
+ Object.defineProperty(HTMLElement.prototype, 'animate', {
+ configurable: true,
+ value: animate,
+ });
+ animate.mockClear();
+});
+afterEach(() => {
+ cleanup();
+ vi.unstubAllGlobals();
+ if (originalAnimate)
+ Object.defineProperty(HTMLElement.prototype, 'animate', originalAnimate);
+ else Reflect.deleteProperty(HTMLElement.prototype, 'animate');
+});
+
+it('animates the bolt and retains its paid-speed preference across mounts', () => {
+ const view = render(() => <SpeedToggle model={Model.gpt6Astra} />);
+ fireEvent.click(screen.getByRole('button'));
+ expect(screen.getByRole('button').getAttribute('aria-pressed')).toBe('true');
+ expect(localStorage.getItem('macro:ai-fast-mode')).toBe('true');
+ expect(animate).toHaveBeenCalledOnce();
+ view.unmount();
+ render(() => <SpeedToggle model={Model.opus55} />);
+ expect(screen.getByRole('button').getAttribute('aria-label')).toContain(
+ 'Fast mode on · 2×'
+ );
+ expect(speedForModel(Model.opus55, true)).toBe('fast');
+});
+
+it('leaves unsupported models at standard speed and respects reduced motion', () => {
+ vi.stubGlobal(
+ 'matchMedia',
+ vi.fn(() => ({ matches: true }))
+ );
+ const view = render(() => <SpeedToggle model={Model.gpt6Astra} />);
+ fireEvent.click(screen.getByRole('button'));
+ expect(animate).not.toHaveBeenCalled();
+ view.unmount();
+ render(() => <SpeedToggle model={Model.gemini38Flash} />);
+ expect(screen.queryByRole('button')).toBeNull();
+ expect(speedForModel(Model.gemini38Flash, true)).toBe('standard');
+});
diff --git a/apps/web/src/lib/core/component/AI/component/input/SpeedToggle.tsx b/apps/web/src/lib/core/component/AI/component/input/SpeedToggle.tsx
new file mode 100644
index 0000000000..91d7ad89eb
--- /dev/null
+++ b/apps/web/src/lib/core/component/AI/component/input/SpeedToggle.tsx
@@ -0,0 +1,58 @@
+import LightningIcon from '@phosphor-fill/lightning-fill.svg';
+import { Button, cn } from '@ui';
+import { onCleanup, Show } from 'solid-js';
+import { acceleratedSpeed } from '../../constant/speed';
+import { fastModeEnabled, setFastModeEnabled } from '../../signal/speed';
+
+/** An explicit paid-speed preference. Changing models never enables unsupported tiers. */
+export function SpeedToggle(props: { model: string }) {
+ const speed = () => acceleratedSpeed(props.model);
+ const enabled = () => !!speed() && fastModeEnabled();
+ const label = () => {
+ if (!speed()) return 'Fast mode is unavailable for this model';
+ const name = speed() === 'ultrafast' ? 'Ultrafast' : 'Fast';
+ const multiplier = speed() === 'ultrafast' ? '6×' : '2×';
+ return `${name} mode ${enabled() ? 'on' : 'off'} · ${multiplier} AI usage`;
+ };
+ let bolt: HTMLSpanElement | undefined;
+ let animation: Animation | undefined;
+ onCleanup(() => animation?.cancel());
+
+ function toggle() {
+ if (!speed()) return;
+ setFastModeEnabled(!fastModeEnabled());
+ animation?.cancel();
+ if (window.matchMedia?.('(prefers-reduced-motion: reduce)').matches) return;
+ animation = bolt?.animate?.(
+ [
+ { transform: 'scale(1) rotate(0deg)', opacity: 1 },
+ { transform: 'scale(0.65) rotate(-18deg)', opacity: 0.5, offset: 0.2 },
+ { transform: 'scale(1.35) rotate(12deg)', opacity: 1, offset: 0.55 },
+ { transform: 'scale(1) rotate(0deg)', opacity: 1 },
+ ],
+ { duration: 360, easing: 'cubic-bezier(0.22, 1, 0.36, 1)' }
+ );
+ }
+
+ return (
+ <Show when={speed()}>
+ <Button
+ variant="ghost"
+ size="icon-composer"
+ label={label()}
+ aria-label={label()}
+ aria-pressed={enabled()}
+ onMouseDown={(event) => event.preventDefault()}
+ onClick={toggle}
+ class={cn(
+ 'shrink-0 transition-colors motion-reduce:transition-none',
+ enabled() ? 'text-accent bg-accent/10' : 'text-ink-muted'
+ )}
+ >
+ <span ref={bolt} class="flex items-center justify-center">
+ <LightningIcon class="size-4" />
+ </span>
+ </Button>
+ </Show>
+ );
+}
diff --git a/apps/web/src/lib/core/component/AI/component/input/buildRequest.ts b/apps/web/src/lib/core/component/AI/component/input/buildRequest.ts
index 4fd792ecd7..6c81626a32 100644
--- a/apps/web/src/lib/core/component/AI/component/input/buildRequest.ts
+++ b/apps/web/src/lib/core/component/AI/component/input/buildRequest.ts
@@ -7,10 +7,12 @@ import { isPaymentError } from '@core/util/handlePaymentError';
import { cognitionApiServiceClient } from '@service-cognition/client';
import type { ChatMessageStream } from '@service-connection/stream';
import { subscribe } from '@service-connection/stream';
+import type { ModelSpeed } from '../../constant/speed';
export type ChatSendInput = {
content: string;
model: Model;
+ speed?: ModelSpeed;
attachments: Attachment[];
toolset: ToolSet;
metaKey?: boolean;
@@ -31,6 +33,7 @@ export function useSendChatMessage() {
return async function sendChatMessage({
content,
model,
+ speed,
chatId,
attachments,
toolset,
@@ -38,6 +41,7 @@ export function useSendChatMessage() {
const response = await cognitionApiServiceClient.sendStreamChatMessage({
content,
model: model ?? DEFAULT_MODEL,
+ speed: speed ?? 'standard',
chat_id: chatId,
attachments: attachments.length > 0 ? attachments : undefined,
toolset,
diff --git a/apps/web/src/lib/core/component/AI/constant/model.ts b/apps/web/src/lib/core/component/AI/constant/model.ts
index d76ec17434..9c8dd41388 100644
--- a/apps/web/src/lib/core/component/AI/constant/model.ts
+++ b/apps/web/src/lib/core/component/AI/constant/model.ts
@@ -18,6 +18,7 @@ export const Model = {
opus55: 'anthropic/claude-opus-5-5',
haiku45: 'anthropic/claude-haiku-4-5',
gpt6Astra: 'openai/gpt-6-astra',
+ gpt61Sol: 'openai/gpt-6.1-sol',
gpt56: 'openai/gpt-5.6',
gpt56Mini: 'openai/gpt-5.6-mini',
gemini38Flash: 'google/gemini-3.8-flash',
@@ -37,6 +38,7 @@ export const MODEL_PRETTYNAME: ExhaustiveMap = {
'anthropic/claude-opus-5-5': 'Opus 5.5',
'anthropic/claude-haiku-4-5': 'Haiku 4.5',
'openai/gpt-6-astra': 'GPT-6 Astra',
+ 'openai/gpt-6.1-sol': 'GPT-6.1 Sol',
'openai/gpt-5.6': 'GPT-5.6',
'openai/gpt-5.6-mini': 'GPT-5.6 mini',
'google/gemini-3.8-flash': 'Gemini 3.8 Flash',
@@ -47,6 +49,7 @@ export const MODEL_PROVIDER_ICON: ExhaustiveMap = {
'anthropic/claude-opus-5-5': ClaudeIcon,
'anthropic/claude-haiku-4-5': ClaudeIcon,
'openai/gpt-6-astra': OpenAiIcon,
+ 'openai/gpt-6.1-sol': OpenAiIcon,
'openai/gpt-5.6': OpenAiIcon,
'openai/gpt-5.6-mini': OpenAiIcon,
'google/gemini-3.8-flash': GoogleIcon,
@@ -101,6 +104,7 @@ export const MODEL_PROVIDER: ExhaustiveMap = {
'anthropic/claude-opus-5-5': 'anthropic',
'anthropic/claude-haiku-4-5': 'anthropic',
'openai/gpt-6-astra': 'openai',
+ 'openai/gpt-6.1-sol': 'openai',
'openai/gpt-5.6': 'openai',
'openai/gpt-5.6-mini': 'openai',
'google/gemini-3.8-flash': 'google',
diff --git a/apps/web/src/lib/core/component/AI/constant/speed.ts b/apps/web/src/lib/core/component/AI/constant/speed.ts
new file mode 100644
index 0000000000..1ed8624e6f
--- /dev/null
+++ b/apps/web/src/lib/core/component/AI/constant/speed.ts
@@ -0,0 +1,16 @@
+import { Model } from './model';
+
+export type ModelSpeed = 'standard' | 'fast' | 'ultrafast';
+
+/** Only advertise speed tiers supported by the provider for this exact model. */
+export function acceleratedSpeed(
+ model: string
+): Exclude<ModelSpeed, 'standard'> | undefined {
+ if (model === Model.gpt6Astra || model === Model.gpt61Sol) return 'ultrafast';
+ if (model === Model.opus55) return 'fast';
+ return undefined;
+}
+
+export function speedForModel(model: string, enabled: boolean): ModelSpeed {
+ return enabled ? (acceleratedSpeed(model) ?? 'standard') : 'standard';
+}
diff --git a/apps/web/src/lib/core/component/AI/signal/pendingSend.ts b/apps/web/src/lib/core/component/AI/signal/pendingSend.ts
index b27998532a..5d98b4f87a 100644
--- a/apps/web/src/lib/core/component/AI/signal/pendingSend.ts
+++ b/apps/web/src/lib/core/component/AI/signal/pendingSend.ts
@@ -1,10 +1,12 @@
import type { Attachment, Model } from '@core/component/AI/types';
import { createSignal } from 'solid-js';
+import type { ModelSpeed } from '../constant/speed';
export type PendingSend = {
content: string;
attachments: Attachment[];
model: Model;
+ speed?: ModelSpeed;
};
const [pendingSend, setPendingSend] = createSignal<PendingSend | null>(null);
diff --git a/apps/web/src/lib/core/component/AI/signal/speed.ts b/apps/web/src/lib/core/component/AI/signal/speed.ts
new file mode 100644
index 0000000000..0e7ef178a8
--- /dev/null
+++ b/apps/web/src/lib/core/component/AI/signal/speed.ts
@@ -0,0 +1,11 @@
+import { makePersisted } from '@solid-primitives/storage';
+import { createSignal } from 'solid-js';
+
+/** Browser preference shared by all composers; unsupported models still use standard speed. */
+export const [fastModeEnabled, setFastModeEnabled] = makePersisted(
+ createSignal(false),
+ {
+ name: 'macro:ai-fast-mode',
+ deserialize: (value) => value === 'true',
+ }
+);
diff --git a/apps/web/src/lib/service-clients/service-cognition/generated/schemas/httpSendChatMessageRequest.ts b/apps/web/src/lib/service-clients/service-cognition/generated/schemas/httpSendChatMessageRequest.ts
index 4419e98fc6..dc59aaf541 100644
--- a/apps/web/src/lib/service-clients/service-cognition/generated/schemas/httpSendChatMessageRequest.ts
+++ b/apps/web/src/lib/service-clients/service-cognition/generated/schemas/httpSendChatMessageRequest.ts
@@ -7,6 +7,7 @@
import type { HttpSendChatMessageRequestAdditionalInstructions } from './httpSendChatMessageRequestAdditionalInstructions';
import type { HttpSendChatMessageRequestAttachments } from './httpSendChatMessageRequestAttachments';
import type { HttpSendChatMessageRequestChatId } from './httpSendChatMessageRequestChatId';
+import type { ModelSpeed } from './modelSpeed';
import type { ToolSet } from './toolSet';
/**
@@ -24,6 +25,8 @@ export interface HttpSendChatMessageRequest {
content: string;
/** The model to respond with (`provider/model` id) */
model: string;
+ /** Inference speed, validated against the selected model. Defaults to standard. */
+ speed?: ModelSpeed;
/** Which toolset to use. Defaults to `all` */
toolset?: ToolSet;
}
diff --git a/apps/web/src/lib/service-clients/service-cognition/generated/schemas/index.ts b/apps/web/src/lib/service-clients/service-cognition/generated/schemas/index.ts
index e12c4b8748..e71ba734d8 100644
--- a/apps/web/src/lib/service-clients/service-cognition/generated/schemas/index.ts
+++ b/apps/web/src/lib/service-clients/service-cognition/generated/schemas/index.ts
@@ -156,6 +156,7 @@ export * from './mcpAuthCallbackParams';
export * from './memoryErrorBody';
export * from './memoryResponse';
export * from './messageWithAttachments';
+export * from './modelSpeed';
export * from './newAttachment';
export * from './newChatMessage';
export * from './newChatMessageAttachments';
diff --git a/apps/web/src/lib/service-clients/service-cognition/generated/schemas/modelSpeed.ts b/apps/web/src/lib/service-clients/service-cognition/generated/schemas/modelSpeed.ts
new file mode 100644
index 0000000000..c42c799b9f
--- /dev/null
+++ b/apps/web/src/lib/service-clients/service-cognition/generated/schemas/modelSpeed.ts
@@ -0,0 +1,18 @@
+/**
+ * Generated by orval v7.21.0 🍺
+ * Do not edit manually.
