fix(agent-poll): forced looks and a failing query no longer launch PowerShell nonstop (core 0.41.1)

core/main/agentPoll.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect, it } from 'vitest'
-import { AGENT_POLL_MAX, AGENT_POLL_MIN, createAgentPoll, type AgentPollDeps } from './agentPoll'
+import { AGENT_FAIL_MAX, AGENT_FORCE_GAP, AGENT_POLL_MAX, AGENT_POLL_MIN, createAgentPoll, type AgentPollDeps } from './agentPoll'
 
 /**
  * The poll's three rules, checked without a process list: ask only when a
@@ -47,9 +47,10 @@ describe('the agent poll', () => {
     w.set({ out: '100 1\n200 100 claude\n', now: 1 }) // claude starts, silently, inside the backoff
     poll()
     expect(w.sent).toEqual([['t1', false, null]])
+    w.set({ now: AGENT_FORCE_GAP })
     poll(true)
     expect(w.sent).toEqual([['t1', false, null], ['t1', true, 'claude']])
-    w.set({ out: '100 1\n', now: 2 }) // and quits
+    w.set({ out: '100 1\n', now: 2 * AGENT_FORCE_GAP }) // and quits
     poll(true)
     expect(w.sent.at(-1)).toEqual(['t1', false, null])
     const before = w.sent.length
@@ -140,10 +141,59 @@ describe('the agent poll', () => {
     // Same tick count, but there is still no answer: it asks again.
     w.set({ out: `100 1\n200 100 node ${CLAUDE}\n` })
     poll()
+    // (Inside the failure's backoff it waits; past it, it asks again.)
+    expect(w.asked()).toBe(1)
+    w.set({ now: 2 * AGENT_POLL_MIN })
+    poll()
     expect(w.asked()).toBe(2)
     expect(w.sent).toEqual([['t1', true, 'claude']])
   })
 
+  // Perf review 2026-10-16: a query that always failed launched PowerShell
+  // every 2.5 s for the app's life, and a Claude the poll cannot see (WSL,
+  // SSH) had every title frame force a new one.
+  it('a failing query backs off, doubling to a minute, forced looks included, and says so once', () => {
+    const w = world()
+    let failing = 0
+    const poll = createAgentPoll(w.send, { ...w.deps, failing: () => (failing += 1) })
+    w.set({ out: null })
+    let t = 0
+    const times: number[] = []
+    for (let i = 0; i < 400; i += 1) {
+      t += 500
+      w.set({ now: t, ticks: i })
+      const before = w.asked()
+      poll(i % 2 === 0)
+      if (w.asked() > before) times.push(t)
+    }
+    const gaps = times.slice(1).map((x, i) => x - times[i])
+    expect(gaps.slice(0, 4)).toEqual([5000, 10000, 20000, 40000])
+    expect(Math.max(...gaps)).toBe(AGENT_FAIL_MAX)
+    expect(failing).toBe(1)
+    // An answer resets it: the next look comes at the usual pace.
+    w.set({ out: '100 1\n', now: t + AGENT_FAIL_MAX })
+    poll()
+    w.set({ now: t + AGENT_FAIL_MAX + AGENT_POLL_MIN, ticks: 999 })
+    const before = w.asked()
+    poll()
+    expect(w.asked()).toBe(before + 1)
+  })
+
+  it('a forced look waits for the gap since the last query, then looks', () => {
+    const w = world()
+    const poll = createAgentPoll(w.send, w.deps)
+    poll()
+    expect(w.asked()).toBe(1)
+    for (let i = 1; i < 10; i += 1) {
+      w.set({ now: i * 400 })
+      poll(true)
+    }
+    expect(w.asked()).toBe(1)
+    w.set({ now: AGENT_FORCE_GAP })
+    poll(true)
+    expect(w.asked()).toBe(2)
+  })
+
   it('forgets a session that ended, so its id starts clean', () => {
     const w = world()
     const poll = createAgentPoll(w.send, w.deps)

core/main/agentPoll.ts

@@ -3,12 +3,26 @@ import { execFile } from 'child_process'
 import type { DetectedAgent } from '../shared/types'
 import { parseProcLines, treeAgentKind } from './agentDetect'
 import { livePids, ptyOutputTicks } from './terminal'
+import { diagMain } from './diagLog'
 
