diff --git a/chatbot/src/App.jsx b/chatbot/src/App.jsx index 644dae2f..0c1daf0f 100644 --- a/chatbot/src/App.jsx +++ b/chatbot/src/App.jsx @@ -342,15 +342,19 @@ export default function App() { [repositorySelections, setRepositorySelections] = useState({}), [uploading, setUploading] = useState(false), [runsRefresh, setRunsRefresh] = useState(0), - [approval, setApprovalState] = useState(storedApprovalMode), + [approvalSelection, setApprovalSelection] = useState(null), [quickAccessOpen, setQuickAccessOpen] = useState( () => new URLSearchParams(location.search).get("quickAccess") === "1", ), [theme, setTheme] = useState(() => readThemePreference(localStorage)), [environmentVisibility, setEnvironmentVisibility] = useState(() => readJSON(environmentVisibilityKey, {})); + const approvalUserId = realChat.user?.isGuest ? null : realChat.user?.id ?? null; + const approval = approvalSelection?.userId === approvalUserId + ? approvalSelection.value + : storedApprovalMode(approvalUserId); function setApproval(value) { - setApprovalState(value); - rememberApprovalMode(value); + setApprovalSelection({ userId: approvalUserId, value }); + rememberApprovalMode(approvalUserId, value); } const environmentPreferenceOwner = String(chat.user?.id || "guest"); const showEnvironmentPicker = environmentVisibility[environmentPreferenceOwner] === true; @@ -460,7 +464,7 @@ export default function App() { chat.projects.find((p) => p.id === chat.projectId)?.name || (chat.projectId && (chat.session?.projectName || chat.projectId)) || "个人对话"; - const selectedModel = chat.models.find((m) => m.id === chat.model); + const selectedModel = chat.models.find((m) => m.id === chat.model) || (!chat.local ? chat.modelDetails : null); const modelLabel = selectedModel?.name?.replace(/^Claude-/, "").replace(/^GPT-/, "") || "模型"; // The CLI switch is only for a conversation that has not started. A reply diff --git a/chatbot/src/PermissionPicker.jsx b/chatbot/src/PermissionPicker.jsx index 658eb595..0a37c830 100644 --- a/chatbot/src/PermissionPicker.jsx +++ b/chatbot/src/PermissionPicker.jsx @@ -1,6 +1,7 @@ import React, { useCallback, useEffect, useId, useLayoutEffect, useRef, useState } from "react"; import { createPortal } from "react-dom"; import HarnessIcon from "./HarnessIcon"; +import { readComposerPreference, writeComposerPreference } from "./composer-preferences.js"; import { usePanelPresence } from "./usePanelPresence"; import "./permission-picker.css"; @@ -10,27 +11,16 @@ const modes = [ { id: "auto", icon: "priority-bolt", label: "完全权限", tone: "danger", detail: "工作区文件和命令自动执行;平台及外部操作使用当前账号权限直接执行。" }, ]; -// The approval mode is a personal preference: new conversations and reloads -// keep the last pick instead of going back to "请求批准". 完全权限 is never -// stored: it lets platform and external writes run unreviewed, so it lasts -// for this page only and a reload returns to the last reviewed mode. -const APPROVAL_KEY = "loopforge.assistant.approvalMode"; -const rememberedMode = (value) => value === "confirm" || value === "assisted"; -export function storedApprovalMode() { - try { - const value = localStorage.getItem(APPROVAL_KEY); - return rememberedMode(value) ? value : "confirm"; - } catch { - return "confirm"; - } +const validApprovalMode = (value) => modes.some((mode) => mode.id === value); + +export function storedApprovalMode(userId) { + if (!userId) return "confirm"; + const value = readComposerPreference(`approvalMode.${userId}`); + return validApprovalMode(value) ? value : "confirm"; } -export function rememberApprovalMode(value) { - if (!rememberedMode(value)) return; - try { - localStorage.setItem(APPROVAL_KEY, value); - } catch { - // Without storage the pick still applies to this page. - } + +export function rememberApprovalMode(userId, value) { + if (userId && validApprovalMode(value)) writeComposerPreference(`approvalMode.${userId}`, value); } // `glyph` is a rendered brand mark. Options that only set `icon` keep the Harness icon. diff --git a/chatbot/src/composer-preferences.js b/chatbot/src/composer-preferences.js new file mode 100644 index 00000000..b03aaa09 --- /dev/null +++ b/chatbot/src/composer-preferences.js @@ -0,0 +1,17 @@ +const prefix = "loopforge.assistant."; + +export function readComposerPreference(name) { + try { + return localStorage.getItem(prefix + name); + } catch { + return null; + } +} + +export function writeComposerPreference(name, value) { + try { + localStorage.setItem(prefix + name, value); + } catch { + // The selection still applies to the current page. + } +} diff --git a/chatbot/src/effort.js b/chatbot/src/effort.js index 42fb577c..bf614014 100644 --- a/chatbot/src/effort.js +++ b/chatbot/src/effort.js @@ -1,5 +1,6 @@ const levels = { none: { label: "关闭", icon: "sparkles", detail: "不使用额外推理。" }, + thinking: { label: "开启", icon: "brain", detail: "开启思考模式。" }, minimal: { label: "最低", icon: "effort-low", detail: "只做最少推理,优先尽快返回。" }, low: { label: "低", icon: "effort-low", detail: "优先响应速度,适合简单问题。" }, medium: { label: "中", icon: "effort-medium", detail: "兼顾响应速度与推理深度。" }, @@ -7,6 +8,7 @@ const levels = { xhigh: { label: "更高", icon: "search", detail: "投入更多推理时间,适合困难任务。" }, max: { label: "最高", icon: "effort-max", detail: "使用模型支持的最高思考强度,耗时可能更长。" }, ultra: { label: "极致", icon: "effort-max", detail: "使用极致推理强度,耗时可能更长。" }, + ultracode: { label: "极致", icon: "effort-max", detail: "使用模型的 ultracode 思考强度,耗时可能更长。" }, }; export function modelEffort(model, value) { diff --git a/chatbot/src/message-queue.js b/chatbot/src/message-queue.js index 6642260f..8d18fa5d 100644 --- a/chatbot/src/message-queue.js +++ b/chatbot/src/message-queue.js @@ -20,6 +20,7 @@ export function queuedMessage({ attachments = [], references = [], approvalMode = "confirm", + approvalUserId, }) { return { id: localId("queued"), @@ -27,10 +28,19 @@ export function queuedMessage({ attachments, references, approvalMode, + approvalUserId, queuedAt: new Date().toISOString(), }; } +export function queuedApproval({ approvalMode, approvalUserId }, userId) { + return { + // Queues saved before account binding require confirmation when resumed. + approvalMode: approvalUserId ? approvalMode || "confirm" : "confirm", + approvalUserId: approvalUserId ?? userId, + }; +} + export function enqueue(queues, sessionId, item) { const entry = entryOf(queues, sessionId); return withEntry(queues, sessionId, { ...entry, items: [...entry.items, item] }); diff --git a/chatbot/src/useChat.js b/chatbot/src/useChat.js index 2d4eb0cb..cbac0a83 100644 --- a/chatbot/src/useChat.js +++ b/chatbot/src/useChat.js @@ -17,10 +17,12 @@ import { useExecutionMachines } from "./useExecutionMachines.js"; import { executionLocked } from "./execution-target.js"; import { modelEffort } from "./effort.js"; import { conversationEngine, engineModels } from "./engines.js"; +import { readComposerPreference, writeComposerPreference } from "./composer-preferences.js"; import { enqueue, holdQueue, nextQueued, + queuedApproval, queuedMessage, readQueues, removeQueued, @@ -40,47 +42,42 @@ export const EXECUTORS = [ { id: "dsh", icon: "terminal", label: "DeepSeek Harness", detail: "使用 DeepSeek Harness 与个人 API Key。" }, { id: "zcode", icon: "terminal", label: "ZCode", detail: "使用 ZCode 与个人 API Key。" }, ]; -const EXECUTOR_KEY = "loopforge.assistant.executor"; function knownExecutor(id) { if (id === "claude") return "claude-code"; return EXECUTORS.some((item) => item.id === id) ? id : "codebuddy"; } function storedExecutor() { - try { - return knownExecutor(localStorage.getItem(EXECUTOR_KEY)); - } catch { - return "codebuddy"; - } + return knownExecutor(readComposerPreference("executor")); } -// The chat model is a personal preference like the executor: new -// conversations and page reloads keep the last pick. Each CLI remembers its -// own, because model IDs differ between CLIs. -const MODEL_KEY = "loopforge.assistant.model."; -function storedModel(executor) { - try { - return localStorage.getItem(MODEL_KEY + executor) || ""; - } catch { - return ""; - } + +function executorModelList(models, executor, providerCatalogs) { + const listed = engineModels(models, executor); + if (executor === "codebuddy") return listed.filter((model) => model.id !== "default"); + return listed.length ? listed : providerCatalogs?.[executor]?.models || []; } -function rememberModel(executor, id) { - if (!id) return; - try { - localStorage.setItem(MODEL_KEY + executor, id); - } catch { - // Without storage the pick still applies to this page. - } + +// A CLI's fallback list uses bare model IDs, while the personal catalog +// prefixes them with the CLI. Both spellings share the same preference. +function unprefixedModel(executor, model) { + const prefix = executor + ":"; + return model?.startsWith(prefix) ? model.slice(prefix.length) : model; } -const codeBuddyModels = (models) => (models || []).filter((m) => !m.engine || m.engine === "codebuddy"); -const executorModelList = (models, executor) => - executor === "codebuddy" ? codeBuddyModels(models) : engineModels(models, executor); -// The model a new conversation on executor starts with: the remembered one -// while list still offers it, else the first listed. Before a CLI's list has -// loaded the remembered model (or that CLI's own default) holds the place. -function preferredModel(list, executor) { - const remembered = storedModel(executor); - if (!list?.length) return remembered || (executor === "codebuddy" ? "" : "default"); - return list.some((m) => m.id === remembered) ? remembered : list[0].id; + +function findExecutorModel(list, executor, model) { + const id = unprefixedModel(executor, model); + return list.find((item) => unprefixedModel(executor, item.id) === id); +} + +function effortPreference(executor, model) { + return `effort.${executor}.${unprefixedModel(executor, model)}`; +} + +// A new conversation restores a listed preference or picks the first model. +// When the list is empty, keep the CLI's fallback. +function preferredModel(list, executor, codeBuddyDefault) { + const remembered = readComposerPreference(`model.${executor}`); + const listed = findExecutorModel(list, executor, remembered); + return listed?.id || list[0]?.id || (executor === "codebuddy" ? codeBuddyDefault : remembered || "default"); } // Mirrors errAssistantSponsoredModel in internal/httpapi/assistant_sponsor.go; @@ -117,6 +114,7 @@ export function useChat() { new URLSearchParams(location.search).get("issueId") || "", ), [models, setModels] = useState([]), + [modelCatalogRefresh, requestModelCatalogRefresh] = useState(0), [sessions, setSessions] = useState([]), [session, setSession] = useState(null), [activeId, setActiveId] = useState( @@ -126,8 +124,9 @@ export function useChat() { // the header reveals that conversation's title once it has loaded. [revealId, setRevealId] = useState(activeId), [projectId, setProjectId] = useState(initialProject), - [model, setModel] = useState(() => storedModel(storedExecutor())), - [effort, setEffort] = useState(""), + [model, setModel] = useState(null), + [codeBuddyDefault, setCodeBuddyDefault] = useState({ id: "default", name: "默认" }), + [effort, setEffort] = useState(null), [sponsoredModels, setSponsoredModels] = useState([]), [executor, setExecutorState] = useState(storedExecutor), [providerCatalogs, setProviderCatalogs] = useState({}), @@ -213,26 +212,26 @@ export function useChat() { setLive({ answer: "", thinking: "", tools: [] }); setSession(null); executionMachines.resetMachine(); - // A new conversation starts on the person's remembered model, not on - // whatever model the previously opened conversation used. - if (!id) { - const preferred = storedExecutor(); - setExecutorState(preferred); - setModel(preferredModel(executorModelList(models, preferred), preferred)); - } - setEffort(""); + if (!id) setExecutorState(storedExecutor()); + setModel(null); + setEffort(null); setSandboxProfileId(""); currentId.current = id; setActiveId(id); if (id) setRevealId(id); changeURL(id); }, - [changeURL, models, executionMachines.resetMachine], + [changeURL, executionMachines.resetMachine], ); const newChat = useCallback(() => { setDraftIssueId(""); open(null); }, [open]); + const applyModelCatalog = useCallback((catalog) => { + setModels(catalog.models || []); + setCodeBuddyDefault({ id: catalog.default || "default", name: "默认", efforts: catalog.defaultEfforts }); + setSponsoredModels(sponsoredModelList(catalog)); + }, []); useEffect(() => { if (!user?.id || user.isGuest) { setSteerOnSend(false); @@ -282,10 +281,8 @@ export function useChat() { setUsers(users || []); setProjects(projectRows || []); setSandboxes(sandboxRows || []); - setModels(catalog.models || []); - setSponsoredModels(sponsoredModelList(catalog)); - const startExecutor = storedExecutor(); - setModel(preferredModel(executorModelList(catalog.models, startExecutor), startExecutor)); + applyModelCatalog(catalog); + if (catalog.cursorModelsPending) requestModelCatalogRefresh((revision) => revision + 1); await refreshSessions(); if (!current) return; const recovered = connectionError; @@ -304,7 +301,7 @@ export function useChat() { current = false; clearTimeout(timer); }; - }, [refreshSessions]); + }, [applyModelCatalog, refreshSessions]); useEffect(() => { if (!ready || !activeId) return; // send/edit already supplied this session. Reloading it here can return an @@ -325,7 +322,7 @@ export function useChat() { const storedEngine = knownExecutor(locked || stored.executor); setExecutorState(storedEngine); // A conversation shows the model it runs on. - setModel(stored.model || preferredModel(executorModelList(models, storedEngine), storedEngine)); + setModel(stored.model || null); setEffort(stored.effort || ""); setProjectId(stored.projectId || ""); syncProject(stored.projectId || ""); @@ -599,17 +596,29 @@ export function useChat() { ); const codeBuddy = executor === "codebuddy"; const listedForEngine = engineModels(models, executor); - const executorModels = listedForEngine.length || codeBuddy - ? (codeBuddy ? models.filter((m) => !m.engine || m.engine === "codebuddy") : listedForEngine) - : providerCatalogs[executor]?.models || []; + const executorModels = executorModelList(models, executor, providerCatalogs); const missingKey = !!user && !user.codeBuddyApiKeyConfigured; // Without their own key users chat on the platform's shared key, limited to // the models the administrator allowed (the first is the default). const sponsoredModel = sponsoredModels[0] || ""; const sponsored = codeBuddy && missingKey && !user.isGuest && !!sponsoredModel; const sharedModelFor = (id) => (sponsoredModels.includes(id) ? id : sponsoredModel); - const activeModel = sponsored ? sharedModelFor(model) : model; - const activeEffort = modelEffort(executorModels.find((m) => m.id === activeModel), effort); + function modelEntry(id) { + return findExecutorModel(executorModels, executor, id) ?? + (codeBuddy && id === codeBuddyDefault.id ? codeBuddyDefault : undefined); + } + const hasSelectedModel = model !== null && + (session || !executorModels.length || findExecutorModel(executorModels, executor, model)); + const selectedModel = hasSelectedModel ? model : preferredModel(executorModels, executor, codeBuddyDefault.id); + const availableModel = sponsored ? sharedModelFor(selectedModel) : selectedModel; + const activeModelEntry = modelEntry(availableModel); + const activeModel = activeModelEntry?.id ?? availableModel; + // null follows the browser preference, while "" is the CLI's default level. + const activeEffort = modelEffort( + activeModelEntry, + ((model === null || hasSelectedModel) ? effort : null) ?? + readComposerPreference(effortPreference(executor, activeModel)), + ); useEffect(() => { if (codeBuddy || listedForEngine.length || providerCatalogs[executor]) return; let current = true; @@ -617,13 +626,6 @@ export function useChat() { .then((catalog) => { if (!current) return; setProviderCatalogs((old) => ({ ...old, [executor]: catalog })); - // An open conversation keeps the model it runs on: the catalog lists - // bare IDs without the CLI prefix, and no gateway models. - setModel((old) => - (catalog.models || []).some((m) => m.id === old) || (old && old === sessionRef.current?.model) - ? old - : preferredModel(catalog.models, executor), - ); }) .catch((e) => current && setError(e.message)); return () => { @@ -635,30 +637,44 @@ export function useChat() { if (id === "claude") id = "claude-code"; if (!EXECUTORS.some((item) => item.id === id) || id === executor) return; const locked = conversationEngine(sessionRef.current?.messages, models, sessionRef.current?.executor); - try { - localStorage.setItem(EXECUTOR_KEY, id); - } catch {} - setEffort(""); - const listed = executorModelList(models, id); - setExecutorState(id); - setModel(preferredModel(listed.length || id === "codebuddy" ? listed : providerCatalogs[id]?.models, id)); + writeComposerPreference("executor", id); if (locked && locked !== id) open(null); + setExecutorState(id); + setModel(null); + setEffort(null); }, - [executor, models, open, providerCatalogs], - ); - const chooseModel = useCallback( - (id) => { - if (sponsored && !sponsoredModels.includes(id)) { - setError(sponsoredModelMessage(sponsoredModels)); - return false; - } - setModel(id); - rememberModel(executor, id); - setEffort((value) => modelEffort(executorModels.find((m) => m.id === id), value)); - return true; - }, - [sponsored, sponsoredModels, executorModels, executor], + [executor, models, open], ); + + function rememberModelChoice(id, level) { + writeComposerPreference("executor", executor); + writeComposerPreference(`model.${executor}`, id); + writeComposerPreference(effortPreference(executor, id), level); + } + + function chooseModel(id) { + if (sponsored && !sponsoredModels.includes(id)) { + setError(sponsoredModelMessage(sponsoredModels)); + return false; + } + const entry = executorModels.find((item) => item.id === id); + if (!entry || entry.disabled || entry.unavailable) return false; + const savedLevel = readComposerPreference(effortPreference(executor, id)); + const level = id === activeModel ? activeEffort : modelEffort(entry, savedLevel ?? activeEffort); + setModel(id); + setEffort(level); + rememberModelChoice(id, level); + return true; + } + + function chooseEffort(value) { + const level = modelEffort(activeModelEntry, value); + setEffort(level); + if (activeModelEntry) { + setModel(activeModel); + rememberModelChoice(activeModel, level); + } + } // Credential flags are omitted when false, so drop the old ones before // merging; otherwise a cleared key would still look configured. const refreshUser = useCallback( @@ -685,22 +701,44 @@ export function useChat() { [], ); useEffect(() => { - if (!ready) return; const onMessage = (e) => { if ( e.source === window.parent && e.origin === location.origin && e.data?.type === "agentflow-credentials-changed" - ) - void refreshUser(); + ) requestModelCatalogRefresh((revision) => revision + 1); }; window.addEventListener("message", onMessage); - if (missingKey) window.addEventListener("focus", refreshUser); + if (ready && missingKey) window.addEventListener("focus", refreshUser); return () => { window.removeEventListener("message", onMessage); window.removeEventListener("focus", refreshUser); }; }, [ready, missingKey, refreshUser]); + useEffect(() => { + if (!ready || !modelCatalogRefresh) return; + void refreshUser(); + const request = new AbortController(); + // Cursor discovery can run for up to 30 seconds in the background. + const deadline = Date.now() + 40000; + let timer; + async function refreshModels() { + try { + const catalog = await api("/api/assistant/models", { signal: request.signal }); + if (request.signal.aborted) return; + applyModelCatalog(catalog); + if (!catalog.cursorModelsPending) return; + } catch (error) { + if (error.status >= 400 && error.status < 500) return; + } + if (!request.signal.aborted && Date.now() < deadline) timer = setTimeout(refreshModels, 1500); + } + void refreshModels(); + return () => { + request.abort(); + clearTimeout(timer); + }; + }, [ready, modelCatalogRefresh, refreshUser, applyModelCatalog]); useEffect(() => writeQueues(queues), [queues]); useEffect(() => { const s = steeringRef.current; @@ -800,7 +838,7 @@ export function useChat() { references: item.references || [], model: activeModel, effort: activeEffort, - approvalMode: item.approvalMode || "confirm", + ...queuedApproval(item, user?.id), }, ); request.pending = false; @@ -824,10 +862,14 @@ export function useChat() { attachments = [], references = [], approvalMode = "confirm", + approvalUserId, messagesOverride = null, fromQueue = false, }) { if (!text.trim() && !attachments.length) return false; + const approval = fromQueue + ? queuedApproval({ approvalMode, approvalUserId }, user?.id) + : { approvalMode, approvalUserId: user?.id }; if (currentId.current && sessionRef.current?.id !