Skip to content

Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.

fix(chat): preserve steering continuity and composer focus

wizardchen5cf093706ebe
diff --git a/docs/chat-steering.md b/docs/chat-steering.md
new file mode 100644
index 000000000..f1307a783
--- /dev/null
+++ b/docs/chat-steering.md
@@ -0,0 +1,44 @@
+# Steering an active conversation
+
+While an agent is running, you can queue a follow-up or guide the current task.
+
+| Action | Behavior |
+| --- | --- |
+| Enter or the send button | Queue the draft to send after the current answer finishes. |
+| Command+Enter or Alt/Option+Enter | Send the draft as guidance for the active task. With an empty draft, promote the first eligible queued message. |
+| Up arrow on a queued message | Send that message as guidance immediately. |
+| Remove on a queued message | Remove the queued message. |
+| Stop button | Stop generation. This replaces the send button while the agent is running and the draft is empty. |
+| Shift+Enter or Ctrl+Enter | Insert a newline. |
+
+Outside an active agent run, Enter sends a normal message. Input methods can
+confirm composed text without accidentally sending it. Modes without steering
+support keep the stop action during generation.
+
+Queued messages appear in a compact strip immediately above the input's focus
+border. Guidance moves into the conversation immediately; delivery failures can
+be retried or removed. Earlier work remains visible, and guidance does not mark
+the ongoing task as complete. Sending and entering a newly created session keep
+the composer ready for further input.
+
+Guidance applies to the task in progress. The agent retains unfinished goals and
+constraints unless the user explicitly changes or cancels them. A short tagged
+reminder accompanies each guidance message; existing context and tool results
+are reused without an additional model call to classify the message. Message IDs
+and execution boundaries preserve ordering across live updates and history reloads.
+
+The conversation outline includes both user and assistant messages and highlights
+all messages intersecting the visible conversation area.
+
+## Component previews
+
+These previews use sample messages and isolated production components; they do
+not contain real conversation data or invoke a model.
+
+Running with an empty draft:
+
+![Queued guidance, visible outline entries, and Stop](images/chat-steer-preview.png)
+
+Running with a draft:
+
+![The same primary control becomes Send](images/chat-steer-send-preview.png)
diff --git a/docs/images/chat-steer-preview.png b/docs/images/chat-steer-preview.png
new file mode 100644
index 000000000..b0d08258b
Binary files /dev/null and b/docs/images/chat-steer-preview.png differ
diff --git a/docs/images/chat-steer-send-preview.png b/docs/images/chat-steer-send-preview.png
new file mode 100644
index 000000000..3fc3b6252
Binary files /dev/null and b/docs/images/chat-steer-send-preview.png differ
diff --git a/frontend/src/api/chat/steer.ts b/frontend/src/api/chat/steer.ts
index 443bc9cf9..e2292e711 100644
--- a/frontend/src/api/chat/steer.ts
+++ b/frontend/src/api/chat/steer.ts
@@ -9,16 +9,17 @@ export type SteerQueueItem = {
mentioned_items?: unknown[]
promoting?: boolean
awaitingIdleSend?: boolean
- // True while POST /steer is in flight. The overlay uses a client-side
- // id until the server replies with the durable one; promote/remove are
- // disabled in that window so they cannot hit /steer/{client_id}.
+ // True while POST /steer is in flight. The stable client UUID is also the
+ // durable server ID; disable actions until its queue entry exists.
pending?: boolean
+ failed?: boolean
+ expected_assistant_message_id?: string
client_id?: string
}
/**
* Append a message to a running agent turn.
- * Returns { success, status: 'queued' | 'new_run', steer_id?, delivery?, assistant_message_id? }.
+ * Returns { success, status: 'queued' | 'already_injected' | 'new_run', steer_id?, delivery?, assistant_message_id? }.
* - delivery 'after' (default): waits until the current run exits, then starts a follow-up turn.
* - delivery 'inject': the running engine injects it at the next round boundary.
* - 'new_run': no run is live; the caller should fall back to a normal send.
@@ -30,9 +31,13 @@ export async function steerSession(
query: string,
mentionedItems: any[] = [],
delivery: SteerDelivery = 'after',
+ expectedAssistantMessageId?: string,
+ steerId?: string,
) {
return post(`/api/v1/sessions/${session_id}/steer`, {
query,
+ expected_assistant_message_id: expectedAssistantMessageId,
+ steer_id: steerId,
mentioned_items: mentionedItems,
channel: 'web',
delivery,
diff --git a/frontend/src/components/Input-field.agent-switch.test.mjs b/frontend/src/components/Input-field.agent-switch.test.mjs
index 67cc2f100..571f0bbdf 100644
--- a/frontend/src/components/Input-field.agent-switch.test.mjs
+++ b/frontend/src/components/Input-field.agent-switch.test.mjs
@@ -1,10 +1,41 @@
import assert from 'node:assert/strict'
import { readFileSync } from 'node:fs'
import test from 'node:test'
+import vm from 'node:vm'
+import ts from 'typescript'
+import { computed, reactive, ref } from 'vue'
+import { chatSubmitShortcut } from '../utils/chatSubmitShortcut.ts'
const inputField = readFileSync(new URL('./Input-field.vue', import.meta.url), 'utf8')
const settingsStore = readFileSync(new URL('../stores/settings.ts', import.meta.url), 'utf8')
+test('composer shortcuts queue Enter, inject drafts, and promote the first queued message when empty', () => {
+ const code = inputField.slice(inputField.indexOf('const steerShortcutLabel ='), inputField.indexOf('const onPaste ='))
+ const sends = [], promotes = []
+ const query = ref('draft')
+ const composing = ref(false)
+ const props = reactive({ isReplying: true, canSteer: true, queuedSteers: [
+ { steer_id: 'pending', delivery: 'after', pending: true },
+ { steer_id: 'first', delivery: 'after' }, { steer_id: 'second', delivery: 'after' },
+ ] })
+ const { onKeydown } = vm.runInNewContext(ts.transpile(`${code}\n({ onKeydown })`), {
+ navigator: { platform: 'MacIntel' }, computed, props, query, chatSubmitShortcut,
+ isComposing: composing, showMention: ref(false),
+ createSession: (...args) => sends.push(args), emit: (...args) => promotes.push(args),
+ })
+ const enter = { key: 'Enter', keyCode: 13, preventDefault() {}, shiftKey: false, ctrlKey: false, altKey: false }
+ onKeydown(query.value, { e: enter })
+ onKeydown(query.value, { e: { ...enter, altKey: true } })
+ onKeydown(query.value, { e: { ...enter, metaKey: true } })
+ assert.deepEqual(sends, [['draft', 'after'], ['draft', 'inject'], ['draft', 'inject']])
+ query.value = ''
+ onKeydown('', { e: { ...enter, metaKey: true } })
+ assert.deepEqual(promotes, [['promote-steer', 'first']])
+ composing.value = true
+ onKeydown('', { e: { ...enter, altKey: true } })
+ assert.equal(promotes.length, 1, 'IME confirmation must not promote a message')
+})
+
test('selecting an agent leaves web search off until the user enables it', () => {
const selectAgentStart = settingsStore.indexOf('selectAgent(agentId: string')
const getSelectedAgentStart = settingsStore.indexOf('getSelectedAgentId()', selectAgentStart)
@@ -36,17 +67,18 @@ test('shared-agent web search button waits for source readiness metadata', () =>
assert.match(showWebSearchButton, /selectedSharedAgent\.value\?\.web_search_ready/)
})
-test('replying keeps stop visible; typed text also shows send', () => {
+test('one composer action switches between stop and send', () => {
const controlsStart = inputField.indexOf('class="control-right"')
- const controls = inputField.slice(controlsStart, controlsStart + 1200)
+ const controls = inputField.slice(controlsStart, controlsStart + 2200)
assert.notEqual(controlsStart, -1)
assert.doesNotMatch(inputField, /steer-delivery-toggle/)
assert.doesNotMatch(inputField, /setSteerDelivery/)
- assert.match(inputField, /emit\('steer-msg', val\.trim\(\), steerMentions, 'after'\)/)
- assert.match(controls, /v-if="isReplying"/)
+ assert.match(inputField, /emit\('steer-msg', val\.trim\(\), steerMentions, delivery\)/)
+ assert.match(controls, /v-if="isReplying && \(!canSteer \|\| !query\.trim\(\)\)"/)
assert.match(controls, /handleStop/)
- assert.match(controls, /v-if="!isReplying \|\| query\.trim\(\)"[\s\S]*createSession\(query\)[\s\S]*send-btn/)
- assert.doesNotMatch(controls, /v-if="isReplying && \(!canSteer \|\| !query\.trim\(\)\)"/)
+ assert.match(controls, /v-else[\s\S]*createSession\(query\)[\s\S]*send-btn/)
+ assert.match(controls, /<t-icon name="arrow-up"/)
+ assert.doesNotMatch(controls, /steer-inject-btn|<t-icon[^>]*'time'/)
})
// Only agent turns have a round boundary to take a mid-run message at, and a
@@ -74,7 +106,7 @@ test('queued follow-ups sit in one block with send-now and remove', () => {
assert.match(inputField, /queuedSteers/)
assert.match(inputField, /emit\('promote-steer', item\.steer_id\)/)
assert.match(inputField, /emit\('remove-steer', item\.steer_id\)/)
- assert.match(inputField, /class="steer-queue-remove"/)
+ assert.match(inputField, /class="steer-queue-action steer-queue-remove"/)
assert.match(inputField, /\$t\('input\.steerQueueSendNow'\)/)
assert.match(inputField, /:disabled="item\.promoting \|\| item\.pending"/)
assert.match(inputField, /\.steer-queue-item \+ \.steer-queue-item/)
diff --git a/frontend/src/components/Input-field.focus.test.mjs b/frontend/src/components/Input-field.focus.test.mjs
new file mode 100644
index 000000000..560c2a3cc
--- /dev/null
+++ b/frontend/src/components/Input-field.focus.test.mjs
@@ -0,0 +1,81 @@
+import assert from 'node:assert/strict'
+import { readFileSync } from 'node:fs'
+import test from 'node:test'
+import vm from 'node:vm'
+import ts from 'typescript'
+
+const source = readFileSync(new URL('./Input-field.vue', import.meta.url), 'utf8')
+const focusCode = source.slice(source.indexOf('const focusInput ='), source.indexOf('const onInput ='))
+const sendCode = source.slice(source.indexOf('const createSession ='), source.indexOf('const updateAgentModeDropdownPosition ='))
+
+for (const mode of ['normal', 'embedded', 'after', 'inject']) {
+ test(`${mode} send clears the draft and restores focus after the DOM update`, async () => {
+ const effects = [], ticks = []
+ const textarea = {
+ isConnected: true,
+ focus: options => effects.push(['focus', options.preventScroll]),
+ blur: () => { throw new Error('sending must not blur the textarea') },
+ }
+ const context = {
+ props: { isReplying: ['after', 'inject'].includes(mode), canSteer: true, embeddedMode: mode === 'embedded' },
+ uploadedAttachments: { value: [] }, uploadedImages: { value: [] },
+ allSelectedItems: { value: [] }, selectedModelId: { value: 'model' },
+ selectedAgent: { value: { config: {} } },
+ settingsStore: {}, chatResources: { isFresh: () => true },
+ collectAgentNotReadyReasons: () => ({ keys: [], labels: [] }),
+ attachmentUploadRef: { value: null },
+ emit: name => effects.push([name]),
+ clearvalue: () => effects.push(['clear']),
+ getTextareaEl: () => textarea,
+ nextTick: () => new Promise(resolve => ticks.push(resolve)),
+ }
+ const send = vm.runInNewContext(ts.transpile(`${focusCode}\n${sendCode}\ncreateSession`), context)
+ await send('hello', mode === 'inject' ? 'inject' : 'after')
+ assert.deepEqual(effects, [[context.props.isReplying ? 'steer-msg' : 'send-msg'], ['clear']])
+ for (const resolve of ticks) resolve()
+ await new Promise(setImmediate)
+ assert.deepEqual(effects.at(-1), ['focus', true])
+ })
+}
+
+test('a queued focus does not focus a textarea removed during navigation', async () => {
+ let resolveTick, connected = true, focused = false
+ const focus = vm.runInNewContext(`${focusCode}\nfocusInput`, {
+ nextTick: () => new Promise(resolve => { resolveTick = resolve }),
+ getTextareaEl: () => ({ isConnected: connected, focus: () => { focused = true } }),
+ })
+ const pending = focus()
+ connected = false
+ resolveTick()
+ await pending
+ assert.equal(focused, false)
+})
+
+test('teardown blurs the active textarea before it is detached', () => {
+ const code = source.slice(source.indexOf('onBeforeUnmount(() =>'), source.indexOf('onUnmounted(() =>'))
+ let teardown, blurred = false
+ const textarea = { isConnected: true, blur: () => { blurred = true } }
+ vm.runInNewContext(code, {
+ onBeforeUnmount: fn => { teardown = fn }, getTextareaEl: () => textarea,
+ document: { activeElement: textarea },
+ })
+ teardown()
+ assert.equal(blurred, true)
+})
+
+test('new-session focus survives consumption of the first query and runs on child mount', () => {
+ const page = readFileSync(new URL('../views/chat/index.vue', import.meta.url), 'utf8')
+ const snapshot = page.match(/const focusComposerOnMount = [^;]+;/)[0]
+ const firstQuery = { value: 'initial question' }
+ const autoFocus = vm.runInNewContext(`${snapshot}\nfirstQuery.value = ''; focusComposerOnMount`, { firstQuery })
+ assert.equal(autoFocus, true)
+ assert.match(page, /:auto-focus="focusComposerOnMount"/)
+ const mountStart = source.slice(source.indexOf('onMounted(() => {'), source.indexOf('// Embed 渠道'))
+ let mounted, focused = false
+ vm.runInNewContext(`${mountStart}\n});`, {
+ props: { autoFocus }, onMounted: fn => { mounted = fn }, focusInput: () => { focused = true },
+ })
+ assert.equal(focused, false)
+ mounted()
+ assert.equal(focused, true)
+})
diff --git a/frontend/src/components/Input-field.vue b/frontend/src/components/Input-field.vue
index 2d476af84..ae9836783 100644
--- a/frontend/src/components/Input-field.vue
+++ b/frontend/src/components/Input-field.vue
@@ -1,5 +1,5 @@
<script setup lang="ts">
-import { ref, onMounted, onUnmounted, computed, watch, nextTick, h, type PropType } from "vue";
+import { ref, onMounted, onBeforeUnmount, onUnmounted, computed, watch, nextTick, h, type PropType } from "vue";
import { storeToRefs } from 'pinia';
import { useRoute, useRouter } from 'vue-router';
import { onBeforeRouteUpdate } from 'vue-router';
@@ -11,6 +11,7 @@ import { listKnowledgeBases, searchKnowledge, batchQueryKnowledge, listKnowledge
import { listMCPServices, type MCPService } from '@/api/mcp-service';
import { stopSession } from '@/api/chat';
import type { SteerQueueItem } from '@/api/chat/steer';
+import { chatSubmitShortcut } from '@/utils/chatSubmitShortcut';
import { useOrganizationStore } from '@/stores/organization';
import KnowledgeBaseSelector from './KnowledgeBaseSelector.vue';
import MentionSelector from './MentionSelector.vue';
@@ -502,6 +503,10 @@ const sharedAgentOrgName = computed(() => {
});
const props = defineProps({
+ autoFocus: {
+ type: Boolean,
+ default: false
+ },
isReplying: {
type: Boolean,
required: false
@@ -1507,6 +1512,12 @@ const getTextareaEl = () => {
return el.querySelector('textarea');
};
+const focusInput = async () => {
+ await nextTick();
+ const textarea = getTextareaEl();
+ if (textarea?.isConnected) textarea.focus({ preventScroll: true });
+};
+
const onInput = (val: string | InputEvent) => {
// 如果正在输入法组合中,不处理搜索逻辑,等待 compositionend
if (isComposing.value) return;
@@ -1794,6 +1805,7 @@ let resizeHandler: (() => void) | null = null;
let scrollHandler: (() => void) | null = null;
onMounted(() => {
+ if (props.autoFocus) void focusInput();
// Embed 渠道由宿主注入 agent/KB,勿拉取需 JWT 的平台资源
if (props.embeddedMode) return;
@@ -1861,6 +1873,12 @@ onMounted(() => {
window.addEventListener('scroll', scrollHandler, { passive: true, capture: true });
});
+onBeforeUnmount(() => {
+ // Let TDesign handle blur while its textarea is still attached to the DOM.
+ const textarea = getTextareaEl();
+ if (textarea?.isConnected && document.activeElement === textarea) textarea.blur();
+});
+
onUnmounted(() => {
window.removeEventListener(CHAT_FILE_DROP_EVENT, handleChatFileDrop as EventListener);
document.removeEventListener('click', closeAgentModeSelector);
@@ -1906,14 +1924,15 @@ const emit = defineEmits<{
(e: 'stop-generation'): void;
(e: 'stop-confirmed'): void;
(e: 'stop-failed'): void;
- // Mid-run send queues until the current run exits (delivery=after).
+ // Running input defaults to after; the explicit shortcut/action steers.
// Empty input while replying shows Stop; typed text also shows Send.
(e: 'steer-msg', query: string, mentionedItems: MentionRequestItem[], delivery: 'inject' | 'after'): void;
(e: 'promote-steer', steerId: string): void;
(e: 'remove-steer', steerId: string): void;
+ (e: 'retry-steer', steerId: string): void;
}>();
-const createSession = async (val: string) => {
+const createSession = async (val: string, delivery: 'inject' | 'after' = 'after') => {
if (!val.trim()) {
MessagePlugin.info(t('input.messages.enterContent'));
return;
@@ -1926,7 +1945,7 @@ const createSession = async (val: string) => {
MessagePlugin.info(t('input.messages.replying'));
return;
}
- // Mid-run steering: queue until the current run exits (delivery=after).
+ // Queue the selected delivery mode until its next safe boundary.
