fix(provider-process): kill a provider group its SIGKILLed supervisor leaves behind (#30997)

* fix(provider-process): kill a provider group its SIGKILLed supervisor leaves behind

* fix(provider-process): re-read a provider start in its read second; poll instead of sleeping in tests

* fix(provider-process): keep the report channel through the Codex session adapter; re-prove same-second starts

* fix(provider-process): never kill a leaderless provider group; list guard files in the CLI project

config/tsconfig.cli.json

@@ -181,6 +181,8 @@
     "../src/main/provider-process/provider-process-exit-deadline.ts",
     "../src/main/provider-process/managed-provider-process.ts",
     "../src/main/provider-process/provider-executable-missing.ts",
+    "../src/main/provider-process/provider-group-orphan-guard.ts",
+    "../src/main/provider-process/provider-process-row-read.ts",
     "../src/main/provider-process/provider-process-close.ts",
     "../src/main/provider-process/provider-process-launch.ts",
     "../src/main/provider-process/provider-process-supervisor.ts",

src/main/acp/acp-agent-connection.test.ts

@@ -81,13 +81,17 @@ describe('ACP process-owning connection', () => {
       expect.objectContaining({
         cwd: start.cwd,
         env: expect.objectContaining({ ACP_ACCOUNT_HOME: '/host/account' }),
-        stdio: ['pipe', 'pipe', 'pipe']
+        stdio: ['pipe', 'pipe', 'pipe', ...(process.platform === 'win32' ? [] : ['ipc'])]
       })
     )
     expect(spawn.mock.calls[0][0].env).not.toHaveProperty('STRIPPED')
-    expect(await connection.start(start)).toMatchObject({ sessionId: 'session-1' })
+    expect(await connection.start(start)).toMatchObject({
+      sessionId: 'session-1'
+    })
     agent.on('session/prompt', (frame) => agent.reply(frame, { stopReason: 'end_turn' }))
-    expect(await connection.prompt([...prompt])).toEqual({ stopReason: 'end_turn' })
+    expect(await connection.prompt([...prompt])).toEqual({
+      stopReason: 'end_turn'
+    })
     connection.pauseReading()
     expect(child.stdout.isPaused()).toBe(true)
     connection.resumeReading()
@@ -161,7 +165,9 @@ describe('ACP process-owning connection', () => {
 
   it('drops an overlong stderr line it could not classify instead of keeping its prefix', async () => {
     const { connection, child } = fixture({
-      signIn: { read: (line) => (line.endsWith('authenticate') ? 'link' : undefined) }
+      signIn: {
+        read: (line) => (line.endsWith('authenticate') ? 'link' : undefined)
+      }
     })
     await connection.start(start)
     const link = `https://accounts.example/auth?state=${'x'.repeat(33_000)}`
@@ -282,7 +288,11 @@ describe('ACP process-owning connection', () => {
     connection.onExit(failing)
     connection.onExit(early)
     expect(() => child.emit('exit', 0, null)).not.toThrow()
-    expect(early).toHaveBeenCalledExactlyOnceWith({ code: 0, signal: null, processless: false })
+    expect(early).toHaveBeenCalledExactlyOnceWith({
+      code: 0,
+      signal: null,
+      processless: false
+    })
     expect(() => connection.onExit(failing)).not.toThrow()
     const late = vi.fn()
     connection.onExit(late)

src/main/codex/codex-app-server-client.test.ts

@@ -18,6 +18,11 @@ import {
 } from '../provider-process/provider-process-supervisor'
 import { runCodexAppServerSession } from './codex-app-server-session'
 
+// The orphan guard's process reads add timers of their own; the timer audit below is the session's.
+vi.mock('../provider-process/provider-group-orphan-guard', () => ({
+  guardSupervisedProviderGroup: () => false
+}))
+
 // Stub codex app-server speaking the same JSONL protocol: initialize →
 // initialized → hooks/list → config/batchWrite → hooks/list. Scenario-driven
 // via STUB_CONFIG so each test controls listings, errors, and hangs.

src/main/codex/codex-app-server-session.ts

@@ -122,10 +122,11 @@ export async function runCodexAppServerSession<T>(
       site: CODEX_APP_SERVER_KILL_SITE,
       inheritedEnv: pairedEnv,
       spawnImpl: (spec) =>
-        // oxlint-disable-next-line typescript/consistent-type-assertions -- SAFETY: every stdio slot is 'pipe', so stdin, stdout and stderr exist.
+        // oxlint-disable-next-line typescript/consistent-type-assertions -- SAFETY: the managed spec's first three stdio slots are 'pipe', so stdin, stdout and stderr exist.
         spawnImpl(spec.program, [...(spec.args ?? [])], {
           env: spec.env,
-          stdio: ['pipe', 'pipe', 'pipe'],
+          // The managed spec's slots, so the supervisor's group-report channel survives.
+          stdio: spec.stdio ?? ['pipe', 'pipe', 'pipe'],
           windowsHide: true,
           ...(spec.detached ? { detached: true } : {})
         }) as ChildProcessWithoutNullStreams

src/main/codex/codex-state-db-backfill-recovery-process.ts

@@ -10,6 +10,7 @@ import {
 } from '../provider-process/provider-process-supervisor'
 import type { CodexAppServerSpawn } from './codex-app-server-process-tree-kill'
 import { stopSupervisedChildProcess } from '../provider-process/supervised-child-process-stop'
+import { guardSupervisedProviderGroup } from '../provider-process/provider-group-orphan-guard'
 import { resolveWslInteropSpawnCwd } from '../wsl-interop-spawn-directory'
 
 const BACKFILL_RECOVERY_KILL_SITE = 'codex-state-db-backfill-recovery'
@@ -17,7 +18,10 @@ const BACKFILL_RECOVERY_KILL_SITE = 'codex-state-db-backfill-recovery'
 // for the spawn spec (a gone Orca) and the stop (Orca stopping it), so they cannot drift.
 const BACKFILL_RECOVERY_CLOSE_REQUEST: ProviderCloseRequest = 'stdin-end'
 
-export type CodexBackfillRecoveryProcess = { child: ChildProcessHandle; supervised: boolean }
+export type CodexBackfillRecoveryProcess = {
+  child: ChildProcessHandle
+  supervised: boolean
+}
 