== currentId.current) { setError("会话仍在加载,请稍候再发送。"); return false; @@ -843,7 +885,7 @@ export function useChat() { return false; } const id = currentId.current; - const item = queuedMessage({ text, attachments, references, approvalMode }); + const item = queuedMessage({ text, attachments, references, ...approval }); if (steerOnSend && generation?.status === "running" && generation.id && !submittingRef.current && !steeringRef.current && !dispatchingRef.current && !stopRequestRef.current) { @@ -903,6 +945,7 @@ export function useChat() { effort: activeEffort, executor, title: text.trim().slice(0, 36) || "图片与文件", + approvalUserId: approval.approvalUserId, messages: [], }, "PUT", @@ -946,7 +989,7 @@ export function useChat() { executor, model: activeModel, effort: activeEffort, - approvalMode, + ...approval, executionTarget, messages, }); @@ -1040,7 +1083,7 @@ export function useChat() { ); open(stored.id); setSession(stored); - setModel(stored.model || preferredModel(executorModels, executor)); + setModel(stored.model || null); setEffort(stored.effort || ""); await refreshSessions(); return stored; @@ -1057,7 +1100,7 @@ export function useChat() { setSubmitting(true); open(stored.id); setSession(stored); - const chosen = sponsored ? sharedModelFor(stored.model || model) : stored.model || model; + const chosen = sponsored ? sharedModelFor(stored.model || activeModel) : stored.model || activeModel; setModel(chosen); setEffort(stored.effort || ""); const g = await write("/api/assistant/generations", { @@ -1066,9 +1109,10 @@ export function useChat() { runId: stored.runId || "", executor, model: chosen, - effort: codeBuddy ? modelEffort(models.find((m) => m.id === chosen), stored.effort || "") : "", + effort: modelEffort(modelEntry(chosen), stored.effort || ""), messages: stored.messages, approvalMode: ["confirm", "assisted", "auto"].includes(approvalMode) ? approvalMode : "confirm", + approvalUserId: user?.id, }); setGeneration(g); await refreshSessions(); @@ -1365,8 +1409,9 @@ export function useChat() { revealId, projectId, model: activeModel, + modelDetails: activeModelEntry, effort: activeEffort, - setEffort, + setEffort: chooseEffort, sponsored, sponsoredModel, sponsoredModels, diff --git a/chatbot/tests/app-render.test.jsx b/chatbot/tests/app-render.test.jsx index f7cb18c4..dcea5e04 100644 --- a/chatbot/tests/app-render.test.jsx +++ b/chatbot/tests/app-render.test.jsx @@ -104,16 +104,30 @@ describe("App", () => { await user.click(screen.getByRole("button", { name: "发送消息" })); expect(chat.send).toHaveBeenLastCalledWith(expect.objectContaining({ approvalMode: "assisted" })); }); - test("keeps full permission for this page only", async () => { + test("restores full permission after a reload and resets it for another account", async () => { const user = userEvent.setup(); - localStorage.setItem("loopforge.assistant.approvalMode", "assisted"); const first = render(); await user.click(screen.getByRole("button", { name: "操作确认方式" })); await user.click(screen.getByRole("menuitemradio", { name: "完全权限" })); - expect(screen.getByRole("button", { name: "操作确认方式" })).toHaveTextContent("完全权限"); first.unmount(); - render(); - expect(screen.getByRole("button", { name: "操作确认方式" })).toHaveTextContent("帮我批准"); + const reloaded = render(); + fireEvent.keyDown(window, { key: "O", ctrlKey: true, shiftKey: true }); + expect(chat.newChat).toHaveBeenCalled(); + expect(screen.getByRole("button", { name: "操作确认方式" })).toHaveTextContent("完全权限"); + await user.type(screen.getByRole("textbox", { name: /消息|输入/ }), "你好"); + await user.click(screen.getByRole("button", { name: "发送消息" })); + expect(chat.send).toHaveBeenLastCalledWith(expect.objectContaining({ approvalMode: "auto" })); + const originalUser = chat.user; + try { + chat.user = { id: "u2", loginName: "another-user" }; + reloaded.rerender(); + expect(screen.getByRole("button", { name: "操作确认方式" })).toHaveTextContent("请求批准"); + await user.type(screen.getByRole("textbox", { name: "消息" }), "新账号"); + await user.click(screen.getByRole("button", { name: "发送消息" })); + expect(chat.send).toHaveBeenLastCalledWith(expect.objectContaining({ approvalMode: "confirm" })); + } finally { + chat.user = originalUser; + } }); test("does not ask for per-command native approval in a working conversation", () => { const previous = {...chat}; diff --git a/chatbot/tests/effort.test.mjs b/chatbot/tests/effort.test.mjs index 1b9bca9a..0f65ce13 100644 --- a/chatbot/tests/effort.test.mjs +++ b/chatbot/tests/effort.test.mjs @@ -10,7 +10,7 @@ test("effort choices follow the current model catalog and keep the native defaul assert.equal(modelEffort(undefined, "high"), ""); }); -test("cursor-only effort levels none, minimal and ultra are offered when the model lists them", () => { +test("effort levels none, minimal and ultra are offered when the model lists them", () => { const options = effortOptions({ efforts: ["none", "minimal", "low", "ultra", "bogus"] }); assert.deepEqual(options.map((v) => v.id), ["", "none", "minimal", "low", "ultra"]); for (const option of options) { @@ -18,3 +18,18 @@ test("cursor-only effort levels none, minimal and ultra are offered when the mod } assert.equal(modelEffort({ efforts: ["none", "ultra"] }, "ultra"), "ultra"); }); + +test("models with a thinking switch show default, off and on", () => { + const model = { efforts: ["none", "thinking"] }; + assert.deepEqual(effortOptions(model).map(({ id, label }) => [id, label]), [ + ["", "默认"], + ["none", "关闭"], + ["thinking", "开启"], + ]); + assert.equal(modelEffort(model, "thinking"), "thinking"); +}); + +test("custom model effort metadata can expose ultracode or only the native default", () => { + assert.deepEqual(effortOptions({ efforts: ["high", "ultracode"] }).map(({ id }) => id), ["", "high", "ultracode"]); + assert.deepEqual(effortOptions({ efforts: [] }).map(({ id }) => id), [""]); +}); diff --git a/chatbot/tests/hooks.test.jsx b/chatbot/tests/hooks.test.jsx index b6171343..59f0a66e 100644 --- a/chatbot/tests/hooks.test.jsx +++ b/chatbot/tests/hooks.test.jsx @@ -17,7 +17,7 @@ const { apiMock, writeMock } = vi.hoisted(() => { return { result: { structuredContent: { data: { id: "saved" } } } }; if (path === "/api/auth/current") return { user: { id: 1, loginName: "tester", codeBuddyApiKeyConfigured: true } }; if (path === "/api/projects") return [{ id: "p1", name: "项目一", ownerId: 1, members: [1] }]; - if (path === "/api/assistant/models?executor=codex") return { default: "default", models: [{ id: "default", name: "CLI default", family: "gpt" }, { id: "gpt-5", name: "GPT-5", family: "gpt" }] }; + if (path === "/api/assistant/models?executor=codex") return { default: "default", models: [{ id: "default", name: "CLI default", family: "gpt" }, { id: "gpt-5", name: "GPT-5", family: "gpt", efforts: ["high", "xhigh"] }] }; if (path === "/api/assistant/models") return { default: "m1", sponsoredModel: "m1", models: [{ id: "m1", name: "模型一", effort: true }] }; if (path === "/api/users") return [{ id: 1, loginName: "tester", codeBuddyApiKeyConfigured: true }]; if (path === "/api/assistant/sessions") return [baseSession]; @@ -44,6 +44,7 @@ vi.mock("../src/api.js", async (original) => ({ })); import { EXECUTORS, useChat } from "../src/useChat.js"; +import { writeComposerPreference } from "../src/composer-preferences.js"; import { isLocalConversation, useLocalConversations, @@ -82,7 +83,6 @@ describe("conversation hooks", () => { expect(result.current.executors.map((item) => item.id)).toEqual(EXECUTORS.map((item) => item.id)); act(() => result.current.setExecutor("codex")); await waitFor(() => expect(result.current.models.map((m) => m.id)).toEqual(["default", "gpt-5"])); - expect(localStorage.getItem("loopforge.assistant.executor")).toBe("codex"); expect(result.current.model).toBe("default"); expect(result.current.effort).toBe(""); expect(result.current.sponsored).toBe(false); @@ -98,6 +98,62 @@ describe("conversation hooks", () => { localStorage.removeItem("loopforge.assistant.executor"); }); + test("restores the CLI and model saved by an earlier version", async () => { + localStorage.setItem("loopforge.assistant.executor", "codex"); + localStorage.setItem("loopforge.assistant.model.codex", "gpt-5"); + const { result } = renderHook(() => useChat()); + await waitFor(() => expect(result.current.ready).toBe(true)); + expect(result.current.executor).toBe("codex"); + await waitFor(() => expect(result.current.model).toBe("gpt-5")); + }); + + test.each(["after saving a key", "during the first open"])("refreshes a pending Cursor catalog %s", async (scenario) => { + const original = apiMock.getMockImplementation(); + let configured = scenario === "during the first open"; + let discovered = false; + let pendingReads = 0; + const fallback = { id: "widget", name: "Widget" }; + writeComposerPreference("executor", "cursor"); + apiMock.mockImplementation(async (path, options) => { + if (path === "/api/assistant/models?executor=cursor") return { models: [fallback] }; + if (path === "/api/assistant/models" && configured) { + if (!discovered) pendingReads++; + return { + default: "m1", + cursorModelsPending: !discovered, + models: [{ ...fallback, id: "cursor:widget", engine: "cursor", efforts: discovered ? ["low", "xhigh"] : [] }], + }; + } + return original(path, options); + }); + try { + const { result } = renderHook(() => useChat()); + await waitFor(() => expect(result.current.modelDetails?.name).toBe("Widget")); + await waitFor(() => expect(result.current.ready).toBe(true)); + expect(result.current.modelDetails.efforts || []).toEqual([]); + act(() => { result.current.setModel(result.current.model); }); + if (!configured) { + configured = true; + act(() => window.dispatchEvent(new MessageEvent("message", { + origin: location.origin, + source: window.parent, + data: { type: "agentflow-credentials-changed" }, + }))); + } + await waitFor(() => expect(pendingReads).toBeGreaterThan(0)); + discovered = true; + await waitFor(() => expect(result.current.modelDetails.efforts).toEqual(["low", "xhigh"]), { timeout: 3500 }); + act(() => result.current.setEffort("xhigh")); + expect(result.current.effort).toBe("xhigh"); + await act(() => result.current.send({ text: "验证强度" })); + expect(writeMock).toHaveBeenCalledWith( + "/api/assistant/generations", expect.objectContaining({ executor: "cursor", model: "cursor:widget", effort: "xhigh" }), + ); + } finally { + apiMock.mockImplementation(original); + } + }); + test("without an own key chat is limited to the administrator's shared models", async () => { localStorage.removeItem("loopforge.assistant.executor"); const original = apiMock.getMockImplementation(); @@ -127,52 +183,86 @@ describe("conversation hooks", () => { } }); - test("remembers each CLI's model across new conversations and reloads", async () => { - localStorage.removeItem("loopforge.assistant.executor"); + test("restores model and effort choices in fresh chats and across catalog sources", async () => { + localStorage.setItem("loopforge.assistant.model.codebuddy", "default"); const original = apiMock.getMockImplementation(); + let personalCodexCatalog = false; apiMock.mockImplementation(async (path, options) => { - if (path === "/api/assistant/models") - return { default: "default", models: [{ id: "default", name: "CLI 默认模型" }, { id: "claude-opus-5.5", name: "Opus" }] }; + if (path === "/api/assistant/models") { + const models = [ + { id: "default", name: "CLI 默认模型" }, + { id: "m1", name: "模型一", efforts: ["low"] }, + { id: "claude-opus-5.5", name: "Opus", efforts: ["max"] }, + ]; + if (personalCodexCatalog) + models.push({ id: "codex:gpt-5", engine: "codex", name: "GPT-5", efforts: ["high"] }); + return { default: "default", defaultEfforts: ["low"], models }; + } return original(path, options); }); try { const first = renderHook(() => useChat()); await waitFor(() => expect(first.result.current.ready).toBe(true)); - // Nothing remembered yet: the first listed model. - expect(first.result.current.model).toBe("default"); + expect(first.result.current.models.map((m) => m.id)).toEqual(["m1", "claude-opus-5.5"]); + expect(first.result.current.model).toBe("m1"); + act(() => first.result.current.setEffort("low")); + await act(() => first.result.current.send({ text: "运行第一个可用模型" })); + expect(writeMock).toHaveBeenCalledWith( + "/api/assistant/generations", + expect.objectContaining({ model: "m1", effort: "low" }), + ); + act(() => first.result.current.newChat()); act(() => { first.result.current.setModel("claude-opus-5.5"); }); - expect(localStorage.getItem("loopforge.assistant.model.codebuddy")).toBe("claude-opus-5.5"); - // A new conversation no longer goes back to the default model. + act(() => first.result.current.setEffort("max")); + act(() => first.result.current.open("s1")); + await waitFor(() => expect(first.result.current.session?.id).toBe("s1")); + expect(first.result.current.model).toBe("m1"); + expect(first.result.current.effort).toBe(""); act(() => first.result.current.newChat()); expect(first.result.current.model).toBe("claude-opus-5.5"); + expect(first.result.current.effort).toBe("max"); await act(() => first.result.current.send({ text: "你好" })); expect(writeMock).toHaveBeenCalledWith( "/api/assistant/generations", - expect.objectContaining({ model: "claude-opus-5.5" }), + expect.objectContaining({ model: "claude-opus-5.5", effort: "max" }), ); - // Another CLI keeps its own pick and does not overwrite CodeBuddy's. act(() => first.result.current.newChat()); act(() => first.result.current.setExecutor("codex")); await waitFor(() => expect(first.result.current.models.map((m) => m.id)).toEqual(["default", "gpt-5"])); act(() => { first.result.current.setModel("gpt-5"); }); + act(() => first.result.current.setEffort("high")); act(() => first.result.current.setExecutor("codebuddy")); expect(first.result.current.model).toBe("claude-opus-5.5"); + expect(first.result.current.effort).toBe("max"); act(() => first.result.current.setExecutor("codex")); + act(() => first.result.current.newChat()); expect(first.result.current.model).toBe("gpt-5"); - act(() => first.result.current.setExecutor("codebuddy")); + expect(first.result.current.effort).toBe("high"); first.unmount(); - // Reloading the home page starts on the remembered model too. + personalCodexCatalog = true; history.replaceState({}, "", "/"); const reloaded = renderHook(() => useChat()); await waitFor(() => expect(reloaded.result.current.ready).toBe(true)); + expect(reloaded.result.current.executor).toBe("codex"); + expect(reloaded.result.current.model).toBe("codex:gpt-5"); + expect(reloaded.result.current.effort).toBe("high"); + act(() => reloaded.result.current.setExecutor("codebuddy")); expect(reloaded.result.current.model).toBe("claude-opus-5.5"); + expect(reloaded.result.current.effort).toBe("max"); + act(() => reloaded.result.current.setModel("m1")); + expect(reloaded.result.current.effort).toBe("low"); + act(() => reloaded.result.current.setModel("claude-opus-5.5")); + expect(reloaded.result.current.effort).toBe("max"); + act(() => reloaded.result.current.setEffort("")); // A remembered model that is no longer offered falls back to the list. - localStorage.setItem("loopforge.assistant.model.codebuddy", "retired-model"); + writeComposerPreference("model.codebuddy", "retired-model"); act(() => reloaded.result.current.newChat()); - expect(reloaded.result.current.model).toBe("default"); + expect(reloaded.result.current.model).toBe("m1"); + expect(reloaded.result.current.effort).toBe("low"); + act(() => reloaded.result.current.setModel("claude-opus-5.5")); + expect(reloaded.result.current.effort).toBe(""); } finally { apiMock.mockImplementation(original); - localStorage.removeItem("loopforge.assistant.executor"); } }); diff --git a/chatbot/tests/message-queue-hook.test.jsx b/chatbot/tests/message-queue-hook.test.jsx index c04fee90..6eda8129 100644 --- a/chatbot/tests/message-queue-hook.test.jsx +++ b/chatbot/tests/message-queue-hook.test.jsx @@ -103,7 +103,7 @@ describe("sending while a reply runs", () => { const { result, unmount } = await openRunning(); let ok; await act(async () => { - ok = await result.current.send({ text: "应该看 CI 而不是 lint" }); + ok = await result.current.send({ text: "应该看 CI 而不是 lint", approvalMode: "auto" }); }); expect(ok).toBe(true); expect(result.current.error).toBe(""); @@ -115,6 +115,7 @@ describe("sending while a reply runs", () => { await finish("failed"); await waitFor(() => expect(generationPosts()).toHaveLength(1)); const body = generationPosts()[0][1]; + expect(body).toMatchObject({ approvalMode: "auto", approvalUserId: 1 }); // The queued question follows the stored unfinished reply. expect(body.messages.map((m) => m.role)).toEqual(["user", "assistant", "user"]); expect(body.messages[1].unfinished).toBe("生成已停止"); @@ -152,7 +153,8 @@ describe("sending while a reply runs", () => { }); expect(steerPosts()).toHaveLength(1); expect(steerPosts()[0][1]).toMatchObject({ - content: "立即看这里", model: "m1", attachments, references, approvalMode: "assisted", + content: "立即看这里", model: "m1", attachments, references, + approvalMode: "assisted", approvalUserId: 1, }); expect(result.current.steering?.text).toBe("立即看这里"); expect(result.current.queue.map((m) => m.text)).toEqual(["下一条补充"]); diff --git a/chatbot/tests/message-queue.test.mjs b/chatbot/tests/message-queue.test.mjs index 80e3a246..b2684d13 100644 --- a/chatbot/tests/message-queue.test.mjs +++ b/chatbot/tests/message-queue.test.mjs @@ -4,6 +4,7 @@ import { enqueue, holdQueue, nextQueued, + queuedApproval, queuedMessage, readQueues, removeQueued, @@ -22,6 +23,12 @@ function memoryStorage() { } describe("message queue", () => { + test("keeps the account that made the approval choice and confirms older queue items", () => { + const message = queuedMessage({ text: "继续", approvalMode: "auto", approvalUserId: 1 }); + expect(queuedApproval(message, 2)).toEqual({ approvalMode: "auto", approvalUserId: 1 }); + expect(queuedApproval({ approvalMode: "auto" }, 2)).toEqual({ approvalMode: "confirm", approvalUserId: 2 }); + }); + test("keeps messages per conversation in order", () => { const a = queuedMessage({ text: " 先看 CI " }); const b = queuedMessage({ text: "再看日志", attachments: [{ id: "f1" }] }); diff --git a/chatbot/tests/pickers-and-activity.test.jsx b/chatbot/tests/pickers-and-activity.test.jsx index 0e24b80f..a27fbdc3 100644 --- a/chatbot/tests/pickers-and-activity.test.jsx +++ b/chatbot/tests/pickers-and-activity.test.jsx @@ -82,22 +82,24 @@ describe("pickers and activity", () => { }); test("remembers the approval mode and ignores unknown values", () => { - expect(storedApprovalMode()).toBe("confirm"); - rememberApprovalMode("assisted"); - expect(storedApprovalMode()).toBe("assisted"); - rememberApprovalMode("bogus"); - expect(storedApprovalMode()).toBe("assisted"); - localStorage.setItem("loopforge.assistant.approvalMode", "bogus"); - expect(storedApprovalMode()).toBe("confirm"); + expect(storedApprovalMode(1)).toBe("confirm"); + rememberApprovalMode(1, "assisted"); + expect(storedApprovalMode(1)).toBe("assisted"); + rememberApprovalMode(1, "bogus"); + expect(storedApprovalMode(1)).toBe("assisted"); + localStorage.setItem("loopforge.assistant.approvalMode.1", "bogus"); + expect(storedApprovalMode(1)).toBe("confirm"); }); - test("never remembers full permission across reloads", () => { - rememberApprovalMode("assisted"); - rememberApprovalMode("auto"); - expect(storedApprovalMode()).toBe("assisted"); - // A value stored by an older build does not come back either. + test("only restores full permissions for the account that selected them", () => { localStorage.setItem("loopforge.assistant.approvalMode", "auto"); - expect(storedApprovalMode()).toBe("confirm"); + expect(storedApprovalMode(1)).toBe("confirm"); + rememberApprovalMode(1, "auto"); + expect(storedApprovalMode(1)).toBe("auto"); + expect(storedApprovalMode(2)).toBe("confirm"); + expect(storedApprovalMode(null)).toBe("confirm"); + rememberApprovalMode(1, "confirm"); + expect(storedApprovalMode(1)).toBe("confirm"); }); test("expands detailed activity including failures", async () => { diff --git a/design.md b/design.md index cb0344bb..c34e2634 100644 --- a/design.md +++ b/design.md @@ -56,6 +56,7 @@ - 凭据定位等跳转应在目标页面挂载后聚焦;用户已切换到其他页面时取消旧的焦点请求。 - 对话失败如果是因为当前服务器没有安装所选 CLI,错误写明「没有安装」,并在同一条提示里提供「安装」按钮。点击后按钮变为「自动安装」,刷新页面也保持这个状态,直到安装结束。安装完成后,原来的「没有安装」或「AI 助手生成失败」改为可以重新发送。按钮只安装平台允许的 CLI。安装在服务器上继续,关闭页面也不会中断。 - 个人配置里,Codex、Claude Code,以及带协议的 Agent CLI,可以维护模型目录:获取可用模型,勾选、排序、修改显示名。目录为空时,未填 Base URL 或协议是 Gemini 官方 API 则使用该 CLI 自带模型,填写了 Base URL 则跟随服务公布的模型。保存过的目录在官方服务和自定义地址上都生效。Cursor CLI 只使用官方模型。 +- 模型目录每行可展开「思考强度」,选择自动或自定义可选档位;自定义全部取消表示仅使用模型默认强度。获取模型时保留服务提供的档位,缺失时沿用已有 CLI 目录能力;用户自定义优先,Cursor 的档位与模型 ID 映射独立保留。相同 ID 补充的能力信息也需要保存。 - 主题入口在左下角头像卡片(`#userProfilePopover`)的「外观」行:「浅色 / 深色 / 跟随系统」三段开关(`.upp-theme-seg`,`role="radiogroup"`),访客与登录卡片都显示,选项文案与 chatbot 设置一致,偏好写入共享键 `agentflow.theme.v3`。三段前分别是 Harness `sun`、`moon`、`monitor` 图标,悬停、聚焦或点击时播放一次动效;选中滑块由 SlidingTabs 驱动,卡片展开时立即对准当前选项。切换时页面颜色约 280ms 渐变,减少动态效果时直接切换。 - 顶栏 `#themeToggle` 只在宽度 ≤1000px(图标栏隐藏)时显示,作为浅色、深色一键切换的兜底;宽屏下仅隐藏,DOM 节点保留给 AI 助手按钮定位。 @@ -88,6 +89,9 @@ - 模型触发器和选项显示品牌图标;未知品牌使用通用图标,不影响模型可选性。 - 品牌图标保留原始颜色及明暗主题变体,悬停时只在图形内部播放短暂光效。 - 模型列表保留服务端顺序、真实模型 ID、名称、系列及倍率信息,品牌推断只用于展示。 +- 对话输入区在当前浏览器记住最近选用的 CLI、每个 CLI 的模型和每个模型的思考强度。操作确认方式在同一浏览器内按登录账号保存,新账号默认使用「请求批准」。刷新或开始新对话时优先恢复上次选用的可用模型,否则选择列表第一项。打开已有对话时模型和强度以会话记录为准。CodeBuddy 的菜单只列出具体模型,列表为空时使用 CLI 默认模型。 +- 发送消息时校验审批选择对应的账号与当前登录账号。切换账号后,从旧标签页发送会提示刷新。 +- 保存 CLI 凭据后自动更新对话输入区的模型和强度。Cursor 后台探测完成时更新账号模型目录。 - 搜索支持名称、ID 和系列;选中态使用固定勾选位置,搜索为空时显示明确结果。 - 浮层根据可用空间向上或向下展开,限制在可视视口内;手机键盘改变 `visualViewport` 时重新定位。 - Escape 关闭并返回触发器,Tab 接回输入区的正常顺序;点击外部关闭时不抢焦点。 diff --git a/docs/deployment.md b/docs/deployment.md index fad11a04..e441ba49 100644 --- a/docs/deployment.md +++ b/docs/deployment.md @@ -10,7 +10,7 @@ For a native installation, use Node.js 22+ on the API host and install the CLIs npm install --global @tencent-ai/codebuddy-code codebuddy --version # Optional executors for Main Bot and SOP roles: -npm install --global @openai/codex@0.162.1 @anthropic-ai/claude-code@2.1.300 +npm install --global @openai/codex@0.162.1 @anthropic-ai/claude-code@2.1.326 ``` Ensure the service can find `codebuddy` and Node.js on its PATH, or set `AGENTFLOW_CODEBUDDY_BIN` to the installed executable. Configure the model key and its network environment in personal settings. The no-key script demo does not need the CLI. The production Docker image includes CodeBuddy, Codex and DeepSeek Harness (`dsh`). Claude Code and the other agent CLIs are not bundled. When a conversation fails because one is missing, the error says it is not installed and offers an install button. That downloads the pinned `@anthropic-ai/claude-code` release, or the public package for OpenCode, Kimi Code and Pi, into `$LOOPFORGE_DATA_DIR/cli`; an installation outside the image gets the pinned `@deepseek-ai/dsh` release the same way. Cursor CLI and Antigravity use their providers' installers and land in the service user's `~/.local/bin`. Closing or reloading the browser does not stop an install that has already started, and the button stays on “自动安装” until it finishes. A finished install replaces the earlier generation failure with a prompt to send the message again. The next message reuses the installation. An SOP sandbox of the published image that selects Claude Code, OpenCode, Kimi Code or Pi installs the same package on first use and shares it with conversations. Cursor CLI and Antigravity are not installed into SOP sandboxes. Connected runners keep their own CLI requirements. diff --git a/e2e/public/executors.spec.mjs b/e2e/public/executors.spec.mjs index 33272037..23bb0d37 100644 --- a/e2e/public/executors.spec.mjs +++ b/e2e/public/executors.spec.mjs @@ -16,17 +16,25 @@ for(const viewport of [{width:1280,height:400},{width:390,height:600}]){ await signInUnconfigured(page); await openHome(page); const frame=chat(page); - for(const [button,menu,name] of [[executorButton(frame),'.permission-menu[role="menu"]','执行器'],[frame.getByRole('button',{name:'选择模型',exact:true}),'.grasp-model-menu','选择模型']]){ - await button.click(); - const open=frame.locator(menu); - await expect(open).toBeVisible(); - await expect.poll(()=>fitsFrame(open),{message:`${name} menu at ${viewport.width}x${viewport.height}`}).toBe(true); - await page.keyboard.press('Escape'); - await expect(open).toBeHidden(); - } + const cli=executorButton(frame),cliMenu=frame.locator('.permission-menu[role="menu"]'); + const fits=async menu=>{ + await expect(menu).toBeVisible(); + await expect.poll(()=>fitsFrame(menu)).toBe(true); + }; + await cli.click(); + await fits(cliMenu); + await page.keyboard.press('Escape'); + await expect(cliMenu).toBeHidden(); + // Codex offers a built-in catalog on installations without any configured models. + await cli.click(); + await cliMenu.getByRole('menuitemradio',{name:'Codex',exact:true}).click(); + const model=frame.getByRole('button',{name:'选择模型',exact:true}),modelMenu=frame.locator('.grasp-model-menu'); + await model.click(); + await fits(modelMenu); + await page.keyboard.press('Escape'); + await expect(modelMenu).toBeHidden(); // Long catalogs are searched inside the menu. - await frame.getByRole('button',{name:'选择模型'}).click(); - const modelMenu=frame.locator('.grasp-model-menu'); + await model.click(); const search=modelMenu.locator('input[aria-label="搜索模型"]'); await expect(search).toBeVisible(); const first=(await modelMenu.locator('[role^="menuitem"]').first().innerText()).split('\n')[0].trim(); @@ -63,7 +71,11 @@ test('executor selection has independent models, credentials and a persistent ch const response=await page.request.post('/api/assistant/generations',{data:{executor:id,model:'default',projectId:'welcome',messages:[{role:'user',content:'No key: reject before executing.'