// Attachments are intentionally not allowed on the steer path — the
// running turn already resolved its own scope.
if (uploadedAttachments.value.some(item => item.status === 'uploading')) {
@@ -1947,8 +1966,9 @@ const createSession = async (val: string) => {
service_id: item.serviceId,
skill_name: item.skillName,
}));
- emit('steer-msg', val.trim(), steerMentions, 'after');
+ emit('steer-msg', val.trim(), steerMentions, delivery);
clearvalue();
+ void focusInput();
return;
}
// Only block while the file is still uploading (no document ID yet). Once
@@ -1969,10 +1989,9 @@ const createSession = async (val: string) => {
// Embed 渠道由后端绑定 agent/KB,勿走平台侧 agent 列表与就绪校验
if (props.embeddedMode) {
- const textarea = getTextareaEl();
- if (textarea) textarea.blur();
emit('send-msg', val, selectedModelId.value || '', [], [], []);
clearvalue();
+ void focusInput();
return;
}
@@ -2033,11 +2052,6 @@ const createSession = async (val: string) => {
const imageFiles = uploadedImages.value.map(img => img.file);
const attachmentFiles = uploadedAttachments.value;
- // Blur the textarea BEFORE emitting, so that when the parent navigates away
- // and Vue unmounts this component, TDesign's blur handler won't fire on a
- // detached DOM element (which causes getComputedStyle to throw).
- const textarea = getTextareaEl();
- if (textarea) textarea.blur();
emit('send-msg', val, selectedModelId.value, mentionedItems, imageFiles, attachmentFiles);
// Clean up image previews
@@ -2049,6 +2063,7 @@ const createSession = async (val: string) => {
uploadedAttachments.value = [];
clearvalue();
+ void focusInput();
}
const updateAgentModeDropdownPosition = () => {
@@ -2272,7 +2287,17 @@ const clearPendingUploads = () => {
uploadedAttachments.value = [];
}
-const onKeydown = (val: string, event: { e: { preventDefault(): unknown; keyCode: number; shiftKey: any; ctrlKey: any; }; }) => {
+const steerShortcutLabel = /Mac|iPhone|iPad/.test(navigator.platform) ? '⌘ Enter' : 'Alt+Enter';
+const firstQueuedSteer = computed(() => props.queuedSteers.find(item =>
+ item.delivery === 'after' && !item.pending && !item.promoting && !item.failed));
+const injectCurrentInput = () => {
+ if (!props.isReplying || !props.canSteer) return;
+ if (query.value.trim()) void createSession(query.value, 'inject');
+ else if (firstQueuedSteer.value) emit('promote-steer', firstQueuedSteer.value.steer_id);
+};
+
+const onKeydown = (val: string, event: { e: KeyboardEvent }) => {
+ if (isComposing.value || event.e.isComposing || event.e.keyCode === 229) return;
if (showMention.value) {
if (event.e.keyCode === 38) { // Up
event.e.preventDefault();
@@ -2312,12 +2337,11 @@ const onKeydown = (val: string, event: { e: { preventDefault(): unknown; keyCode
}
}
- if ((event.e.keyCode == 13 && event.e.shiftKey) || (event.e.keyCode == 13 && event.e.ctrlKey)) {
- return;
- }
- if (event.e.keyCode == 13) {
+ const delivery = chatSubmitShortcut(event.e, props.isReplying && props.canSteer);
+ if (delivery) {
event.e.preventDefault();
- createSession(val)
+ if (delivery === 'inject' && props.isReplying && props.canSteer) injectCurrentInput();
+ else void createSession(val, delivery);
}
}
@@ -2557,6 +2581,7 @@ onBeforeRouteUpdate((to, from, next) => {
})
defineExpose({
+ focusInput,
triggerSend(text: string) {
if (!text.trim()) return;
query.value = text;
@@ -2570,29 +2595,31 @@ defineExpose({
<!-- Hidden file input for image upload -->
<input ref="imageInputRef" type="file" accept="image/jpeg,image/png,image/gif,image/webp" multiple
style="display:none" @change="handleImageSelect" />
+ <!-- 队列紧贴输入框上方,不参与输入区的焦点高亮 -->
<div v-if="queuedSteers.length" class="steer-queue" role="list" :aria-label="$t('input.steerQueueWaiting')">
<div v-for="item in queuedSteers" :key="item.steer_id" class="steer-queue-item" role="listitem">
- <span class="steer-queue-text">{{ item.content }}</span>
+ <t-tooltip :content="$t('input.steerAfter')">
+ <t-icon name="time" class="steer-queue-icon" :aria-label="$t('input.steerAfter')" />
+ </t-tooltip>
+ <span class="steer-queue-text" :title="item.content">{{ item.content }}</span>
<div class="steer-queue-actions">
- <button
- v-if="item.delivery !== 'inject'"
- type="button"
- class="steer-queue-send-now"
- :disabled="item.promoting || item.pending"
- @click="emit('promote-steer', item.steer_id)"
- >{{ $t('input.steerQueueSendNow') }}</button>
- <span v-else class="steer-queue-hint">{{ $t('input.steerQueueInjecting') }}</span>
- <button
- type="button"
- class="steer-queue-remove"
- :aria-label="$t('common.remove')"
- :disabled="item.promoting || item.pending"
- @click="emit('remove-steer', item.steer_id)"
- >×</button>
+ <t-tooltip v-if="item.failed" :content="$t('input.steerRetry')">
+ <button type="button" class="steer-queue-action" :aria-label="$t('input.steerRetry')" @click="emit('retry-steer', item.steer_id)"><t-icon name="refresh" /></button>
+ </t-tooltip>
+ <t-icon v-else-if="item.pending" name="loading" class="steer-sending" :aria-label="$t('common.loading')" />
+ <t-tooltip v-else :content="`${$t('input.steerQueueSendNow')}${item.steer_id === firstQueuedSteer?.steer_id ? ` · ${steerShortcutLabel}` : ''}`">
+ <button type="button" class="steer-queue-action"
+ :aria-label="$t('input.steerQueueSendNow')" :disabled="item.promoting || item.pending"
+ @click="emit('promote-steer', item.steer_id)"><t-icon name="arrow-up" /></button>
+ </t-tooltip>
+ <t-tooltip :content="$t('common.remove')">
+ <button type="button" class="steer-queue-action steer-queue-remove"
+ :aria-label="$t('common.remove')" :disabled="item.promoting || item.pending"
+ @click="emit('remove-steer', item.steer_id)"><t-icon name="close" /></button>
+ </t-tooltip>
</div>
</div>
</div>
- <!-- 富文本输入框容器 -->
<div class="rich-input-container" data-guide="chat-input">
<!-- 图片预览区域 -->
<div v-if="uploadedImages.length > 0" class="image-preview-bar">
@@ -2635,7 +2662,7 @@ defineExpose({
@keydown="onKeydown" @input="onInput" @compositionstart="onCompositionStart" @compositionend="onCompositionEnd"
@paste="onPaste" />
- <!-- 控制栏(放在 rich-input-container 内,相对输入框边框定位) -->
+ <!-- 控制栏按文档流排列,换行时自动撑开容器 -->
<div class="control-bar" :class="{ 'is-embedded': embeddedMode }">
<!-- 左侧控制按钮 -->
<div class="control-left" v-if="!embeddedMode">
@@ -2815,19 +2842,18 @@ defineExpose({
<!-- 右侧控制:回复中且输入为空是停止,一旦输入新内容同一位置变成发送 -->
<div class="control-right">
- <t-tooltip v-if="isReplying" :content="$t('input.stopGeneration')"
- placement="top">
- <div @click="handleStop" class="control-btn stop-btn">
- <svg width="16" height="16" viewBox="0 0 16 16" fill="currentColor">
- <rect x="5" y="5" width="6" height="6" rx="1" />
- </svg>
- </div>
+ <t-tooltip v-if="isReplying && (!canSteer || !query.trim())" :content="$t('input.stopGeneration')" placement="top">
+ <button type="button" @click="handleStop" class="control-btn stop-btn" :aria-label="$t('input.stopGeneration')">
+ <t-icon name="stop" />
+ </button>
+ </t-tooltip>
+ <t-tooltip v-else :content="`${isReplying && canSteer ? $t('input.steerAfter') : $t('input.send')} · Enter`">
+ <button type="button" @click="createSession(query)" class="control-btn send-btn" data-guide="chat-send"
+ :disabled="!query.trim()" :class="{ 'disabled': !query.trim() }"
+ :aria-label="isReplying && canSteer ? $t('input.steerAfter') : $t('input.send')">
+ <t-icon name="arrow-up" />
+ </button>
</t-tooltip>
-
- <div v-if="!isReplying || query.trim()" @click="createSession(query)" class="control-btn send-btn" data-guide="chat-send"
- :class="{ 'disabled': !query.length }">
- <img src="../assets/img/sending-aircraft.svg" :alt="$t('input.send')" />
- </div>
</div>
</div>
</div>
@@ -2879,21 +2905,22 @@ const getImgSrc = (url: string) => {
}
.steer-queue {
- width: 100%;
- max-width: 960px;
- margin-bottom: 8px;
- background: var(--td-bg-color-container, #fff);
+ width: calc(100% - 24px);
+ max-width: 936px;
+ box-sizing: border-box;
+ max-height: 140px;
+ overflow-y: auto;
+ background: var(--td-bg-color-secondarycontainer, #f5f5f5);
border: 1px solid var(--td-component-stroke, #dcdcdc);
- border-radius: 12px;
- box-shadow: 0 2px 8px rgba(0, 0, 0, 0.04);
- overflow: hidden;
+ border-bottom: 0;
+ border-radius: 10px 10px 0 0;
}
.steer-queue-item {
display: flex;
- align-items: flex-start;
- gap: 10px;
- padding: 8px 12px;
+ align-items: center;
+ gap: 8px;
+ padding: 4px 12px;
}
.steer-queue-item + .steer-queue-item {
@@ -2906,8 +2933,9 @@ const getImgSrc = (url: string) => {
font-size: 13px;
line-height: 1.5;
color: var(--td-text-color-primary);
- white-space: pre-wrap;
- word-break: break-word;
+ white-space: nowrap;
+ overflow: hidden;
+ text-overflow: ellipsis;
}
.steer-queue-actions {
@@ -2915,60 +2943,35 @@ const getImgSrc = (url: string) => {
align-items: center;
flex-shrink: 0;
gap: 4px;
- padding-top: 1px;
-}
-
-.steer-queue-send-now {
- flex-shrink: 0;
- margin: 0;
- padding: 0 4px;
- border: 0;
- background: transparent;
- color: var(--td-brand-color, #07C05F);
- font-size: 12px;
- line-height: 22px;
- cursor: pointer;
-
- &:disabled {
- opacity: 0.5;
- cursor: default;
- }
}
-.steer-queue-hint {
+.steer-queue-icon {
flex-shrink: 0;
- font-size: 12px;
- line-height: 22px;
+ font-size: 14px;
color: var(--td-text-color-secondary);
}
-.steer-queue-remove {
+.steer-queue-action {
display: inline-flex;
align-items: center;
justify-content: center;
- width: 22px;
- height: 22px;
- margin: 0;
+ width: 26px;
+ height: 26px;
padding: 0;
border: 0;
- border-radius: 50%;
+ border-radius: 6px;
background: transparent;
- color: var(--td-text-color-placeholder, #999);
+ color: var(--td-text-color-secondary);
font-size: 16px;
- line-height: 1;
cursor: pointer;
-
- &:hover:not(:disabled) {
- background: var(--td-bg-color-secondarycontainer, #f3f3f3);
- color: var(--td-text-color-primary);
- }
-
- &:disabled {
- opacity: 0.4;
- cursor: default;
- }
+ &:hover:not(:disabled) { background: var(--td-bg-color-secondarycontainer); color: var(--td-text-color-primary); }
+ &:disabled { opacity: 0.4; cursor: default; }
}
+.steer-sending { animation: steer-spin 1s linear infinite; }
+@keyframes steer-spin { to { transform: rotate(360deg); } }
+@media (prefers-reduced-motion: reduce) { .steer-sending { animation: none; } }
+
/* 富文本输入框容器 */
.rich-input-container {
position: relative;
@@ -3145,15 +3148,15 @@ const getImgSrc = (url: string) => {
:deep(.t-textarea__inner) {
width: 100%;
- max-height: 200px !important;
- min-height: 120px !important;
+ max-height: 152px !important;
+ min-height: 72px !important;
resize: none;
color: var(--td-text-color-primary, #000000e6);
font-size: 16px;
font-weight: 400;
line-height: 24px;
font-family: var(--app-font-family);
- padding: 12px 16px 56px 16px;
+ padding: 12px 16px;
border-radius: 0 0 12px 12px;
border: none;
box-sizing: border-box;
@@ -3182,18 +3185,14 @@ const getImgSrc = (url: string) => {
/* 控制栏 */
.control-bar {
- position: absolute;
- bottom: 12px;
- left: 16px;
- right: 16px;
+ position: relative;
+ margin: 0 16px 12px;
display: flex;
align-items: center;
justify-content: space-between;
gap: 8px;
flex-wrap: wrap;
- max-height: 56px;
z-index: 10;
- background: linear-gradient(to bottom, rgba(255, 255, 255, 0) 0%, var(--td-bg-color-container, #fff) 40%, var(--td-bg-color-container, #fff) 100%);
pointer-events: auto;
padding-top: 8px;
@@ -3212,6 +3211,8 @@ const getImgSrc = (url: string) => {
}
.control-btn {
+ border: 0;
+ font: inherit;
display: flex;
align-items: center;
justify-content: center;
@@ -3558,61 +3559,23 @@ const getImgSrc = (url: string) => {
gap: 8px;
}
-.stop-btn {
+.stop-btn, .send-btn {
width: 28px;
height: 28px;
padding: 0;
- background: rgba(16, 185, 129, 0.08);
- color: var(--td-brand-color);
- border: 1.5px solid rgba(16, 185, 129, 0.2);
- position: relative;
- display: flex;
- align-items: center;
- justify-content: center;
-
- &:hover {
- background: rgba(16, 185, 129, 0.12);
- border-color: var(--td-brand-color);
- }
-
- &:active {
- background: rgba(16, 185, 129, 0.15);
- }
-
- svg {
- display: none;
- }
-
- &::before {
- content: '';
- width: 12px;
- height: 12px;
- background: var(--td-brand-color);
- border-radius: 50%;
- display: block;
- animation: stopBtnPulse 1.5s ease-in-out infinite;
- }
-}
-
-@keyframes stopBtnPulse {
-
- 0%,
- 100% {
- transform: scale(1);
- opacity: 1;
- }
+ box-sizing: border-box;
+ font-size: 16px;
+ line-height: 1;
- 50% {
- transform: scale(0.75);
- opacity: 0.6;
+ &:focus-visible {
+ outline: 2px solid var(--td-brand-color);
+ outline-offset: 2px;
}
}
-.send-btn {
- width: 28px;
- height: 28px;
- padding: 0;
+.stop-btn, .send-btn {
background-color: var(--td-brand-color);
+ color: #fff;
&:hover:not(.disabled) {
background-color: var(--td-brand-color-active);
diff --git a/frontend/src/components/chat/ChatQuestionMinimap.test.ts b/frontend/src/components/chat/ChatQuestionMinimap.test.ts
index bbda23aa1..0d0aa2b50 100644
--- a/frontend/src/components/chat/ChatQuestionMinimap.test.ts
+++ b/frontend/src/components/chat/ChatQuestionMinimap.test.ts
@@ -76,9 +76,11 @@ test('does not paint the first question green on hover-open', () => {
assert.doesNotMatch(component, /question-minimap__row:hover,\s*\.question-minimap__row--active/)
})
-test('highlights the matching tick for the peaked question', () => {
+test('highlights visible messages without losing them when previewing another tick', () => {
assert.match(component, /peakId/)
- assert.match(component, /tick\.id === peakId/)
+ assert.match(component, /highlightedIds\.has\(tick\.id\)/)
+ assert.match(component, /new Set\(visibleIds\.value\)/)
+ assert.match(composable, /visibleMessageIds\(measured, el\.scrollTop, el\.clientHeight\)/)
})
test('keeps the preview card at 13px instead of inheriting the chat 20px type', () => {
diff --git a/frontend/src/components/chat/ChatQuestionMinimap.vue b/frontend/src/components/chat/ChatQuestionMinimap.vue
index 20d0ec432..e53781e3b 100644
--- a/frontend/src/components/chat/ChatQuestionMinimap.vue
+++ b/frontend/src/components/chat/ChatQuestionMinimap.vue
@@ -23,10 +23,10 @@
v-for="tick in ticks"
:key="tick.id"
class="question-minimap__tick"
- :class="{ 'question-minimap__tick--active': tick.id === peakId }"
+ :class="{ 'question-minimap__tick--active': highlightedIds.has(tick.id) }"
:style="{
top: `${tick.yPx}px`,
- transform: `translateY(-50%) scaleX(${tickDisplayScale(tick.yPx, mountainPointerY, tick.id === peakId)})`,
+ transform: `translateY(-50%) scaleX(${tickDisplayScale(tick.yPx, mountainPointerY, highlightedIds.has(tick.id))})`,
}"
/>
</button>
@@ -65,7 +65,7 @@ import {
questionDisplayText,
tickDisplayScale,
type ChatMessageLike,
- type UserQuestion,
+ type OutlineMessage,
} from '@/utils/chatQuestionMinimap'
const CLOSE_DELAY_MS = 150
@@ -94,6 +94,7 @@ const {
questions,
ticks,
activeId,
+ visibleIds,
anchoredIds,
trackHeight,
} = useChatQuestionMinimap({
@@ -114,6 +115,11 @@ const peakId = computed(() => {
const peakTurn = computed(() => (
questions.value.find((question) => question.id === peakId.value) ?? null
))