+ * Document Cognition Service
+ * OpenAPI spec version: 1.0.0
+ */
+
+/**
+ * The requested inference speed. Standard is backward compatible with old clients.
+ */
+export type ModelSpeed = (typeof ModelSpeed)[keyof typeof ModelSpeed];
+
+// eslint-disable-next-line @typescript-eslint/no-redeclare
+export const ModelSpeed = {
+ standard: 'standard',
+ fast: 'fast',
+ ultrafast: 'ultrafast',
+} as const;
diff --git a/apps/web/src/lib/service-clients/service-cognition/openapi.json b/apps/web/src/lib/service-clients/service-cognition/openapi.json
index cb22dea994..91fcefa5ab 100644
--- a/apps/web/src/lib/service-clients/service-cognition/openapi.json
+++ b/apps/web/src/lib/service-clients/service-cognition/openapi.json
@@ -3761,6 +3761,10 @@
"type": "string",
"description": "The model to respond with (`provider/model` id)"
},
+ "speed": {
+ "$ref": "#/components/schemas/ModelSpeed",
+ "description": "Inference speed, validated against the selected model. Defaults to standard."
+ },
"toolset": {
"$ref": "#/components/schemas/ToolSet",
"description": "Which toolset to use. Defaults to `all`"
@@ -4003,6 +4007,11 @@
}
}
},
+ "ModelSpeed": {
+ "type": "string",
+ "description": "The requested inference speed. Standard is backward compatible with old clients.",
+ "enum": ["standard", "fast", "ultrafast"]
+ },
"NewAttachment": {
"type": "object",
"required": ["attachment_type", "attachment_id"],
diff --git a/crates/agent/src/agent_loop.rs b/crates/agent/src/agent_loop.rs
index 67b25611bf..74e256644a 100644
--- a/crates/agent/src/agent_loop.rs
+++ b/crates/agent/src/agent_loop.rs
@@ -3,7 +3,7 @@ use crate::error::AgentError;
use crate::hook::{BridgeInputs, RegisterFn, ToolLoads, ToolRouter, UserToolFinisher};
use crate::model::metering::MeteringContext;
use crate::model::router::{ModelRouter, ProviderAgent};
-use crate::model::{PredefinedModel, ReasoningEffort};
+use crate::model::{ModelSpeed, PredefinedModel, ReasoningEffort};
use crate::stream::ChatCompletionStream;
use crate::telemetry::GenAiContext;
use crate::tool_adapter::DynToolSetAdapter;
@@ -103,6 +103,7 @@ impl From<&str> for SystemPrompt {
pub struct AgentLoop {
model: String,
reasoning_effort: Option<ReasoningEffort>,
+ speed: ModelSpeed,
max_turns: usize,
max_tokens: u64,
recorder: Arc<dyn UsageRecorder>,
@@ -129,6 +130,7 @@ impl AgentLoop {
Self {
model: PredefinedModel::default().to_string(),
reasoning_effort: None,
+ speed: ModelSpeed::Standard,
max_turns: DEFAULT_MAX_TURNS,
max_tokens: DEFAULT_MAX_TOKENS,
recorder,
@@ -139,6 +141,17 @@ impl AgentLoop {
}
}
+ /// Select a supported inference speed without changing the model or effort.
+ pub fn with_speed(mut self, speed: ModelSpeed) -> Result<Self, AgentError> {
+ if !speed.supported(&self.model) {
+ return Err(AgentError::Other(anyhow::anyhow!(
+ "speed is not supported by this model"
+ )));
+ }
+ self.speed = speed;
+ Ok(self)
+ }
+
/// Finish user tools inside the turn.
///
/// A user tool (`ai_toolset::UserTool`) answers `"PendingUserExecution"`
@@ -159,6 +172,9 @@ impl AgentLoop {
/// api-id string (frontend).
pub fn with_model<M: ToString>(mut self, model: M) -> Self {
self.model = model.to_string();
+ if !self.speed.supported(&self.model) {
+ self.speed = ModelSpeed::Standard;
+ }
self
}
@@ -245,6 +261,7 @@ impl AgentLoop {
telemetry.set_model(routed.provider(), routed.model_name());
routed.into_agent(
self.reasoning_effort,
+ self.speed,
handle,
prompt,
tool_search,
@@ -410,7 +427,12 @@ impl AgentLoop {
telemetry.clone(),
);
- let financial_context = MeteringContext::for_operation(self.recorder.as_ref(), &usage_ctx);
+ let financial_context = MeteringContext::for_session(
+ self.recorder.clone(),
+ &usage_ctx,
+ &self.model,
+ self.speed,
+ );
Session {
agent,
history: Vec::new(),
@@ -499,6 +521,7 @@ impl AgentLoop {
telemetry.set_model(routed.provider(), routed.model_name());
ProviderAgent::Test(Box::new(routed.into_agent(
self.reasoning_effort,
+ self.speed,
handle,
prompt,
tool_search,
@@ -594,7 +617,7 @@ impl Session {
} else if self
.financial_context
.as_ref()
- .is_some_and(MeteringContext::activated)
+ .is_some_and(|context| context.activated() || context.records_per_call())
{
self.financial_context.clone()
} else if current.is_some() {
diff --git a/crates/agent/src/lib.rs b/crates/agent/src/lib.rs
index 5a1898b6b3..8f9dc6ddf8 100644
--- a/crates/agent/src/lib.rs
+++ b/crates/agent/src/lib.rs
@@ -26,7 +26,7 @@ pub use convert::{merge_consecutive_parts, to_rig_messages};
pub use error::AgentError;
pub use hook::{FinishedUserTool, PendingUserTool, UserToolFinisher};
pub use model::metering::{MeteringContext, MeteringError, ProviderSupport, WireProtocol};
-pub use model::{PredefinedModel, ReasoningEffort};
+pub use model::{ModelSpeed, PredefinedModel, ReasoningEffort};
pub use stream::{ChatCompletionStream, McpInfo, StreamPart, ToolCall, ToolResponse, Usage};
pub use tool_adapter::{DynToolSetAdapter, ToolsetToolAdapter, normalize_request_schema};
diff --git a/crates/agent/src/model/metering.rs b/crates/agent/src/model/metering.rs
index df56e32c9b..d814183469 100644
--- a/crates/agent/src/model/metering.rs
+++ b/crates/agent/src/model/metering.rs
@@ -5,6 +5,10 @@
//! trusted policy facts. Raw unscoped HTTP calls retain legacy behavior and are NOT
//! covered producers. Spawned tasks must explicitly carry the scope.
+mod per_call;
+
+use super::ModelSpeed;
+use per_call::PerCallUsage;
use std::future::Future;
use std::sync::Arc;
use std::sync::atomic::{AtomicBool, Ordering};
@@ -201,6 +205,7 @@ pub struct MeteringContext {
mode: FinancialMode,
capability: FinancialCapability,
tracking: Option<Arc<dyn UsageTracking>>,
+ per_call: Option<PerCallUsage>,
usage: UsageContext,
support: Arc<Vec<ProviderSupport>>,
stopped: Arc<AtomicBool>,
@@ -219,6 +224,7 @@ impl MeteringContext {
mode,
capability,
tracking: None,
+ per_call: None,
usage,
support: Arc::new(support),
stopped: Arc::new(AtomicBool::new(false)),
@@ -241,10 +247,11 @@ impl MeteringContext {
/// Bridge analytics injection to per-attempt tracking. Financial scopes remain
/// strict; observational child operations bind their own user/feature/entity.
pub fn for_operation(recorder: &dyn UsageRecorder, usage: &UsageContext) -> Option<Self> {
- if let Some(current) = Self::current() {
+ if let Some(mut current) = Self::current() {
if current.activated() {
return Some(current);
}
+ current.per_call = None;
if let Some(tracking) = &current.tracking {
if current.usage.user == usage.user
&& current.usage.feature == usage.feature
@@ -255,12 +262,45 @@ impl MeteringContext {
let tracking = recorder.tracking().unwrap_or_else(|| tracking.clone());
return Some(Self::tracking_only(tracking, usage.clone()));
}
+ // Mask a parent session's per-call scope even without observational tracking.
+ current.usage = usage.clone();
+ current.tracking = recorder.tracking();
+ return Some(current);
}
recorder
.tracking()
.map(|tracking| Self::tracking_only(tracking, usage.clone()))
}
+ /// Bind per-call cost attribution for models with speed or context-dependent rates.
+ /// This uses the existing usage recorder; it never enables financial admission.
+ pub(crate) fn for_session(
+ recorder: Arc<dyn UsageRecorder>,
+ usage: &UsageContext,
+ model: &str,
+ speed: ModelSpeed,
+ ) -> Option<Self> {
+ let context = Self::for_operation(recorder.as_ref(), usage);
+ if !ModelSpeed::has_per_call_pricing(model) || context.as_ref().is_some_and(Self::activated)
+ {
+ return context;
+ }
+ let mut context = context.unwrap_or_else(|| {
+ Self::new(
+ FinancialMode::Legacy,
+ FinancialCapability::Unavailable,
+ usage.clone(),
+ vec![],
+ )
+ });
+ context.per_call = Some(PerCallUsage::new(recorder, usage.clone(), model, speed));
+ Some(context)
+ }
+
+ pub(crate) fn records_per_call(&self) -> bool {
+ self.per_call.is_some()
+ }
+
/// Run a completion, structured generation or session operation in this scope.
/// This does not automatically propagate into arbitrary `tokio::spawn` calls.
pub async fn scope<F: Future>(&self, future: F) -> F::Output {
@@ -299,7 +339,7 @@ impl MeteringContext {
}
pub(crate) fn enabled(&self) -> bool {
- self.activated() || self.tracking.is_some()
+ self.activated() || self.tracking.is_some() || self.per_call.is_some()
}
pub(crate) async fn begin(
@@ -308,6 +348,14 @@ impl MeteringContext {
protocol: WireProtocol,
body: &mut Value,
) -> Result<Attempt, MeteringError> {
+ let pricing = self
+ .per_call
+ .as_ref()
+ .filter(|pricing| pricing.matches(&model, body))
+ .cloned();
+ if self.per_call.is_some() && pricing.is_none() {
+ return Err(MeteringError::Unsupported);
+ }
if let Some(service) = &self.tracking {
let request = TrackedInvocation {
run_id: self.run_id,
@@ -329,10 +377,21 @@ impl MeteringContext {
},
context: self.clone(),
protocol,
+ pricing,
// Unreviewed compatible endpoints may return zero sentinels.
zero_usage_is_missing: protocol == WireProtocol::ChatCompletions,
});
}
+ if self.per_call.is_some() && !self.activated() {
+ return Ok(Attempt {
+ id: InvocationId::new(),
+ service: AttemptRecorder::Legacy,
+ context: self.clone(),
+ protocol,
+ pricing,
+ zero_usage_is_missing: false,
+ });
+ }
if self.stopped.load(Ordering::Acquire) {
return Err(MeteringError::Stopped);
}
@@ -377,12 +436,14 @@ impl MeteringContext {
service: AttemptRecorder::Financial(service.clone()),
context: self.clone(),
protocol: support.protocol,
+ pricing: None,
zero_usage_is_missing: support.zero_usage_is_missing,
})
}
}
enum AttemptRecorder {
+ Legacy,
Financial(Arc<dyn FinancialUsage>),
Tracking {
service: Arc<dyn UsageTracking>,
@@ -396,9 +457,26 @@ pub(crate) struct Attempt {
context: MeteringContext,
pub(crate) protocol: WireProtocol,
pub(crate) zero_usage_is_missing: bool,
+ pricing: Option<PerCallUsage>,
}
impl Attempt {
+ pub(crate) fn record_cost(
+ &self,
+ usage: UsageEvidence,
+ delivered_speed: Option<&str>,
+ ) -> Result<(), MeteringError> {
+ if let Some(pricing) = &self.pricing
+ && let Err(error) = pricing.record(usage, delivered_speed)
+ {
+ if self.activated() {
+ return Err(error);
+ }
+ tracing::error!(invocation_id = ?self.id, ?error, "per-call usage not recorded; provider result preserved");
+ }
+ Ok(())
+ }
+
pub(crate) fn activated(&self) -> bool {
self.context.activated()
}
@@ -433,6 +511,7 @@ impl Attempt {
let delivered = tokio::spawn(async move {
for retry in 0..3 {
let result = match &self.service {
+ AttemptRecorder::Legacy => Ok(()),
AttemptRecorder::Financial(service) => service.finalize(evidence.clone()).await.map(|_| ()),
AttemptRecorder::Tracking { service, request } => {
// Retry begin as well: a lost acknowledgement must not lose
diff --git a/crates/agent/src/model/metering/per_call.rs b/crates/agent/src/model/metering/per_call.rs
new file mode 100644
index 0000000000..067d8ce779
--- /dev/null
+++ b/crates/agent/src/model/metering/per_call.rs
@@ -0,0 +1,105 @@
+//! Per-provider-call pricing for speed tiers and long-context requests.
