diff --git a/frontend/src/lib/aiStore.test.ts b/frontend/src/lib/aiStore.test.ts index be8980e93a..22c841c91a 100644 --- a/frontend/src/lib/aiStore.test.ts +++ b/frontend/src/lib/aiStore.test.ts @@ -1,16 +1,53 @@ import { describe, expect, it, vi } from 'vitest' import { get } from 'svelte/store' -// aiStore pulls in the (heavy) copilot lib for workspaceAIClients and the gen -// client for WorkspaceService; neither is exercised by setCopilotInfo. +let requests: { workspace: string; resolve: (config: any) => void }[] = [] + +// aiStore pulls in the (heavy) copilot lib for workspaceAIClients; setCopilotInfo does not +// exercise it. vi.mock('./components/copilot/lib', () => ({ workspaceAIClients: { init: () => {} } })) vi.mock('./gen', () => ({ - WorkspaceService: { getCopilotInfo: async () => ({}) } + WorkspaceService: { + getCopilotInfo: ({ workspace }: { workspace: string }) => + new Promise((resolve) => requests.push({ workspace, resolve })) + } })) -import { isWebSearchEnabledForProvider, setCopilotInfo, copilotInfo } from './aiStore' +import { + isWebSearchEnabledForProvider, + setCopilotInfo, + copilotInfo, + getWorkspaceCopilotInfo, + invalidateWorkspaceCopilotInfo, + loadWorkspaceCopilotInfo +} from './aiStore' + +const withModel = (model: string) => ({ + providers: { anthropic: { resource_path: 'u/admin/anthropic', models: [model] } } +}) +const modelOf = (info: { aiModels: { model: string }[] } | undefined) => info?.aiModels[0]?.model + +// A config saved, or one every workspace inherits, while a session's load is in flight: the load +// read the config from before the save and must not keep it. +describe.each([ + ['its workspace', () => invalidateWorkspaceCopilotInfo('ws')], + ['every workspace', () => invalidateWorkspaceCopilotInfo()] +])('loadWorkspaceCopilotInfo invalidated for %s mid-fetch', (_, invalidate) => { + it('refetches instead of keeping the outdated config', async () => { + requests = [] + invalidateWorkspaceCopilotInfo() + const load = loadWorkspaceCopilotInfo('ws') + invalidate() + requests[0].resolve(withModel('before-save')) + await vi.waitFor(() => expect(requests).toHaveLength(2)) + requests[1].resolve(withModel('saved')) + + expect(modelOf(await load)).toBe('saved') + expect(modelOf(getWorkspaceCopilotInfo('ws'))).toBe('saved') + }) +}) describe('setCopilotInfo legacy /thinking migration', () => { it('strips the /thinking suffix from configured model slots', () => { diff --git a/frontend/src/lib/aiStore.ts b/frontend/src/lib/aiStore.ts index ac9320f8a3..246fd0a1f0 100644 --- a/frontend/src/lib/aiStore.ts +++ b/frontend/src/lib/aiStore.ts @@ -3,13 +3,15 @@ // import aiStore back, and such a cycle crashes the app once the bundler splits it across // chunks (docs/frontend-import-cycles.md; the build fails on the chunk cycle, not on this). import { writable, get } from 'svelte/store' +import { SvelteMap } from 'svelte/reactivity' import { type AIProviderModel, type AIProvider, type AIConfig, type AIWebSearchConfig, type FreeTierInfo, - type ModelPriceOverride + type ModelPriceOverride, + WorkspaceService } from './gen' import { aiUserDisabled, @@ -40,7 +42,7 @@ export const copilotSessionModel = writable( : undefined ) -export const copilotInfo = writable<{ +export type CopilotInfo = { enabled: boolean // The workspace hid the assistant (`ai_config.copilot_disabled`). `enabled` is then false // whatever the providers say, and the AI entry points that nudge "configure AI" when @@ -64,7 +66,9 @@ export const copilotInfo = writable<{ // Windmill's free tier. `exhausted` means the grant is spent: there is no model, but // that is a different state from "never configured" and the UI must say so. freeTier?: FreeTierInfo -}>({ +} + +export const copilotInfo = writable({ enabled: false, workspaceDisabled: false, codeCompletionModel: undefined, @@ -78,13 +82,27 @@ export const copilotInfo = writable<{ webSearchEnabledProviders: {} }) +// Each workspace's AI config, read by AI sessions, which send with their own workspace's config +// rather than the tab's. Only `loadWorkspaceCopilotInfo` fills it: a save invalidates instead. +const workspaceCopilotInfos = new SvelteMap() +const workspaceLoads = new Map>() +// Bumped by invalidation: a load started before it refetches instead of keeping its answer. +let allWorkspacesVersion = 0 +const workspaceVersions = new Map() +const versionOf = (workspace: string) => + `${allWorkspacesVersion}:${workspaceVersions.get(workspace) ?? 0}` + // Apply the per-user opt-out live: toggling it flips `enabled` without re-fetching. // Only enable when providers exist (aiModels is populated whenever the config has any). aiUserDisabled.subscribe((disabled) => { - copilotInfo.update((info) => ({ + const withOptOut = (info: CopilotInfo): CopilotInfo => ({ ...info, enabled: info.aiModels.length > 0 && !disabled && !info.workspaceDisabled - })) + }) + copilotInfo.update(withOptOut) + for (const [workspace, info] of workspaceCopilotInfos) { + workspaceCopilotInfos.set(workspace, withOptOut(info)) + } }) /** Strip the deprecated /thinking suffix from a configured model slot, if present. */ @@ -110,7 +128,7 @@ function dedupeModels(models: AIProviderModel[]): AIProviderModel[] { // workspace's provider/model while the scoped load is still in flight. export const copilotWorkspace = writable(undefined) -export function setCopilotInfo(aiConfig: AIConfig) { +function buildCopilotInfo(aiConfig: AIConfig, userDisabled: boolean): CopilotInfo { if (Object.keys(aiConfig.providers ?? {}).length > 0) { const aiModels = dedupeModels( Object.entries(aiConfig.providers ?? {}).flatMap(([provider, providerConfig]) => @@ -128,20 +146,10 @@ export function setCopilotInfo(aiConfig: AIConfig) { ]) ) as Partial> - copilotSessionModel.update((model) => { - if ( - model && - !aiModels.some((m) => m.model === model.model && m.provider === model.provider) - ) { - return undefined - } - return model - }) - const workspaceDisabled = aiConfig.copilot_disabled === true - copilotInfo.set({ + return { // Providers are configured; only the workspace or per-user opt-outs can gate it off. - enabled: !workspaceDisabled && !get(aiUserDisabled), + enabled: !workspaceDisabled && !userDisabled, workspaceDisabled, // Strip the deprecated /thinking suffix from the configured model slots too, // otherwise a workspace whose default still carries it sends an invalid model id. @@ -156,11 +164,9 @@ export function setCopilotInfo(aiConfig: AIConfig) { webSearchBackend: aiConfig.web_search?.provider, modelPricing: aiConfig.model_pricing ?? {}, freeTier: aiConfig.free_tier - }) + } } else { - copilotSessionModel.set(undefined) - - copilotInfo.set({ + return { enabled: false, workspaceDisabled: aiConfig.copilot_disabled === true, codeCompletionModel: undefined, @@ -175,20 +181,26 @@ export function setCopilotInfo(aiConfig: AIConfig) { // An exhausted free grant lands here — no providers, but the reason AI is off // is "you used it up", not "you never set it up". freeTier: aiConfig.free_tier - }) + } } } -export function isWebSearchEnabledForProvider(provider: AIProvider | undefined): boolean { +export function isWebSearchEnabledForProvider( + provider: AIProvider | undefined, + info: CopilotInfo = get(copilotInfo) +): boolean { if (!provider) { return false } - return get(copilotInfo).webSearchEnabledProviders?.[provider] ?? true + return info.webSearchEnabledProviders?.