+const highlightedIds = computed(() => {
+ const ids = new Set(visibleIds.value)
+ if ((hoveredId.value || keyboardIndex.value >= 0) && peakId.value) ids.add(peakId.value)
+ return ids
+})
const peakYPx = computed(() => {
const tick = ticks.value.find((item) => item.id === peakId.value)
return tick?.yPx ?? 0
@@ -125,11 +131,13 @@ const mountainPointerY = computed(() => {
let closeTimer: number | null = null
-const questionText = (question: UserQuestion) => (
- questionDisplayText(question.content, t('chat.questionMinimapAttachmentPlaceholder'))
+const questionText = (question: OutlineMessage) => (
+ question.role === 'assistant'
+ ? answerPreviewText(question.content) || t('chat.thinking')
+ : questionDisplayText(question.content, t('chat.questionMinimapAttachmentPlaceholder'))
)
-const answerText = (question: UserQuestion) => (
+const answerText = (question: OutlineMessage) => (
answerPreviewText(question.answerContent)
)
diff --git a/frontend/src/composables/useChatQuestionMinimap.ts b/frontend/src/composables/useChatQuestionMinimap.ts
index 45952ef94..74d04f2fb 100644
--- a/frontend/src/composables/useChatQuestionMinimap.ts
+++ b/frontend/src/composables/useChatQuestionMinimap.ts
@@ -10,16 +10,17 @@ import {
} from 'vue'
import {
activeQuestionId,
- collectUserQuestions,
+ collectOutlineMessages,
isChatOverflowing,
mapQuestionTicks,
offsetFromScrollContent,
questionMinimapTrackHeight,
shouldShowQuestionMinimap,
viewportBand,
+ visibleMessageIds,
type ChatMessageLike,
type QuestionTick,
- type UserQuestion,
+ type OutlineMessage,
type ViewportBand,
} from '@/utils/chatQuestionMinimap'
@@ -28,19 +29,21 @@ export function useChatQuestionMinimap(options: {
messages: ChatMessageLike[] | Ref<ChatMessageLike[]>
}): {
visible: ComputedRef<boolean>
- questions: ComputedRef<UserQuestion[]>
+ questions: ComputedRef<OutlineMessage[]>
ticks: Ref<QuestionTick[]>
viewport: Ref<ViewportBand>
activeId: Ref<string | null>
+ visibleIds: Ref<Set<string>>
anchoredIds: Ref<Set<string>>
scrollbarGutterPx: Ref<number>
trackHeight: Ref<number>
} {
- const questions = computed(() => collectUserQuestions(toValue(options.messages)))
+ const questions = computed(() => collectOutlineMessages(toValue(options.messages)))
const overflowing = ref(false)
const ticks = ref<QuestionTick[]>([])
const viewport = ref<ViewportBand>({ topPx: 0, heightPx: 0 })
const activeId = ref<string | null>(null)
+ const visibleIds = ref<Set<string>>(new Set())
const anchoredIds = ref<Set<string>>(new Set())
const scrollbarGutterPx = ref(0)
const trackHeight = ref(0)
@@ -59,6 +62,7 @@ export function useChatQuestionMinimap(options: {
ticks.value = []
viewport.value = { topPx: 0, heightPx: 0 }
activeId.value = null
+ visibleIds.value = new Set()
anchoredIds.value = new Set()
scrollbarGutterPx.value = 0
trackHeight.value = 0
@@ -90,14 +94,17 @@ export function useChatQuestionMinimap(options: {
const measured = questions.value.flatMap((question) => {
const anchor = anchorByMessageId.get(question.id)
if (!anchor) return []
+ const rect = anchor.getBoundingClientRect()
+ if (rect.height <= 0) return []
return [{
id: question.id,
offsetTop: offsetFromScrollContent(
- anchor.getBoundingClientRect().top,
+ rect.top,
containerTop,
el.scrollTop,
),
+ offsetBottom: offsetFromScrollContent(rect.bottom, containerTop, el.scrollTop),
}]
})
@@ -105,6 +112,7 @@ export function useChatQuestionMinimap(options: {
ticks.value = mapQuestionTicks(measured, trackHeight.value)
viewport.value = viewportBand(el.scrollTop, el.clientHeight, el.scrollHeight, trackHeight.value)
activeId.value = activeQuestionId(measured, el.scrollTop)
+ visibleIds.value = visibleMessageIds(measured, el.scrollTop, el.clientHeight)
anchoredIds.value = new Set(measured.map((item) => item.id))
}
@@ -179,6 +187,7 @@ export function useChatQuestionMinimap(options: {
ticks,
viewport,
activeId,
+ visibleIds,
anchoredIds,
scrollbarGutterPx,
trackHeight,
diff --git a/frontend/src/composables/useChatStreamHandler.test.mjs b/frontend/src/composables/useChatStreamHandler.test.mjs
index 369f497a4..c862b0695 100644
--- a/frontend/src/composables/useChatStreamHandler.test.mjs
+++ b/frontend/src/composables/useChatStreamHandler.test.mjs
@@ -1,9 +1,33 @@
import assert from 'node:assert/strict'
import { readFileSync } from 'node:fs'
import test from 'node:test'
+import vm from 'node:vm'
+import ts from 'typescript'
+import { resetSteerTurnForReplay } from '../utils/steerStreamFork.ts'
const source = readFileSync(new URL('./useChatStreamHandler.ts', import.meta.url), 'utf8')
+test('replaying agent_query binds the first segment and preserves distinct row IDs', () => {
+ const messagesList = [
+ { id: 'a', role: 'assistant', request_id: 'r', steerForked: true, is_completed: true },
+ { id: 'u', role: 'user', request_id: 'r' },
+ { id: 'a:steer:1', assistant_message_id: 'a', role: 'assistant', request_id: 'r', is_completed: false },
+ ]
+ const start = source.indexOf("if (data.response_type === 'agent_query')")
+ const block = source.slice(start, source.indexOf('const isAgentOnlyResponse', start))
+ const replaySegments = new Map()
+ const process = vm.runInNewContext(ts.transpile(`(data) => { ${block} }`), {
+ messagesList, replaySegments, resetSteerTurnForReplay,
+ currentAssistantMessageId: { value: 'a' },
+ getTrailingIncompleteAssistant: () => [...messagesList].reverse().find(m => m.role === 'assistant' && !m.is_completed),
+ findLastMessage: fn => [...messagesList].reverse().find(fn),
+ log() {}, ensureAgentMessageShell() {}, bindServerTurnTimestamps() {}, onAgentQuery() {},
+ })
+ process({ response_type: 'agent_query', id: 'r', assistant_message_id: 'a' })
+ assert.deepEqual(messagesList.map(m => m.id), ['a', 'u', 'a:steer:1'])
+ assert.equal(replaySegments.get('r'), messagesList[0])
+})
+
test('failed tool results keep stdout/output instead of replacing it with the short error', () => {
assert.match(source, /toolCallEvent\.output = dataPayload\.output \|\| data\.content/)
assert.doesNotMatch(
diff --git a/frontend/src/composables/useChatStreamHandler.ts b/frontend/src/composables/useChatStreamHandler.ts
index 1fa61d143..21300413a 100644
--- a/frontend/src/composables/useChatStreamHandler.ts
+++ b/frontend/src/composables/useChatStreamHandler.ts
@@ -3,7 +3,7 @@ import { markRaw, nextTick, type Ref } from 'vue'
import { useI18n } from 'vue-i18n'
import { ensureRagPipelineHistoryStream } from '@/utils/rag-pipeline-history'
import { applyMessageCreatedAt, bindServerTurnTimestamps, ensureMessageCreatedAt } from '@/utils/messageTimestamp'
-import { expandSteerForksInHistory, forkAfterInjectedUser } from '@/utils/steerStreamFork'
+import { expandSteerForksInHistory, forkAfterInjectedUser, steerStepEvents, resetSteerTurnForReplay } from '@/utils/steerStreamFork'
export type ChatMessage = Record<string, unknown>
@@ -163,6 +163,8 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
return Array.isArray(refs) ? refs : []
}
+ const replaySegments = new Map<string, ChatMessage>()
+
/** Match the in-flight assistant row by request id or assistant message id. */
const resolveActiveAssistantMessage = (data: ChatMessage) => {
const dataId = data.id as string | undefined
@@ -171,6 +173,8 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
currentAssistantMessageId.value ||
undefined
+ if (dataId && replaySegments.has(dataId)) return replaySegments.get(dataId)
+
const matched = findLastMessage((item) => {
if (item.role !== 'assistant') return false
if (dataId && (item.request_id === dataId || item.id === dataId)) return true
@@ -326,6 +330,7 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
if (agentSteps && Array.isArray(agentSteps) && agentSteps.length > 0) {
agentSteps.forEach((rawStep) => {
const step = rawStep as ChatMessage
+ events.push(...steerStepEvents(step))
const stepTimestamp = step.timestamp ? new Date(String(step.timestamp)).getTime() : 0
const toolCalls = step.tool_calls
const hasToolCalls = toolCalls && Array.isArray(toolCalls) && toolCalls.length > 0
@@ -346,6 +351,10 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
})
}
const preambleText = step.thought && String(step.thought).trim() ? String(step.thought) : ''
+ if (preambleText && step.intermediate_answer) {
+ events.push({ type: 'answer', event_id: `step-${step.iteration}-answer`,
+ content: preambleText, done: true, intermediate_answer: true, timestamp: stepTimestamp || undefined })
+ }
if (preambleText && hasToolCalls) {
events.push({
type: 'answer',
@@ -941,8 +950,7 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
)
alreadyInList = Boolean(injectedUser)
if (!injectedUser) {
- // Queued messages live in the composer overlay, never as chat
- // rows, so the first delivery builds the bubble here.
+ // A restored event or promoted follow-up may arrive before its preview.
injectedUser = {
id: injectedId || steerId,
role: 'user',
@@ -953,6 +961,8 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
}
if (data.created_at) applyMessageCreatedAt(injectedUser, data.created_at)
}
+ if (injectedId) injectedUser.id = injectedId
+ if (data.created_at) applyMessageCreatedAt(injectedUser, data.created_at)
}
if (injectedUser) {
if (dataId && !injectedUser.request_id) injectedUser.request_id = dataId
@@ -962,6 +972,7 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
injectedUser,
steerId,
)
+ if (dataId && replaySegments.has(dataId)) replaySegments.set(dataId, continuation)
if (!alreadyInList) emitMessageCreated(injectedUser)
if (continuation !== message) {
emitMessageCreated(continuation)
@@ -972,6 +983,7 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
break
}
case 'complete': {
+ if (dataId) replaySegments.delete(dataId)
log('[Agent] Complete event received')
applyFinalArtifactContent(message, (dataPayload as any)?.final_content)
loading.value = false
@@ -1040,6 +1052,9 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
})
if (data.response_type === 'agent_query') {
+ if (data.id) replaySegments.delete(String(data.id))
+ const replay = resetSteerTurnForReplay(messagesList, String(data.id || ''))
+ if (replay) replaySegments.set(String(data.id), replay)
if (data.id) {
const earlyMsg = getTrailingIncompleteAssistant()
if (earlyMsg) earlyMsg.request_id = data.id
@@ -1055,7 +1070,7 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
request_id: (data.data as ChatMessage | undefined)?.request_id,
})
- let existingMessage = findLastMessage(
+ let existingMessage = replay || findLastMessage(
(item) =>
item.role === 'assistant' &&
(item.id === data.id || item.request_id === data.id),
@@ -1144,6 +1159,7 @@ export function useChatStreamHandler(options: UseChatStreamHandlerOptions) {
log('[Stop Event] Generation stopped')
const stoppedMessage = resolveActiveAssistantMessage(data)
if (stoppedMessage) markAssistantStopped(stoppedMessage)
+ if (data.id) replaySegments.delete(String(data.id))
loading.value = false
isReplying.value = false
currentAssistantMessageId.value = ''
diff --git a/frontend/src/i18n/locales/en-US.ts b/frontend/src/i18n/locales/en-US.ts
index a75e7b6ac..e0f74e051 100755
--- a/frontend/src/i18n/locales/en-US.ts
+++ b/frontend/src/i18n/locales/en-US.ts
@@ -2445,9 +2445,13 @@ export default {
noModel: 'No available models',
stopGeneration: 'Stop Generation',
send: 'Send',
- steerQueueSendNow: 'Send now',
+ steerCurrent: 'Supplement current task',
+ steerAfter: 'Send after completion',
+ steerAccepted: 'Added during this task',
+ steerRetry: 'Retry sending',
+ steerQueueSendNow: 'Supplement current task',
steerQueueWaiting: 'Will send after the current answer finishes',
- steerQueueInjecting: 'Waiting for the next thinking round',
+ steerQueueInjecting: 'Waiting to apply update',
messages: {
enterContent: 'Please enter content first!',
replying: 'Currently replying, please try again later!',
diff --git a/frontend/src/i18n/locales/ja-JP.ts b/frontend/src/i18n/locales/ja-JP.ts
index f086d3b49..d752683f1 100644
--- a/frontend/src/i18n/locales/ja-JP.ts
+++ b/frontend/src/i18n/locales/ja-JP.ts
@@ -2445,9 +2445,13 @@ export default {
noModel: '利用可能なモデルがありません',
stopGeneration: '生成を停止',
send: '送信',
- steerQueueSendNow: '今すぐ送信',
+ steerCurrent: '現在のタスクに追加',
+ steerAfter: '完了後に送信',
+ steerAccepted: 'このタスクへの追加',
+ steerRetry: '再送信',
+ steerQueueSendNow: '現在のタスクに追加',
steerQueueWaiting: '現在の回答が終わってから送信されます',
- steerQueueInjecting: '次の思考ラウンドを待っています',
+ steerQueueInjecting: '追加内容の適用待ち',
messages: {
enterContent: '先に内容を入力してください。',
replying: '回答中です。しばらくしてから再試行してください。',
diff --git a/frontend/src/i18n/locales/ko-KR.ts b/frontend/src/i18n/locales/ko-KR.ts
index f7ccff4eb..8b5fdbf5f 100755
--- a/frontend/src/i18n/locales/ko-KR.ts
+++ b/frontend/src/i18n/locales/ko-KR.ts
@@ -4511,9 +4511,13 @@ export default {
noModel: '사용 가능한 모델 없음',
stopGeneration: '생성 중지',
send: '전송',
- steerQueueSendNow: '지금 보내기',
+ steerCurrent: '현재 작업에 추가',
+ steerAfter: '완료 후 보내기',
+ steerAccepted: '현재 작업에 추가됨',
+ steerRetry: '다시 보내기',
+ steerQueueSendNow: '현재 작업에 추가',
steerQueueWaiting: '현재 답변이 끝난 뒤에 전송됩니다',
- steerQueueInjecting: '다음 사고 라운드를 기다리는 중',
+ steerQueueInjecting: '추가 내용 반영 대기 중',
webSearchDisabledByAgent: '현재 에이전트에 대해서는 네트워크 검색이 비활성화되어 있습니다.',
kbLockedByAgent: '현재 에이전트가 지식베이스 구성을 잠갔습니다.',
kbDisabledByAgent: '현재 에이전트가 지식베이스 기능을 비활성화했습니다.',
diff --git a/frontend/src/i18n/locales/ru-RU.ts b/frontend/src/i18n/locales/ru-RU.ts
index 1e6df4431..8d1b3f16b 100755
--- a/frontend/src/i18n/locales/ru-RU.ts
+++ b/frontend/src/i18n/locales/ru-RU.ts
@@ -4511,9 +4511,13 @@ export default {
noModel: 'Нет доступных моделей',
stopGeneration: 'Остановить генерацию',
send: 'Отправить',
- steerQueueSendNow: 'Отправить сейчас',
+ steerCurrent: 'Дополнить текущую задачу',
+ steerAfter: 'Отправить после завершения',
+ steerAccepted: 'Дополнение к текущей задаче',
+ steerRetry: 'Повторить отправку',
+ steerQueueSendNow: 'Дополнить текущую задачу',
steerQueueWaiting: 'Отправится после завершения текущего ответа',
- steerQueueInjecting: 'Ожидание следующего раунда рассуждения',
+ steerQueueInjecting: 'Ожидание применения дополнения',
webSearchDisabledByAgent: 'Web search is disabled by the current agent',
kbLockedByAgent: 'Knowledge base configuration is locked by the current agent',
kbDisabledByAgent: 'Knowledge base is disabled by the current agent',
diff --git a/frontend/src/i18n/locales/zh-CN.ts b/frontend/src/i18n/locales/zh-CN.ts
index 89b9188e6..b00f390ba 100755
--- a/frontend/src/i18n/locales/zh-CN.ts
+++ b/frontend/src/i18n/locales/zh-CN.ts
@@ -4513,9 +4513,13 @@ export default {
noModel: '暂无可用模型',
stopGeneration: '停止生成',
send: '发送',
- steerQueueSendNow: '立即发送',
+ steerCurrent: '补充当前任务',
+ steerAfter: '完成后发送',
+ steerAccepted: '本轮补充',
+ steerRetry: '重试发送',
+ steerQueueSendNow: '补充当前任务',
steerQueueWaiting: '当前回答结束后发送',
- steerQueueInjecting: '等待下一轮思考',
+ steerQueueInjecting: '等待接收补充',
webSearchDisabledByAgent: '当前智能体已禁用网络搜索',
kbLockedByAgent: '当前智能体已锁定知识库配置',
kbDisabledByAgent: '当前智能体已禁用知识库功能',
diff --git a/frontend/src/utils/chatQuestionMinimap.test.ts b/frontend/src/utils/chatQuestionMinimap.test.ts