+
+use super::{MeteringError, ModelSpeed};
+use ai_usage::financial::{ProviderModel, UsageEvidence};
+use ai_usage::{UsageAmount, UsageContext, UsageRecorder};
+use serde_json::Value;
+use std::sync::Arc;
+
+#[derive(Clone)]
+pub(super) struct PerCallUsage {
+ recorder: Arc<dyn UsageRecorder>,
+ usage: UsageContext,
+ model: String,
+ requested: ModelSpeed,
+}
+
+impl PerCallUsage {
+ pub(super) fn new(
+ recorder: Arc<dyn UsageRecorder>,
+ usage: UsageContext,
+ model: &str,
+ requested: ModelSpeed,
+ ) -> Self {
+ Self {
+ recorder,
+ usage,
+ model: model.to_owned(),
+ requested,
+ }
+ }
+
+ pub(super) fn matches(&self, model: &ProviderModel, body: &Value) -> bool {
+ self.model == format!("{}/{}", model.provider(), model.model())
+ && match self.requested {
+ ModelSpeed::Standard => {
+ body.get("speed").is_none()
+ && body
+ .get("service_tier")
+ .is_none_or(|tier| tier == "default")
+ }
+ ModelSpeed::Fast => body.get("speed").is_some_and(|speed| speed == "fast"),
+ ModelSpeed::Ultrafast => body
+ .get("service_tier")
+ .is_some_and(|tier| tier == "ultrafast"),
+ }
+ }
+
+ pub(super) fn record(
+ &self,
+ evidence: UsageEvidence,
+ delivered: Option<&str>,
+ ) -> Result<(), MeteringError> {
+ let UsageEvidence::Reported(tokens) = evidence else {
+ return Ok(());
+ };
+ let speed = match delivered {
+ Some("default" | "standard") => ModelSpeed::Standard,
+ Some("fast") if self.model.starts_with("anthropic/") => ModelSpeed::Fast,
+ Some("ultrafast") if self.model.starts_with("openai/") => ModelSpeed::Ultrafast,
+ None if self.requested == ModelSpeed::Standard => ModelSpeed::Standard,
+ _ => {
+ tracing::error!(model = %self.model, "provider did not report a supported billing speed");
+ return Err(MeteringError::Unsupported);
+ }
+ };
+ if !speed.supported(&self.model) {
+ return Err(MeteringError::Unsupported);
+ }
+ let bare = self
+ .model
+ .split_once('/')
+ .ok_or(MeteringError::Unsupported)?
+ .1;
+ let mut price_key = bare.to_owned();
+ match speed {
+ ModelSpeed::Standard => {}
+ ModelSpeed::Fast => price_key.push_str(":fast"),
+ ModelSpeed::Ultrafast => price_key.push_str(":ultrafast"),
+ }
+ // OpenAI's threshold uses inclusive prompt tokens, per call, not the sum
+ // of all prompts in a tool loop. Reasoning is billed as output once.
+ if self.model.starts_with("openai/")
+ && tokens
+ .input()
+ .saturating_add(tokens.cache_read())
+ .saturating_add(tokens.cache_write())
+ > 272_000
+ {
+ price_key.push_str(":long");
+ }
+ self.recorder.record(self.usage.clone().into_event(
+ price_key,
+ UsageAmount::Tokens {
+ input: tokens.input(),
+ output: tokens.output() + tokens.reasoning(),
+ cache_read: tokens.cache_read(),
+ cache_write: tokens.cache_write(),
+ },
+ ));
+ Ok(())
+ }
+}
+
+#[cfg(test)]
+mod test;
diff --git a/crates/agent/src/model/metering/per_call/test.rs b/crates/agent/src/model/metering/per_call/test.rs
new file mode 100644
index 0000000000..4867f94302
--- /dev/null
+++ b/crates/agent/src/model/metering/per_call/test.rs
@@ -0,0 +1,80 @@
+use super::*;
+use ai_usage::financial::{TrustedTokenUsage, UnresolvedReason};
+use ai_usage::{AiFeature, UsageEvent};
+use std::sync::Mutex;
+
+#[derive(Default)]
+struct Recorder(Mutex<Vec<UsageEvent>>);
+impl UsageRecorder for Recorder {
+ fn record(&self, event: UsageEvent) {
+ self.0.lock().unwrap().push(event);
+ }
+}
+
+fn setup(model: &str, speed: ModelSpeed) -> (PerCallUsage, Arc<Recorder>) {
+ let recorder = Arc::new(Recorder::default());
+ let usage = UsageContext::new(
+ AiFeature::Chat,
+ "macro|[email protected]".to_owned().try_into().unwrap(),
+ );
+ (
+ PerCallUsage::new(recorder.clone(), usage, model, speed),
+ recorder,
+ )
+}
+
+#[test]
+fn prices_delivered_tier_and_each_calls_inclusive_prompt() {
+ let (pricing, recorder) = setup("openai/gpt-6-astra", ModelSpeed::Ultrafast);
+ for (input, cached, tier) in [
+ (270_000, 2_000, "ultrafast"),
+ (270_000, 2_001, "ultrafast"),
+ (1, 0, "default"),
+ ] {
+ pricing
+ .record(
+ UsageEvidence::Reported(TrustedTokenUsage::from_disjoint(input, 10, cached, 0, 5)),
+ Some(tier),
+ )
+ .unwrap();
+ }
+ let events = recorder.0.lock().unwrap();
+ assert_eq!(
+ events.iter().map(|e| e.model.as_str()).collect::<Vec<_>>(),
+ [
+ "gpt-6-astra:ultrafast",
+ "gpt-6-astra:ultrafast:long",
+ "gpt-6-astra"
+ ]
+ );
+ assert!(matches!(
+ events[0].amount,
+ UsageAmount::Tokens {
+ output: 15,
+ cache_read: 2_000,
+ ..
+ }
+ ));
+}
+
+#[test]
+fn fast_claude_records_cache_writes_and_rejects_unknown_tiers() {
+ let (pricing, recorder) = setup("anthropic/claude-opus-5-5", ModelSpeed::Fast);
+ let usage = UsageEvidence::Reported(TrustedTokenUsage::from_disjoint(5, 10, 20, 30, 0));
+ pricing.record(usage.clone(), Some("fast")).unwrap();
+ assert!(pricing.record(usage.clone(), None).is_err());
+ assert!(pricing.record(usage, Some("priority")).is_err());
+ pricing
+ .record(UsageEvidence::Missing(UnresolvedReason::Interrupted), None)
+ .unwrap();
+ let events = recorder.0.lock().unwrap();
+ assert_eq!(events.len(), 1);
+ assert_eq!(events[0].model, "claude-opus-5-5:fast");
+ assert!(matches!(
+ events[0].amount,
+ UsageAmount::Tokens {
+ cache_write: 30,
+ ..
+ }
+ ));
+}
diff --git a/crates/agent/src/model/metering_http.rs b/crates/agent/src/model/metering_http.rs
index e45bd35724..aa318e5f02 100644
--- a/crates/agent/src/model/metering_http.rs
+++ b/crates/agent/src/model/metering_http.rs
@@ -56,8 +56,33 @@ impl<H> MeteredHttpClient<H> {
async fn prepare(
&self,
- request: Request<Bytes>,
+ mut request: Request<Bytes>,
) -> http_client::Result<(Request<Bytes>, Option<Attempt>)> {
+ if self.protocol == WireProtocol::Anthropic
+ && serde_json::from_slice::<Value>(request.body())
+ .ok()
+ .is_some_and(|body| body.get("speed").is_some_and(|speed| speed == "fast"))
+ {
+ let beta = "fast-mode-2026-02-01";
+ let existing = request
+ .headers()
+ .get("anthropic-beta")
+ .and_then(|value| value.to_str().ok())
+ .unwrap_or_default();
+ if !existing.split(',').any(|value| value.trim() == beta) {
+ let value = if existing.is_empty() {
+ beta.to_owned()
+ } else {
+ format!("{existing},{beta}")
+ };
+ request.headers_mut().insert(
+ "anthropic-beta",
+ value
+ .parse()
+ .map_err(|_| http_error(MeteringError::Unsupported))?,
+ );
+ }
+ }
let Some(context) = MeteringContext::current().filter(MeteringContext::enabled) else {
return Ok((request, None));
};
@@ -79,7 +104,7 @@ impl<H> MeteredHttpClient<H> {
})();
let (mut body, model) = match parsed {
Ok(parsed) => parsed,
- Err(error) if !context.activated() => {
+ Err(error) if !context.activated() && !context.records_per_call() => {
tracing::error!(error = ?error, "cannot identify provider attempt; observation not persisted");
return Ok((request, None));
}
@@ -160,10 +185,12 @@ impl<H: HttpClientExt + Clone + 'static> HttpClientExt for MeteredHttpClient<H>
request_id(&parts.headers),
);
facts.response(&bytes);
- attempt
+ let pricing =
+ attempt.record_cost(facts.evidence(), facts.delivered_speed());
+ let finalized = attempt
.finish(facts.outcome, facts.evidence(), facts.request_id)
- .await
- .map_err(http_error)?;
+ .await;
+ pricing.and(finalized).map_err(http_error)?;
let body: LazyBody<U> = Box::pin(async move { Ok(U::from(bytes)) });
Ok(Response::from_parts(parts, body))
}
@@ -256,16 +283,20 @@ impl<H: HttpClientExt + Clone + 'static> HttpClientExt for MeteredHttpClient<H>
decoder.push(bytes, &mut facts);
if facts.terminal
&& let Some(attempt) = attempt.take()
- && let Err(error) = attempt
+ {
+ let pricing = attempt
+ .record_cost(facts.evidence(), facts.delivered_speed());
+ let finalized = attempt
.finish(
facts.outcome,
facts.evidence(),
facts.request_id.clone(),
)
- .await
- {
- let _ = tx.send(Err(http_error(error))).await;
- return;
+ .await;
+ if let Err(error) = pricing.and(finalized) {
+ let _ = tx.send(Err(http_error(error))).await;
+ return;
+ }
}
}
Err(_) => {
@@ -340,10 +371,9 @@ async fn finish_failure(attempt: Attempt, error: &http_client::Error) -> http_cl
evidence = facts.evidence();
}
}
- attempt
- .finish(outcome, evidence, facts.request_id)
- .await
- .map_err(http_error)
+ let pricing = attempt.record_cost(evidence, facts.delivered_speed());
+ let finalized = attempt.finish(outcome, evidence, facts.request_id).await;
+ pricing.and(finalized).map_err(http_error)
}
/// Raw provider facts: never use Rig's generic Usage defaults, which replace
@@ -353,6 +383,7 @@ struct Facts {
zero_usage_is_missing: bool,
request_id: Option<ProviderRequestId>,
usage: Option<Value>,
+ service_tier: Option<String>,
terminal: bool,
invalid: bool,
outcome: ProviderOutcome,
@@ -369,6 +400,7 @@ impl Facts {
zero_usage_is_missing,
request_id,
usage: None,
+ service_tier: None,
terminal: false,
invalid: false,
outcome: ProviderOutcome::Succeeded,
@@ -389,6 +421,10 @@ impl Facts {
match serde_json::from_slice::<Value>(bytes) {
Ok(value) => {
self.identify(&value);
+ self.service_tier = value
+ .get("service_tier")
+ .and_then(Value::as_str)
+ .map(str::to_owned);
self.usage = value
.get(if self.protocol == WireProtocol::Gemini {
"usageMetadata"
@@ -463,6 +499,10 @@ impl Facts {
value.get("type").and_then(Value::as_str),
Some("response.completed" | "response.incomplete" | "response.failed")
) {
+ self.service_tier = value
+ .pointer("/response/service_tier")
+ .and_then(Value::as_str)
+ .map(str::to_owned);
self.usage = value.pointer("/response/usage").cloned();
self.terminal = true;
if value["type"] == "response.failed" {
@@ -496,6 +536,17 @@ impl Facts {
}
}
+ fn delivered_speed(&self) -> Option<&str> {
+ if self.protocol == WireProtocol::Anthropic {
+ self.usage
+ .as_ref()
+ .and_then(|usage| usage.get("speed"))
+ .and_then(Value::as_str)
+ } else {
+ self.service_tier.as_deref()
+ }
+ }
+
fn evidence(&self) -> UsageEvidence {
if self.invalid {
return UsageEvidence::Missing(UnresolvedReason::UnsupportedDimensions);
diff --git a/crates/agent/src/model/metering_http/test.rs b/crates/agent/src/model/metering_http/test.rs
index 1e8582e8c6..f459de388d 100644
--- a/crates/agent/src/model/metering_http/test.rs
+++ b/crates/agent/src/model/metering_http/test.rs
@@ -120,6 +120,7 @@ fn admission(request: BeginInvocation) -> AuthorizedInvocation {
struct Transport {
calls: Arc<AtomicUsize>,
bodies: Arc<Mutex<Vec<Value>>>,
+ headers: Arc<Mutex<Vec<http_client::HeaderMap>>>,
response: Bytes,
remaining: Arc<Mutex<std::collections::VecDeque<Bytes>>>,
fail: bool,
@@ -146,6 +147,7 @@ impl Transport {
fn called<T: Into<Bytes>>(&self, request: Request<T>) {
self.calls.fetch_add(1, Ordering::SeqCst);
+ self.headers.lock().unwrap().push(request.headers().clone());
self.bodies
.lock()
.unwrap()
@@ -929,3 +931,5 @@ async fn a_reshaped_anthropic_request_is_metered_with_its_cache_usage() {
UsageEvidence::Reported(TrustedTokenUsage::from_disjoint(4, 7, 30000, 250, 0))
);
}
+
+mod speed;
diff --git a/crates/agent/src/model/metering_http/test/speed.rs b/crates/agent/src/model/metering_http/test/speed.rs
new file mode 100644
index 0000000000..4c4c44abc7
--- /dev/null
+++ b/crates/agent/src/model/metering_http/test/speed.rs
@@ -0,0 +1,215 @@
+use super::*;
+use crate::ModelSpeed;
+use ai_usage::{UsageAmount, UsageEvent, UsageRecorder};
+
+#[derive(Default)]
+struct Recorder(Mutex<Vec<UsageEvent>>);
+impl UsageRecorder for Recorder {
+ fn record(&self, event: UsageEvent) {
+ self.0.lock().unwrap().push(event);
+ }
+}
+
+#[tokio::test]
+async fn responses_transport_prices_actual_tier_and_rejects_mismatched_request() {
+ let recorder = Arc::new(Recorder::default());
+ let context = MeteringContext::for_session(
+ recorder.clone(),
+ &UsageContext::new(
+ AiFeature::Chat,
+ "macro|[email protected]".to_owned().try_into().unwrap(),
+ ),
+ "openai/gpt-6.1-sol",
+ ModelSpeed::Ultrafast,
+ )
+ .unwrap();
+ let transport = Transport::new(json!({"status":"completed", "service_tier":"ultrafast", "usage":{"input_tokens":100,"output_tokens":20,"input_tokens_details":{"cached_tokens":30},"output_tokens_details":{"reasoning_tokens":5}}}).to_string());
+ let client = MeteredHttpClient::new(transport.clone(), "openai", WireProtocol::Responses);
+ context
+ .scope(async {
+ for tier in ["ultrafast", "default"] {
+ let request = Request::builder()
+ .uri("https://example.test/responses")
+ .body(Bytes::from(
+ json!({"model":"gpt-6.1-sol","service_tier":tier}).to_string(),
+ ))
+ .unwrap();
+ let result = client.send::<_, Bytes>(request).await;
+ assert_eq!(result.is_ok(), tier == "ultrafast");
+ }
+ })
+ .await;
+ assert_eq!(transport.calls.load(Ordering::SeqCst), 1);
+ let events = recorder.0.lock().unwrap();
+ assert_eq!(events.len(), 1);
+ assert_eq!(events[0].model, "gpt-6.1-sol:ultrafast");
+ assert!(matches!(
+ events[0].amount,
+ UsageAmount::Tokens {
+ input: 70,
+ output: 20,
+ cache_read: 30,
+ ..
