chore(llm): drop unused imports, add XIAOMI_MIMO cross-turn cache tests

This commit is contained in:
matevip 2026-05-21 17:24:53 +08:00
parent 7861f603eb
commit 9c5ad29d42
7 changed files with 380 additions and 27 deletions

View File

@ -1,9 +1,6 @@
package vip.mate.llm.chatmodel;
import lombok.extern.slf4j.Slf4j;
import java.util.List;
import java.util.Map;
import java.util.concurrent.ConcurrentHashMap;
/**
@ -34,7 +31,6 @@ import java.util.concurrent.ConcurrentHashMap;
*
* @author MateClaw Team
*/
@Slf4j
public final class ReasoningContentCache {
private static final long DEFAULT_MAX_AGE_MS = 24 * 60 * 60 * 1000L; // 24 hours

View File

@ -82,11 +82,13 @@ class PatchReasoningContentTest {
@BeforeEach
void clearRelay() {
AssistantThinkingRelay.clearAll();
ReasoningContentCache.clear();
}
@AfterEach
void clearRelayAfter() {
AssistantThinkingRelay.clearAll();
ReasoningContentCache.clear();
}
// ---------- No-relay, no-thinking-mode path ----------
@ -427,4 +429,73 @@ class PatchReasoningContentTest {
assertEquals(" ", out.messages().get(3).reasoningContent(),
"DEEPSEEK plain in-turn assistant gets ' ' as before");
}
// ---------- XIAOMI_MIMO policy + cross-turn cache replay ----------
@Test
@DisplayName("XIAOMI_MIMO cross-turn tool_call: cache hit replays real reasoning_content")
void xiaomiMimoCrossTurn_replaysCachedReasoning() {
// Prior turn produced a tool_call with real thinking; NodeStreamingChatHelper
// stored it in the cache keyed by tool_call_id. On the next turn, the same
// assistant message is replayed as history with reasoning_content=null
// resolveCrossTurnReasoning must fetch the cached value before falling
// back to the policy's empty " ".
ReasoningContentCache.store(List.of("call_1"), "real-prior-thinking");
// Empty relay: no in-turn thinking (current turn hasn't produced one yet).
String token = AssistantThinkingRelay.stash(List.of(""), null);
ChatCompletionRequest req = request(List.of(
user("q1"),
assistantToolCall("a1", null), // i=1, cross-turn (1 <= 2), tool_call id="call_1"
user("q2") // i=2, lastUserIdx
), token);
ChatCompletionRequest out = OpenAiRequestRewriter.patchReasoningContent(req, provider("xiaomi-mimo"));
assertEquals("real-prior-thinking", out.messages().get(1).reasoningContent(),
"XIAOMI_MIMO cross-turn tool_call must replay cached reasoning_content over the ' ' fallback");
}
@Test
@DisplayName("XIAOMI_MIMO cross-turn tool_call: cache miss falls back to ' '")
void xiaomiMimoCrossTurnCacheMiss_fallsBackToSpace() {
// No cache entry for call_1 the multi-turn path must still validate by
// injecting the policy's emptyFallback so MiMo doesn't 400.
String token = AssistantThinkingRelay.stash(List.of(""), null);
ChatCompletionRequest req = request(List.of(
user("q1"),
assistantToolCall("a1", null), // i=1, cross-turn, no cache entry
user("q2")
), token);
ChatCompletionRequest out = OpenAiRequestRewriter.patchReasoningContent(req, provider("xiaomi-mimo"));
assertEquals(" ", out.messages().get(1).reasoningContent(),
"XIAOMI_MIMO cross-turn cache miss falls back to ' ' so the request still validates");
}
@Test
@DisplayName("XIAOMI_MIMO plain cross-turn assistant (no tool_calls) also patched via patchNonToolCall=true")
void xiaomiMimoCrossTurnPlainAssistant_patchedWithSpace() {
// XIAOMI_MIMO mirrors DEEPSEEK: patchNonToolCall=true means even plain
// text assistants in prior turns must carry reasoning_content. Cache
// can't help here (no tool_call_ids to key on) fallback is " ".
String token = AssistantThinkingRelay.stash(List.of("", ""), null);
ChatCompletionRequest req = request(List.of(
user("q1"),
new ChatCompletionMessage("plain a1", Role.ASSISTANT), // i=1, cross-turn, no tool_calls
user("q2"),
new ChatCompletionMessage("plain a2", Role.ASSISTANT) // i=3, in-turn, no tool_calls
), token);
ChatCompletionRequest out = OpenAiRequestRewriter.patchReasoningContent(req, provider("xiaomi-mimo"));
assertEquals(" ", out.messages().get(1).reasoningContent(),
"XIAOMI_MIMO plain prior-turn assistant gets ' ' (patchNonToolCall=true + patchCrossTurn=true)");
assertEquals(" ", out.messages().get(3).reasoningContent(),
"XIAOMI_MIMO plain in-turn assistant gets ' ' (patchNonToolCall=true)");
}
}