[provider] ?? true } export function getCurrentModel(): ReasoningProviderModel { - const model = - get(copilotSessionModel) ?? get(copilotInfo).defaultModel ?? get(copilotInfo).aiModels[0] + return resolveModelFor(get(copilotInfo)) +} + +export function resolveModelFor(info: CopilotInfo): ReasoningProviderModel { + const model = tryResolveModelFor(info) if (!model) { throw new Error('No model selected') } @@ -196,7 +208,10 @@ export function getCurrentModel(): ReasoningProviderModel { } export function getMetadataModel(): AIProviderModel { - const info = get(copilotInfo) + return resolveMetadataModelFor(get(copilotInfo)) +} + +export function resolveMetadataModelFor(info: CopilotInfo): AIProviderModel { const model = info.metadataModel ?? info.defaultModel ?? info.aiModels[0] if (!model) { throw new Error('No model selected') @@ -205,9 +220,85 @@ export function getMetadataModel(): AIProviderModel { } export function tryGetCurrentModel(): ReasoningProviderModel | undefined { - return get(copilotSessionModel) ?? get(copilotInfo).defaultModel ?? get(copilotInfo).aiModels[0] + return tryResolveModelFor(get(copilotInfo)) } +/** `workspace`'s AI config if loaded; otherwise starts loading it and returns undefined. */ +export function getWorkspaceCopilotInfo(workspace: string): CopilotInfo | undefined { + const info = workspaceCopilotInfos.get(workspace) + if (!info) { + loadWorkspaceCopilotInfo(workspace).catch((e) => + console.error(`Failed to load the AI config of workspace ${workspace}`, e) + ) + } + return info +} + +/** `workspace`'s AI config for a request made on its behalf, without touching the tab's. + * Fetched once and kept until invalidated. `fresh` refetches a config that may have changed + * elsewhere: unlike after a save, the one held stays readable until the new one lands, so what + * shows it does not blank meanwhile, and a load already in flight is as fresh and is shared. */ +export function loadWorkspaceCopilotInfo(workspace: string, fresh = false): Promise { + const known = workspaceCopilotInfos.get(workspace) + if (known && !fresh) return Promise.resolve(known) + let load = workspaceLoads.get(workspace) + if (!load) { + const version = versionOf(workspace) + load = WorkspaceService.getCopilotInfo({ workspace }) + // Settled before the retry below, which would otherwise be handed this same load. + .finally(() => workspaceLoads.delete(workspace)) + .then((config) => { + if (version !== versionOf(workspace)) return loadWorkspaceCopilotInfo(workspace) + const info = buildCopilotInfo(config, get(aiUserDisabled)) + workspaceCopilotInfos.set(workspace, info) + return info + }) + workspaceLoads.set(workspace, load) + } + return load +} + +/** After a save of `workspace`'s AI config, or with none of a config every workspace inherits + * (the instance's): the next request refetches. */ +export function invalidateWorkspaceCopilotInfo(workspace?: string) { + if (workspace) { + workspaceVersions.set(workspace, (workspaceVersions.get(workspace) ?? 0) + 1) + workspaceCopilotInfos.delete(workspace) + } else { + allWorkspacesVersion++ + workspaceCopilotInfos.clear() + } +} + +export function setCopilotInfo(aiConfig: AIConfig) { + applyCopilotInfo(buildCopilotInfo(aiConfig, get(aiUserDisabled))) +} + +export function applyCopilotInfo(info: CopilotInfo) { + copilotSessionModel.update((model) => (model && isOffered(info, model) ? model : undefined)) + copilotInfo.set(info) +} + +function isOffered(info: CopilotInfo, model: AIProviderModel): boolean { + return info.aiModels.some((m) => m.model === model.model && m.provider === model.provider) +} + +/** The model a request for `info`'s workspace sends: the shared pick when that workspace offers + * it, else its default, else its first model. Never one the workspace does not offer. */ +export function tryResolveModelFor(info: CopilotInfo): ReasoningProviderModel | undefined { + const picked = get(copilotSessionModel) + if (picked && isOffered(info, picked)) return picked + return info.defaultModel ?? info.aiModels[0] +} + +/** A workspace a request is made for and that workspace's AI config, kept together so a request + * never pairs one workspace's credentials with another's settings. */ +export type AiTarget = { workspace: string; info: CopilotInfo } + +/** What an AI session sends with, for code that otherwise reads the tab's model and config. + * Both are undefined while the session's config is loading. */ +export type SessionAi = { model: ReasoningProviderModel | undefined; info: CopilotInfo | undefined } + export function getUserCustomPrompts(): Record { const stored = getLocalSetting(USER_CUSTOM_PROMPTS_KEY) if (stored) { @@ -242,8 +333,11 @@ export function getCombinedCustomPrompt(mode: string): string | undefined { // Like getCombinedCustomPrompt but keeps the workspace and user slices separate so the // Global system prompt can label them distinctly — only the user slice is editable by the // update_user_instructions tool. -export function getCustomPromptParts(mode: string): { workspace?: string; user?: string } { - const workspace = get(copilotInfo).customPrompts?.[mode]?.trim() || undefined +export function getCustomPromptParts( + mode: string, + customPrompts: Record | undefined = get(copilotInfo).customPrompts +): { workspace?: string; user?: string } { + const workspace = customPrompts?.[mode]?.trim() || undefined const user = getUserCustomPrompts()[mode]?.trim() || undefined return { workspace, user } } diff --git a/frontend/src/lib/components/Editor.svelte b/frontend/src/lib/components/Editor.svelte index b103e7ddf8..7e9b192578 100644 --- a/frontend/src/lib/components/Editor.svelte +++ b/frontend/src/lib/components/Editor.svelte @@ -276,6 +276,10 @@ // rather than via getAiChatManager(), which collapses "no session" into the // singleton — the distinction is what tells ⌘L whether a pane exists to open. const sessionScopedChatManager = getContext('aiChatManager') + // Autocomplete in a session's editor follows the session's workspace, not the tab's. + const completionInfo = $derived( + sessionScopedChatManager ? sessionScopedChatManager.aiInfo : $copilotInfo + ) // let graphqlService: MonacoGraphQLAPI | undefined = undefined let fetchedDbSchema: DBSchema | undefined = $state(undefined) @@ -1008,7 +1012,11 @@ if (autocompletor) { autocompletor.dispose() } - autocompletor = new Autocompletor(editor, scriptLang, { workflowAsCode }) + const session = sessionScopedChatManager + autocompletor = new Autocompletor(editor, scriptLang, { + workflowAsCode, + sessionTarget: session ? () => session.loadedSessionAiTarget : undefined + }) } const outputChannel = { @@ -2082,9 +2090,9 @@ }) $effect(() => { const currentWorkflowAsCode = workflowAsCode - $copilotInfo.enabled && + completionInfo?.enabled && $codeCompletionSessionEnabled && - Autocompletor.isProviderModelSupported($copilotInfo.codeCompletionModel) && + Autocompletor.isProviderModelSupported(completionInfo.codeCompletionModel) && initialized && editor && scriptLang && @@ -2094,7 +2102,7 @@ $copilotInfo.enabled && initialized && editor && untrack(() => editor && addChatHandler(editor)) }) $effect(() => { - ;(!$codeCompletionSessionEnabled || !$copilotInfo.enabled) && autocompletor?.dispose() + ;(!$codeCompletionSessionEnabled || !completionInfo?.enabled) && autocompletor?.dispose() }) $effect(() => { if (yContent && awareness && model && editor) { diff --git a/frontend/src/lib/components/copilot/autocomplete/Autocompletor.ts b/frontend/src/lib/components/copilot/autocomplete/Autocompletor.ts index 5c9519ba1e..ce1f9bc3f2 100644 --- a/frontend/src/lib/components/copilot/autocomplete/Autocompletor.ts +++ b/frontend/src/lib/components/copilot/autocomplete/Autocompletor.ts @@ -9,7 +9,7 @@ import { setGlobalCSS } from '../shared' import { supportsAutocomplete } from '../utils' import { get } from 'svelte/store' import type { MonacoLanguageClient } from 'monaco-languageclient' -import { copilotInfo } from '$lib/aiStore' +import { copilotInfo, type AiTarget } from '$lib/aiStore' import { getTypeScriptWorker } from '@codingame/monaco-vscode-standalone-typescript-language-features' // max ratio of completions to context window @@ -72,15 +72,21 @@ export class Autocompletor { #completionDisposable: IDisposable #cursorDisposable: IDisposable #lastCompletions: string[] = [] - #contextWindow: number = 0 + #sessionTarget: (() => AiTarget | undefined) | undefined #shouldShowDeletionCue = false #languageClient: MonacoLanguageClient | undefined = undefined constructor( editor: meditor.IStandaloneCodeEditor, scriptLang: ScriptLang | 'bunnative' | 'jsx' | 'tsx' | 'json', - options: { workflowAsCode?: boolean } = {} + options: { + workflowAsCode?: boolean + /** Set in an AI session's editor: completions go to the session's workspace with its + * AI config. Undefined while that config is loading. */ + sessionTarget?: () => AiTarget | undefined + } = {} ) { + this.