+ }
+ ));
+}
+
+#[tokio::test]
+async fn streaming_prices_completed_openai_and_claude_tiers_once() {
+ for (model, protocol, speed, body, wire, expected) in [
+ (
+ "openai/gpt-6-astra",
+ WireProtocol::Responses,
+ ModelSpeed::Ultrafast,
+ json!({"model":"gpt-6-astra","service_tier":"ultrafast"}),
+ "data: {\"type\":\"response.completed\",\"response\":{\"service_tier\":\"ultrafast\",\"usage\":{\"input_tokens\":10,\"output_tokens\":20}}}\n\n",
+ "gpt-6-astra:ultrafast",
+ ),
+ (
+ "anthropic/claude-opus-5-5",
+ WireProtocol::Anthropic,
+ ModelSpeed::Fast,
+ json!({"model":"claude-opus-5-5","speed":"fast"}),
+ "data: {\"type\":\"message_start\",\"message\":{\"usage\":{\"speed\":\"fast\",\"input_tokens\":10,\"output_tokens\":1,\"cache_read_input_tokens\":5,\"cache_creation_input_tokens\":3}}}\n\ndata: {\"type\":\"message_delta\",\"usage\":{\"output_tokens\":20}}\n\ndata: {\"type\":\"message_stop\"}\n\n",
+ "claude-opus-5-5:fast",
+ ),
+ (
+ "anthropic/claude-opus-5-5",
+ WireProtocol::Anthropic,
+ ModelSpeed::Fast,
+ json!({"model":"claude-opus-5-5","speed":"fast"}),
+ "data: {\"type\":\"message_start\",\"message\":{\"usage\":{\"speed\":\"standard\",\"input_tokens\":10,\"output_tokens\":1}}}\n\ndata: {\"type\":\"message_delta\",\"usage\":{\"output_tokens\":20}}\n\ndata: {\"type\":\"message_stop\"}\n\n",
+ "claude-opus-5-5",
+ ),
+ ] {
+ let recorder = Arc::new(Recorder::default());
+ let context = MeteringContext::for_session(
+ recorder.clone(),
+ &UsageContext::system(AiFeature::Chat),
+ model,
+ speed,
+ )
+ .unwrap();
+ let provider = model.split_once('/').unwrap().0;
+ let transport = Transport::new(wire);
+ let client = MeteredHttpClient::new(transport.clone(), provider, protocol);
+ context
+ .scope(async {
+ let request = Request::builder()
+ .uri("https://example.test/api")
+ .header("anthropic-beta", "existing-beta")
+ .body(Bytes::from(body.to_string()))
+ .unwrap();
+ let mut stream = client.send_streaming(request).await.unwrap().into_body();
+ while let Some(chunk) = stream.next().await {
+ chunk.unwrap();
+ }
+ })
+ .await;
+ let events = recorder.0.lock().unwrap();
+ assert_eq!(events.len(), 1);
+ assert_eq!(events[0].model, expected);
+ assert!(matches!(
+ events[0].amount,
+ UsageAmount::Tokens {
+ input: 10,
+ output: 20,
+ ..
+ }
+ ));
+ if expected.ends_with(":fast") {
+ assert!(matches!(
+ events[0].amount,
+ UsageAmount::Tokens {
+ cache_read: 5,
+ cache_write: 3,
+ ..
+ }
+ ));
+ }
+ if provider == "anthropic" {
+ let headers = transport.headers.lock().unwrap();
+ let beta = headers[0]["anthropic-beta"].to_str().unwrap();
+ assert!(beta.contains("existing-beta"));
+ assert!(beta.contains("fast-mode-2026-02-01"));
+ }
+ }
+}
+
+#[tokio::test]
+async fn child_completion_does_not_inherit_parent_speed_pricing() {
+ let recorder = Arc::new(Recorder::default());
+ let usage = UsageContext::system(AiFeature::Chat);
+ let parent = MeteringContext::for_session(
+ recorder.clone(),
+ &usage,
+ "openai/gpt-6-astra",
+ ModelSpeed::Ultrafast,
+ )
+ .unwrap();
+ let transport = Transport::new(anthropic_response());
+ let client = MeteredHttpClient::new(transport.clone(), "anthropic", WireProtocol::Anthropic);
+ parent
+ .scope(async {
+ let child = MeteringContext::for_operation(recorder.as_ref(), &usage);
+ MeteringContext::carry(child, async {
+ client
+ .send::<_, Bytes>(request())
+ .await
+ .unwrap()
+ .into_body()
+ .await
+ .unwrap();
+ })
+ .await;
+ })
+ .await;
+ assert_eq!(transport.calls.load(Ordering::SeqCst), 1);
+ assert!(recorder.0.lock().unwrap().is_empty());
+}
+
+#[tokio::test]
+async fn missing_delivered_tier_preserves_legacy_provider_results() {
+ for (protocol, model, speed, request_body, response) in [
+ (
+ WireProtocol::Responses,
+ "openai/gpt-6-astra",
+ ModelSpeed::Ultrafast,
+ json!({"model":"gpt-6-astra","service_tier":"ultrafast"}),
+ json!({"status":"completed","usage":{"input_tokens":10,"output_tokens":20}}),
+ ),
+ (
+ WireProtocol::Anthropic,
+ "anthropic/claude-opus-5-5",
+ ModelSpeed::Fast,
+ json!({"model":"claude-opus-5-5","speed":"fast"}),
+ json!({"usage":{"input_tokens":10,"output_tokens":20}}),
+ ),
+ ] {
+ let recorder = Arc::new(Recorder::default());
+ let context = MeteringContext::for_session(
+ recorder.clone(),
+ &UsageContext::system(AiFeature::Chat),
+ model,
+ speed,
+ )
+ .unwrap();
+ let provider = model.split_once('/').unwrap().0;
+ let client =
+ MeteredHttpClient::new(Transport::new(response.to_string()), provider, protocol);
+ context
+ .scope(async {
+ let request = Request::builder()
+ .uri("https://example.test/api")
+ .body(Bytes::from(request_body.to_string()))
+ .unwrap();
+ let delivered = client.send::<_, Bytes>(request).await.unwrap();
+ assert_eq!(
+ delivered.into_body().await.unwrap(),
+ Bytes::from(response.to_string())
+ );
+ })
+ .await;
+ assert!(recorder.0.lock().unwrap().is_empty());
+ }
+}
diff --git a/crates/agent/src/model/mod.rs b/crates/agent/src/model/mod.rs
index bacdf46aaa..35c40cc14c 100644
--- a/crates/agent/src/model/mod.rs
+++ b/crates/agent/src/model/mod.rs
@@ -3,9 +3,11 @@
//! Model routing takes a string and returns the appropriate client
mod predefined_model;
mod reasoning_effort;
+mod speed;
pub(crate) mod types;
pub use predefined_model::*;
pub use reasoning_effort::*;
+pub use speed::*;
pub(crate) mod anthropic;
pub mod anthropic_prompt_layout;
mod gemini;
diff --git a/crates/agent/src/model/router.rs b/crates/agent/src/model/router.rs
index 51e63443cc..954e1c45fd 100644
--- a/crates/agent/src/model/router.rs
+++ b/crates/agent/src/model/router.rs
@@ -42,7 +42,7 @@ use super::metering_http::MeteredHttpClient;
use super::openai::{OpenAiChatCompletionsModel, OpenAiResponsesModel};
use super::types::Model;
use super::usage_amount::usage_amount;
-use super::{PredefinedModel, ReasoningEffort};
+use super::{ModelSpeed, PredefinedModel, ReasoningEffort};
use crate::agent_loop::{SystemPrompt, ToolSearch};
use crate::error::AgentError;
use crate::hook::{BridgeInputs, StreamBridge};
@@ -131,6 +131,7 @@ impl<'a, H: HttpClientExt + Clone + Default + std::fmt::Debug + 'static> RoutedM
pub(crate) fn into_agent(
self,
reasoning_effort: Option<ReasoningEffort>,
+ speed: ModelSpeed,
handle: ToolServerHandle,
system_prompt: &SystemPrompt,
tool_search: &ToolSearch,
@@ -142,7 +143,11 @@ impl<'a, H: HttpClientExt + Clone + Default + std::fmt::Debug + 'static> RoutedM
let system_prompt_text = joined.as_str();
match self {
RoutedModel::Anthropic(m) => {
- let thinking = m.thinking_params(reasoning_effort);
+ let mut params = m
+ .thinking_params(reasoning_effort)
+ .unwrap_or_else(|| serde_json::json!({}));
+ speed.apply(&m.model().to_string(), &mut params);
+ let thinking = Some(params);
let layout = SessionLayout {
shared_system: system_prompt.shared().is_some(),
deferred: (m.loads_tools_by_reference() && !tool_search.catalog.is_empty())
@@ -180,7 +185,11 @@ impl<'a, H: HttpClientExt + Clone + Default + std::fmt::Debug + 'static> RoutedM
))
}
RoutedModel::OpenAiChatCompletions(m) => {
- let thinking = m.thinking_params(reasoning_effort);
+ let mut params = m
+ .thinking_params(reasoning_effort)
+ .unwrap_or_else(|| serde_json::json!({}));
+ speed.apply(&m.model().to_string(), &mut params);
+ let thinking = Some(params);
ProviderAgent::OpenAiChatCompletions(build_agent(
m.completion(),
thinking,
@@ -192,7 +201,11 @@ impl<'a, H: HttpClientExt + Clone + Default + std::fmt::Debug + 'static> RoutedM
))
}
RoutedModel::OpenAiResponses(m) => {
- let thinking = m.thinking_params(reasoning_effort);
+ let mut params = m
+ .thinking_params(reasoning_effort)
+ .unwrap_or_else(|| serde_json::json!({}));
+ speed.apply(&m.model().to_string(), &mut params);
+ let thinking = Some(params);
ProviderAgent::OpenAiResponses(build_agent(
m.completion(),
thinking,
@@ -736,6 +749,9 @@ where
};
let driver_span = agent_span.clone();
let financial_context = MeteringContext::current();
+ let records_per_call = financial_context
+ .as_ref()
+ .is_some_and(MeteringContext::records_per_call);
let driver = tokio::spawn(
MeteringContext::carry(financial_context, async move {
let mut liveness = StreamLiveness::new(agent_span.clone(), telemetry.clone());
@@ -770,11 +786,14 @@ where
);
// Aggregate analytics only. Financial evidence is
// persisted per HTTP attempt before SDK parsing.
- recorder.record(
- usage_ctx
- .clone()
- .into_event(model.clone(), usage_amount(protocol, &usage)),
- );
+ if !records_per_call {
+ recorder.record(
+ usage_ctx.clone().into_event(
+ model.clone(),
+ usage_amount(protocol, &usage),
+ ),
+ );
+ }
let _ =
driver_tx.send(Ok(StreamPart::Usage(crate::stream::Usage {
input_tokens: usage.input_tokens,
diff --git a/crates/agent/src/model/speed.rs b/crates/agent/src/model/speed.rs
new file mode 100644
index 0000000000..8e33d4d13b
--- /dev/null
+++ b/crates/agent/src/model/speed.rs
@@ -0,0 +1,70 @@
+//! Provider speed selection, independent of model identity and reasoning effort.
+
+use serde::{Deserialize, Serialize};
+use utoipa::ToSchema;
+
+/// The requested inference speed. Standard is backward compatible with old clients.