}]}}); expect(response.status()).toBe(400);expect((await response.json()).error).toContain(label); const models=await page.request.get('/api/assistant/models?executor='+id); - expect(models.ok()).toBe(true);expect((await models.json()).models.length).toBeGreaterThan(0); + expect(models.ok()).toBe(true); + const catalog=await models.json(); + expect(Array.isArray(catalog.models)).toBe(true); + if(id==='codebuddy')expect(catalog.default).toBeTruthy(); + else expect(catalog.models.length).toBeGreaterThan(0); } await choose('Codex'); await page.reload(); diff --git a/internal/domain/agent.go b/internal/domain/agent.go index bca6dcb5..8e6d1529 100644 --- a/internal/domain/agent.go +++ b/internal/domain/agent.go @@ -7,8 +7,8 @@ import ( "time" ) -// ExecutorEffortLevels lists the reasoning-effort levels each built-in CLI -// accepts on its command line. An empty level keeps the CLI default. +// ExecutorEffortLevels lists the reasoning-effort values each built-in CLI +// can apply. An empty level uses the CLI default. var ExecutorEffortLevels = map[string][]string{ "claude-code": {"low", "medium", "high", "xhigh", "max", "ultracode"}, "codebuddy": {"low", "medium", "high", "xhigh", "max", "ultracode"}, @@ -16,7 +16,10 @@ var ExecutorEffortLevels = map[string][]string{ "opencode": {"low", "medium", "high"}, "cursor": {"none", "minimal", "low", "medium", "high", "xhigh", "max", "ultra"}, "agy": {"low", "medium", "high", "xhigh", "max"}, + "kimi": {"none", "thinking", "low", "high", "max"}, "pi": {"minimal", "low", "medium", "high", "xhigh", "max"}, + "dsh": {"none", "low", "high", "max"}, + "zcode": {"low", "high", "max"}, } // AgentCLIExecutorKinds are the agent CLIs beyond CodeBuddy / Codex / Claude diff --git a/internal/domain/agent_kind_effort_test.go b/internal/domain/agent_kind_effort_test.go index f9f75b8f..bf40f64e 100644 --- a/internal/domain/agent_kind_effort_test.go +++ b/internal/domain/agent_kind_effort_test.go @@ -30,7 +30,14 @@ func TestValidateAgentEffortChecksPublishedLevels(t *testing.T) { {effort: "ultra", kind: "codex"}, {effort: " max ", kind: "claude-code"}, {effort: "minimal", kind: " pi "}, - {effort: "high", kind: "kimi", wantErr: "当前 CLI 不支持设置思考强度"}, + {effort: "high", kind: "kimi"}, + {effort: "thinking", kind: "kimi"}, + {effort: "none", kind: "kimi"}, + {effort: "none", kind: "dsh"}, + {effort: "max", kind: "zcode"}, + {effort: "medium", kind: "kimi", wantErr: "思考强度取值不被当前 CLI 支持"}, + {effort: "medium", kind: "dsh", wantErr: "思考强度取值不被当前 CLI 支持"}, + {effort: "none", kind: "zcode", wantErr: "思考强度取值不被当前 CLI 支持"}, {effort: "high", kind: "unknown", wantErr: "当前 CLI 不支持设置思考强度"}, {effort: "ultracode", kind: "codex", wantErr: "思考强度取值不被当前 CLI 支持"}, {effort: "HIGH", kind: "codex", wantErr: "思考强度取值不被当前 CLI 支持"}, diff --git a/internal/domain/user.go b/internal/domain/user.go index 0b417e56..d8f38f7f 100644 --- a/internal/domain/user.go +++ b/internal/domain/user.go @@ -203,6 +203,10 @@ type CLIModel struct { Name string `json:"name,omitempty"` Efforts map[string]string `json:"efforts,omitempty"` DefaultEffort string `json:"defaultEffort,omitempty"` + // ReasoningEfforts is the user's override. An empty list permits only the native default. + ReasoningEfforts *[]string `json:"reasoningEfforts,omitempty"` + // DiscoveredReasoningEfforts is supplied by model discovery and used in automatic mode. + DiscoveredReasoningEfforts *[]string `json:"discoveredReasoningEfforts,omitempty"` } // AgentCLICredential is the personal configuration of one agent CLI. An diff --git a/internal/httpapi/assistant.go b/internal/httpapi/assistant.go index 5af36283..af9bbea3 100644 --- a/internal/httpapi/assistant.go +++ b/internal/httpapi/assistant.go @@ -61,6 +61,7 @@ type assistantRequest struct { Effort string `json:"effort"` Executor string `json:"executor,omitempty"` ApprovalMode string `json:"approvalMode"` + ApprovalUserID int64 `json:"approvalUserId,omitempty"` RunID string `json:"runId,omitempty"` SpaceID string `json:"spaceId,omitempty"` workspaceToken string @@ -115,6 +116,9 @@ func (h *assistantHandlers) chat(c *gin.Context) { if !decodeLimit(c, &req, assistantRequestBodyLimit) { return } + if !requireAssistantApprovalUser(c, current.ID, req.ApprovalUserID) { + return + } req.Executor = assistantExecutor(req.Executor) req.Model = assistantRequestModel(req.Executor, req.Model) if len(req.Messages) == 0 { @@ -222,7 +226,7 @@ func (h *assistantHandlers) chat(c *gin.Context) { } cliEngine, _ = assistantModelEngine(req.Model) req.Executor = cliEngine - if err := validateAssistantEffort(user, req.Model, req.Effort); err != nil { + if err := validateAssistantEffort(c.Request.Context(), user, req.Model, req.Effort); err != nil { c.JSON(400, gin.H{"error": err.Error()}) return } diff --git a/internal/httpapi/assistant_agentcli_bin_test.go b/internal/httpapi/assistant_agentcli_bin_test.go index 0c4060ea..7ecee340 100644 --- a/internal/httpapi/assistant_agentcli_bin_test.go +++ b/internal/httpapi/assistant_agentcli_bin_test.go @@ -78,10 +78,10 @@ func TestCursorModelFamilies(t *testing.T) { if !reflect.DeepEqual(got, want) { t.Fatalf("families=%+v", got) } - if efforts := assistantModelEfforts(domain.User{}, "cursor:gpt-5.6-sol"); !slices.Equal(efforts, []string{"none", "low", "medium", "high", "xhigh", "max"}) { + if efforts := assistantModelEfforts(context.Background(), domain.User{}, "cursor:gpt-5.6-sol"); !slices.Equal(efforts, []string{"none", "low", "medium", "high", "xhigh", "max"}) { t.Fatalf("cursor family efforts: %v", efforts) } - if err := validateAssistantEffort(domain.User{}, "cursor:kimi-k3", "medium"); err == nil { + if err := validateAssistantEffort(context.Background(), domain.User{}, "cursor:kimi-k3", "medium"); err == nil { t.Fatal("chat must offer only the family's own efforts") } } @@ -95,16 +95,16 @@ func TestCachedCursorModelsRefreshInBackground(t *testing.T) { } t.Cleanup(func() { assistantCursorModels = previous }) key := "crsr_test_" + t.Name() - if got := cachedCursorModels(key); got != nil { - t.Fatalf("cold cache must answer at once with the builtin catalog: %+v", got) + if got, pending := cachedCursorModels(key); got != nil || !pending { + t.Fatalf("cold cache must answer at once with discovery pending: %+v, pending=%v", got, pending) } if got := <-calls; got != key { t.Fatalf("discovery used key %q", got) } deadline := time.Now().Add(2 * time.Second) for { - got := cachedCursorModels(key) - if len(got) == 1 && got[0].ID == "claude-opus-5-5-high" { + got, pending := cachedCursorModels(key) + if !pending && len(got) == 1 && got[0].ID == "claude-opus-5-5-high" { break } if time.Now().After(deadline) { diff --git a/internal/httpapi/assistant_approval.go b/internal/httpapi/assistant_approval.go index 4c9d9f5a..8164fd27 100644 --- a/internal/httpapi/assistant_approval.go +++ b/internal/httpapi/assistant_approval.go @@ -4,11 +4,14 @@ import ( "context" "encoding/json" "fmt" + "net/http" "regexp" "strings" "agentflow/internal/domain" "agentflow/internal/repo" + + "github.com/gin-gonic/gin" ) // 平台写入与沙箱外写入(外部 MCP、git push 等命令)的审批模式;工作区沙箱内的开发操作默认执行。 @@ -34,6 +37,16 @@ func normalizeAssistantApprovalMode(mode string) string { } } +// Check the account attached by the browser against the account that +// authenticated this request. Existing API clients may omit the field. +func requireAssistantApprovalUser(c *gin.Context, currentID, approvalUserID int64) bool { + if approvalUserID != 0 && approvalUserID != currentID { + c.JSON(http.StatusConflict, gin.H{"error": "登录账号已变更,请刷新页面后重试"}) + return false + } + return true +} + var assistantRiskyToolPattern = regexp.MustCompile(`(_delete|_publish|_dispatch|_start|_cancel|_reject|_run|_restore|_pass|_retry|_pause|_resume|_reopen)$`) // assistantRiskyToolName 判断平台写入工具是否属于危险操作(删除/发布/ diff --git a/internal/httpapi/assistant_approval_test.go b/internal/httpapi/assistant_approval_test.go index bc559e07..372d9870 100644 --- a/internal/httpapi/assistant_approval_test.go +++ b/internal/httpapi/assistant_approval_test.go @@ -1,13 +1,17 @@ package httpapi import ( + "bytes" "context" "encoding/json" + "net/http" "strings" "testing" "agentflow/internal/domain" "agentflow/internal/repo" + + "github.com/gin-gonic/gin" ) func TestNormalizeAssistantApprovalMode(t *testing.T) { @@ -25,6 +29,39 @@ func TestNormalizeAssistantApprovalMode(t *testing.T) { } } +func TestAssistantRejectsApprovalFromAnotherAccount(t *testing.T) { + h, first := workbenchFixture(t) + if err := h.repo.PutUser(domain.User{LoginName: "different-account", CodeBuddyAPIKey: "test-only"}); err != nil { + t.Fatal(err) + } + current, _ := h.repo.UserByLoginName("different-account") + for _, route := range []struct { + method, path string + handler gin.HandlerFunc + }{ + {http.MethodPut, "/api/assistant/sessions/stale-tab", h.session}, + {http.MethodPost, "/api/assistant/generations", h.startGeneration}, + {http.MethodPost, "/api/assistant/sessions/stale-tab/generation/steer", h.steerGeneration}, + {http.MethodPost, "/api/assistant/chat", h.chat}, + } { + t.Run(route.method+" "+route.path, func(t *testing.T) { + raw, _ := json.Marshal(map[string]any{ + "sessionId": "stale-tab", "approvalMode": "auto", "approvalUserId": first.ID, + "content": "Continue", "messages": []assistantMessage{{Role: "user", Content: "Continue"}}, + }) + c, rec := workbenchContext(current, route.method, route.path, bytes.NewBuffer(raw)) + c.Params = gin.Params{{Key: "id", Value: "stale-tab"}} + route.handler(c) + if rec.Code != http.StatusConflict || !strings.Contains(rec.Body.String(), "登录账号已变更") { + t.Fatalf("stale approval under a new account: %d %s", rec.Code, rec.Body) + } + }) + } + if _, ok := h.repo.AssistantSession(current.ID, "stale-tab"); ok { + t.Fatal("the stale tab created a session for the newly authenticated account") + } +} + func TestAssistantExecModeForApproval(t *testing.T) { for _, tc := range []struct { mode string diff --git a/internal/httpapi/assistant_effort.go b/internal/httpapi/assistant_effort.go index c712878e..339471cd 100644 --- a/internal/httpapi/assistant_effort.go +++ b/internal/httpapi/assistant_effort.go @@ -1,6 +1,7 @@ package httpapi import ( + "context" "fmt" "slices" "strings" @@ -9,29 +10,89 @@ import ( "agentflow/internal/runtime" ) -// Mirrors CodeBuddy Code 2.180.0's product reasoning.supportedEfforts and -// catalog effort capability. Unknown models use the conservative low/medium/high -// fallback: the CLI clamps xhigh/max to high when no capability is known. -// Keep this snapshot alongside the fixed chat model catalog when updating it. -// Empty effort always leaves the model's native default untouched. -// -// user's saved Cursor catalog resolves the Cursor families folded out of -// their CLI's own model list (runtime.FoldCursorCLIModels). -func assistantModelEfforts(user domain.User, model string) []string { - return assistantModelEffortsIn(model, assistantCursorCatalog(user)) +// assistantModelEfforts resolves the same capabilities as the picker for +// custom gateways. Cursor uses the saved family mapping that drives its CLI. +func assistantModelEfforts(ctx context.Context, user domain.User, model string) []string { + engine, _ := assistantModelEngine(model) + if engine == assistantEngineCodeBuddy || engine == runtime.AgentCLICursor { + return assistantModelEffortsIn(model, assistantEngineCustomModels(user, engine)) + } + _, baseURL := assistantEngineCredential(user, engine) + if engine == runtime.AgentCLIKimi && baseURL != "" { + return nil + } + models, _ := assistantDiscoveredModelsFor(ctx, user, engine) + return assistantModelEffortsIn(model, cliModelsOf(models)) +} + +func assistantModelEffortsForBaseURL(model string, catalog []domain.CLIModel, baseURL string) []string { + if baseURL != "" && strings.HasPrefix(strings.TrimSpace(model), runtime.AgentCLIKimi+":") { + return nil + } + return assistantModelEffortsIn(model, catalog) } -// assistantModelEffortsIn is assistantModelEfforts with the Cursor catalog -// given directly. -func assistantModelEffortsIn(model string, cursor []domain.CLIModel) []string { - if engine, id := assistantModelEngine(model); engine == runtime.AgentCLICursor { - return runtime.CursorModelEfforts(id, cursor) +// assistantModelEffortsIn prefers configured levels, then discovered levels, +// then the built-in model or CLI capabilities. +// The CodeBuddy fallback mirrors CodeBuddy Code 2.180.0's model capabilities. +// Unknown CodeBuddy, Codex and Claude Code models fall back to low/medium/high. +func assistantModelEffortsIn(model string, catalog []domain.CLIModel) []string { + engine, id := assistantModelEngine(model) + supported := domain.ExecutorEffortLevels[engine] + builtins := builtinExecutorModels[engine] + builtin := slices.IndexFunc(builtins, func(model builtinExecutorModel) bool { return model.id == id }) + if engine == runtime.AgentCLICursor { + supported = runtime.CursorModelEfforts(id, catalog) } else if engine == runtime.AgentCLIAntigravity { - return runtime.AgyModelEfforts(id) - } else if runtime.IsAgentCLI(engine) { - return domain.ExecutorEffortLevels[engine] + supported = runtime.AgyModelEfforts(id) + } else if engine == runtime.AgentCLIKimi && builtin >= 0 { + supported = builtins[builtin].efforts + } + for _, item := range catalog { + if item.ID != id { + continue + } + configured := item.ReasoningEfforts + if configured == nil { + configured = item.DiscoveredReasoningEfforts + } + if configured == nil { + continue + } + advertised := *configured + if engine == runtime.AgentCLIKimi || engine == runtime.AgentCLIDeepSeek { + advertised = slices.Clone(advertised) + for i, level := range advertised { + switch level { + case "off": + advertised[i] = "none" + case "on": + advertised[i] = "thinking" + } + } + } + levels := []string{} + for _, level := range supported { + if slices.Contains(advertised, level) { + levels = append(levels, level) + } + } + return levels + } + if builtin >= 0 { + return builtins[builtin].efforts + } + if runtime.IsAgentCLI(engine) { + if len(builtins) > 0 { + return nil + } + return supported + } + if engine == assistantEngineCodex { + return []string{"low", "medium", "high"} } - switch model { + // Keep the legacy fallback for CodeBuddy and custom Claude gateways. + switch id { case "hy4std", "hy4-preview-ioa": return []string{"high"} case "glm-5.3-ioa", "glm-5.3-flash-ioa", "glm-5.3-flashx", "kimi-k3.1-pro", "kimi-k2.8-preview": @@ -47,12 +108,15 @@ func assistantModelEffortsIn(model string, cursor []domain.CLIModel) []string { } } -func validateAssistantEffort(user domain.User, model, effort string) error { +func validateAssistantEffort(ctx context.Context, user domain.User, model, effort string) error { if effort == "" { return nil } - efforts := assistantModelEfforts(user, model) + efforts := assistantModelEfforts(ctx, user, model) if !slices.Contains(efforts, effort) { + if len(efforts) == 0 { + return fmt.Errorf("模型 %s 仅支持默认思考强度", model) + } return fmt.Errorf("模型 %s 不支持思考强度 %q,可选:默认、%s", model, effort, strings.Join(efforts, "、")) } return nil diff --git a/internal/httpapi/assistant_effort_test.go b/internal/httpapi/assistant_effort_test.go index c7d33c21..f846f962 100644 --- a/internal/httpapi/assistant_effort_test.go +++ b/internal/httpapi/assistant_effort_test.go @@ -1,12 +1,16 @@ package httpapi import ( - "agentflow/internal/domain" "bytes" "encoding/json" - "github.com/gin-gonic/gin" + "net/http" "strings" "testing" + + "agentflow/internal/domain" + "agentflow/internal/repo" + + "github.com/gin-gonic/gin" ) func TestAssistantEffortReachesNativeCLIAndPersists(t *testing.T) { @@ -64,6 +68,58 @@ func TestAssistantEffortSessionPutPreservesAbsentValue(t *testing.T) { } } +func TestAssistantEffortSessionRenameAfterCatalogChange(t *testing.T) { + const sessionID = "effort-historical" + const model = "deepseek-v4.1-flash" + t.Setenv("AGENTFLOW_ASSISTANT_MODELS", model) + h, user := workbenchFixture(t) + nativeCLIRecorder(t) + stubCLIModelDiscovery(t) + if err := h.repo.PutAssistantSession(user.ID, domain.AssistantSession{ + ID: sessionID, Title: "original", Executor: "codebuddy", Model: model, Effort: "low", + }); err != nil { + t.Fatal(err) + } + defaultOnly := []string{} + models := []domain.CLIModel{{ID: model, ReasoningEfforts: &defaultOnly}} + user, err := h.repo.UpdateUserCredentials(user.ID, repo.UserCredentialUpdate{CodeBuddyModels: &models}) + if err != nil { + t.Fatal(err) + } + + for _, tc := range []struct { + name string + body string + status int + title string + }{ + {"title only", `{"title":"renamed"}`, http.StatusOK, "renamed"}, + {"unchanged settings", `{"title":"new title","executor":"codebuddy","model":"deepseek-v4.1-flash","effort":"low"}`, http.StatusOK, "new title"}, + {"effort change", `{"effort":"high"}`, http.StatusBadRequest, "new title"}, + {"model change", `{"model":"hy4-preview-ioa"}`, http.StatusBadRequest, "new title"}, + } { + t.Run(tc.name, func(t *testing.T) { + c, rec := workbenchContext(user, http.MethodPut, "/api/assistant/sessions/"+sessionID, bytes.NewBufferString(tc.body)) + c.Params = gin.Params{{Key: "id", Value: sessionID}} + h.session(c) + saved, ok := h.repo.AssistantSession(user.ID, sessionID) + if rec.Code != tc.status || !ok || saved.Title != tc.title || saved.Model != model || saved.Effort != "low" { + t.Fatalf("status %d, session %s %s/%s, body %s", rec.Code, saved.Title, saved.Model, saved.Effort, rec.Body) + } + }) + } + + body, _ := json.Marshal(assistantRequest{ + SessionID: sessionID, Model: model, Effort: "low", + Messages: []assistantMessage{{Role: "user", Content: "继续"}}, + }) + c, rec := workbenchContext(user, http.MethodPost, "/api/assistant/generations", bytes.NewBuffer(body)) + h.startGeneration(c) + if rec.Code != http.StatusBadRequest || !strings.Contains(rec.Body.String(), "仅支持默认思考强度") { + t.Fatalf("generation with historical effort returned %d: %s", rec.Code, rec.Body) + } +} + func TestAssistantReviewCommandHasNoToolsOrPersistentSession(t *testing.T) { command, err := prepareAssistantCommand(assistantCommandOpts{Bin: "codebuddy", Config: "/tmp/empty-review-mcp.json", Prompt: "review data", Model: assistantChatDefaultModel, Effort: "high", ReviewOnly: true}) if err != nil { diff --git a/internal/httpapi/assistant_engine_models_test.go b/internal/httpapi/assistant_engine_models_test.go index c882f11a..af1af0d5 100644 --- a/internal/httpapi/assistant_engine_models_test.go +++ b/internal/httpapi/assistant_engine_models_test.go @@ -18,8 +18,9 @@ func TestAssistantEngineModelsKeepPersonalCatalogOnOfficialKey(t *testing.T) { return []assistantDiscoveredModel{{ID: "remote-model", Name: "Remote"}}, nil } ids := func(user domain.User) []string { + models, _ := assistantEngineModels(context.Background(), user, nil) var out []string - for _, model := range assistantEngineModels(context.Background(), user, nil) { + for _, model := range models { out = append(out, model.ID) } return out diff --git a/internal/httpapi/assistant_engines.go b/internal/httpapi/assistant_engines.go index b0d00d90..d51cf9ec 100644 --- a/internal/httpapi/assistant_engines.go +++ b/internal/httpapi/assistant_engines.go @@ -152,7 +152,7 @@ func assistantEngineBuiltinModels(engine string) []assistantDiscoveredModel { catalogKey = "claude" } var out []assistantDiscoveredModel - for _, model := range executorModelCatalogs()[catalogKey].Models { + for _, model := range executorModelCatalogs("")[catalogKey].Models { out = append(out, assistantDiscoveredModel{ID: model.ID, Name: model.Name}) } return out @@ -310,7 +310,8 @@ func assistantScheduledModel(id string, catalog assistantModelCatalog) string { // assistantDiscoveredModelsFor resolves the model list shared by the chat // composer and role editor. A saved catalog takes precedence over discovery. -func assistantDiscoveredModelsFor(ctx context.Context, user domain.User, engine string) []assistantDiscoveredModel { +// The second result reports whether Cursor is refreshing its automatic catalog. +func assistantDiscoveredModelsFor(ctx context.Context, user domain.User, engine string) ([]assistantDiscoveredModel, bool) { key, baseURL := assistantEngineCredential(user, engine) protocol := assistantEngineProtocol(user, engine) listed := assistantSavedModels(user, engine) @@ -318,33 +319,39 @@ func assistantDiscoveredModelsFor(ctx context.Context, user domain.User, engine listed = cachedAssistantEngineDiscovery(ctx, assistantEngineAnthropicModels(engine, protocol), key, baseURL) } if engine == runtime.AgentCLICursor && key != "" && !user.AgentCLICredentials[engine].ModelsCustomized { - if fresh := cachedCursorModels(key); len(fresh) > 0 { + fresh, pending := cachedCursorModels(key) + if len(fresh) > 0 { listed = fresh } + return listed, pending } - return listed + return listed, false } func assistantSavedModels(user domain.User, engine string) []assistantDiscoveredModel { var listed []assistantDiscoveredModel for _, model := range assistantEngineCustomModels(user, engine) { if assistantValidEngineModel(model.ID) { - listed = append(listed, assistantDiscoveredModel{ID: model.ID, Name: model.Name, Efforts: model.Efforts, DefaultEffort: model.DefaultEffort}) + listed = append(listed, assistantDiscoveredModel{ + ID: model.ID, Name: model.Name, Efforts: model.Efforts, DefaultEffort: model.DefaultEffort, + ReasoningEfforts: model.ReasoningEfforts, DiscoveredReasoningEfforts: model.DiscoveredReasoningEfforts, + }) } } return listed } -func assistantEngineModels(ctx context.Context, user domain.User, runnerLogin map[string]bool) []assistantModel { - var out []assistantModel +func assistantEngineModels(ctx context.Context, user domain.User, runnerLogin map[string]bool) (models []assistantModel, cursorModelsPending bool) { for _, engine := range assistantCLIEngines() { - key, _ := assistantEngineCredential(user, engine) + key, baseURL := assistantEngineCredential(user, engine) listed := assistantSavedModels(user, engine) if key == "" && !runnerLogin[engine] && len(listed) == 0 { continue } if !runnerLogin[engine] { - listed = assistantDiscoveredModelsFor(ctx, user, engine) + var pending bool + listed, pending = assistantDiscoveredModelsFor(ctx, user, engine) + cursorModelsPending = cursorModelsPending || pending } if len(listed) == 0 { listed = assistantEngineBuiltinModels(engine) @@ -356,16 +363,16 @@ func assistantEngineModels(ctx context.Context, user domain.User, runnerLogin ma name = model.ID } id := engine + ":" + model.ID - out = append(out, assistantModel{ + models = append(models, assistantModel{ ID: id, Name: name, Family: assistantModelFamily(model.ID), Engine: engine, - Efforts: assistantModelEffortsIn(id, catalog), + Efforts: assistantModelEffortsForBaseURL(id, catalog, baseURL), }) } } - return out + return models, cursorModelsPending } type assistantDiscoveredModel struct { @@ -373,8 +380,10 @@ type assistantDiscoveredModel struct { Name string // Efforts and DefaultEffort describe a Cursor family folded out of the // CLI's list (domain.CLIModel). - Efforts map[string]string - DefaultEffort string + Efforts map[string]string + DefaultEffort string + ReasoningEfforts *[]string + DiscoveredReasoningEfforts *[]string } type assistantDiscoveryEntry struct { @@ -395,9 +404,13 @@ var ( assistantEngineDiscover = discoverAssistantEngineModels ) -func cachedAssistantEngineDiscovery(ctx context.Context, anthropic bool, apiKey, baseURL string) []assistantDiscoveredModel { +func assistantDiscoveryKey(anthropic bool, apiKey, baseURL string) string { sum := sha256.Sum256([]byte(fmt.Sprint(anthropic) + "\x00" + baseURL + "\x00" + apiKey)) - cacheKey := hex.EncodeToString(sum[:]) + return hex.EncodeToString(sum[:]) +} + +func cachedAssistantEngineDiscovery(ctx context.Context, anthropic bool, apiKey, baseURL string) []assistantDiscoveredModel { + cacheKey := assistantDiscoveryKey(anthropic, apiKey, baseURL) now := time.Now() if cached, ok := assistantDiscoveryCache.Load(cacheKey); ok { if entry := cached.