 /** Starts the read-only app-server whose presence lets Codex finish its own backfill. */
 export function spawnCodexBackfillRecoveryProcess(
@@ -47,17 +51,24 @@ export function spawnCodexBackfillRecoveryProcess(
   // Kept alive up to an hour; on POSIX the supervisor stops it with Orca, however Orca exits.
   const spawnSpec = createProviderSpawnSpec(
     { command, args: [...CODEX_READ_ONLY_APP_SERVER_ARGS], cwd: codexHomePath },
-    withCliRuntimeOnPath(command, { ...process.env, CODEX_HOME: codexHomePath }),
+    withCliRuntimeOnPath(command, {
+      ...process.env,
+      CODEX_HOME: codexHomePath
+    }),
     process.platform,
     { lifetime: 'session', closeRequest: BACKFILL_RECOVERY_CLOSE_REQUEST }
   )
   const child = spawnProcess(spawnSpec.program, spawnSpec.args, {
     cwd: codexHomePath,
-    stdio: ['pipe', 'ignore', 'ignore'],
+    stdio: ['pipe', 'ignore', 'ignore', ...spawnSpec.groupReportStdio],
     windowsHide: true,
     env: spawnSpec.env,
     ...(spawnSpec.supervised ? { detached: true } : {})
   })
+  guardSupervisedProviderGroup(child, {
+    site: BACKFILL_RECOVERY_KILL_SITE,
+    supervised: spawnSpec.supervised
+  })
   return { child, supervised: spawnSpec.supervised }
 }
 

src/main/jsonl-rpc/agent-connection.test.ts

@@ -64,7 +64,7 @@ describe('JSON-lines RPC process ownership', () => {
       expect.objectContaining({
         cwd: '/execution-host/folder',
         env: expect.objectContaining({ PI_CODING_AGENT_DIR: '/host/account' }),
-        stdio: ['pipe', 'pipe', 'pipe']
+        stdio: ['pipe', 'pipe', 'pipe', ...(process.platform === 'win32' ? [] : ['ipc'])]
       })
     )
     expect(spawn.mock.calls[0][0].env).not.toHaveProperty('STRIPPED')
@@ -173,7 +173,10 @@ describe('JSON-lines RPC process ownership', () => {
     await close
     child.emit('exit', null, 'SIGKILL')
     await connection.close()
-    expect(connection.lastCloseResult).toEqual({ root: 'live', tree: 'unverifiable' })
+    expect(connection.lastCloseResult).toEqual({
+      root: 'live',
+      tree: 'unverifiable'
+    })
     expect(connection.rootVerdict).toBe('exited')
   })
 

src/main/opencode/opencode-launch-model-context.ts

@@ -13,6 +13,7 @@ import { cancelUnreadResponseBody } from '../lib/unread-response-body'
 import type { ChildProcessHandle } from '@orca/process-host/process-spec'
 import { createProviderSpawnSpec } from '../provider-process/provider-process-supervisor'
 import { stopSupervisedChildProcess } from '../provider-process/supervised-child-process-stop'
+import { guardSupervisedProviderGroup } from '../provider-process/provider-group-orphan-guard'
 
 const OPENCODE_PREFLIGHT_KILL_SITE = 'opencode-launch-model-preflight'
 // A supervisor's pipes are its own, so they close as it exits.
@@ -29,7 +30,10 @@ const agent = z.object({
 const location = z.object({ directory: z.string() })
 const modelResponse = z.object({ location, data: availableModel.array() })
 const agentResponse = z.object({ location, data: agent.array() })
-const defaultModelResponse = z.object({ location, data: availableModel.nullable() })
+const defaultModelResponse = z.object({
+  location,
+  data: availableModel.nullable()
+})
 const configResponse = z.array(
   z.object({
     type: z.enum(['directory', 'document']),
@@ -107,7 +111,11 @@ export async function probeOpenCodeLaunchModelContext(options: {
       args: ['serve', '--hostname', '127.0.0.1', '--port', '0'],
       cwd: options.cwd
     },
-    { ...options.env, OPENCODE_SERVER_USERNAME: 'opencode', OPENCODE_SERVER_PASSWORD: password },
+    {
+      ...options.env,
+      OPENCODE_SERVER_USERNAME: 'opencode',
+      OPENCODE_SERVER_PASSWORD: password
+    },
     process.platform,
     // `opencode serve` never exits on stdin end, so on POSIX only its supervisor stops it with Orca.
     { lifetime: 'one-shot' }
@@ -117,7 +125,12 @@ export async function probeOpenCodeLaunchModelContext(options: {
     args: spawnSpec.args,
     cwd: options.cwd,
     env: spawnSpec.env,
-    detached: true
+    detached: true,
+    stdio: ['pipe', 'pipe', 'pipe', ...spawnSpec.groupReportStdio]
+  })
+  guardSupervisedProviderGroup(child, {
+    site: OPENCODE_PREFLIGHT_KILL_SITE,
+    supervised: spawnSpec.supervised
   })
   const output = createOutputSink(4096)
   let childClosed = false

src/main/provider-process/managed-provider-process.ts

@@ -14,6 +14,7 @@ import {
 import type { DescendantTreeVerdict } from '../pty-descendant-exit-verification'
 import { isMissingProviderExecutable } from './provider-executable-missing'
 import { supervisedProviderSpawnFailure } from './provider-spawn-failure-report'
+import { guardSupervisedProviderGroup } from './provider-group-orphan-guard'
 import {
   acceptProviderRootExit,
   closeProviderProcess,
@@ -96,8 +97,9 @@ export function spawnManagedProviderProcess(
     cwd: spec.cwd,
     env: spec.env,
     detached: spec.detached,
-    stdio: ['pipe', 'pipe', 'pipe']
+    stdio: ['pipe', 'pipe', 'pipe', ...spec.groupReportStdio]
   })
+  guardSupervisedProviderGroup(child, { site: options.site, supervised: spec.supervised })
   const listeners = new Set<(exit: ProviderProcessExit) => void>()
   let observed: ProviderProcessExit | null = null
   let spawnFailed = false
@@ -199,7 +201,10 @@ export function spawnManagedProviderProcess(
       return exitProof.run(async () => {
         // The one already-exited guard. It observed nothing, so an earlier close's findings stay.
         if (observed && !tree) {
-          const result: ProviderProcessCloseResult = { root: 'exited', tree: null }
+          const result: ProviderProcessCloseResult = {
+            root: 'exited',
+            tree: null
+          }
           lastCloseResult ??= result
           return result
         }

