Add the ability to stop inference

app.js

@@ -338,6 +338,7 @@ const qvacView = createQvacView({
     messages: $('qvac-messages'),
     input: /** @type {HTMLTextAreaElement|null} */ ($('qvac-input')),
     sendBtn: $('qvac-send'),
+    stopBtn: $('qvac-stop'),
     status: $('qvac-status'),
     modelChip: $('qvac-model-chip'),
     samples: $('qvac-samples'),

electron/qvac-service.cjs

@@ -19,6 +19,10 @@ let modelId = null
 let loadedChatModel = null
 /** @type {Error|string|null} */
 let lastError = null
+/** Active completion requestId for cancel() */
+let activeRequestId = null
+/** Soft flag — event loop checks between iterations */
+let cancelRequested = false
 
 function appRoot() {
   try {
@@ -162,6 +166,7 @@ async function complete(opts, onToken, onThinking) {
   const s = await ensureSdk()
   if (!modelId) throw new Error('No model loaded')
 
+  cancelRequested = false
   const run = s.completion({
     modelId,
     history: opts.history || [],
@@ -170,14 +175,21 @@ async function complete(opts, onToken, onThinking) {
     // Best-effort: emit thinkingDelta when model uses <think> blocks
     captureThinking: true,
   })
+  activeRequestId = run.requestId || run.id || null
 
   let content = ''
   let thinking = ''
   /** @type {any[]} */
   const toolCalls = []
+  let stopped = false
 
+  try {
     if (run.events) {
       for await (const ev of run.events) {
+        if (cancelRequested) {
+          stopped = true
+          break
+        }
         if (ev.type === 'thinkingDelta' && (ev.text || ev.delta)) {
           const t = ev.text || ev.delta || ''
           thinking += t
@@ -201,11 +213,26 @@ async function complete(opts, onToken, onThinking) {
       }
     } else if (run.tokenStream) {
       for await (const token of run.tokenStream) {
+        if (cancelRequested) {
+          stopped = true
+          break
+        }
         content += token
         onToken?.(token)
       }
     }
 
+    if (stopped || cancelRequested) {
+      return {
+        contentText: content,
+        thinkingText: thinking,
+        toolCalls: [],
+        mode: 'qvac',
+        stopped: true,
+        error: null,
+      }
+    }
+
     const final = run.final ? await run.final : { contentText: content, toolCalls }
     const calls = (final.toolCalls?.length ? final.toolCalls : toolCalls).filter(Boolean)
     // Prefer full text that may still include think tags if captureThinking missed
@@ -221,7 +248,42 @@ async function complete(opts, onToken, onThinking) {
       toolCalls: calls,
       stats: final.stats,
       mode: 'qvac',
+      stopped: Boolean(cancelRequested),
     }
+  } finally {
+    activeRequestId = null
+    cancelRequested = false
+  }
+}
+
+/**
+ * Cancel in-flight completion (and optional model-wide inference).
+ */
+async function cancelInference() {
+  cancelRequested = true
+  const rid = activeRequestId
+  try {
+    const s = sdk || (await ensureSdk().catch(() => null))
+    if (s?.cancel) {
+      if (rid) {
+        try {
+          await s.cancel({ requestId: rid })
+        } catch {
+          // fall through to broad cancel
+        }
+      }
+      if (modelId) {
+        try {
+          await s.cancel({ modelId, kind: 'completion' })
+        } catch {
+          // ignore
+        }
+      }
+    }
+  } catch {
+    // ignore
+  }
+  return { ok: true, requestId: rid }
 }
 