View File

@ -106,6 +106,11 @@
/>
<div v-else class="conv-title" @dblclick.stop="startRename(conv)">
<span>{{ conv.title }}</span>
<span
v-if="convGoalStatus(conv.conversationId) === 'active'"
class="conv-goal-dot"
:title="t('goal.sidebarActive', '此对话有正在进行的目标')"
></span>
<span
v-if="hasUnread(conv)"
class="conv-unread-dot"
@ -193,6 +198,16 @@ import { channelIconUrl } from '@/utils/channelSource'
import { mcToast } from '@/composables/useMcToast'
import { mcConfirm } from '@/components/common/useConfirm'
import type { Conversation, Agent } from '@/types'
import { useGoalStore } from '@/stores/useGoalStore'
const goalStore = useGoalStore()
/** Sidebar marker: a 6 px dot next to the conv title when that conv
* has an active goal. Color tracks status (orange = in progress).
* Falls back silently when the store hasn't loaded this conv yet. */
function convGoalStatus(cid: string): string | null {
const g = goalStore.activeGoalByConv?.[cid]
return g?.status ?? null
}
const props = defineProps<{
conversations: Conversation[]
@ -684,6 +699,22 @@ function onMenuSelect(item: DropdownMenuItem) {
vertical-align: middle;
}
/* Sidebar marker for "this conversation has an active goal". Same hue
* family as the avatar ring; sits at 60% opacity so it whispers rather
* than competes with unread / running indicators on the same row. */
.conv-goal-dot {
display: inline-block;
width: 7px;
height: 7px;
border-radius: 50%;
background: var(--mc-primary, #d97757);
opacity: 0.7;
margin-left: 6px;
flex-shrink: 0;
vertical-align: middle;
box-shadow: 0 0 0 2px color-mix(in srgb, var(--mc-primary, #d97757) 18%, transparent);
}
.conv-item.is-running {
background: color-mix(in srgb, #fbbf24 8%, transparent);
}