src/main/provider-process/provider-group-guard-launch-paths.test.ts

@@ -0,0 +1,99 @@
+import { chmodSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+import type { ChildProcess } from 'node:child_process'
+import { afterAll, beforeEach, describe, expect, it, vi } from 'vitest'
+import type * as OrphanGuard from './provider-group-orphan-guard'
+import { spawnManagedProviderProcess } from './managed-provider-process'
+import { runCodexAppServerSession } from '../codex/codex-app-server-session'
+import { spawnCodexAppServerProcess } from '../codex/codex-app-server-process-tree-kill'
+import { spawnCodexBackfillRecoveryProcess } from '../codex/codex-state-db-backfill-recovery-process'
+import { probeOpenCodeLaunchModelContext } from '../opencode/opencode-launch-model-context'
+import { spawnSourceControlAgent } from '../text-generation/source-control-agent-launch'
+
+// Every production launch path must hand the supervisor its report channel; an adapter that
+// rewrites stdio drops it, and the guard then never watches the group.
+
+const { registrations } = vi.hoisted(() => {
+  const recorded: { site: string; armed: boolean }[] = []
+  return { registrations: recorded }
+})
+
+vi.mock('./provider-group-orphan-guard', async (importOriginal) => {
+  const actual = await importOriginal<typeof OrphanGuard>()
+  return {
+    ...actual,
+    guardSupervisedProviderGroup: (
+      ...args: Parameters<typeof actual.guardSupervisedProviderGroup>
+    ) => {
+      const armed = actual.guardSupervisedProviderGroup(...args)
+      registrations.push({ site: args[1].site, armed })
+      return armed
+    }
+  }
+})
+
+const dir = mkdtempSync(join(tmpdir(), 'orca-guard-launch-paths-'))
+afterAll(() => rmSync(dir, { recursive: true, force: true }))
+// Exits at once, so each supervisor relays its exit and stops on its own.
+const QUICK_EXIT = ['-e', 'setTimeout(() => {}, 50)']
+
+function exited(child: Pick<ChildProcess, 'exitCode' | 'signalCode' | 'once'>): Promise<void> {
+  return child.exitCode !== null || child.signalCode !== null
+    ? Promise.resolve()
+    : new Promise((resolve) => child.once('exit', () => resolve()))
+}
+
+beforeEach(() => {
+  registrations.length = 0
+})
+
+describe.runIf(process.platform !== 'win32')('provider group guard on every launch path', () => {
+  it('managed provider processes (Codex, Claude, ACP, JSON-lines, model catalog)', async () => {
+    const managed = spawnManagedProviderProcess(
+      { command: process.execPath, args: QUICK_EXIT },
+      { site: 'managed' }
+    )
+    expect(registrations).toEqual([{ site: 'managed', armed: true }])
+    await managed.close()
+  })
+
+  it('Codex app-server sessions, whose adapter re-maps the spawn options', async () => {
+    await runCodexAppServerSession(
+      { command: process.execPath, cliPath: null, args: QUICK_EXIT, timeoutMs: 2_000 },
+      async () => null
+    ).catch(() => null)
+    expect(registrations).toEqual([{ site: expect.any(String), armed: true }])
+  })
+
+  it('the Codex backfill recovery app-server', async () => {
+    const { child } = spawnCodexBackfillRecoveryProcess(
+      dir,
+      spawnCodexAppServerProcess,
+      () => process.execPath
+    )
+    expect(registrations).toEqual([{ site: 'codex-state-db-backfill-recovery', armed: true }])
+    await exited(child)
+  })
+
+  it('the OpenCode launch-context probe', async () => {
+    const standIn = join(dir, 'opencode')
+    writeFileSync(standIn, '#!/bin/sh\nexit 0\n')
+    chmodSync(standIn, 0o755)
+    await probeOpenCodeLaunchModelContext({ executable: standIn, cwd: dir, env: process.env })
+    expect(registrations).toEqual([{ site: 'opencode-launch-model-preflight', armed: true }])
+  })
+
+  it('source-control text generation', async () => {
+    const child = spawnSourceControlAgent({
+      binary: process.execPath,
+      args: QUICK_EXIT,
+      env: process.env,
+      stdinMode: 'pipe',
+      useCwdForNative: false
+    })
+    expect(registrations).toEqual([{ site: 'source-control-text-generation', armed: true }])
+    child.stdin?.end()
+    await exited(child)
+  })
+})

src/main/provider-process/provider-group-orphan-guard.integration.test.ts

@@ -0,0 +1,136 @@
+import { afterEach, describe, expect, it } from 'vitest'
+import { spawnProcess } from '@orca/process-host'
+import type { ChildProcess } from 'node:child_process'
+import { createProviderSpawnSpec } from './provider-process-supervisor'
+import {
+  guardSupervisedProviderGroup,
+  type ProviderGroupSweepVerdict
+} from './provider-group-orphan-guard'
+
+// Leads its own group with a SIGTERM-ignoring grandchild, as a server with helpers would.
+const PROVIDER = String.raw`
+  const { spawn } = require('node:child_process')
+  process.stdout.on('error', () => {})
+  const grandchild = spawn(
+    process.execPath,
+    ['-e', "process.on('SIGTERM', () => {}); process.stdout.write('armed'); setInterval(() => {}, 60000)"],
+    { stdio: ['ignore', 'pipe', 'ignore'] }
+  )
+  grandchild.stdout.once('data', () =>
+    process.stdout.write(JSON.stringify({ provider: process.pid, grandchild: grandchild.pid }) + '\n')
+  )
+  setInterval(() => {}, 60000)
+`
+
+const IDENTITY_TIMEOUT_MS = 15_000
+const recordedPids = new Set<number>()
+
+function alive(pid: number): boolean {
+  try {
+    process.kill(pid, 0)
+    return true
+  } catch (error) {
+    return !(error instanceof Error && 'code' in error && error.code === 'ESRCH')
+  }
+}
+
+async function waitFor(predicate: () => boolean, timeoutMs: number): Promise<boolean> {
+  const deadline = Date.now() + timeoutMs
+  while (!predicate()) {
+    if (Date.now() >= deadline) {
+      return false
+    }
+    await new Promise((resolve) => setTimeout(resolve, 25))
+  }
+  return true
+}
+
+function launch(): {
+  supervisor: ChildProcess
+  proven: () => boolean
+  swept: Promise<ProviderGroupSweepVerdict | 'unproven'>
+  pids: Promise<{ provider: number; grandchild: number }>
+} {
+  const spec = createProviderSpawnSpec(
+    { command: process.execPath, args: ['-e', PROVIDER] },
+    process.env,
+    process.platform,
+    { lifetime: 'session' }
+  )
+  const supervisor = spawnProcess({
+    program: spec.program,
+    args: spec.args,
+    env: spec.env,
+    cwd: spec.cwd,
+    detached: spec.detached,
+    stdio: ['pipe', 'pipe', 'ignore', ...spec.groupReportStdio]
+  })
+  recordedPids.add(supervisor.pid!)
+  let identityProven = false
+  const swept = new Promise<ProviderGroupSweepVerdict | 'unproven'>((resolve) =>
+    guardSupervisedProviderGroup(supervisor, {
+      site: 'test',
+      supervised: spec.supervised,
+      onProven: () => {
+        identityProven = true
+      },
+      onSwept: resolve
+    })
+  )
+  const pids = new Promise<{ provider: number; grandchild: number }>((resolve) => {
+    let buffered = ''
+    supervisor.stdout?.on('data', (chunk: Buffer) => {
+      buffered += chunk.toString()
+      const line = buffered.split('\n')[0]
+      if (buffered.includes('\n')) {
+        const parsed = JSON.parse(line)
+        recordedPids.add(parsed.provider)
+        recordedPids.add(parsed.grandchild)
+        resolve(parsed)
+      }
+    })
+  })
+  return { supervisor, proven: () => identityProven, swept, pids }
+}
+
+afterEach(() => {
+  for (const pid of recordedPids) {
+    if (alive(pid)) {
+      process.kill(pid, 'SIGKILL')
+    }
+  }
+  recordedPids.clear()
+})
+
+describe.runIf(process.platform !== 'win32')('provider group orphan guard', () => {
+  it('kills the provider group a SIGKILLed supervisor leaves behind', async () => {
+    const { supervisor, proven, swept, pids } = launch()
+    const { provider, grandchild } = await pids
+    // A start in the read's own second is re-read after it, so this can take over a second.
+    expect(await waitFor(proven, IDENTITY_TIMEOUT_MS)).toBe(true)
+
+    supervisor.kill('SIGKILL')
+
+    await expect(swept).resolves.toBe('killed')
+    expect(await waitFor(() => !alive(provider) && !alive(grandchild), 10_000)).toBe(true)
+  })
+
+  it('leaves nothing to sweep when the supervisor stops its own group', async () => {
+    const { supervisor, proven, swept, pids } = launch()
+    await pids
+    expect(await waitFor(proven, IDENTITY_TIMEOUT_MS)).toBe(true)
+
+    supervisor.kill('SIGTERM')
+
+    await expect(swept).resolves.toBe('absent')
+  })
+
+  it('closes the report channel while the provider still runs', async () => {
+    const { supervisor, pids } = launch()
+    const disconnected = new Promise<void>((resolve) => supervisor.once('disconnect', resolve))
+    const { provider } = await pids
+
+    await disconnected
+    expect(alive(provider)).toBe(true)
+  })
+})