 function getStatus() {
@@ -288,15 +350,34 @@ function registerIpc(ipcMain) {
       )
     } catch (err) {
       lastError = err
+      const msg = err?.message || String(err)
+      // Cancel often surfaces as inference cancelled
+      if (/cancel|abort|stopped/i.test(msg)) {
         return {
           contentText: '',
           thinkingText: '',
           toolCalls: [],
-        error: err?.message || String(err),
+          stopped: true,
+          mode: 'qvac',
+        }
+      }
+      return {
+        contentText: '',
+        thinkingText: '',
+        toolCalls: [],
+        error: msg,
         mode: 'error',
       }
     }
   })
+
+  ipcMain.handle('peardata:qvac-cancel', async () => {
+    try {
+      return await cancelInference()
+    } catch (err) {
+      return { ok: false, error: err?.message || String(err) }
+    }
+  })
 }
 
 module.exports = {

index.html

@@ -483,6 +483,7 @@
           <form id="qvac-form" class="qvac-composer" onsubmit="return false">
             <textarea id="qvac-input" rows="2" placeholder="Ask about host health, metrics, anomalies, processes…" autocomplete="off"></textarea>
             <button type="button" id="qvac-send" class="btn btn-primary">Send</button>
+            <button type="button" id="qvac-stop" class="btn btn-ghost qvac-stop-btn hidden" title="Stop inference (Esc)">Stop</button>
           </form>
         </div>
       </section>

ui/qvac/agents.js

@@ -227,6 +227,7 @@ export function pickAgents(userText, maxAgents = 3) {
  *   tools: { run: (name: string, args?: object) => Promise<any> },
  *   query: string,
  *   maxAgents?: number,
+ *   isAborted?: () => boolean,
  *   onAgent?: (ev: {
  *     id: string,
  *     label: string,
@@ -239,15 +240,36 @@ export function pickAgents(userText, maxAgents = 3) {
 export async function runSubAgents(opts) {
   const specs = pickAgents(opts.query, opts.maxAgents ?? 3)
   const tools = opts.tools
+  const aborted = () => Boolean(opts.isAborted?.())
 
   /** @type {Array<{ id: string, label: string, status: string, summary: string, result?: any, error?: string }>} */
   const agents = []
 
   await Promise.all(
     specs.map(async (spec) => {
+      if (aborted()) {
+        agents.push({
+          id: spec.id,
+          label: spec.label,
+          status: 'error',
+          summary: 'stopped',
+          error: 'stopped',
+        })
+        return
+      }
       opts.onAgent?.({ id: spec.id, label: spec.label, status: 'start' })
       try {
         const result = await spec.run(tools)
+        if (aborted()) {
+          agents.push({
+            id: spec.id,
+            label: spec.label,
+            status: 'error',
+            summary: 'stopped',
+            error: 'stopped',
+          })
+          return
+        }
         const summary = spec.summarize(result)
         const row = {
           id: spec.id,

ui/qvac/engine.js

@@ -95,6 +95,47 @@ function getElectronIpc() {
 }
 
 export function createQvacEngine(deps) {
+  /** Soft cancel for current complete() / multi-round tool loop */
+  let abortGeneration = false
+  /** @type {string|null} */
+  let activeRequestId = null
+
+  /**
+   * Stop in-flight inference (and pending tool rounds).
+   * Safe to call when idle.
+   */
+  async function stop() {
+    abortGeneration = true
+    const ipc = getElectronIpc()
+    try {
+      if (ipc && sdkMode === 'main') {
+        await ipc.invoke('peardata:qvac-cancel')
+      } else if (sdk?.cancel) {
+        if (activeRequestId) {
+          try {
+            await sdk.cancel({ requestId: activeRequestId })
+          } catch {
+            // ignore
+          }
+        }
+        if (modelId) {
+          try {
+            await sdk.cancel({ modelId, kind: 'completion' })
+          } catch {
+            // ignore
+          }
+        }
+      }
+    } catch (err) {
+      deps.log?.(`QVAC stop: ${err?.message || err}`)
+    }
+    return { ok: true }
+  }
+
+  function isAborted() {
+    return abortGeneration
+  }
+
   /** @type {any} */
   let sdk = null
   /** @type {'direct'|'main'|null} */
@@ -442,6 +483,7 @@ export function createQvacEngine(deps) {
    */
   async function complete(history, opts = {}) {
     touchActivity()
+    abortGeneration = false
     const profile = getProfile(profileId || 'recommended')
     const prefs = deps.getPrefs?.() || {}
     const userLast = [...history].reverse().find((m) => m.role === 'user')
@@ -471,6 +513,10 @@ export function createQvacEngine(deps) {
       ...history.filter((m) => m.role !== 'system'),
     ]
 