 /* ------------------------------------------------------------------ *
  * The agent poll's shape. See the poll itself for why it is like this.
  * ------------------------------------------------------------------ */
 export const AGENT_POLL_MIN = 2500
 export const AGENT_POLL_MAX = 20000
+/**
+ * A FORCED LOOK WAITS THIS LONG after the last query started, and a query
+ * that FAILS backs off doubling up to the ceiling below, forced looks
+ * included (perf review 2026-10-16). Before, a title or hook signal forced a
+ * PowerShell + CIM query on every frame of a Claude the poll cannot see
+ * (Claude under WSL or SSH: its process is not in the Windows tree, so the
+ * poll never reports it and the renderer asks again on the next frame), half
+ * to a whole core while it worked; and a query that always fails (WMI broken,
+ * AppLocker, Constrained Language) launched PowerShell every 2.5 s for the
+ * app's whole life.
+ */
+export const AGENT_FORCE_GAP = 5000
+export const AGENT_FAIL_MAX = 60000
 /**
  * "pid ppid" for everything, plus the command line only where a cheap word
  * match hits. The full dump with every command line was megabytes; this is a
@@ -42,13 +56,17 @@ export interface AgentPollDeps {
   ticks: () => number
   query: (done: (stdout: string | null) => void) => void
   now: () => number
+  /** Said once when the query has failed three times in a row: why agent
+   *  marks are missing, for the diagnostics log. */
+  failing?: () => void
 }
 
 const REAL: AgentPollDeps = {
   pids: livePids,
   ticks: ptyOutputTicks,
   query: queryProcesses,
-  now: () => Date.now()
+  now: () => Date.now(),
+  failing: () => diagMain().write('main', 'agent-poll-failing', { note: 'the process query keeps failing: agent marks rely on titles and hooks alone' })
 }
 
 /**
@@ -90,11 +108,20 @@ export function createAgentPoll(send: AgentSend, deps: AgentPollDeps = REAL): Ag
   // started again between two looks. Held until a look STARTS, so a look
   // already in flight, taken before the ask, does not spend it.
   const asked = new Set<string>()
+  /** When the last query started, and the failures in a row since one answered. */
+  let lastStart = Number.NEGATIVE_INFINITY
+  let fails = 0
+  let failNext = 0
   // `force` skips the two waits below, never the answer (#73): a look the
   // renderer asked for sends only what CHANGED, like any other.
-  const poll = (force = false): void => {
+  const poll = (asForced = false): void => {
     const pids = deps.pids()
     if (!pids.length || agentBusy) return
+    const startAt = deps.now()
+    // A failing query backs off for everyone, forced looks too.
+    if (fails && startAt < failNext) return
+    // A forced look inside the gap is an ordinary one: its waits apply.
+    const force = asForced && startAt - lastStart >= AGENT_FORCE_GAP
     const ticks = deps.ticks()
     const quiet = ticks === agentSeenTicks
     const known = pids.every((s) => agentState.has(s.id) && !asked.has(s.id))
@@ -107,6 +134,7 @@ export function createAgentPoll(send: AgentSend, deps: AgentPollDeps = REAL): Ag
     const now = deps.now()
     if (known && now < agentNext && !force) return
     agentBusy = true
+    lastStart = startAt
     const asking = new Set(asked)
     asked.clear()
     deps.query((stdout) => {
@@ -118,8 +146,12 @@ export function createAgentPoll(send: AgentSend, deps: AgentPollDeps = REAL): Ag
       const rows = stdout ? parseProcLines(stdout) : []
       if (!rows.length) {
         asking.forEach((id) => asked.add(id))
+        fails += 1
+        failNext = deps.now() + Math.min(AGENT_FAIL_MAX, AGENT_POLL_MIN * 2 ** fails)
+        if (fails === 3) deps.failing?.()
         return
       }
+      fails = 0
       agentSeenTicks = ticks
       let changed = false
       for (const { id, pid } of deps.pids()) {

core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "prism-term-core",
-  "version": "0.41.0",
+  "version": "0.41.1",
   "description": "What Prism Terminal and Prism share: the terminal (pty, shells, agent detection and indicator, themes, links, the panel, dictation) and the update chip with its window. TypeScript source, compiled by the host.",
   "license": "MIT",
   "private": true,