index 2f89cccb4..2f23c74a6 100644
--- a/frontend/src/utils/chatQuestionMinimap.test.ts
+++ b/frontend/src/utils/chatQuestionMinimap.test.ts
@@ -10,7 +10,7 @@ import {
CURRENT_TICK_SCALE,
activeQuestionId,
answerPreviewText,
- collectUserQuestions,
+ collectOutlineMessages,
isChatOverflowing,
mapQuestionTicks,
nearestTickId,
@@ -21,6 +21,7 @@ import {
tickDisplayScale,
tickMountainScale,
viewportBand,
+ visibleMessageIds,
} from './chatQuestionMinimap.ts'
test('isChatOverflowing is false when content fits or matches the viewport', () => {
@@ -48,8 +49,8 @@ test('shouldShowQuestionMinimap requires overflow and at least two questions', (
assert.equal(shouldShowQuestionMinimap(false, 4), false)
})
-test('collectUserQuestions keeps loaded user turns with ids, in list order', () => {
- const questions = collectUserQuestions([
+test('collectOutlineMessages includes both user and assistant messages in order', () => {
+ const questions = collectOutlineMessages([
{ role: 'assistant', id: 'a1', content: 'hi' },
{ role: 'user', content: 'no id yet' },
{ role: 'user', id: 'u1', content: 'first', images: [{}] },
@@ -59,10 +60,29 @@ test('collectUserQuestions keeps loaded user turns with ids, in list order', ()
assert.deepEqual(
questions.map((q) => ({ id: q.id, hasAttachments: q.hasAttachments, answerContent: q.answerContent })),
[
+ { id: 'a1', hasAttachments: false, answerContent: '' },
{ id: 'u1', hasAttachments: true, answerContent: 'answer one' },
+ { id: 'a2', hasAttachments: false, answerContent: '' },
{ id: 'u2', hasAttachments: true, answerContent: '' },
],
)
+ assert.deepEqual(questions.map(q => q.role), ['assistant', 'user', 'assistant', 'user'])
+})
+
+test('all messages intersecting the viewport are highlighted, including partial and long answers', () => {
+ const items = [
+ { id: 'above', offsetTop: 0, offsetBottom: 100 },
+ { id: 'user', offsetTop: 90, offsetBottom: 125 },
+ { id: 'assistant', offsetTop: 140, offsetBottom: 280 },
+ { id: 'next-user', offsetTop: 290, offsetBottom: 330 },
+ { id: 'below', offsetTop: 300, offsetBottom: 400 },
+ { id: 'hidden', offsetTop: 150, offsetBottom: 150 },
+ ]
+ assert.deepEqual([...visibleMessageIds(items, 100, 200)], ['user', 'assistant', 'next-user'])
+ assert.deepEqual([...visibleMessageIds([
+ { id: 'long-answer', offsetTop: 0, offsetBottom: 2000 },
+ ], 500, 200)], ['long-answer'])
+ assert.equal(visibleMessageIds(items, 100, 0).size, 0)
})
test('answerPreviewText strips markdown and stays empty when there is no answer yet', () => {
diff --git a/frontend/src/utils/chatQuestionMinimap.ts b/frontend/src/utils/chatQuestionMinimap.ts
index fd6e667d8..0f1dcde5f 100644
--- a/frontend/src/utils/chatQuestionMinimap.ts
+++ b/frontend/src/utils/chatQuestionMinimap.ts
@@ -30,8 +30,9 @@ export type ChatMessageLike = {
attachments?: unknown[]
}
-export type UserQuestion = {
+export type OutlineMessage = {
id: string
+ role: 'user' | 'assistant'
content: string
hasAttachments: boolean
answerContent: string
@@ -56,15 +57,15 @@ export function shouldShowQuestionMinimap(overflowing: boolean, questionCount: n
return overflowing && questionCount >= 2
}
-export function collectUserQuestions(messages: ChatMessageLike[]): UserQuestion[] {
- const questions: UserQuestion[] = []
+export function collectOutlineMessages(messages: ChatMessageLike[]): OutlineMessage[] {
+ const questions: OutlineMessage[] = []
for (let index = 0; index < messages.length; index++) {
const message = messages[index]
- if (message.role !== 'user' || !message.id) continue
+ if ((message.role !== 'user' && message.role !== 'assistant') || !message.id) continue
let answerContent = ''
- for (let next = index + 1; next < messages.length; next++) {
+ for (let next = index + 1; message.role === 'user' && next < messages.length; next++) {
const following = messages[next]
if (following.role === 'user') break
if (following.role === 'assistant') {
@@ -75,6 +76,7 @@ export function collectUserQuestions(messages: ChatMessageLike[]): UserQuestion[
questions.push({
id: message.id,
+ role: message.role,
content: message.content ?? '',
hasAttachments:
(message.images?.length ?? 0) > 0 || (message.attachments?.length ?? 0) > 0,
@@ -85,6 +87,19 @@ export function collectUserQuestions(messages: ChatMessageLike[]): UserQuestion[
return questions
}
+/** Include partially visible messages and long answers spanning the viewport. */
+export function visibleMessageIds(
+ items: Array<{ id: string; offsetTop: number; offsetBottom: number }>,
+ scrollTop: number,
+ clientHeight: number,
+): Set<string> {
+ if (clientHeight <= 0) return new Set()
+ const bottom = scrollTop + clientHeight
+ return new Set(items.filter(item => (
+ item.offsetBottom > item.offsetTop && item.offsetBottom > scrollTop && item.offsetTop < bottom
+ )).map(item => item.id))
+}
+
export function questionDisplayText(
content: string | undefined,
attachmentPlaceholder: string,
diff --git a/frontend/src/utils/chatSubmitShortcut.test.ts b/frontend/src/utils/chatSubmitShortcut.test.ts
new file mode 100644
index 000000000..e8b1867eb
--- /dev/null
+++ b/frontend/src/utils/chatSubmitShortcut.test.ts
@@ -0,0 +1,25 @@
+import assert from 'node:assert/strict'
+import test from 'node:test'
+import { chatSubmitShortcut } from './chatSubmitShortcut'
+
+const enter = { key: 'Enter', keyCode: 13, isComposing: false, shiftKey: false, ctrlKey: false, altKey: false, metaKey: false }
+test('Enter queues during a run and Alt+Enter injects only while steering is available', () => {
+ assert.equal(chatSubmitShortcut(enter, true), 'after')
+ assert.equal(chatSubmitShortcut(enter, false), 'inject')
+ assert.equal(chatSubmitShortcut({ ...enter, altKey: true }, true), 'inject')
+ assert.equal(chatSubmitShortcut({ ...enter, altKey: true }, false), undefined)
+})
+test('Cmd+Enter injects during a run and sends normally when idle', () => {
+ const cmdEnter = { ...enter, metaKey: true }
+ assert.equal(chatSubmitShortcut(cmdEnter, true), 'inject')
+ assert.equal(chatSubmitShortcut(cmdEnter, false), 'inject')
+ for (const modifiers of [{ shiftKey: true }, { ctrlKey: true }, { isComposing: true }, { keyCode: 229 }]) {
+ assert.equal(chatSubmitShortcut({ ...cmdEnter, ...modifiers }, true), undefined)
+ }
+})
+test('newlines and IME confirmation never submit', () => {
+ for (const modifiers of [{ shiftKey: true }, { ctrlKey: true }, { isComposing: true }, { keyCode: 229 }]) {
+ assert.equal(chatSubmitShortcut({ ...enter, ...modifiers }, true), undefined)
+ }
+ assert.equal(chatSubmitShortcut({ ...enter, key: 'a', keyCode: 65 }, true), undefined)
+})
diff --git a/frontend/src/utils/chatSubmitShortcut.ts b/frontend/src/utils/chatSubmitShortcut.ts
new file mode 100644
index 000000000..6fd1551ce
--- /dev/null
+++ b/frontend/src/utils/chatSubmitShortcut.ts
@@ -0,0 +1,11 @@
+/** Enter queues during a run; Cmd/Alt+Enter steers. Preserve newline and IME behavior. */
+export function chatSubmitShortcut(
+ event: Pick<KeyboardEvent, 'key' | 'keyCode' | 'isComposing' | 'shiftKey' | 'ctrlKey' | 'altKey' | 'metaKey'>,
+ canSteer: boolean,
+): 'inject' | 'after' | undefined {
+ if (event.isComposing || event.keyCode === 229 || (event.key !== 'Enter' && event.keyCode !== 13)) return
+ if (event.shiftKey || event.ctrlKey) return
+ if (event.metaKey) return 'inject'
+ if (event.altKey) return canSteer ? 'inject' : undefined
+ return canSteer ? 'after' : 'inject'
+}
diff --git a/frontend/src/utils/steerStreamFork.test.mjs b/frontend/src/utils/steerStreamFork.test.mjs
index a0b7dae7f..aed475a49 100644
--- a/frontend/src/utils/steerStreamFork.test.mjs
+++ b/frontend/src/utils/steerStreamFork.test.mjs
@@ -1,6 +1,6 @@
import assert from 'node:assert/strict'
import test from 'node:test'
-import { expandSteerForksInHistory, forkAfterInjectedUser } from './steerStreamFork.ts'
+import { expandSteerForksInHistory, forkAfterInjectedUser, resetSteerTurnForReplay, steerStepEvents, isAssistantTurnComplete, previewSteerMessage, discardSteerPreview } from './steerStreamFork.ts'
test('inject forks later events onto a new assistant below the user bubble', () => {
const assistant = {
@@ -219,3 +219,92 @@ test('expanding an already split transcript is a no-op', () => {
assert.equal(twice[3].assistant_message_id, 'a0')
assert.equal(twice[3].id, once[3].id)
})
+
+test('explicit boundaries preserve drafts, repeated inputs and delivery order despite equal timestamps', () => {
+ const users = ['u2', 'u1'].map(id => ({ id, role: 'user', request_id: 'r', content: 'revise', created_at: '1970-01-01T00:00:00Z' }))
+ const assistant = { id: 'a', role: 'assistant', request_id: 'r', content: 'final', is_completed: true,
+ used_memories: [{ id: 'memory' }], agentEventStream: [
+ { type: 'answer', content: 'first draft', done: true, intermediate_answer: true },
+ { type: 'user_message_injected', user_message_id: 'u1' },
+ { type: 'answer', content: 'second draft', done: true, intermediate_answer: true },
+ { type: 'user_message_injected', user_message_id: 'u2' },
+ { type: 'answer', content: 'final', done: true },
+ ] }
+ const out = expandSteerForksInHistory([assistant, ...users])
+ assert.deepEqual(out.map(m => m.content), ['first draft', 'revise', 'second draft', 'revise', 'final'])
+ assert.equal(out[1].id, 'u1')
+ assert.equal(out[3].id, 'u2')
+ assert.equal(out[1].isSteer, true)
+ assert.equal(out.filter(m => m.used_memories?.length).length, 1)
+})
+
+test('sealing a draft finishes its text stream without discarding the draft', () => {
+ const source = { id: 'a', role: 'assistant', request_id: 'r', content: 'draft', agentEventStream: [{ type: 'answer', content: 'draft', done: false }] }
+ const list = [source]
+ forkAfterInjectedUser(list, source, { id: 'u1', role: 'user', request_id: 'r', content: 'revise' }, 's1')
+ assert.equal(source.content, 'draft')
+ assert.equal(source.agentEventStream[0].done, true)
+})
+
+test('replay walks successive existing segments rather than duplicating earlier events on the tail', () => {
+ const a = { id: 'a', role: 'assistant', request_id: 'r', steerForked: true, content: 'draft', agentEventStream: [{ type: 'answer', content: 'draft' }] }
+ const u1 = { id: 'u1', role: 'user', request_id: 'r', content: 'one' }
+ const middle = { id: 'a:1', role: 'assistant', request_id: 'r', steerForked: true, agentEventStream: [] }
+ const u2 = { id: 'u2', role: 'user', request_id: 'r', content: 'two' }
+ const tail = { id: 'a:2', role: 'assistant', request_id: 'r', is_completed: false, agentEventStream: [] }
+ const list = [a, u1, middle, u2, tail]
+ let cursor = resetSteerTurnForReplay(list, 'r')
+ assert.equal(cursor, a)
+ cursor.agentEventStream.push({ type: 'answer', content: 'draft' })
+ cursor = forkAfterInjectedUser(list, cursor, u1, 's1')
+ assert.equal(cursor, middle)
+ cursor.agentEventStream.push({ type: 'thinking', content: 'work' })
+ cursor = forkAfterInjectedUser(list, cursor, u2, 's2')
+ assert.equal(cursor, tail)
+ assert.equal(list.length, 5)
+ assert.equal(a.agentEventStream.length, 1)
+ assert.equal(middle.agentEventStream.length, 1)
+ assert.equal(tail.agentEventStream.length, 0)
+ assert.equal(a.is_completed, true)
+ assert.equal(middle.is_completed, true)
+ assert.equal(tail.is_completed, false)
+})
+
+test('step boundary events retain the server delivery order', () => {
+ assert.deepEqual(steerStepEvents({ user_messages_before: ['u2', 'u1'] }).map(e => e.user_message_id), ['u2', 'u1'])
+})
+
+test('only run completion ends the task; draft completion and sealed prefixes do not', () => {
+ assert.equal(isAssistantTurnComplete({ is_completed: false, agentEventStream: [{ type: 'answer', done: true }] }), false)
+ assert.equal(isAssistantTurnComplete({ is_completed: true, steerForked: true }), false)
+ assert.equal(isAssistantTurnComplete({ agentEventStream: [{ type: 'agent_complete' }] }), true)
+ assert.equal(isAssistantTurnComplete({ agentEventStream: [{ type: 'stop' }] }), true)
+ assert.equal(isAssistantTurnComplete({ is_completed: true }), true)
+})
+
+
+test('inject appears immediately and the delivery receipt reuses its bubble', () => {
+ const assistant = { id: 'a', role: 'assistant', request_id: 'r', is_completed: false, agentEventStream: [{ type: 'thinking', done: false }] }
+ const list = [assistant]
+ const preview = previewSteerMessage(list, { steer_id: 's', content: '补充' })
+ assert.equal(list[1], preview)
+ assert.equal(assistant.is_completed, false)
+ assert.equal(assistant.agentEventStream[0].done, false)
+ assert.equal(previewSteerMessage(list, { steer_id: 's', content: '补充' }), preview)
+ assert.equal(list.length, 2)
+ forkAfterInjectedUser(list, assistant, preview, 's')
+ assert.equal(list.length, 3)
+ assert.equal(list[1], preview)
+ assert.equal(preview._steerPending, undefined)
+ discardSteerPreview(list, 's')
+ assert.equal(list.length, 3, 'removal must never discard a consumed message')
+})
+
+test('failed promotion can restore its queue without leaving a duplicate bubble', () => {
+ const list = []
+ previewSteerMessage(list, { steer_id: 's', content: '补充' })
+ discardSteerPreview(list, 's')
+ assert.equal(list.length, 0)
+ previewSteerMessage(list, { steer_id: 's', content: '补充' })
+ assert.equal(list.length, 1)
+})
diff --git a/frontend/src/utils/steerStreamFork.ts b/frontend/src/utils/steerStreamFork.ts
index 8592255d6..23c396780 100644
--- a/frontend/src/utils/steerStreamFork.ts
+++ b/frontend/src/utils/steerStreamFork.ts
@@ -2,6 +2,24 @@ import { markRaw } from 'vue'
export type ChatMessage = Record<string, unknown>
+/** Display a steer immediately; the receipt, not this preview, splits execution. */
+export function previewSteerMessage(list: ChatMessage[], item: { steer_id: string; content: string; mentioned_items?: unknown[] }): ChatMessage {
+ const existing = list.find(m => m.role === 'user' && m.steer_id === item.steer_id)
+ if (existing) return existing
+ const message: ChatMessage = {
+ id: `steer-user-${item.steer_id}`, steer_id: item.steer_id,
+ role: 'user', content: item.content, mentioned_items: item.mentioned_items || [],
+ isSteer: true, is_completed: true, _steerPending: true,
+ }
+ list.push(message)
+ return message
+}
+
+export function discardSteerPreview(list: ChatMessage[], steerId: string): void {
+ const index = list.findIndex(m => m.steer_id === steerId && m._steerPending)
+ if (index >= 0) list.splice(index, 1)
+}
+
/**
* The row id the server knows this assistant message by.
*
@@ -28,7 +46,7 @@ export function sealAssistantSegment(message: ChatMessage): void {
for (const raw of stream) {
if (!raw || typeof raw !== 'object') continue
const event = raw as ChatMessage
- if (event.type === 'thinking') {
+ if (event.type === 'thinking' || event.type === 'answer') {
event.done = true
event.thinking = false
}
@@ -48,6 +66,9 @@ export function forkAfterInjectedUser(
// "Already split" means the bubble sits *between* two segments of this turn.
// A bubble merely sitting above the live assistant is the opposite case —
// it still has to be moved underneath it.
