fix(agent-status): persist Claude idle-row hydration cleanup

src/main/agent-hooks/server.test.ts

@@ -6093,6 +6093,10 @@ describe('Last-status persistence', () => {
           subagents: undefined
         })
       ])
+      // Why: make the migration one-time; otherwise every launch reparses and
+      // re-prunes the same persisted idle rows.
+      const persisted = JSON.parse(readFileSync(lastStatusPath(), 'utf8'))
+      expect(persisted.entries[PANE].payload.subagents).toBeUndefined()
     } finally {
       server.stop()
     }

src/main/agent-hooks/server.ts

@@ -244,11 +244,10 @@ function sanitizeHydratedEntry(
   } else {
     return null
   }
-  const normalizedPayload = normalizeAgentStatusPayload(record.payload)
-  if (!normalizedPayload) {
+  const payload = normalizeAgentStatusPayload(record.payload)
+  if (!payload) {
     return null
   }
-  const payload = dropHydratedIdleClaudeSubagents(normalizedPayload)
   return {
     paneKey,
     launchToken: typeof record.launchToken === 'string' ? record.launchToken : undefined,
@@ -1777,6 +1776,7 @@ export class AgentHookServer {
     }
     let hydrated = 0
     let dropped = 0
+    let prunedLegacyClaudeSubagents = 0
     // Why: bound disk growth — drop anything older than HYDRATE_MAX_AGE_MS so
     // entries from worktrees archived weeks ago do not pile up forever. Use
     // Date.now() once to keep the cutoff consistent across all entries this
@@ -1790,11 +1790,15 @@ export class AgentHookServer {
           : { ...(rawEntry as Record<string, unknown>), paneKey: resolvedPaneKey }
       const entry = sanitizeHydratedEntry(resolvedPaneKey, rawResolvedEntry)
       if (entry && entry.receivedAt >= ttlCutoff) {
+        const hydratedPayload = dropHydratedIdleClaudeSubagents(entry.payload)
+        if (hydratedPayload !== entry.payload) {
+          prunedLegacyClaudeSubagents +=
+            (entry.payload.subagents?.length ?? 0) - (hydratedPayload.subagents?.length ?? 0)
+          entry.payload = hydratedPayload
+        }
         this.state.lastStatusByPaneKey.set(resolvedPaneKey, entry)
-        // Why: the in-memory subagent roster died with the previous process.
-        // Reseed it from the persisted snapshot so the next teammate-bearing
-        // Stop (whose task ids never match lifecycle ids) doesn't silently
-        // drop the replayed child rows.
+        // Why: preserve only working children across restart. Live activity
+        // confirms them; a later complete inventory may reap stale seeds.
         if (entry.payload.subagents) {
           seedClaudeSubagentRosterFromSnapshots(
             this.state,
@@ -1812,18 +1816,16 @@ export class AgentHookServer {
         `[agent-hooks] last-status hydrate dropped ${dropped} entries (kept ${hydrated})`
       )
     }
-    if (hydrated > 0 && dropped === 0) {
+    if (dropped > 0 || prunedLegacyClaudeSubagents > 0) {
+      // Why: persist load-time pruning once so legacy idle rows do not consume
+      // parse/filter work again on every launch.
+      this.runStatusPersist()
+    } else if (hydrated > 0) {
       // Why: prime from the raw on-disk bytes (not a re-serialization) so the
       // dedup is robust against future shape drift in serializeStatusFile.
       // Only prime when hydration was lossless — if entries were dropped
       // during sanitize, the in-memory map diverges from the on-disk bytes.
       this.lastWrittenJson = raw
-    } else if (dropped > 0) {
-      // Why: clean the stale on-disk file now so a user who has not run an
-      // agent in 8+ days does not re-drop the same entries on every cold
-      // boot. Synchronous variant is safe at start time and avoids
-      // unref'd-timer-during-quit edge cases.
-      this.runStatusPersist()
     }
   }
 

src/shared/agent-hook-listener.ts

@@ -2477,11 +2477,9 @@ export function markClaudeLeadTurnInterrupted(state: HookListenerState, paneKey:
   state.claudeLeadStateByPaneKey.set(paneKey, { state: 'done', interrupted: true })
 }
 
-/** Rebuild a pane's roster from a persisted status snapshot during hydration.
- *  Restart loses the in-memory roster while the renderer replays the child
- *  rows from disk; without reseeding, the next teammate-bearing Stop (whose
- *  task ids never match lifecycle ids) would silently drop those rows. Stale
- *  seeds self-heal: an empty background_tasks clears, activity refreshes. */
+/** Rebuild a pane's working roster from a persisted status snapshot. Live
+ *  activity confirms a seed after restart; a complete task inventory may reap
+ *  an unconfirmed seed whose finish hook arrived while Orca was offline. */
 export function seedClaudeSubagentRosterFromSnapshots(
   state: HookListenerState,
   paneKey: string,

src/shared/claude-subagent-roster.test.ts

@@ -46,7 +46,7 @@ describe('claude-subagent-roster', () => {
     expect(roster.has('aprobe1-6d3cb5b5')).toBe(false)
   })
 
-  it('re-adds a resumed agent as working, keeping its original startedAt', () => {
+  it('re-adds a resumed agent as working with a fresh startedAt', () => {
     const roster: ClaudeSubagentRoster = new Map()
     upsertWorkingClaudeSubagent(roster, 'aprobe1-6d3cb5b5', { agentType: 'probe1' }, 100)
     finishClaudeSubagent(roster, 'aprobe1-6d3cb5b5')