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(agents): offer Fireworks open-weight models on inmem Macro (#6954)

Wolf Mermelstein0e4892aa5b92
diff --git a/apps/web/src/features/agents-view/primitives/preferred-inmem-model.test.ts b/apps/web/src/features/agents-view/primitives/preferred-inmem-model.test.ts
new file mode 100644
index 0000000000..7094e1f483
--- /dev/null
+++ b/apps/web/src/features/agents-view/primitives/preferred-inmem-model.test.ts
@@ -0,0 +1,47 @@
+import { createRoot } from 'solid-js';
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+import { createPreferredInmemModel } from './preferred-inmem-model';
+
+beforeEach(() => window.localStorage.clear());
+
+describe('preferred in-memory model', () => {
+ it('remembers the last Models choice across composer mounts', () => {
+ createRoot((dispose) => {
+ const preferred = createPreferredInmemModel('user-a');
+ preferred.remember('fireworks/kimi-k3');
+ preferred.remember('fireworks/muse-glimmer-30b');
+ expect(preferred.model()).toBe('fireworks/muse-glimmer-30b');
+ dispose();
+ });
+ createRoot((dispose) => {
+ expect(createPreferredInmemModel('user-a').model()).toBe(
+ 'fireworks/muse-glimmer-30b'
+ );
+ expect(createPreferredInmemModel('user-b').model()).toBeUndefined();
+ dispose();
+ });
+ });
+
+ it('ignores blank stored values', () => {
+ window.localStorage.setItem('agents-view-inmem-model-v1:user-a', ' ');
+ createRoot((dispose) => {
+ expect(createPreferredInmemModel('user-a').model()).toBeUndefined();
+ dispose();
+ });
+ });
+
+ it('keeps working when storage cannot be written', () => {
+ const write = vi
+ .spyOn(Storage.prototype, 'setItem')
+ .mockImplementation(() => {
+ throw new Error('Storage unavailable');
+ });
+ createRoot((dispose) => {
+ const preferred = createPreferredInmemModel('user-a');
+ preferred.remember('fireworks/kimi-k3');
+ expect(preferred.model()).toBe('fireworks/kimi-k3');
+ dispose();
+ });
+ write.mockRestore();
+ });
+});
diff --git a/apps/web/src/features/agents-view/primitives/preferred-inmem-model.ts b/apps/web/src/features/agents-view/primitives/preferred-inmem-model.ts
new file mode 100644
index 0000000000..afba58136e
--- /dev/null
+++ b/apps/web/src/features/agents-view/primitives/preferred-inmem-model.ts
@@ -0,0 +1,34 @@
+import { makePersistedState } from '@app/lib/persistence';
+import { createUserScopedStorage } from '@core/util/userScopedStorage';
+import { createSignal } from 'solid-js';
+
+/**
+ * The last model chosen from the New conversation **Models** group (Macro's
+ * in-memory catalog). Scoped to the signed-in user so a reload keeps that
+ * choice as the default until another Models entry is picked.
+ */
+export function createPreferredInmemModel(userId: string | undefined) {
+ const storage = createUserScopedStorage('agents-view-inmem-model-v1');
+ const [model, setModel] = makePersistedState(
+ createSignal<string | undefined>(undefined),
+ {
+ storages: {
+ restore: () => {
+ if (!userId) return;
+ const stored = storage.read(userId)?.trim();
+ return stored || undefined;
+ },
+ write: (value) => {
+ if (!userId) return;
+ if (value) storage.write(userId, value);
+ else storage.write(userId, '');
+ },
+ },
+ }
+ );
+
+ return {
+ model,
+ remember: (id: string) => setModel(id.trim() || undefined),
+ };
+}
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 819ac546c7..4a5caaadf7 100644
--- a/apps/web/src/features/agents-view/views/NewChatPage.test.tsx
+++ b/apps/web/src/features/agents-view/views/NewChatPage.test.tsx
@@ -19,6 +19,10 @@ const mocks = vi.hoisted(() => ({
attachments: [] as InputAttachmentData[],
recentIds: [] as string[],
recentUrls: [] as string[],
+ preferredInmemModel: undefined as string | undefined,
+ rememberInmemModel: vi.fn((id: string) => {
+ mocks.preferredInmemModel = id;
+ }),
repositories: [] as { url: string; defaultBranch?: string }[],
}));
vi.mock('@core/util/upload', () => ({ uploadFile: vi.fn() }));
@@ -42,6 +46,12 @@ vi.mock('../primitives/recent-repositories', () => ({
remember: vi.fn(),
}),
}));
+vi.mock('../primitives/preferred-inmem-model', () => ({
+ createPreferredInmemModel: () => ({
+ model: () => mocks.preferredInmemModel,
+ remember: mocks.rememberInmemModel,
+ }),
+}));
vi.mock('../queries/reachable-repositories', () => ({
createReachableRepositories: () => ({
repositories: () => mocks.repositories,
@@ -177,10 +187,14 @@ describe('agent-led new conversation', () => {
mocks.attachments = [];
mocks.recentIds = [MACRO_CODER_BOT_ID];
mocks.recentUrls = [];
+ mocks.preferredInmemModel = undefined;
mocks.repositories = [
{ url: 'https://github.com/macro-inc/macro', defaultBranch: 'develop' },
];
vi.clearAllMocks();
+ mocks.rememberInmemModel.mockImplementation((id: string) => {
+ mocks.preferredInmemModel = id;
+ });
motionStyles = document.createElement('style');
motionStyles.textContent =
'[role="menu"] { animation-name: none; transition-duration: 0s; }';
@@ -471,6 +485,9 @@ describe('agent-led new conversation', () => {
expect(screen.queryByText('anthropic/claude-sonnet-5')).toBeNull();
fireEvent.keyDown(sonnet, { key: 'Enter' });
await waitFor(() => expect(screen.queryByRole('menu')).toBeNull());
+ expect(mocks.rememberInmemModel).toHaveBeenCalledWith(
+ 'anthropic/claude-sonnet-5'
+ );
expect(screen.getByRole('button', { name: 'Agent' }).textContent).toBe(
'Sonnet 5'
);
@@ -487,6 +504,21 @@ describe('agent-led new conversation', () => {
repoUrl: undefined,
modelOverride: 'anthropic/claude-sonnet-5',
});
+ // Macro Models picks stick: a second send still uses the preferred model.