+ userMessage.isSteer = true
+ delete userMessage._steerPending
+ delete userMessage._steerFailed
const existingIdx = list.indexOf(userMessage)
if (existingIdx > 0) {
const before = list[existingIdx - 1]
@@ -56,9 +77,14 @@ export function forkAfterInjectedUser(
before?.role === 'assistant' &&
before.request_id === sourceAssistant.request_id &&
after?.role === 'assistant' &&
- !after.is_completed &&
after.request_id === sourceAssistant.request_id
) {
+ sealAssistantSegment(before)
+ if (after._steerReplayPending) {
+ delete after._steerReplayPending
+ after.is_completed = false
+ after.steerForked = false
+ }
return after
}
}
@@ -122,7 +148,7 @@ function eventTime(event: ChatMessage): number {
function isTrailingTurnEvent(event: ChatMessage): boolean {
if (event.type === 'agent_complete' || event.type === 'stop') return true
- return event.type === 'answer' && event.done === true && !event.superseded
+ return event.type === 'answer' && event.done === true && !event.superseded && !event.intermediate_answer
}
export function expandSteerForksInHistory(messages: ChatMessage[]): ChatMessage[] {
@@ -159,20 +185,34 @@ export function expandSteerForksInHistory(messages: ChatMessage[]): ChatMessage[
}
const stream = Array.isArray(item.agentEventStream) ? (item.agentEventStream as ChatMessage[]) : []
+ const boundaryIDs = stream.filter(e => e.type === 'user_message_injected').map(e => e.user_message_id)
+ const hasBoundaries = boundaryIDs.length > 0
+ if (hasBoundaries) {
+ forks.sort((a, b) => {
+ const ai = boundaryIDs.indexOf(a.id), bi = boundaryIDs.indexOf(b.id)
+ return (ai < 0 ? Infinity : ai) - (bi < 0 ? Infinity : bi)
+ })
+ }
const cuts = forks.map((user) => {
const parsed = Date.parse(String(user.created_at || ''))
return Number.isNaN(parsed) ? 0 : parsed
})
const buckets: ChatMessage[][] = Array.from({ length: forks.length + 1 }, () => [])
const trailing: ChatMessage[] = []
+ let segmentIndex = 0
for (const event of stream) {
+ if (event.type === 'user_message_injected') {
+ const index = forks.findIndex(user => user.id === event.user_message_id)
+ if (index >= 0) segmentIndex = index + 1
+ continue
+ }
if (isTrailingTurnEvent(event)) {
trailing.push(event)
continue
}
const t = eventTime(event)
- let idx = 0
- while (idx < cuts.length && t >= cuts[idx]) idx += 1
+ let idx = hasBoundaries ? segmentIndex : 0
+ while (!hasBoundaries && idx < cuts.length && t >= cuts[idx]) idx += 1
buckets[idx].push(event)
}
buckets[buckets.length - 1].push(...trailing)
@@ -197,6 +237,7 @@ export function expandSteerForksInHistory(messages: ChatMessage[]): ChatMessage[
assistant_message_id: persistedId,
}
segment.agentEventStream = markRaw(buckets[s])
+ if (s > 0) segment.used_memories = undefined
if (isLast) {
segment.content = original.content
segment.artifacts = original.artifacts
@@ -207,15 +248,54 @@ export function expandSteerForksInHistory(messages: ChatMessage[]): ChatMessage[
} else {
segment.steerForked = true
segment.is_completed = true
- segment.content = ''
+ segment.content = buckets[s].filter(e => e.type === 'answer' && !e.superseded).map(e => e.content || '').join('')
segment.artifacts = undefined
segment.usage = undefined
segment.knowledge_references = []
}
out.push(segment)
- if (s < forks.length) out.push(forks[s])
+ if (s < forks.length) {
+ forks[s].isSteer = true
+ out.push(forks[s])
+ }
}
i = j - 1
}
return out
}
+
+export function steerStepEvents(step: ChatMessage): ChatMessage[] {
+ const events: ChatMessage[] = []
+ if (Array.isArray(step.user_messages_before)) {
+ for (const id of step.user_messages_before) {
+ events.push({ type: 'user_message_injected', user_message_id: id })
+ }
+ }
+ return events
+}
+
+/** A continue-stream replays from offset zero. Keep row identities, but rebuild
+ * every segment from that log rather than appending old events onto the tail. */
+export function resetSteerTurnForReplay(list: ChatMessage[], requestId: string): ChatMessage | undefined {
+ const segments = list.filter(m => m.role === 'assistant' && m.request_id === requestId)
+ if (!segments.some(m => m.steerForked)) return undefined
+ for (const segment of segments) {
+ segment.content = ''
+ segment.agentEventStream = []
+ segment._eventMap = new Map()
+ segment._pendingToolCalls = new Map()
+ segment.is_completed = true
+ segment._steerReplayPending = true
+ }
+ segments[0].is_completed = false
+ segments[0].steerForked = false
+ delete segments[0]._steerReplayPending
+ return segments[0]
+}
+
+/** Individual answer.done events close a message, not the running task. */
+export function isAssistantTurnComplete(message: ChatMessage | undefined): boolean {
+ if (!message || message.steerForked) return false
+ return Boolean(message.is_completed) || (Array.isArray(message.agentEventStream) &&
+ message.agentEventStream.some(e => e.type === 'agent_complete' || e.type === 'stop'))
+}
diff --git a/frontend/src/views/chat/components/AgentStreamDisplay.style.test.mjs b/frontend/src/views/chat/components/AgentStreamDisplay.style.test.mjs
index ddb85de20..0c05e9a4c 100644
--- a/frontend/src/views/chat/components/AgentStreamDisplay.style.test.mjs
+++ b/frontend/src/views/chat/components/AgentStreamDisplay.style.test.mjs
@@ -3,10 +3,32 @@ import { readFileSync } from 'node:fs'
import { dirname, join } from 'node:path'
import { fileURLToPath } from 'node:url'
import test from 'node:test'
+import vm from 'node:vm'
+import { computed, reactive } from 'vue'
+import { isAssistantTurnComplete } from '../../../utils/steerStreamFork.ts'
const here = dirname(fileURLToPath(import.meta.url))
const source = readFileSync(join(here, 'AgentStreamDisplay.vue'), 'utf8')
+test('streamed responses do not mount a hidden Wiki drawer that steals composer focus', () => {
+ assert.match(source, /<t-drawer v-if="wikiDrawerVisible" v-model:visible="wikiDrawerVisible"/)
+})
+
+test('steer seals animation while keeping the preceding timeline open without completion', () => {
+ const stateBlock = source.slice(source.indexOf('const isConversationDone ='), source.indexOf('const streamingMermaidSvgCache ='))
+ const collapseBlock = source.slice(source.indexOf('const shouldShowCollapsedSteps ='), source.indexOf('// Once the steps collapse'))
+ const props = reactive({ ragMode: false, session: { steerForked: true, is_completed: true } })
+ const state = vm.runInNewContext(`${stateBlock}\n${collapseBlock}\n({ isConversationDone, isSegmentDone, shouldShowCollapsedSteps })`, {
+ props, computed, isAssistantTurnComplete, intermediateStepsCount: { value: 2 },
+ })
+ assert.equal(state.isSegmentDone.value, true)
+ assert.equal(state.isConversationDone.value, false)
+ assert.equal(state.shouldShowCollapsedSteps.value, false)
+ props.session.steerForked = false
+ assert.equal(state.isConversationDone.value, true)
+ assert.equal(state.shouldShowCollapsedSteps.value, true)
+})
+
test('agent steps use compact muted timeline styling', () => {
assert.match(source, /--agent-step-text-size:\s*14px/)
assert.match(source, /--agent-step-summary-size:\s*13px/)
@@ -50,7 +72,7 @@ test('streaming tool log uses the same timeline structure', () => {
test('final done row uses an existing common translation key', () => {
assert.match(source, /t\('common\.finish'\)/)
assert.doesNotMatch(source, /\$t\('common\.done'\)/)
- assert.match(source, /'tree-child-last': !isConversationDone && index === visibleIntermediateEvents\.length - 1/)
+ assert.match(source, /'tree-child-last': !isSegmentDone && index === visibleIntermediateEvents\.length - 1/)
})
test('tool rows use line icon names instead of legacy asset masks', () => {
@@ -71,7 +93,7 @@ test('tool rows use line icon names instead of legacy asset masks', () => {
test('rag mode delegates pre-answer loading to pipeline and adds no row after answer starts', () => {
assert.match(source, /if \(props\.ragMode \|\| hasAnswerStarted\.value\) return false/)
- assert.match(source, /v-if="!ragMode \|\| displayEvents\.length > 0 \|\| showAgentActivityIndicator"/)
+ assert.match(source, /v-if="displayEvents\.length > 0 \|\| showMemoryRow \|\| showAgentActivityIndicator"/)
assert.doesNotMatch(source, /ChatActivityIndicator/)
})
@@ -119,7 +141,7 @@ test('pending tool rows do not render an extra axis dot', () => {
})
test('agent mode shows a native placeholder before answer whenever nothing is pending', () => {
- assert.match(source, /if \(isConversationDone\.value\) return false/)
+ assert.match(source, /if \(isSegmentDone\.value\) return false/)
assert.match(source, /return !hasPendingStreamingActivity\.value/)
assert.match(source, /const hasPendingStreamingActivity = computed/)
assert.match(source, /event\.thinking === true \|\| isThinkingActive\(event\.event_id\)/)
diff --git a/frontend/src/views/chat/components/AgentStreamDisplay.vue b/frontend/src/views/chat/components/AgentStreamDisplay.vue
index cebe24a5b..3df5311cb 100644
--- a/frontend/src/views/chat/components/AgentStreamDisplay.vue
+++ b/frontend/src/views/chat/components/AgentStreamDisplay.vue
@@ -1,5 +1,5 @@
<template>
- <div ref="rootElement" class="agent-stream-display" :class="{ 'is-embedded': embeddedMode, 'is-rag-mode': ragMode }">
+ <div ref="rootElement" class="agent-stream-display" :class="{ 'is-embedded': embeddedMode, 'is-rag-mode': ragMode, 'is-steer-prefix': session?.steerForked }">
<!-- Collapsed intermediate steps (tree root) -->
<div v-if="shouldShowCollapsedSteps" class="tree-container">
@@ -28,7 +28,7 @@
/>
<template v-for="(event, index) in visibleIntermediateEvents" :key="getEventKey(event, index)">
<div v-if="event && event.type" class="tree-child"
- :class="{ 'tree-child-last': !isConversationDone && index === visibleIntermediateEvents.length - 1 }">
+ :class="{ 'tree-child-last': !isSegmentDone && index === visibleIntermediateEvents.length - 1 }">
<div class="tree-branch"></div>
<div class="tree-child-content">
<!-- Plan Task Change Event -->
@@ -246,9 +246,9 @@
</div>
<!-- Event Stream (non-tree mode: before answer starts, or answer events) -->
- <div v-if="!ragMode || displayEvents.length > 0 || showAgentActivityIndicator" ref="streamingStepsContainer"
+ <div v-if="displayEvents.length > 0 || showMemoryRow || showAgentActivityIndicator" ref="streamingStepsContainer"
class="streaming-steps-container" :class="{
- 'streaming-steps-constrained': !answerEverStarted && !isConversationDone,
+ 'streaming-steps-constrained': !answerEverStarted && !isSegmentDone,
'is-streaming-timeline': showStreamingTimeline
}">
<!-- Recalled memory leads the timeline: it is what the turn knew before it
@@ -546,7 +546,8 @@
<picturePreview :reviewImg="imagePreviewVisible" :reviewUrl="imagePreviewUrl" @closePreImg="closeImagePreview" />
<!-- Wiki Page Detail Drawer -->
- <t-drawer v-model:visible="wikiDrawerVisible" :header="wikiDrawerPage?.title || ''" size="480px" :footer="false"
+ <!-- Mount on demand: a hidden TDesign drawer steals focus when each response mounts. -->
+ <t-drawer v-if="wikiDrawerVisible" v-model:visible="wikiDrawerVisible" :header="wikiDrawerPage?.title || ''" size="480px" :footer="false"
placement="right" attach="body" :show-overlay="true" :close-btn="true" :close-on-overlay-click="true"
class="wiki-graph-drawer">
<template v-if="wikiDrawerPage">
@@ -587,6 +588,7 @@
</template>
<script setup lang="ts">
+import { isAssistantTurnComplete } from '@/utils/steerStreamFork';
import { ref, computed, watch, onMounted, onBeforeUnmount, onUpdated, nextTick } from 'vue';
import { useRouter, useRoute } from 'vue-router';
import { marked } from 'marked';
@@ -1547,36 +1549,9 @@ watch(eventStream, (stream) => {
}, { deep: true, immediate: true });
-// Check if conversation is done (based on answer event with done=true or stop event)
-const isConversationDone = computed(() => {
- if (props.session?.steerForked) return true
- const stream = eventStream.value;
- if (!stream || stream.length === 0) {
- console.log('[Collapse] No stream or empty stream');
- return false;
- }
-
- // Check for stop event (user cancelled)
- const stopEvent = stream.find((e: any) => e.type === 'stop');
- if (stopEvent) {
- console.log('[Collapse] Found stop event, conversation done');
- return true;
- }
-
- const completeEvent = stream.find((e: any) => e.type === 'agent_complete');
- if (completeEvent) {
- console.log('[Collapse] Found complete event, conversation done');
- return true;
- }
-
- // Check for answer event with done=true. Exclude superseded preambles: a
- // retracted tool-round preamble is also closed with done=true, but the agent
- // keeps running, so it must not mark the whole conversation as finished.
- const answerEvents = stream.filter((e: any) => e.type === 'answer' && !e.superseded);
- const doneAnswer = answerEvents.find((e: any) => e.done === true);
-
- return !!doneAnswer;
-});
+// A steer boundary stops local animation without completing or folding the task.
+const isConversationDone = computed(() => isAssistantTurnComplete(props.session));
+const isSegmentDone = computed(() => Boolean(props.session?.steerForked) || isConversationDone.value);
const streamingMermaidSvgCache = ref<string[]>([]);
let streamingMermaidRenderTask: Promise<void> | null = null;
@@ -1603,7 +1578,7 @@ const activeAnswerEventRef = computed(() => {
// full instead of replaying.
const { displayed: typedAnswer } = useTypewriter(
() => activeAnswerMarkdown.value,
- () => isConversationDone.value,
+ () => isSegmentDone.value,
);
const cacheStreamingMermaidSvg = async () => {
@@ -1629,7 +1604,7 @@ const cacheStreamingMermaidSvg = async () => {
}
};
-watch(isConversationDone, (done) => {
+watch(isSegmentDone, (done) => {
if (!done) {
streamingMermaidSvgCache.value = [];
streamingMermaidRenderTask = null;
@@ -1656,7 +1631,7 @@ watch(activeAnswerMarkdown, () => {
const answerFullyRendered = computed(
() =>
!props.session?.steerForked &&
- isConversationDone.value &&
+ isSegmentDone.value &&
typedAnswer.value.length >= activeAnswerMarkdown.value.length,
);
watch(answerFullyRendered, (ready) => {
@@ -1696,7 +1671,7 @@ const hasPendingStreamingActivity = computed(() => {
// feedback-less timeline. Once a real pending step exists it carries its own
// shimmer, and once answer text starts the stream itself is enough feedback.
const showAgentActivityIndicator = computed(() => {
- if (isConversationDone.value) return false;
+ if (isSegmentDone.value) return false;
if (props.ragMode || hasAnswerStarted.value) return false;
return !hasPendingStreamingActivity.value;
});
@@ -1734,7 +1709,7 @@ const finalContent = computed(() => {
return null;
}
- if (!isConversationDone.value) {
+ if (!isSegmentDone.value) {
return null;
}
@@ -1779,7 +1754,7 @@ const finalContent = computed(() => {
// Count intermediate steps (after merging consecutive thinking events, matching what user sees in tree)
const intermediateStepsCount = computed(() => {
- if (!hasAnswerStarted.value && !isConversationDone.value) return 0;
+ if (!hasAnswerStarted.value && !isSegmentDone.value) return 0;
// Count only thinking and tool_call events (exclude plan_task_change, etc.)
return intermediateEvents.value.filter(
(e: any) => e.type === 'thinking' || e.type === 'tool_call'
@@ -1791,12 +1766,12 @@ const intermediateStepsCount = computed(() => {
// over-counts what the user perceives as agent loops (a single loop emits one
// thinking card plus its tool calls).
const reasoningRoundsCount = computed(() => {
- if (!hasAnswerStarted.value && !isConversationDone.value) return 0;
+ if (!hasAnswerStarted.value && !isSegmentDone.value) return 0;
return intermediateEvents.value.filter((e: any) => e.type === 'thinking').length;
});
const toolCallsCount = computed(() => {
- if (!hasAnswerStarted.value && !isConversationDone.value) return 0;
+ if (!hasAnswerStarted.value && !isSegmentDone.value) return 0;
return intermediateEvents.value.filter((e: any) => e.type === 'tool_call').length;
});
@@ -2545,7 +2520,7 @@ agentRenderer.image = function agentImageRenderer(token) {
artifacts: artifactList.value,
labels: artifactRefLabels.value,
context: artifactRefContext.value,
- streaming: !isConversationDone.value,
+ streaming: !isSegmentDone.value,
});
if (artifactHtml !== null) return artifactHtml;
return defaultImageRenderer.call(this, token);
@@ -2561,7 +2536,7 @@ const prepareAgentMarkdown = (markdown: string, cachedSvgHtml?: CachedMermaidSvg
const cache = cachedSvgHtml ?? streamingMermaidSvgCache.value;
// Keep masking after the turn ends when SVG is already cached so v-html /
// v-stable-html cannot replace a painted diagram with mermaid source.
- const mermaidSafe = !isConversationDone.value || hasCachedMermaidSvg(cache)
+ const mermaidSafe = !isSegmentDone.value || hasCachedMermaidSvg(cache)
? prepareStreamingMermaidMarkdown(markdown, cache)
: replaceIncompleteMermaidWithPlaceholder(markdown);
return mermaidSafe.replace(/<(?:kb|web)\b[^>]*$/i, '');
@@ -2578,7 +2553,7 @@ const renderAgentMarkdown = (
renderer: agentRenderer,
escapeMarkdown,
sanitizeHtml: sanitizeMarkdownHTML,
- streaming: !isConversationDone.value,
+ streaming: !isSegmentDone.value,
knowledgeReferences: getReferencesForDrawer(),
cachedMermaidSvgHtml: streamingMermaidSvgCache.value,
prepareMarkdown: prepareAgentMarkdown,
@@ -3157,6 +3132,11 @@ const handleAddToKnowledge = (answerEvent: any) => {
--stream-brand-15: color-mix(in srgb, var(--td-brand-color) 15%, transparent);
--stream-brand-20: color-mix(in srgb, var(--td-brand-color) 20%, transparent);
+ &.is-steer-prefix {
+ margin-bottom: 0;
+ .tree-container { margin-bottom: 0; }
+ }
+
&.is-rag-mode {
margin-top: 0;
}
diff --git a/frontend/src/views/chat/components/usermsg.vue b/frontend/src/views/chat/components/usermsg.vue
index c85c6e0e5..93801de80 100644
--- a/frontend/src/views/chat/components/usermsg.vue
+++ b/frontend/src/views/chat/components/usermsg.vue
@@ -43,6 +43,11 @@
<div class="user_msg">
{{ content }}
</div>
+ <div v-if="steerFailed" class="steer-failure" role="status">
+ <span>{{ t('input.messages.steerFailed') }}</span>
+ <t-tooltip :content="t('input.steerRetry')"><button type="button" :aria-label="t('input.steerRetry')" @click="emit('retry-steer')"><t-icon name="refresh" /></button></t-tooltip>
+ <t-tooltip :content="t('common.remove')"><button type="button" :aria-label="t('common.remove')" @click="emit('remove-steer')"><t-icon name="close" /></button></t-tooltip>
+ </div>
<picturePreview :reviewImg="reviewImg" :reviewUrl="reviewUrl" @closePreImg="closePreImg" />
</div>
</template>
@@ -54,6 +59,7 @@ import { useI18n } from 'vue-i18n';
import { useChatAttachmentPreviewDrawer } from '@/composables/useChatAttachmentPreviewDrawer';
import { isPreviewableAttachment, resolveAttachmentFileType } from '@/utils/attachmentPreview';
import { SKILL_ICON } from '@/types/mention';
+const emit = defineEmits(['retry-steer', 'remove-steer']);
const { t } = useI18n();
@@ -70,6 +76,7 @@ const mentionTagIcon = (item) => {
};
const props = defineProps({
+ steerFailed: { type: Boolean, default: false },
content: {
type: String,
required: false
@@ -357,3 +364,10 @@ html[theme-mode="dark"] {
}
}
</style>
+
+<style scoped>
+.steer-failure { display: flex; align-items: center; justify-content: flex-end; gap: 4px; font-size: 12px; color: var(--td-text-color-secondary); margin-bottom: 4px; }
+.steer-failure { margin-top: 6px; color: var(--td-error-color); }
+.steer-failure button { display: inline-flex; align-items: center; justify-content: center; width: 26px; height: 26px; border: 0; border-radius: 6px; background: transparent; color: inherit; cursor: pointer; }
+.steer-failure button:hover { background: var(--td-bg-color-secondarycontainer); }
+</style>
diff --git a/frontend/src/views/chat/index.vue b/frontend/src/views/chat/index.vue
index 19ae3bb5f..44b2c435b 100644
--- a/frontend/src/views/chat/index.vue
+++ b/frontend/src/views/chat/index.vue
@@ -98,7 +98,8 @@
仅对极少数尚未拿到 id 的本地占位消息 fallback 到 role+created_at+index。
-->
<div v-for="(session, index) in messagesList"
- :key="session.id || `${session.role}-${session.created_at}-${index}`" class="msg-item-wrapper">
+ :key="session.id || `${session.role}-${session.created_at}-${index}`" class="msg-item-wrapper"
+ :class="{ 'is-steer-prefix': session.steerForked, 'is-empty-segment': session.role === 'assistant' && !shouldRenderAssistantMessage(session) }">
<MessageTimestamp v-if="shouldShowConversationTimestamp(messagesList, index)"
:value="session.created_at" />
@@ -107,7 +108,10 @@
:class="{ 'is-minimap-target': session.id && session.id === minimapTargetId }">
<usermsg :content="session.content" :mentioned_items="session.mentioned_items"
:images="session.images" :attachments="session.attachments" :embeddedMode="embeddedMode"
- :session-id="session_id">
+ :session-id="session_id"
+ :steer-failed="Boolean(session._steerFailed)"
+ @retry-steer="handleRetrySteer(session.steer_id)"
+ @remove-steer="handleRemoveSteer(session.steer_id)">
</usermsg>
</div>
<div v-if="session.role == 'assistant' && shouldRenderAssistantMessage(session)"
@@ -143,16 +147,17 @@
</div>
</transition>
<div class="input-container" :class="{ 'is-embedded': embeddedMode }">
- <InputField ref="inputFieldRef"
+ <InputField ref="inputFieldRef" :auto-focus="focusComposerOnMount"
@send-msg="(query, modelId, mentionedItems, imageFiles, attachmentFiles) => sendMsg(query, modelId, mentionedItems, imageFiles, attachmentFiles)"
@steer-msg="(query, mentionedItems, delivery) => handleSteerMsg(query, mentionedItems, delivery)"
@promote-steer="handlePromoteSteer"
@remove-steer="handleRemoveSteer"
+ @retry-steer="handleRetrySteer"
@stop-generation="handleStopGeneration"
@stop-confirmed="handleStopConfirmed"
@stop-failed="handleStopFailed" :isReplying="isReplying" :sessionId="session_id"
:assistantMessageId="currentAssistantMessageId" :embeddedMode="embeddedMode"
- :queuedSteers="steerQueue" :canSteer="isAgentStreamSession()"></InputField>
+ :queuedSteers="steerQueue.filter(item => item.delivery === 'after')" :canSteer="isAgentStreamSession()"></InputField>
</div>
</div>
<KnowledgeBaseEditorModal :visible="uiStore.showKBEditorModal" :mode="uiStore.kbEditorMode"
@@ -178,7 +183,7 @@ import { getSuggestedQuestions } from "@/api/agent/index";
import { deleteTemporaryAttachment, uploadTemporaryAttachment } from '@/api/chat/temporary-attachments';
import { useStream } from '../../api/chat/streame'
import { listSteerSession, promoteSteerSession, removeSteerSession, steerSession } from '@/api/chat/steer';
-import { persistedAssistantId } from '@/utils/steerStreamFork';
+import { persistedAssistantId, previewSteerMessage, discardSteerPreview } from '@/utils/steerStreamFork';
import { useMenuStore } from '@/stores/menu';
import { useSettingsStore } from '@/stores/settings';
import { MessagePlugin } from 'tdesign-vue-next';
@@ -239,6 +244,8 @@ const uiStore = useUIStore();
const { navigateToKnowledgeBaseList } = useKnowledgeBaseCreationNavigation();
const { t } = useI18n();
const { firstQuery, firstMentionedItems, firstModelId, firstImageFiles, firstAttachmentFiles } = storeToRefs(usemenuStore);
+// Capture before the initial send consumes firstQuery; the child focuses after mounting.