src/main/provider-process/provider-group-orphan-guard.test.ts

@@ -0,0 +1,270 @@
+import { EventEmitter } from 'node:events'
+import { describe, expect, it, vi } from 'vitest'
+import {
+  guardSupervisedProviderGroup,
+  sweepOrphanedProviderGroup,
+  type ProcessRowRead,
+  type ProviderGroupSweepVerdict
+} from './provider-group-orphan-guard'
+
+const STARTED = 'Mon Oct 19 10:00:00 2026'
+const identity = { pgid: 4242, startedAt: STARTED }
+
+function errno(code: string): Error {
+  return Object.assign(new Error(code), { code })
+}
+
+function row(overrides: Partial<Exclude<ProcessRowRead, null | 'unverifiable'>> = {}) {
+  return {
+    pid: 4242,
+    ppid: 1,
+    pgid: 4242,
+    startedAt: STARTED,
+    startResolution: 'second' as const,
+    ...overrides
+  }
+}
+
+describe('sweepOrphanedProviderGroup', () => {
+  it('does nothing when the group is already gone', async () => {
+    const signalGroup = vi.fn(() => {
+      throw errno('ESRCH')
+    })
+    const readProcessRow = vi.fn()
+    await expect(
+      sweepOrphanedProviderGroup(identity, {
+        site: 'test',
+        supervised: true,
+        signalGroup,
+        readProcessRow
+      })
+    ).resolves.toBe('absent')
+    expect(readProcessRow).not.toHaveBeenCalled()
+  })
+
+  it('kills a surviving group whose leader is the recorded process', async () => {
+    const signalGroup = vi.fn()
+    await expect(
+      sweepOrphanedProviderGroup(identity, {
+        site: 'test',
+        supervised: true,
+        signalGroup,
+        readProcessRow: async () => row()
+      })
+    ).resolves.toBe('killed')
+    expect(signalGroup).toHaveBeenLastCalledWith(4242, 'SIGKILL')
+  })
+
+  it('refuses a leaderless group, which nothing can prove is still ours', async () => {
+    const signalGroup = vi.fn()
+    await expect(
+      sweepOrphanedProviderGroup(identity, {
+        site: 'test',
+        supervised: true,
+        signalGroup,
+        readProcessRow: async () => null
+      })
+    ).resolves.toBe('refused')
+    expect(signalGroup).not.toHaveBeenCalledWith(4242, 'SIGKILL')
+  })
+
+  it.each<[string, ProcessRowRead]>([
+    ['a reused pid with another start time', row({ startedAt: 'Mon Oct 19 11:00:00 2026' })],
+    ['a reused pid leading another group', row({ pgid: 99 })],
+    ['a failed process read', 'unverifiable']
+  ])('refuses %s', async (_label, leader) => {
+    const signalGroup = vi.fn()
+    await expect(
+      sweepOrphanedProviderGroup(identity, {
+        site: 'test',
+        supervised: true,
+        signalGroup,
+        readProcessRow: async () => leader
+      })
+    ).resolves.toBe('refused')
+    expect(signalGroup).not.toHaveBeenCalledWith(4242, 'SIGKILL')
+  })
+
+  it('refuses a group this user cannot signal', async () => {
+    const signalGroup = vi.fn(() => {
+      throw errno('EPERM')
+    })
+    await expect(
+      sweepOrphanedProviderGroup(identity, { site: 'test', supervised: true, signalGroup })
+    ).resolves.toBe('refused')
+  })
+})
+
+class FakeSupervisor extends EventEmitter {
+  pid = 777
+  connected = true
+  report(message: unknown): void {
+    this.emit('message', message)
+    this.connected = false
+    this.emit('disconnect')
+  }
+}
+
+function guard(
+  child: FakeSupervisor,
+  readProcessRow: (pid: number) => Promise<ProcessRowRead>,
+  signalGroup = vi.fn()
+): Promise<ProviderGroupSweepVerdict | 'unproven'> {
+  return new Promise((resolve) =>
+    guardSupervisedProviderGroup(child, {
+      site: 'test',
+      supervised: true,
+      readProcessRow,
+      signalGroup,
+      onSwept: resolve
+    })
+  )
+}
+
+describe('guardSupervisedProviderGroup', () => {
+  it('sweeps the reported group once the supervisor exits', async () => {
+    const child = new FakeSupervisor()
+    const signalGroup = vi.fn()
+    const swept = guard(child, async () => row({ ppid: 777 }), signalGroup)
+    child.report({ pgid: 4242 })
+    child.emit('exit', null, 'SIGKILL')
+    await expect(swept).resolves.toBe('killed')
+    expect(signalGroup).toHaveBeenLastCalledWith(4242, 'SIGKILL')
+  })
+
+  it('never records a reported pid that its supervisor no longer parents', async () => {
+    const child = new FakeSupervisor()
+    const signalGroup = vi.fn()
+    const swept = guard(child, async () => row({ ppid: 1 }), signalGroup)
+    child.report({ pgid: 4242 })
+    child.emit('exit', null, 'SIGKILL')
+    await expect(swept).resolves.toBe('unproven')
+    expect(signalGroup).not.toHaveBeenCalled()
+  })
+
+  it('re-reads a leader started in the read second, so a same-second reuse cannot match', async () => {