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize, ToSchema)]
+#[serde(rename_all = "snake_case")]
+pub enum ModelSpeed {
+ /// Use the model's standard processing tier.
+ #[default]
+ Standard,
+ /// Use Anthropic's faster Opus inference.
+ Fast,
+ /// Use OpenAI's fastest Responses processing tier.
+ Ultrafast,
+}
+
+impl ModelSpeed {
+ /// Whether this model supports the requested speed.
+ pub fn supported(self, model: &str) -> bool {
+ match self {
+ Self::Standard => true,
+ Self::Fast => model == "anthropic/claude-opus-5-5",
+ Self::Ultrafast => matches!(model, "openai/gpt-6-astra" | "openai/gpt-6.1-sol"),
+ }
+ }
+
+ /// Models whose usage must be priced per provider call, including long context.
+ pub fn has_per_call_pricing(model: &str) -> bool {
+ matches!(
+ model,
+ "openai/gpt-6-astra" | "openai/gpt-6.1-sol" | "anthropic/claude-opus-5-5"
+ )
+ }
+
+ /// Stable serialized value used by native session configuration.
+ pub const fn as_str(self) -> &'static str {
+ match self {
+ Self::Standard => "standard",
+ Self::Fast => "fast",
+ Self::Ultrafast => "ultrafast",
+ }
+ }
+
+ /// Parse an advertised native session value.
+ pub fn parse(value: &str) -> Option<Self> {
+ match value {
+ "standard" => Some(Self::Standard),
+ "fast" => Some(Self::Fast),
+ "ultrafast" => Some(Self::Ultrafast),
+ _ => None,
+ }
+ }
+
+ /// Add the provider-specific setting without changing reasoning parameters.
+ pub(crate) fn apply(self, model: &str, params: &mut serde_json::Value) {
+ match self {
+ Self::Standard if model.starts_with("openai/") => {
+ params["service_tier"] = "default".into()
+ }
+ Self::Standard => {}
+ Self::Fast => params["speed"] = "fast".into(),
+ Self::Ultrafast => params["service_tier"] = "ultrafast".into(),
+ }
+ }
+}
+
+#[cfg(test)]
+mod test;
diff --git a/crates/agent/src/model/speed/test.rs b/crates/agent/src/model/speed/test.rs
new file mode 100644
index 0000000000..295fe6f5b5
--- /dev/null
+++ b/crates/agent/src/model/speed/test.rs
@@ -0,0 +1,27 @@
+use super::*;
+
+#[test]
+fn premium_speeds_are_model_specific() {
+ assert!(ModelSpeed::Ultrafast.supported("openai/gpt-6-astra"));
+ assert!(ModelSpeed::Ultrafast.supported("openai/gpt-6.1-sol"));
+ assert!(ModelSpeed::Fast.supported("anthropic/claude-opus-5-5"));
+ for model in [
+ "openai/gpt-6-sol",
+ "anthropic/claude-sonnet-5-5",
+ "google/gemini-3.8-flash",
+ ] {
+ assert!(!ModelSpeed::Fast.supported(model));
+ assert!(!ModelSpeed::Ultrafast.supported(model));
+ assert!(ModelSpeed::Standard.supported(model));
+ }
+}
+
+#[test]
+fn speed_preserves_reasoning_parameters() {
+ let mut params = serde_json::json!({"reasoning": {"effort": "high"}});
+ ModelSpeed::Ultrafast.apply("openai/gpt-6-astra", &mut params);
+ assert_eq!(params["reasoning"]["effort"], "high");
+ assert_eq!(params["service_tier"], "ultrafast");
+ ModelSpeed::Standard.apply("openai/gpt-6-astra", &mut params);
+ assert_eq!(params["service_tier"], "default");
+}
diff --git a/crates/agent_inmem/src/domain/agent.rs b/crates/agent_inmem/src/domain/agent.rs
index 99dfb1ca36..6d4f71407f 100644
--- a/crates/agent_inmem/src/domain/agent.rs
+++ b/crates/agent_inmem/src/domain/agent.rs
@@ -14,8 +14,8 @@ use std::sync::{Arc, Mutex};
use std::time::{Duration, Instant};
use crate::domain::model_access::{InMemModelAccess, ModelAccess, ModelAccessError};
-use agent::ReasoningEffort;
use agent::types::{AssistantMessagePart, ChatMessage};
+use agent::{ModelSpeed, ReasoningEffort};
use agent::{StreamAccumulator, StreamPart, ToolResponse};
use agent_client_protocol::schema::v1::{
AgentCapabilities, AvailableCommand, AvailableCommandInput, AvailableCommandsUpdate,
@@ -105,6 +105,7 @@ struct TurnInput {
model: String,
/// Reasoning effort the turn runs with.
reasoning_effort: ReasoningEffort,
+ speed: ModelSpeed,
/// Who this agent is, for the engine's system prompt.
identity: Option<AgentIdentity>,
/// The session's instructions, for the engine's system prompt.
@@ -249,6 +250,9 @@ impl AgentState {
if !ReasoningEffort::supported(&model).contains(&state.reasoning_effort) {
state.reasoning_effort = ReasoningEffort::default();
}
+ if !state.speed.supported(&model) {
+ state.speed = ModelSpeed::Standard;
+ }
state.model = model;
}
}
@@ -269,6 +273,7 @@ impl AgentState {
&session.model,
&access.models(self.engine.supported_models()),
session.reasoning_effort,
+ session.speed,
)
}
@@ -280,6 +285,7 @@ impl AgentState {
messages: messages_for_turn(&[], prompt),
model: String::new(),
reasoning_effort: ReasoningEffort::default(),
+ speed: ModelSpeed::Standard,
identity: None,
instructions: None,
},
@@ -287,6 +293,7 @@ impl AgentState {
messages: messages_for_turn(&state.history, prompt),
model: state.model.clone(),
reasoning_effort: state.reasoning_effort,
+ speed: state.speed,
identity: state.identity.clone(),
instructions: state.instructions.clone(),
},
@@ -747,6 +754,25 @@ pub async fn serve(state: Arc<AgentState>, acp: AcpChannel) -> Result<(), AcpErr
}
state.set_model(value.to_string());
}
+ "speed" => {
+ let speed = ModelSpeed::parse(&value.to_string());
+ let Some(speed) = speed else {
+ return responder.respond_with_error(
+ AcpError::invalid_params().data("unknown speed"),
+ );
+ };
+ let mut session = state
+ .store
+ .get_mut(&state.session_id)
+ .expect("active session exists");
+ if !speed.supported(&session.model) || !access.allows(&session.model) {
+ return responder.respond_with_error(
+ AcpError::invalid_params()
+ .data("speed is not available for this model"),
+ );
+ }
+ session.speed = speed;
+ }
REASONING_EFFORT_CONFIG_ID => {
let Ok(effort) = value.to_string().parse() else {
return responder.respond_with_error(
@@ -851,6 +877,7 @@ async fn run_turn(
messages,
model,
reasoning_effort,
+ speed,
identity,
instructions,
} = state.turn_input(&prompt);
@@ -870,6 +897,7 @@ async fn run_turn(
owner: state.owner.clone(),
model,
reasoning_effort,
+ speed,
identity,
instructions,
messages,
diff --git a/crates/agent_inmem/src/domain/agent/test.rs b/crates/agent_inmem/src/domain/agent/test.rs
index cc8905e4e4..f8fd717c83 100644
--- a/crates/agent_inmem/src/domain/agent/test.rs
+++ b/crates/agent_inmem/src/domain/agent/test.rs
@@ -2022,3 +2022,74 @@ async fn downgrade_and_permission_failure_cannot_run_a_paid_model() {
.await;
}
}
+
+#[tokio::test]
+async fn changing_speed_applies_to_next_turn_and_rejects_unsupported_models() {
+ struct SpeedEngine(ScriptedEngine);
+ impl TurnEngine for SpeedEngine {
+ fn supported_models(&self) -> &[&str] {
+ &["anthropic/claude-sonnet-5-5", "anthropic/claude-opus-5-5"]
+ }
+ fn run_turn(
+ &self,
+ request: TurnRequest,
+ ) -> tokio::sync::mpsc::Receiver<Result<StreamPart, agent::AgentError>> {
+ self.0.run_turn(request)
+ }
+ }
+ let engine = Arc::new(SpeedEngine(ScriptedEngine::new(vec![])));
+ with_agent(Arc::clone(&engine), async |connection, session| {
+ assert!(
+ connection
+ .send_request(SetSessionConfigOptionRequest::new(
+ session.clone(),
+ "speed",
+ "fast"
+ ))
+ .block_task()
+ .await
+ .is_err()
+ );
+ connection
+ .send_request(SetSessionConfigOptionRequest::new(
+ session.clone(),
+ MODEL_CONFIG_ID,
+ "anthropic/claude-opus-5-5",
+ ))
+ .block_task()
+ .await
+ .unwrap();
+ connection
+ .send_request(SetSessionConfigOptionRequest::new(
+ session.clone(),
+ "speed",
+ "fast",
+ ))
+ .block_task()
+ .await
+ .unwrap();
+ connection
+ .send_request(text_prompt(&session, "fast please"))
+ .block_task()
+ .await
+ .unwrap();
+ connection
+ .send_request(SetSessionConfigOptionRequest::new(
+ session.clone(),
+ MODEL_CONFIG_ID,
+ "anthropic/claude-sonnet-5-5",
+ ))
+ .block_task()
+ .await
+ .unwrap();
+ connection
+ .send_request(text_prompt(&session, "standard please"))
+ .block_task()
+ .await
+ .unwrap();
+ })
+ .await;
+ let requests = engine.0.requests();
+ assert_eq!(requests[0].speed, agent::ModelSpeed::Fast);
+ assert_eq!(requests[1].speed, agent::ModelSpeed::Standard);
+}
diff --git a/crates/agent_inmem/src/domain/engine.rs b/crates/agent_inmem/src/domain/engine.rs
index 4fe138c4a1..c49826992e 100644
--- a/crates/agent_inmem/src/domain/engine.rs
+++ b/crates/agent_inmem/src/domain/engine.rs
@@ -48,6 +48,8 @@ pub struct TurnRequest {
pub model: String,
/// Provider-independent effort selected for this session.
pub reasoning_effort: ReasoningEffort,
+ /// Inference speed applied without changing model or reasoning effort.
+ pub speed: agent::ModelSpeed,
/// Who this agent is. Folded into every turn's system prompt so the
/// model can answer "who are you" even when the session has no
/// instructions. `None` leaves the standing prompt unnamed.
diff --git a/crates/agent_inmem/src/domain/model_options.rs b/crates/agent_inmem/src/domain/model_options.rs
index 2a61156f82..89c1c4c0a2 100644
--- a/crates/agent_inmem/src/domain/model_options.rs
+++ b/crates/agent_inmem/src/domain/model_options.rs
@@ -1,6 +1,6 @@
//! ACP session configuration advertised by the in-memory agent.
-use agent::ReasoningEffort;
+use agent::{ModelSpeed, ReasoningEffort};
use agent_client_protocol::schema::v1::{
SessionConfigOption, SessionConfigOptionCategory, SessionConfigSelectOption,
SessionConfigValueId,
@@ -13,6 +13,7 @@ pub fn session_config_options(
current_model: &str,
models: &[&str],
current_effort: ReasoningEffort,
+ current_speed: ModelSpeed,
) -> Vec<SessionConfigOption> {
let options: Vec<_> = models
.iter()
@@ -38,9 +39,35 @@ pub fn session_config_options(
current_effort,
));
}
+ if let Some(speed) = speed_config_option(current_model, current_speed) {
+ config.push(speed);
+ }
config
}
+/// ACP option for inference speed, available only for supported native models.
+pub fn speed_config_option(model: &str, current: ModelSpeed) -> Option<SessionConfigOption> {
+ let accelerated = [ModelSpeed::Fast, ModelSpeed::Ultrafast]
+ .into_iter()
+ .find(|speed| speed.supported(model))?;
+ Some(SessionConfigOption::select(
+ "speed",
+ "Speed",
+ SessionConfigValueId::new(current.as_str()),
+ vec![
+ SessionConfigSelectOption::new("standard", "Standard"),
+ SessionConfigSelectOption::new(
+ accelerated.as_str(),
+ if accelerated == ModelSpeed::Fast {
+ "Fast · 2× usage"
+ } else {
+ "Ultrafast · 6× usage"
+ },
+ ),
+ ],
+ ))
+}
+
/// ACP id for Macro's portable reasoning-effort selector.
pub const REASONING_EFFORT_CONFIG_ID: &str = "reasoning_effort";
@@ -70,7 +97,12 @@ pub fn reasoning_effort_config_option(
/// Model discovery uses the same default session configuration as session/new.