(assistantDiscoveryEntry); now.Before(entry.expires) { @@ -424,24 +437,26 @@ var ( // cachedCursorModels lists the models the key's Cursor plan offers. The CLI // takes seconds, so a miss answers with the built-in catalog (nil) while a -// background refresh fills the cache; a stale list is served meanwhile. -func cachedCursorModels(apiKey string) []assistantDiscoveredModel { +// background refresh fills the cache. The second result identifies snapshots +// that should be fetched again after discovery completes. +func cachedCursorModels(apiKey string) ([]assistantDiscoveredModel, bool) { sum := sha256.Sum256([]byte("cursor\x00" + apiKey)) cacheKey := hex.EncodeToString(sum[:]) var stale []assistantDiscoveredModel if cached, ok := assistantCursorModelsCache.Load(cacheKey); ok { entry := cached.(assistantDiscoveryEntry) if time.Now().Before(entry.expires) { - return entry.models + return entry.models, false } stale = entry.models } if _, running := assistantCursorModelsInflight.LoadOrStore(cacheKey, true); !running { + discover := assistantCursorModels go func() { defer assistantCursorModelsInflight.Delete(cacheKey) ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second) defer cancel() - found, err := assistantCursorModels(ctx, apiKey) + found, err := discover(ctx, apiKey) var models []assistantDiscoveredModel for _, model := range found { if assistantValidEngineModel(model.ID) && len(models) < assistantDiscoveryMaxModels { @@ -455,7 +470,7 @@ func cachedCursorModels(apiKey string) []assistantDiscoveredModel { assistantCursorModelsCache.Store(cacheKey, assistantDiscoveryEntry{models: models, expires: time.Now().Add(ttl)}) }() } - return stale + return stale, true } // saveCursorCatalog keeps the user's freshly listed Cursor models, folded @@ -471,7 +486,7 @@ func (h *assistantHandlers) saveCursorCatalog(user domain.User) { if key == "" { return } - fresh := cachedCursorModels(key) + fresh, _ := cachedCursorModels(key) if len(fresh) == 0 { return } @@ -551,9 +566,11 @@ func discoverAssistantEngineModels(ctx context.Context, anthropic bool, apiKey, } var payload struct { Data []struct { - ID string `json:"id"` - Name string `json:"name"` - DisplayName string `json:"display_name"` + ID string `json:"id"` + Name string `json:"name"` + DisplayName string `json:"display_name"` + ReasoningEfforts *[]string `json:"reasoningEfforts"` + SupportedReasoningEfforts *[]string `json:"supported_reasoning_efforts"` } `json:"data"` } if err := json.Unmarshal(body, &payload); err != nil { @@ -571,7 +588,11 @@ func discoverAssistantEngineModels(ctx context.Context, anthropic bool, apiKey, if name == "" { name = strings.TrimSpace(item.Name) } - out = append(out, assistantDiscoveredModel{ID: id, Name: truncateRunes(name, 64)}) + efforts := item.ReasoningEfforts + if efforts == nil { + efforts = item.SupportedReasoningEfforts + } + out = append(out, assistantDiscoveredModel{ID: id, Name: truncateRunes(name, 64), DiscoveredReasoningEfforts: efforts}) if len(out) == assistantDiscoveryMaxModels { break } diff --git a/internal/httpapi/assistant_generation.go b/internal/httpapi/assistant_generation.go index 1865e621..987dad72 100644 --- a/internal/httpapi/assistant_generation.go +++ b/internal/httpapi/assistant_generation.go @@ -34,6 +34,9 @@ func (h *assistantHandlers) startGeneration(c *gin.Context) { if !decodeLimit(c, &req, assistantRequestBodyLimit) { return } + if !requireAssistantApprovalUser(c, current.ID, req.ApprovalUserID) { + return + } req.Model = assistantRequestModel(req.Executor, req.Model) if len(req.Messages) == 0 || req.Messages[len(req.Messages)-1].Role != "user" { c.JSON(400, gin.H{"error": "请提供用户消息"}) @@ -119,7 +122,7 @@ func (h *assistantHandlers) startGeneration(c *gin.Context) { } cliEngine, _ = assistantModelEngine(req.Model) req.Executor = cliEngine - if err := validateAssistantEffort(user, req.Model, req.Effort); err != nil { + if err := validateAssistantEffort(c.Request.Context(), user, req.Model, req.Effort); err != nil { c.JSON(400, gin.H{"error": err.Error()}) return } @@ -460,6 +463,13 @@ func (h *assistantHandlers) steerGeneration(c *gin.Context) { c.JSON(401, gin.H{"error": "请先登录"}) return } + var body struct { + domain.AssistantSteer + ApprovalUserID int64 `json:"approvalUserId"` + } + if !decode(c, &body) || !requireAssistantApprovalUser(c, current.ID, body.ApprovalUserID) { + return + } stored, ok := h.repo.User(current.ID) if !ok { c.JSON(400, gin.H{"error": "请先配置个人 CodeBuddy API Key"}) @@ -470,10 +480,6 @@ func (h *assistantHandlers) steerGeneration(c *gin.Context) { c.JSON(403, gin.H{"error": err.Error()}) return } - var body domain.AssistantSteer - if !decode(c, &body) { - return - } content := strings.TrimSpace(body.Content) if content == "" && len(body.Attachments) == 0 { c.JSON(400, gin.H{"error": "请提供用户消息"}) @@ -539,7 +545,7 @@ func (h *assistantHandlers) steerGeneration(c *gin.Context) { c.JSON(409, gin.H{"error": refusal}) return } - if err := validateAssistantEffort(user, model, body.Effort); err != nil { + if err := validateAssistantEffort(c.Request.Context(), user, model, body.Effort); err != nil { c.JSON(400, gin.H{"error": err.Error()}) return } diff --git a/internal/httpapi/assistant_models.go b/internal/httpapi/assistant_models.go index 1dbc7bf7..60587067 100644 --- a/internal/httpapi/assistant_models.go +++ b/internal/httpapi/assistant_models.go @@ -12,6 +12,7 @@ import ( "time" "agentflow/internal/domain" + "agentflow/internal/runtime" "github.com/gin-gonic/gin" ) @@ -35,15 +36,17 @@ type assistantModel struct { } type assistantModelCatalog struct { - Default string `json:"default"` - Models []assistantModel `json:"models"` + Default string `json:"default"` + DefaultEfforts []string `json:"defaultEfforts,omitempty"` + Models []assistantModel `json:"models"` // SponsoredModels is set when users without their own CodeBuddy API Key // may still chat on the platform's shared key, limited to these models; // SponsoredModel is the first (the default). SponsoredModel string `json:"sponsoredModel,omitempty"` SponsoredModels []string `json:"sponsoredModels,omitempty"` // Engines lists the CLIs a signed-in user can switch between. - Engines []assistantEngineOption `json:"engines,omitempty"` + Engines []assistantEngineOption `json:"engines,omitempty"` + CursorModelsPending bool `json:"cursorModelsPending,omitempty"` } func prefetchAssistantModels() {} @@ -182,15 +185,33 @@ func cachedCLIModels(environment string) []string { // requires login plus a configured or sponsored CodeBuddy key. func (h *assistantHandlers) models(c *gin.Context) { if raw := strings.TrimSpace(c.Query("executor")); raw != "" && assistantExecutor(raw) != "codebuddy" { - id := assistantExecutor(raw) - catalog, ok := executorModelCatalogs()[id] + engine := assistantExecutor(raw) + id := engine + if engine == assistantEngineClaude { + id = "claude" + } else if id == "claude" { + engine = assistantEngineClaude + } + var kimiBaseURL string + if engine == runtime.AgentCLIKimi { + if current, signedIn := h.requestUser(c.Request, true); signedIn { + if stored, found := h.repo.User(current.ID); found { + _, kimiBaseURL = assistantEngineCredential(stored, engine) + } + } + } + catalog, ok := executorModelCatalogs(kimiBaseURL)[id] if !ok { c.JSON(400, gin.H{"error": "不支持的执行器"}) return } - result := assistantModelCatalog{Default: "default", Models: []assistantModel{{ID: "default", Name: "CLI default", Family: map[string]string{"codex": "gpt", "claude": "claude"}[id]}}} + result := assistantModelCatalog{Default: "default", Models: []assistantModel{{ID: "default", Name: "CLI default", Family: map[string]string{"codex": "gpt", "claude": "claude"}[id], Efforts: assistantModelEffortsIn(engine+":default", nil)}}} for _, item := range catalog.Models { - result.Models = append(result.Models, assistantModel{ID: item.ID, Name: item.Name, Family: map[string]string{"codex": "gpt", "claude": "claude"}[id]}) + levels := make([]string, 0, len(item.Efforts)) + for _, option := range item.Efforts { + levels = append(levels, option.Value) + } + result.Models = append(result.Models, assistantModel{ID: item.ID, Name: item.Name, Family: map[string]string{"codex": "gpt", "claude": "claude"}[id], Efforts: levels}) } c.JSON(200, result) return @@ -202,8 +223,9 @@ func (h *assistantHandlers) models(c *gin.Context) { if current, ok := h.requestUser(c.Request, true); ok { user, signedIn = h.repo.User(current.ID) } + usingSharedKey := signedIn && shared && strings.TrimSpace(user.CodeBuddyAPIKey) == "" switch { - case signedIn && shared && strings.TrimSpace(user.CodeBuddyAPIKey) == "": + case usingSharedKey: // Chat runs on the shared key, so list what that key's environment offers. refreshCLIModels(c.Request.Context(), key.InternetEnvironment, key.APIKey) catalog = restrictAssistantCatalog(assistantChatCatalog(c.Request.Context(), key.InternetEnvironment, key.Models), key.Models) @@ -213,20 +235,30 @@ func (h *assistantHandlers) models(c *gin.Context) { refreshCLIModels(c.Request.Context(), user.CodeBuddyInternetEnvironment, user.CodeBuddyAPIKey) catalog = assistantChatCatalog(c.Request.Context(), user.CodeBuddyInternetEnvironment, nil) if saved := codeBuddyPersonalModels(user); len(saved) > 0 { - catalog.Models = saved catalog.Default = saved[0].ID + if listed := selectableCodeBuddyModels(saved); len(listed) > 0 { + catalog.Models = listed + } } default: catalog = assistantChatCatalog(c.Request.Context(), "", nil) } + if catalog.Default == assistantDefaultModel { + models := user.CodeBuddyModels + if usingSharedKey { + models = nil + } + catalog.DefaultEfforts = assistantModelEffortsIn(catalog.Default, models) + } + catalog.Models = selectableCodeBuddyModels(catalog.Models) if shared { catalog.SponsoredModel, catalog.SponsoredModels = key.Models[0], key.Models } if signedIn { - catalog.Engines = assistantEngineOptions(user, shared && strings.TrimSpace(user.CodeBuddyAPIKey) == "") + c.Header("Cache-Control", "private, no-store") + catalog.Engines = assistantEngineOptions(user, usingSharedKey) var runnerLogin map[string]bool if c.Query("schedule") == "1" { - c.Header("Cache-Control", "private, no-store") runnerLogin = make(map[string]bool) for i := range catalog.Engines { engine := catalog.Engines[i].ID @@ -236,16 +268,29 @@ func (h *assistantHandlers) models(c *gin.Context) { } } } - catalog.Models = append(catalog.Models, assistantEngineModels(c.Request.Context(), user, runnerLogin)...) + models, pending := assistantEngineModels(c.Request.Context(), user, runnerLogin) + catalog.Models = append(catalog.Models, models...) + catalog.CursorModelsPending = pending h.saveCursorCatalog(user) } c.JSON(200, catalog) } -// withCodeBuddyPersonalModels also accepts the user's saved CodeBuddy models -// and, like the picker (/api/assistant/models), defaults to the first of them, -// so a request without a model (API, bots, scheduled runs) runs on the model -// the UI shows as default. The environment's models stay accepted. +// Keep the CLI default as an execution fallback and for older saved records. +// The model pickers offer the concrete models available to the caller. +func selectableCodeBuddyModels(models []assistantModel) []assistantModel { + listed := make([]assistantModel, 0, len(models)) + for _, model := range models { + if model.ID != assistantDefaultModel { + listed = append(listed, model) + } + } + return listed +} + +// withCodeBuddyPersonalModels accepts the user's saved CodeBuddy models and +// defaults requests without a model to the first of them. The environment's +// models stay accepted. func withCodeBuddyPersonalModels(catalog assistantModelCatalog, user domain.User) assistantModelCatalog { saved := codeBuddyPersonalModels(user) if len(saved) == 0 { @@ -266,7 +311,9 @@ func codeBuddyPersonalModels(user domain.User) []assistantModel { if name == "" { name = item.ID } - models = append(models, assistantListedModel(item.ID, name, "")) + model := assistantListedModel(item.ID, name, "") + model.Efforts = assistantModelEffortsIn(item.ID, user.CodeBuddyModels) + models = append(models, model) } return models } @@ -341,7 +388,7 @@ func assistantFixedCatalog() assistantModelCatalog { } func assistantListedModel(id, name, credits string) assistantModel { - return assistantModel{ID: id, Name: name, Family: assistantModelFamily(id), Credits: credits, Efforts: assistantModelEfforts(domain.User{}, id)} + return assistantModel{ID: id, Name: name, Family: assistantModelFamily(id), Credits: credits, Efforts: assistantModelEffortsIn(id, nil)} } func assistantModelFamily(id string) string { diff --git a/internal/httpapi/assistant_models_test.go b/internal/httpapi/assistant_models_test.go index 6426f6c1..8baf4b6c 100644 --- a/internal/httpapi/assistant_models_test.go +++ b/internal/httpapi/assistant_models_test.go @@ -1,11 +1,43 @@ package httpapi import ( + "bytes" "context" + "encoding/json" + "net/http" "reflect" "testing" ) +func TestAssistantModelPickerOmitsCodeBuddyCLIDefault(t *testing.T) { + t.Setenv("AGENTFLOW_ASSISTANT_MODEL", "") + t.Setenv("AGENTFLOW_ASSISTANT_MODELS", "model-a,model-b") + h, user := workbenchFixture(t) + stubCLIModelDiscovery(t) + + c, rec := workbenchContext(user, http.MethodGet, "/api/assistant/models", bytes.NewBuffer(nil)) + h.models(c) + var catalog assistantModelCatalog + if rec.Code != http.StatusOK || json.Unmarshal(rec.Body.Bytes(), &catalog) != nil { + t.Fatalf("model picker: %d %s", rec.Code, rec.Body) + } + if catalog.Default != assistantDefaultModel || len(catalog.DefaultEfforts) == 0 || + assistantCatalogLists(catalog, assistantDefaultModel) || !assistantCatalogLists(catalog, "model-b") { + t.Fatalf("model picker shows the wrong choices: %+v", catalog) + } + if got := assistantAllowedModel("", h.chatCatalog(context.Background(), user)); got != assistantDefaultModel { + t.Fatalf("an omitted request model should still use the CLI fallback, got %q", got) + } + t.Setenv("AGENTFLOW_ASSISTANT_MODELS", "") + c, rec = workbenchContext(user, http.MethodGet, "/api/assistant/models", bytes.NewBuffer(nil)) + h.models(c) + var unconfigured assistantModelCatalog + if rec.Code != http.StatusOK || json.Unmarshal(rec.Body.Bytes(), &unconfigured) != nil || + unconfigured.Default != assistantDefaultModel || unconfigured.Models == nil || len(unconfigured.Models) != 0 { + t.Fatalf("unconfigured picker: %d %s", rec.Code, rec.Body) + } +} + func TestCodeBuddyModelDiscoveryPerEnvironment(t *testing.T) { help := " --model Model for the current session. Currently supported: (gpt-5.5, glm-5.3, echo, custom-local:glm-5.2, kimi-k3)\n" if got := parseCodeBuddyModels(help); !reflect.DeepEqual(got, []string{"gpt-5.5", "glm-5.3", "kimi-k3"}) { diff --git a/internal/httpapi/assistant_process.go b/internal/httpapi/assistant_process.go index 3f2f15ce..725643a6 100644 --- a/internal/httpapi/assistant_process.go +++ b/internal/httpapi/assistant_process.go @@ -490,7 +490,7 @@ func (h *assistantHandlers) testRunnerCLI(ctx context.Context, runnerID, provide if provider == assistantEngineCodeBuddy { model = loadAssistantModels(ctx, user.CodeBuddyInternetEnvironment).Default } else if !runnerLogin { - models := assistantDiscoveredModelsFor(ctx, user, provider) + models, _ := assistantDiscoveredModelsFor(ctx, user, provider) if len(models) == 0 { models = assistantEngineBuiltinModels(provider) } diff --git a/internal/httpapi/assistant_session.go b/internal/httpapi/assistant_session.go index 9660e270..fe2f8ffd 100644 --- a/internal/httpapi/assistant_session.go +++ b/internal/httpapi/assistant_session.go @@ -292,9 +292,10 @@ func (h *assistantHandlers) session(c *gin.Context) { existing, _ := loadAssistantConversation(h.repo, ownerID, id) var body struct { assistantConversation - Effort *string `json:"effort"` + Effort *string `json:"effort"` + ApprovalUserID int64 `json:"approvalUserId"` } - if !decodeLimit(c, &body, assistantRequestBodyLimit) { + if !decodeLimit(c, &body, assistantRequestBodyLimit) || !requireAssistantApprovalUser(c, current.ID, body.ApprovalUserID) { return } conv := body.assistantConversation @@ -401,10 +402,19 @@ func (h *assistantHandlers) session(c *gin.Context) { if conv.Model == "" { conv.Model = existing.Model } - editor, _ := h.repo.User(current.ID) - if err := validateAssistantEffort(editor, conv.Model, conv.Effort); err != nil { - c.JSON(400, gin.H{"error": err.Error()}) - return + model := assistantRequestModel(conv.Executor, conv.Model) + engine, modelID := assistantModelEngine(model) + oldEngine, oldModelID := assistantModelEngine(assistantRequestModel(existing.Executor, existing.Model)) + // Metadata edits can keep historical settings after the model catalog changes. + if engine != oldEngine || modelID != oldModelID || conv.Effort != existing.Effort { + editor, _ := h.repo.User(current.ID) + if engine == assistantEngineCodeBuddy { + editor, _, _ = h.assistantChatUser(editor) + } + if err := validateAssistantEffort(c.Request.Context(), editor, model, conv.Effort); err != nil { + c.JSON(400, gin.H{"error": err.Error()}) + return + } } var saveErr error if shared { diff --git a/internal/httpapi/assistant_sponsor.go b/internal/httpapi/assistant_sponsor.go index 4f7f3865..9b3f0f44 100644 --- a/internal/httpapi/assistant_sponsor.go +++ b/internal/httpapi/assistant_sponsor.go @@ -32,10 +32,9 @@ func (h *assistantHandlers) assistantSharedKey() (repo.SharedCodeBuddyKey, bool) } // assistantChatUser returns the identity a chat turn runs as. Users with -// their own key are returned unchanged. Otherwise the shared key and its -// network environment are substituted and shared lists the models the turn -// may use; everything else (ID, permissions, Git credentials) stays the -// caller's own. +// their own key are returned unchanged. Other turns use the shared key, +// its environment and catalog while keeping the caller's identity, +// permissions and Git credentials. func (h *assistantHandlers) assistantChatUser(user domain.User) (chatUser domain.User, shared []string, ok bool) { if strings.TrimSpace(user.CodeBuddyAPIKey) != "" { return user, nil, true @@ -46,6 +45,7 @@ func (h *assistantHandlers) assistantChatUser(user domain.User) (chatUser domain } user.CodeBuddyAPIKey = key.APIKey user.CodeBuddyInternetEnvironment = key.InternetEnvironment + user.CodeBuddyModels = nil return user, key.Models, true } diff --git a/internal/httpapi/assistant_sponsor_test.go b/internal/httpapi/assistant_sponsor_test.go index 1a432c6e..0adc9504 100644 --- a/internal/httpapi/assistant_sponsor_test.go +++ b/internal/httpapi/assistant_sponsor_test.go @@ -1,11 +1,13 @@ package httpapi import ( + "bytes" "context" "encoding/json" "net/http" "net/http/httptest" "path/filepath" + "slices" "strings" "testing" @@ -18,6 +20,56 @@ import ( "github.com/gin-gonic/gin" ) +func TestSharedCodeBuddyEffortsIgnoreAnOldPersonalOverride(t *testing.T) { + st, member := sponsorFixture(t) + stubCLIModelDiscovery(t) + h := &assistantHandlers{ + deps: &deps{repo: st}, + generate: func(_ context.Context, _ string, user domain.User) (string, error) { + if len(user.CodeBuddyModels) != 0 || user.ID != member.ID { + t.Error("shared-key turn must use its own catalog and keep the caller's identity") + } + return `{"message":"ok","drafts":[]}`, nil + }, + titleSummary: func(context.Context, domain.User, string) (string, error) { return "", context.Canceled }, + } + const model = "deepseek-v4.1-flash" + defaultOnly := []string{} + models := []domain.CLIModel{{ID: model, ReasoningEfforts: &defaultOnly}} + if err := st.SetPlatformCodeBuddyKey(repo.PlatformCodeBuddyKeyUpdate{APIKey: "shared-fixture", Models: []string{model}}, 0); err != nil { + t.Fatal(err) + } + personal, err := st.UpdateUserCredentials(member.ID, repo.UserCredentialUpdate{CodeBuddyAPIKey: "own-fixture", CodeBuddyModels: &models}) + if err != nil { + t.Fatal(err) + } + if err := validateAssistantEffort(context.Background(), personal, model, "low"); err == nil { + t.Fatal("the personal key must still use its explicit override") + } + afterClear, err := st.UpdateUserCredentials(member.ID, repo.UserCredentialUpdate{CodeBuddyAPIKeyClear: true}) + if err != nil || len(afterClear.CodeBuddyModels) != 1 { + t.Fatalf("expected the personal model catalog to remain saved after clearing its key: %v", err) + } + c, rec := workbenchContext(afterClear, http.MethodGet, "/api/assistant/models", &bytes.Buffer{}) + h.models(c) + var catalog assistantModelCatalog + if err := json.Unmarshal(rec.Body.Bytes(), &catalog); rec.Code != http.StatusOK || err != nil || + len(catalog.Models) != 1 || catalog.Models[0].ID != model || !slices.Contains(catalog.Models[0].Efforts, "low") { + t.Fatalf("shared catalog: %d %s", rec.Code, rec.Body) + } + c, rec = workbenchContext(afterClear, http.MethodPut, "/api/assistant/sessions/shared-effort", bytes.NewBufferString(`{"model":"deepseek-v4.1-flash","effort":"low","executor":"codebuddy"}`)) + c.Params = gin.Params{{Key: "id", Value: "shared-effort"}} + h.session(c) + if rec.Code != http.StatusOK { + t.Fatalf("session rejected the shared effort: %d %s", rec.Code, rec.Body) + } + c, rec = workbenchContext(afterClear, http.MethodPost, "/api/assistant/chat", bytes.NewBufferString(`{"sessionId":"shared-effort","model":"deepseek-v4.1-flash","effort":"low","messages":[{"role":"user","content":"Continue"}]}`)) + h.chat(c) + if rec.Code != http.StatusOK { + t.Fatalf("turn rejected the shared effort: %d %s", rec.Code, rec.Body) + } +} + func sponsorFixture(t *testing.T) (*repo.Store, domain.User) { t.Helper() st, err := repo.OpenTest() @@ -286,7 +338,7 @@ func TestAssistantModelsPerExecutor(t *testing.T) { _ = json.Unmarshal(rec.Body.Bytes(), &catalog) return rec, catalog } - for executor, family := range map[string]string{"codex": "gpt", "claude": "claude"} { + for executor, family := range map[string]string{"codex": "gpt", "claude": "claude", "claude-code": "claude"} { rec, catalog := get(executor) if rec.Code != 200 || catalog.Default != "default" || len(catalog.Models) < 2 || catalog.Models[0].ID != "default" { t.Fatalf("%s catalog: %d %s", executor, rec.Code, rec.Body.String()) @@ -295,6 +347,9 @@ func TestAssistantModelsPerExecutor(t *testing.T) { if model.Family != family { t.Fatalf("%s model %s family = %q", executor, model.ID, model.Family) } + if len(model.Efforts) == 0 { + t.Fatalf("%s model %s lost reasoning efforts", executor, model.ID) + } } if catalog.SponsoredModel != "" { t.Fatalf("%s must not offer the CodeBuddy sponsored model", executor) diff --git a/internal/httpapi/assistant_title.go b/internal/httpapi/assistant_title.go index 281ef276..3c114da9 100644 --- a/internal/httpapi/assistant_title.go +++ b/internal/httpapi/assistant_title.go @@ -120,7 +120,7 @@ func (h *assistantHandlers) scheduleAssistantTitle(ctx context.Context, ownerID // effort; the gateway behind a Base URL may serve nothing else. func assistantTitleModel(ctx context.Context, user domain.User, model string, shared []string) (string, string) { if engine, _ := assistantModelEngine(model); engine != assistantEngineCodeBuddy { - if efforts := assistantModelEfforts(user, model); len(efforts) > 0 { + if efforts := assistantModelEfforts(ctx, user, model); len(efforts) > 0 { return model, efforts[0] } return model, "" diff --git a/internal/httpapi/assistant_title_test.go b/internal/httpapi/assistant_title_test.go index b62aaac2..181f15b6 100644 --- a/internal/httpapi/assistant_title_test.go +++ b/internal/httpapi/assistant_title_test.go @@ -63,8 +63,9 @@ func TestAssistantTitleModelReusesSessionModelOnOtherCLIs(t *testing.T) { {"pi:deepseek-v4-flash", "minimal"}, {"opencode:glm-5", "low"}, {"kimi:kimi-k3", ""}, - {"dsh:deepseek-v4-pro", ""}, - {"zcode:glm-5.3", ""}, + {"kimi:k3.1-pro", "low"}, + {"dsh:deepseek-v4-pro", "none"}, + {"zcode:glm-5.3", "low"}, } { model, effort := assistantTitleModel(context.Background(), domain.User{}, tc.model, nil) if model != tc.model || effort != tc.wantEffort { diff --git a/internal/httpapi/catalog_test.go b/internal/httpapi/catalog_test.go index 60e04cba..768f3dd7 100644 --- a/internal/httpapi/catalog_test.go +++ b/internal/httpapi/catalog_test.go @@ -42,6 +42,8 @@ func TestPutAgentValidatesEffortLevel(t *testing.T) { } func TestExecutorModelsCatalogListsPerCLI(t *testing.T) { + t.Setenv("AGENTFLOW_ASSISTANT_MODEL", "") + t.Setenv("AGENTFLOW_ASSISTANT_MODELS", "test-model") s, err := repo.OpenTest() if err != nil { t.Fatal(err) @@ -57,11 +59,29 @@ func TestExecutorModelsCatalogListsPerCLI(t *testing.T) { if cb.Models[0].Credits != "" { t.Fatal("unconfigured public model must not claim a private price") } + for _, model := range cb.Models { + if model.ID == assistantDefaultModel { + t.Fatal("the CodeBuddy CLI default must not be listed as a model") + } + } codex := catalogs["codex"] if len(codex.Models) < 2 || codex.Models[0].ID != "gpt-6.1-sol" || codex.Models[1].ID != "gpt-6.1-terra" || codex.Models[1].Name != "GPT-6.1 Terra" { t.Fatalf("codex catalog = %+v", codex.Models) } + codexNames := make(map[string]string, len(codex.Models)) + for _, model := range codex.Models { + codexNames[model.ID] = model.Name + } + for id, name := range map[string]string{"gpt-6.2-sol": "GPT-6.2 Sol", "gpt-6.2-astra": "GPT-6.2 Astra"} { + if got := codexNames[id]; got != name { + t.Errorf("Codex model %s has name %q, want %q", id, got, name) + } + } claude := catalogs["claude"] + if len(claude.Models) < 2 || claude.Models[0].ID != "claude-fable-5-1" || + claude.Models[1].ID != "claude-fable-5-2" || claude.Models[1].Name != "Claude Fable 5.2" { + t.Fatalf("claude catalog = %+v", claude.Models) + } haiku := -1 for i, model := range claude.Models { if model.ID == "claude-haiku-5-5" { diff --git a/internal/httpapi/cli_model_efforts_test.go b/internal/httpapi/cli_model_efforts_test.go new file mode 100644 index 00000000..8f07a8f2 --- /dev/null +++ b/internal/httpapi/cli_model_efforts_test.go @@ -0,0 +1,291 @@ +package httpapi + +import ( + "bytes" + "context" + "encoding/json" + "io" + "net/http" + "net/http/httptest" + "slices" + "strconv" + "sync/atomic" + "testing" + + "agentflow/internal/domain" + "agentflow/internal/repo" + + "github.com/gin-gonic/gin" +) + +func TestModelEffortsSurviveProbeSaveAndSelection(t *testing.T) { + h, owner := workbenchFixture(t) + stubCLIModelDiscovery(t) + service := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { + if r.URL.Path != "/models" || r.Header.Get("Authorization") != "Bearer fixture-key" { + t.Errorf("unexpected discovery request: %s", r.URL.Path) + } + _, _ = io.WriteString(w, `{"data":[ + {"id":"gateway-reasoner","supported_reasoning_efforts":["xhigh","low","invalid","xhigh"]}, + {"id":"gateway-chat","reasoningEfforts":[]}, + {"id":"gpt-5.5"} + ]}`) + })) + t.Cleanup(service.Close) + body, _ := json.Marshal(map[string]string{"engine": "codex", "apiKey": "fixture-key", "baseUrl": service.URL}) + c, response := workbenchContext(owner, http.MethodPost, "/", bytes.NewBuffer(body)) + c.Params = gin.Params{{Key: "id", Value: strconv.FormatInt(owner.ID, 10)}} + (&usersHandlers{deps: h.deps}).probeCLIModels(c) + var result cliModelProbeResult + if err := json.Unmarshal(response.Body.Bytes(), &result); err != nil || response.Code != http.StatusOK || result.Error != "" || len(result.Models) != 3 { + t.Fatalf("probe=%s err=%v", response.Body, err) + } + for i, want := range [][]string{{"low", "xhigh"}, {}, {"low", "medium", "high", "xhigh"}} { + if got := result.Models[i].DiscoveredReasoningEfforts; got == nil || !slices.Equal(*got, want) { + t.Fatalf("model %s lost efforts: %+v, want %v", result.Models[i].ID, got, want) + } + if result.Models[i].ReasoningEfforts != nil { + t.Fatalf("model %s returned an override instead of discovered capabilities", result.Models[i].ID) + } + } + if _, err := h.repo.UpdateUserCredentials(owner.ID, repo.UserCredentialUpdate{ + CodexAPIKey: "fixture-key", CodexBaseURL: &service.URL, CodexModels: &result.Models, + }); err != nil { + t.Fatal(err) + } + owner, _ = h.repo.User(owner.ID) + models, _ := assistantEngineModels(context.Background(), owner, nil) + for id, efforts := range map[string][]string{ + "codex:gateway-reasoner": {"low", "xhigh"}, + "codex:gpt-5.5": {"low", "medium", "high", "xhigh"}, + } { + if !slices.ContainsFunc(models, func(model assistantModel) bool { + return model.ID == id && slices.Equal(model.Efforts, efforts) + }) { + t.Errorf("chat catalog missing %s with efforts %v: %+v", id, efforts, models) + } + } + for _, check := range []struct { + model, effort string + valid bool + }{ + {"codex:gateway-reasoner", "xhigh", true}, + {"codex:gateway-reasoner", "high", false}, + {"codex:gateway-chat", "high", false}, + {"codex:gateway-chat", "", true}, + {"codex:gpt-5.5", "xhigh", true}, + {"codex:gpt-5.5", "max", false}, + } { + if err := validateAssistantEffort(context.Background(), owner, check.model, check.effort); (err == nil) != check.valid { + t.Errorf("validate %s/%s: %v", check.model, check.effort, err) + } + } + c, response = workbenchContext(owner, http.MethodGet, "/api/executor-models", bytes.NewBuffer(nil)) + (&catalogHandlers{deps: h.deps}).executorModels(c) + var catalogs map[string]executorModelCatalog + if err := json.Unmarshal(response.Body.Bytes(), &catalogs); err != nil { + t.Fatal(err) + } + if got := catalogs["codex"].Models; len(got) != 3 || + !slices.Equal(got[0].Efforts, effortOptions([]string{"low", "xhigh"})) || len(got[1].Efforts) != 0 || + !slices.Equal(got[2].Efforts, effortOptions([]string{"low", "medium", "high", "xhigh"})) { + t.Fatalf("role catalog lost per-model efforts: %+v", got) + } +} + +func TestModelEffortsKeepLegacyCatalogsAndCLIConstraints(t *testing.T) { + custom := []string{"low", "xhigh", "ultra"} + discovered := []string{"high"} + explicitDefault := []string{} + switchLevels := []string{"off", "on"} + deepseekLevels := []string{"off", "max"} + for _, check := range []struct { + model string + entry domain.CLIModel + want []string + }{ + {"codex:custom", domain.CLIModel{ID: "custom"}, []string{"low", "medium", "high"}}, + {"kimi:unknown", domain.CLIModel{ID: "unknown"}, nil}, + {"kimi:kimi-for-coding-highspeed", domain.CLIModel{ID: "kimi-for-coding-highspeed", DiscoveredReasoningEfforts: &switchLevels}, []string{"none", "thinking"}}, + {"kimi:k3.1-pro", domain.CLIModel{ID: "k3.1-pro", DiscoveredReasoningEfforts: &deepseekLevels}, []string{"max"}}, + {"dsh:deepseek-flash", domain.CLIModel{ID: "deepseek-flash", DiscoveredReasoningEfforts: &deepseekLevels}, []string{"none", "max"}}, + {"dsh:deepseek-flash", domain.CLIModel{ID: "deepseek-flash", ReasoningEfforts: &explicitDefault, DiscoveredReasoningEfforts: &deepseekLevels}, nil}, + {"zcode:glm-5.3", domain.CLIModel{ID: "glm-5.3", DiscoveredReasoningEfforts: &discovered}, discovered}, + {"codex:custom", domain.CLIModel{ID: "custom", DiscoveredReasoningEfforts: &discovered}, discovered}, + {"codex:custom", domain.CLIModel{ID: "custom", ReasoningEfforts: &custom, DiscoveredReasoningEfforts: &discovered}, custom}, + {"codex:custom", domain.CLIModel{ID: "custom", ReasoningEfforts: &custom}, custom}, + {"codex:gpt-5.5", domain.CLIModel{ID: "gpt-5.5"}, []string{"low", "medium", "high", "xhigh"}}, + {"codex:gpt-5.5", domain.CLIModel{ID: "gpt-5.5", DiscoveredReasoningEfforts: &discovered}, discovered}, + {"codex:gpt-5.5", domain.CLIModel{ID: "gpt-5.5", ReasoningEfforts: &custom, DiscoveredReasoningEfforts: &discovered}, custom}, + {"codex:gpt-5.5", domain.CLIModel{ID: "gpt-5.5", ReasoningEfforts: &explicitDefault, DiscoveredReasoningEfforts: &discovered}, nil}, + {"claude-code:claude-fable-5-2", domain.CLIModel{ID: "claude-fable-5-2"}, []string{"low", "medium", "high", "xhigh", "max"}}, + {"claude-code:claude-fable-5-2", domain.CLIModel{ID: "claude-fable-5-2", DiscoveredReasoningEfforts: &discovered}, discovered}, + {"claude-code:claude-fable-5-2", domain.CLIModel{ID: "claude-fable-5-2", ReasoningEfforts: &custom, DiscoveredReasoningEfforts: &discovered}, []string{"low", "xhigh"}}, + {"claude-code:claude-fable-5-2", domain.CLIModel{ID: "claude-fable-5-2", ReasoningEfforts: &explicitDefault, DiscoveredReasoningEfforts: &discovered}, nil}, + {"claude-code:claude-opus-5.5", domain.CLIModel{ID: "claude-opus-5.5"}, []string{"low", "medium", "high", "xhigh", "max"}}, + {"claude-fable-5-2", domain.CLIModel{ID: "claude-fable-5-2"}, []string{"low", "medium", "high"}}, + {"gpt-5.5", domain.CLIModel{ID: "gpt-5.5"}, []string{"low", "medium", "high"}}, + {"opencode:custom", domain.CLIModel{ID: "custom", ReasoningEfforts: &custom}, []string{"low"}}, + {"agy:claude-custom", domain.CLIModel{ID: "claude-custom", ReasoningEfforts: &custom}, nil}, + {"cursor:unknown", domain.CLIModel{ID: "unknown", ReasoningEfforts: &custom}, nil}, + {"cursor:widget", domain.CLIModel{ID: "widget", Efforts: map[string]string{"low": "widget-low", "xhigh": "widget-extra-high"}, ReasoningEfforts: &custom}, []string{"low", "xhigh"}}, + } { + if got := assistantModelEffortsIn(check.model, []domain.CLIModel{check.entry}); !slices.Equal(got, check.want) { + t.Errorf("%s: %v, want %v", check.model, got, check.want) + } + } +} + +func TestAutomaticDiscoveryUsesSameEffortsForColdValidation(t *testing.T) { + var requests atomic.Int32 + gateway := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { + if r.URL.Path != "/models" || r.Header.Get("Authorization") != "Bearer fixture-key" { + t.Error("unexpected model discovery request") + } + requests.Add(1) + _, _ = io.WriteString(w, `{"data":[{"id":"remote","supported_reasoning_efforts":["xhigh"]}]}`) + })) + t.Cleanup(gateway.Close) + ctx := context.Background() + user := domain.User{CodexAPIKey: "fixture-key", CodexBaseURL: gateway.URL} + key := assistantDiscoveryKey(false, user.CodexAPIKey, user.CodexBaseURL) + t.Cleanup(func() { assistantDiscoveryCache.Delete(key) }) + catalog, _ := assistantEngineModels(ctx, user, nil) + if !slices.ContainsFunc(catalog, func(model assistantModel) bool { + return model.ID == "codex:remote" && slices.Equal(model.Efforts, []string{"xhigh"}) + }) { + t.Fatalf("the picker did not expose the gateway capability: %+v", catalog) + } + + assistantDiscoveryCache.Delete(key) + if err := validateAssistantEffort(ctx, user, "codex:remote", "xhigh"); err != nil { + t.Fatal(err) + } + if err := validateAssistantEffort(ctx, user, "codex:remote", "high"); err == nil { + t.Fatal("validation ignored the discovered model's capabilities") + } + if err := validateAssistantEffort(ctx, domain.User{}, "codex:remote", "xhigh"); err == nil { + t.Fatal("another user's discovered capabilities were used") + } + if got := requests.Load(); got != 2 { + t.Fatalf("expected one discovery per process cache, got %d", got) + } +} + +func TestKimiThinkingSwitchSurvivesSavedCatalog(t *testing.T) { + h, owner := workbenchFixture(t) + levels := []string{"none", "thinking"} + models := []domain.CLIModel{{ID: "kimi-for-coding-highspeed", ReasoningEfforts: &levels}} + if _, err := h.repo.UpdateUserCredentials(owner.ID, repo.UserCredentialUpdate{ + AgentCLICredentials: map[string]repo.AgentCLICredentialUpdate{ + "kimi": {APIKey: "fixture-key", Models: &models}, + }, + }); err != nil { + t.Fatal(err) + } + owner, _ = h.repo.User(owner.ID) + chat, _ := assistantEngineModels(context.Background(), owner, nil) + if !slices.ContainsFunc(chat, func(model assistantModel) bool { + return model.ID == "kimi:kimi-for-coding-highspeed" && slices.Equal(model.Efforts, levels) + }) { + t.Fatalf("saved Kimi thinking options not shown: %+v", chat) + } + + baseURL := "https://gateway.example/v1" + if _, err := h.repo.UpdateUserCredentials(owner.ID, repo.UserCredentialUpdate{ + AgentCLICredentials: map[string]repo.AgentCLICredentialUpdate{"kimi": {BaseURL: &baseURL, Models: &models}}, + }); err != nil { + t.Fatal(err) + } + owner, _ = h.repo.User(owner.ID) + chat, _ = assistantEngineModels(context.Background(), owner, nil) + if !slices.ContainsFunc(chat, func(model assistantModel) bool { + return model.ID == "kimi:kimi-for-coding-highspeed" && len(model.Efforts) == 0 + }) { + t.Fatalf("gateway catalog advertised a Kimi override: %+v", chat) + } + if err := validateAssistantEffort(context.Background(), owner, "kimi:kimi-for-coding-highspeed", "thinking"); err == nil { + t.Fatal("Kimi gateway accepted an unsupported per-turn override") + } +} + +func TestBuiltinModelEffortsMatchTheCatalogAndTurnValidation(t *testing.T) { + h, owner := workbenchFixture(t) + chatCatalogs := make(map[string]assistantModelCatalog) + for _, engine := range []string{"codex", "claude-code", "kimi", "dsh", "zcode"} { + c, response := workbenchContext(owner, http.MethodGet, "/api/assistant/models?executor="+engine, &bytes.Buffer{}) + h.models(c) + var chat assistantModelCatalog + if err := json.Unmarshal(response.Body.Bytes(), &chat); response.Code != http.StatusOK || err != nil { + t.Fatalf("%s chat catalog: %d %s", engine, response.Code, response.Body) + } + chatCatalogs[engine] = chat + } + c, response := workbenchContext(owner, http.MethodGet, "/api/executor-models?builtin=1", &bytes.Buffer{}) + (&catalogHandlers{deps: h.deps}).executorModels(c) + var role map[string]executorModelCatalog + if err := json.Unmarshal(response.Body.Bytes(), &role); response.Code != http.StatusOK || err != nil { + t.Fatalf("role catalog: %d %s", response.Code, response.Body) + } + xhigh := []string{"low", "medium", "high", "xhigh"} + max := []string{"low", "medium", "high", "xhigh", "max"} + ultra := []string{"low", "medium", "high", "xhigh", "max", "ultra"} + levels := append([]string{"none", "thinking"}, ultra...) + for _, model := range []struct { + engine string + id string + levels []string + }{ + {"codex", "gpt-5.5", xhigh}, + {"codex", "gpt-5.6-luna", max}, + {"codex", "gpt-6-luna", max}, + {"codex", "gpt-6.1-terra", ultra}, + {"codex", "gpt-6.2-sol", ultra}, + {"codex", "gpt-6.2-astra", ultra}, + {"claude-code", "claude-fable-5-2", max}, + {"claude-code", "claude-opus-5-5", max}, + {"claude-code", "claude-sonnet-4-6", []string{"low", "medium", "high", "max"}}, + {"kimi", "k3.1-pro", []string{"low", "high", "max"}}, + {"kimi", "k3", []string{"none", "low", "high", "max"}}, + {"kimi", "kimi-for-coding-highspeed", []string{"none", "thinking"}}, + {"dsh", "deepseek-flash", []string{"none", "low", "high", "max"}}, + {"dsh", "deepseek-v4-pro", []string{"none", "low", "high", "max"}}, + {"zcode", "glm-5.3", []string{"low", "high", "max"}}, + {"zcode", "glm-5.3-flash", []string{"low", "high", "max"}}, + } { + t.Run(model.engine+"/"+model.id, func(t *testing.T) { + chat := chatCatalogs[model.engine] + chatIndex := slices.IndexFunc(chat.Models, func(m assistantModel) bool { return m.ID == model.id }) + if chatIndex < 0 || !slices.Equal(chat.Models[chatIndex].Efforts, model.levels) { + t.Errorf("chat picker missing %s with efforts %v", model.id, model.levels) + } + roleKey := model.engine + if roleKey == "claude-code" { + roleKey = "claude" + } + roleModels := role[roleKey].Models + roleIndex := slices.IndexFunc(roleModels, func(m executorModelInfo) bool { return m.ID == model.id }) + if roleIndex < 0 || !slices.Equal(roleModels[roleIndex].Efforts, effortOptions(model.levels)) { + t.Errorf("role picker missing %s with efforts %v", model.id, model.levels) + } + for _, effort := range levels { + err := validateAssistantEffort(context.Background(), owner, model.engine+":"+model.id, effort) + if want := slices.Contains(model.levels, effort); (err == nil) != want { + t.Errorf("%s effort %s: validation returned %v, want accepted=%t", model.id, effort, err, want) + } + } + last := model.levels[len(model.levels)-1] + body, err := json.Marshal(map[string]string{"model": model.id, "executor": model.engine, "effort": last}) + if err != nil { + t.Fatal(err) + } + sessionID := model.engine + "-" + model.id + c, response := workbenchContext(owner, http.MethodPut, "/api/assistant/sessions/"+sessionID, bytes.NewBuffer(body)) + c.Params = gin.Params{{Key: "id", Value: sessionID}} + h.session(c) + if response.Code != http.StatusOK { + t.Fatalf("session rejected the bare model ID with executor %s: %d %s", model.engine, response.Code, response.Body) + } + }) + } +} diff --git a/internal/httpapi/cli_models_probe.go b/internal/httpapi/cli_models_probe.go index 398bb685..29e3125e 100644 --- a/internal/httpapi/cli_models_probe.go +++ b/internal/httpapi/cli_models_probe.go @@ -81,6 +81,7 @@ func (h *usersHandlers) probeCLIModels(c *gin.Context) { if req.BaseURL != nil { baseURL = strings.TrimSpace(*req.BaseURL) } + configuredBaseURL := baseURL if req.Protocol != nil && runtime.IsAgentCLI(engine) { protocol = runtime.AgentCLIProtocol(engine, *req.Protocol) } @@ -98,7 +99,7 @@ func (h *usersHandlers) probeCLIModels(c *gin.Context) { defer cancel() var models []domain.CLIModel for _, id := range discoverCLIModels(ctx, environment, apiKey) { - if assistantValidEngineModel(id) { + if id != assistantDefaultModel && assistantValidEngineModel(id) { models = append(models, domain.CLIModel{ID: id}) } } @@ -110,7 +111,7 @@ func (h *usersHandlers) probeCLIModels(c *gin.Context) { if apiKey == "" { source = "builtin" } - c.JSON(http.StatusOK, cliModelProbeResult{Models: models, Source: source}) + c.JSON(http.StatusOK, cliModelProbeResult{Models: cliModelsWithEfforts(engine, models, ""), Source: source}) return } if engine == runtime.AgentCLICursor { @@ -118,7 +119,7 @@ func (h *usersHandlers) probeCLIModels(c *gin.Context) { return } if baseURL == "" && apiKey == "" { - c.JSON(http.StatusOK, cliModelProbeResult{Models: cliModelsOf(assistantEngineBuiltinModels(engine)), Source: "builtin"}) + c.JSON(http.StatusOK, cliModelProbeResult{Models: cliModelsWithEfforts(engine, cliModelsOf(assistantEngineBuiltinModels(engine)), ""), Source: "builtin"}) return } if apiKey == "" { @@ -164,7 +165,7 @@ func (h *usersHandlers) probeCLIModels(c *gin.Context) { c.JSON(http.StatusOK, cliModelProbeResult{Error: "服务没有返回可用模型,可以手动添加模型"}) return } - c.JSON(http.StatusOK, cliModelProbeResult{Models: cliModelsOf(models), Source: source}) + c.JSON(http.StatusOK, cliModelProbeResult{Models: cliModelsWithEfforts(engine, cliModelsOf(models), configuredBaseURL), Source: source}) } func discoverGeminiModels(ctx context.Context, key, baseURL string) ([]assistantDiscoveredModel, error) { @@ -218,7 +219,7 @@ func (h *usersHandlers) probeCursorModels(c *gin.Context, apiKey string) { defer cancel() models, err := cursorAgentModels(ctx, apiKey) if errors.Is(err, errAssistantCLIMissing) { - c.JSON(http.StatusOK, cliModelProbeResult{Models: cliModelsOf(assistantEngineBuiltinModels(runtime.AgentCLICursor)), Source: "builtin"}) + c.JSON(http.StatusOK, cliModelProbeResult{Models: cliModelsWithEfforts(runtime.AgentCLICursor, cliModelsOf(assistantEngineBuiltinModels(runtime.AgentCLICursor)), ""), Source: "builtin"}) return } if err != nil { @@ -229,7 +230,7 @@ func (h *usersHandlers) probeCursorModels(c *gin.Context, apiKey string) { c.JSON(http.StatusOK, cliModelProbeResult{Error: "Cursor CLI 没有返回可用模型,可以手动添加模型"}) return } - c.JSON(http.StatusOK, cliModelProbeResult{Models: models, Source: "official"}) + c.JSON(http.StatusOK, cliModelProbeResult{Models: cliModelsWithEfforts(runtime.AgentCLICursor, models, ""), Source: "official"}) } var cursorModelLine = regexp.MustCompile(`^([A-Za-z0-9][A-Za-z0-9._-]*) - (.