View File

@ -10,14 +10,16 @@
<!-- 头像 -->
<div class="msg-avatar" :class="`${role}-avatar`">
<slot name="avatar">
<!-- RFC 48 Jobs-cut: when the assistant has an active goal, wrap
the logo in GoalAvatarRing so the progress ring + breathing
halo + hover tooltip all sit naturally around the avatar.
The component renders only the slot content when no goal
exists, so non-goal turns look identical to before. -->
<!-- When the assistant has an active goal, wrap the logo in
GoalAvatarRing so the progress ring + breathing halo + hover
tooltip all sit naturally around the avatar. The component
renders only the slot content when no goal exists, so non-
goal turns look identical to before. The followup glyph
appears on messages that came from an auto-followup turn. -->
<GoalAvatarRing
v-if="role === 'assistant'"
:conversation-id="message.conversationId"
:show-followup-mark="isFollowupTurn"
>
<img src="/logo/mateclaw_logo_s.png" alt="" class="avatar-logo" />
</GoalAvatarRing>
@ -444,6 +446,7 @@ import ToolCallSegment from './ToolCallSegment.vue'
import ThinkingSegment from './ThinkingSegment.vue'
import ContentSegment from './ContentSegment.vue'
import GoalAvatarRing from '@/components/goal/GoalAvatarRing.vue'
import { useGoalStore } from '@/stores/useGoalStore'
import PlanStepsPanel from './PlanStepsPanel.vue'
import UserMessageContent from './UserMessageContent.vue'
import type { BrowserAction } from './BrowserTimeline.vue'
@ -487,6 +490,19 @@ const avatarIcon = computed(() => {
return role.value === 'user' ? props.userIcon : props.assistantIcon
})
// Followup attribution: an assistant message that opened right after a
// `goal_followup` SSE event belongs to an auto-followup turn. The chat
// composable stamps the message via goalStore on `message_start`; this
// computed reads it back so the glyph renders on exactly those turns.
const goalStore = useGoalStore()
const isFollowupTurn = computed(() => {
if (role.value !== 'assistant') return false
const cid = props.message.conversationId
const mid = props.message.id
if (!cid || mid == null) return false
return goalStore.isFollowupMessage(String(cid), String(mid))
})
// --- ---
const errorInfo = computed<ChatErrorInfo | undefined>(() => props.message.errorInfo)

View File

@ -345,6 +345,12 @@ export function useChat(options: UseChatOptions): UseChatReturn {
headers: streamHeaders,
})
// Goal store is referenced from several stream handlers (message_start
// for followup attribution, message_complete for the evaluating halo,
// plus the dedicated goal_* events below). Resolve once up front so
// the handlers don't each pull their own copy.
const goalStore = useGoalStore()
// ===== Async-task lifecycle bridge =====
// Generative tools (music / video / image) return a taskId synchronously and
// finish asynchronously via `async_task_completed`. If the upstream provider
@ -456,6 +462,13 @@ export function useChat(options: UseChatOptions): UseChatReturn {
const assistantMessage = createAssistantMessage('', streamConversationId)
;(assistantMessage as any)._turnId = activeTurnId
currentAssistantId.value = assistantMessage.id as string
// Auto-followup attribution: if the goal evaluator just decided to
// inject a followup, the message that just opened belongs to that
// turn. Stamp it so MessageBubble can render the small ↻ glyph.
if (streamConversationId && goalStore.consumePendingFollowup(streamConversationId)) {
goalStore.markFollowupMessage(streamConversationId, String(assistantMessage.id))
}
})
stream.on('warning', (data) => {
@ -529,6 +542,19 @@ export function useChat(options: UseChatOptions): UseChatReturn {
triggerAutoTts(streamConversationId, msg.content)
}
}
// Goal-evaluator breathing halo: when an assistant message finishes
// and this conversation has an active goal, the backend's evaluation
// node runs next. Flip the per-conv flag so GoalAvatarRing paints the
// breathing halo until `goal_evaluated` resets it. Skip when no goal
// is active — the halo should be quiet for ordinary turns.
if (
data.status === 'completed'
&& streamConversationId
&& goalStore.activeGoal(streamConversationId)
) {
goalStore.markEvaluating(streamConversationId, true)
}
})
stream.on('done', (data) => {
@ -1590,11 +1616,10 @@ export function useChat(options: UseChatOptions): UseChatReturn {
}
})
// ===== Goal events (RFC 48) =====
// Forward GoalEvaluationNode emissions to the goal store. The store
// owns active-goal cache + the per-conv "evaluating" flag that drives
// ===== Goal events =====
// Forward goal evaluator emissions to the goal store. The store owns
// the active-goal cache + the per-conv "evaluating" flag that drives
// the avatar ring's breathing halo.
const goalStore = useGoalStore()
stream.on('goal_evaluated', (data) => {
if (isStaleEvent(data)) return