#sessionTarget = options.sessionTarget setGlobalCSS( 'ai-chat-autocomplete', ` @@ -94,15 +100,6 @@ export class Autocompletor { const deletionsCues = editor.createDecorationsCollection() - const { codeCompletionModel: completionModel, contextWindowPerModel } = get(copilotInfo) - this.#contextWindow = completionModel - ? getEffectiveModelContextWindow( - completionModel.provider, - completionModel.model, - contextWindowPerModel - ) - : ASSUMED_CONTEXT_WINDOW - this.#completionDisposable = languages.registerInlineCompletionsProvider( { pattern: '**' }, { @@ -520,9 +517,22 @@ export class Autocompletor { const markers = meditor.getModelMarkers({ resource: model.uri }) const markersAtCursor = this.#markersAtCursor(position, markers) + const target = this.#sessionTarget?.() + // A session's editor completes only once its own workspace's config is known. + if (this.#sessionTarget && !target) { + return + } + const { codeCompletionModel, contextWindowPerModel } = target?.info ?? get(copilotInfo) + const contextWindow = codeCompletionModel + ? getEffectiveModelContextWindow( + codeCompletionModel.provider, + codeCompletionModel.model, + contextWindowPerModel + ) + : ASSUMED_CONTEXT_WINDOW const librariesCompletions: string = completions .join('\n') - .slice(0, Math.floor(this.#contextWindow * COMPLETIONS_MAX_RATIO)) + .slice(0, Math.floor(contextWindow * COMPLETIONS_MAX_RATIO)) const completion = await autocompleteRequest( { @@ -533,7 +543,8 @@ export class Autocompletor { libraries: librariesCompletions, workflowAsCode: this.#workflowAsCode }, - this.#abortController + this.#abortController, + target ) if (!completion) { diff --git a/frontend/src/lib/components/copilot/autocomplete/request.ts b/frontend/src/lib/components/copilot/autocomplete/request.ts index cc3987ebe5..651b212ec2 100644 --- a/frontend/src/lib/components/copilot/autocomplete/request.ts +++ b/frontend/src/lib/components/copilot/autocomplete/request.ts @@ -4,7 +4,7 @@ import { getLangContext } from '../chat/script/core' import { type ScriptLang } from '$lib/gen/types.gen' import type { editor } from 'monaco-editor' import { getCommentSymbol } from '../utils' -import { copilotInfo } from '$lib/aiStore' +import { copilotInfo, type AiTarget } from '$lib/aiStore' function comment(commentSymbol: string, text: string) { return text @@ -22,9 +22,10 @@ export async function autocompleteRequest( libraries: string workflowAsCode?: boolean }, - abortController: AbortController + abortController: AbortController, + target?: AiTarget ) { - const providerModel = get(copilotInfo).codeCompletionModel + const providerModel = (target?.info ?? get(copilotInfo)).codeCompletionModel if (!providerModel) { throw new Error('No code completion model selected') @@ -56,7 +57,8 @@ export async function autocompleteRequest( context.prefix, context.suffix, providerModel, - abortController + abortController, + target?.workspace ) return completion diff --git a/frontend/src/lib/components/copilot/chat/AIChatInput.svelte b/frontend/src/lib/components/copilot/chat/AIChatInput.svelte index f3b49d680b..f3dd73bc46 100644 --- a/frontend/src/lib/components/copilot/chat/AIChatInput.svelte +++ b/frontend/src/lib/components/copilot/chat/AIChatInput.svelte @@ -44,7 +44,6 @@ type AttachedImage } from './imageUtils' import { modelSupportsVision } from '../modelConfig' - import { tryGetCurrentModel } from '$lib/aiStore' import { createLongHash } from '$lib/editorLangUtils' import { fileToAttachedTextFile, @@ -334,11 +333,9 @@ if (imageFiles.length === 0) return // The vision check is about the model this composer's own turn will hit, so it // only applies to a host that picks that model. Elsewhere the model is chosen - // in the flow and tryGetCurrentModel would answer for the wrong one. + // in the flow and the host's model would answer for the wrong one. if (chatHost.supportsModelSettings) { - // tryGetCurrentModel returns undefined instead of throwing: this runs from a - // drop/paste handler that can't surface a rejection. - const model = tryGetCurrentModel() + const model = chatHost.currentModel // Only known text-only models fail this, so attaching would certainly 400 the // next turn — refuse rather than warn and send it anyway. if (model && !modelSupportsVision(model.provider, model.model)) { diff --git a/frontend/src/lib/components/copilot/chat/AIChatManager.svelte.ts b/frontend/src/lib/components/copilot/chat/AIChatManager.svelte.ts index 6c8142fa72..9910bb1e87 100644 --- a/frontend/src/lib/components/copilot/chat/AIChatManager.svelte.ts +++ b/frontend/src/lib/components/copilot/chat/AIChatManager.svelte.ts @@ -98,7 +98,17 @@ import { fromStore, get } from 'svelte/store' import type { ProviderToolSummary } from './agentContext' import { BROWSER } from 'esm-env' import { roleCapabilities, workspaceStore, type DBSchemas } from '$lib/stores' -import { copilotInfo, copilotSessionModel } from '$lib/aiStore' +import { + copilotInfo, + copilotSessionModel, + getWorkspaceCopilotInfo, + loadWorkspaceCopilotInfo, + resolveModelFor, + tryResolveModelFor, + type AiTarget, + type CopilotInfo +} from '$lib/aiStore' +import type { ReasoningProviderModel } from '../reasoningRegistry' import { copilotWorkspaceRequested, loadCopilot } from '$lib/components/copilot/loadCopilot' import { askTools, prepareAskSystemMessage, prepareAskUserMessage } from './ask/core' import { readDocsPageTool, searchDocsTool } from './docs/core' @@ -707,8 +717,11 @@ export class AIChatManager implements ChatViewHost { if (this.reasoningSummaryUnavailableFor.length === 0) { return undefined } - const provider = getCurrentModel().provider - if (!this.reasoningSummaryUnavailableFor.includes(this.reasoningSummaryKey(provider))) { + const provider = this.currentModel?.provider + if ( + !provider || + !this.reasoningSummaryUnavailableFor.includes(this.reasoningSummaryKey(provider)) + ) { return undefined } return `Thinking (hidden, ${providerDisplayName(provider)} org not verified)` @@ -860,16 +873,12 @@ export class AIChatManager implements ChatViewHost { #copilotInfo = fromStore(copilotInfo) #sessionModel = fromStore(copilotSessionModel) /** Capabilities the provider runs on its own servers, attached by the adapter at send - * time. Never part of `tools`: the request path would try to dispatch them. The two - * `void` reads only subscribe: the model helpers read through `get()`, which tracks - * nothing, so a change of model or of the workspace's AI config would otherwise not - * re-derive. A workspace switch re-derives through the AI config it loads. */ + * time. Never part of `tools`: the request path would try to dispatch them. */ get providerTools(): ProviderToolSummary[] { - void this.#copilotInfo.current - void this.#sessionModel.current - const model = tryGetCurrentModel() - if (!model || (model.provider === 'openai' && this.skipResponsesApi)) return [] - const enabled = isWebSearchEnabledForProvider(model.provider) + const model = this.currentModel + const info = this.aiInfo + if (!model || !info || (model.provider === 'openai' && this.skipResponsesApi)) return [] + const enabled = isWebSearchEnabledForProvider(model.provider, info) return sendsWebSearch(this.operatingWorkspace ?? '', model, enabled) ? [ { @@ -1061,6 +1070,37 @@ export class AIChatManager implements ChatViewHost { return this.workspaceResolver?.() ?? get(workspaceStore) } + /** The AI config this chat sends with. A session's is its own workspace's, never the tab's, + * which belongs to whatever workspace is on screen; other chats use the tab's. Undefined while + * a session's is still loading, and reading it starts the load. */ + get aiInfo(): CopilotInfo | undefined { + if (!this.isSessionChat) return this.