perf(agent-manager): reduce GitHub API usage in PR status polling (#8582)

Poll only the active worktree on each 15s timer tick instead of all
worktrees. Inactive worktrees refresh when selected or on manual
refresh. On first poll and visibility resume, all worktrees are fetched
once to populate badges.

Add a 60s TTL cache on PR lookup results so repeated polls skip the
expensive multi-strategy gh pr view cascade and only re-fetch checks
and comments (the fast-changing data).

With 10 worktrees this reduces API calls from ~5,000-9,800/hr to
~480-720/hr, staying well within GitHub's 5,000/hr rate limit.

packages/kilo-vscode/src/agent-manager/PRStatusPoller.ts

@@ -14,6 +14,7 @@ interface PRStatusPollerOptions {
 const GH_PROBE_TTL = 300_000 // 5 minutes — gh installation state rarely changes at runtime
 const MAX_BACKOFF = 120_000 // 2 minutes — cap for exponential backoff on repeated errors
 const BACKOFF_MULTIPLIER = 2
+const PR_LOOKUP_TTL = 60_000 // 1 minute — PR number for a branch rarely changes
 
 export class PRStatusPoller {
   private timer: ReturnType<typeof setTimeout> | undefined
@@ -27,6 +28,7 @@ export class PRStatusPoller {
   private ghProbeTime = 0
   private activeWorktreeId: string | undefined
   private cachedRepo: { owner: string; name: string; cwd: string } | undefined
+  private prCache = new Map<string, { result: PRResult | null; expires: number }>()
   private readonly intervalMs: number
 
   constructor(private readonly options: PRStatusPollerOptions) {
@@ -48,9 +50,12 @@ export class PRStatusPoller {
     this.visible = visible
     if (!this.active) return
     if (visible) {
-      // Resume — poll immediately then schedule normally
+      // Resume — expire all PR caches and fetch all worktrees once to catch up,
+      // then resume the normal active-only poll cycle.
       if (this.timer) clearTimeout(this.timer)
       this.timer = undefined
+      this.prCache.clear()
+      this.lastHash.clear()
       void this.poll()
       return
     }
@@ -74,16 +79,22 @@ export class PRStatusPoller {
     this.ghAvailable = undefined
     this.ghProbeTime = 0
     this.cachedRepo = undefined
+    this.prCache.clear()
   }
 
-  /** Force-refresh a specific worktree immediately. */
+  /** Force-refresh a specific worktree immediately, bypassing the PR cache. */
   refresh(worktreeId: string): void {
     if (!this.active) return
+    this.prCache.delete(worktreeId)
     void this.fetchOne(worktreeId)
   }
 
   setActiveWorktreeId(id: string | undefined): void {
+    const prev = this.activeWorktreeId
     this.activeWorktreeId = id
+    // When switching to a different worktree, fetch it immediately so the
+    // badge updates without waiting for the next poll cycle.
+    if (id && id !== prev && this.active) void this.fetchOne(id)
   }
 
   private start(): void {
@@ -146,8 +157,20 @@ export class PRStatusPoller {
     }
 
     this.lastError = undefined
+
+    // Only poll the active worktree on each timer tick. Inactive worktrees
+    // refresh when selected (setActiveWorktreeId) or manually (refresh()).
+    // On first poll (lastHash empty) fetch all worktrees once to populate badges.
     const worktrees = this.options.getWorktrees()
-    const results = await Promise.allSettled(worktrees.map((wt) => this.fetchOne(wt.id)))
+    const initial = this.lastHash.size === 0
+    const targets = initial ? worktrees : worktrees.filter((wt) => wt.id === this.activeWorktreeId)
+
+    if (targets.length === 0) {
+      this.failures = 0
+      return
+    }
+
+    const results = await Promise.allSettled(targets.map((wt) => this.fetchOne(wt.id)))
     const ok = results.every((r) => r.status === "fulfilled")
     if (ok) {
       this.failures = 0
@@ -164,7 +187,7 @@ export class PRStatusPoller {
     if (!this.options.getWorkspaceRoot()) return
 
     try {
-      const pr = await this.fetchPRForBranch(wt.branch, wt.path)
+      const pr = await this.cachedFetchPR(worktreeId, wt.branch, wt.path)
       if (!pr) {
         const hash = `${worktreeId}:none`
         if (this.lastHash.get(worktreeId) === hash) return
@@ -217,6 +240,15 @@ export class PRStatusPoller {
   private static readonly PR_JSON_FIELDS =
     "number,title,url,state,isDraft,reviewDecision,additions,deletions,changedFiles,headRefName,headRefOid"
 
+  /** Return a cached PR lookup if still fresh, otherwise fetch and cache. */
+  private async cachedFetchPR(worktreeId: string, branch: string, cwd: string): Promise<PRResult | null> {
+    const cached = this.prCache.get(worktreeId)
+    if (cached && Date.now() < cached.expires) return cached.result
+    const result = await this.fetchPRForBranch(branch, cwd)
+    this.prCache.set(worktreeId, { result, expires: Date.now() + PR_LOOKUP_TTL })
+    return result
+  }
+
   private async fetchPRForBranch(branch: string, cwd: string): Promise<PRResult | null> {
     // Strategy 1: bare `gh pr view` — resolves via the branch's tracking ref.
     // Works for fork PRs checked out with `gh pr checkout` (tracking ref = refs/pull/N/head).