+    if (abortGeneration) {
+      return { contentText: '_(Stopped.)_', toolCalls: [], mode: 'stopped', stopped: true }
+    }
+
     if (modelId && (sdkMode === 'main' || (sdk && sdk.completion))) {
       return completeWithSdk(fullHistory, profile, opts, prefs)
     }
@@ -478,18 +524,40 @@ export function createQvacEngine(deps) {
     status = status === 'ready' ? status : 'fallback'
     // If multi-agent pack already ran, prefer its synthesis over naive fallback
     if (opts.agentFallbackText) {
+      if (abortGeneration) {
+        return { contentText: '_(Stopped.)_', toolCalls: [], mode: 'stopped', stopped: true }
+      }
       const text = String(opts.agentFallbackText)
       if (opts.onToken) {
         const chunk = 24
-        for (let i = 0; i < text.length; i += chunk) opts.onToken(text.slice(i, i + chunk))
+        for (let i = 0; i < text.length; i += chunk) {
+          if (abortGeneration) break
+          opts.onToken(text.slice(i, i + chunk))
         }
-      opts.onEvent?.({ type: 'completionDone', stopReason: 'eos' })
+      }
-      return { contentText: text, toolCalls: opts.agentToolCalls || [], mode: 'fallback' }
+      opts.onEvent?.({ type: 'completionDone', stopReason: abortGeneration ? 'cancelled' : 'eos' })
+      return {
+        contentText: abortGeneration ? `${text}\n\n_(Stopped.)_` : text,
+        toolCalls: opts.agentToolCalls || [],
+        mode: abortGeneration ? 'stopped' : 'fallback',
+        stopped: abortGeneration,
+      }
+    }
+    if (abortGeneration) {
+      return { contentText: '_(Stopped.)_', toolCalls: [], mode: 'stopped', stopped: true }
     }
     const result = await fallbackComplete(userLast?.content || '', deps.tools, {
       catalog: deps.getCatalog?.() || {},
       rag: prefs.rag !== false,
     })
+    if (abortGeneration) {
+      return {
+        contentText: (result.contentText || '') + '\n\n_(Stopped.)_',
+        toolCalls: result.toolCalls || [],
+        mode: 'stopped',
+        stopped: true,
+      }
+    }
     for (const c of result.toolCalls || []) {
       opts.onTool?.(c.name, c.args, c.result)
     }
@@ -497,11 +565,16 @@ export function createQvacEngine(deps) {
       const text = result.contentText || ''
       const chunk = 24
       for (let i = 0; i < text.length; i += chunk) {
+        if (abortGeneration) break
         opts.onToken(text.slice(i, i + chunk))
       }
     }
-    opts.onEvent?.({ type: 'completionDone', stopReason: 'eos' })
+    opts.onEvent?.({ type: 'completionDone', stopReason: abortGeneration ? 'cancelled' : 'eos' })
-    return result
+    return {
+      ...result,
+      stopped: abortGeneration,
+      mode: abortGeneration ? 'stopped' : result.mode,
+    }
   }
 
   async function completeWithSdk(history, profile, opts, prefs = {}) {
@@ -523,6 +596,14 @@ export function createQvacEngine(deps) {
     )
 