+const focusComposerOnMount = Boolean(firstQuery.value);
const { onChunk, error, isStreaming, startStream, stopStream, lastStreamRequest } = useStream();
/** Snapshot of the in-flight HTTP request for attaching to the next assistant message. */
const pendingStreamDebug = ref(null);
@@ -639,7 +646,11 @@ const hydrateSteerQueue = async ({ onlyWhenLive = false } = {}) => {
content: item.content || '',
delivery: item.delivery === 'inject' ? 'inject' : 'after',
mentioned_items: item.mentioned_items || [],
- }));
+ expected_assistant_message_id: res.assistant_message_id,
+ })).concat(steerQueue.value.filter(item => item.failed && !items.some(remote => remote.steer_id === item.steer_id)));
+ for (const item of steerQueue.value) {
+ if (item.delivery === 'inject') previewSteerMessage(messagesList, item);
+ }
} catch (e) {
console.warn('[Steer] Failed to restore queue:', e);
}
@@ -738,6 +749,7 @@ const {
dropSteerQueueItem(steerId);
},
onGenerationStopped: () => {
+ for (const item of steerQueue.value) discardSteerPreview(messagesList, item.steer_id);
steerQueue.value = [];
},
onTurnComplete: (message) => {
@@ -803,6 +815,7 @@ const handleStopGeneration = () => {
};
const handleStopConfirmed = () => {
+ for (const item of steerQueue.value) discardSteerPreview(messagesList, item.steer_id);
steerQueue.value = [];
};
@@ -820,97 +833,135 @@ const dropSteerQueueItem = (steerId) => {
const findSteerQueueItem = (steerId) =>
steerQueue.value.find((item) => item.steer_id === steerId);
-const makeSteerClientId = () =>
- `steer-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
+const makeSteerClientId = () => {
+ // getRandomValues also works on HTTP deployments without randomUUID.
+ const bytes = crypto.getRandomValues(new Uint8Array(16));
+ bytes[6] = (bytes[6] & 15) | 64;
+ bytes[8] = (bytes[8] & 63) | 128;
+ const hex = Array.from(bytes, b => b.toString(16).padStart(2, '0')).join('');
+ return `${hex.slice(0,8)}-${hex.slice(8,12)}-${hex.slice(12,16)}-${hex.slice(16,20)}-${hex.slice(20)}`;
+};
-// 追加消息到正在运行的 agent turn。默认 delivery=after,排队显示在输入框上方;
-// 点「立即发送」再 promote 为 inject。注入后的时间线分叉复用 forkAfterInjectedUser。
-const handleSteerMsg = async (value, mentionedItems = [], delivery = 'after') => {
+// Enter queues a follow-up; an explicit inject appears in the transcript immediately.
+const handleSteerMsg = async (value, mentionedItems = [], delivery = 'after', retryId = '') => {
if (!session_id.value || !value?.trim()) return;
- if (!isReplying.value) {
+ if (!isReplying.value && !retryId) {
// 空闲时没有运行中的 turn 可排队:直接走正常发送,而不是把
// steering(服务端为 handleSteer/指定事务)当隐形 sendMsg 用。
await sendMsg(value, '', mentionedItems);
return;
}
- const clientId = makeSteerClientId();
- steerQueue.value.push({
+ const requestSessionId = session_id.value;
+ const clientId = retryId || makeSteerClientId();
+ const retryItem = retryId ? findSteerQueueItem(retryId) : null;
+ const expectedId = retryItem?.expected_assistant_message_id || currentAssistantMessageId.value;
+ if (retryItem) { retryItem.pending = true; retryItem.failed = false; }
+ else steerQueue.value.push({
steer_id: clientId,
client_id: clientId,
+ expected_assistant_message_id: expectedId,
content: value,
delivery,
mentioned_items: mentionedItems,
pending: true,
});
+ if (delivery === 'inject') {
+ const preview = previewSteerMessage(messagesList, findSteerQueueItem(clientId));
+ delete preview._steerFailed;
+ scrollToBottom(true);
+ }
try {
- const res = await steerSession(session_id.value, value, mentionedItems, delivery);
+ const res = await steerSession(requestSessionId, value, mentionedItems, delivery, expectedId, clientId);
+ if (session_id.value !== requestSessionId) return;
+ if (res?.status === 'already_injected') {
+ dropSteerQueueItem(clientId);
+ const preview = messagesList.find(m => m.steer_id === clientId);
+ if (preview) delete preview._steerPending;
+ MessagePlugin.info(t('input.messages.steerAlreadyInjected'));
+ return;
+ }
if (res?.status === 'new_run') {
const item = findSteerQueueItem(clientId);
// Still attached to a stream: aborting it to POST AgentQA races the
- // finishing turn and can start a second engine. Keep the chip and
+ // finishing turn and can start a second engine. Keep the message and
// send once the current SSE completes.
if (isReplying.value || isStreaming.value) {
if (item) {
item.pending = false;
item.awaitingIdleSend = true;
- item.delivery = 'after';
}
return;
}
dropSteerQueueItem(clientId);
+ discardSteerPreview(messagesList, clientId);
await sendMsg(value, '', mentionedItems);
return;
}
const item = findSteerQueueItem(clientId);
if (item) {
- // The server id is the durable one — promote/remove and the
- // injected-event correlation all go through it. client_id stays
- // so handlePromoteSteer can still find the row while the swap
- // is rendering.
+ // The server echoes the stable client UUID used by SSE receipts.
if (res?.steer_id) item.steer_id = res.steer_id;
item.pending = false;
}
} catch (e) {
console.error('[Steer] Failed to queue message:', e);
- dropSteerQueueItem(clientId);
+ if (session_id.value !== requestSessionId) return;
+ const item = findSteerQueueItem(clientId);
+ if (!item) return; // The delivery receipt may have already consumed it.
+ item.pending = false;
+ item.failed = true;
+ const preview = messagesList.find(m => m.steer_id === clientId && m._steerPending);
+ if (preview) preview._steerFailed = true;
+ if (e?.status === 409) item.expected_assistant_message_id = currentAssistantMessageId.value;
MessagePlugin.error(e?.message || t('input.messages.steerFailed'));
}
};
+const handleRetrySteer = async (steerId) => {
+ const item = findSteerQueueItem(steerId);
+ if (!item || item.pending) return;
+ await handleSteerMsg(item.content, item.mentioned_items || [], item.delivery, steerId);
+};
+
const handlePromoteSteer = async (steerId) => {
if (!session_id.value || !steerId) return;
const item = findSteerQueueItem(steerId) || steerQueue.value.find((entry) => entry.client_id === steerId);
if (!item || item.delivery === 'inject') return;
+ if (item.pending || item.promoting || item.failed) return;
+ const requestSessionId = session_id.value;
item.promoting = true;
+ item.delivery = 'inject';
+ previewSteerMessage(messagesList, item);
+ scrollToBottom(true);
try {
- // Promote needs the durable server id; a pending item only has the
- // client-side one. isPending doesn't disable the button yet, so guard
- // here to keep the queue handlers total.
- if (item.pending) {
- MessagePlugin.info(t('input.messages.steerPromoteFailed'));
- return;
- }
- const res = await promoteSteerSession(session_id.value, item.steer_id);
+ const res = await promoteSteerSession(requestSessionId, item.steer_id);
+ if (session_id.value !== requestSessionId) return;
if (res?.status === 'already_injected') {
dropSteerQueueItem(item.steer_id);
+ const preview = messagesList.find(m => m.steer_id === item.steer_id);
+ if (preview) delete preview._steerPending;
MessagePlugin.info(t('input.messages.steerAlreadyInjected'));
return;
}
if (res?.status === 'new_run') {
if (isReplying.value || isStreaming.value) {
item.awaitingIdleSend = true;
- item.delivery = 'after';
return;
}
const content = item.content;
const mentions = item.mentioned_items || [];
dropSteerQueueItem(steerId);
+ discardSteerPreview(messagesList, steerId);
await sendMsg(content, '', mentions);
return;
}
item.delivery = 'inject';
} catch (e) {
console.error('[Steer] Failed to promote queued message:', e);
+ if (session_id.value !== requestSessionId) return;
+ if (!findSteerQueueItem(steerId)) return;
+ item.delivery = 'after';
+ discardSteerPreview(messagesList, steerId);
MessagePlugin.error(e?.message || t('input.messages.steerPromoteFailed'));
} finally {
item.promoting = false;
@@ -932,14 +983,16 @@ const handleRemoveSteer = async (steerId) => {
return;
}
if (res?.status === 'gone') {
+ discardSteerPreview(messagesList, steerId);
dropSteerQueueItem(steerId);
return;
}
- if (res && res.removed === false) {
+ if (res && res.removed === false && !item.failed) {
MessagePlugin.error(t('input.messages.steerRemoveFailed'));
return;
}
}
+ discardSteerPreview(messagesList, steerId);
dropSteerQueueItem(steerId);
} catch (e) {
console.error('[Steer] Failed to remove queued message:', e);
@@ -955,6 +1008,7 @@ const flushSteerAfterTurn = async (completedAssistantId) => {
const awaiting = steerQueue.value.filter((item) => item.awaitingIdleSend);
if (awaiting.length) {
const batch = awaiting.slice();
+ for (const item of batch) discardSteerPreview(messagesList, item.steer_id);
steerQueue.value = steerQueue.value.filter((item) => !item.awaitingIdleSend);
const first = batch[0];
await sendMsg(first.content, '', first.mentioned_items || []);
@@ -967,7 +1021,7 @@ const flushSteerAfterTurn = async (completedAssistantId) => {
};
const attachSteerFollowUp = async (completedAssistantId) => {
- const queued = steerQueue.value.slice();
+ const queued = steerQueue.value.filter(item => !item.failed);
if (!queued.length || attachingSteerFollowUp || !session_id.value) return;
const sessionId = session_id.value;
attachingSteerFollowUp = true;
@@ -999,12 +1053,19 @@ const attachSteerFollowUp = async (completedAssistantId) => {
(q) => q.content === persisted.content && !claimed.has(q.steer_id)
);
if (queuedMatch) claimed.add(queuedMatch.steer_id);
- messagesList.push({
+ const userRow = {
...persisted,
mentioned_items: queuedMatch?.mentioned_items?.length
? queuedMatch.mentioned_items
: persisted.mentioned_items,
- });
+ };
+ const preview = queuedMatch && messagesList.find(m => m.steer_id === queuedMatch.steer_id && m._steerPending);
+ if (preview) {
+ delete preview._steerPending;
+ delete preview._steerFailed;
+ delete preview.isSteer;
+ Object.assign(preview, userRow);
+ } else messagesList.push(userRow);
}
// Rows that made it into the transcript are no longer queued.
for (const steerId of claimed) dropSteerQueueItem(steerId);
@@ -1038,7 +1099,7 @@ const attachSteerFollowUp = async (completedAssistantId) => {
loading.value = false;
isReplying.value = false;
MessagePlugin.error(t('input.messages.steerFollowUpTimeout'));
- } else if (steerQueue.value.length) {
+ } else if (steerQueue.value.some(item => !item.failed)) {
// startStream awaits the whole SSE. The follow-up's onTurnComplete
// therefore runs while attachingSteerFollowUp is still true and
// no-ops. Chain remaining after-items once that guard drops.
@@ -1521,7 +1582,7 @@ onBeforeRouteUpdate((to, from, next) => {
&.is-embedded :deep(.answers-input) .t-textarea__inner {
width: 100% !important;
min-height: 48px !important;
- padding: 10px 14px 48px 14px;
+ padding: 10px 14px;
}
}
@@ -1721,6 +1782,8 @@ onBeforeRouteUpdate((to, from, next) => {
*/
.msg-item-wrapper {
contain: layout style;
+ &.is-empty-segment { display: none; }
+ &.is-steer-prefix { margin-bottom: -4px; }
}
.message-row {
diff --git a/frontend/src/views/chat/steerFollowUp.test.mjs b/frontend/src/views/chat/steerFollowUp.test.mjs
index 30598db75..5a932ef8d 100644
--- a/frontend/src/views/chat/steerFollowUp.test.mjs
+++ b/frontend/src/views/chat/steerFollowUp.test.mjs
@@ -84,7 +84,7 @@ test('follow-up attaches persisted users by request_id', () => {
// an explicit new_run (no live turn) starts a normal AgentQA.
test('steer lookup failure does not start a second turn', () => {
const start = source.indexOf('const handleSteerMsg = async')
- const end = source.indexOf('const handlePromoteSteer = async', start)
+ const end = source.indexOf('const handleRetrySteer = async', start)
assert.notEqual(start, -1)
assert.notEqual(end, -1)
const fn = source.slice(start, end)
@@ -102,7 +102,7 @@ test('steer lookup failure does not start a second turn', () => {
test('new_run while a stream is live does not abort it', () => {
const start = source.indexOf('const handleSteerMsg = async')
- const end = source.indexOf('const handlePromoteSteer = async', start)
+ const end = source.indexOf('const handleRetrySteer = async', start)
const fn = source.slice(start, end)
const newRun = fn.indexOf("if (res?.status === 'new_run')")
const catchIdx = fn.indexOf('} catch (e) {')
@@ -130,7 +130,7 @@ test('remaining overlay after-items chain after a follow-up stream ends', () =>
// startStream awaits the whole SSE. The follow-up's onTurnComplete therefore
// runs while attachingSteerFollowUp is still true and would no-op. Remaining
// after-items have to be retried once that guard drops.