+ expect(screen.getByRole('button', { name: 'Agent' }).textContent).toBe(
+ 'Sonnet 5'
+ );
+ fireEvent.click(screen.getByRole('button', { name: 'Send' }));
+ expect(send.mock.calls[1][0]).toMatchObject({
+ modelOverride: 'anthropic/claude-sonnet-5',
+ });
+ });
+ it('restores a preferred Macro model on a fresh composer', () => {
+ mocks.preferredInmemModel = 'anthropic/claude-sonnet-5';
+ page();
+ expect(screen.getByRole('button', { name: 'Agent' }).textContent).toBe(
+ 'Sonnet 5'
+ );
});
it('restores the most recently used supported agent', async () => {
mocks.recentIds = [CURSOR_BOT_ID];
diff --git a/apps/web/src/features/agents-view/views/NewChatPage.tsx b/apps/web/src/features/agents-view/views/NewChatPage.tsx
index 46aecd604c..7086b9accb 100644
--- a/apps/web/src/features/agents-view/views/NewChatPage.tsx
+++ b/apps/web/src/features/agents-view/views/NewChatPage.tsx
@@ -14,7 +14,9 @@ import { ChatComposer } from '../components/ChatComposer';
import type { AgentKind } from '../core/agent-kind';
import { defaultBranchFor } from '../core/repository';
import { MACRO_PERSONA_ID, type RosterAgent } from '../core/roster';
+import { createPreferredInmemModel } from '../primitives/preferred-inmem-model';
import { createRecentRepositories } from '../primitives/recent-repositories';
+import { createComposerModels } from '../queries/composer-models';
import { createReachableRepositories } from '../queries/reachable-repositories';
import { createRepositoryBranches } from '../queries/repository-branches';
import { AgentPicker } from './AgentPicker';
@@ -47,8 +49,10 @@ export function NewChatPage(props: {
const { openSettings } = useSettingsState();
const recentAgents = createRecentAgentSelections(userId());
const repositories = createRecentRepositories(userId());
+ const preferredInmem = createPreferredInmemModel(userId());
const options = () => props.roster;
const [agentId, setAgentId] = createSignal<string>();
+ /** One-shot model from a coding agent's submenu; Macro uses {@link preferredInmem}. */
const [modelOverride, setModelOverride] = createSignal<string>();
// A new conversation starts on Automatic until the caller picks a repository.
const [repoUrl, setRepoUrl] = createSignal<string | undefined>();
@@ -68,6 +72,28 @@ export function NewChatPage(props: {
MACRO_PERSONA_ID;
return options().find((agent) => agent.id === wanted) ?? options()[0];
});
+ const macro = () => options().find((agent) => agent.id === MACRO_PERSONA_ID);
+ const macroCatalog = createComposerModels(macro);
+ /** Preferred Macro model when it is still in the live in-memory catalog. */
+ const preferredInmemModel = () => {
+ const id = preferredInmem.model();
+ if (!id) return undefined;
+ const catalog = macroCatalog.models();
+ // Until discovery returns, keep the stored id so the trigger can label it.
+ if (catalog.length === 0) return id;
+ return catalog.some((option) => option.id === id) ? id : undefined;
+ };
+ /**
+ * Model shown on the agent control and sent with the next start. Coding
+ * agents use a one-shot override; Macro prefers an in-session pick, then
+ * the remembered Models choice.
+ */
+ const composerModelOverride = () => {
+ if (selected()?.id === MACRO_PERSONA_ID) {
+ return modelOverride() ?? preferredInmemModel();
+ }
+ return modelOverride();
+ };
const coding = () => selected()?.kind === 'coder';
// The create-session API accepts explicit repositories only for Cursor.
const canSelectRepository = () => selected()?.harness === 'cursor';
@@ -116,6 +142,7 @@ export function NewChatPage(props: {
recentAgents.remember(persona.id);
const repo = canSelectRepository() ? repoUrl() : undefined;
if (repo) repositories.remember(repo);
+ const model = composerModelOverride();
props.onStart({
prompt,
...(attachments.length > 0
@@ -124,9 +151,10 @@ export function NewChatPage(props: {
botId: persona.botId,
repoUrl: repo,
...(repo ? { repoBranch: repoBranch() } : {}),
- ...(modelOverride() ? { modelOverride: modelOverride() } : {}),
+ ...(model ? { modelOverride: model } : {}),
});
attachmentTracker.clearAttachments();
+ // Macro's preferred model stays; coding-agent submenu picks are one-shot.
setModelOverride(undefined);
};
@@ -134,10 +162,17 @@ export function NewChatPage(props: {
<AgentPicker
agents={options()}
selected={selected()}
- modelOverride={modelOverride()}
+ modelOverride={composerModelOverride()}
loading={props.rosterLoading}
onSelect={(agent, model) => {
setAgentId(agent.id);
+ if (agent.id === MACRO_PERSONA_ID) {
+ if (model) preferredInmem.remember(model);
+ // Still set the override so the trigger updates when Macro was
+ // already selected (agent id unchanged would otherwise skip a render).