core/renderer/lib/useAgentIndicator.ts

@@ -101,6 +101,13 @@ export function useAgentIndicator(activeId: string | null): AgentIndicator {
    */
   const [moved, setMoved] = useState(0)
   const movedAt = useRef(0)
+  // Whether a look by the pointer could take anything down: only a Finished
+  // or Failed mark on the tab in front (perf review 2026-10-16: every move
+  // over the unfocused window re-rendered App twice a second for nothing).
+  const frontMarked = useRef(false)
+  useEffect(() => {
+    frontMarked.current = !!activeId && (doneIds.has(activeId) || failedIds.has(activeId))
+  }, [activeId, doneIds, failedIds])
   useEffect(() => {
     const MOVE_LOOK_MS = 1000
     let last = 0
@@ -111,6 +118,7 @@ export function useAgentIndicator(activeId: string | null): AgentIndicator {
       if (looking.current.focused) return
       const now = Date.now()
       movedAt.current = now
+      if (!frontMarked.current) return
       // A look every half second at most: a moving pointer fires constantly.
       if (now - last < MOVE_LOOK_MS / 2) return
       last = now
@@ -156,6 +164,21 @@ export function useAgentIndicator(activeId: string | null): AgentIndicator {
   const titled = useRef(new Set<string>())
   /** Sessions the process poll has said host an agent. */
   const polled = useRef(new Set<string>())
+  /**
+   * ASK THE POLL TO LOOK, AT MOST ONCE IN 15 S PER TAB (perf review
+   * 2026-10-16). Every title frame and hook signal of an agent the poll has
+   * not reported asked for a look, and a Claude the poll can never see (under
+   * WSL or SSH) kept PowerShell queries running back to back while it worked.
+   * Main also holds forced looks 5 s apart (agentPoll).
+   */
+  const lookAsked = useRef(new Map<string, number>())
+  const askLook = useCallback((id: string): void => {
+    if (polled.current.has(id)) return
+    const now = Date.now()
+    if (now - (lookAsked.current.get(id) ?? Number.NEGATIVE_INFINITY) < 15_000) return
+    lookAsked.current.set(id, now)
+    termApi().termAgentLook?.()
+  }, [])
   /** The one clearing timer per fallback-scored session. */
   const fallbackTimers = useRef(new Map<string, number>())
   const stopFallback = useCallback((id: string): void => {
@@ -389,7 +412,7 @@ export function useAgentIndicator(activeId: string | null): AgentIndicator {
         // Present by its title alone: have the poll look now, so it sees the
         // agent while it runs and can say when it leaves (#73). Only the poll
         // ever takes a title's claim back.
-        if (!polled.current.has(id)) termApi().termAgentLook?.()
+        askLook(id)
         if (!titled.current.has(id)) {
           titled.current.add(id)
           outputRuns.current.delete(id)
@@ -424,7 +447,7 @@ export function useAgentIndicator(activeId: string | null): AgentIndicator {
           return next
         })
       }),
-    [stopFallback, checkQuestion, applyHook, stopAnswerCheck]
+    [stopFallback, checkQuestion, applyHook, stopAnswerCheck, askLook]
   )
 
   /**
@@ -437,7 +460,7 @@ export function useAgentIndicator(activeId: string | null): AgentIndicator {
     () =>
       onAgentSignal((id, signal) => {
         logAgentHook(id, signal.state, signal.kind)
-        if (!polled.current.has(id)) termApi().termAgentLook?.()
+        askLook(id)
         if (!hooked.current.has(id)) {
           outputRuns.current.delete(id)
           stopFallback(id)
@@ -449,7 +472,7 @@ export function useAgentIndicator(activeId: string | null): AgentIndicator {
         setAgentIds((prev) => (prev.has(id) ? prev : new Set(prev).add(id)))
         applyHook(id, signal)
       }),
-    [stopFallback, applyHook]
+    [stopFallback, applyHook, askLook]
   )
 
   /**
@@ -557,6 +580,7 @@ export function useAgentIndicator(activeId: string | null): AgentIndicator {
     outputRuns.current.delete(id)
     titled.current.delete(id)
     polled.current.delete(id)
+    lookAsked.current.delete(id)
     agentKinds.current.delete(id)
     forgetAgentTitle(id)
     stopFallback(id)

core/shared/termVersion.ts

@@ -7,5 +7,5 @@
  * so a core release that bumps `core/package.json` (or xterm) without this
  * fails `npm test` (core/README.md, "Releasing the core").
  */
-export const TERM_CORE_VERSION = '0.41.0'
+export const TERM_CORE_VERSION = '0.41.1'
 export const XTERM_VERSION = '6.0.0'

package.json

@@ -1,7 +1,7 @@
 {
   "name": "prism-terminal",
   "productName": "Prism Terminal",
-  "version": "0.55.0",
+  "version": "0.55.1",
   "description": "A tabbed Windows terminal for AI CLIs.",
   "main": "./out/main/index.js",
   "author": "Max",