+    const child = new FakeSupervisor()
+    const startedAtMs = Date.parse(STARTED)
+    const clock = { now: startedAtMs + 300 }
+    const readProcessRow = vi.fn(async () => row({ ppid: 777 }))
+    const sleep = vi.fn(async (ms: number) => {
+      clock.now += ms
+    })
+    const proven = new Promise((resolve) =>
+      guardSupervisedProviderGroup(child, {
+        site: 'test',
+        supervised: true,
+        readProcessRow,
+        now: () => clock.now,
+        sleep,
+        onProven: resolve
+      })
+    )
+    child.report({ pgid: 4242 })
+    await expect(proven).resolves.toEqual(identity)
+    expect(readProcessRow).toHaveBeenCalledTimes(2)
+    expect(clock.now).toBeGreaterThanOrEqual(startedAtMs + 1_000)
+  })
+
+  it('arms nothing when a same-second replacement holds the pid once the second ends', async () => {
+    const child = new FakeSupervisor()
+    const clock = { now: Date.parse(STARTED) + 300 }
+    // The first read is our leader; by the re-read its pid names a stranger with the same lstart.
+    const readProcessRow = vi
+      .fn<(pid: number) => Promise<ProcessRowRead>>()
+      .mockResolvedValueOnce(row({ ppid: 777 }))
+      .mockResolvedValueOnce(row({ ppid: 1 }))
+    const signalGroup = vi.fn()
+    const swept = new Promise((resolve) =>
+      guardSupervisedProviderGroup(child, {
+        site: 'test',
+        supervised: true,
+        readProcessRow,
+        signalGroup,
+        now: () => clock.now,
+        sleep: async (ms) => {
+          clock.now += ms
+        },
+        onSwept: resolve
+      })
+    )
+    child.report({ pgid: 4242 })
+    await vi.waitFor(() => expect(readProcessRow).toHaveBeenCalledTimes(2))
+    child.emit('exit', null, 'SIGKILL')
+    await expect(swept).resolves.toBe('unproven')
+    expect(signalGroup).not.toHaveBeenCalled()
+  })
+
+  it('trusts a Linux clock-tick start without waiting out the second', async () => {
+    const child = new FakeSupervisor()
+    const startedAt = 'clock-tick:123456'
+    const readProcessRow = vi.fn(async () =>
+      row({ ppid: 777, startedAt, startResolution: 'clock-tick' })
+    )
+    const sleep = vi.fn(async () => {})
+    const proven = new Promise((resolve) =>
+      guardSupervisedProviderGroup(child, {
+        site: 'test',
+        supervised: true,
+        readProcessRow,
+        now: () => Date.now(),
+        sleep,
+        onProven: resolve
+      })
+    )
+    child.report({ pgid: 4242 })
+    await expect(proven).resolves.toEqual({ pgid: 4242, startedAt })
+    expect(sleep).not.toHaveBeenCalled()
+  })
+
+  it('records nothing when every read stays inside the start second', async () => {
+    const child = new FakeSupervisor()
+    const now = Date.parse(STARTED) + 100
+    const signalGroup = vi.fn()
+    const swept = new Promise((resolve) =>
+      guardSupervisedProviderGroup(child, {
+        site: 'test',
+        supervised: true,
+        readProcessRow: async () => row({ ppid: 777 }),
+        signalGroup,
+        now: () => now,
+        sleep: async () => {},
+        onSwept: resolve
+      })
+    )
+    child.report({ pgid: 4242 })
+    child.emit('exit', null, 'SIGKILL')
+    await expect(swept).resolves.toBe('unproven')
+    expect(signalGroup).not.toHaveBeenCalled()
+  })
+
+  it('warns once and watches nothing for a supervised spawn without the report channel', () => {
+    const warn = vi.spyOn(console, 'warn').mockImplementation(() => {})
+    const child = new FakeSupervisor()
+    child.connected = false
+    expect(guardSupervisedProviderGroup(child, { site: 'test', supervised: true })).toBe(false)
+    expect(guardSupervisedProviderGroup(child, { site: 'test', supervised: true })).toBe(false)
+    expect(child.listenerCount('exit')).toBe(0)
+    expect(warn).toHaveBeenCalledTimes(1)
+    warn.mockRestore()
+  })
+
+  it('stays quiet for an unsupervised spawn', () => {
+    const warn = vi.spyOn(console, 'warn').mockImplementation(() => {})
+    expect(
+      guardSupervisedProviderGroup(new FakeSupervisor(), { site: 'test', supervised: false })
+    ).toBe(false)
+    expect(warn).not.toHaveBeenCalled()
+    warn.mockRestore()
+  })
+
+  it('ignores a malformed report', async () => {
+    const child = new FakeSupervisor()
+    const readProcessRow = vi.fn(async () => row({ ppid: 777 }))
+    const swept = guard(child, readProcessRow)
+    child.report({ pgid: 1 })
+    child.emit('exit', 0, null)
+    await expect(swept).resolves.toBe('unproven')
+    expect(readProcessRow).not.toHaveBeenCalled()
+  })
+})