#copilotInfo.current + const workspace = this.operatingWorkspace + return workspace ? getWorkspaceCopilotInfo(workspace) : undefined + } + + /** The session's workspace and loaded config, for code that cannot wait for the load. */ + get loadedSessionAiTarget(): AiTarget | undefined { + const info = this.isSessionChat ? this.aiInfo : undefined + const workspace = this.operatingWorkspace + return info && workspace ? { workspace, info } : undefined + } + + get currentModel(): ReasoningProviderModel | undefined { + // Subscribes only: the model helpers read through `get()`, which tracks nothing. + void this.#sessionModel.current + const info = this.aiInfo + return info && tryResolveModelFor(info) + } + + /** Resolved at each request, after any fork commit. Undefined for other chats, whose requests + * use the tab's. */ + sessionAiTarget = async (): Promise => { + if (!this.isSessionChat) return undefined + const workspace = this.operatingWorkspace ?? '' + return { workspace, info: await loadWorkspaceCopilotInfo(workspace) } + } + // Fired whenever the active chat id changes away from the one the consumer // knows (a "/clear" rotation or a history switch). Session runtimes wire this // to keep the session record's chatId aligned — the compare-page handoff @@ -1792,6 +1832,7 @@ export class AIChatManager implements ChatViewHost { ): Promise<'ok' | 'empty' | 'aborted' | 'error'> => { this.compacting = true try { + const target = await this.sessionAiTarget() // Cap the summarizer's output at the budget already reserved for the // summary: without a cap the model's default max_tokens applies. const raw = await getStreamedCompletionText( @@ -1802,7 +1843,7 @@ export class AIChatManager implements ChatViewHost { { role: 'user', content: getCompactionSummaryPrompt() } ], abortController, - { maxTokensCap: SUMMARY_OUTPUT_RESERVE_TOKENS } + { maxTokensCap: SUMMARY_OUTPUT_RESERVE_TOKENS, ...target } ) const formatted = formatCompactSummary(raw) if (!formatted) { @@ -1991,12 +2032,9 @@ export class AIChatManager implements ChatViewHost { model: ChatLoopConfig['modelProvider'] | undefined, inFlightTokens = 0 ): Promise => { + const info = (await this.sessionAiTarget())?.info ?? get(copilotInfo) const contextWindow = model - ? getEffectiveModelContextWindow( - model.provider, - model.model, - get(copilotInfo).contextWindowPerModel - ) + ? getEffectiveModelContextWindow(model.provider, model.model, info.contextWindowPerModel) : undefined if (contextWindow === undefined || projectedTokens < contextWindow * COMPACTION_TRIGGER_RATIO) { return false @@ -2893,6 +2931,7 @@ export class AIChatManager implements ChatViewHost { ? { sessionId: this.sessionId, operatingWorkspace: this.operatingWorkspace, + sessionAi: () => ({ model: this.currentModel, info: this.aiInfo }), artifacts: this.artifacts, access: this.#toolAccess, getChatId: () => this.historyManager.getCurrentChatId(), @@ -2923,7 +2962,10 @@ export class AIChatManager implements ChatViewHost { ensurePipeline: this.#ensurePipeline } : baseHelpers - this.systemMessage = assembleGlobalSystemMessage(getCustomPromptParts(AIMode.GLOBAL), opts) + this.systemMessage = assembleGlobalSystemMessage( + getCustomPromptParts(AIMode.GLOBAL, this.aiInfo?.customPrompts ?? {}), + opts + ) this.syncArtifactsSession() } @@ -3103,7 +3145,7 @@ export class AIChatManager implements ChatViewHost { return } this.systemMessage = assembleGlobalSystemMessage( - getCustomPromptParts(AIMode.GLOBAL), + getCustomPromptParts(AIMode.GLOBAL, this.aiInfo?.customPrompts ?? {}), this.globalAssemblyOpts() ) } @@ -3435,7 +3477,8 @@ export class AIChatManager implements ChatViewHost { } private notifyReasoningSummaryUnavailable = () => { - const provider = getCurrentModel().provider + const provider = this.currentModel?.provider + if (!provider) return const key = this.reasoningSummaryKey(provider) if (!this.reasoningSummaryUnavailableFor.includes(key)) { this.reasoningSummaryUnavailableFor = [...this.reasoningSummaryUnavailableFor, key] @@ -3482,6 +3525,8 @@ export class AIChatManager implements ChatViewHost { // while a response streams would bill it to the workspace the user landed // on rather than the one whose credentials and proxy served it. const usageWorkspace = this.operatingWorkspace + const target = await this.sessionAiTarget() + const info = target?.info const result = await runChatLoop({ messages, addedMessages, @@ -3502,7 +3547,7 @@ export class AIChatManager implements ChatViewHost { get tools() { return withWebSearchFallback( self.tools, - !!get(copilotInfo).webSearchBackend, + !!(info ?? get(copilotInfo)).webSearchBackend, self.providerTools.length > 0 ) }, @@ -3512,10 +3557,10 @@ export class AIChatManager implements ChatViewHost { abortController, callbacks, get modelProvider() { - return getCurrentModel() + return info ? resolveModelFor(info) : getCurrentModel() }, get webSearch() { - return isWebSearchEnabledForProvider(getCurrentModel().provider) + return isWebSearchEnabledForProvider(this.modelProvider.provider, info) }, // Build the proxy clients against the operating workspace, not the global // singleton: a session deliberately leaves workspaceStore untouched, so the @@ -3523,13 +3568,14 @@ export class AIChatManager implements ChatViewHost { // request through the navigation workspace's /ai/proxy instead of the // session's — sending it to the wrong workspace's AI credentials. get clients() { - const ws = self.operatingWorkspace ?? '' + const ws = target?.workspace ?? self.operatingWorkspace ?? '' return { openai: workspaceAIClients.createOpenaiClient(ws), anthropic: workspaceAIClients.createAnthropicClient(ws) } }, - workspace: this.operatingWorkspace ?? '', + workspace: target?.workspace ?? this.operatingWorkspace ?? '', + info, skipResponsesApi: this.skipResponsesApi, onSkipResponsesApi: () => { this.skipResponsesApi = true @@ -3973,7 +4019,7 @@ export class AIChatManager implements ChatViewHost { // only edit/retry of a pre-id transcript can arrive without one, and the // hash is deterministic, so hydration reproduces the original id. const files = withAttachedTextFileIds(options.files ?? []) - const sendModel = tryGetCurrentModel() + const sendModel = this.currentModel const modelIsBlind = !!sendModel && !modelSupportsVision(sendModel.provider, sendModel.model) if (requestedImages.length > 0 && modelIsBlind) { // An image-only message has nothing left once the images are dropped — @@ -4111,8 +4157,13 @@ export class AIChatManager implements ChatViewHost { this.refreshSessionAccess(this.operatingWorkspace ?? ''), resolveResourceTypeSync(this.operatingWorkspace ?? '', this.messages).then((sync) => { this.resourceTypeSync = sync - }) + }), + // Its workspace prompt goes into the system prompt rebuilt below. + this.sessionAiTarget().catch(() => {}) ]) + if (this.isSessionChat) { + this.rebuildGlobalSystemMessage() + } } // Stop/Escape during the beforeSend pre-flight aborted this send before any // request went out. Mirror the main "cancelled before usable output" recovery: @@ -4268,7 +4319,7 @@ export class AIChatManager implements ChatViewHost { // loading + abortController were set optimistically before beforeSend, above. this.#automaticScroll = true - const model = tryGetCurrentModel() + const model = this.currentModel if (model) { const chatId = this.historyManager.getCurrentChatId() logFeatureUsage('ai_chat', 'message', { @@ -4792,15 +4843,11 @@ export class AIChatManager implements ChatViewHost { if ( requestCarriedImages && messagesHaveImageParts(this.messages) && - isImageRejection(err, [ - sendModel?.model, - tryGetCurrentModel()?.model, - failingModel?.model - ]) + isImageRejection(err, [sendModel?.model, this.currentModel?.model, failingModel?.model]) ) { this.messages = stripImagePartsFromMessages(this.messages) sendUserToast( - `${tryGetCurrentModel()?.model ?? 