#[must_use]
pub fn model_config_options(current: &str, models: &[&str]) -> Vec<SessionConfigOption> {
- session_config_options(current, models, ReasoningEffort::default())
+ session_config_options(
+ current,
+ models,
+ ReasoningEffort::default(),
+ ModelSpeed::Standard,
+ )
}
#[cfg(test)]
diff --git a/crates/agent_inmem/src/domain/model_options/test.rs b/crates/agent_inmem/src/domain/model_options/test.rs
index dd339e07de..ee2f71869a 100644
--- a/crates/agent_inmem/src/domain/model_options/test.rs
+++ b/crates/agent_inmem/src/domain/model_options/test.rs
@@ -21,3 +21,28 @@ fn routed_options_use_house_names() {
]
);
}
+
+#[test]
+fn speed_is_advertised_only_for_supported_models_with_current_value() {
+ use agent::{ModelSpeed, ReasoningEffort};
+ use agent_client_protocol::schema::v1::SessionConfigKind;
+ for (model, speed) in [
+ ("openai/gpt-6-astra", ModelSpeed::Ultrafast),
+ ("openai/gpt-6.1-sol", ModelSpeed::Ultrafast),
+ ("anthropic/claude-opus-5-5", ModelSpeed::Fast),
+ ] {
+ let options =
+ super::session_config_options(model, &[model], ReasoningEffort::Default, speed);
+ let option = options
+ .iter()
+ .find(|option| option.id.to_string() == "speed")
+ .unwrap();
+ let SessionConfigKind::Select(select) = &option.kind else {
+ panic!("speed must be selectable")
+ };
+ assert_eq!(select.current_value.to_string(), speed.as_str());
+ }
+ assert!(
+ super::speed_config_option("anthropic/claude-sonnet-5-5", ModelSpeed::Standard).is_none()
+ );
+}
diff --git a/crates/agent_inmem/src/domain/replay.rs b/crates/agent_inmem/src/domain/replay.rs
index 41eaa651e3..7931387975 100644
--- a/crates/agent_inmem/src/domain/replay.rs
+++ b/crates/agent_inmem/src/domain/replay.rs
@@ -111,8 +111,30 @@ pub fn replay_history(frames: impl IntoIterator<Item = Message>) -> Vec<HistoryE
/// Rebuild the last valid reasoning-effort selection from durable frames.
#[must_use]
pub fn replay_reasoning_effort(frames: &[Message]) -> ReasoningEffort {
+ replay_config(frames, effort_from_options)
+}
+
+/// Restore the last confirmed speed selection after a runtime restart.
+pub fn replay_speed(frames: &[Message]) -> agent::ModelSpeed {
+ replay_config(frames, |options| {
+ options
+ .iter()
+ .find_map(|option| {
+ if option.id.to_string() != "speed" {
+ return None;
+ }
+ let SessionConfigKind::Select(select) = &option.kind else {
+ return None;
+ };
+ agent::ModelSpeed::parse(&select.current_value.to_string())
+ })
+ .unwrap_or_default()
+ })
+}
+
+fn replay_config<T: Default>(frames: &[Message], read: impl Fn(&[SessionConfigOption]) -> T) -> T {
let mut pending = HashSet::new();
- let mut current = ReasoningEffort::default();
+ let mut current = T::default();
for frame in frames {
match frame {
Message::ToRuntime(ToRuntimeMessage::Acp(acp)) => {
@@ -135,7 +157,7 @@ pub fn replay_reasoning_effort(frames: &[Message]) -> ReasoningEffort {
if let Some(options) = result.get("configOptions").and_then(|options| {
serde_json::from_value::<Vec<SessionConfigOption>>(options.clone()).ok()
}) {
- current = effort_from_options(&options);
+ current = read(&options);
}
}
RawJsonRpcMessage::Response(Response::Error { id, .. }) => {
@@ -147,7 +169,7 @@ pub fn replay_reasoning_effort(frames: &[Message]) -> ReasoningEffort {
..
}) = deserialize_params::<SessionNotification>(notification.params.as_ref())
{
- current = effort_from_options(&update.config_options);
+ current = read(&update.config_options);
}
}
_ => {}
diff --git a/crates/agent_inmem/src/domain/replay/test.rs b/crates/agent_inmem/src/domain/replay/test.rs
index 478221150b..4d8888cb50 100644
--- a/crates/agent_inmem/src/domain/replay/test.rs
+++ b/crates/agent_inmem/src/domain/replay/test.rs
@@ -242,6 +242,7 @@ fn only_confirmed_effort_replays_and_removal_resets_it() {
"anthropic/claude-sonnet-5-5",
&["anthropic/claude-sonnet-5-5"],
ReasoningEffort::Low,
+ agent::ModelSpeed::Standard,
);
frames.push(config_response(
serde_json::json!({ "configOptions": options }),
@@ -258,3 +259,23 @@ fn only_confirmed_effort_replays_and_removal_resets_it() {
frames.push(config_response(serde_json::json!({ "configOptions": [] })));
assert_eq!(replay_reasoning_effort(&frames), ReasoningEffort::Default);
}
+
+#[test]
+fn speed_replays_only_from_confirmed_configuration() {
+ use agent::ModelSpeed;
+ let mut frames = vec![config_frame("speed", "ultrafast")];
+ assert_eq!(replay_speed(&frames), ModelSpeed::Standard);
+ let options = crate::domain::model_options::session_config_options(
+ "openai/gpt-6-astra",
+ &["openai/gpt-6-astra"],
+ ReasoningEffort::Default,
+ ModelSpeed::Ultrafast,
+ );
+ frames.push(config_response(
+ serde_json::json!({"configOptions":options}),
+ ));
+ assert_eq!(replay_speed(&frames), ModelSpeed::Ultrafast);
+ frames.push(config_frame("model", "other-model"));
+ frames.push(config_response(serde_json::json!({"configOptions":[]})));
+ assert_eq!(replay_speed(&frames), ModelSpeed::Standard);
+}
diff --git a/crates/agent_inmem/src/domain/session.rs b/crates/agent_inmem/src/domain/session.rs
index d52be80c53..ed1c35951f 100644
--- a/crates/agent_inmem/src/domain/session.rs
+++ b/crates/agent_inmem/src/domain/session.rs
@@ -212,6 +212,8 @@ pub struct SessionState {
pub model: String,
/// Reasoning effort applied to subsequent turns.
pub reasoning_effort: ReasoningEffort,
+ /// Inference speed applied to subsequent turns.
+ pub speed: agent::ModelSpeed,
/// Who this agent is, snapshotted from the session's bot at attach.
pub identity: Option<AgentIdentity>,
/// Instructions every turn runs under, snapshotted from the session row
@@ -231,6 +233,7 @@ impl SessionState {
acp_session_id: None,
model,
reasoning_effort: ReasoningEffort::default(),
+ speed: agent::ModelSpeed::Standard,
identity: None,
instructions: None,
history: Vec::new(),
diff --git a/crates/agent_inmem/src/outbound/local_attachments/test.rs b/crates/agent_inmem/src/outbound/local_attachments/test.rs
index 6ce415c6ed..6b71e0dcdc 100644
--- a/crates/agent_inmem/src/outbound/local_attachments/test.rs
+++ b/crates/agent_inmem/src/outbound/local_attachments/test.rs
@@ -150,6 +150,7 @@ fn request(messages: Vec<ChatMessage>) -> TurnRequest {
owner: Owner::from_principal_str("macro|[email protected]").expect("user"),
model: "test-model".to_owned(),
reasoning_effort: agent::ReasoningEffort::default(),
+ speed: agent::ModelSpeed::Standard,
identity: None,
instructions: None,
messages,
diff --git a/crates/agent_inmem/src/outbound/manager.rs b/crates/agent_inmem/src/outbound/manager.rs
index 0112ce95af..fcb3511093 100644
--- a/crates/agent_inmem/src/outbound/manager.rs
+++ b/crates/agent_inmem/src/outbound/manager.rs
@@ -20,7 +20,7 @@ use model_owner::Owner;
use crate::domain::agent::{AgentState, serve};
use crate::domain::engine::{AgentIdentity, TurnEngine};
use crate::domain::mcp::DynMcpToolConnector;
-use crate::domain::replay::{FrameSource, replay_history, replay_reasoning_effort};
+use crate::domain::replay::{FrameSource, replay_history, replay_reasoning_effort, replay_speed};
use crate::domain::session::{SessionState, SessionStore};
#[cfg(test)]
@@ -161,11 +161,18 @@ impl InMemAgentManager {
} else {
agent::ReasoningEffort::default()
};
+ let restored_speed = replay_speed(&frames);
+ let speed = if restored_speed.supported(&facts.model) {
+ restored_speed
+ } else {
+ agent::ModelSpeed::Standard
+ };
let history = replay_history(frames);
self.store.entry(facts.id).or_insert_with(|| SessionState {
acp_session_id: facts.acp_session_id.clone(),
model: facts.model.clone(),
reasoning_effort,
+ speed,
identity: facts.identity.clone(),
instructions: facts.instructions.clone(),
history,
diff --git a/crates/agent_inmem/src/rig_engine.rs b/crates/agent_inmem/src/rig_engine.rs
index a8151f26e7..eac29e9da2 100644
--- a/crates/agent_inmem/src/rig_engine.rs
+++ b/crates/agent_inmem/src/rig_engine.rs
@@ -269,6 +269,7 @@ async fn drive_turn(
owner,
model,
reasoning_effort,
+ speed,
identity,
instructions,
messages,
@@ -362,6 +363,7 @@ async fn drive_turn(
let mut agent_loop = AgentLoop::new(base_context.recorder.clone())
.with_model(&model)
.with_reasoning_effort(reasoning_effort)
+ .with_speed(speed)?
.with_genai_telemetry(false);
if let Some(reviewer) = reviewer {
agent_loop = agent_loop.with_user_tool_finisher(user_tool_finisher(
diff --git a/crates/agent_inmem/src/testing.rs b/crates/agent_inmem/src/testing.rs
index 4595b5aa8b..692f27ab65 100644
--- a/crates/agent_inmem/src/testing.rs
+++ b/crates/agent_inmem/src/testing.rs
@@ -142,6 +142,8 @@ pub(crate) struct RecordedTurn {
pub(crate) model: String,
/// Reasoning effort the turn was to use.
pub(crate) reasoning_effort: agent::ReasoningEffort,
+ /// Provider speed selected for this turn.
+ pub(crate) speed: agent::ModelSpeed,
/// The conversation, flattened to text per message.
pub(crate) messages: Vec<String>,
/// Every image URL attached across the conversation, in order.
@@ -177,6 +179,7 @@ impl TurnEngine for ScriptedEngine {
.push(RecordedTurn {
model: request.model.clone(),
reasoning_effort: request.reasoning_effort,
+ speed: request.speed,
messages: request
.messages
.iter()
diff --git a/crates/ai_usage/src/outbound/pg_usage_repo/test.rs b/crates/ai_usage/src/outbound/pg_usage_repo/test.rs
index f804623280..56ac2a54db 100644
--- a/crates/ai_usage/src/outbound/pg_usage_repo/test.rs
+++ b/crates/ai_usage/src/outbound/pg_usage_repo/test.rs
@@ -499,3 +499,29 @@ async fn repricing_bills_cache_tokens_and_unprices_them_without_a_cache_rate(poo
"cache writes without a write rate must stay unpriced"
);
}
+
+#[sqlx::test(migrator = "MACRO_DB_MIGRATIONS")]
+async fn accelerated_and_long_context_prices_remain_distinct(pool: PgPool) {
+ let repo = PgUsageRepo::new(pool);
+ for (model, input, output, cache_read, cache_write) in [
+ ("gpt-6.1-sol", 2.0, 10.0, 0.1, 2.5),
+ ("gpt-6.1-sol:long", 4.0, 15.0, 0.2, 5.0),
+ ("gpt-6.1-sol:ultrafast", 12.0, 60.0, 0.6, 15.0),
+ ("gpt-6.1-sol:ultrafast:long", 24.0, 90.0, 1.2, 30.0),
+ ("gpt-6-astra:long", 20.0, 75.0, 2.0, 25.0),
+ ("gpt-6-astra:ultrafast", 60.0, 300.0, 6.0, 75.0),
+ ("gpt-6-astra:ultrafast:long", 120.0, 450.0, 12.0, 150.0),
+ ("claude-opus-5-5:fast", 8.0, 40.0, 0.4, 10.0),
+ ] {
+ assert_eq!(
+ repo.get_pricing(model).await.unwrap(),
+ Some(ModelPricing::Tokens {
+ input,
+ output,
+ cache_read: Some(cache_read),
+ cache_write: Some(cache_write),
+ }),
+ "{model}"
+ );
+ }
+}
diff --git a/crates/chat/src/domain/models/model_access.rs b/crates/chat/src/domain/models/model_access.rs
index b869ba1f2f..bcffc425bc 100644
--- a/crates/chat/src/domain/models/model_access.rs
+++ b/crates/chat/src/domain/models/model_access.rs
@@ -12,6 +12,7 @@ pub const CHAT_MODELS: &[&str] = &[
"anthropic/claude-opus-5-5",
"anthropic/claude-haiku-4-5",
"openai/gpt-6-astra",
+ "openai/gpt-6.1-sol",
"openai/gpt-5.6",
"openai/gpt-5.6-mini",
"openai/gpt-5.5",
diff --git a/crates/macro_db_client/migrations/20261008210853_ai_speed_pricing.down.sql b/crates/macro_db_client/migrations/20261008210853_ai_speed_pricing.down.sql
new file mode 100644
index 0000000000..1c5dfada03
--- /dev/null
+++ b/crates/macro_db_client/migrations/20261008210853_ai_speed_pricing.down.sql
@@ -0,0 +1,3 @@
+-- Retain pricing on rollback: usage recorded under these keys must remain priced.
+-- Old application versions do not request the new accounting identities.
+SELECT 1;
diff --git a/crates/macro_db_client/migrations/20261008210853_ai_speed_pricing.up.sql b/crates/macro_db_client/migrations/20261008210853_ai_speed_pricing.up.sql
new file mode 100644
index 0000000000..49c441f823
--- /dev/null
+++ b/crates/macro_db_client/migrations/20261008210853_ai_speed_pricing.up.sql
@@ -0,0 +1,19 @@
+-- USD per million tokens, verified 2026-10-08:
+-- https://developers.openai.com/api/docs/pricing?latest-pricing=ultrafast
+-- https://platform.claude.com/docs/en/about-claude/pricing
+-- Speed keys are accounting identities only; providers receive the real model id.