src/main/provider-process/provider-group-orphan-guard.ts

@@ -0,0 +1,224 @@
+import { hasUnambiguousStartTime } from '../pty-descendant-termination'
+import { recordSelfInitiatedTreeKill } from '../crash-reporting/self-initiated-tree-kill-log'
+import { readPosixProcessRow, type ProcessRowRead } from './provider-process-row-read'
+
+export type { ProcessRowRead } from './provider-process-row-read'
+
+// Covers the one stop the supervisor cannot make itself: its own SIGKILL leaves the provider's
+// detached group running, so Orca, which outlives it, finishes the stop. POSIX only: Windows
+// providers have no supervisor.
+
+// The supervisor disconnects right after reporting, so this bounds only a wedged channel.
+const REPORT_END_WAIT_MS = 1_000
+const IDENTITY_READ_ATTEMPTS = 3
+const SECOND_BOUNDARY_MARGIN_MS = 50
+
+/** A provider group proven ours while its supervisor still parented the leader. */
+export type ProviderGroupIdentity = { pgid: number; startedAt: string }
+
+export type ProviderGroupSweepVerdict = 'absent' | 'killed' | 'refused'
+
+export type ProviderGroupGuardDeps = {
+  site: string
+  /** Only a supervised POSIX spawn has a group to guard. */
+  supervised: boolean
+  readProcessRow?: (pid: number) => Promise<ProcessRowRead>
+  signalGroup?: (pgid: number, signal: NodeJS.Signals | 0) => void
+  now?: () => number
+  sleep?: (ms: number) => Promise<void>
+  onProven?: (identity: ProviderGroupIdentity) => void
+  /** `unproven`: no group was reported and proven ours, so nothing was swept. */
+  onSwept?: (verdict: ProviderGroupSweepVerdict | 'unproven') => void
+}
+
+let warnedMissingChannel = false
+
+type GuardedChild = {
+  readonly pid?: number
+  readonly connected: boolean
+  once(event: 'exit' | 'disconnect', listener: () => void): unknown
+  once(event: 'message', listener: (message: unknown) => void): unknown
+}
+
+function defaultSignalGroup(pgid: number, signal: NodeJS.Signals | 0): void {
+  process.kill(-pgid, signal)
+}
+
+/** SIGKILLs a surviving provider group, never one whose leader pid now names another process. */
+export async function sweepOrphanedProviderGroup(
+  identity: ProviderGroupIdentity,
+  deps: ProviderGroupGuardDeps
+): Promise<ProviderGroupSweepVerdict> {
+  const signalGroup = deps.signalGroup ?? defaultSignalGroup
+  try {
+    signalGroup(identity.pgid, 0)
+  } catch (error) {
+    // EPERM: the group exists but is not this user's, so it is not ours.
+    return isErrnoCode(error, 'ESRCH') ? 'absent' : 'refused'
+  }
+  const leader = await (deps.readProcessRow ?? readPosixProcessRow)(identity.pgid)
+  // Only a leader that re-proves its recorded identity names our group; a leaderless group has
+  // nothing to prove it by, so its survivors are left alone. The recorded start predates its
+  // read's second, so a later reuse cannot share it.
+  if (
+    !leader ||
+    leader === 'unverifiable' ||
+    leader.pgid !== identity.pgid ||
+    leader.startedAt !== identity.startedAt
+  ) {
+    return 'refused'
+  }
+  try {
+    signalGroup(identity.pgid, 'SIGKILL')
+  } catch (error) {
+    return isErrnoCode(error, 'ESRCH') ? 'absent' : 'refused'
+  }
+  recordSelfInitiatedTreeKill({
+    pid: identity.pgid,
+    site: deps.site,
+    scope: 'posix-process-group'
+  })
+  return 'killed'
+}
+
+/**
+ * Reads the group a supervisor reports on its IPC channel, proves it by the leader's parent and
+ * start time, and sweeps that group once the supervisor exits. True when it is watching.
+ */
+export function guardSupervisedProviderGroup(
+  child: GuardedChild,
+  deps: ProviderGroupGuardDeps
+): boolean {
+  const supervisorPid = child.pid
+  if (!supervisorPid || !deps.supervised) {
+    return false
+  }
+  if (!hasReportChannel(child)) {
+    // A launch adapter that rewrites stdio drops the channel; say so once instead of silently.
+    if (!warnedMissingChannel) {
+      warnedMissingChannel = true
+      console.warn(`[provider-group-guard] ${deps.site}: supervised spawn has no report channel`)
+    }
+    return false
+  }
+  let identity: Promise<ProviderGroupIdentity | null> = Promise.resolve(null)
+  // A gone supervisor parents nothing, so a re-read it would wait for can only fail; skip it.
+  const proof = createProofWait(deps.sleep)
+  let channelClosed = false
+  let onChannelClosed = (): void => {}
+  child.once('disconnect', () => {
+    channelClosed = true
+    onChannelClosed()
+  })
+  child.once('message', (message: unknown) => {
+    const pgid = reportedGroup(message)
+    if (pgid !== null) {
+      identity = proveReportedGroup(pgid, supervisorPid, deps, proof)
+      void identity.then((proven) => proven && deps.onProven?.(proven))
+    }
+  })
+  const untilChannelClosed = (): Promise<void> =>
+    channelClosed
+      ? Promise.resolve()
+      : new Promise((resolve) => {
+          const timer = setTimeout(resolve, REPORT_END_WAIT_MS)
+          timer.unref?.()
+          onChannelClosed = () => {
+            clearTimeout(timer)
+            resolve()
+          }
+        })
+  child.once('exit', () => {
+    proof.stop()
+    // Exit can be observed before the report is delivered.
+    void untilChannelClosed()
+      .then(() => identity)
+      .then((proven) => (proven ? sweepOrphanedProviderGroup(proven, deps) : ('unproven' as const)))
+      .then(deps.onSwept, () => deps.onSwept?.('refused'))
+  })
+  return true
+}
+
+/**
+ * The leader's identity, read while this supervisor still parents it (so its pid was not reused).
+ * lstart has one-second resolution, so a start in the read's own second is re-proven after that
+ * second ends; otherwise a pid reused within that second would carry the same start time. A
+ * supervisor gone by then no longer parents it, so nothing is armed.
+ */
+async function proveReportedGroup(
+  pgid: number,
+  supervisorPid: number,
+  deps: ProviderGroupGuardDeps,
+  proof: ProofWait
+): Promise<ProviderGroupIdentity | null> {
+  const readRow = deps.readProcessRow ?? readPosixProcessRow
+  const now = deps.now ?? Date.now
+  for (let attempt = 0; attempt < IDENTITY_READ_ATTEMPTS; attempt += 1) {
+    // The boundary before ps starts, as the descendant sweep takes it.
+    const capturedAtMs = now()
+    const row = await readRow(pgid)
+    if (!row || row === 'unverifiable' || row.ppid !== supervisorPid || row.pgid !== pgid) {
+      return null
+    }
+    if (
+      row.startResolution === 'clock-tick' ||
+      hasUnambiguousStartTime(row.startedAt, capturedAtMs)
+    ) {
+      return { pgid, startedAt: row.startedAt }
+    }
+    if (!(await proof.wait(1_000 - (now() % 1_000) + SECOND_BOUNDARY_MARGIN_MS))) {
+      return null
+    }
+  }
+  return null
+}
+
+type ProofWait = { wait(ms: number): Promise<boolean>; stop(): void }
+
+/** Waits that end early, false, once stopped; their timers never outlive the stop. */
+function createProofWait(sleep?: (ms: number) => Promise<void>): ProofWait {
+  let stopped = false
+  let cancel = (): void => {}
+  return {
+    async wait(ms) {
+      if (stopped) {
+        return false
+      }
+      if (sleep) {
+        await sleep(ms)
+        return !stopped
+      }
+      return new Promise<boolean>((resolve) => {
+        const timer = setTimeout(() => resolve(!stopped), ms)
+        timer.unref?.()
+        cancel = () => {
+          clearTimeout(timer)
+          resolve(false)
+        }
+      })
+    },
+    stop() {
+      stopped = true
+      cancel()
+    }
+  }
+}
+
+function hasReportChannel(child: GuardedChild): boolean {
+  try {
+    return child.connected
+  } catch {
+    // A handle-less child can throw here; the guard must never fail the spawn it watches.
+    return false
+  }
+}
+
+function reportedGroup(message: unknown): number | null {
+  const pgid =
+    typeof message === 'object' && message !== null && 'pgid' in message ? message.pgid : null
+  return typeof pgid === 'number' && Number.isSafeInteger(pgid) && pgid > 1 ? pgid : null
+}
+
+function isErrnoCode(error: unknown, code: string): boolean {
+  return error instanceof Error && 'code' in error && error.code === code
+}