+ setModelOverride(model);
+ return;
+ }
setModelOverride(model);
}}
onConnect={connect}
diff --git a/apps/web/src/features/block-agent/ui/AgentModelSelector.test.tsx b/apps/web/src/features/block-agent/ui/AgentModelSelector.test.tsx
index 4f200821b6..e2cfca87f1 100644
--- a/apps/web/src/features/block-agent/ui/AgentModelSelector.test.tsx
+++ b/apps/web/src/features/block-agent/ui/AgentModelSelector.test.tsx
@@ -22,6 +22,7 @@ const INMEM_MODELS: ModelOption[] = [
'anthropic/claude-sonnet-5',
'anthropic/claude-haiku-4-5',
'openai/gpt-5.5',
+ 'fireworks/kimi-k3',
].map((id) => ({ id, name: id, description: null, group: null }));
let motionStyles: HTMLStyleElement;
@@ -74,6 +75,9 @@ describe('agent model selector', () => {
expect(providerOf(screen.getByRole('menuitem', { name: /^GPT-5.5/ }))).toBe(
'openai'
);
+ expect(providerOf(screen.getByRole('menuitem', { name: /^Kimi K3/ }))).toBe(
+ 'kimi'
+ );
expect(screen.queryByText('anthropic/claude-haiku-4-5')).toBeNull();
fireEvent.keyDown(haiku, { key: 'Enter' });
diff --git a/apps/web/src/lib/core/component/AI/assets/deepseek.svg b/apps/web/src/lib/core/component/AI/assets/deepseek.svg
new file mode 100644
index 0000000000..f01bd509ab
--- /dev/null
+++ b/apps/web/src/lib/core/component/AI/assets/deepseek.svg
@@ -0,0 +1,3 @@
+<svg viewBox="0 0 24 24" fill="currentColor" xmlns="http://www.w3.org/2000/svg">
+ <path d="M23.748 4.651c-.254-.124-.364.113-.512.233-.051.04-.094.09-.137.137-.372.397-.806.657-1.373.626-.829-.046-1.537.214-2.163.848-.133-.782-.575-1.248-1.247-1.548-.352-.155-.708-.311-.955-.65-.172-.24-.219-.509-.305-.774-.055-.16-.11-.323-.293-.35-.2-.031-.278.136-.356.276-.313.572-.434 1.202-.422 1.84.027 1.436.633 2.58 1.838 3.393.137.094.172.187.129.323-.082.28-.18.553-.266.833-.055.179-.137.218-.328.14a5.5 5.5 0 0 1-1.737-1.179c-.857-.828-1.631-1.743-2.597-2.46a12 12 0 0 0-.689-.47c-.985-.957.13-1.743.387-1.836.27-.098.094-.433-.778-.428-.872.003-1.67.295-2.687.685a3 3 0 0 1-.465.136 9.6 9.6 0 0 0-2.883-.101c-1.885.21-3.39 1.1-4.497 2.622C.082 8.776-.231 10.854.152 13.02c.403 2.284 1.568 4.175 3.36 5.653 1.857 1.533 3.997 2.284 6.438 2.14 1.482-.085 3.132-.284 4.994-1.86.47.234.962.328 1.78.398.629.058 1.235-.031 1.705-.129.735-.155.684-.836.418-.961-2.155-1.004-1.682-.595-2.112-.926 1.095-1.295 2.768-3.598 3.284-6.733.05-.346.115-.834.108-1.114-.004-.171.035-.238.23-.257a4.2 4.2 0 0 0 1.545-.475c1.397-.763 1.96-2.016 2.093-3.517.02-.23-.004-.467-.247-.588M11.58 18.168c-2.088-1.642-3.101-2.183-3.52-2.16-.39.024-.32.472-.234.763.09.288.207.487.371.74.114.167.192.416-.113.603-.673.416-1.842-.14-1.897-.168-1.361-.801-2.5-1.86-3.301-3.306-.775-1.393-1.225-2.888-1.299-4.482-.02-.385.094-.522.477-.592a4.7 4.7 0 0 1 1.53-.038c2.131.311 3.946 1.264 5.467 2.774.868.86 1.525 1.887 2.202 2.89.72 1.066 1.494 2.082 2.48 2.915.348.291.626.513.892.677-.802.09-2.14.109-3.055-.615zm1.001-6.44a.306.306 0 0 1 .415-.287.3.3 0 0 1 .113.074.3.3 0 0 1 .086.214c0 .17-.136.307-.308.307a.303.303 0 0 1-.306-.307m3.11 1.596c-.2.081-.4.151-.591.16a1.25 1.25 0 0 1-.798-.254c-.274-.23-.47-.358-.551-.758a1.7 1.7 0 0 1 .015-.588c.07-.327-.007-.537-.238-.727-.188-.156-.426-.199-.689-.199a.6.6 0 0 1-.254-.078.253.253 0 0 1-.114-.358 1 1 0 0 1 .192-.21c.356-.202.767-.136 1.146.016.352.144.618.408 1.001.782.392.451.462.576.685.915.176.264.336.536.446.848.066.194-.02.353-.25.45" />
+</svg>
diff --git a/apps/web/src/lib/core/component/AI/assets/kimi.svg b/apps/web/src/lib/core/component/AI/assets/kimi.svg
new file mode 100644
index 0000000000..7b6adb370d
--- /dev/null
+++ b/apps/web/src/lib/core/component/AI/assets/kimi.svg
@@ -0,0 +1,4 @@
+<svg viewBox="0 0 24 25" fill="currentColor" xmlns="http://www.w3.org/2000/svg">
+ <path d="M21.7202 0.939941C22.9502 0.939941 23.9502 1.93994 23.9502 3.16994C23.9502 4.39994 22.9502 5.39994 21.7202 5.39994H19.7502C19.6002 5.39994 19.4902 5.27994 19.4902 5.13994V3.16994C19.4902 1.93994 20.4902 0.939941 21.7202 0.939941Z" />
+ <path d="M9.39 13.9501L17.82 5.59012C17.98 5.43012 17.89 5.12012 17.68 5.12012H13.14C13.14 5.12012 13.04 5.14012 13 5.18012L3.92 14.1901C3.78 14.3301 3.57 14.2101 3.57 13.9801V5.39012C3.57 5.24012 3.47 5.12012 3.35 5.12012H0.219999C0.0999993 5.12012 0 5.24012 0 5.39012V23.9201C0 24.0701 0.0999993 24.1901 0.219999 24.1901H3.35C3.47 24.1901 3.57 24.0701 3.57 23.9201V20.1401C3.57 20.0601 3.6 19.9801 3.65 19.9301L6.47 17.1401C6.54 17.0701 6.63 17.0601 6.71 17.1101L14.24 22.6501C15.47 23.4801 16.85 23.9901 18.25 24.1401C18.37 24.1501 18.48 24.0301 18.48 23.8701V20.3101C18.48 20.1701 18.4 20.0601 18.29 20.0501C17.47 19.9201 16.66 19.6001 15.94 19.1101L9.42 14.3901C9.28 14.3001 9.27 14.0701 9.39 13.9501Z" />
+</svg>
diff --git a/apps/web/src/lib/core/component/AI/assets/muse.svg b/apps/web/src/lib/core/component/AI/assets/muse.svg
new file mode 100644
index 0000000000..2238b797cf
--- /dev/null
+++ b/apps/web/src/lib/core/component/AI/assets/muse.svg
@@ -0,0 +1,3 @@
+<svg viewBox="0 0 24 24" fill="currentColor" xmlns="http://www.w3.org/2000/svg">