     for (let round = 0; round < 4; round++) {
+      if (abortGeneration) {
+        return {
+          contentText: '_(Stopped.)_',
+          toolCalls: [],
+          mode: 'stopped',
+          stopped: true,
+        }
+      }
       touchActivity()
       // Re-compact each tool round (payloads grow fast)
       messages = compactMessages(messages, {
@@ -539,14 +620,32 @@ export function createQvacEngine(deps) {
         /** @type {(( _e: any, t: string) => void)|null} */
         let onThink = null
         try {
-          onTok = (_e, t) => opts.onToken?.(t)
+          onTok = (_e, t) => {
-          onThink = (_e, t) => opts.onThinking?.(t)
+            if (!abortGeneration) opts.onToken?.(t)
+          }
+          onThink = (_e, t) => {
+            if (!abortGeneration) opts.onThinking?.(t)
+          }
           ipc.on('peardata:qvac-token', onTok)
           ipc.on('peardata:qvac-thinking', onThink)
           const final = await ipc.invoke('peardata:qvac-complete', {
             history: messages,
             tools: toolDefs,
           })
+          if (final?.stopped || abortGeneration) {
+            let text = final?.contentText || ''
+            if (final?.thinkingText && !/<think/i.test(text)) {
+              text = `<think>\n${final.thinkingText}\n</think>\n${text}`
+            }
+            if (!text.trim()) text = '_(Stopped.)_'
+            else if (!/stopped/i.test(text)) text += '\n\n_(Stopped.)_'
+            return {
+              contentText: text,
+              toolCalls: [],
+              mode: 'stopped',
+              stopped: true,
+            }
+          }
           if (final?.error && !final.contentText) {
             if (isContextOverflowError(final.error) && overflowRetries < 3) {
               overflowRetries++
@@ -602,6 +701,14 @@ export function createQvacEngine(deps) {
           }
           messages = [...messages, { role: 'assistant', content: text }]
           for (const tc of calls) {
+            if (abortGeneration) {
+              return {
+                contentText: (text || '') + '\n\n_(Stopped.)_',
+                toolCalls: [],
+                mode: 'stopped',
+                stopped: true,
+              }
+            }
             const name = tc.name || tc.function?.name
             let args = tc.arguments || tc.function?.arguments || {}
             if (typeof args === 'string') {
@@ -647,13 +754,20 @@ export function createQvacEngine(deps) {
         tools: toolDefs,
         captureThinking: true,
       })
+      activeRequestId = run.requestId || run.id || null
 
       let content = ''
       /** @type {Array<{ name: string, arguments?: any, id?: string }>} */
       const toolCalls = []
+      let streamStopped = false
 
+      try {
         if (run.events) {
           for await (const ev of run.events) {
+            if (abortGeneration) {
+              streamStopped = true
+              break
+            }
             opts.onEvent?.(ev)
             if (ev.type === 'thinkingDelta' && (ev.text || ev.delta)) {
               opts.onThinking?.(ev.text || ev.delta)
@@ -671,11 +785,24 @@ export function createQvacEngine(deps) {
           }
         } else if (run.tokenStream) {
           for await (const token of run.tokenStream) {
+            if (abortGeneration) {
+              streamStopped = true
+              break
+            }
             content += token
             opts.onToken?.(token)
           }
         }
 
+        if (streamStopped || abortGeneration) {
+          return {
+            contentText: (content || '') + (content ? '\n\n' : '') + '_(Stopped.)_',
+            toolCalls: [],
+            mode: 'stopped',
+            stopped: true,
+          }
+        }
+
         const final = run.final ? await run.final : { contentText: content, toolCalls }
 
         const calls = (final.toolCalls?.length ? final.toolCalls : toolCalls).filter(Boolean)
@@ -688,8 +815,19 @@ export function createQvacEngine(deps) {
           }
         }
 