'The model'} could not read the attached image(s), so they were removed from the conversation. Your message was kept.`, + `${this.currentModel?.model ?? 'The model'} could not read the attached image(s), so they were removed from the conversation. Your message was kept.`, true ) } diff --git a/frontend/src/lib/components/copilot/chat/AIChatManager.test.ts b/frontend/src/lib/components/copilot/chat/AIChatManager.test.ts index 9f6179a538..ff64eb07fc 100644 --- a/frontend/src/lib/components/copilot/chat/AIChatManager.test.ts +++ b/frontend/src/lib/components/copilot/chat/AIChatManager.test.ts @@ -141,6 +141,11 @@ vi.mock('$lib/aiStore', () => ({ copilotSessionModel: writable(undefined), getCurrentModel: mocks.getCurrentModel, tryGetCurrentModel: mocks.tryGetCurrentModel, + // A session reads its own workspace's config; here every workspace has the same one. + getWorkspaceCopilotInfo: () => ({ enabled: true, workspaceDisabled: false, aiModels: [] }), + loadWorkspaceCopilotInfo: async () => ({ enabled: true, workspaceDisabled: false, aiModels: [] }), + resolveModelFor: mocks.getCurrentModel, + tryResolveModelFor: mocks.tryGetCurrentModel, getCombinedCustomPrompt: () => '', getCustomPromptParts: () => ({}), getUserCustomPrompts: () => ({}), diff --git a/frontend/src/lib/components/copilot/chat/AIChatModelSettings.svelte b/frontend/src/lib/components/copilot/chat/AIChatModelSettings.svelte index 5f4b4a5040..fb65460f2c 100644 --- a/frontend/src/lib/components/copilot/chat/AIChatModelSettings.svelte +++ b/frontend/src/lib/components/copilot/chat/AIChatModelSettings.svelte @@ -17,6 +17,7 @@ copilotInfo, copilotSessionModel, getUserCustomPrompts, + invalidateWorkspaceCopilotInfo, setCopilotInfo, setUserCustomPrompts } from '$lib/aiStore' @@ -192,6 +193,7 @@ requestBody: { ...config, custom_prompts } }) setCopilotInfo(response.effective_ai_config) + invalidateWorkspaceCopilotInfo(workspace) initialPrompt = value customPrompts = { [activeMode]: value } sendUserToast('Workspace AI prompt saved') diff --git a/frontend/src/lib/components/copilot/chat/AssistantInstructionsSection.svelte b/frontend/src/lib/components/copilot/chat/AssistantInstructionsSection.svelte index 547940bde9..ebea3dbb6a 100644 --- a/frontend/src/lib/components/copilot/chat/AssistantInstructionsSection.svelte +++ b/frontend/src/lib/components/copilot/chat/AssistantInstructionsSection.svelte @@ -23,6 +23,7 @@ its block on a detail page laid over the list, like a skill in the Skills sectio import { copilotInfo, getUserCustomPrompts, + invalidateWorkspaceCopilotInfo, setCopilotInfo, setUserCustomPrompts } from '$lib/aiStore' @@ -596,6 +597,7 @@ its block on a detail page laid over the list, like a skill in the Skills sectio requestBody: { ...config, custom_prompts } }) setCopilotInfo(response.effective_ai_config) + invalidateWorkspaceCopilotInfo(target) workspaceSaved = value workspaceDraft = value sendUserToast('Workspace instructions were saved') diff --git a/frontend/src/lib/components/copilot/chat/anthropic.ts b/frontend/src/lib/components/copilot/chat/anthropic.ts index ce11b0dade..937fe47ed4 100644 --- a/frontend/src/lib/components/copilot/chat/anthropic.ts +++ b/frontend/src/lib/components/copilot/chat/anthropic.ts @@ -14,6 +14,7 @@ import type { } from '@anthropic-ai/sdk/resources' import type { MessageStream } from '@anthropic-ai/sdk/lib/MessageStream' import type { AIProviderModel } from '$lib/gen' +import type { CopilotInfo } from '$lib/aiStore' import { getProviderAndCompletionConfig, workspaceAIClients } from '../lib' import { applyReasoningToConfig } from '../reasoningRegistry' import { @@ -137,13 +138,15 @@ export async function getAnthropicCompletion( anthropicClient?: Anthropic webSearch?: boolean reasoningEffort?: string + info?: CopilotInfo } ): Promise { const { provider, config } = getProviderAndCompletionConfig({ messages, stream: true, forceModelProvider: options?.forceModelProvider, - reasoningEffort: options?.reasoningEffort + reasoningEffort: options?.reasoningEffort, + info: options?.info }) const { system, messages: anthropicMessages } = convertOpenAIToAnthropicMessages(messages) let anthropicTools = convertOpenAIToolsToAnthropic(tools) diff --git a/frontend/src/lib/components/copilot/chat/chatLoop.ts b/frontend/src/lib/components/copilot/chat/chatLoop.ts index 14de05cd74..e6b9e5c060 100644 --- a/frontend/src/lib/components/copilot/chat/chatLoop.ts +++ b/frontend/src/lib/components/copilot/chat/chatLoop.ts @@ -14,6 +14,7 @@ import { type ReasoningProviderModel } from '../reasoningRegistry' import { getAnthropicCompletion, parseAnthropicCompletion } from './anthropic' +import type { CopilotInfo } from '$lib/aiStore' import { modelSupportsVision, usesAnthropicMessagesApi } from '../modelConfig' import { boundImagePartBytes, stripImagePartsFromMessages } from './imageUtils' import { @@ -49,6 +50,8 @@ export interface ChatLoopConfig { modelProvider: ReasoningProviderModel clients: ChatClients workspace: string + /** `workspace`'s AI config when it is not the tab's (an AI session). */ + info?: CopilotInfo /** * Enable provider-native web search. Defaults to true for compatible providers * and is re-read each iteration so model/provider changes take effect. Explicit @@ -371,6 +374,7 @@ export async function runChatLoop(config: ChatLoopConfig): Promise[] = [ // A known text-only model would reject the follow-up image message and fail // the turn, so refuse before capturing rather than buffer an image it can // never read. The model is re-read here because it can change between turns. - const model = tryGetCurrentModel() + const sessionAi = (ctx.helpers as GlobalToolHelpers | undefined)?.sessionAi + const model = sessionAi ? sessionAi().model : tryGetCurrentModel() if (model && !modelSupportsVision(model.provider, model.model)) { const cannotSee = `${model.model} cannot read images, so a screenshot would be discarded. Ask the user to describe what looks wrong, or to switch to a model that supports images.` ctx.toolCallbacks.setToolStatus(ctx.toolId, { @@ -4694,6 +4695,8 @@ export type GlobalToolHelpers = SessionToolHelpers & { // (possibly forked) workspace while $workspaceStore stays on the navigation workspace, // so permission gating (open_page) must read this, not the global store. operatingWorkspace?: string + // The session's model and its workspace's AI config, rather than the tab's. + sessionAi?: () => SessionAi // Wired only for session chats (see AIChatManager): the artifact tools are session-gated. artifacts?: SessionArtifactsStore getChatId?: () => string | undefined diff --git a/frontend/src/lib/components/copilot/chat/openai-responses.ts b/frontend/src/lib/components/copilot/chat/openai-responses.ts index 2071ff0d74..eccda3bc1a 100644 --- a/frontend/src/lib/components/copilot/chat/openai-responses.ts +++ b/frontend/src/lib/components/copilot/chat/openai-responses.ts @@ -28,6 +28,7 @@ import { OutputTokenLimitError } from './outputTokenLimit' import { buildAssistantToolCallMessage } from './openaiReasoning' import type { ResponseStream } from 'openai/lib/responses/ResponseStream.mjs' import type { AIProviderModel } from '$lib/gen' +import type { CopilotInfo } from '$lib/aiStore' import { openAIResponsesUsageToChatTokenUsage, type ChatTokenUsage } from './tokenUsage' interface ParsedCompletionResult { @@ -308,6 +309,7 @@ export async function getOpenAIResponsesCompletion( reasoningEffort?: string reasoningSummary?: boolean promptCacheKey?: string + info?: CopilotInfo } ) { const { provider, config } = getProviderAndCompletionConfig({ @@ -315,7 +317,8 @@ export async function getOpenAIResponsesCompletion( stream: true, tools, forceModelProvider: options?.forceModelProvider, - reasoningEffort: options?.reasoningEffort + reasoningEffort: options?.reasoningEffort, + info: options?.info }) const { instructions, input } = convertMessagesToResponsesInput(messages, provider === 'deepseek') const responsesConfig = applyReasoningToConfig( @@ -379,6 +382,7 @@ export async function* getOpenAIResponsesCompletionStream( openaiClient?: OpenAI reasoningEffort?: string maxTokensCap?: number + info?: CopilotInfo } ): AsyncGenerator { const { provider, config } = getProviderAndCompletionConfig({ @@ -387,7 +391,8 @@ export async function* getOpenAIResponsesCompletionStream( tools, forceModelProvider: options?.forceModelProvider, maxTokensCap: options?.maxTokensCap, - reasoningEffort: options?.reasoningEffort + reasoningEffort: options?.reasoningEffort, + info: options?.info }) const { instructions, input } = convertMessagesToResponsesInput(messages) // No prompt cache key here: a rejected key has to be retried without it, and this is @@ -717,13 +722,15 @@ export async function getNonStreamingOpenAIResponsesCompletion( resourceValue?: Record forceModelProvider?: AIProviderModel maxTokensCap?: number + info?: CopilotInfo } ): Promise { const { provider, config } = getProviderAndCompletionConfig({ messages, stream: false, forceModelProvider: options?.forceModelProvider, - maxTokensCap: options?.maxTokensCap + maxTokensCap: options?.maxTokensCap, + info: options?.info }) const { instructions, input } = convertMessagesToResponsesInput(messages) diff --git a/frontend/src/lib/components/copilot/chat/script/core.ts b/frontend/src/lib/components/copilot/chat/script/core.ts index e7053ce141..79f8621e40 100644 --- a/frontend/src/lib/components/copilot/chat/script/core.ts +++ b/frontend/src/lib/components/copilot/chat/script/core.ts @@ -26,7 +26,7 @@ import { createWorkspaceMutationTools } from '../workspaceTools' import { setupTypeAcquisition, type DepsToGet } from '$lib/ata' import { getEffectiveModelContextWindow } from '../../modelConfig' import type { ReviewChangesOpts } from '../monaco-adapter' -import { copilotInfo, getCurrentModel } from '$lib/aiStore' +import { copilotInfo, getCurrentModel, type SessionAi } from '$lib/aiStore' import { get } from 'svelte/store' import { getDbSchemas } from '$lib/components/apps/components/display/dbtable/metadata' import { scriptLangToEditorLang } from '$lib/scripts' @@ -604,7 +604,10 @@ type PackageSearchResult = { } const packagesSearchCache = new Map() -export async function searchExternalIntegrationResources(args: { query: string }): Promise { +export async function searchExternalIntegrationResources( + args: { query: string }, + sessionAi?: SessionAi +): Promise { try { if (packagesSearchCache.has(args.query)) { return JSON.stringify(packagesSearchCache.get(args.query)) @@ -619,11 +622,14 @@ export async function searchExternalIntegrationResources(args: { query: string } (r: PackageSearchQuery) => r.searchScore >= best * NPM_RUNNER_UP_SHARE ) - const model = getCurrentModel() + const model = sessionAi ? sessionAi.model : getCurrentModel() + if (!model) { + throw new Error('No model selected') + } const modelContextWindow = getEffectiveModelContextWindow( model.provider, model.model, - get(copilotInfo).contextWindowPerModel + (sessionAi ? sessionAi.info : get(copilotInfo))?.contextWindowPerModel ) const results: PackageSearchResult[] = await Promise.all( filtered.map(async (r: PackageSearchQuery) => { @@ -684,9 +690,10 @@ export const searchNpmPackagesTool: SessionTool<{}> = { requires: NONE, def: SEARCH_NPM_PACKAGES_TOOL, planModeSafe: true, - fn: async ({ args, toolId, toolCallbacks }) => { + fn: async ({ args, toolId, toolCallbacks, helpers }) => { toolCallbacks.setToolStatus(toolId, { content: 'Searching for relevant packages...' }) - const result = await searchExternalIntegrationResources(args) + const sessionAi = (helpers as { sessionAi?: () => SessionAi } | undefined)?.sessionAi + const result = await searchExternalIntegrationResources(args, sessionAi?.()) toolCallbacks.setToolStatus(toolId, { content: 'Retrieved relevant packages' }) return result } diff --git a/frontend/src/lib/components/copilot/chat/sessionAiConfig.test.ts b/frontend/src/lib/components/copilot/chat/sessionAiConfig.test.ts new file mode 100644 index 0000000000..afe3dc773a --- /dev/null +++ b/frontend/src/lib/components/copilot/chat/sessionAiConfig.test.ts @@ -0,0 +1,145 @@ +import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest' +import type { AIConfig } from '$lib/gen' +import { copilotSessionModel, setCopilotInfo } from '$lib/aiStore' +import { workspaceStore } from '$lib/stores' +import { workspaceAIClients } from '../lib' +import { AIChatManager } from './AIChatManager.svelte' +import { generateSessionSummary } from '$lib/components/sessions/sessionTitle' + +vi.mock('$app/navigation', () => ({ + goto: vi.fn(), + afterNavigate: vi.fn(), + beforeNavigate: vi.fn() +})) +vi.mock('monaco-editor', () => ({ Selection: class Selection {} })) +vi.mock('$lib/utils/featureUsage', () => ({ logFeatureUsage: vi.fn() })) + +// The session works in `staging` while the tab shows `prod`: every request the session +// makes must carry staging's proxy, model and limits. +const CONFIGS: Record = { + staging: { + providers: { mistral: { resource_path: 'u/admin/mistral', models: ['staging-model'] } }, + default_model: { provider: 'mistral', model: 'staging-model' }, + max_tokens_per_model: { 'mistral:staging-model': 3000 }, + context_window_per_model: { 'mistral:staging-model': 30000 } + }, + prod: { + providers: { deepseek: { resource_path: 'u/admin/deepseek', models: ['prod-model'] } }, + default_model: { provider: 'deepseek', model: 'prod-model' } + } +} + +vi.mock('$lib/gen', async (importOriginal) => ({ + ...(await importOriginal()), + WorkspaceService: { + getCopilotInfo: async ({ workspace }: { workspace: string }) => CONFIGS[workspace] + } +})) + +type SentRequest = { url: string; body: any } +let sent: SentRequest[] = [] + +function completionResponse(body: any): Response { + const text = 'Earlier work' + if (!body.stream) { + return Response.json({ + id: 'c', + object: 'chat.completion', + created: 0, + model: body.model, + choices: [ + { + index: 0, + message: { role: 'assistant', content: 'Staging title' }, + finish_reason: 'stop' + } + ] + }) + } + const chunk = (delta: object, finish_reason: string | null) => + `data: ${JSON.stringify({ + id: 'c', + object: 'chat.completion.chunk', + created: 0, + model: body.model, + choices: [{ index: 0, delta, finish_reason }] + })}\n\n` + return new Response(chunk({ content: text }, null) + chunk({}, 'stop') + 'data: [DONE]\n\n', { + headers: { 'content-type': 'text/event-stream' } + }) +} + +function sessionManager(): AIChatManager { + const manager = new AIChatManager() + manager.isSessionChat = true + manager.workspaceResolver = () => 'staging' + manager.messages = [ + { role: 'user', content: 'q1' }, + { role: 'assistant', content: 'a1' }, + { role: 'user', content: 'q2' }, + { role: 'assistant', content: 'a2' }, + { role: 'user', content: 'q3' } + ] + manager.displayMessages = [ + { role: 'user', content: 'q1', index: 0 }, + { role: 'assistant', content: 'a1' }, + { role: 'user', content: 'q2', index: 2 }, + { role: 'assistant', content: 'a2' }, + { role: 'user', content: 'q3', index: 4 } + ] as AIChatManager['displayMessages'] + return manager +} + +function expectSentForStaging(request: SentRequest | undefined) { + expect(request?.url).toContain('/w/staging/ai/proxy/') + expect(request?.body.model).toBe('staging-model') + expect(request?.body.max_tokens).toBe(3000) +} + +beforeEach(() => { + sent = [] + vi.stubGlobal('location', { origin: 'http://localhost' }) + vi.stubGlobal('fetch', async (input: RequestInfo | URL, init?: RequestInit) => { + const url = input instanceof Request ? input.url : String(input) + const body = JSON.parse( + String(init?.body ?? (input instanceof Request ? await input.text() : '{}')) + ) + sent.push({ url, body }) + return completionResponse(body) + }) + // The tab is on prod, with prod's model picked. + workspaceStore.set('prod') + workspaceAIClients.init('prod') + setCopilotInfo(CONFIGS.prod) + copilotSessionModel.set({ provider: 'deepseek', model: 'prod-model' }) +}) + +afterEach(() => { + vi.unstubAllGlobals() +}) + +describe('an AI session in another workspace than the tab', () => { + it('names the session with its own workspace', async () => { + expect(await generateSessionSummary(sessionManager())).toBe('Staging title') + expect(sent).toHaveLength(1) + expectSentForStaging(sent[0]) + }) + + it('summarizes for compaction with its own workspace', async () => { + const manager = sessionManager() + await manager.compactManually() + expect(sent).toHaveLength(1) + expectSentForStaging(sent[0]) + }) + + it("compacts at its own workspace's context window", async () => { + const manager = sessionManager() + manager.abortController = new AbortController() + // 80% of staging's 30k window; far below the window prod's model would get. + const compacted = await (manager as any).compactIfOverTrigger(25000, { + provider: 'mistral', + model: 'staging-model' + }) + expect(compacted).toBe(true) + }) +}) diff --git a/frontend/src/lib/components/copilot/lib.ts b/frontend/src/lib/components/copilot/lib.ts index d2eae56bcd..5975fc3546 100644 --- a/frontend/src/lib/components/copilot/lib.ts +++ b/frontend/src/lib/components/copilot/lib.ts @@ -47,7 +47,15 @@ import { import { convertOpenAIToAnthropicMessages } from './chat/anthropic' import type { Stream } from 'openai/core/streaming.mjs' import { generateRandomString } from '$lib/utils' -import { copilotInfo, getCurrentModel, getMetadataModel } from '$lib/aiStore' +import { + copilotInfo, + getCurrentModel, + getMetadataModel, + resolveMetadataModelFor, + resolveModelFor, + type AiTarget, + type CopilotInfo +} from '$lib/aiStore' import { emptyChatTokenUsage, openAICompletionsUsageToChatTokenUsage, @@ -279,7 +287,8 @@ export function getModelMaxTokens( function resolveMaxTokens( modelProvider: AIProviderModel, cap?: number, - reasoningEffort?: ReasoningEffort + reasoningEffort?: ReasoningEffort, + info?: CopilotInfo ): number { const defaultMaxTokens = getModelMaxTokens( modelProvider.provider, @@ -289,7 +298,7 @@ function resolveMaxTokens( const modelKey = `${modelProvider.provider}:${modelProvider.model}` let customMaxTokensStore: Record | undefined try { - customMaxTokensStore = get(copilotInfo)?.maxTokensPerModel + customMaxTokensStore = (info ?? get(copilotInfo))?.maxTokensPerModel } catch { // copilotInfo store may not be initialized in vitest } @@ -308,9 +317,10 @@ function getModelSpecificConfig( modelProvider: AIProviderModel, tools?: OpenAI.Chat.Completions.ChatCompletionTool[], maxTokensCap?: number, - reasoningEffort?: ReasoningEffort + reasoningEffort?: ReasoningEffort, + info?: CopilotInfo ) { - const maxTokens = resolveMaxTokens(modelProvider, maxTokensCap, reasoningEffort) + const maxTokens = resolveMaxTokens(modelProvider, maxTokensCap, reasoningEffort, info) if ( (modelProvider.provider === 'openai' || modelProvider.provider === 'azure_openai' || @@ -594,6 +604,7 @@ interface AnthropicCompletionParams { resourcePath?: string resourceValue?: Record maxTokensCap?: number + info?: CopilotInfo } function buildAnthropicProxyRequest({ @@ -603,7 +614,8 @@ function buildAnthropicProxyRequest({ workspace, resourcePath, resourceValue, - maxTokensCap + maxTokensCap, + info }: Omit) { const { system, messages: anthropicMessages } = convertOpenAIToAnthropicMessages(messages) @@ -623,7 +635,7 @@ function buildAnthropicProxyRequest({ const body = { model: modelProvider.model, - max_tokens: resolveMaxTokens(modelProvider, maxTokensCap), + max_tokens: resolveMaxTokens(modelProvider, maxTokensCap, undefined, info), messages: anthropicMessages, ...(system && { system }) } @@ -887,6 +899,11 @@ export function modelSupportsWebSearch(provider: AIProvider | undefined, model: return providerSupportsWebSearch(provider) } +/** The model a request sends: for `info`'s workspace when given, else the tab's. */ +function modelFor(info: CopilotInfo | undefined): AIProviderModel { + return info ? resolveModelFor(info) : getCurrentModel() +} + export function getProviderAndCompletionConfig({ messages, stream, @@ -894,7 +911,8 @@ export function getProviderAndCompletionConfig({ forceModelProvider, maxTokensCap, promptCaching, - reasoningEffort + reasoningEffort, + info }: { messages: ChatCompletionMessageParam[] stream: K @@ -907,13 +925,15 @@ export function getProviderAndCompletionConfig({ // The effort the caller will add with `applyReasoningToConfig`. Only sizes the // output budget here: a request that reasons needs room for the thinking. reasoningEffort?: ReasoningEffort + // The AI config of the workspace the request is made for, when it is not the tab's. + info?: CopilotInfo }): { provider: AIProvider config: K extends true ? ChatCompletionCreateParamsStreaming : ChatCompletionCreateParamsNonStreaming } { - const modelProvider = forceModelProvider ?? getCurrentModel() + const modelProvider = forceModelProvider ?? modelFor(info) const providerConfig = PROVIDER_COMPLETION_CONFIG_MAP[modelProvider.provider] const processedMessages = prepareMessages(modelProvider.provider, messages, { model: modelProvider.model, @@ -923,7 +943,7 @@ export function getProviderAndCompletionConfig({ provider: modelProvider.provider, config: { ...providerConfig, - ...getModelSpecificConfig(modelProvider, tools, maxTokensCap, reasoningEffort), + ...getModelSpecificConfig(modelProvider, tools, maxTokensCap, reasoningEffort, info), messages: processedMessages, stream } as any @@ -940,9 +960,10 @@ export async function getNonStreamingCompletion( workspace?: string // use a specific workspace proxy when testing a workspace resource forceModelProvider?: AIProviderModel maxTokensCap?: number // hard ceiling on output tokens (see METADATA_MAX_TOKENS) + info?: CopilotInfo // the AI config of `workspace`, when it is not the tab's } ) { - const modelProvider = options?.forceModelProvider ?? getCurrentModel() + const modelProvider = options?.forceModelProvider ?? modelFor(options?.info) if (usesAnthropicMessagesApi(modelProvider.provider, modelProvider.model)) { return getAnthropicNonStreamingCompletion({ @@ -953,7 +974,8 @@ export async function getNonStreamingCompletion( workspace: options?.workspace, resourcePath: options?.resourcePath, resourceValue: options?.resourceValue, - maxTokensCap: options?.maxTokensCap + maxTokensCap: options?.maxTokensCap, + info: options?.info }) } @@ -961,18 +983,18 @@ export async function getNonStreamingCompletion( const { provider, config } = getProviderAndCompletionConfig({ messages, stream: false, - forceModelProvider: options?.forceModelProvider, - maxTokensCap: options?.maxTokensCap + forceModelProvider: modelProvider, + maxTokensCap: options?.maxTokensCap, + info: options?.info }) // Use Responses API for OpenAI and Azure OpenAI if (provider === 'openai' || provider === 'azure_openai') { try { - const response = await getNonStreamingOpenAIResponsesCompletion( - messages, - abortController, - options - ) + const response = await getNonStreamingOpenAIResponsesCompletion(messages, abortController, { + ...options, + forceModelProvider: modelProvider + }) return response } catch (error) { console.error('Error using Responses API:', error) @@ -1003,13 +1025,17 @@ export async function getNonStreamingCompletion( return response } +/** `target` sends it for that workspace with its config; without one it uses the tab's. */ export async function getNonStreamingMetadataCompletion( messages: ChatCompletionMessageParam[], - abortController: AbortController + abortController: AbortController, + target?: AiTarget ) { return getNonStreamingCompletion(messages, abortController, { - forceModelProvider: getMetadataModel(), - maxTokensCap: METADATA_MAX_TOKENS + forceModelProvider: target ? resolveMetadataModelFor(target.info) : getMetadataModel(), + maxTokensCap: METADATA_MAX_TOKENS, + workspace: target?.workspace, + info: target?.info }) } @@ -1019,7 +1045,8 @@ export async function getFimCompletion( prompt: string, suffix: string, providerModel: AIProviderModel, - abortController: AbortController + abortController: AbortController, + workspace: string = get(workspaceStore)! ): Promise { // The Anthropic Messages API has no fill-in-the-middle endpoint, and Foundry // Claude deployments don't expose the OpenAI-compatible completions surface the @@ -1039,8 +1066,6 @@ export async function getFimCompletion( } } - const workspace = get(workspaceStore) - const response = await fetch( `${location.origin}${OpenAPI.BASE}/w/${workspace}/ai/proxy/fim/completions`, { @@ -1108,36 +1133,46 @@ export async function getCompletion( reasoningEffort?: string promptCaching?: boolean maxTokensCap?: number + // Send for this workspace, with its AI config, rather than the tab's. + workspace?: string + info?: CopilotInfo } ): Promise> { - const modelProvider = options?.forceModelProvider ?? getCurrentModel() + const modelProvider = options?.forceModelProvider ?? modelFor(options?.info) if (usesAnthropicMessagesApi(modelProvider.provider, modelProvider.model)) { return getAnthropicStreamingCompletion({ messages, modelProvider, abortController, - maxTokensCap: options?.maxTokensCap + maxTokensCap: options?.maxTokensCap, + workspace: options?.workspace, + info: options?.info }) } + const openaiClient = + options?.openaiClient ?? + (options?.workspace ? workspaceAIClients.createOpenaiClient(options.workspace) : undefined) const { provider, config } = getProviderAndCompletionConfig({ messages, stream: true, tools, - forceModelProvider: options?.forceModelProvider, + forceModelProvider: modelProvider, maxTokensCap: options?.maxTokensCap, promptCaching: options?.promptCaching, - reasoningEffort: options?.reasoningEffort + reasoningEffort: options?.reasoningEffort, + info: options?.info }) if (usesOpenAIResponsesApi(provider, modelProvider.model) && !options?.forceCompletions) { try { const stream = getOpenAIResponsesCompletionStream(messages, abortController, tools, { - forceModelProvider: options?.forceModelProvider, - openaiClient: options?.openaiClient, + forceModelProvider: modelProvider, + openaiClient, reasoningEffort: options?.reasoningEffort, - maxTokensCap: options?.maxTokensCap + maxTokensCap: options?.maxTokensCap, + info: options?.info }) as any return stream } catch (error) { @@ -1146,7 +1181,7 @@ export async function getCompletion( } // Use Completions API for other providers - const client = options?.openaiClient ?? workspaceAIClients.getOpenaiClient() + const client = openaiClient ?? workspaceAIClients.getOpenaiClient() // gpt-5.5+ refuse function tools here unless reasoning is off, so the Responses API // fallback turns it off where the model can rather than failing the turn. const reasoningEffort = @@ -1187,12 +1222,17 @@ export async function getCompletion( export async function getStreamedCompletionText( messages: ChatCompletionMessageParam[], abortController: AbortController, - options?: { maxTokensCap?: number } + // `workspace` and `info` send it for that workspace, with its AI config, rather than the tab's. + options?: { maxTokensCap?: number; workspace?: string; info?: CopilotInfo } ): Promise { + const modelProvider = modelFor(options?.info) const drain = async (forceCompletions: boolean) => { const stream = await getCompletion(messages, abortController, undefined, { forceCompletions, - maxTokensCap: options?.maxTokensCap + forceModelProvider: modelProvider, + maxTokensCap: options?.maxTokensCap, + workspace: options?.workspace, + info: options?.info }) let text = '' let finished = false @@ -1215,7 +1255,7 @@ export async function getStreamedCompletionText( } catch (error) { // The Responses API stream only fails once iterated, so the fallback to // chat completions for a deployment that doesn't serve it lives here. - const { provider, model } = getCurrentModel() + const { provider, model } = modelProvider if (abortController.signal.aborted || !usesOpenAIResponsesApi(provider, model)) { throw error } diff --git a/frontend/src/lib/components/copilot/loadCopilot.test.ts b/frontend/src/lib/components/copilot/loadCopilot.test.ts index 5bfcf1f547..a54d3be0ec 100644 --- a/frontend/src/lib/components/copilot/loadCopilot.test.ts +++ b/frontend/src/lib/components/copilot/loadCopilot.test.ts @@ -11,7 +11,12 @@ vi.mock('$lib/gen', () => ({ })) import { loadCopilot } from './loadCopilot' -import { copilotWorkspace } from '$lib/aiStore' +import { + copilotInfo, + copilotWorkspace, + getWorkspaceCopilotInfo, + loadWorkspaceCopilotInfo +} from '$lib/aiStore' const config = { providers: { anthropic: { resource_path: 'u/admin/anthropic', models: ['m'] } } } @@ -34,16 +39,34 @@ describe('loadCopilot', () => { it('lets a re-requested workspace win back over the one that superseded it', async () => { const a1 = loadCopilot('a') const b = loadCopilot('b') - // A per-workspace cache would hand back `a1` here — which has already lost the - // apply race to `b`, so `a` would silently never be written. + // `a2` may share `a1`'s download but must apply it on its own token: `a1` already lost + // the apply race to `b`, so handing back `a1` itself would leave `a` never written. const a2 = loadCopilot('a') - expect(requests.map((r) => r.workspace)).toEqual(['a', 'b', 'a']) + expect(requests.map((r) => r.workspace)).toEqual(['a', 'b']) requests[0].resolve(config) requests[1].resolve(config) - requests[2].resolve(config) await Promise.all([a1, b, a2]) expect(get(copilotWorkspace)).toBe('a') }) + + // The config can change elsewhere (another tab, another admin): a refresh of the tab's config + // must also be what a session in that workspace sends with. + it('refreshes the workspace config sessions send with, keeping the old one until then', async () => { + const withModel = (model: string) => ({ + providers: { anthropic: { resource_path: 'u/admin/anthropic', models: [model] } } + }) + const sessionLoad = loadWorkspaceCopilotInfo('refreshed') + requests[0].resolve(withModel('old')) + await sessionLoad + const refresh = loadCopilot('refreshed') + await vi.waitFor(() => expect(requests).toHaveLength(2)) + expect(getWorkspaceCopilotInfo('refreshed')?.aiModels[0]?.model).toBe('old') + requests[1].resolve(withModel('new')) + await refresh + + expect(get(copilotInfo).aiModels[0]?.model).toBe('new') + expect(getWorkspaceCopilotInfo('refreshed')?.aiModels[0]?.model).toBe('new') + }) }) diff --git a/frontend/src/lib/components/copilot/loadCopilot.ts b/frontend/src/lib/components/copilot/loadCopilot.ts index 4b1e53a686..62a97e8dbf 100644 --- a/frontend/src/lib/components/copilot/loadCopilot.ts +++ b/frontend/src/lib/components/copilot/loadCopilot.ts @@ -1,5 +1,9 @@ -import { WorkspaceService } from '$lib/gen' -import { copilotWorkspace, setCopilotInfo } from '$lib/aiStore' +import { + applyCopilotInfo, + copilotWorkspace, + loadWorkspaceCopilotInfo, + setCopilotInfo +} from '$lib/aiStore' import { setAiClientWorkspace } from './aiClientWorkspace' import { writable } from 'svelte/store' @@ -44,9 +48,11 @@ export function loadCopilot(workspace: string): Promise { async function fetchAndApply(workspace: string, token: number) { setAiClientWorkspace(workspace) try { - const info = await WorkspaceService.getCopilotInfo({ workspace }) + // Through the per-workspace cache: a session in this workspace then sends with the config + // the tab shows. + const info = await loadWorkspaceCopilotInfo(workspace, true) if (token !== loadCopilotToken) return - setCopilotInfo(info) + applyCopilotInfo(info) copilotWorkspace.set(workspace) } catch (err) { if (token !== loadCopilotToken) return diff --git a/frontend/src/lib/components/deploymentRequest/AiReview.svelte b/frontend/src/lib/components/deploymentRequest/AiReview.svelte index 0af4488718..33964f60cc 100644 --- a/frontend/src/lib/components/deploymentRequest/AiReview.svelte +++ b/frontend/src/lib/components/deploymentRequest/AiReview.svelte @@ -1,7 +1,12 @@