+ <path d="M6.915 4.03c-1.968 0-3.683 1.28-4.871 3.113C.704 9.208 0 11.883 0 14.449c0 .706.07 1.369.21 1.973a6.624 6.624 0 0 0 .265.86 5.297 5.297 0 0 0 .371.761c.696 1.159 1.818 1.927 3.593 1.927 1.497 0 2.633-.671 3.965-2.444.76-1.012 1.144-1.626 2.663-4.32l.756-1.339.186-.325c.061.1.121.196.183.3l2.152 3.595c.724 1.21 1.665 2.556 2.47 3.314 1.046.987 1.992 1.22 3.06 1.22 1.075 0 1.876-.355 2.455-.843a3.743 3.743 0 0 0 .81-.973c.542-.939.861-2.127.861-3.745 0-2.72-.681-5.357-2.084-7.45-1.282-1.912-2.957-2.93-4.716-2.93-1.047 0-2.088.467-3.053 1.308-.652.57-1.257 1.29-1.82 2.05-.69-.875-1.335-1.547-1.958-2.056-1.182-.966-2.315-1.303-3.454-1.303zm10.16 2.053c1.147 0 2.188.758 2.992 1.999 1.132 1.748 1.647 4.195 1.647 6.4 0 1.548-.368 2.9-1.839 2.9-.58 0-1.027-.23-1.664-1.004-.496-.601-1.343-1.878-2.832-4.358l-.617-1.028a44.908 44.908 0 0 0-1.255-1.98c.07-.109.141-.224.211-.327 1.12-1.667 2.118-2.602 3.358-2.602zm-10.201.553c1.265 0 2.058.791 2.675 1.446.307.327.737.871 1.234 1.579l-1.02 1.566c-.757 1.163-1.882 3.017-2.837 4.338-1.191 1.649-1.81 1.817-2.486 1.817-.524 0-1.038-.237-1.383-.794-.263-.426-.464-1.13-.464-2.046 0-2.221.63-4.535 1.66-6.088.454-.687.964-1.226 1.533-1.533a2.264 2.264 0 0 1 1.088-.285z" />
+</svg>
diff --git a/apps/web/src/lib/core/component/AI/component/ProviderIcon.test.tsx b/apps/web/src/lib/core/component/AI/component/ProviderIcon.test.tsx
index 408cad3901..b67dbb96c1 100644
--- a/apps/web/src/lib/core/component/AI/component/ProviderIcon.test.tsx
+++ b/apps/web/src/lib/core/component/AI/component/ProviderIcon.test.tsx
@@ -17,6 +17,14 @@ describe('AI provider icons', () => {
['o3', 'openai'],
['google/gemini-pro', 'google'],
['gemini-2.5-pro', 'google'],
+ ['fireworks/kimi-k3', 'kimi'],
+ ['kimi-k3', 'kimi'],
+ ['fireworks/deepseek-v4-pro-0813', 'deepseek'],
+ ['deepseek-v4-pro-0813', 'deepseek'],
+ ['fireworks/muse-glimmer-30b', 'muse'],
+ ['muse-glimmer-30b', 'muse'],
+ ['google/gemini-3.8-flash', 'google'],
+ ['gemini-3.8-flash', 'google'],
])('recognizes %s as %s', (model, provider) => {
expect(modelProvider(model)).toBe(provider);
});
diff --git a/apps/web/src/lib/core/component/AI/component/ProviderIcon.tsx b/apps/web/src/lib/core/component/AI/component/ProviderIcon.tsx
index 276f17d51c..fd35801b94 100644
--- a/apps/web/src/lib/core/component/AI/component/ProviderIcon.tsx
+++ b/apps/web/src/lib/core/component/AI/component/ProviderIcon.tsx
@@ -3,9 +3,18 @@ import SparkleIcon from '@phosphor/sparkle.svg';
import GoogleIcon from '@phosphor-fill/google-logo-fill.svg';
import { type Component, type JSX, Show } from 'solid-js';
import { Dynamic } from 'solid-js/web';
+import DeepSeekIcon from '../assets/deepseek.svg';
+import KimiIcon from '../assets/kimi.svg';
+import MuseIcon from '../assets/muse.svg';
import OpenAiIcon from '../assets/openai.svg';
-type Provider = 'anthropic' | 'openai' | 'google';
+type Provider =
+ | 'anthropic'
+ | 'openai'
+ | 'google'
+ | 'kimi'
+ | 'deepseek'
+ | 'muse';
/** Accept both routed model ids and the bare ids reported by agent runtimes. */
export function modelProvider(
@@ -13,6 +22,11 @@ export function modelProvider(
): Provider | undefined {
const id = model?.trim().toLowerCase();
if (!id) return undefined;
+ const slug = id.replace(/^[\w.-]+\//, '');
+ if (slug.startsWith('kimi') || slug.includes('kimi-k')) return 'kimi';
+ if (slug.startsWith('deepseek') || slug.includes('deepseek'))
+ return 'deepseek';
+ if (slug.startsWith('muse')) return 'muse';
if (
id.startsWith('anthropic/') ||
/^(claude|sonnet|opus|haiku)(-|$)/.test(id)
@@ -34,6 +48,9 @@ const icons: Record<
anthropic: ClaudeIcon,
openai: OpenAiIcon,
google: GoogleIcon,
+ kimi: KimiIcon,
+ deepseek: DeepSeekIcon,
+ muse: MuseIcon,
};
/** Unknown/loading providers reserve space instead of showing a misleading logo. */
diff --git a/apps/web/src/lib/core/component/AI/constant/model-label.test.ts b/apps/web/src/lib/core/component/AI/constant/model-label.test.ts
index 13b692589e..61f2178f89 100644
--- a/apps/web/src/lib/core/component/AI/constant/model-label.test.ts
+++ b/apps/web/src/lib/core/component/AI/constant/model-label.test.ts
@@ -7,6 +7,14 @@ describe('modelLabel', () => {
expect(modelLabel('openai/gpt-5.6-mini')).toBe('GPT-5.6 mini');
// Runtimes that drop the provider segment still get the house name.
expect(modelLabel('claude-haiku-4-5')).toBe('Haiku 4.5');
+ expect(modelLabel('fireworks/kimi-k3')).toBe('Kimi K3');
+ expect(modelLabel('kimi-k3')).toBe('Kimi K3');
+ expect(modelLabel('fireworks/deepseek-v4-pro-0813')).toBe(
+ 'DeepSeek V4 Pro'
+ );
+ expect(modelLabel('fireworks/muse-glimmer-30b')).toBe('Muse Glimmer');
+ expect(modelLabel('google/gemini-3.8-flash')).toBe('Gemini 3.8 Flash');
+ expect(modelLabel('gemini-3.8-flash')).toBe('Gemini 3.8 Flash');
});
it('reads an unknown slug as a name rather than showing it raw', () => {
diff --git a/apps/web/src/lib/core/component/AI/constant/model-label.ts b/apps/web/src/lib/core/component/AI/constant/model-label.ts
index a0e6798cce..bcf088d3b7 100644
--- a/apps/web/src/lib/core/component/AI/constant/model-label.ts
+++ b/apps/web/src/lib/core/component/AI/constant/model-label.ts
@@ -13,6 +13,18 @@
import { MODEL_PRETTYNAME, type Model } from './model';
+/**
+ * House names for in-memory Macro models that are not in the closed chat
+ * {@link Model} set. Keyed by the bare slug so a routed id and the bare id a
+ * session row may store both resolve to one entry.