+-- :long applies above 272,000 inclusive prompt tokens for each OpenAI request.
+-- Claude cache writes use the existing 5-minute TTL. Opus 5.5 cache reads are 5%.
+INSERT INTO ai_pricing (
+ model, price_per_million_in, price_per_million_out,
+ price_per_million_cache_read, price_per_million_cache_write
+) VALUES
+ ('gpt-6.1-sol', 2, 10, 0.10, 2.50),
+ ('gpt-6.1-sol:long', 4, 15, 0.20, 5),
+ ('gpt-6.1-sol:ultrafast', 12, 60, 0.60, 15),
+ ('gpt-6.1-sol:ultrafast:long', 24, 90, 1.20, 30),
+ ('gpt-6-astra:long', 20, 75, 2, 25),
+ ('gpt-6-astra:ultrafast', 60, 300, 6, 75),
+ ('gpt-6-astra:ultrafast:long', 120, 450, 12, 150),
+ ('claude-opus-5-5:fast', 8, 40, 0.40, 10)
+ON CONFLICT (model) DO NOTHING;
diff --git a/docs/AGENT_GUIDE/ai-chat.md b/docs/AGENT_GUIDE/ai-chat.md
index f47e7e48d1..48ceb737f7 100644
--- a/docs/AGENT_GUIDE/ai-chat.md
+++ b/docs/AGENT_GUIDE/ai-chat.md
@@ -1558,3 +1558,24 @@ only that link. If another edit makes the revision stale, the AI reads the lates
settings and confirms a fresh proposal. Identical retries reuse the saved link.
Old pending booking review requests can be dismissed so the agent can resume with
conversational confirmation.
+
+## Accelerated model speed
+
+The lightning button beside the model selector enables Ultrafast for GPT-6 Astra
+and GPT-6.1 Sol (6× token pricing), or Fast for Claude Opus 5.5 (2× token pricing).
+Its tooltip shows the mode, state, and usage multiplier before sending. Other
+models keep standard speed. The preference is saved in browser local storage and
+shared by legacy chat and native Macro agent composers. External coding agents
+use their own speed controls.
+
+Preview the controls at `/app/debug/ui?ui=speed-toggle`.
+Click the bolt and verify its brief scale/rotation animation and highlighted state.
+Reload, change between supported models, and confirm the preference survives.
+Select an unsupported model and verify the bolt is hidden; switching back restores
+the preference. With reduced motion enabled, toggling changes the state without
+animation. Native sessions confirm speed configuration before sending the prompt;
+a rejected configuration must retain the unsent draft and attachments.
+
+Usage records follow the provider's delivered speed for each call, including tool
+loops. OpenAI prompts exceeding 272,000 tokens use the long-context rate for that
+call. Standard and accelerated rates have separate pricing keys.
diff --git a/packages/sdk/generated/cognition/index.ts b/packages/sdk/generated/cognition/index.ts
index ef043f9f47..83f9443206 100644
--- a/packages/sdk/generated/cognition/index.ts
+++ b/packages/sdk/generated/cognition/index.ts
@@ -1,4 +1,4 @@
// This file is auto-generated by @hey-api/openapi-ts
export { type Options, Sdk } from './sdk.gen';
-export type { AccessLevel, AddMcpServerData, AddMcpServerError, AddMcpServerErrors, AddMcpServerResponse, AddMcpServerResponses, AddServerRequest, AiAdmissionErrorBody, AiFeature, AssistantMessagePart, AttachmentMetadata, AttachmentType, BrowsePipedreamMcpCatalogData, BrowsePipedreamMcpCatalogError, BrowsePipedreamMcpCatalogErrors, BrowsePipedreamMcpCatalogResponse, BrowsePipedreamMcpCatalogResponses, CallToolData, CallToolError, CallToolErrors, CallToolRequest, CallToolResponse, CallToolResponse2, CallToolResponses, ChannelSharePermission, ChannelType, Chat, ChatAttachment, ChatHistory, ChatHistoryBatchMessagesRequest, ChatMessage, ChatMessageContent, ChatMessageError, ChatMessageWithAttachments, ChatPreview, ChatPreviewData, ChatResponse, ChatStream, ClientOptions, CompleteHandlerData, CompleteHandlerErrors, CompleteHandlerResponse, CompleteHandlerResponses, CompleteOnboardingRequest, CompletePipedreamMcpConnectionData, CompletePipedreamMcpConnectionError, CompletePipedreamMcpConnectionErrors, CompletePipedreamMcpConnectionResponse, CompletePipedreamMcpConnectionResponses, CompletionUsage, ConnectedServer, ConversationRecord, CopyChatData, CopyChatError, CopyChatErrors, CopyChatResponse, CopyChatResponses, CreateChatData, CreateChatError, CreateChatErrors, CreateChatRequest, CreateChatResponse, CreateChatResponses, CreatePipedreamMcpTokenData, CreatePipedreamMcpTokenError, CreatePipedreamMcpTokenErrors, CreatePipedreamMcpTokenResponse, CreatePipedreamMcpTokenResponses, DatabaseChange, DeleteChatData, DeleteChatError, DeleteChatErrors, DeleteChatResponses, DeleteMcpServerData, DeleteMcpServerError, DeleteMcpServerErrors, DeleteMcpServerResponse, DeleteMcpServerResponses, DeletePipedreamMcpConnectionData, DeletePipedreamMcpConnectionError, DeletePipedreamMcpConnectionErrors, DeletePipedreamMcpConnectionResponse, DeletePipedreamMcpConnectionResponses, DiscoverHandlerData, DiscoverHandlerErrors, DiscoverHandlerResponse, DiscoverHandlerResponses, DismissRunHandlerData, DismissRunHandlerErrors, DismissRunHandlerResponse, DismissRunHandlerResponses, DocumentCognitionServiceApiVersion, DocumentReference, DocumentTextPart, DynamicSchema, Entity, EntityType, ErrorBody, ErrorResponse, Expiry, FeatureUsage, FileType, GenericErrorResponse, GetBatchPreviewData, GetBatchPreviewError, GetBatchPreviewErrors, GetBatchPreviewRequest, GetBatchPreviewResponse, GetBatchPreviewResponse2, GetBatchPreviewResponses, GetChatData, GetChatError, GetChatErrors, GetChatHistoryBatchMessagesHandlerData, GetChatHistoryBatchMessagesHandlerError, GetChatHistoryBatchMessagesHandlerErrors, GetChatHistoryBatchMessagesHandlerResponse, GetChatHistoryBatchMessagesHandlerResponses, GetChatHistoryHandlerData, GetChatHistoryHandlerError, GetChatHistoryHandlerErrors, GetChatHistoryHandlerResponse, GetChatHistoryHandlerResponses, GetChatPermissionsData, GetChatPermissionsError, GetChatPermissionsErrors, GetChatPermissionsResponse, GetChatPermissionsResponse2, GetChatPermissionsResponses, GetChatResponse, GetChatResponse2, GetChatResponses, GetChatsForAttachmentHandlerData, GetChatsForAttachmentHandlerError, GetChatsForAttachmentHandlerErrors, GetChatsForAttachmentHandlerResponse, GetChatsForAttachmentHandlerResponses, GetChatsForAttachmentResponse, GetCitationHandlerData, GetCitationHandlerError, GetCitationHandlerErrors, GetCitationHandlerResponse, GetCitationHandlerResponses, GetMemoryHandlerData, GetMemoryHandlerError, GetMemoryHandlerErrors, GetMemoryHandlerResponse, GetMemoryHandlerResponses, GetStateHandler2Data, GetStateHandler2Errors, GetStateHandler2Response, GetStateHandler2Responses, GetStateHandlerData, GetStateHandlerErrors, GetStateHandlerResponse, GetStateHandlerResponses, GetUsageHandlerData, GetUsageHandlerError, GetUsageHandlerErrors, GetUsageHandlerResponse, GetUsageHandlerResponses, HealthHandlerData, HealthHandlerResponse, HealthHandlerResponses, HttpSendChatMessageRequest, ImportEntity, ImportRun, ImportSource, ImportState, ImportStatus, Initiator, JwtPayload, LinearIssueMeta, LinkShare, ListMcpServersData, ListMcpServersError, ListMcpServersErrors, ListMcpServersResponse, ListMcpServersResponses, ListPipedreamMcpConnectionsData, ListPipedreamMcpConnectionsError, ListPipedreamMcpConnectionsErrors, ListPipedreamMcpConnectionsResponse, ListPipedreamMcpConnectionsResponses, MacroUserIdStr, McpAuthCallbackData, McpAuthCallbackError, McpAuthCallbackErrors, McpAuthCallbackResponses, McpOauthClientMetadataData, McpOauthClientMetadataResponses, MemoryErrorBody, MemoryResponse, MessageWithAttachments, NewAttachment, NewChatMessage, NewMessageAttachment, NotionDocMeta, OnboardingRow, OnboardingState, OnboardingStatus, PatchChatData, PatchChatError, PatchChatErrors, PatchChatRequest, PatchChatResponses, PermanentlyDeleteChatData, PermanentlyDeleteChatError, PermanentlyDeleteChatErrors, PermanentlyDeleteChatResponses, PipedreamCatalogEntryResponse, PipedreamCatalogResponse, PipedreamCompleteRequest, PipedreamConnectionResponse, PipedreamTokenResponse, PipedreamUpdateRequest, Price, ProjectionStateResponse, ProjectionStatus, RefreshCadence, RejectToolCallData, RejectToolCallError, RejectToolCallErrors, RejectToolCallRequest, RejectToolCallResponses, RetryGatherHandlerData, RetryGatherHandlerError, RetryGatherHandlerErrors, RetryGatherHandlerResponse, RetryGatherHandlerResponses, RevertDeleteChatData, RevertDeleteChatError, RevertDeleteChatErrors, RevertDeleteChatResponses, Role, RunImportHandlerData, RunImportHandlerErrors, RunImportHandlerResponse, RunImportHandlerResponses, RunImportOutcome, RunImportRequest, RunStatus, SendChatMessageData, SendChatMessageError, SendChatMessageErrors, SendChatMessagePayload, SendChatMessageResponse, SendChatMessageResponse2, SendChatMessageResponses, ServerResponse, SetPricingHandlerData, SetPricingHandlerError, SetPricingHandlerErrors, SetPricingHandlerResponses, SetPricingRequest, SharePermissionV2, SlackChannelMeta, SlackParticipant, StartAuthRequest, StartAuthResponse, StartMcpAuthData, StartMcpAuthError, StartMcpAuthErrors, StartMcpAuthResponse, StartMcpAuthResponses, StopChatStreamData, StopChatStreamError, StopChatStreamError2, StopChatStreamErrors, StopChatStreamRequest, StopChatStreamResponse, StopChatStreamResponse2, StopChatStreamResponses, StreamError, StringIdResponse, StructuredCompletionData, StructuredCompletionError, StructuredCompletionError2, StructuredCompletionErrors, StructuredCompletionOutcome, StructuredCompletionRequest, StructuredCompletionResponse, StructuredCompletionResponse2, StructuredCompletionResponses, StructuredToolActivity, TargetType, ToolOutcome, ToolSet, UpdateChannelSharePermission, UpdateMcpServerData, UpdateMcpServerError, UpdateMcpServerErrors, UpdateMcpServerResponse, UpdateMcpServerResponses, UpdateOperation, UpdatePipedreamMcpConnectionData, UpdatePipedreamMcpConnectionError, UpdatePipedreamMcpConnectionErrors, UpdatePipedreamMcpConnectionResponse, UpdatePipedreamMcpConnectionResponses, UpdateServerRequest, UpdateSharePermissionRequestV2, UpdateToolCallData, UpdateToolCallError, UpdateToolCallErrors, UpdateToolCallRequest, UpdateToolCallResponses, UpdateToolResponseData, UpdateToolResponseError, UpdateToolResponseErrors, UpdateToolResponseRequest, UpdateToolResponseResponses, UpsertAiProjectionData, UpsertAiProjectionError, UpsertAiProjectionErrors, UpsertAiProjectionResponse, UpsertAiProjectionResponses, UpsertProjectionRequest, Usage, UsageRequest, UsageSummary, UserPdfRect, UserToolResponseValue, WebCitation, WithChatId } from './types.gen';