View File

@ -5,11 +5,11 @@ import { goalApi, type Goal, type GoalEvent } from '@/api/index'
/**
* Per-conversation active goal cache + SSE event handlers.
*
* RFC 48 v3 Jobs-cut UI: the only visible affordances are the ring on the
* assistant avatar (driven by `activeGoalByConv[cid]`) and the inline
* "set goal" prompt that appears after the first assistant reply. There
* is no banner, no modal dialog, no drawer in the chat view. A separate
* /goals admin page (PR4b, optional) shows the full timeline.
* The visible affordances are the ring on the assistant avatar
* (driven by `activeGoalByConv[cid]`) and the inline "set goal" prompt
* that appears after the first assistant reply. There is no banner,
* no modal dialog, no drawer in the chat view. A separate /goals admin
* page shows the full timeline.
*/
export const useGoalStore = defineStore('goal', () => {
// Map of conversationId -> active goal (or null).
@ -23,6 +23,37 @@ export const useGoalStore = defineStore('goal', () => {
// Cached event timelines, keyed by goalId.
const eventsByGoal = ref<Record<string, GoalEvent[]>>({})
// Per-conversation "the user said 'No thanks' to setting a goal" — keeps
// GoalSetInlinePrompt from re-appearing every turn. In-memory only
// (intentionally not persisted: a fresh session re-prompts so the user
// can change their mind on the next visit).
const dismissedPromptByConv = ref<Record<string, boolean>>({})
// Per-conversation snapshot of the most recent terminal event
// (completed / exhausted). ChatConsole reads this to render the
// GoalSystemLine in the message stream. Cleared when the user dismisses
// it or starts a new goal on the same conversation.
const recentTerminalByConv = ref<Record<string, {
status: 'completed' | 'exhausted'
title: string
score?: number | null
reason?: string
at: number
} | null>>({})
// Per-conversation flag: "the goal evaluator just chose to inject a
// followup prompt, and the next assistant message that opens belongs
// to that followup turn." Consumed (cleared) by the chat composable's
// `message_start` handler so the message gets stamped exactly once.
const pendingFollowupByConv = ref<Record<string, boolean>>({})
// Assistant message IDs that came from auto-followup turns, grouped by
// conversation. MessageBubble reads this to show the small ↻ glyph on
// the avatar — the only visible signal that a turn was auto-triggered.
// Kept in memory only; on refetch the metadata persists server-side via
// the message's `metadata.fromFollowup` flag (handled by ChatHistory).
const followupMessageIdsByConv = ref<Record<string, Set<string>>>({})
const loading = ref(false)
async function loadActiveForConversation(conversationId: string) {
@ -134,25 +165,41 @@ export const useGoalStore = defineStore('goal', () => {
break
}
case 'goal_followup': {
// The next assistant turn will land soon; nothing to do for the ring.
// The next assistant turn will land soon. Flag the conversation
// so the chat composable can stamp the upcoming message as a
// followup turn when its `message_start` arrives. The ring keeps
// its evaluating state until message_complete fires for that
// followup turn — so the user sees breathe → still → breathe.
pendingFollowupByConv.value[conversationId] = true
break
}
case 'goal_completed': {
evaluatingByConv.value[conversationId] = false
if (goal) {
goal.status = 'completed'