+ */
+const HARNESS_PRETTYNAME: Record<string, string> = {
+ 'kimi-k3': 'Kimi K3',
+ 'deepseek-v4-pro-0813': 'DeepSeek V4 Pro',
+ 'muse-glimmer-30b': 'Muse Glimmer',
+ 'gemini-3.8-flash': 'Gemini 3.8 Flash',
+};
+
/** Vendor acronyms that read as shouting only when they are not shouted. */
const ACRONYMS = new Set(['gpt', 'ai', 'llm']);
@@ -67,12 +79,13 @@ export function modelLabel(id: string | undefined, name?: string): string {
if (!id) return 'Model';
const pretty =
MODEL_PRETTYNAME[id as Model] ??
- MODEL_PRETTYNAME[`anthropic/${id}` as Model];
+ MODEL_PRETTYNAME[`anthropic/${id}` as Model] ??
+ HARNESS_PRETTYNAME[id.replace(/^[\w.-]+\//, '')];
if (pretty) return pretty;
const reported = name?.trim();
if (reported && reported !== id.trim())
return reported
- .replace(/^(anthropic|openai|google)\//, '')
+ .replace(/^(anthropic|openai|google|fireworks)\//, '')
.replace(/^Claude /, '');
return humanizeModelId(id);
}
diff --git a/crates/agent/src/model/openai.rs b/crates/agent/src/model/openai.rs
index 0724b33f44..de3b67f94a 100644
--- a/crates/agent/src/model/openai.rs
+++ b/crates/agent/src/model/openai.rs
@@ -3,13 +3,26 @@ use std::sync::Arc;
use crate::model::types::Model;
use rig_core::{client::CompletionClient, providers::openai};
+/// Fireworks' Chat Completions `model` field is the account-scoped path, not
+/// the short catalog slug. Routing ids stay `fireworks/<slug>` so the picker,
+/// usage, and pricing can use the bare name.
+fn completion_model_id(model: &Model<'_>) -> String {
+ if model.provider() == "fireworks" && !model.name().starts_with("accounts/") {
+ format!("accounts/fireworks/models/{}", model.name())
+ } else {
+ model.name().to_string()
+ }
+}
+
/// A model served over the OpenAI-compatible **Chat Completions** API.
///
/// OpenAI-compatible providers differ only by the
/// [`CompletionsClient`](openai::CompletionsClient)'s base URL and key, never
/// by Rust type, so routing can hold any number of them without new variants.
/// Which provider serves an id is decided by routing (the `provider/…`
-/// segment), so there is no id classification here.
+/// segment). Fireworks is the one exception on the wire: its API wants the
+/// account-scoped path, which [`completion_model_id`] derives from the
+/// catalog slug.
pub struct OpenAiChatCompletionsModel<'a> {
model: Model<'a>,
client: Arc<openai::CompletionsClient>,
@@ -31,7 +44,8 @@ impl<'a> OpenAiChatCompletionsModel<'a> {
/// Unlike the Responses API, the Chat Completions API does not coerce tools
/// into a strict subset, so tools are already sent verbatim.
pub fn completion(&self) -> openai::completion::CompletionModel {
- self.client.completion_model(self.model.name().to_string())
+ self.client
+ .completion_model(completion_model_id(&self.model))
}
/// Best-effort reasoning config, flattened into the request by rig, or
@@ -124,3 +138,6 @@ impl<'a> OpenAiResponsesModel<'a> {
}))
}
}
+
+#[cfg(test)]
+mod test;
diff --git a/crates/agent/src/model/openai/test.rs b/crates/agent/src/model/openai/test.rs
new file mode 100644
index 0000000000..1edf6360cf
--- /dev/null
+++ b/crates/agent/src/model/openai/test.rs
@@ -0,0 +1,30 @@
+use super::completion_model_id;
+use crate::model::types::Model;
+
+#[test]
+fn fireworks_short_name_becomes_the_account_path() {
+ let model = Model::try_from("fireworks/kimi-k3").expect("routable id");
+ assert_eq!(
+ completion_model_id(&model),
+ "accounts/fireworks/models/kimi-k3"
+ );
+}
+
+#[test]
+fn fireworks_account_path_is_sent_verbatim() {
+ let model = Model::try_from("fireworks/accounts/fireworks/models/kimi-k3")
+ .expect("full path is still one provider segment");
+ assert_eq!(
+ completion_model_id(&model),
+ "accounts/fireworks/models/kimi-k3"
+ );
+}
+
+#[test]
+fn other_compatible_providers_keep_the_catalog_name() {
+ let model = Model::try_from("cerebras/gpt-oss-120b").expect("routable id");
+ assert_eq!(completion_model_id(&model), "gpt-oss-120b");
+
+ let gemini = Model::try_from("google/gemini-3.8-flash").expect("routable id");
+ assert_eq!(completion_model_id(&gemini), "gemini-3.8-flash");
+}
diff --git a/crates/agent/src/model/router.rs b/crates/agent/src/model/router.rs
index 0ce53a704d..478bdc3d8e 100644
--- a/crates/agent/src/model/router.rs
+++ b/crates/agent/src/model/router.rs
@@ -12,7 +12,7 @@
//! (e.g. `agent_loop`) hold one type and never fan out.
//!
//! Ids are addressed as `provider/model` (e.g. `anthropic/claude-opus-4-8`,
-//! `groq/llama-3.3-70b`); routing picks the provider from the segment, never by
+//! `fireworks/kimi-k3`); routing picks the provider from the segment, never by
//! sniffing the id. Unroutable ids fall back to the default model.
use std::collections::HashMap;
@@ -45,7 +45,10 @@ env_var! {
struct ApiKeys {
AnthropicApiKey,
OpenaiApiKey,
- CerebrasApiKey
+ CerebrasApiKey,
+ /// Doppler name is `FIREWORK_API_KEY` (singular), from `shared_ai`.
+ FireworkApiKey,
+ GoogleGenerativeAiApiKey
}
}
@@ -58,6 +61,17 @@ const OPENAI_PROVIDER: &str = "openai";
const CEREBRAS_PROVIDER: &str = "cerebras";
/// Cerebras inference endpoint (OpenAI-compatible Chat Completions API).
const CEREBRAS_BASE_URL: &str = "https://api.cerebras.ai/v1";
+/// Provider segment Fireworks is registered under (OpenAI-compatible Chat
+/// Completions). Open-weight models on the in-memory Macro agent route here.