+export type { AccessLevel, AddMcpServerData, AddMcpServerError, AddMcpServerErrors, AddMcpServerResponse, AddMcpServerResponses, AddServerRequest, AiAdmissionErrorBody, AiFeature, AssistantMessagePart, AttachmentMetadata, AttachmentType, BrowsePipedreamMcpCatalogData, BrowsePipedreamMcpCatalogError, BrowsePipedreamMcpCatalogErrors, BrowsePipedreamMcpCatalogResponse, BrowsePipedreamMcpCatalogResponses, CallToolData, CallToolError, CallToolErrors, CallToolRequest, CallToolResponse, CallToolResponse2, CallToolResponses, ChannelSharePermission, ChannelType, Chat, ChatAttachment, ChatHistory, ChatHistoryBatchMessagesRequest, ChatMessage, ChatMessageContent, ChatMessageError, ChatMessageWithAttachments, ChatPreview, ChatPreviewData, ChatResponse, ChatStream, ClientOptions, CompleteHandlerData, CompleteHandlerErrors, CompleteHandlerResponse, CompleteHandlerResponses, CompleteOnboardingRequest, CompletePipedreamMcpConnectionData, CompletePipedreamMcpConnectionError, CompletePipedreamMcpConnectionErrors, CompletePipedreamMcpConnectionResponse, CompletePipedreamMcpConnectionResponses, CompletionUsage, ConnectedServer, ConversationRecord, CopyChatData, CopyChatError, CopyChatErrors, CopyChatResponse, CopyChatResponses, CreateChatData, CreateChatError, CreateChatErrors, CreateChatRequest, CreateChatResponse, CreateChatResponses, CreatePipedreamMcpTokenData, CreatePipedreamMcpTokenError, CreatePipedreamMcpTokenErrors, CreatePipedreamMcpTokenResponse, CreatePipedreamMcpTokenResponses, DatabaseChange, DeleteChatData, DeleteChatError, DeleteChatErrors, DeleteChatResponses, DeleteMcpServerData, DeleteMcpServerError, DeleteMcpServerErrors, DeleteMcpServerResponse, DeleteMcpServerResponses, DeletePipedreamMcpConnectionData, DeletePipedreamMcpConnectionError, DeletePipedreamMcpConnectionErrors, DeletePipedreamMcpConnectionResponse, DeletePipedreamMcpConnectionResponses, DiscoverHandlerData, DiscoverHandlerErrors, DiscoverHandlerResponse, DiscoverHandlerResponses, DismissRunHandlerData, DismissRunHandlerErrors, DismissRunHandlerResponse, DismissRunHandlerResponses, DocumentCognitionServiceApiVersion, DocumentReference, DocumentTextPart, DynamicSchema, Entity, EntityType, ErrorBody, ErrorResponse, Expiry, FeatureUsage, FileType, GenericErrorResponse, GetBatchPreviewData, GetBatchPreviewError, GetBatchPreviewErrors, GetBatchPreviewRequest, GetBatchPreviewResponse, GetBatchPreviewResponse2, GetBatchPreviewResponses, GetChatData, GetChatError, GetChatErrors, GetChatHistoryBatchMessagesHandlerData, GetChatHistoryBatchMessagesHandlerError, GetChatHistoryBatchMessagesHandlerErrors, GetChatHistoryBatchMessagesHandlerResponse, GetChatHistoryBatchMessagesHandlerResponses, GetChatHistoryHandlerData, GetChatHistoryHandlerError, GetChatHistoryHandlerErrors, GetChatHistoryHandlerResponse, GetChatHistoryHandlerResponses, GetChatPermissionsData, GetChatPermissionsError, GetChatPermissionsErrors, GetChatPermissionsResponse, GetChatPermissionsResponse2, GetChatPermissionsResponses, GetChatResponse, GetChatResponse2, GetChatResponses, GetChatsForAttachmentHandlerData, GetChatsForAttachmentHandlerError, GetChatsForAttachmentHandlerErrors, GetChatsForAttachmentHandlerResponse, GetChatsForAttachmentHandlerResponses, GetChatsForAttachmentResponse, GetCitationHandlerData, GetCitationHandlerError, GetCitationHandlerErrors, GetCitationHandlerResponse, GetCitationHandlerResponses, GetMemoryHandlerData, GetMemoryHandlerError, GetMemoryHandlerErrors, GetMemoryHandlerResponse, GetMemoryHandlerResponses, GetStateHandler2Data, GetStateHandler2Errors, GetStateHandler2Response, GetStateHandler2Responses, GetStateHandlerData, GetStateHandlerErrors, GetStateHandlerResponse, GetStateHandlerResponses, GetUsageHandlerData, GetUsageHandlerError, GetUsageHandlerErrors, GetUsageHandlerResponse, GetUsageHandlerResponses, HealthHandlerData, HealthHandlerResponse, HealthHandlerResponses, HttpSendChatMessageRequest, ImportEntity, ImportRun, ImportSource, ImportState, ImportStatus, Initiator, JwtPayload, LinearIssueMeta, LinkShare, ListMcpServersData, ListMcpServersError, ListMcpServersErrors, ListMcpServersResponse, ListMcpServersResponses, ListPipedreamMcpConnectionsData, ListPipedreamMcpConnectionsError, ListPipedreamMcpConnectionsErrors, ListPipedreamMcpConnectionsResponse, ListPipedreamMcpConnectionsResponses, MacroUserIdStr, McpAuthCallbackData, McpAuthCallbackError, McpAuthCallbackErrors, McpAuthCallbackResponses, McpOauthClientMetadataData, McpOauthClientMetadataResponses, MemoryErrorBody, MemoryResponse, MessageWithAttachments, ModelSpeed, NewAttachment, NewChatMessage, NewMessageAttachment, NotionDocMeta, OnboardingRow, OnboardingState, OnboardingStatus, PatchChatData, PatchChatError, PatchChatErrors, PatchChatRequest, PatchChatResponses, PermanentlyDeleteChatData, PermanentlyDeleteChatError, PermanentlyDeleteChatErrors, PermanentlyDeleteChatResponses, PipedreamCatalogEntryResponse, PipedreamCatalogResponse, PipedreamCompleteRequest, PipedreamConnectionResponse, PipedreamTokenResponse, PipedreamUpdateRequest, Price, ProjectionStateResponse, ProjectionStatus, RefreshCadence, RejectToolCallData, RejectToolCallError, RejectToolCallErrors, RejectToolCallRequest, RejectToolCallResponses, RetryGatherHandlerData, RetryGatherHandlerError, RetryGatherHandlerErrors, RetryGatherHandlerResponse, RetryGatherHandlerResponses, RevertDeleteChatData, RevertDeleteChatError, RevertDeleteChatErrors, RevertDeleteChatResponses, Role, RunImportHandlerData, RunImportHandlerErrors, RunImportHandlerResponse, RunImportHandlerResponses, RunImportOutcome, RunImportRequest, RunStatus, SendChatMessageData, SendChatMessageError, SendChatMessageErrors, SendChatMessagePayload, SendChatMessageResponse, SendChatMessageResponse2, SendChatMessageResponses, ServerResponse, SetPricingHandlerData, SetPricingHandlerError, SetPricingHandlerErrors, SetPricingHandlerResponses, SetPricingRequest, SharePermissionV2, SlackChannelMeta, SlackParticipant, StartAuthRequest, StartAuthResponse, StartMcpAuthData, StartMcpAuthError, StartMcpAuthErrors, StartMcpAuthResponse, StartMcpAuthResponses, StopChatStreamData, StopChatStreamError, StopChatStreamError2, StopChatStreamErrors, StopChatStreamRequest, StopChatStreamResponse, StopChatStreamResponse2, StopChatStreamResponses, StreamError, StringIdResponse, StructuredCompletionData, StructuredCompletionError, StructuredCompletionError2, StructuredCompletionErrors, StructuredCompletionOutcome, StructuredCompletionRequest, StructuredCompletionResponse, StructuredCompletionResponse2, StructuredCompletionResponses, StructuredToolActivity, TargetType, ToolOutcome, ToolSet, UpdateChannelSharePermission, UpdateMcpServerData, UpdateMcpServerError, UpdateMcpServerErrors, UpdateMcpServerResponse, UpdateMcpServerResponses, UpdateOperation, UpdatePipedreamMcpConnectionData, UpdatePipedreamMcpConnectionError, UpdatePipedreamMcpConnectionErrors, UpdatePipedreamMcpConnectionResponse, UpdatePipedreamMcpConnectionResponses, UpdateServerRequest, UpdateSharePermissionRequestV2, UpdateToolCallData, UpdateToolCallError, UpdateToolCallErrors, UpdateToolCallRequest, UpdateToolCallResponses, UpdateToolResponseData, UpdateToolResponseError, UpdateToolResponseErrors, UpdateToolResponseRequest, UpdateToolResponseResponses, UpsertAiProjectionData, UpsertAiProjectionError, UpsertAiProjectionErrors, UpsertAiProjectionResponse, UpsertAiProjectionResponses, UpsertProjectionRequest, Usage, UsageRequest, UsageSummary, UserPdfRect, UserToolResponseValue, WebCitation, WithChatId } from './types.gen';
diff --git a/packages/sdk/generated/cognition/types.gen.ts b/packages/sdk/generated/cognition/types.gen.ts
index 6445845c2b..660675d83f 100644
--- a/packages/sdk/generated/cognition/types.gen.ts
+++ b/packages/sdk/generated/cognition/types.gen.ts
@@ -675,6 +675,10 @@ export type HttpSendChatMessageRequest = {
* The model to respond with (`provider/model` id)
*/
model: string;
+ /**
+ * Inference speed, validated against the selected model. Defaults to standard.
+ */
+ speed?: ModelSpeed;
/**
* Which toolset to use. Defaults to `all`
*/
@@ -877,6 +881,11 @@ export type MessageWithAttachments = {
date: string;
};
+/**
+ * The requested inference speed. Standard is backward compatible with old clients.
+ */
+export type ModelSpeed = 'standard' | 'fast' | 'ultrafast';
+
export type NewAttachment = {
attachment_id: string;
attachment_type: AttachmentType;
diff --git a/packages/sdk/specs/cognition.json b/packages/sdk/specs/cognition.json
index cb22dea994..91fcefa5ab 100644
--- a/packages/sdk/specs/cognition.json
+++ b/packages/sdk/specs/cognition.json
@@ -3761,6 +3761,10 @@
"type": "string",
"description": "The model to respond with (`provider/model` id)"
},
+ "speed": {
+ "$ref": "#/components/schemas/ModelSpeed",
+ "description": "Inference speed, validated against the selected model. Defaults to standard."
+ },
"toolset": {
"$ref": "#/components/schemas/ToolSet",
"description": "Which toolset to use. Defaults to `all`"
@@ -4003,6 +4007,11 @@
}
}
},
+ "ModelSpeed": {
+ "type": "string",
+ "description": "The requested inference speed. Standard is backward compatible with old clients.",
+ "enum": ["standard", "fast", "ultrafast"]
+ },
"NewAttachment": {
"type": "object",
"required": ["attachment_type", "attachment_id"],
diff --git a/services/document_cognition_service/src/api/stream/chat_message.rs b/services/document_cognition_service/src/api/stream/chat_message.rs
index 537d721256..4a74a08fc2 100644
--- a/services/document_cognition_service/src/api/stream/chat_message.rs
+++ b/services/document_cognition_service/src/api/stream/chat_message.rs
@@ -14,7 +14,7 @@ use crate::service::chat_renamer::spawn_initial_chat_rename;
use crate::service::get_chat::get_chat;
use crate::service::notification::notify;
use agent::types::{AssistantMessagePart, ChatMessage, ChatMessageContent};
-use agent::{AgentLoop, StreamAccumulator};
+use agent::{AgentLoop, ModelSpeed, StreamAccumulator};
use ai_billing::inbound::admission::admission_status;
use async_stream::stream;
use attachment::FormattedParts;
@@ -77,6 +77,9 @@ pub struct HttpSendChatMessageRequest {
pub chat_id: Option<String>,
/// The model to respond with (`provider/model` id)
pub model: String,
+ /// Inference speed, validated against the selected model. Defaults to standard.
+ #[serde(default)]
+ pub speed: ModelSpeed,
/// Additional system instructions appended to the base system prompt
#[serde(skip_serializing_if = "Option::is_none")]
pub additional_instructions: Option<String>,
@@ -189,6 +192,14 @@ async fn send_chat_message_inner(
code: None,
});
}
+ if !request.speed.supported(&request.model) {
+ return Err(ChatMessageError {
+ error: "Speed is not supported by the selected model".to_owned(),
+ stream_id: Some(stream_id),
+ status: Some(StatusCode::BAD_REQUEST),
+ code: None,
+ });
+ }
let model = request.model.clone();
// Resolve existing-chat permissions before admission, without creating anything.
@@ -366,6 +377,7 @@ async fn send_chat_message_inner(
message_id.clone(),
stream_id.clone(),
model,
+ request.speed,
now,
request.content.clone(),
user_message_id,
@@ -522,6 +534,7 @@ fn stream_and_save_message(
message_id: String,
stream_id: String,
model: String,
+ speed: ModelSpeed,
now: std::time::Instant,
user_message_content: String,
user_message_id: String,
@@ -571,7 +584,9 @@ fn stream_and_save_message(
// in the observability backend.
let agent_loop = AgentLoop::new(tool_context.recorder.clone())
.with_model(&model)
- .with_conversation_id(&chat_id);
+ .with_conversation_id(&chat_id)
+ .with_speed(speed)
+ .expect("chat speed was validated before starting the stream");
let rig_messages = agent::to_rig_messages(&request);
diff --git a/services/document_cognition_service/src/api/stream/chat_message/test.rs b/services/document_cognition_service/src/api/stream/chat_message/test.rs
index 4208740863..bb3328fac1 100644
--- a/services/document_cognition_service/src/api/stream/chat_message/test.rs
+++ b/services/document_cognition_service/src/api/stream/chat_message/test.rs
@@ -6,6 +6,7 @@ use macro_db_migrator::MACRO_DB_MIGRATIONS;
fn admission_request(chat_id: Option<String>) -> HttpSendChatMessageRequest {
HttpSendChatMessageRequest {
+ speed: agent::ModelSpeed::Standard,
content: "hello".into(),
chat_id,
model: chat::domain::models::FREE_MODEL.into(),