-      messages = [...messages, { role: 'assistant', content: final.contentText || content || '' }]
+        messages = [
+          ...messages,
+          { role: 'assistant', content: final.contentText || content || '' },
+        ]
         for (const tc of calls) {
+          if (abortGeneration) {
+            return {
+              contentText: (final.contentText || content || '') + '\n\n_(Stopped.)_',
+              toolCalls: [],
+              mode: 'stopped',
+              stopped: true,
+            }
+          }
           const name = tc.name || tc.function?.name
           let args = tc.arguments || tc.function?.arguments || {}
           if (typeof args === 'string') {
@@ -707,6 +845,18 @@ export function createQvacEngine(deps) {
             name,
           })
         }
+      } finally {
+        activeRequestId = null
+      }
+    }
+
+    if (abortGeneration) {
+      return {
+        contentText: '_(Stopped.)_',
+        toolCalls: [],
+        mode: 'stopped',
+        stopped: true,
+      }
     }
 
     return {
@@ -721,6 +871,8 @@ export function createQvacEngine(deps) {
     loadProfile,
     unload,
     complete,
+    stop,
+    isAborted,
     getStatus,
     tryLoadSdk,
     touchActivity,

ui/qvac/index.js

@@ -22,6 +22,7 @@ import { chartIdsFromToolLog, mountChartEmbeds } from './chart-embed.js'
  *     messages: HTMLElement|null,
  *     input: HTMLTextAreaElement|null,
  *     sendBtn: HTMLElement|null,
+ *     stopBtn?: HTMLElement|null,
  *     status: HTMLElement|null,
  *     modelChip: HTMLElement|null,
  *     samples: HTMLElement|null,
@@ -792,13 +793,36 @@ export function createQvacView(opts) {
     }
   }
 
+  function setBusyUi(on) {
+    busy = on
+    const send = opts.els.sendBtn
+    const stop = opts.els.stopBtn
+    if (send) {
+      send.disabled = on
+      send.classList.toggle('hidden', on)
+    }
+    if (stop) {
+      stop.classList.toggle('hidden', !on)
+      stop.disabled = false
+    }
+  }
+
+  async function stopInference() {
+    if (!busy) return
+    setStatus('Stopping…', 'busy')
+    try {
+      await engine.stop?.()
+    } catch (err) {
+      opts.log?.(`QVAC stop: ${err?.message || err}`)
+    }
+  }
+
   async function send() {
     const input = opts.els.input
     const text = (input?.value || '').trim()
     if (!text || busy) return
     if (input) input.value = ''
-    busy = true
+    setBusyUi(true)
-    opts.els.sendBtn && (opts.els.sendBtn.disabled = true)
     appendMsg('user', text)
     const assistantUi = appendMsg('assistant', '…', { streaming: true })
     const streamBody = assistantUi?.body
@@ -826,7 +850,11 @@ export function createQvacView(opts) {
         maxSubAgents: Number(settings().qvacMaxSubAgents) || 3,
       }
 
-      if (shouldUseSubAgents(text, prefs) && opts.isConnected?.()) {
+      if (
+        shouldUseSubAgents(text, prefs) &&
+        opts.isConnected?.() &&
+        !engine.isAborted?.()
+      ) {
         setStatus('Spinning up agents…', 'busy')
         if (streamBody) {
           streamBody.innerHTML = formatMdLite('_Dispatching multi-agent investigation…_')
@@ -835,11 +863,19 @@ export function createQvacView(opts) {
           tools,
           query: text,
           maxAgents: prefs.maxSubAgents,
+          isAborted: () => Boolean(engine.isAborted?.()),
           onAgent: (ev) => {
+            if (engine.isAborted?.()) return
             paintAgentEvent(ev)
             setStatus(`Agent: ${ev.label}…`, 'busy')
           },
         })
+        if (engine.isAborted?.()) {
+          acc = '_(Stopped during multi-agent investigation.)_'
+          paintAssistant(streamBody, acc, false)
+          setStatus('Stopped', 'warn')
+          return
+        }
         completeOpts.agentBriefing = pack.contextText
         completeOpts.agentFallbackText = synthesizeFromAgents(text, pack)
         completeOpts.agentToolCalls = pack.agents.map((a) => ({
@@ -862,6 +898,13 @@ export function createQvacView(opts) {
         setStatus('Synthesizing…', 'busy')
       }
 