- assert.match(finallyBlock, /steerQueue\.value\.length/)
+ assert.match(finallyBlock, /steerQueue\.value\.some\(item => !item\.failed\)/)
assert.match(finallyBlock, /void attachSteerFollowUp\(/)
})
diff --git a/frontend/src/views/chat/steerInteraction.test.mjs b/frontend/src/views/chat/steerInteraction.test.mjs
new file mode 100644
index 000000000..fedea335a
--- /dev/null
+++ b/frontend/src/views/chat/steerInteraction.test.mjs
@@ -0,0 +1,93 @@
+import assert from 'node:assert/strict'
+import { readFileSync } from 'node:fs'
+import { webcrypto } from 'node:crypto'
+import vm from 'node:vm'
+import test from 'node:test'
+import { previewSteerMessage, discardSteerPreview, forkAfterInjectedUser } from '../../utils/steerStreamFork.ts'
+
+const source = readFileSync(new URL('./index.vue', import.meta.url), 'utf8')
+const handlers = source.slice(source.indexOf('const dropSteerQueueItem ='), source.indexOf('let attachingSteerFollowUp ='))
+const deferred = () => {
+ let resolve, reject
+ const promise = new Promise((yes, no) => { resolve = yes; reject = no })
+ return { promise, resolve, reject }
+}
+function harness(overrides = {}) {
+ const state = {
+ session_id: { value: 'session' }, currentAssistantMessageId: { value: 'assistant' },
+ isReplying: { value: true }, isStreaming: { value: true },
+ steerQueue: { value: [] }, messagesList: [], crypto: webcrypto,
+ previewSteerMessage, discardSteerPreview,
+ scrollToBottom() {}, sendMsg() { throw new Error('must not start a second run') },
+ MessagePlugin: { info() {}, error() {} }, t: key => key,
+ console: { error() {} }, ...overrides,
+ }
+ const actions = vm.runInNewContext(`${handlers}\n({ handleSteerMsg, handlePromoteSteer, handleRetrySteer, handleRemoveSteer })`, state)
+ return { ...actions, state }
+}
+
+test('default after stays below; promotion immediately moves it into the transcript', async () => {
+ const request = deferred(), promote = deferred()
+ const h = harness({ steerSession: () => request.promise, promoteSteerSession: () => promote.promise })
+ const sending = h.handleSteerMsg('queued')
+ const item = h.state.steerQueue.value[0]
+ assert.equal(item.delivery, 'after')
+ assert.equal(h.state.messagesList.length, 0)
+ request.resolve({ status: 'queued', steer_id: item.steer_id })
+ await sending
+ const promoting = h.handlePromoteSteer(item.steer_id)
+ assert.equal(item.delivery, 'inject')
+ assert.equal(h.state.steerQueue.value.filter(q => q.delivery === 'after').length, 0)
+ assert.equal(h.state.messagesList[0].content, 'queued')
+ promote.resolve({ status: 'queued' })
+ await promoting
+})
+
+test('direct inject shows immediately and an early SSE receipt does not duplicate it', async () => {
+ const request = deferred()
+ const h = harness({ steerSession: () => request.promise })
+ const assistant = { id: 'assistant', role: 'assistant', request_id: 'request', is_completed: false }
+ h.state.messagesList.push(assistant)
+ const sending = h.handleSteerMsg('补充', [], 'inject')
+ const item = h.state.steerQueue.value[0]
+ assert.equal(h.state.messagesList[1].content, '补充')
+ forkAfterInjectedUser(h.state.messagesList, assistant, h.state.messagesList[1], item.steer_id)
+ h.state.steerQueue.value.splice(0, 1) // onUserMessageInjected receipt
+ request.resolve({ status: 'queued', steer_id: item.steer_id })
+ await sending
+ assert.equal(h.state.messagesList.filter(m => m.role === 'user').length, 1)
+ assert.equal(h.state.steerQueue.value.length, 0)
+})
+
+test('failed inject retries with the same ID and bubble', async () => {
+ const ids = []
+ const h = harness({ steerSession: async (...args) => {
+ ids.push(args[5])
+ if (ids.length === 1) throw new Error('network')
+ return { status: 'queued', steer_id: args[5] }
+ } })
+ await h.handleSteerMsg('补充', [], 'inject')
+ assert.equal(h.state.messagesList[0]._steerFailed, true)
+ await h.handleRetrySteer(ids[0])
+ assert.deepEqual(ids, [ids[0], ids[0]])
+ assert.equal(h.state.messagesList.length, 1)
+ assert.equal(h.state.messagesList[0]._steerFailed, undefined)
+})
+
+test('a lost HTTP response after a delivery receipt does not report a failed send', async () => {
+ const request = deferred(), errors = []
+ const h = harness({ steerSession: () => request.promise, MessagePlugin: { info() {}, error: message => errors.push(message) } })
+ const sending = h.handleSteerMsg('补充', [], 'inject')
+ h.state.steerQueue.value.splice(0, 1) // receipt already removed the item
+ request.reject(new Error('response lost'))
+ await sending
+ assert.equal(errors.length, 0)
+})
+
+test('failed promotion restores the after queue and removes only its optimistic row', async () => {
+ const h = harness({ promoteSteerSession: async () => { throw new Error('network') } })
+ h.state.steerQueue.value.push({ steer_id: 'queued', content: '补充', delivery: 'after' })
+ await h.handlePromoteSteer('queued')
+ assert.equal(h.state.steerQueue.value[0].delivery, 'after')
+ assert.equal(h.state.messagesList.length, 0)
+})
diff --git a/internal/agent/engine.go b/internal/agent/engine.go
index 3a3e87e4a..4ab34b81f 100644
--- a/internal/agent/engine.go
+++ b/internal/agent/engine.go
@@ -731,12 +731,14 @@ func (e *AgentEngine) runReActIteration(
// Create agent step
step := types.AgentStep{
- Iteration: state.CurrentRound,
- Thought: response.Content,
- ReasoningContent: response.ReasoningContent,
- ToolCalls: make([]types.ToolCall, 0),
- Timestamp: time.Now(),
+ UserMessagesBefore: state.PendingSteerMessages,
+ Iteration: state.CurrentRound,
+ Thought: response.Content,
+ ReasoningContent: response.ReasoningContent,
+ ToolCalls: make([]types.ToolCall, 0),
+ Timestamp: time.Now(),
}
+ state.PendingSteerMessages = nil
// If the request was cancelled while the LLM was streaming (e.g. the
// user pressed "stop"), the stream driver still returns a usable
@@ -751,7 +753,7 @@ func (e *AgentEngine) runReActIteration(
if ctx.Err() != nil {
logger.Warnf(ctx, "[Agent][Round-%d] Context cancelled during LLM call; preserving partial step",
round)
- if step.Thought != "" || len(step.ToolCalls) > 0 {
+ if step.Thought != "" || len(step.ToolCalls) > 0 || len(step.UserMessagesBefore) > 0 {
state.RoundSteps = append(state.RoundSteps, step)
}
return iterOutcomeBreak, nil
@@ -766,6 +768,7 @@ func (e *AgentEngine) runReActIteration(
if verdict.emptyContent {
*emptyRetries++
if *emptyRetries <= maxEmptyResponseRetries {
+ state.PendingSteerMessages = step.UserMessagesBefore
logger.Warnf(ctx, "[Agent][Round-%d] Empty content with stop - retrying (%d/%d)",
round, *emptyRetries, maxEmptyResponseRetries)
*messagesPtr = append(*messagesPtr, chat.Message{
@@ -799,6 +802,7 @@ func (e *AgentEngine) runReActIteration(
})
injected := e.drainSteerMessages(ctx, state, messagesPtr, sessionID, assistantMessageID)
if injected > 0 {
+ verdict.step.IntermediateAnswer = true
state.RoundSteps = append(state.RoundSteps, verdict.step)
if !e.withinIterationBudget(nextRound) {
e.steerOverruns++
diff --git a/internal/agent/finalize.go b/internal/agent/finalize.go
index cab3aa51b..ca2028ac4 100644
--- a/internal/agent/finalize.go
+++ b/internal/agent/finalize.go
@@ -192,6 +192,14 @@ func (e *AgentEngine) handleMaxIterations(
func (e *AgentEngine) emitCompletionEvent(
ctx context.Context, state *types.AgentState, sessionID, messageID string, startTime time.Time,
) {
+ steps := state.RoundSteps
+ if len(state.PendingSteerMessages) > 0 {
+ // A stop or model failure can arrive after delivery but before the next
+ // response exists. Preserve that boundary without inventing an answer.
+ steps = append(append([]types.AgentStep(nil), steps...), types.AgentStep{
+ Iteration: state.CurrentRound, UserMessagesBefore: state.PendingSteerMessages,
+ })
+ }
// Convert knowledge refs to interface{} slice for event data
knowledgeRefsInterface := make([]interface{}, 0, len(state.KnowledgeRefs))
for _, ref := range state.KnowledgeRefs {
@@ -205,7 +213,7 @@ func (e *AgentEngine) emitCompletionEvent(
Data: event.AgentCompleteData{
FinalAnswer: state.FinalAnswer,
KnowledgeRefs: knowledgeRefsInterface,
- AgentSteps: state.RoundSteps, // Include detailed execution steps for message storage
+ AgentSteps: steps,
Usage: turnUsage(state),
TotalSteps: len(state.RoundSteps),
TotalDurationMs: time.Since(startTime).Milliseconds(),
diff --git a/internal/agent/prompts.go b/internal/agent/prompts.go
index 16727c4a5..b46ad59d4 100644
--- a/internal/agent/prompts.go
+++ b/internal/agent/prompts.go
@@ -390,6 +390,7 @@ func BuildSystemPromptWithOptions(
language = options.Language
}
basePrompt = renderPromptPlaceholdersWithStatus(template, knowledgeBases, webSearchEnabled, currentTime, language)
+ basePrompt += "\n\n" + steerGuidance
if options != nil {
basePrompt += formatGroundingGuidance(options.SelectedTools)
@@ -406,6 +407,14 @@ func BuildSystemPromptWithOptions(
return basePrompt
}
+// Apply to custom prompts too: mid-run delivery is a harness capability.
+const steerGuidance = "<steering_guidance>\n" +
+ "Messages in <steer_message> guide the task in progress. Apply them in context; " +
+ "respond briefly when appropriate, then continue unfinished work. Preserve unfinished objectives, " +
+ "accepted constraints and useful tool results unless explicitly changed. " +
+ "Acknowledging guidance alone does not complete the task. Follow explicit cancellation or replacement requests. " +
+ "Hide delivery tags. Untagged subsequent requests are ordinary user messages.\n</steering_guidance>"
+
// GetPureAgentSystemPrompt returns the Pure Agent system prompt from config templates.
// The template must be defined in config/prompt_templates/agent_system_prompt.yaml
// with mode "pure". Returns empty string if config is nil or template not found.
diff --git a/internal/agent/steer.go b/internal/agent/steer.go
index 6a4f26dde..42e907984 100644
--- a/internal/agent/steer.go
+++ b/internal/agent/steer.go
@@ -66,7 +66,8 @@ func (e *AgentEngine) drainSteerMessages(
logger.Warnf(ctx, "[Agent] Steer persist failed for %s, leaving event pending", steerID)
continue
}
- *messagesPtr = append(*messagesPtr, chat.Message{Role: "user", Content: content})
+ *messagesPtr = append(*messagesPtr, chat.Message{Role: "user", Content: types.SteerMessageContent(content)})
+ state.PendingSteerMessages = append(state.PendingSteerMessages, userMessageID)
_ = e.eventBus.Emit(ctx, event.Event{
ID: generateEventID("injected"),
Type: event.EventUserMessageInjected,
diff --git a/internal/agent/steer_test.go b/internal/agent/steer_test.go
index f8c02100a..cf16b72c9 100644
--- a/internal/agent/steer_test.go
+++ b/internal/agent/steer_test.go
@@ -98,9 +98,10 @@ func TestDrainSteerMessagesInjectsIntoTail(t *testing.T) {
require.Len(t, messages, 5)
assert.Equal(t, "user", messages[4].Role)
- assert.Equal(t, "再补充一点:也对比一下成本", messages[4].Content)
+ assert.Equal(t, types.SteerMessageContent("再补充一点:也对比一下成本"), messages[4].Content)
// Tool result pairing is untouched — the tool message stays where it was.
assert.Equal(t, "tool", messages[3].Role)
+ assert.Equal(t, []string{"user-row-for-再补充一点:也对比一下成本"}, state.PendingSteerMessages)
require.Len(t, sink.persisted, 1)
assert.Equal(t, "再补充一点:也对比一下成本", sink.persisted[0])
@@ -136,8 +137,8 @@ func TestDrainSteerMessagesEmptyAndOffset(t *testing.T) {
}
engine.drainSteerMessages(context.Background(), state, &messages, "sess", "msg")
assert.Len(t, messages, 3)
- assert.Equal(t, "first", messages[1].Content)
- assert.Equal(t, "second", messages[2].Content)
+ assert.Equal(t, types.SteerMessageContent("first"), messages[1].Content)
+ assert.Equal(t, types.SteerMessageContent("second"), messages[2].Content)
require.Len(t, sink.persisted, 2)
engine.drainSteerMessages(context.Background(), state, &messages, "sess", "msg")
@@ -178,7 +179,7 @@ func TestDrainSteerMessagesBlankContentSkipped(t *testing.T) {
engine.drainSteerMessages(context.Background(), state, &messages, "sess", "msg")
require.Len(t, messages, 2)
- assert.Equal(t, "real instruction", messages[1].Content)
+ assert.Equal(t, types.SteerMessageContent("real instruction"), messages[1].Content)
require.Len(t, sink.persisted, 1)
}
@@ -206,14 +207,15 @@ func TestExecuteLoopInjectsBeforeNextLLMCall(t *testing.T) {
engine := newTestEngine(t, model)
engine.eventBus = event.NewEventBus()
- sink := &fakeSteerSink{}
- engine.SetSteerSink(sink)
-
- // Simulate the user queueing a message while round 1 runs: it is in the
- // queue before the loop's second drain point executes.
- sink.queued = []map[string]interface{}{
- steerEntry("s1", "focus on the cost angle"),
+ // Deliver only after round 1 so this exercises continuation with existing
+ // tool results, rather than adding guidance before any work has started.
+ sink := &delayedSteerSink{
+ fakeSteerSink: fakeSteerSink{
+ queued: []map[string]interface{}{steerEntry("s1", "focus on the cost angle")},
+ },
+ hideUntil: 2,
}
+ engine.SetSteerSink(sink)
engine.toolRegistry = agentRegistryForTest(t, "counting_tool")
@@ -230,20 +232,28 @@ func TestExecuteLoopInjectsBeforeNextLLMCall(t *testing.T) {
require.NoError(t, err)
require.Len(t, model.calls, 2)
- // Round 2's call must contain the injected steer text. SanitizeMessages
- // merges consecutive same-role messages, so the injected user message
- // may be concatenated onto the original "start" user turn — that is the
- // desired end state (the model reads it as the user adding instructions
- // mid-task).
+ for _, msg := range model.calls[0] {
+ assert.NotContains(t, msg.Content, "focus on the cost angle")
+ }
+ // Continue with the original task and tool result exactly once; steering
+ // needs neither a restarted run nor an extra model call to interpret it.
second := model.calls[1]
- found := false
+ steerCount, taskCount, toolCount := 0, 0, 0
for _, msg := range second {
+ if msg.Role == "user" && msg.Content == "start" {
+ taskCount++
+ }
+ if msg.Role == "tool" && msg.ToolCallID == "tc1" {
+ toolCount++
+ }
if msg.Role == "user" && strings.Contains(msg.Content, "focus on the cost angle") {
- found = true
- break
+ steerCount++
+ assert.Contains(t, msg.Content, "<continue_task>")
}
}
- assert.True(t, found, "injected steer text missing from round 2 LLM call: %+v", second)
+ assert.Equal(t, 1, taskCount)
+ assert.Equal(t, 1, toolCount)
+ assert.Equal(t, 1, steerCount)
require.Len(t, sink.persisted, 1)
}
@@ -486,3 +496,10 @@ func agentRegistryForTest(t *testing.T, name string) *agenttools.ToolRegistry {
registry.RegisterTool(tool)
return registry
}
+
+func TestSteerGuidanceAppliesToCustomAgentPrompt(t *testing.T) {
+ prompt := BuildSystemPromptWithOptions(nil, false, nil, "Custom agent instructions.")
+ assert.Contains(t, prompt, "Custom agent instructions.")
+ assert.Contains(t, prompt, steerGuidance)
+ assert.Contains(t, prompt, "Preserve unfinished objectives")
+}
diff --git a/internal/agent/tools/mcp_catalog.go b/internal/agent/tools/mcp_catalog.go
index 6d57b97e7..ed23816cb 100644
--- a/internal/agent/tools/mcp_catalog.go
+++ b/internal/agent/tools/mcp_catalog.go
@@ -18,10 +18,11 @@ import (
const mcpDiscoveryDescription = "" +
"Discover authorized MCP tools without loading every schema. If a server_id is " +
- "already listed in this tool's source summaries, call list_tools or describe " +
+ "already listed in this tool's source summaries, call list_tools or search " +
"directly; do not call list_servers first. Use list_servers only when this " +
- "description says further services are available, or to paginate. After you " +
- "have a server_id, list_tools (or search), then describe an exact tool. " +
+ "description says further services are available, or to paginate. Describe tools, " +
+ "not servers; use describe directly only with an exact tool name already returned " +
+ "by this directory. Never infer tool names from server summaries. " +
"Server IDs and tool names must come from this directory. Only describe returns " +
"a callable tool_ref. Call call_mcp_tool with that tool_ref and arguments " +
"matching input_schema. Wait for each discovery result before issuing dependent calls. " +
diff --git a/internal/application/service/agent_history.go b/internal/application/service/agent_history.go
index c24affcb4..5b93eb718 100644
--- a/internal/application/service/agent_history.go
+++ b/internal/application/service/agent_history.go
@@ -137,22 +137,39 @@ func buildTurnBodyMessages(assistant *types.Message, midRunUsers []*types.Messag
}
out := make([]chat.Message, 0, len(assistant.AgentSteps)*2+len(midRunUsers)+1)
+ usersByID := make(map[string]*types.Message, len(midRunUsers))
+ for _, user := range midRunUsers {
+ usersByID[user.ID] = user
+ }
+ hasBoundaries := false
+ for _, step := range assistant.AgentSteps {
+ hasBoundaries = hasBoundaries || len(step.UserMessagesBefore) > 0
+ }
+ appendUser := func(user *types.Message) {
+ msg := buildUserHistoryMessage(user)
+ msg.Content = types.SteerMessageContent(msg.Content)
+ out = append(out, msg)
+ delete(usersByID, user.ID)
+ }
next := 0
for _, step := range assistant.AgentSteps {
- stepMsgs := buildAgentStepMessages(step)
- if len(stepMsgs) == 0 {
- continue
+ for _, id := range step.UserMessagesBefore {
+ if user := usersByID[id]; user != nil {
+ appendUser(user)
+ }
}
- for next < len(midRunUsers) &&
+ for !hasBoundaries && next < len(midRunUsers) &&
!step.Timestamp.IsZero() &&
midRunUsers[next].CreatedAt.Before(step.Timestamp) {
- out = append(out, buildUserHistoryMessage(midRunUsers[next]))
+ appendUser(midRunUsers[next])
next++
}
- out = append(out, stepMsgs...)
+ out = append(out, buildAgentStepMessages(step)...)