+)$`) @@ -288,7 +289,25 @@ func parseCursorModels(output string) []domain.CLIModel { func cliModelsOf(models []assistantDiscoveredModel) []domain.CLIModel { out := make([]domain.CLIModel, 0, len(models)) for _, model := range models { - out = append(out, domain.CLIModel{ID: model.ID, Name: model.Name, Efforts: model.Efforts, DefaultEffort: model.DefaultEffort}) + out = append(out, domain.CLIModel{ + ID: model.ID, Name: model.Name, Efforts: model.Efforts, DefaultEffort: model.DefaultEffort, + ReasoningEfforts: model.ReasoningEfforts, DiscoveredReasoningEfforts: model.DiscoveredReasoningEfforts, + }) } return out } + +// Standard model listings often contain only IDs. Supply the CLI's catalog +// levels when the endpoint omits them, and constrain advertised metadata to +// levels this CLI can actually pass to its provider. +func cliModelsWithEfforts(engine string, models []domain.CLIModel, baseURL string) []domain.CLIModel { + for i := range models { + id := models[i].ID + if engine != assistantEngineCodeBuddy { + id = engine + ":" + id + } + levels := append([]string{}, assistantModelEffortsForBaseURL(id, models, baseURL)...) + models[i].DiscoveredReasoningEfforts = &levels + } + return models +} diff --git a/internal/httpapi/cursor_catalog_test.go b/internal/httpapi/cursor_catalog_test.go index ea275427..7cc2e467 100644 --- a/internal/httpapi/cursor_catalog_test.go +++ b/internal/httpapi/cursor_catalog_test.go @@ -1,7 +1,10 @@ package httpapi import ( + "bytes" "context" + "encoding/json" + "net/http" "slices" "testing" "time" @@ -48,45 +51,74 @@ func TestCursorFamiliesSurviveRestart(t *testing.T) { a, _ = st.User(a.ID) b, _ = st.User(b.ID) + stubCLIModelDiscovery(t) + release := make(chan struct{}) previous := assistantCursorModels t.Cleanup(func() { assistantCursorModels = previous; clearCursorModelsCache() }) assistantCursorModels = func(_ context.Context, apiKey string) ([]domain.CLIModel, error) { if apiKey != "crsr_cursor-a" { return nil, nil } + select { + case <-release: + case <-t.Context().Done(): + return nil, t.Context().Err() + } return cursorModelFamilies(parseCursorModels(cursorCatalogTestOutput)), nil } clearCursorModelsCache() - cachedCursorModels("crsr_cursor-a") + h := &assistantHandlers{deps: &deps{repo: st}} + readCatalog := func() assistantModelCatalog { + t.Helper() + c, rec := workbenchContext(a, http.MethodGet, "/api/assistant/models", &bytes.Buffer{}) + h.models(c) + var catalog assistantModelCatalog + if err := json.Unmarshal(rec.Body.Bytes(), &catalog); rec.Code != http.StatusOK || err != nil { + t.Fatalf("model catalog: %d %s", rec.Code, rec.Body) + } + return catalog + } + if catalog := readCatalog(); !catalog.CursorModelsPending { + t.Fatal("cold Cursor catalog must report that discovery is pending") + } + close(release) deadline := time.Now().Add(5 * time.Second) - for len(cachedCursorModels("crsr_cursor-a")) == 0 { + for { + catalog := readCatalog() + if !catalog.CursorModelsPending { + if !slices.ContainsFunc(catalog.Models, func(model assistantModel) bool { + return model.ID == "cursor:widget" && slices.Contains(model.Efforts, "xhigh") + }) { + t.Fatal("completed discovery did not expose the model's efforts") + } + break + } if time.Now().After(deadline) { t.Fatal("cursor model list was never cached") } time.Sleep(10 * time.Millisecond) } - h := &assistantHandlers{deps: &deps{repo: st}} - h.saveCursorCatalog(a) // Restart: the process forgets every listed model. clearCursorModelsCache() assistantCursorModels = func(context.Context, string) ([]domain.CLIModel, error) { return nil, nil } a, _ = st.User(a.ID) - if err := validateAssistantEffort(a, "cursor:widget", "xhigh"); err != nil { + if err := validateAssistantEffort(context.Background(), a, "cursor:widget", "xhigh"); err != nil { t.Fatalf("saved family rejected its effort: %v", err) } - if err := validateAssistantEffort(a, "cursor:nova", "max"); err == nil { + if err := validateAssistantEffort(context.Background(), a, "cursor:nova", "max"); err == nil { t.Fatal("effort the family lacks was accepted") } - if err := validateAssistantEffort(b, "cursor:widget", "xhigh"); err == nil { + if err := validateAssistantEffort(context.Background(), b, "cursor:widget", "xhigh"); err == nil { t.Fatal("another user's family resolved") } - if err := validateAssistantEffort(a, "cursor:grok-5-code", "high"); err == nil { + if err := validateAssistantEffort(context.Background(), a, "cursor:grok-5-code", "high"); err == nil { t.Fatal("unknown model accepted an effort") } var listed []assistantModel - for _, model := range assistantEngineModels(context.Background(), a, nil) { + models, _ := assistantEngineModels(context.Background(), a, nil) + for _, model := range models { if model.Engine == runtime.AgentCLICursor { listed = append(listed, model) } diff --git a/internal/httpapi/executor_models.go b/internal/httpapi/executor_models.go index f59dba1f..2939325d 100644 --- a/internal/httpapi/executor_models.go +++ b/internal/httpapi/executor_models.go @@ -12,9 +12,10 @@ import ( // executorModelInfo describes one selectable model of a built-in CLI. type executorModelInfo struct { - ID string `json:"id"` - Name string `json:"name"` - Credits string `json:"credits,omitempty"` + ID string `json:"id"` + Name string `json:"name"` + Credits string `json:"credits,omitempty"` + Efforts []executorEffortOption `json:"efforts"` } // executorEffortOption is one reasoning-effort level of a built-in CLI. @@ -34,8 +35,74 @@ type executorModelCatalog struct { Efforts []executorEffortOption `json:"efforts"` } +type builtinExecutorModel struct { + id string + name string + efforts []string +} + +// Reasoning levels offered by the built-in models on each CLI. +var builtinExecutorModels = func() map[string][]builtinExecutorModel { + legacy := []string{"low", "medium", "high"} + lowHighMax := []string{"low", "high", "max"} + offLowHighMax := []string{"none", "low", "high", "max"} + xhigh := []string{"low", "medium", "high", "xhigh"} + max := []string{"low", "medium", "high", "xhigh", "max"} + maxWithoutXHigh := []string{"low", "medium", "high", "max"} + ultra := []string{"low", "medium", "high", "xhigh", "max", "ultra"} + return map[string][]builtinExecutorModel{ + assistantEngineCodex: { + {"gpt-6.1-sol", "GPT-6.1 Sol", ultra}, + {"gpt-6.1-terra", "GPT-6.1 Terra", ultra}, + {"gpt-6.2-sol", "GPT-6.2 Sol", ultra}, + {"gpt-6.2-astra", "GPT-6.2 Astra", ultra}, + {"gpt-6-astra", "GPT-6 Astra", ultra}, + {"gpt-6-sol", "GPT-6 Sol", ultra}, + {"gpt-6-luna", "GPT-6 Luna", max}, + {"gpt-5.6-sol", "GPT-5.6 Sol", ultra}, + {"gpt-5.6-terra", "GPT-5.6 Terra", ultra}, + {"gpt-5.6-luna", "GPT-5.6 Luna", max}, + {"gpt-5.5", "GPT-5.5", xhigh}, + }, + assistantEngineClaude: { + {"claude-fable-5-1", "Claude Fable 5.1", max}, + {"claude-fable-5-2", "Claude Fable 5.2", max}, + {"claude-fable-5", "Claude Fable 5", max}, + {"claude-opus-5-5", "Claude Opus 5.5", max}, + {"claude-opus-5", "Claude Opus 5", max}, + {"claude-opus-4-8", "Claude Opus 4.8", max}, + {"claude-opus-4-7", "Claude Opus 4.7", max}, + {"claude-opus-4-6", "Claude Opus 4.6", maxWithoutXHigh}, + {"claude-sonnet-5-5", "Claude Sonnet 5.5", max}, + {"claude-sonnet-5", "Claude Sonnet 5", max}, + {"claude-sonnet-4-6", "Claude Sonnet 4.6", maxWithoutXHigh}, + // Keep accepting effort values saved in existing 4.5 conversations. + {"claude-sonnet-4-5", "Claude Sonnet 4.5", legacy}, + {"claude-haiku-5-5", "Claude Haiku 5.5", max}, + {"claude-haiku-4-5", "Claude Haiku 4.5", legacy}, + }, + runtime.AgentCLIKimi: { + {"k3.1-pro", "K3.1 Pro", lowHighMax}, + {"k3.1-pro-highspeed", "K3.1 Pro Highspeed", lowHighMax}, + {"k3", "K3", offLowHighMax}, + {"k3-256k", "K3 256K", offLowHighMax}, + {"kimi-for-coding", "K2.8 Preview", offLowHighMax}, + {"kimi-for-coding-highspeed", "K2.7 Code Highspeed", []string{"none", "thinking"}}, + }, + runtime.AgentCLIDeepSeek: { + {"deepseek-flash", "DeepSeek V4.1 Flash", domain.ExecutorEffortLevels[runtime.AgentCLIDeepSeek]}, + {"deepseek-v4-pro", "DeepSeek V4 Pro", domain.ExecutorEffortLevels[runtime.AgentCLIDeepSeek]}, + }, + runtime.AgentCLIZCode: { + {"glm-5.3", "GLM-5.3", lowHighMax}, + {"glm-5.3-flash", "GLM-5.3 Flash", lowHighMax}, + }, + } +}() + var effortOptionLabels = map[string]string{ "none": "关闭 none", + "thinking": "开启 thinking", "minimal": "最低 minimal", "low": "低 low", "medium": "中 medium", @@ -60,13 +127,13 @@ func effortOptions(levels []string) []executorEffortOption { // executorModelCatalogs returns the fixed per-CLI catalog. The CodeBuddy list // reuses the assistant catalog so the wizard, AI 助手 and 澄清 stay in sync. -func executorModelCatalogs() map[string]executorModelCatalog { +func executorModelCatalogs(kimiBaseURL string) map[string]executorModelCatalog { codebuddyModels := make([]executorModelInfo, 0) known := assistantKnownCatalog() - for _, model := range known.Models { + for _, model := range selectableCodeBuddyModels(known.Models) { codebuddyModels = append(codebuddyModels, executorModelInfo{ID: model.ID, Name: model.Name, Credits: model.Credits}) } - return map[string]executorModelCatalog{ + catalogs := map[string]executorModelCatalog{ "codebuddy": { Default: known.Default, Models: codebuddyModels, @@ -74,37 +141,11 @@ func executorModelCatalogs() map[string]executorModelCatalog { }, "codex": { CustomModel: true, - Models: []executorModelInfo{ - {ID: "gpt-6.1-sol", Name: "GPT-6.1 Sol"}, - {ID: "gpt-6.1-terra", Name: "GPT-6.1 Terra"}, - {ID: "gpt-6-astra", Name: "GPT-6 Astra"}, - {ID: "gpt-6-sol", Name: "GPT-6 Sol"}, - {ID: "gpt-6-luna", Name: "GPT-6 Luna"}, - {ID: "gpt-5.6-sol", Name: "GPT-5.6 Sol"}, - {ID: "gpt-5.6-terra", Name: "GPT-5.6 Terra"}, - {ID: "gpt-5.6-luna", Name: "GPT-5.6 Luna"}, - {ID: "gpt-5.5", Name: "GPT-5.5"}, - }, - Efforts: effortOptions(domain.ExecutorEffortLevels["codex"]), + Efforts: effortOptions(domain.ExecutorEffortLevels["codex"]), }, "claude": { CustomModel: true, - Models: []executorModelInfo{ - {ID: "claude-fable-5-1", Name: "Claude Fable 5.1"}, - {ID: "claude-fable-5", Name: "Claude Fable 5"}, - {ID: "claude-opus-5-5", Name: "Claude Opus 5.5"}, - {ID: "claude-opus-5", Name: "Claude Opus 5"}, - {ID: "claude-opus-4-8", Name: "Claude Opus 4.8"}, - {ID: "claude-opus-4-7", Name: "Claude Opus 4.7"}, - {ID: "claude-opus-4-6", Name: "Claude Opus 4.6"}, - {ID: "claude-sonnet-5-5", Name: "Claude Sonnet 5.5"}, - {ID: "claude-sonnet-5", Name: "Claude Sonnet 5"}, - {ID: "claude-sonnet-4-6", Name: "Claude Sonnet 4.6"}, - {ID: "claude-sonnet-4-5", Name: "Claude Sonnet 4.5"}, - {ID: "claude-haiku-5-5", Name: "Claude Haiku 5.5"}, - {ID: "claude-haiku-4-5", Name: "Claude Haiku 4.5"}, - }, - Efforts: effortOptions(domain.ExecutorEffortLevels["claude-code"]), + Efforts: effortOptions(domain.ExecutorEffortLevels["claude-code"]), }, // Official-endpoint catalogs. OpenCode and Pi have none, so their // models come from the Base URL. @@ -122,33 +163,32 @@ func executorModelCatalogs() map[string]executorModelCatalog { }, Efforts: effortOptions(domain.ExecutorEffortLevels["agy"]), }, - "kimi": { - CustomModel: true, - Models: []executorModelInfo{ - {ID: "k3.1-pro", Name: "K3.1 Pro"}, - {ID: "k3.1-pro-highspeed", Name: "K3.1 Pro Highspeed"}, - {ID: "k3", Name: "K3"}, - {ID: "k3-256k", Name: "K3 256K"}, - {ID: "kimi-for-coding", Name: "K2.8 Preview"}, - {ID: "kimi-for-coding-highspeed", Name: "K2.7 Code Highspeed"}, - }, - }, - "pi": {CustomModel: true, Efforts: effortOptions(domain.ExecutorEffortLevels["pi"])}, - "dsh": { - CustomModel: true, - Models: []executorModelInfo{ - {ID: "deepseek-flash", Name: "DeepSeek V4.1 Flash"}, - {ID: "deepseek-v4-pro", Name: "DeepSeek V4 Pro"}, - }, - }, - "zcode": { - CustomModel: true, - Models: []executorModelInfo{ - {ID: "glm-5.3", Name: "GLM-5.3"}, - {ID: "glm-5.3-flash", Name: "GLM-5.3 Flash"}, - }, - }, + "kimi": {CustomModel: true, Efforts: effortOptions(domain.ExecutorEffortLevels["kimi"])}, + "pi": {CustomModel: true, Efforts: effortOptions(domain.ExecutorEffortLevels["pi"])}, + "dsh": {CustomModel: true, Efforts: effortOptions(domain.ExecutorEffortLevels["dsh"])}, + "zcode": {CustomModel: true, Efforts: effortOptions(domain.ExecutorEffortLevels["zcode"])}, } + for engine, catalog := range catalogs { + modelEngine := engine + if engine == "claude" { + modelEngine = assistantEngineClaude + } + if engine == runtime.AgentCLIKimi && kimiBaseURL != "" { + catalog.Efforts = effortOptions(nil) + } + for _, model := range builtinExecutorModels[modelEngine] { + catalog.Models = append(catalog.Models, executorModelInfo{ID: model.id, Name: model.name}) + } + for i := range catalog.Models { + id := catalog.Models[i].ID + if modelEngine != assistantEngineCodeBuddy { + id = modelEngine + ":" + id + } + catalog.Models[i].Efforts = effortOptions(assistantModelEffortsForBaseURL(id, nil, kimiBaseURL)) + } + catalogs[engine] = catalog + } + return catalogs } func cursorCatalogModels() []executorModelInfo { @@ -164,7 +204,7 @@ func cursorCatalogModels() []executorModelInfo { func (h *catalogHandlers) executorModels(c *gin.Context) { c.Header("Cache-Control", "private, no-store") if c.Query("builtin") == "1" { - c.JSON(http.StatusOK, executorModelCatalogs()) + c.JSON(http.StatusOK, executorModelCatalogs("")) return } var user domain.User @@ -174,9 +214,10 @@ func (h *catalogHandlers) executorModels(c *gin.Context) { refreshCLIModels(c.Request.Context(), user.CodeBuddyInternetEnvironment, user.CodeBuddyAPIKey) } } - catalogs := executorModelCatalogs() + _, kimiBaseURL := assistantEngineCredential(user, runtime.AgentCLIKimi) + catalogs := executorModelCatalogs(kimiBaseURL) for _, engine := range append([]string{assistantEngineCodeBuddy}, assistantCLIEngines()...) { - models := assistantDiscoveredModelsFor(c.Request.Context(), user, engine) + models, _ := assistantDiscoveredModelsFor(c.Request.Context(), user, engine) if len(models) == 0 { continue } @@ -187,14 +228,24 @@ func (h *catalogHandlers) executorModels(c *gin.Context) { catalog := catalogs[id] catalog.Source = "personal" catalog.Models = make([]executorModelInfo, 0, len(models)) + personal := cliModelsOf(models) for _, model := range models { + if engine == assistantEngineCodeBuddy && model.ID == assistantDefaultModel { + continue + } name := strings.TrimSpace(model.Name) if name == "" { name = model.ID } - catalog.Models = append(catalog.Models, executorModelInfo{ID: model.ID, Name: name}) + modelID := model.ID + if engine != assistantEngineCodeBuddy { + modelID = engine + ":" + modelID + } + catalog.Models = append(catalog.Models, executorModelInfo{ID: model.ID, Name: name, Efforts: effortOptions(assistantModelEffortsForBaseURL(modelID, personal, kimiBaseURL))}) + } + if len(catalog.Models) > 0 { + catalogs[id] = catalog } - catalogs[id] = catalog } c.JSON(http.StatusOK, catalogs) } diff --git a/internal/httpapi/platform_settings_test.go b/internal/httpapi/platform_settings_test.go index 0501519e..e54150cb 100644 --- a/internal/httpapi/platform_settings_test.go +++ b/internal/httpapi/platform_settings_test.go @@ -240,6 +240,9 @@ func stubCLIModelDiscovery(t *testing.T) { t.Helper() old := discoverCLIModels discoverCLIModels = func(context.Context, string, string) []string { return nil } + cliModelMu.Lock() + cliModelCache = map[string]cliModelEntry{} + cliModelMu.Unlock() t.Cleanup(func() { discoverCLIModels = old cliModelMu.Lock() diff --git a/internal/httpapi/runner_models.go b/internal/httpapi/runner_models.go index 6b2d729e..729c522d 100644 --- a/internal/httpapi/runner_models.go +++ b/internal/httpapi/runner_models.go @@ -70,5 +70,5 @@ func (h *runnersHandlers) probeRunnerModels(c *gin.Context) { c.JSON(http.StatusOK, cliModelProbeResult{Error: "账号没有返回可用模型,可以手动添加模型"}) return } - c.JSON(http.StatusOK, cliModelProbeResult{Models: models, Source: "runner"}) + c.JSON(http.StatusOK, cliModelProbeResult{Models: cliModelsWithEfforts(provider, models, ""), Source: "runner"}) } diff --git a/internal/repo/provider_credentials_test.go b/internal/repo/provider_credentials_test.go index b157895c..84cfd14a 100644 --- a/internal/repo/provider_credentials_test.go +++ b/internal/repo/provider_credentials_test.go @@ -5,6 +5,7 @@ import ( "encoding/json" "fmt" "os" + "slices" "strings" "testing" "time" @@ -94,8 +95,9 @@ func TestProviderUserCredentials(t *testing.T) { } u, _ := st.UserByLoginName(login) codexURL, claudeURL := "https://codex.example/v1", "https://claude.example" - codexModels := []domain.CLIModel{{ID: "gpt-5.5", Name: "gpt-5.5"}} - claudeModels := []domain.CLIModel{{ID: "claude-sonnet", Name: "Sonnet"}} + levels, defaultOnly := []string{" low ", "xhigh", "low"}, []string{} + codexModels := []domain.CLIModel{{ID: "gpt-5.5", Name: "gpt-5.5", ReasoningEfforts: &levels}} + claudeModels := []domain.CLIModel{{ID: "claude-sonnet", Name: "Sonnet", ReasoningEfforts: &defaultOnly}} _, err = st.UpdateUserCredentials(u.ID, UserCredentialUpdate{CodexAPIKey: "codex-secret", CodexBaseURL: &codexURL, CodexModels: &codexModels, ClaudeCodeAPIKey: "claude-secret", ClaudeCodeBaseURL: &claudeURL, ClaudeCodeModels: &claudeModels}) if err != nil { t.Fatal(err) @@ -112,6 +114,12 @@ func TestProviderUserCredentials(t *testing.T) { if !ok || got.CodexAPIKey != "codex-secret" || got.ClaudeCodeAPIKey != "claude-secret" || got.CodexBaseURL != codexURL || got.ClaudeCodeBaseURL != claudeURL || len(got.CodexModels) != 1 || got.CodexModels[0].ID != "gpt-5.5" || got.CodexModels[0].Name != "" || len(got.ClaudeCodeModels) != 1 || got.ClaudeCodeModels[0].Name != "Sonnet" { t.Fatalf("credentials did not survive refresh and partial update: %+v %+v", got.CodexModels, got.ClaudeCodeModels) } + if efforts := got.CodexModels[0].ReasoningEfforts; efforts == nil || !slices.Equal(*efforts, []string{"low", "xhigh"}) { + t.Fatalf("reasoning levels lost on refresh: %+v", got.CodexModels) + } + if efforts := got.ClaudeCodeModels[0].ReasoningEfforts; efforts == nil || len(*efforts) != 0 { + t.Fatal("explicit default-only model reverted to automatic reasoning levels") + } { var configRows int if err := st.DB.Raw(`SELECT COUNT(*) FROM user_configs WHERE user_id=?`, u.ID).Scan(&configRows).Error; err != nil { @@ -192,8 +200,9 @@ func TestCursorFamilyEffortsRoundTrip(t *testing.T) { t.Fatalf("new key kept the old catalog: %+v", got.AgentCLICredentials["cursor"]) } for name, bad := range map[string]domain.CLIModel{ - "unknown effort": {ID: "widget", Efforts: map[string]string{"turbo": "widget-turbo"}}, - "bad target": {ID: "widget", Efforts: map[string]string{"low": "widget low"}}, + "unknown effort": {ID: "widget", Efforts: map[string]string{"turbo": "widget-turbo"}}, + "bad target": {ID: "widget", Efforts: map[string]string{"low": "widget low"}}, + "unknown reasoning level": {ID: "widget", ReasoningEfforts: &[]string{"turbo"}}, } { if _, err := st.UpdateUserCredentials(user.ID, UserCredentialUpdate{AgentCLICredentials: map[string]AgentCLICredentialUpdate{ "cursor": {Models: &[]domain.CLIModel{bad}}, diff --git a/internal/repo/users.go b/internal/repo/users.go index c1739918..f1fdba34 100644 --- a/internal/repo/users.go +++ b/internal/repo/users.go @@ -534,11 +534,46 @@ func normalizeCLIModels(models []domain.CLIModel) ([]domain.CLIModel, error) { if err != nil { return nil, fmt.Errorf("模型 %s:%w", id, err) } - out = append(out, domain.CLIModel{ID: id, Name: name, Efforts: efforts, DefaultEffort: defaultEffort}) + reasoningEfforts, err := normalizeCLIReasoningEfforts(model.ReasoningEfforts) + if err != nil { + return nil, fmt.Errorf("模型 %s:%w", id, err) + } + discoveredEfforts, err := normalizeCLIReasoningEfforts(model.DiscoveredReasoningEfforts) + if err != nil { + return nil, fmt.Errorf("模型 %s:%w", id, err) + } + out = append(out, domain.CLIModel{ + ID: id, Name: name, Efforts: efforts, DefaultEffort: defaultEffort, + ReasoningEfforts: reasoningEfforts, DiscoveredReasoningEfforts: discoveredEfforts, + }) } return out, nil } +func normalizeCLIReasoningEfforts(raw *[]string) (*[]string, error) { + if raw == nil { + return nil, nil + } + levels := make([]string, 0, len(*raw)) + for _, level := range *raw { + level = strings.TrimSpace(level) + supported := false + for _, allowed := range domain.ExecutorEffortLevels { + if slices.Contains(allowed, level) { + supported = true + break + } + } + if !supported { + return nil, fmt.Errorf("不支持思考强度 %q", level) + } + if !slices.Contains(levels, level) { + levels = append(levels, level) + } + } + return &levels, nil +} + func validCLIModelID(id string) error { if id == "" { return errors.New("模型 ID 不能为空") diff --git a/internal/runtime/agent_cli.go b/internal/runtime/agent_cli.go index 4a1b4705..d683043f 100644 --- a/internal/runtime/agent_cli.go +++ b/internal/runtime/agent_cli.go @@ -80,8 +80,7 @@ func agentCLIModelOptional(kind, baseURL string) bool { type AgentCLIOptions struct { ApprovalCommand string Model string - // Effort is the reasoning effort chosen for Model; Cursor resolves the - // pair to one model ID with the owner's catalog. + // Effort is the reasoning effort chosen for Model. Effort string // SystemPromptFile is the system prompt path as the process sees it // (relative paths resolve against the working directory). @@ -93,6 +92,7 @@ type AgentCLIOptions struct { // agentCLIKeyEnv holds the user's key for CLIs whose configuration names // the variable instead of the secret. const agentCLIKeyEnv = "AGENTFLOW_CLI_API_KEY" +const kimiThinkingEffortEnv = "KIMI_MODEL_THINKING_EFFORT" // agentCLIHostEnv is cleared for every child: these variables would select // the host's own login, key, gateway or configuration directory. @@ -101,7 +101,7 @@ var agentCLIHostEnv = append([]string{ "AGY_ADC_AUTH", "GOOGLE_APPLICATION_CREDENTIALS", "GOOGLE_GENAI_USE_VERTEXAI", "GEMINI_HOME", "AGY_ACP_FORCE_FILE_STORAGE", "AGY_LLM_GATEWAY_URL", "AGY_LLM_GATEWAY_API_KEY", "AGY_LLM_GATEWAY_WIRE_PROTOCOL", "AGY_LLM_GATEWAY_MODELS", "DEEPSEEK_API_KEY", "DSH_HOME", "MOONSHOT_API_KEY", "KIMI_API_KEY", "KIMI_BASE_URL", "KIMI_CODE_HOME", - "KIMI_MODEL_NAME", "KIMI_MODEL_API_KEY", "KIMI_MODEL_BASE_URL", "KIMI_MODEL_PROVIDER_TYPE", + "KIMI_MODEL_NAME", "KIMI_MODEL_API_KEY", "KIMI_MODEL_BASE_URL", "KIMI_MODEL_PROVIDER_TYPE", kimiThinkingEffortEnv, "OPENCODE_CONFIG", "OPENCODE_CONFIG_CONTENT", "OPENCODE_CONFIG_DIR", "PI_CODING_AGENT_DIR", "ZCODE_DATA_BASE_DIR", zcodePersonalFileEnv, }, zcodeHostEnv...) @@ -462,7 +462,27 @@ func ApplyAgentCLICredential(kind string, cred ProviderCredential, options Agent } lines = append(lines, "api_key_env = "+strconv.Quote(agentCLIKeyEnv), "", "[models.agentflow]", `provider = "agentflow"`, "model = "+strconv.Quote(model), - "max_context_size = 262144", `capabilities = ["thinking", "tool_use"]`, "") + "max_context_size = 262144") + effort := strings.TrimSpace(options.Effort) + if err := domain.ValidateAgentEffort(effort, kind); err != nil { + return nil, nil, err + } + if effort != "" { + if baseURL != "" { + return nil, nil, fmt.Errorf("%s 的第三方网关使用默认思考强度", AgentCLILabel(kind)) + } + // Kimi's native provider applies this override on every turn, + // including resumes, without changing the session's saved default. + nativeEffort := effort + switch effort { + case "none": + nativeEffort = "off" + case "thinking": + nativeEffort = "on" + } + env[kimiThinkingEffortEnv] = nativeEffort + } + lines = append(lines, `capabilities = ["thinking", "tool_use"]`, "") if hook := options.ApprovalCommand; hook != "" { // Kimi validates 1–600 seconds and allows the tool on timeout; the // hook's own retry budget ends well before that. @@ -486,9 +506,9 @@ func ApplyAgentCLICredential(kind string, cred ProviderCredential, options Agent env[agentCLIKeyEnv] = key env["PI_CODING_AGENT_DIR"] = homeEnv + "/.pi/agent" case AgentCLIDeepSeek: - return applyDeepSeekCredential(home, homeEnv, key, baseURL, protocol, model, argv) + return applyDeepSeekCredential(home, homeEnv, key, baseURL, protocol, model, strings.TrimSpace(options.Effort), argv) case AgentCLIZCode: - return applyZCodeCredential(home, homeEnv, key, baseURL, protocol, model, argv) + return applyZCodeCredential(home, homeEnv, key, baseURL, protocol, model, strings.TrimSpace(options.Effort), argv) default: return nil, nil, fmt.Errorf("unknown agent CLI %q", kind) } @@ -503,24 +523,57 @@ func agentCLIPiAPI(protocol string) string { // applyDeepSeekCredential writes the DeepSeek Harness model patch and puts the // launcher --patch flags ahead of the profile's own arguments. -func applyDeepSeekCredential(home AgentCLIHome, homeEnv, key, baseURL, protocol, model string, argv []string) ([]string, map[string]string, error) { +func applyDeepSeekCredential(home AgentCLIHome, homeEnv, key, baseURL, protocol, model, effort string, argv []string) ([]string, map[string]string, error) { env := map[string]string{} patch := []any{} - if baseURL == "" { + selection := map[string]any{} + deepSeekModel := model == "deepseek-flash" || model == "deepseek-v4-pro" + if baseURL == "" && (deepSeekModel || model == "") && effort == "" { + // Headless resumes reuse the saved selection unless the patch names + // an effort. Restore the built-in route's default for a blank choice. + effort = "high" + } + if effort != "" { + if err := domain.ValidateAgentEffort(effort, AgentCLIDeepSeek); err != nil { + return nil, nil, err + } + if effort == "none" { + effort = "off" + } + selection["reasoningEffort"] = effort + } + if baseURL == "" || (deepSeekModel && protocol == AgentCLIProtocolMessages && effort != "") { env["DEEPSEEK_API_KEY"] = key + if baseURL != "" { + patch = append(patch, map[string]any{"id": "llm-deepseek", "config": map[string]any{"baseURL": baseURL}}) + } if model != "" { - patch = append(patch, map[string]any{"id": "agent-default-model", "config": map[string]any{"provider": "deepseek-official", "model": model}}) + selection["provider"] = "deepseek-official" + selection["model"] = model } } else { api := agentCLIPiAPI(protocol) env[agentCLIKeyEnv] = key - patch = []any{ - map[string]any{"id": "llm-pi-ai", "config": map[string]any{"providers": map[string]any{"agentflow": map[string]any{ - "displayName": "AgentFlow", "apiKeyEnv": agentCLIKeyEnv, "api": api, "baseURL": baseURL, - "models": []map[string]any{{"id": model}}, - }}}}, - map[string]any{"id": "agent-default-model", "config": map[string]any{"provider": "agentflow", "model": model}}, + provider := "agentflow" + modelConfig := map[string]any{"id": model} + if effort != "" { + // Keep the gateway's unconfigured route distinct so reverting to + // it on resume does not inherit the explicitly selected effort. + provider = "agentflow-reasoning" + modelConfig["reasoningEfforts"] = map[string]any{"off": nil, "low": "low", "high": "high", "max": "max"} + if api == "openai-completions" && deepSeekModel { + modelConfig["compat"] = map[string]any{"thinkingFormat": "deepseek"} + } } + patch = append(patch, map[string]any{"id": "llm-pi-ai", "config": map[string]any{"providers": map[string]any{provider: map[string]any{ + "displayName": "AgentFlow", "apiKeyEnv": agentCLIKeyEnv, "api": api, "baseURL": baseURL, + "models": []map[string]any{modelConfig}, + }}}}) + selection["provider"] = provider + selection["model"] = model + } + if len(selection) > 0 { + patch = append(patch, map[string]any{"id": "agent-default-model", "config": selection}) } data, _ := json.MarshalIndent(patch, "", " ") if err := writeAgentCLIFile(home, ".dsh/agentflow-model.json", data); err != nil { diff --git a/internal/runtime/agent_cli_deepseek_test.go b/internal/runtime/agent_cli_deepseek_test.go index c119be6e..2bc1c44f 100644 --- a/internal/runtime/agent_cli_deepseek_test.go +++ b/internal/runtime/agent_cli_deepseek_test.go @@ -69,3 +69,39 @@ func TestApplyDeepSeekCredentialRejectsShortArgv(t *testing.T) { t.Fatalf("err = %v", err) } } + +func TestApplyDeepSeekCredentialSetsTurnEffort(t *testing.T) { + home := t.TempDir() + for _, tc := range []struct { + baseURL, effort string + provider string + want any + }{ + {"", "none", "deepseek-official", "off"}, + {"", "max", "deepseek-official", "max"}, + {"", "", "deepseek-official", "high"}, + {"https://gw.example/v1", "max", "agentflow-reasoning", "max"}, + {"https://gw.example/v1", "", "agentflow", nil}, + } { + command := AgentCLICommand(AgentCLIDeepSeek, AgentCLICommandOptions{Model: "deepseek-flash", Effort: tc.effort, Resume: "session", Prompt: "probe"}) + if _, _, err := ApplyAgentCLICredential(AgentCLIDeepSeek, ProviderCredential{APIKey: "k", BaseURL: tc.baseURL}, + command.AgentCLI, HostAgentCLIHome(home), home, command.Argv); err != nil { + t.Fatal(err) + } + data, err := os.ReadFile(filepath.Join(home, ".dsh", "agentflow-model.json")) + if err != nil { + t.Fatal(err) + } + var patch []struct { + ID string `json:"id"` + Config map[string]any `json:"config"` + } + if err = json.Unmarshal(data, &patch); err != nil || len(patch) == 0 { + t.Fatalf("invalid model patch: %s (%v)", data, err) + } + selection := patch[len(patch)-1] + if selection.ID != "agent-default-model" || selection.Config["provider"] != tc.provider || selection.Config["reasoningEffort"] != tc.want { + t.Errorf("base %q and effort %q: wrong model selection in %s", tc.baseURL, tc.effort, data) + } + } +} diff --git a/internal/runtime/agent_cli_test.go b/internal/runtime/agent_cli_test.go index 9bd1f3f9..c5f6f25b 100644 --- a/internal/runtime/agent_cli_test.go +++ b/internal/runtime/agent_cli_test.go @@ -71,6 +71,36 @@ func TestAgentCLICommandChatRestrictions(t *testing.T) { } } +func TestKimiThinkingConfiguration(t *testing.T) { + home := t.TempDir() + for _, tc := range []struct { + effort string + thinking string + }{ + {"none", "off"}, + {"thinking", "on"}, + {"max", "max"}, + {"", ""}, + } { + command := AgentCLICommand(AgentCLIKimi, AgentCLICommandOptions{Model: "fixture-model", Effort: tc.effort, Resume: "session", Prompt: "test"}) + _, env, err := ApplyAgentCLICredential(AgentCLIKimi, ProviderCredential{APIKey: "fixture-key"}, + command.AgentCLI, HostAgentCLIHome(home), home, command.Argv) + if err != nil { + t.Fatal(err) + } + if got := env[kimiThinkingEffortEnv]; got != tc.thinking { + t.Errorf("effort %q yields per-turn override %q, want %q", tc.effort, got, tc.thinking) + } + config, err := os.ReadFile(filepath.Join(home, ".kimi-code", "config.toml")) + if err != nil { + t.Fatal(err) + } + if strings.Contains(string(config), "[thinking]") { + t.Errorf("effort %q was written into the Kimi session defaults:\n%s", tc.effort, config) + } + } +} + func TestAgentCLICommandAgyEffortSuffix(t *testing.T) { for _, tc := range []struct{ model, effort, want string }{ {"gemini-3.8-flash-medium", "low", "gemini-3.8-flash-low"}, diff --git a/internal/runtime/agent_cli_zcode.go b/internal/runtime/agent_cli_zcode.go index 02b95469..2fccfcc5 100644 --- a/internal/runtime/agent_cli_zcode.go +++ b/internal/runtime/agent_cli_zcode.go @@ -7,6 +7,8 @@ import ( "io/fs" "reflect" "strings" + + "agentflow/internal/domain" ) const ( @@ -71,7 +73,7 @@ func nonEmpty(value any) bool { // applyZCodeCredential gives ZCode the key as a personal provider file that // exists only until ZCode has read it: ZCode accepts a literal key and // nothing else, and the key must not stay on disk. -func applyZCodeCredential(home AgentCLIHome, homeEnv, key, baseURL, protocol, model string, argv []string) ([]string, map[string]string, error) { +func applyZCodeCredential(home AgentCLIHome, homeEnv, key, baseURL, protocol, model, effort string, argv []string) ([]string, map[string]string, error) { // The private provider has no model of its own, so "CLI default" (sent as // "default") must name a real one: BigModel's coding model on the default // endpoint, while a custom endpoint has no model we could guess. @@ -85,6 +87,32 @@ func applyZCodeCredential(home AgentCLIHome, homeEnv, key, baseURL, protocol, mo if baseURL == "" { baseURL, apiType = zcodeDefaultBaseURL, "anthropic-messages" } + modelRules := []any{} + if effort != "" { + if err := domain.ValidateAgentEffort(effort, AgentCLIZCode); err != nil { + return nil, nil, err + } + // A resumed session keeps its original selection. Override the + // model's per-call mapping so the current turn uses the chosen effort, + // and removing the rule restores that session's native default. + wire := map[string]any{} + switch apiType { + case "anthropic-messages": + wire["thinking"] = map[string]string{"type": "enabled"} + wire["output_config"] = map[string]string{"effort": effort} + case "openai-chat-completions": + wire["thinking"] = map[string]string{"type": "enabled"} + wire["reasoning_effort"] = effort + wire["reasoning"] = map[string]string{"effort": effort} + case "openai-responses": + wire["reasoning"] = map[string]string{"effort": effort} + } + mapping, _ := json.Marshal(wire) + modelRules = append(modelRules, map[string]any{ + "providerId": "agentflow", "modelId": model, + "config": map[string]any{"optionSpecs": map[string]any{"reasoningLevel": map[string]string{"map": string(mapping)}}}, + }) + } providerConfig, err := json.MarshalIndent(map[string]any{"schemaVersion": 1, "config": map[string]any{ "providerConfigRules": map[string]any{"providerRules": []any{map[string]any{"providerId": "agentflow", "config": map[string]any{ "group": "standard-personal", @@ -92,7 +120,7 @@ func applyZCodeCredential(home AgentCLIHome, homeEnv, key, baseURL, protocol, mo "api": map[string]any{"type": apiType, "baseUrl": baseURL}, "personalModelIds": []string{model}, }}}}, - "modelConfigRules": map[string]any{"providerModelRules": []any{}, "manualProviderModelRules": []any{}}, + "modelConfigRules": map[string]any{"providerModelRules": modelRules, "manualProviderModelRules": []any{}}, "defaultModelSelection": map[string]any{"providerId": "agentflow", "modelId": model}, }}, "", " ") if err != nil { diff --git a/internal/runtime/agent_cli_zcode_test.go b/internal/runtime/agent_cli_zcode_test.go index 7b8d04cc..e9c8ace0 100644 --- a/internal/runtime/agent_cli_zcode_test.go +++ b/internal/runtime/agent_cli_zcode_test.go @@ -82,7 +82,7 @@ func TestApplyZCodeCredential(t *testing.T) { t.Fatal(err) } argv := []string{"zcode", "--prompt", "it's \"quoted\" $HOME"} - bound, env, err := ApplyAgentCLICredential(AgentCLIZCode, ProviderCredential{APIKey: " secret-key "}, AgentCLIOptions{Model: "glm-5.3"}, HostAgentCLIHome(home), "/w/home", argv) + bound, env, err := ApplyAgentCLICredential(AgentCLIZCode, ProviderCredential{APIKey: " secret-key "}, AgentCLIOptions{Model: "glm-5.3", Effort: "max"}, HostAgentCLIHome(home), "/w/home", argv) if err != nil { t.Fatal(err) } @@ -110,9 +110,15 @@ func TestApplyZCodeCredential(t *testing.T) { config["personalModelIds"].([]any)[0] != "glm-5.3" { t.Fatalf("provider config: %v", provider) } - if selection := rules["defaultModelSelection"].(map[string]any); selection["providerId"] != "agentflow" || selection["modelId"] != "glm-5.3" { + selection := rules["defaultModelSelection"].(map[string]any) + if selection["providerId"] != "agentflow" || selection["modelId"] != "glm-5.3" || selection["options"] != nil { t.Fatalf("selection: %v", selection) } + modelRule := rules["modelConfigRules"].(map[string]any)["providerModelRules"].([]any)[0].(map[string]any) + configured := modelRule["config"].(map[string]any)["optionSpecs"].(map[string]any)["reasoningLevel"].(map[string]any) + if modelRule["modelId"] != "glm-5.3" || configured["map"] != `{"output_config":{"effort":"max"},"thinking":{"type":"enabled"}}` { + t.Fatalf("reasoning rule: %v", modelRule) + } // The cleanup hook joins the MCP configuration without replacing it, and // the key itself never reaches config.json. diff --git a/internal/runtime/claude_install.go b/internal/runtime/claude_install.go index ae06ae65..40d3e19d 100644 --- a/internal/runtime/claude_install.go +++ b/internal/runtime/claude_install.go @@ -13,7 +13,7 @@ import ( // ClaudeCodeVersion is installed from npm on first use. Published platform // images do not redistribute Claude Code, so each installation downloads it -// from Anthropic's registry package itself. +// from Anthropic's registry package itself. Fable 5.2 requires 2.1.321 or later. const ClaudeCodeVersion = "2.1.326" var claudeInstallMu sync.Mutex diff --git a/internal/runtime/claude_install_test.go b/internal/runtime/claude_install_test.go index f5837c32..d1a85f0b 100644 --- a/internal/runtime/claude_install_test.go +++ b/internal/runtime/claude_install_test.go @@ -16,7 +16,7 @@ printf '%s\n' "$*" >> "$NPM_TEST_LOG" [ "$NPM_TEST_FAIL" = 1 ] && { echo 'npm ERR! registry unreachable'; exit 1; } prefix=$3 mkdir -p "$prefix/node_modules/@anthropic-ai/claude-code/bin" "$prefix/node_modules/.bin" -printf '#!/bin/sh\necho 2.1.300\n' > "$prefix/node_modules/@anthropic-ai/claude-code/bin/claude.exe" +printf '#!/bin/sh\necho ` + ClaudeCodeVersion + `\n' > "$prefix/node_modules/@anthropic-ai/claude-code/bin/claude.exe" chmod 755 "$prefix/node_modules/@anthropic-ai/claude-code/bin/claude.exe" ln -s ../@anthropic-ai/claude-code/bin/claude.exe "$prefix/node_modules/.bin/claude" ` diff --git a/internal/runtime/provider_agentcli.go b/internal/runtime/provider_agentcli.go index a728d013..372419aa 100644 --- a/internal/runtime/provider_agentcli.go +++ b/internal/runtime/provider_agentcli.go @@ -119,7 +119,7 @@ func AgentCLICommand(kind string, o AgentCLICommandOptions) CommandSpec { extra := append([]string{}, o.Extra...) var argv []string readOnly := o.Chat && !o.Workspace - options := AgentCLIOptions{Model: model, ReadOnly: readOnly} + options := AgentCLIOptions{Model: model, Effort: effort, ReadOnly: readOnly} switch kind { case AgentCLIOpenCode: argv = []string{binary, "run", "--format", "json", "--thinking"} @@ -142,7 +142,6 @@ func AgentCLICommand(kind string, o AgentCLICommandOptions) CommandSpec { // ApplyAgentCLICredential re-resolves this against the owner's // saved catalog, which only the credential carries. argv = append(argv, "--model", CursorModelID(model, effort, nil)) - options.Effort = effort } if resume != "" { argv = append(argv, "--resume", resume) diff --git a/web/app/agents.js b/web/app/agents.js index 1371913d..0a3bf917 100644 --- a/web/app/agents.js +++ b/web/app/agents.js @@ -149,7 +149,9 @@ function agentModelItems(executorId,current){ } function agentEffortItems(executorId){ const cat=executorModelCatalogs()[executorId]||{}; - return [{id:'',name:'默认',icon:false}].concat((cat.efforts||[]).map(e=>({id:e.value,name:e.label,icon:false}))); + const choice=$('#agentModelPicker [name="modelChoice"]')?.value,model=choice===AGENT_CUSTOM_MODEL?$('#agentModelCustomField input[name="model"]')?.value:choice; + const selected=(cat.models||[]).find(item=>item.id===model),efforts=Array.isArray(selected?.efforts)?selected.efforts:(cat.efforts||[]); + return [{id:'',name:'默认',icon:false}].concat(efforts.map(e=>({id:e.value,name:e.label,icon:false}))); } function paintAgentPicker(pickerId,item){ const picker=$(pickerId);if(!picker)return; @@ -175,6 +177,7 @@ function renderAgentModelPicker(){ picker.querySelector('[name="modelChoice"]').value=id; syncAgentModelCustomField(); paintAgentPicker('#agentModelPicker',items.find(x=>x.id===id)); + renderAgentEffortPicker(id!==current); menu.hidden=true;picker.querySelector('[data-model-toggle]').setAttribute('aria-expanded','false'); }); syncAgentModelCustomField(); @@ -185,17 +188,19 @@ function syncAgentModelCustomField(){ if(!picker||!field)return; field.hidden=picker.querySelector('[name="modelChoice"]').value!==AGENT_CUSTOM_MODEL; } -function renderAgentEffortPicker(){ +function renderAgentEffortPicker(modelChanged=false){ const picker=$('#agentEffortPicker');if(!picker)return; const items=agentEffortItems(currentAgentExecutorId()); - const current=picker.querySelector('[name="effort"]').value; + const field=picker.querySelector('[name="effort"]'); + if(modelChanged&&executorModelCatalogs()[currentAgentExecutorId()]&&!items.some(item=>item.id===field.value))field.value=''; + const current=field.value; const menu=picker.querySelector('.ai-model-menu'); agentModelAPI()?.renderOptions?.(menu,items,current,id=>{ picker.querySelector('[name="effort"]').value=id; paintAgentPicker('#agentEffortPicker',items.find(x=>x.id===id)); menu.hidden=true;picker.querySelector('[data-model-toggle]').setAttribute('aria-expanded','false'); }); - paintAgentPicker('#agentEffortPicker',items.find(x=>x.id===current)||items[0]); + paintAgentPicker('#agentEffortPicker',items.find(x=>x.id===current)||{id:current,name:current,icon:false}); } function applyAgentModelUI(executorId,model,effort){ const picker=$('#agentModelPicker');if(!picker||!$('#agentEffortPicker'))return; @@ -206,8 +211,7 @@ function applyAgentModelUI(executorId,model,effort){ const customInput=$('#agentModelCustomField')?.querySelector('input[name="model"]'); if(customInput&&choice===AGENT_CUSTOM_MODEL)customInput.value=model||''; renderAgentModelPicker(); - const efforts=cat.efforts||[]; - $('#agentEffortPicker').querySelector('[name="effort"]').value=(!cat.efforts||efforts.some(e=>e.value===effort))?(effort||''):''; + $('#agentEffortPicker').querySelector('[name="effort"]').value=effort||''; renderAgentEffortPicker(); } function bindAgentPicker(pickerId,render){ diff --git a/web/app/assistant-models.js b/web/app/assistant-models.js index 4d7dbfec..e31b75fd 100644 --- a/web/app/assistant-models.js +++ b/web/app/assistant-models.js @@ -20,7 +20,7 @@ }; let catalog = { default: DEFAULT_MODEL, - models: [{id: DEFAULT_MODEL, name: 'CLI 默认模型', family: 'other'}] + models: [] }; let loading = null; @@ -48,7 +48,8 @@ return (source.engines || []).some(e => e.id === prefix) ? prefix : 'codebuddy'; } function modelMeta(id) { - return (catalog.models || []).find(m => m.id === id) || {id: id || DEFAULT_MODEL, family: familyOf(id), name: id || DEFAULT_MODEL}; + const value = id || DEFAULT_MODEL; + return catalog.models.find(m => m.id === value) || {id: value, family: familyOf(value), name: value === DEFAULT_MODEL ? '默认' : value}; } function fillModelIcon(el, family, id) { if (!el) return; @@ -145,12 +146,14 @@ if (!resp.ok) throw new Error(`模型列表加载失败(${resp.status})`); return resp.json(); } - // load keeps one in-flight request and keeps the built-in fallback when the - // catalog cannot be read, so pickers never spin forever on a cold start. + // Share the in-flight request and retain the last catalog on a refresh failure. function load() { if (!loading) { loading = fetchCatalog().then(data => { - if (data?.models?.length) catalog = {default: data.default || DEFAULT_MODEL, models: data.models, engines: data.engines || []}; + if (Array.isArray(data?.models)) { + const models = data.models.filter(m => m.id !== DEFAULT_MODEL || (m.engine && m.engine !== 'codebuddy')); + catalog = {default: data.default || DEFAULT_MODEL, models, engines: data.engines || []}; + } return catalog; }).catch(() => catalog).finally(() => { loading = null; }); } diff --git a/web/app/runners.js b/web/app/runners.js index 4f193f8e..38449d44 100644 --- a/web/app/runners.js +++ b/web/app/runners.js @@ -382,6 +382,7 @@ function showRunnerModels(runner,runtime){ modal.querySelector('.confirm-modal-footer').before(panel); card.classList.add('runner-model-card'); const catalog=panel.querySelector('[data-credential-catalog]'); + loadCredentialEffortCatalogs().then(()=>refreshCredentialCatalogEfforts(catalog)); if(local){ catalog.probeModels=()=>api(`/api/runners/${encodeURIComponent(runner.id)}/models/probe`,{ method:'POST',body:JSON.stringify({provider}),silent:true @@ -396,6 +397,7 @@ function showRunnerModels(runner,runtime){ catalog.oninput=event=>{ if(event.target.closest('[data-catalog-probe]'))return syncCredentialCatalogProbe(catalog); event.target.closest('.credential-input')?.classList.remove('is-bad'); + syncCredentialCatalogEfforts(event.target.closest('[data-catalog-row]')); syncSave(); }; catalog.onclick=event=>{ diff --git a/web/app/sync-schedule.js b/web/app/sync-schedule.js index e24d844f..98928edd 100644 --- a/web/app/sync-schedule.js +++ b/web/app/sync-schedule.js @@ -169,10 +169,10 @@ function paintScheduleEngines(selected){ const engines=scheduleModelCatalog().engines||[]; const list=engines.length?[...engines]:[{id:'codebuddy',name:'CodeBuddy'}]; if(!list.some(e=>e.id===current))list.push({id:current,name:current}); - // A CLI without listed models lacks credentials in this account. A saved task - // keeps its CLI selectable so editing it does not move it elsewhere. + // CodeBuddy can run on its native default with no concrete model in the + // list. A saved task keeps its CLI selectable during edits. select.replaceChildren(...list.map(e=>{ - const usable=e.id===current||scheduleEngineModels(e.id).length>0; + const usable=e.id===current||(e.id==='codebuddy'&&e.configured)||scheduleEngineModels(e.id).length>0; const option=new Option(usable?e.name:`${e.name}(缺少凭据)`,e.id); option.disabled=!usable; return option; @@ -228,8 +228,7 @@ function bindScheduleModelPicker(){ if(engine)engine.onchange=()=>{ const api=scheduleModelAPI(); const models=scheduleEngineModels(engine.value); - const fallback=engine.value==='codebuddy'?(scheduleModelCatalog().default||api?.DEFAULT):''; - const next=models.some(m=>m.id===fallback)?fallback:models[0]?.id; + const next=models[0]?.id||(engine.value==='codebuddy'?(scheduleModelCatalog().default||api?.DEFAULT):''); if(next)paintScheduleModel(next); }; document.addEventListener('click',e=>{if(!host.contains(e.target))closeScheduleModelMenu()}); diff --git a/web/app/workbench.js b/web/app/workbench.js index 2ef068a5..b00a85f6 100644 --- a/web/app/workbench.js +++ b/web/app/workbench.js @@ -143,15 +143,32 @@ function credentialModelsEndpoint(kind,protocol){ // (or the protocol is Gemini), and the Base URL's published models otherwise. // Edits stay in the DOM until 保存. const CREDENTIAL_CATALOG_MAX=100; +let credentialEffortCatalogs,credentialEffortCatalogRequest; +async function loadCredentialEffortCatalogs(){ + if(credentialEffortCatalogs)return; + if(!credentialEffortCatalogRequest)credentialEffortCatalogRequest=api('/api/executor-models?builtin=1',{silent:true}).then(data=>{credentialEffortCatalogs=data}).catch(()=>{}).finally(()=>{credentialEffortCatalogRequest=null}); + await credentialEffortCatalogRequest; +} function credentialCatalogModels(value){ return (Array.isArray(value)?value:[]).map(model=>{ const id=String(model?.id||'').trim(),name=String(model?.name||'').trim(); - const metadata=model?.efforts?{efforts:model.efforts,defaultEffort:model.defaultEffort}:{}; + const metadata={ + ...(model?.efforts?{efforts:model.efforts,defaultEffort:model.defaultEffort}:{}), + ...(Array.isArray(model?.reasoningEfforts)?{reasoningEfforts:model.reasoningEfforts}:{}), + ...(Array.isArray(model?.discoveredReasoningEfforts)?{discoveredReasoningEfforts:model.discoveredReasoningEfforts}:{}), + }; return {id,name:name===id?'':name,...metadata}; }).filter(model=>model.id||model.name); } function credentialCatalogStored(models){return credentialCatalogModels(models).map(({name,...model})=>name?{...model,name}:model)} -function credentialCatalogKey(models){return JSON.stringify(credentialCatalogModels(models).map(model=>[model.id,model.name]))} +function credentialCatalogKey(models){ + return JSON.stringify(credentialCatalogModels(models).map(model=>[ + model.id,model.name, + model.reasoningEfforts?.slice().sort()??null, + model.discoveredReasoningEfforts?.slice().sort()??null, + Object.entries(model.efforts||{}).sort(([a],[b])=>a.localeCompare(b)),model.defaultEffort||'', + ])); +} function credentialCatalogError(models){ const seen=new Set(); if(models.length>CREDENTIAL_CATALOG_MAX)return {index:CREDENTIAL_CATALOG_MAX,error:`模型目录最多 ${CREDENTIAL_CATALOG_MAX} 个`}; @@ -166,9 +183,46 @@ function credentialCatalogError(models){ } return null; } -function credentialCatalogRowHTML(model={}){ - const metadata=model.efforts?` data-catalog-metadata="${escapeHTML(JSON.stringify({id:model.id,efforts:model.efforts,defaultEffort:model.defaultEffort}))}"`:''; - return `
  • `; +function credentialCatalogEffortSummary(model){ + if(model.reasoningEfforts)return model.reasoningEfforts.join(' / ')||'仅使用默认'; + if(!model.discoveredReasoningEfforts)return '自动'; + return '自动 · '+(model.discoveredReasoningEfforts.join(' / ')||'仅使用默认'); +} +function credentialCatalogEffortHTML(model,engine){ + const available=credentialEffortCatalogs?.[engine==='claude-code'?'claude':engine]?.efforts||[]; + const knownLevels=new Set([...(model.reasoningEfforts||[]),...(model.discoveredReasoningEfforts||[])]); + const additional=[...knownLevels].filter(value=>!available.some(option=>option.value===value)); + const options=available.concat(additional.map(value=>({value,label:value}))); + if(!options.length&&!model.discoveredReasoningEfforts)return ''; + const custom=Array.isArray(model.reasoningEfforts),selected=model.reasoningEfforts??model.discoveredReasoningEfforts??[]; + return `
    思考强度${escapeHTML(credentialCatalogEffortSummary(model))}
    ${options.map(({value,label})=>``).join('')}勾选此模型支持的强度;全部取消时仅使用模型默认强度。