+      if (engine.isAborted?.()) {
+        acc = '_(Stopped.)_'
+        paintAssistant(streamBody, acc, false)
+        setStatus('Stopped', 'warn')
+        return
+      }
+
       const result = await engine.complete(hist, {
         ...completeOpts,
         onToken: (t) => {
@@ -876,6 +919,7 @@ export function createQvacView(opts) {
           paintAssistant(streamBody, raw, true)
         },
         onTool: (name, args, res) => {
+          if (engine.isAborted?.()) return
           toolLog.push({ name, args, result: res })
           setStatus(`Tool: ${name}…`, 'busy')
           if (assistantUi?.div) {
@@ -905,6 +949,7 @@ export function createQvacView(opts) {
         assistantUi.div.appendChild(renderToolChips(toolLog))
       }
       // Live sparklines under the answer for chart-related tools
+      if (!result.stopped) {
         const embedIds = chartIdsFromToolLog(toolLog)
         if (embedIds.length && assistantUi?.div && opts.manager?.request) {
           mountChartEmbeds(assistantUi.div, embedIds, {
@@ -926,14 +971,27 @@ export function createQvacView(opts) {
             },
           }).catch(() => {})
         }
+      }
+      if (result.stopped || result.mode === 'stopped') {
+        setStatus('Stopped', 'warn')
+      } else {
         setStatus(result.mode === 'fallback' ? 'Ready (tools-only)' : 'Ready', 'ok')
+      }
     } catch (err) {
       const msg = err?.message || String(err)
+      if (/cancel|abort|stopped/i.test(msg)) {
+        if (streamBody && !acc) {
+          paintAssistant(streamBody, '_(Stopped.)_', false)
+        } else if (streamBody && acc) {
+          paintAssistant(streamBody, acc + '\n\n_(Stopped.)_', false)
+        }
+        setStatus('Stopped', 'warn')
+      } else {
         if (streamBody) streamBody.innerHTML = formatMdLite(`Error: ${msg}`)
         setStatus(msg, 'error')
+      }
     } finally {
-      busy = false
+      setBusyUi(false)
-      if (opts.els.sendBtn) opts.els.sendBtn.disabled = false
       syncModelChip()
     }
   }
@@ -1003,12 +1061,25 @@ export function createQvacView(opts) {
 
   function bind() {
     opts.els.sendBtn?.addEventListener('click', () => send())
+    opts.els.stopBtn?.addEventListener('click', () => stopInference())
     opts.els.input?.addEventListener('keydown', (ev) => {
+      if (ev.key === 'Escape' && busy) {
+        ev.preventDefault()
+        stopInference()
+        return
+      }
       if (ev.key === 'Enter' && !ev.shiftKey) {
         ev.preventDefault()
         send()
       }
     })
+    // Global Esc while QVAC tab is active
+    opts.els.root?.addEventListener('keydown', (ev) => {
+      if (ev.key === 'Escape' && busy) {
+        ev.preventDefault()
+        stopInference()
+      }
+    })
     opts.els.resetBtn?.addEventListener('click', () => {
       closeSettings()
       resetOnboarding()

ui/styles.css

@@ -4329,12 +4329,27 @@ html[data-theme='light'] .proc-detail-cmd {
 
 .qvac-composer {
   display: grid;
-  grid-template-columns: 1fr auto;
+  grid-template-columns: 1fr auto auto;
   gap: 10px;
   align-items: end;
   padding-bottom: 8px;
 }
 
+.qvac-stop-btn {
+  flex-shrink: 0;
+  border-color: color-mix(in srgb, #f87171 50%, var(--border-color)) !important;
+  color: #fca5a5 !important;
+}
+
+.qvac-stop-btn:hover:not(:disabled) {
+  background: color-mix(in srgb, #f87171 18%, transparent) !important;
+}
+
+.qvac-stop-btn.hidden,
+#qvac-send.hidden {
+  display: none !important;
+}
+
 .qvac-composer textarea {
   width: 100%;
   resize: vertical;