src/main/provider-process/provider-process-row-read.test.ts

@@ -0,0 +1,32 @@
+import { describe, expect, it } from 'vitest'
+import { parseLinuxProcStat, readPosixProcessRow } from './provider-process-row-read'
+
+describe('parseLinuxProcStat', () => {
+  it('reads parent, group and start ticks past a comm with spaces and parentheses', () => {
+    const fields = ['S', '41', '4242', '4242', ...Array<string>(15).fill('0'), '987654', '0']
+    expect(parseLinuxProcStat(4242, `4242 (open (code) serve) ${fields.join(' ')}`)).toEqual({
+      pid: 4242,
+      ppid: 41,
+      pgid: 4242,
+      startedAt: 'clock-tick:987654',
+      startResolution: 'clock-tick'
+    })
+  })
+
+  it('rejects a truncated stat line', () => {
+    expect(parseLinuxProcStat(4242, '4242 (opencode) S 41')).toBeNull()
+  })
+})
+
+describe.runIf(process.platform !== 'win32')('readPosixProcessRow', () => {
+  it('reads this process with the same start marker every time', async () => {
+    const first = await readPosixProcessRow(process.pid)
+    const second = await readPosixProcessRow(process.pid)
+    expect(first).toMatchObject({ pid: process.pid, ppid: process.ppid })
+    expect(second).toEqual(first)
+  })
+
+  it('reports a pid with no process as absent, not unverifiable', async () => {
+    await expect(readPosixProcessRow(2 ** 22 + 12_345)).resolves.toBeNull()
+  })
+})

src/main/provider-process/provider-process-row-read.ts

@@ -0,0 +1,66 @@
+import { readFile } from 'node:fs/promises'
+import { runProcess } from '@orca/process-host'
+import { parseProcessTable, type ProcessTableRow } from '../pty-process-table-parser'
+
+const PROCESS_ROW_TIMEOUT_MS = 2_000
+
+/**
+ * One process's parent, group and start marker. `clock-tick` starts come from Linux /proc, where a
+ * pid reused within one tick would need the whole pid space to cycle; `second` starts are ps lstart.
+ */
+export type ProviderProcessRow = Pick<ProcessTableRow, 'pid' | 'ppid' | 'pgid' | 'startedAt'> & {
+  startResolution: 'clock-tick' | 'second'
+}
+
+/** `null`: no such process. `'unverifiable'`: the read failed, which proves nothing. */
+export type ProcessRowRead = ProviderProcessRow | null | 'unverifiable'
+
+export function readPosixProcessRow(
+  pid: number,
+  platform: NodeJS.Platform = process.platform
+): Promise<ProcessRowRead> {
+  return platform === 'linux' ? readLinuxProcStatRow(pid) : readPsProcessRow(pid)
+}
+
+/** Parses /proc/<pid>/stat; the comm field can hold spaces and parentheses, so cut past ") ". */
+export function parseLinuxProcStat(pid: number, stat: string): ProviderProcessRow | null {
+  const fields = stat.slice(stat.lastIndexOf(') ') + 2).split(' ')
+  const ppid = Number(fields[1])
+  const pgid = Number(fields[2])
+  const ticks = fields[19]
+  if (!Number.isSafeInteger(ppid) || !Number.isSafeInteger(pgid) || !/^\d+$/.test(ticks ?? '')) {
+    return null
+  }
+  return { pid, ppid, pgid, startedAt: `clock-tick:${ticks}`, startResolution: 'clock-tick' }
+}
+
+async function readLinuxProcStatRow(pid: number): Promise<ProcessRowRead> {
+  try {
+    return parseLinuxProcStat(pid, await readFile(`/proc/${pid}/stat`, 'utf8')) ?? 'unverifiable'
+  } catch (error) {
+    return error instanceof Error && 'code' in error && error.code === 'ENOENT'
+      ? null
+      : 'unverifiable'
+  }
+}
+
+async function readPsProcessRow(pid: number): Promise<ProcessRowRead> {
+  try {
+    const result = await runProcess({
+      program: 'ps',
+      args: ['-o', 'pid=,ppid=,pgid=,lstart=', '-p', String(pid)],
+      // The same locale on every read, so two reads of one process compare equal.
+      env: { ...process.env, LANG: 'C', LC_ALL: 'C' },
+      timeoutMs: PROCESS_ROW_TIMEOUT_MS,
+      maxOutputBytes: 4096
+    })
+    const row = parseProcessTable(result.stdout).find((entry) => entry.pid === pid)
+    if (row) {
+      return { ...row, startResolution: 'second' }
+    }
+    // ps exits 1 with no rows for a pid that does not exist; anything else is a failed read.
+    return !result.timedOut && result.code === 1 ? null : 'unverifiable'
+  } catch {
+    return 'unverifiable'
+  }
+}

src/main/provider-process/provider-process-supervisor.test.ts

@@ -62,7 +62,11 @@ describe('structured provider supervision', () => {
       lifetime: 'one-shot'
     })
 
-    expect(spec).toMatchObject({ program: process.execPath, detached: true, supervised: true })
+    expect(spec).toMatchObject({
+      program: process.execPath,
+      detached: true,
+      supervised: true
+    })
     expect(
       JSON.parse(Buffer.from(spec.env.ORCA_PROVIDER_SUPERVISOR_SPEC!, 'base64').toString())
     ).toMatchObject({ lifetime: 'one-shot' })
@@ -72,7 +76,11 @@ describe('structured provider supervision', () => {
     // The launch's override goes through the one env rule, then only the provider gets it.
     const spec = createProviderSpawnSpec(
       { ...command, env: { NODE_OPTIONS: '--require /missing.js' } },
-      { PATH: '/bin', NODE_OPTIONS: '--inherited', NODE_REPL_EXTERNAL_MODULE: '/repl.js' },
+      {
+        PATH: '/bin',
+        NODE_OPTIONS: '--inherited',
+        NODE_REPL_EXTERNAL_MODULE: '/repl.js'
+      },
       'linux'
     )
 