if (data?.score != null) goal.completionScore = Number(data.score)
// Capture the terminal snapshot BEFORE we null out the cache so
// the GoalSystemLine has a title + score to render.
recentTerminalByConv.value[conversationId] = {
status: 'completed',
title: goal?.title || '目标',
score: data?.score != null ? Number(data.score) : (goal?.completionScore ?? null),
at: Date.now(),
}
// Refresh active so the empty-state prompt comes back into view.
// Clear active so the inline "set goal" prompt can come back.
activeGoalByConv.value[conversationId] = null
// Re-enable the inline prompt on completion (the user just closed
// a goal; a follow-up task is plausible).
dismissedPromptByConv.value[conversationId] = false
break
}
case 'goal_exhausted': {
evaluatingByConv.value[conversationId] = false
if (goal) {
goal.status = 'exhausted'
recentTerminalByConv.value[conversationId] = {
status: 'exhausted',
title: goal?.title || '目标',
reason: data?.reason,
at: Date.now(),
}
activeGoalByConv.value[conversationId] = null
dismissedPromptByConv.value[conversationId] = false
break
}
case 'goal_created':
@ -195,10 +242,69 @@ export const useGoalStore = defineStore('goal', () => {
return Math.max(0, Math.min(1, g.completionScore))
}
// ==================== Inline prompt + system line helpers ====================
function isPromptDismissed(conversationId: string): boolean {
return Boolean(dismissedPromptByConv.value[conversationId])
}
function dismissPrompt(conversationId: string) {
dismissedPromptByConv.value[conversationId] = true
}
function clearDismissedPrompt(conversationId: string) {
dismissedPromptByConv.value[conversationId] = false
}
function recentTerminal(conversationId: string) {
return recentTerminalByConv.value[conversationId] ?? null
}
function clearRecentTerminal(conversationId: string) {
recentTerminalByConv.value[conversationId] = null
}
// ==================== Followup attribution helpers ====================
/**
* Consume the pending-followup flag for this conversation if it's
* set, returning true when the caller should stamp the just-opened
* assistant message as a followup turn. Idempotent calling twice
* returns false the second time.
*/
function consumePendingFollowup(conversationId: string): boolean {
if (!conversationId) return false
const pending = pendingFollowupByConv.value[conversationId]
if (pending) {
pendingFollowupByConv.value[conversationId] = false
return true
}
return false
}
function markFollowupMessage(conversationId: string, messageId: string) {
if (!conversationId || !messageId) return
let set = followupMessageIdsByConv.value[conversationId]
if (!set) {
set = new Set<string>()
followupMessageIdsByConv.value[conversationId] = set
}
set.add(messageId)
}
function isFollowupMessage(conversationId: string, messageId: string): boolean {
if (!conversationId || !messageId) return false
return followupMessageIdsByConv.value[conversationId]?.has(messageId) ?? false
}
return {
activeGoalByConv,
evaluatingByConv,
eventsByGoal,
dismissedPromptByConv,
recentTerminalByConv,
pendingFollowupByConv,
followupMessageIdsByConv,
loading,
loadActiveForConversation,
create,
@ -211,6 +317,14 @@ export const useGoalStore = defineStore('goal', () => {
isEvaluating,
activeGoal,
progressFraction,
isPromptDismissed,
dismissPrompt,
clearDismissedPrompt,
recentTerminal,
clearRecentTerminal,
consumePendingFollowup,
markFollowupMessage,
isFollowupMessage,
}
})