+const FIREWORKS_PROVIDER: &str = "fireworks";
+/// Fireworks inference endpoint (OpenAI-compatible Chat Completions API).
+const FIREWORKS_BASE_URL: &str = "https://api.fireworks.ai/inference/v1";
+/// Provider segment Google Gemini is registered under (OpenAI-compatible Chat
+/// Completions).
+const GOOGLE_PROVIDER: &str = "google";
+/// Gemini's OpenAI-compatible Chat Completions endpoint. The native
+/// `generativelanguage` API is a different wire format and is not used here.
+const GOOGLE_BASE_URL: &str = "https://generativelanguage.googleapis.com/v1beta/openai";
/// A routed model id bound to the provider client that serves it.
pub(crate) enum RoutedModel<'a> {
@@ -266,8 +280,9 @@ impl ModelRouter {
/// Build a router with the built-in providers from the environment.
///
- /// Requires `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, and `CEREBRAS_API_KEY`.
- /// Chain [`with_openai_provider`](Self::with_openai_provider) to add more.
+ /// Requires `ANTHROPIC_API_KEY`, `OPENAI_API_KEY`, `CEREBRAS_API_KEY`,
+ /// `FIREWORK_API_KEY`, and `GOOGLE_GENERATIVE_AI_API_KEY`. Chain
+ /// [`with_openai_provider`](Self::with_openai_provider) to add more.
pub fn try_from_env() -> Result<Self, AgentError> {
let env = ApiKeys::new()?;
let anthropic = anthropic::Client::builder()
@@ -278,13 +293,21 @@ impl ModelRouter {
let openai = openai::Client::builder()
.api_key(env.openai_api_key.to_string())
.build()?;
- // Cerebras speaks the OpenAI Chat Completions API, so it rides the
- // compatible-provider registry: `cerebras/<model>` ids route to it.
- Self::new(anthropic, openai).with_openai_provider(
- CEREBRAS_PROVIDER,
- CEREBRAS_BASE_URL,
- &env.cerebras_api_key,
- )
+ // Cerebras, Fireworks, and Gemini all speak the OpenAI Chat Completions
+ // API, so they ride the compatible-provider registry: a `<provider>/`
+ // segment routes to the client registered under that name.
+ Self::new(anthropic, openai)
+ .with_openai_provider(CEREBRAS_PROVIDER, CEREBRAS_BASE_URL, &env.cerebras_api_key)?
+ .with_openai_provider(
+ FIREWORKS_PROVIDER,
+ FIREWORKS_BASE_URL,
+ &env.firework_api_key,
+ )?
+ .with_openai_provider(
+ GOOGLE_PROVIDER,
+ GOOGLE_BASE_URL,
+ &env.google_generative_ai_api_key,
+ )
}
/// The process-wide full router, built from the environment on first use.
diff --git a/crates/agent/src/test/test_model_types.rs b/crates/agent/src/test/test_model_types.rs
index 1d9e7dc694..4679efb97a 100644
--- a/crates/agent/src/test/test_model_types.rs
+++ b/crates/agent/src/test/test_model_types.rs
@@ -8,4 +8,8 @@ fn test_from_string() {
let bad = "claude-big-burger";
assert!(Model::try_from(bad).is_err());
+
+ let fireworks = Model::try_from("fireworks/kimi-k3").expect("short fireworks id");
+ assert_eq!(fireworks.provider(), "fireworks");
+ assert_eq!(fireworks.name(), "kimi-k3");
}
diff --git a/crates/agent/src/test/test_router.rs b/crates/agent/src/test/test_router.rs
index 128d8e4ba6..ef87b73263 100644
--- a/crates/agent/src/test/test_router.rs
+++ b/crates/agent/src/test/test_router.rs
@@ -73,3 +73,37 @@ fn registered_openai_compatible_provider_routes_to_chat_completions() {
RoutedModel::OpenAiChatCompletions(_)
));
}
+
+#[test]
+fn fireworks_provider_routes_to_chat_completions() {
+ let router = test_router().with_openai_client(
+ "fireworks",
+ openai::CompletionsClient::builder()
+ .api_key("test-fireworks-key")
+ .base_url("https://api.fireworks.ai/inference/v1")
+ .build()
+ .unwrap(),
+ );
+
+ let routed = router.route("fireworks/kimi-k3").unwrap();
+ assert!(matches!(routed, RoutedModel::OpenAiChatCompletions(_)));
+ assert_eq!(routed.provider(), "fireworks");
+ assert_eq!(routed.model_name(), "kimi-k3");
+}
+
+#[test]
+fn google_provider_routes_to_chat_completions() {
+ let router = test_router().with_openai_client(
+ "google",
+ openai::CompletionsClient::builder()
+ .api_key("test-google-key")
+ .base_url("https://generativelanguage.googleapis.com/v1beta/openai")
+ .build()
+ .unwrap(),
+ );
+
+ let routed = router.route("google/gemini-3.8-flash").unwrap();
+ assert!(matches!(routed, RoutedModel::OpenAiChatCompletions(_)));
+ assert_eq!(routed.provider(), "google");
+ assert_eq!(routed.model_name(), "gemini-3.8-flash");
+}
diff --git a/crates/agent_inmem/src/domain/mod.rs b/crates/agent_inmem/src/domain/mod.rs
index ea26c77042..669b5bd193 100644
--- a/crates/agent_inmem/src/domain/mod.rs
+++ b/crates/agent_inmem/src/domain/mod.rs
@@ -6,6 +6,8 @@ pub mod engine;
pub mod mcp;
/// ACP model configuration generated from the turn engine catalog.
pub mod model_options;
+/// Models this runtime advertises, including routed Fireworks and Gemini ids.