    `; +} +function syncCredentialCatalogEfforts(row){ + if(!row)return; + const mode=row.querySelector('[data-catalog-effort-mode]');if(!mode)return; + row.querySelector('[data-catalog-effort-levels]').hidden=mode.value!=='custom'; + row.querySelector('[data-catalog-effort-summary]').textContent=credentialCatalogEffortSummary(credentialCatalogRowModel(row)); +} +function refreshCredentialCatalogEfforts(catalog){ + if(!catalog?.isConnected)return; + credentialCatalogRows(catalog).forEach(row=>{ + const html=credentialCatalogEffortHTML(credentialCatalogRowModel(row),catalog.dataset.catalogEngine),levels=row.querySelector('[data-catalog-effort-levels]'); + if(!levels){row.insertAdjacentHTML('beforeend',html);return} + // Capabilities may arrive after a row was expanded. Add missing choices + // without replacing focused controls or the user's unsaved selections. + const holder=document.createElement('div');holder.innerHTML=html; + const options=[...holder.querySelectorAll('[data-catalog-effort]')]; + const labels=new Map([...levels.querySelectorAll('[data-catalog-effort]')].map(input=>[input.dataset.catalogEffort,input.closest('label')])); + options.forEach((input,index)=>{ + if(labels.has(input.dataset.catalogEffort))return; + const next=options.slice(index+1).map(option=>labels.get(option.dataset.catalogEffort)).find(Boolean)||levels.querySelector('small'); + const label=input.closest('label');levels.insertBefore(label,next);labels.set(input.dataset.catalogEffort,label); + }); + }); +} +function credentialCatalogRowHTML(model={},engine=''){ + const metadata=` data-catalog-metadata="${escapeHTML(JSON.stringify(credentialCatalogModels([model])[0]||{}))}"`; + return `
  • ${credentialCatalogEffortHTML(model,engine)}
  • `; } function credentialCatalogCopy(mode,kind){ const name=kind?.name||'该 CLI'; @@ -187,11 +241,11 @@ function credentialCatalogHTML(kind,user,{runnerLogin=false,label='模型目录' const models=credentialCatalogModels(cursorAutomatic?[]:credentialValue(user,kind.models.name)),empty=!models.length; const mode=runnerLogin?'runner':!String(baseUrl||'').trim()||protocol==='gemini'?'official':'remote'; const copy=credentialCatalogCopy(mode,kind); - return `
    + return `
    ${label?`${escapeHTML(label)}`:''}${credentialCatalogCountText(models.length)}
    -
      ${models.map(credentialCatalogRowHTML).join('')}
    +
      ${models.map(model=>credentialCatalogRowHTML(model,kind.models.engine)).join('')}
    ${credentialGlyph('layers',16)}
    ${escapeHTML(copy.title)}${credentialCatalogDetailHTML(mode,kind,protocol)}
    ${escapeHTML(copy.reset)}
    @@ -204,7 +258,12 @@ function credentialCatalogRows(catalog){return [...catalog.querySelectorAll('[da function credentialCatalogRowModel(row){ const id=row.querySelector('[data-catalog-field="id"]').value,name=row.querySelector('[data-catalog-field="name"]').value; const metadata=row.dataset.catalogMetadata?JSON.parse(row.dataset.catalogMetadata):null; - return metadata?.id===id.trim()?{...metadata,id,name}:{id,name}; + const model=metadata?.id===id.trim()?{...metadata,id,name}:{id,name},mode=row.querySelector('[data-catalog-effort-mode]'); + if(mode){ + delete model.reasoningEfforts; + if(mode.value==='custom')model.reasoningEfforts=[...row.querySelectorAll('[data-catalog-effort]:checked')].map(input=>input.dataset.catalogEffort); + } + return model; } function credentialCatalogFilledRows(catalog){return credentialCatalogRows(catalog).filter(row=>credentialCatalogModels([credentialCatalogRowModel(row)]).length)} function readCredentialCatalog(catalog){return credentialCatalogModels(credentialCatalogRows(catalog).map(credentialCatalogRowModel))} @@ -251,7 +310,7 @@ function resetCredentialCatalog(row){ } function addCredentialCatalogRow(catalog,model={}){ const list=catalog.querySelector('[data-catalog-list]'); - list.insertAdjacentHTML('beforeend',credentialCatalogRowHTML(model)); + list.insertAdjacentHTML('beforeend',credentialCatalogRowHTML(model,catalog.dataset.catalogEngine)); const row=list.lastElementChild; hydrateCredentialIcons(row); credentialCatalogChanged(catalog); @@ -367,14 +426,21 @@ function applyCredentialCatalogProbe(catalog){ const probe=catalog.probe;if(!probe)return; const picked=new Set([...catalog.querySelectorAll('[data-catalog-probe] input[type="checkbox"]:checked')].map(input=>input.value)); const offered=new Set(probe.models.map(model=>model.id)),kept=new Set(); + const discovered=new Map(probe.models.map(model=>[model.id,model])); let removed=0; credentialCatalogRows(catalog).forEach(row=>{ const id=row.querySelector('[data-catalog-field="id"]').value.trim(); - if(offered.has(id)&&!picked.has(id)){row.remove();removed++}else kept.add(id); + if(offered.has(id)&&!picked.has(id)){row.remove();removed++}else{ + kept.add(id); + if(picked.has(id)&&discovered.has(id)){ + const current=credentialCatalogRowModel(row),fresh=discovered.get(id); + row.outerHTML=credentialCatalogRowHTML({...current,...fresh,name:current.name,reasoningEfforts:current.reasoningEfforts},catalog.dataset.catalogEngine); + } + } }); const added=probe.models.filter(model=>picked.has(model.id)&&!kept.has(model.id)); const room=Math.max(0,CREDENTIAL_CATALOG_MAX-kept.size),list=catalog.querySelector('[data-catalog-list]'); - list.insertAdjacentHTML('beforeend',credentialCatalogModels(added.slice(0,room)).map(credentialCatalogRowHTML).join('')); + list.insertAdjacentHTML('beforeend',credentialCatalogModels(added.slice(0,room)).map(model=>credentialCatalogRowHTML(model,catalog.dataset.catalogEngine)).join('')); hydrateCredentialIcons(list); closeCredentialCatalogProbe(catalog); credentialCatalogChanged(catalog); @@ -910,6 +976,7 @@ function bindCredentialPanel(panel){ if(catalog){ if(event.target.closest('[data-catalog-probe]'))return syncCredentialCatalogProbe(catalog); event.target.closest('.credential-input')?.classList.remove('is-bad'); + syncCredentialCatalogEfforts(event.target.closest('[data-catalog-row]')); } if(event.target.name==='extra'||event.target.name==='protocol'){const catalogBox=form.querySelector('[data-credential-catalog]');if(catalogBox)syncCredentialCatalog(catalogBox)} if(event.target.name==='secret'){ @@ -1237,6 +1304,7 @@ async function renderPersonalSettings(){ const panel=host.querySelector('#credentialPanel'); bindCredentialPanel(panel); renderCredentialPanel(); + loadCredentialEffortCatalogs().then(()=>panel.querySelectorAll('[data-credential-catalog]').forEach(refreshCredentialCatalogEfforts)); hydrateCredentialIcons(host); setTimeout(()=>panel.classList.remove('is-entering'),900); // Offer the OAuth shortcut only when the server has a Gongfeng application. diff --git a/web/assistant_models_test.js b/web/assistant_models_test.js index 76b4285b..f6fea0e4 100644 --- a/web/assistant_models_test.js +++ b/web/assistant_models_test.js @@ -70,6 +70,7 @@ const ctx = { json: async () => ({ default: 'model-b', models: [ + { id: 'default', name: 'CLI 默认模型', family: 'other' }, { id: 'model-a', name: 'Model A', family: 'claude', credits: 'x1.00' }, { id: 'model-b', name: 'Model B', family: 'glm', credits: 'x0.06' }, { id: 'codex:gpt-5.5', name: 'GPT-5.5', family: 'gpt', engine: 'codex' }, @@ -95,8 +96,9 @@ vm.runInContext(moduleSource, ctx); // must already render a usable label. assert.equal(api.codeBuddyModel('codex:gpt-5.5'), 'default', 'codex ids fall back before the catalog loads'); api.paint(toggle, api.DEFAULT); - assert.equal(label.textContent, 'CLI 默认模型'); + assert.equal(label.textContent, '默认'); assert.equal(icon.dataset.family, 'other'); + assert.equal(api.codeBuddyModels().length, 0, 'the placeholder is not selectable before loading'); await api.load(); assert.equal(api.catalog().default, 'model-b'); diff --git a/web/orchestrate.js b/web/orchestrate.js index aa718bfb..4cbab417 100644 --- a/web/orchestrate.js +++ b/web/orchestrate.js @@ -11,7 +11,10 @@ function aiOrchestratorEngine(model) { } function selectedAIOrchestratorModel() { - return localStorage.getItem('agentflow.orchestratorModel') || localStorage.getItem('agentflow.assistantModel') || window.AgentFlowAssistantModels?.catalog().default || ''; + const api = window.AgentFlowAssistantModels; + const remembered = localStorage.getItem('agentflow.orchestratorModel') || localStorage.getItem('agentflow.assistantModel'); + if (remembered && remembered !== api?.DEFAULT) return remembered; + return api?.codeBuddyModels()[0]?.id || api?.catalog().default || remembered || ''; } function aiOrchestratorModels(engine = aiOrchestratorEngine(selectedAIOrchestratorModel())) { @@ -53,10 +56,14 @@ function bindAIOrchestratorModelPicker() { const search = host.querySelector('[data-model-search]'); const list = host.querySelector('[role="listbox"]'); const selected = () => cli ? aiOrchestratorEngine(selectedAIOrchestratorModel()) : selectedAIOrchestratorModel(); - const items = () => cli ? aiOrchestratorEngines().map(engine => ({...engine, icon: false, disabled: engine.configured === false || !aiOrchestratorModels(engine.id).length})) : aiOrchestratorModels(); + const items = () => cli ? aiOrchestratorEngines().map(engine => ({ + ...engine, + icon: false, + disabled: engine.configured === false || (engine.id !== 'codebuddy' && !aiOrchestratorModels(engine.id).length), + })) : aiOrchestratorModels(); const paint = () => { const id = selected(); - const item = items().find(item => item.id === id) || {id, name: id}; + const item = items().find(item => item.id === id) || {id, name: !cli && id === api.DEFAULT ? '默认' : id}; api.paintItem(toggle, item); if (cli) paintCLIIcon(toggle.querySelector('.ai-model-icon'), id); }; @@ -79,7 +86,9 @@ function bindAIOrchestratorModelPicker() { if (cli) { const models = aiOrchestratorModels(id); const remembered = localStorage.getItem('agentflow.orchestratorModel.' + id); - model = models.find(item => item.id === (id === selected() ? selectedAIOrchestratorModel() : remembered))?.id || models.find(item => item.id === api.catalog().default)?.id || models[0]?.id; + const preference = id === selected() ? selectedAIOrchestratorModel() : remembered; + const fallback = models[0]?.id || (id === 'codebuddy' ? api.catalog().default || api.DEFAULT : ''); + model = models.find(item => item.id === preference)?.id || fallback; } if (!model) return; selectAIOrchestratorModel(model); diff --git a/web/orchestrate_models_test.js b/web/orchestrate_models_test.js index 5c815b42..bcb0178b 100644 --- a/web/orchestrate_models_test.js +++ b/web/orchestrate_models_test.js @@ -143,5 +143,13 @@ dom.run(` request = await send(); assert.equal(request.sessionId, ''); assert.equal(request.messages.length, 1); + + catalog.default = 'default'; + await dom.run('AgentFlowAssistantModels.load()'); + store.delete('agentflow.orchestratorModel.codebuddy'); + choose(cli, 'codebuddy'); + assert.equal((await send()).model, 'cb-model', 'switching to CodeBuddy picks the first offered model'); + open(model); + assert.deepEqual([...model.list.querySelectorAll('[role="option"]')].map(option => option.dataset.id), ['cb-model']); console.log('orchestrator multi-CLI models ok'); })().catch(error => { console.error(error); process.exit(1); }); diff --git a/web/platform.css b/web/platform.css index 89de6401..7f191418 100644 --- a/web/platform.css +++ b/web/platform.css @@ -2472,6 +2472,15 @@ html[data-theme="dark"] .credential-logo.monochrome img{filter:invert(1)} .credential-catalog-row:hover{border-color:var(--cr-line)} .credential-catalog-row .credential-input input{height:32px;font-size:12.5px} .credential-catalog-row .credential-input input.credential-mono{font-size:12px} +.credential-catalog-effort{grid-column:2/4;min-width:0;padding:0 2px 5px;color:var(--cr-muted);font-size:11.5px} +.credential-catalog-effort summary{cursor:pointer;line-height:24px} +.credential-catalog-effort summary span{margin-left:8px;color:var(--cr-tertiary);overflow-wrap:anywhere} +.credential-catalog-effort-mode{display:flex;align-items:center;gap:10px;margin:4px 0 8px} +.credential-catalog-effort-mode select{padding:4px 8px;border:1px solid var(--cr-line);border-radius:6px;background:var(--cr-surface);color:var(--cr-text);font:inherit} +.credential-catalog-effort-levels{display:flex;flex-wrap:wrap;gap:8px 14px;padding-bottom:4px} +.credential-catalog-effort-levels label{display:flex;align-items:center;gap:5px;cursor:pointer;color:var(--cr-text-2)} +.credential-catalog-effort-levels input{margin:0;accent-color:var(--cr-accent)} +.credential-catalog-effort-levels small{flex-basis:100%;color:var(--cr-tertiary);font-size:11px;line-height:1.5} .credential-catalog-grip{display:grid;place-items:center;width:24px;height:30px;margin:0;padding:0;border:0;border-radius:6px;background:transparent;color:var(--cr-tertiary);cursor:grab;touch-action:none} .credential-catalog-grip:hover{background:var(--cr-hover);color:var(--cr-text)} .credential-catalog-grip:focus-visible{outline:2px solid color-mix(in srgb,var(--cr-accent) 65%,transparent);outline-offset:1px;color:var(--cr-text)} diff --git a/webtest/tests/assets-agents.test.mjs b/webtest/tests/assets-agents.test.mjs index d767b208..07042836 100644 --- a/webtest/tests/assets-agents.test.mjs +++ b/webtest/tests/assets-agents.test.mjs @@ -129,3 +129,49 @@ test('role model picker reloads the editor catalog and preserves a CLI change wh assert.deepEqual(menu().map(row => row.dataset.id), ['', 'private/second', '__custom__']); assert.equal(app.document.querySelector('#agentModelPicker [data-model-label]').textContent, 'Renamed model'); }); + +test('role effort picker follows each model and resets an unsupported level on model changes', async (t) => { + const app = await loadApp(); + t.after(app.close); + const xhigh = { value:'xhigh', label:'超高 xhigh' }; + app.router.get('/api/executor-models', { + codebuddy:{models:[],efforts:[]}, claude:{models:[],efforts:[]}, + codex:{customModel:true,models:[{id:'reasoner',name:'Reasoner',efforts:[xhigh]}, {id:'plain',name:'Plain',efforts:[]}],efforts:[{value:'low',label:'低 low'},xhigh]}, + }); + app.evaluate(` + state.data.executors=[{id:'codebuddy',name:'CodeBuddy'}, {id:'codex',name:'Codex'}]; + openAgentWizard({id:'test-role',name:'Test',prompt:'Test',executorId:'codex',model:'reasoner',effort:'xhigh'}); + `); + await settle(); + const choices = () => [...app.document.querySelectorAll('#agentEffortPicker [role=option]')].map(row=>row.dataset.id); + assert.deepEqual(choices(), ['', 'xhigh']); + assert.equal(app.document.querySelector('#agentEffortPicker [name="effort"]').value, 'xhigh'); + app.document.querySelector('#agentModelPicker [data-id="plain"]').click(); + assert.deepEqual(choices(), ['']); + assert.equal(app.document.querySelector('#agentEffortPicker [name="effort"]').value, ''); +}); + +test('a personal catalog does not rewrite an existing role effort during a name edit', async (t) => { + const app = await loadApp(); + t.after(app.close); + let loaded; + app.router.get('/api/executor-models', () => new Promise(resolve => { loaded = resolve; })); + app.router.put('/api/agents', call => ({ status: 200, body: call.body })); + app.evaluate(` + state.data.executors=[{id:'codebuddy',name:'CodeBuddy'}, {id:'codex',name:'Codex'}]; + openAgentWizard({id:'agent-1',name:'Developer',prompt:'Review',executorId:'codex',model:'reasoner',effort:'xhigh'}); + `); + await settle(); + loaded({codex:{customModel:true,models:[{id:'reasoner',name:'Reasoner',efforts:[]}],efforts:[]}}); + await settle(); + const picker = app.document.querySelector('#agentEffortPicker'); + assert.deepEqual([...picker.querySelectorAll('[role=option]')].map(item=>item.dataset.id), ['']); + assert.equal(picker.querySelector('[data-model-label]').textContent, 'xhigh'); + await app.evaluate(` + $('#agentForm').elements.name.value='Renamed'; + persistAgent($('#agentForm')); + `); + const saved = app.router.calls.find(call=>call.path==='/api/agents'&&call.method==='PUT').body; + assert.equal(saved.name, 'Renamed'); + assert.equal(saved.effort, 'xhigh'); +}); diff --git a/webtest/tests/model-catalog-efforts.test.mjs b/webtest/tests/model-catalog-efforts.test.mjs new file mode 100644 index 00000000..c176b77f --- /dev/null +++ b/webtest/tests/model-catalog-efforts.test.mjs @@ -0,0 +1,120 @@ +import test from 'node:test'; +import assert from 'node:assert/strict'; +import { loadApp, settle } from '../lib/load-app.mjs'; + +const levels = ['low', 'medium', 'high', 'xhigh', 'max', 'ultra']; +const catalogs = { + codebuddy: { models: [], efforts: levels.map(value => ({ value, label: value })) }, + codex: { models: [], efforts: levels.map(value => ({ value, label: value })) }, + claude: { models: [], efforts: levels.map(value => ({ value, label: value })) }, + cursor: { models: [], efforts: levels.map(value => ({ value, label: value })) }, +}; + +async function editor(t, models, engine='codex', catalogReply=catalogs) { + const app = await loadApp(); + t.after(app.close); + app.router.get('/api/executor-models?builtin=1', catalogReply); + const credential = engine==='cursor' + ? {agentCliCredentials:{cursor:{apiKeyConfigured:true,models,modelsCustomized:true}}} + : {codexApiKeyConfigured:true,codexModels:models}; + app.evaluate(` + state.currentUser={...state.data.users[0],id:1,isGuest:false,...${JSON.stringify(credential)}}; + state.data.users[0]=state.currentUser; + `); + app.router.put('/api/users/1/credentials', ({body}) => ({ id:1, codexApiKeyConfigured:true, ...body })); + await app.evaluate('renderPersonalSettings()'); + await settle(); + const kind=engine==='cursor'?'agentCli.cursor':'codexApiKey'; + app.evaluate(`toggleCredentialRow(${JSON.stringify(kind)},true)`); + const row = app.document.querySelector(`[data-credential-row="${kind}"]`); + return { app, row, catalog: row.querySelector('[data-credential-catalog]') }; +} + +function input(app, element, value) { + if(element.type === 'checkbox') element.checked = value; + else element.value = value; + element.dispatchEvent(new app.window.Event('input', { bubbles:true })); +} + +test('model effort edits save independently of ID and name, and automatic differs from default-only', async t => { + const {app,row,catalog} = await editor(t, [{id:'gateway',reasoningEfforts:['high','low']}, {id:'legacy'}]); + const save = () => row.querySelector('[data-credential-act="save"]'); + assert.equal(save().disabled, true, 'opening a stored catalog never reorders it into a dirty state'); + const model = catalog.querySelector('[data-catalog-row]'); + model.querySelector('details').open = true; + input(app, model.querySelector('[data-catalog-effort="high"]'), false); + assert.equal(save().disabled, false, 'capability-only edits enable save'); + assert.equal(model.querySelector('[data-catalog-effort-summary]').textContent, 'low'); + save().click(); + await settle(); + const saved = app.router.calls.find(call => call.method === 'PUT').body; + assert.deepEqual(saved.codexModels, [{id:'gateway',reasoningEfforts:['low']}, {id:'legacy'}]); + + app.evaluate(`toggleCredentialRow('codexApiKey',true)`); + const fresh = app.document.querySelector('[data-credential-row="codexApiKey"] [data-catalog-row]'); + input(app, fresh.querySelector('[data-catalog-effort="low"]'), false); + assert.deepEqual(JSON.parse(app.evaluate(`JSON.stringify(readCredentialCatalog(document.querySelector('[data-credential-row="codexApiKey"] [data-credential-catalog]')))`))[0].reasoningEfforts, []); + input(app, fresh.querySelector('[data-catalog-effort-mode]'), 'auto'); + const automatic = JSON.parse(app.evaluate(`JSON.stringify(readCredentialCatalog(document.querySelector('[data-credential-row="codexApiKey"] [data-credential-catalog]')))`)); + assert.equal(Object.hasOwn(automatic[0], 'reasoningEfforts'), false); + assert.equal(fresh.querySelector('[data-catalog-effort-levels]').hidden, true); +}); + +test('re-probing changes automatic capabilities and keeps explicit overrides', async t => { + const {app,row,catalog} = await editor(t, [ + {id:'automatic',name:'My name',discoveredReasoningEfforts:['low']}, + {id:'custom',discoveredReasoningEfforts:['low'],reasoningEfforts:['low']}, + {id:'default-only',reasoningEfforts:[]}, + ]); + app.router.post('/api/users/1/cli-models/probe', { + source:'remote', models:[ + {id:'automatic',name:'Remote name',discoveredReasoningEfforts:['high','xhigh']}, + {id:'custom',discoveredReasoningEfforts:['high']}, + {id:'default-only',discoveredReasoningEfforts:['high']}, + ], + }); + catalog.querySelector('[data-catalog-act="probe"]').click(); + await settle(); + catalog.querySelector('[data-catalog-act="probe-apply"]').click(); + const modes = [...catalog.querySelectorAll('[data-catalog-effort-mode]')]; + assert.deepEqual(modes.map(mode=>mode.value), ['auto','custom','custom']); + assert.equal(catalog.querySelector('[data-catalog-effort-summary]').textContent, '自动 · high / xhigh'); + const save = row.querySelector('[data-credential-act="save"]'); + assert.equal(save.disabled, false); + save.click(); + await settle(); + assert.deepEqual(app.router.calls.find(call => call.method === 'PUT').body.codexModels, [ + {id:'automatic',name:'My name',discoveredReasoningEfforts:['high','xhigh']}, + {id:'custom',discoveredReasoningEfforts:['high'],reasoningEfforts:['low']}, + {id:'default-only',discoveredReasoningEfforts:['high'],reasoningEfforts:[]}, + ]); +}); + +test('metadata-only refresh is dirty and leaves Cursor model-ID mappings intact', async t => { + const {app,catalog} = await editor(t, [{id:'family',efforts:{high:'family-high',low:'family-low'},defaultEffort:'low'}], 'cursor'); + const row = catalog.querySelector('[data-catalog-row]'); + input(app, row.querySelector('[data-catalog-effort-mode]'), 'custom'); + input(app, row.querySelector('[data-catalog-effort="high"]'), true); + const payload = JSON.parse(app.evaluate(`JSON.stringify(credentialCatalogPayload(document.querySelector('[data-credential-row="agentCli.cursor"] [data-credential-catalog]')))`)); + assert.deepEqual(payload[0].efforts, {high:'family-high',low:'family-low'}); + assert.equal(payload[0].defaultEffort, 'low'); + assert.deepEqual(payload[0].reasoningEfforts, ['high']); +}); + +test('late CLI capabilities fill in missing choices without replacing unsaved effort edits', async t => { + let finish; + const pending=new Promise(resolve=>{finish=resolve}); + const {app,catalog}=await editor(t,[{id:'slow-model',reasoningEfforts:['high']}],'codex',()=>pending); + const high=catalog.querySelector('[data-catalog-effort="high"]'); + assert.ok(high); + assert.equal(catalog.querySelector('[data-catalog-effort="xhigh"]'),null); + input(app,high,false); + high.focus(); + finish(catalogs); + await settle(); + assert.ok(catalog.querySelector('[data-catalog-effort="xhigh"]')); + assert.equal(high.checked,false); + assert.equal(app.document.activeElement,high); + const payload=JSON.parse(app.evaluate(`JSON.stringify(credentialCatalogPayload(document.querySelector('[data-credential-row="codexApiKey"] [data-credential-catalog]')))`)); + assert.deepEqual(payload[0].reasoningEfforts,[]); +}); diff --git a/webtest/tests/sync-schedule.test.mjs b/webtest/tests/sync-schedule.test.mjs index 9b398502..cc40ab05 100644 --- a/webtest/tests/sync-schedule.test.mjs +++ b/webtest/tests/sync-schedule.test.mjs @@ -541,7 +541,7 @@ test('AI schedule model picker includes account login CLI models and keeps a sav const app = await loadApp(); t.after(app.close); app.router.get('/api/assistant/models', call => ({ - default: 'glm-test', + default: 'default', models: [ { id: 'glm-test', name: 'GLM Test', family: 'glm' }, { id: 'claude-test', name: 'Claude Test', family: 'claude' },