}
- for ; next < len(midRunUsers); next++ {
- out = append(out, buildUserHistoryMessage(midRunUsers[next]))
+ for _, user := range midRunUsers {
+ if usersByID[user.ID] != nil {
+ appendUser(user)
+ }
}
if final := finalAnswerHistoryMessage(assistant); final != nil {
@@ -203,6 +220,9 @@ func buildAssistantHistoryMessages(m *types.Message) []chat.Message {
func buildAgentStepMessages(step types.AgentStep) []chat.Message {
nonTerminalCalls := filterNonTerminalToolCalls(step.ToolCalls)
if len(nonTerminalCalls) == 0 {
+ if step.IntermediateAnswer && strings.TrimSpace(step.Thought) != "" {
+ return []chat.Message{{Role: "assistant", Content: step.Thought, ReasoningContent: step.ReasoningContent}}
+ }
return nil
}
assistantMsg := chat.Message{
diff --git a/internal/application/service/agent_history_test.go b/internal/application/service/agent_history_test.go
index 5f45af07b..0e5c463d5 100644
--- a/internal/application/service/agent_history_test.go
+++ b/internal/application/service/agent_history_test.go
@@ -356,7 +356,7 @@ func TestBuildTurnBodyMessages_KeepsMidRunUsersInPlace(t *testing.T) {
// The injected message belongs between the round it interrupted and the
// round it caused, not bolted onto either end.
- assert.Equal(t, chat.Message{Role: "user", Content: "also check B"}, got[2])
+ assert.Equal(t, chat.Message{Role: "user", Content: types.SteerMessageContent("also check B")}, got[2])
assert.Equal(t, "assistant", got[3].Role)
assert.Equal(t, "tool", got[4].Role)
@@ -390,7 +390,7 @@ func TestBuildTurnBodyMessages_LateMidRunUserPrecedesAnswer(t *testing.T) {
got := buildTurnBodyMessages(assistant, midRun)
require.Len(t, got, 4)
- assert.Equal(t, chat.Message{Role: "user", Content: "and keep it short"}, got[2])
+ assert.Equal(t, chat.Message{Role: "user", Content: types.SteerMessageContent("and keep it short")}, got[2])
assert.Equal(t, chat.Message{Role: "assistant", Content: "Here you go."}, got[3])
}
@@ -474,3 +474,27 @@ func TestMCPProxyHistoryRetainsProtocolCallAndTarget(t *testing.T) {
require.Contains(t, history[0].ToolCalls[0].Function.Arguments, "tool_ref")
require.Equal(t, "proxy-id", history[1].ToolCallID)
}
+
+func TestBuildTurnBodyMessages_ReplaysExplicitBoundariesAndIntermediateAnswers(t *testing.T) {
+ // Identical (and deliberately misleading) timestamps must not move updates
+ // ahead of work the model already did. IDs, not text, distinguish repeats.
+ now := time.Now()
+ users := []*types.Message{
+ {ID: "u2", Content: "revise", CreatedAt: now},
+ {ID: "u1", Content: "revise", CreatedAt: now},
+ }
+ assistant := &types.Message{Content: "final", AgentSteps: types.AgentSteps{
+ {Thought: "first draft", IntermediateAnswer: true, Timestamp: now.Add(time.Second)},
+ {Thought: "second draft", IntermediateAnswer: true, UserMessagesBefore: []string{"u1"}},
+ {Thought: "final", UserMessagesBefore: []string{"u2"}},
+ }}
+ got := buildTurnBodyMessages(assistant, users)
+ require.Len(t, got, 5)
+ assert.Equal(t, []chat.Message{
+ {Role: "assistant", Content: "first draft"},
+ {Role: "user", Content: types.SteerMessageContent("revise")},
+ {Role: "assistant", Content: "second draft"},
+ {Role: "user", Content: types.SteerMessageContent("revise")},
+ {Role: "assistant", Content: "final"},
+ }, got)
+}
diff --git a/internal/handler/session/steer.go b/internal/handler/session/steer.go
index bef55be53..19a36ad80 100644
--- a/internal/handler/session/steer.go
+++ b/internal/handler/session/steer.go
@@ -39,9 +39,13 @@ const steerDataUserMessageID = "user_message_id"
// SteerMessageRequest is the payload of POST /sessions/:session_id/steer.
type SteerMessageRequest struct {
- Query string `json:"query" binding:"required"`
- MentionedItems []MentionedItemRequest `json:"mentioned_items,omitempty"`
- Channel string `json:"channel,omitempty"`
+ // Optional for older clients. New clients pin delivery to the run they see
+ // and supply a stable ID so a consume event may precede the HTTP response.
+ ExpectedAssistantMessageID string `json:"expected_assistant_message_id,omitempty"`
+ SteerID string `json:"steer_id,omitempty"`
+ Query string `json:"query" binding:"required"`
+ MentionedItems []MentionedItemRequest `json:"mentioned_items,omitempty"`
+ Channel string `json:"channel,omitempty"`
// Delivery is "after" (default) or "inject". See the constants above.
Delivery string `json:"delivery,omitempty"`
}
@@ -495,6 +499,27 @@ func (h *Handler) SteerMessage(c *gin.Context) {
if !ok {
return
}
+ if req.SteerID != "" && req.ExpectedAssistantMessageID != "" {
+ previous, _, err := h.streamManager.GetSteerEvents(ctx, sessionID, req.ExpectedAssistantMessageID, 0)
+ if err != nil {
+ _ = c.Error(errors.NewServiceUnavailableError("Failed to look up previous delivery"))
+ return
+ }
+ for _, delivered := range previous {
+ if delivered.ID == req.SteerID && steerEventConsumed(delivered) {
+ if delivered.Content != query {
+ _ = c.Error(errors.NewConflictError("steer_id already belongs to another message"))
+ return
+ }
+ c.JSON(200, gin.H{"success": true, "status": "already_injected", "steer_id": req.SteerID})
+ return
+ }
+ }
+ }
+ if req.ExpectedAssistantMessageID != "" && assistantID != "" && req.ExpectedAssistantMessageID != assistantID {
+ _ = c.Error(errors.NewConflictError("The active turn changed; retry the message"))
+ return
+ }
if assistantID == "" {
// No run is live: the message starts a brand-new run via the normal
// AgentQA path. The request returns immediately and the SSE stream
@@ -513,13 +538,33 @@ func (h *Handler) SteerMessage(c *gin.Context) {
_ = c.Error(errors.NewInternalServerError("Failed to check steer queue"))
return
}
+
+ steerID := req.SteerID
+ if steerID == "" {
+ steerID = uuid.New().String()
+ } else if _, err := uuid.Parse(steerID); err != nil {
+ _ = c.Error(errors.NewBadRequestError("invalid steer_id"))
+ return
+ }
+ for _, existingEvent := range existing {
+ if existingEvent.ID == steerID {
+ if existingEvent.Content != query {
+ _ = c.Error(errors.NewConflictError("steer_id already belongs to another message"))
+ return
+ }
+ c.JSON(200, gin.H{
+ "success": true, "status": "queued", "steer_id": steerID,
+ "assistant_message_id": assistantID, "delivery": steerDeliveryOfEvent(existingEvent),
+ })
+ return
+ }
+ }
pending := len(selectSteerBacklog(existing, nil))
if pending >= maxSteerQueueDepth {
_ = c.Error(errors.NewBadRequestError("too many queued messages for the running turn"))
return
}
- steerID := uuid.New().String()
evt := steerEvent(steerID, query, convertMentionedItems(req.MentionedItems), req.Channel)
evt.Data["delivery"] = delivery
if err := h.streamManager.AppendSteerEvents(ctx, sessionID, assistantID,
diff --git a/internal/handler/session/steer_live_run_test.go b/internal/handler/session/steer_live_run_test.go
index fecd08d4b..67971201e 100644
--- a/internal/handler/session/steer_live_run_test.go
+++ b/internal/handler/session/steer_live_run_test.go
@@ -460,3 +460,31 @@ func TestRebindSteerKeepsEventWhenAppendFails(t *testing.T) {
require.Len(t, old, 1)
assert.Equal(t, "late-1", old[0].ID)
}
+
+func TestSteerMessageRejectsDifferentExpectedRun(t *testing.T) {
+ h := &Handler{
+ sessionService: &steerOwnedSessionStub{},
+ messageService: &steerMessageLookupStub{msg: &types.Message{ID: "new-run", IsCompleted: false}},
+ streamManager: &steerLiveRunLookupStub{assistantID: "new-run"},
+ }
+ w := postSteer(t, newSteerLiveRunRouter(h),
+ `{"query":"keep going","delivery":"inject","expected_assistant_message_id":"old-run"}`)
+ assert.Equal(t, http.StatusConflict, w.Code)
+}
+
+func TestSteerRetryFindsConsumedReceiptAfterRunCompleted(t *testing.T) {
+ mgr := stream.NewMemoryStreamManager()
+ ctx := context.Background()
+ id := "7c2f7062-c80a-480e-ac21-c509b6210936"
+ evt := steerEventWithDelivery(id, "more", steerDeliveryInject)
+ evt.Data[steerDataConsumed] = true
+ require.NoError(t, mgr.AppendSteerEvents(ctx, "sess-1", "old-run", []interfaces.StreamEvent{evt}))
+ h := &Handler{
+ sessionService: &steerOwnedSessionStub{}, messageService: &steerMessageLookupStub{}, streamManager: mgr,
+ }
+ w := postSteer(t, newSteerLiveRunRouter(h),
+ `{"query":"more","delivery":"inject","expected_assistant_message_id":"old-run","steer_id":"`+id+`"}`)
+ assert.Equal(t, http.StatusOK, w.Code)
+ assert.Contains(t, w.Body.String(), "already_injected")
+ assert.NotContains(t, w.Body.String(), "new_run")
+}
diff --git a/internal/stream/memory_manager.go b/internal/stream/memory_manager.go
index 03aa48367..5c79cab70 100644
--- a/internal/stream/memory_manager.go
+++ b/internal/stream/memory_manager.go
@@ -137,7 +137,15 @@ func (m *MemoryStreamManager) AppendSteerEvents(
stream.mu.Lock()
defer stream.mu.Unlock()
+ seen := make(map[string]bool, len(stream.steerEvents))
+ for _, event := range stream.steerEvents {
+ seen[event.ID] = true
+ }
for i := range events {
+ if events[i].ID != "" && seen[events[i].ID] {
+ continue
+ }
+ seen[events[i].ID] = true
if events[i].Timestamp.IsZero() {
events[i].Timestamp = time.Now()
}
diff --git a/internal/stream/redis_live_run_test.go b/internal/stream/redis_live_run_test.go
index 8545a4316..21ade7d19 100644
--- a/internal/stream/redis_live_run_test.go
+++ b/internal/stream/redis_live_run_test.go
@@ -150,3 +150,35 @@ func TestMemorySetLiveRunRejectsADifferentAssistant(t *testing.T) {
require.NoError(t, err)
require.Equal(t, "assist-2", id)
}
+
+func TestAppendSteerEventsDeduplicatesClientIDs(t *testing.T) {
+ redisManager, _ := newTestRedisStreamManager(t, time.Hour)
+ for name, manager := range map[string]interfaces.StreamManager{
+ "memory": NewMemoryStreamManager(), "redis": redisManager,
+ } {
+ t.Run(name, func(t *testing.T) {
+ ctx := context.Background()
+ evt := interfaces.StreamEvent{
+ ID: "same-client-id", Content: "do this next", Data: map[string]interface{}{"delivery": "after"},
+ }
+ errs := make(chan error, 8)
+ for i := 0; i < 8; i++ {
+ go func() { errs <- manager.AppendSteerEvents(ctx, "session", "run", []interfaces.StreamEvent{evt}) }()
+ }
+ for i := 0; i < 8; i++ {
+ require.NoError(t, <-errs)
+ }
+ events, _, err := manager.GetSteerEvents(ctx, "session", "run", 0)
+ require.NoError(t, err)
+ require.Len(t, events, 1)
+ _, err = manager.UpdateSteerEventData(ctx, "session", "run", evt.ID,
+ map[string]interface{}{"consumed": true})
+ require.NoError(t, err)
+ require.NoError(t, manager.AppendSteerEvents(ctx, "session", "run", []interfaces.StreamEvent{evt}))
+ events, _, err = manager.GetSteerEvents(ctx, "session", "run", 0)
+ require.NoError(t, err)
+ require.Len(t, events, 1)
+ require.Equal(t, true, events[0].Data["consumed"])
+ })
+ }
+}
diff --git a/internal/stream/redis_manager.go b/internal/stream/redis_manager.go
index f6b976e2f..f40d21d15 100644
--- a/internal/stream/redis_manager.go
+++ b/internal/stream/redis_manager.go
@@ -170,16 +170,33 @@ func (r *RedisStreamManager) AppendSteerEvents(
payloads = append(payloads, eventJSON)
}
- pipe := r.client.Pipeline()
- pipe.RPush(ctx, key, payloads...)
- pipe.Expire(ctx, key, r.ttl)
- pipe.Expire(ctx, r.buildLiveRunKey(sessionID), r.ttl)
- if _, err := pipe.Exec(ctx); err != nil {
+ // Deduplication is atomic with append: a timed-out POST can be retried
+ // while another replica is still accepting the same client ID.
+ args := append([]interface{}{int64(r.ttl / time.Second)}, payloads...)
+ if err := steerAppendUnique.Run(ctx, r.client, []string{key}, args...).Err(); err != nil {
return fmt.Errorf("failed to append steer events to Redis: %w", err)
}
+ _ = r.client.Expire(ctx, r.buildLiveRunKey(sessionID), r.ttl).Err()
return nil
}
+var steerAppendUnique = redis.NewScript(`
+local seen = {}
+for _, raw in ipairs(redis.call('LRANGE', KEYS[1], 0, -1)) do
+ local ok, event = pcall(cjson.decode, raw)
+ if ok and event.id then seen[event.id] = true end
+end
+for i = 2, #ARGV do
+ local event = cjson.decode(ARGV[i])
+ if not event.id or event.id == '' or not seen[event.id] then
+ redis.call('RPUSH', KEYS[1], ARGV[i])
+ if event.id then seen[event.id] = true end
+ end
+end
+redis.call('EXPIRE', KEYS[1], ARGV[1])
+return 1
+`)
+
// GetSteerEvents drains the steer sub-list starting fromOffset.
func (r *RedisStreamManager) GetSteerEvents(
ctx context.Context,
diff --git a/internal/types/agent.go b/internal/types/agent.go
index b7398e2d6..68fd8a93a 100644
--- a/internal/types/agent.go
+++ b/internal/types/agent.go
@@ -413,6 +413,12 @@ func IsPipelineToolCallID(id string) bool {
type AgentStep struct {
Iteration int `json:"iteration"` // Iteration number (0-indexed)
Thought string `json:"thought"` // LLM's reasoning/thinking (Think phase)
+ // UserMessagesBefore records consumed steer rows in delivery order, before
+ // this model response. Unlike timestamps, this remains unambiguous on replay.
+ UserMessagesBefore []string `json:"user_messages_before,omitempty"`
+ // IntermediateAnswer preserves a plain answer followed by a loop-end steer.
+ // The canonical final answer is still stored in Message.Content.
+ IntermediateAnswer bool `json:"intermediate_answer,omitempty"`
// ReasoningContent stores the OpenAI-protocol reasoning_content emitted by the
// model in this round. Persisted on AgentStep so cross-turn replay can put it
// back on the assistant message — required by MiMo / DeepSeek V3.2+ thinking
@@ -439,12 +445,13 @@ func (s *AgentStep) GetObservations() []string {
// AgentState tracks the execution state of an agent across iterations
type AgentState struct {
- CurrentRound int `json:"current_round"` // Current round number
- RoundSteps []AgentStep `json:"round_steps"` // All steps taken so far in the current round
- IsComplete bool `json:"is_complete"` // Whether agent has finished
- FinalAnswer string `json:"final_answer"` // The final answer to the query
- KnowledgeRefs []*SearchResult `json:"knowledge_refs"` // Collected knowledge references
- TurnUsage TokenUsage `json:"turn_usage"` // LLM token usage accumulated across every round of this turn
+ PendingSteerMessages []string `json:"-"`
+ CurrentRound int `json:"current_round"` // Current round number
+ RoundSteps []AgentStep `json:"round_steps"` // All steps taken so far in the current round
+ IsComplete bool `json:"is_complete"` // Whether agent has finished
+ FinalAnswer string `json:"final_answer"` // The final answer to the query
+ KnowledgeRefs []*SearchResult `json:"knowledge_refs"` // Collected knowledge references
+ TurnUsage TokenUsage `json:"turn_usage"` // LLM usage accumulated across this turn
}
// FunctionDefinition represents a function definition for LLM function calling
diff --git a/internal/types/qa_request.go b/internal/types/qa_request.go
index 5c2eb3b4a..d23fa7473 100644
--- a/internal/types/qa_request.go
+++ b/internal/types/qa_request.go
@@ -4,6 +4,14 @@ import (
"context"
)
+// SteerMessageContent adds delivery context only to model input. The persisted
+// user message and the UI always retain the user's original text.
+func SteerMessageContent(content string) string {
+ return "<steer_message>\n" + content + "\n</steer_message>\n<continue_task>\n" +
+ "This is guidance for the task in progress. Apply it and continue unfinished work " +
+ "unless the user explicitly changes or cancels the task.\n</continue_task>"
+}
+
// SteerSink abstracts the handler-side back half of mid-run message
// injection: draining the queued steer events and persisting accepted
// messages. Declared in types (not agent/interfaces) because both packages
diff --git a/internal/types/qa_request_test.go b/internal/types/qa_request_test.go
new file mode 100644
index 000000000..90364567e
--- /dev/null
+++ b/internal/types/qa_request_test.go
@@ -0,0 +1,26 @@
+package types
+
+import (
+ "strings"
+ "testing"
+)
+
+func TestSteerMessageContentKeepsContinuationHintSmall(t *testing.T) {
+ content := strings.Repeat("补充说明\n", 1000)
+ wrapped := SteerMessageContent(content)
+ if strings.Count(wrapped, content) != 1 {
+ t.Fatal("user update must appear exactly once")
+ }
+ if len(wrapped)-len(content) > 210 {
+ t.Fatal("steering delivery overhead exceeded 210 bytes")
+ }
+ if !strings.Contains(wrapped, "guidance for the task in progress") ||
+ !strings.Contains(wrapped, "explicitly changes or cancels") {
+ t.Fatal("steering must preserve the task while respecting explicit redirection")
+ }
+ for _, marker := range []string{"<steer_message>", "</steer_message>", "<continue_task>", "</continue_task>"} {
+ if strings.Count(wrapped, marker) != 1 {
+ t.Fatalf("expected a single %s", marker)
+ }
+ }
+}