+pub mod models;
pub mod replay;
pub mod session;
pub mod user_input;
diff --git a/crates/agent_inmem/src/domain/model_options.rs b/crates/agent_inmem/src/domain/model_options.rs
index c306562e6b..afe4bc5dd1 100644
--- a/crates/agent_inmem/src/domain/model_options.rs
+++ b/crates/agent_inmem/src/domain/model_options.rs
@@ -11,7 +11,10 @@ pub fn model_config_options(current: &str, models: &[&str]) -> Vec<SessionConfig
let options: Vec<_> = models
.iter()
.map(|model| {
- SessionConfigSelectOption::new(SessionConfigValueId::new((*model).to_owned()), *model)
+ SessionConfigSelectOption::new(
+ SessionConfigValueId::new((*model).to_owned()),
+ super::models::display_name(model),
+ )
})
.collect();
vec![SessionConfigOption::select(
@@ -21,3 +24,6 @@ pub fn model_config_options(current: &str, models: &[&str]) -> Vec<SessionConfig
options,
)]
}
+
+#[cfg(test)]
+mod test;
diff --git a/crates/agent_inmem/src/domain/model_options/test.rs b/crates/agent_inmem/src/domain/model_options/test.rs
new file mode 100644
index 0000000000..71f0e672ac
--- /dev/null
+++ b/crates/agent_inmem/src/domain/model_options/test.rs
@@ -0,0 +1,23 @@
+use super::model_config_options;
+use agent_fold::domain::model_selection::model_selection;
+
+#[test]
+fn routed_options_use_house_names() {
+ let options = model_config_options(
+ "fireworks/kimi-k3",
+ &["fireworks/kimi-k3", "anthropic/claude-sonnet-5"],
+ );
+ let selection = model_selection(&options).expect("a model select");
+ assert_eq!(selection.current, "fireworks/kimi-k3");
+ assert_eq!(
+ selection
+ .options
+ .iter()
+ .map(|model| (model.id.as_str(), model.name.as_str()))
+ .collect::<Vec<_>>(),
+ vec![
+ ("fireworks/kimi-k3", "Kimi K3"),
+ ("anthropic/claude-sonnet-5", "anthropic/claude-sonnet-5"),
+ ]
+ );
+}
diff --git a/crates/agent_inmem/src/domain/models.rs b/crates/agent_inmem/src/domain/models.rs
new file mode 100644
index 0000000000..f5d4d54700
--- /dev/null
+++ b/crates/agent_inmem/src/domain/models.rs
@@ -0,0 +1,46 @@
+//! Models the in-memory Macro agent advertises over ACP.
+
+use std::sync::OnceLock;
+
+use chat::domain::models::CHAT_MODELS;
+
+/// Models the in-memory agent offers beyond the closed chat catalog, paired
+/// with the house name the picker shows.
+///
+/// Catalog ids are `<provider>/<slug>`, matching the router's provider
+/// registry. The router expands a Fireworks slug to its
+/// `accounts/fireworks/models/<slug>` wire id; other providers send the slug
+/// verbatim.
+pub const ROUTED_MODELS: &[(&str, &str)] = &[
+ ("fireworks/kimi-k3", "Kimi K3"),
+ ("fireworks/deepseek-v4-pro-0813", "DeepSeek V4 Pro"),
+ ("fireworks/muse-glimmer-30b", "Muse Glimmer"),
+ ("google/gemini-3.8-flash", "Gemini 3.8 Flash"),
+];
+
+/// House name for an advertised id, or the id itself when the catalog has
+/// none (closed chat models keep their frontend names).
+#[must_use]
+pub fn display_name(id: &str) -> &str {
+ ROUTED_MODELS
+ .iter()
+ .find_map(|(model, name)| (*model == id).then_some(*name))
+ .unwrap_or(id)
+}
+
+/// Closed chat models plus the routed extras, in picker order.
+#[must_use]
+pub fn advertised_models() -> &'static [&'static str] {
+ static MODELS: OnceLock<Vec<&'static str>> = OnceLock::new();
+ MODELS
+ .get_or_init(|| {
+ let mut models = Vec::with_capacity(CHAT_MODELS.len() + ROUTED_MODELS.len());
+ models.extend_from_slice(CHAT_MODELS);
+ models.extend(ROUTED_MODELS.iter().map(|(model, _)| *model));
+ models
+ })
+ .as_slice()
+}
+
+#[cfg(test)]
+mod test;
diff --git a/crates/agent_inmem/src/domain/models/test.rs b/crates/agent_inmem/src/domain/models/test.rs
new file mode 100644
index 0000000000..13612d961b
--- /dev/null
+++ b/crates/agent_inmem/src/domain/models/test.rs
@@ -0,0 +1,44 @@
+use super::{ROUTED_MODELS, advertised_models, display_name};
+use chat::domain::models::CHAT_MODELS;
+
+#[test]
+fn advertised_models_keep_chat_then_append_routed() {
+ let models = advertised_models();
+ assert_eq!(&models[..CHAT_MODELS.len()], CHAT_MODELS);
+ assert_eq!(
+ &models[CHAT_MODELS.len()..],
+ ROUTED_MODELS
+ .iter()
+ .map(|(model, _)| *model)
+ .collect::<Vec<_>>()
+ .as_slice()
+ );
+}
+
+#[test]
+fn routed_models_have_house_names() {
+ assert_eq!(display_name("fireworks/kimi-k3"), "Kimi K3");
+ assert_eq!(
+ display_name("fireworks/deepseek-v4-pro-0813"),
+ "DeepSeek V4 Pro"
+ );
+ assert_eq!(display_name("fireworks/muse-glimmer-30b"), "Muse Glimmer");
+ assert_eq!(display_name("google/gemini-3.8-flash"), "Gemini 3.8 Flash");
+ assert_eq!(
+ display_name("anthropic/claude-sonnet-5"),
+ "anthropic/claude-sonnet-5"
+ );
+}
+
+/// Every advertised id must carry a provider segment the router registers,
+/// or the picker would offer an id that silently falls back.
+#[test]
+fn routed_models_use_known_providers() {
+ for (model, _) in ROUTED_MODELS {
+ let provider = model.split('/').next().expect("a provider segment");
+ assert!(
+ matches!(provider, "fireworks" | "google"),
+ "unregistered provider in {model}"
+ );
+ }
+}
diff --git a/crates/agent_inmem/src/rig_engine.rs b/crates/agent_inmem/src/rig_engine.rs
index b1f8f56c92..a04deaca8f 100644
--- a/crates/agent_inmem/src/rig_engine.rs
+++ b/crates/agent_inmem/src/rig_engine.rs
@@ -97,7 +97,7 @@ fn tools_for_turn(
impl TurnEngine for RigTurnEngine {
fn supported_models(&self) -> &[&str] {
- chat::domain::models::CHAT_MODELS
+ crate::domain::models::advertised_models()
}
fn run_turn(&self, request: TurnRequest) -> mpsc::Receiver<Result<StreamPart, AgentError>> {
diff --git a/crates/ai_usage/src/domain/service/test.rs b/crates/ai_usage/src/domain/service/test.rs
index ceb88ccbfd..9efe6406de 100644
--- a/crates/ai_usage/src/domain/service/test.rs
+++ b/crates/ai_usage/src/domain/service/test.rs
@@ -105,6 +105,11 @@ fn normalize_model_id_strips_provider_prefix() {
"claude-opus-5"
);
assert_eq!(normalize_model_id("openai/gpt-5.6"), "gpt-5.6");
+ assert_eq!(normalize_model_id("fireworks/kimi-k3"), "kimi-k3");
+ assert_eq!(
+ normalize_model_id("google/gemini-3.8-flash"),
+ "gemini-3.8-flash"
+ );
assert_eq!(normalize_model_id("claude-haiku-4-5"), "claude-haiku-4-5");
// A trailing slash is not a provider prefix; leave the id alone.