View File

@ -164,6 +164,31 @@
</div>
</div>
<!-- Terminal-state announcement after a goal completed or exhausted
in this conversation. Auto-dismisses when the user clicks × or
starts a new goal. -->
<GoalSystemLine
v-if="goalTerminalForCurrent && currentConversationId"
:variant="goalTerminalForCurrent.status"
:title="goalSystemLineTitle"
:detail="goalSystemLineDetail"
class="goal-system-line-slot"
@click.stop="onGoalSystemLineDismiss"
/>
<!-- Inline "set a goal?" invitation shown after the first assistant
reply when the conversation has no active goal and the user
hasn't dismissed it for this conv. -->
<GoalSetInlinePrompt
v-if="showGoalSetPrompt"
:conversation-id="currentConversationId"
:agent-id="String(selectedAgentId)"
:workspace-id="String(currentWorkspaceId || '1')"
:suggested-title="goalSuggestedTitle"
class="goal-set-prompt-slot"
@dismiss="onGoalPromptDismiss"
/>
<!-- 流式处理 Loading 消息和输入框之间 -->
<StreamLoadingBar
:is-loading="isGenerating && !blockingPrompt"
@ -259,6 +284,9 @@ import { useEChartsRenderer } from '@/composables/useEChartsRenderer'
import { useKatexRenderer } from '@/composables/useKatexRenderer'
import { useMermaidRenderer, handleMermaidDownload } from '@/composables/useMermaidRenderer'
import { useGoalStore } from '@/stores/useGoalStore'
import { useWorkspaceStore } from '@/stores/useWorkspaceStore'
import GoalSetInlinePrompt from '@/components/goal/GoalSetInlinePrompt.vue'
import GoalSystemLine from '@/components/goal/GoalSystemLine.vue'
// ============ Talk Mode ============
const showTalkMode = ref(false)
@ -1009,16 +1037,88 @@ watch([selectedAgentId, currentConversationId], () => {
syncRouteState()
})
// RFC 48 load the active goal whenever the user switches conversation.
// The avatar ring listens on goalStore.activeGoalByConv[cid]; without this
// Load the active goal whenever the user switches conversation. The
// avatar ring listens on goalStore.activeGoalByConv[cid]; without this
// fetch the ring would only appear after an SSE event mutated the store.
const goalStore = useGoalStore()
const workspaceStoreForGoal = useWorkspaceStore()
const currentWorkspaceId = computed(() => workspaceStoreForGoal.currentWorkspaceId ?? '1')
watch(currentConversationId, async (cid) => {
if (cid) {
await goalStore.loadActiveForConversation(cid)
}
}, { immediate: true })
// Derive props for the inline prompt + system-line slots that sit
// between MessageList and ChatInput. The prompt shows only when:
// 1) there's a current conversation, agent, and at least one assistant
// reply (otherwise the prompt is premature);
// 2) there's no active goal (the ring already covers active state);
// 3) the user hasn't dismissed the prompt on this conv;
// 4) we're not mid-stream (don't pop suggestions while the agent
// is still typing).
const goalTerminalForCurrent = computed(() =>
currentConversationId.value
? goalStore.recentTerminal(currentConversationId.value)
: null,
)
const goalSystemLineTitle = computed(() => {
const t = goalTerminalForCurrent.value
if (!t) return ''
// The leading icon is owned by GoalSystemLine ( / ) so we don't
// prepend one here doing so produced " 🎉 " double-glyph titles.
return t.status === 'completed' ? `目标达成 · ${t.title}` : `这次的预算用完了 · ${t.title}`
})
const goalSystemLineDetail = computed(() => {
const t = goalTerminalForCurrent.value
if (!t) return ''
if (t.status === 'completed') {
return t.score != null
? `已完成 · final score ${t.score.toFixed(2)}`
: '已完成 · 总结已存入长期记忆'
}
// exhausted
if (t.reason === 'turn_budget') return '预算轮数用完。'
if (t.reason === 'llm_call_budget') return 'LLM 调用预算用完。'
return '预算耗尽。'
})
const showGoalSetPrompt = computed(() => {
if (!currentConversationId.value || !selectedAgentId.value) return false
if (isGenerating.value) return false
// Active goal? The ring covers that no need for a prompt.
if (goalStore.activeGoal(currentConversationId.value)) return false
// Recent terminal still showing? Let the user dismiss that first.
if (goalTerminalForCurrent.value) return false
if (goalStore.isPromptDismissed(currentConversationId.value)) return false
// Need at least one user assistant exchange so the prompt has
// context to derive a suggested title from.
const hasAssistantReply = messages.value.some(m => m.role === 'assistant')
return hasAssistantReply
})
// Build a sensible default title from the conversation's first user
// message. The user can always edit later via the goal page.
const goalSuggestedTitle = computed(() => {
const firstUser = messages.value.find(m => m.role === 'user')
const raw = (firstUser?.content || '').trim()
if (!raw) return '新目标'
// 80 char clip mirrors GoalController.create validation.
return raw.length > 80 ? raw.slice(0, 77) + '...' : raw
})
function onGoalPromptDismiss() {
if (currentConversationId.value) {
goalStore.dismissPrompt(currentConversationId.value)
}
}
function onGoalSystemLineDismiss() {
if (currentConversationId.value) {
goalStore.clearRecentTerminal(currentConversationId.value)
}
}
// Refetch agent capabilities (modalities + sidecar config) on agent change so
// the multimodal routing hint above the input box can react synchronously when
// the user attaches an image / video.