assert_eq!(normalize_model_id("weird/"), "weird/");
diff --git a/crates/macro_db_client/migrations/20260924221223_seed_fireworks_open_weight_pricing.sql b/crates/macro_db_client/migrations/20260924221223_seed_fireworks_open_weight_pricing.sql
new file mode 100644
index 0000000000..df6191648b
--- /dev/null
+++ b/crates/macro_db_client/migrations/20260924221223_seed_fireworks_open_weight_pricing.sql
@@ -0,0 +1,10 @@
+-- Seed Fireworks serverless prices (USD per million tokens) for the
+-- open-weight models advertised on the in-memory Macro agent.
+-- Standard-tier rates from Fireworks' serverless pricing page:
+-- Kimi K3 $3 / $15, DeepSeek V4 Pro $1.32 / $3.96, Muse Glimmer $0.35 / $1.50.
+-- ON CONFLICT keeps any price already set at runtime via set_pricing.
+INSERT INTO ai_pricing (model, price_per_million_in, price_per_million_out) VALUES
+ ('kimi-k3', 3.0, 15.0),
+ ('deepseek-v4-pro-0813', 1.32, 3.96),
+ ('muse-glimmer-30b', 0.35, 1.50)
+ON CONFLICT (model) DO NOTHING;
diff --git a/crates/macro_db_client/migrations/20260925001529_seed_gemini_3_8_flash_pricing.sql b/crates/macro_db_client/migrations/20260925001529_seed_gemini_3_8_flash_pricing.sql
new file mode 100644
index 0000000000..4d737f1947
--- /dev/null
+++ b/crates/macro_db_client/migrations/20260925001529_seed_gemini_3_8_flash_pricing.sql
@@ -0,0 +1,8 @@
+-- Seed pricing for Gemini 3.8 Flash (USD per million tokens), offered on the
+-- in-memory Macro agent. Standard-tier rates from Google's Gemini API pricing
+-- page: $0.75 in / $3.75 out through 2026-12-31, rising to $1.50 / $7.50 on
+-- 2027-01-01. ON CONFLICT keeps any price already set at runtime via the
+-- set_pricing endpoint, which is also how the January increase gets applied.
+INSERT INTO ai_pricing (model, price_per_million_in, price_per_million_out) VALUES
+ ('gemini-3.8-flash', 0.75, 3.75)
+ON CONFLICT (model) DO NOTHING;
diff --git a/docs/AGENT_GUIDE/ai-chat.md b/docs/AGENT_GUIDE/ai-chat.md
index b14e447d32..9b8cc83db9 100644
--- a/docs/AGENT_GUIDE/ai-chat.md
+++ b/docs/AGENT_GUIDE/ai-chat.md
@@ -66,10 +66,14 @@ permission failures should display a failed tool call without a successful resul
- The agent dropdown includes every saved agent regardless of runtime, plus Cursor,
grouped in **Models**, **Agents**, and **Coding agents** sections. **Models** lists
Macro’s available models with readable names (for example, **Sonnet 5**) and
- provider icons aligned with the agent icons. The chat catalog offers Sonnet 5,
- Opus 5, and Haiku 4.5; older Sonnet and Opus versions are not offered.
+ provider or model icons aligned with the agent icons. The in-memory catalog
+ offers the closed Anthropic and OpenAI chat models, Fireworks-hosted
+ **Kimi K3**, **DeepSeek V4 Pro**, and **Muse Glimmer**, and Google's
+ **Gemini 3.8 Flash**; older Sonnet and Opus versions are not offered.
Selecting a model here selects
the default runtime and applies that model to the next send, retracting the repository drawer.
+ A model chosen from that **Models** list is remembered in local storage as the
+ default for Macro's in-memory agent until another Models entry is picked.
The built-in Macro agent is the only agent excluded from these sections; its models remain available.
Unavailable paired agents stay visible with a reason. Model discovery uses the
selected runtime, including Claude Cloud. Every coding agent opens the repository
diff --git a/tooling/xtask/crates/xtask_local/src/local/local_env.rs b/tooling/xtask/crates/xtask_local/src/local/local_env.rs
index c1985d4a54..3049646afe 100644
--- a/tooling/xtask/crates/xtask_local/src/local/local_env.rs
+++ b/tooling/xtask/crates/xtask_local/src/local/local_env.rs
@@ -707,6 +707,14 @@ impl BootStubEnv {
env.insert("LIVEKIT_API_KEY".into(), "local-livekit-key".into());
env.insert("LIVEKIT_API_SECRET".into(), "local-livekit-secret".into());
env.insert("OPENAI_API_KEY".into(), "local-openai-key".into());
+ // Required by the agent router. Present so it builds on a stack with no
+ // Doppler; real provider calls still fail on the dummy keys.
+ // Doppler's shared_ai name is singular: FIREWORK_API_KEY.
+ env.insert("FIREWORK_API_KEY".into(), "local-firework-key".into());
+ env.insert(
+ "GOOGLE_GENERATIVE_AI_API_KEY".into(),
+ "local-google-generative-ai-key".into(),
+ );
env.insert("COHERE_API_KEY".into(), "local-cohere-key".into());
env.insert(
"CAL_WEBHOOK_SECRET_KEY".into(),
diff --git a/tooling/xtask/crates/xtask_local/src/local/local_env/test.rs b/tooling/xtask/crates/xtask_local/src/local/local_env/test.rs
index ce13efc45d..0099834127 100644
--- a/tooling/xtask/crates/xtask_local/src/local/local_env/test.rs
+++ b/tooling/xtask/crates/xtask_local/src/local/local_env/test.rs
@@ -84,6 +84,8 @@ fn emits_required_keys() {
"LIVEKIT_API_KEY",
"LIVEKIT_API_SECRET",
"OPENAI_API_KEY",
+ "FIREWORK_API_KEY",
+ "GOOGLE_GENERATIVE_AI_API_KEY",
"COHERE_API_KEY",
"CAL_WEBHOOK_SECRET_KEY",
"CAL_EVENT_TYPE_CONTENT_NAMES_KEY",