@@ -80,7 +88,10 @@ describe('structured provider supervision', () => {
     expect(spec.env).not.toHaveProperty('NODE_REPL_EXTERNAL_MODULE')
     expect(
       JSON.parse(Buffer.from(spec.env.ORCA_PROVIDER_SUPERVISOR_SPEC!, 'base64').toString()).nodeEnv
-    ).toEqual({ NODE_OPTIONS: '--require /missing.js', NODE_REPL_EXTERNAL_MODULE: '/repl.js' })
+    ).toEqual({
+      NODE_OPTIONS: '--require /missing.js',
+      NODE_REPL_EXTERNAL_MODULE: '/repl.js'
+    })
   })
 
   it('keeps every argv and env string of a 120 KiB argv prompt under the Linux 128 KiB cap', () => {
@@ -128,22 +139,40 @@ describe('structured provider supervision', () => {
       env: { PATH: '/bin', CODEX_HOME: '/tmp/codex' },
       cwd: '/work/repo',
       detached: false,
-      supervised: false
+      supervised: false,
+      groupReportStdio: []
     })
   })
 
   it.each(['win32', 'darwin', 'linux'] as const)(
     'applies launch environment overrides and deletions on %s',
     (platform) => {
-      const baseEnv = { PATH: '/bin', AGENT_HOME: '/inherited', PARENT_AGENT: 'parent' }
-      const overlay = { AGENT_HOME: '/pinned', LAUNCH_ONLY: 'added', PARENT_AGENT: 'overlay' }
+      const baseEnv = {
+        PATH: '/bin',
+        AGENT_HOME: '/inherited',
+        PARENT_AGENT: 'parent'
+      }
+      const overlay = {
+        AGENT_HOME: '/pinned',
+        LAUNCH_ONLY: 'added',
+        PARENT_AGENT: 'overlay'
+      }
       const spec = createProviderSpawnSpec(
-        { command: 'provider', args: [], env: overlay, envToDelete: ['PARENT_AGENT'] },
+        {
+          command: 'provider',
+          args: [],
+          env: overlay,
+          envToDelete: ['PARENT_AGENT']
+        },
         baseEnv,
         platform
       )
 
-      expect(spec.env).toMatchObject({ PATH: '/bin', AGENT_HOME: '/pinned', LAUNCH_ONLY: 'added' })
+      expect(spec.env).toMatchObject({
+        PATH: '/bin',
+        AGENT_HOME: '/pinned',
+        LAUNCH_ONLY: 'added'
+      })
       expect(spec.env).not.toHaveProperty('PARENT_AGENT')
       expect(baseEnv.PARENT_AGENT).toBe('parent')
       expect(baseEnv.AGENT_HOME).toBe('/inherited')

src/main/provider-process/provider-process-supervisor.ts

@@ -42,18 +42,31 @@ const exitWithSpawnFailure = (error, thrown) => {
   try { require('node:fs').writeSync(2, ${JSON.stringify(PROVIDER_SPAWN_FAILURE_MARKER)} + JSON.stringify(report) + '\\n') } catch {}
   process.exit(127)
 }
+// Orca opens the report channel only when it will keep the group's identity. Linux children
+// inherit its fd, so /dev/null takes that slot and only this pid holds the channel.
+const reports = typeof process.send === 'function'
+const reportShield = reports ? require('node:fs').openSync('/dev/null', 'r') : null
 let child
 try {
   child = spawn(providerCommand, providerArgs, {
     cwd: spec.cwd,
     env: childEnv,
-    stdio: ['pipe', 'pipe', 'pipe'],
+    stdio: ['pipe', 'pipe', 'pipe', ...(reportShield === null ? [] : [reportShield])],
     detached: true
   })
 } catch (error) {
   // Node throws most spawn failures (ENOEXEC, ENOTDIR, ...) rather than emitting them.
   exitWithSpawnFailure(error, true)
 }
+// With the group's identity in Orca, a SIGKILL of this pid cannot orphan the group.
+if (reports) {
+  try { require('node:fs').closeSync(reportShield) } catch {}
+  const disconnect = () => { try { process.disconnect() } catch {} }
+  try {
+    if (child.pid) process.send({ pgid: child.pid }, disconnect)
+    else disconnect()
+  } catch { disconnect() }
+}
 let timer
 let ownerShutdownTimer
 let settling = false
@@ -269,6 +282,8 @@ export function createProviderSpawnSpec(
   detached: boolean
   /** The child is the supervisor, whose SIGTERM stops the provider and then itself. */
   supervised: boolean
+  /** Appended after the three standard slots: the channel the supervisor reports its group on. */
+  groupReportStdio: [] | ['ipc']
 } {
   const childEnv = resolveProviderChildEnv(launch, baseEnv)
   const supervisor =
@@ -285,6 +300,7 @@ export function createProviderSpawnSpec(
     env: supervisor?.env ?? childEnv,
     cwd: launch.cwd ?? process.cwd(),
     detached: platform !== 'win32',
-    supervised: supervisor !== null
+    supervised: supervisor !== null,
+    groupReportStdio: supervisor ? ['ipc'] : []
   }
 }

src/main/text-generation/commit-message-text-generation-model-discovery.test.ts

@@ -40,12 +40,18 @@ const expectChildTerminated = createChildTerminationExpectation(terminateWindows
 const hostPlatform = process.platform
 
 afterEach(() => {
-  Object.defineProperty(process, 'platform', { configurable: true, value: hostPlatform })
+  Object.defineProperty(process, 'platform', {
+    configurable: true,
+    value: hostPlatform
+  })
   vi.unstubAllEnvs()
 })
 
 beforeEach(() => {
-  Object.defineProperty(process, 'platform', { configurable: true, value: 'win32' })
+  Object.defineProperty(process, 'platform', {
+    configurable: true,
+    value: 'win32'
+  })
   // Windows resolves a bare agent name on PATH; the host's own installs must not answer it.
   vi.stubEnv('PATH', '')
   terminateWindowsProcessTreeMock.mockClear()
@@ -67,13 +73,20 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
 
   it('discovers dynamic models through the agent CLI', async () => {
     // The host's own spawn shape: supervised on POSIX, direct on Windows.
-    Object.defineProperty(process, 'platform', { configurable: true, value: hostPlatform })
+    Object.defineProperty(process, 'platform', {
+      configurable: true,
+      value: hostPlatform
+    })
     const listeners = new Map<string, (value: unknown) => void>()
     const child = {
       pid: 123,
       kill: vi.fn(),
-      stdout: { on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback)) },
-      stderr: { on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback)) },
+      stdout: {
+        on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback))
+      },
+      stderr: {
+        on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback))
+      },
       stdin: { end: vi.fn() },
       on: vi.fn((event, callback) => listeners.set(event, callback))
     }
@@ -98,13 +111,20 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
 
   it('writes the Claude list_models request to stdin and parses the control response', async () => {
     // The host's own spawn shape: supervised on POSIX, direct on Windows.
-    Object.defineProperty(process, 'platform', { configurable: true, value: hostPlatform })
+    Object.defineProperty(process, 'platform', {
+      configurable: true,
+      value: hostPlatform
+    })
     const listeners = new Map<string, (value: unknown) => void>()
     const child = {
       pid: 123,
       kill: vi.fn(),
-      stdout: { on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback)) },
-      stderr: { on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback)) },
+      stdout: {
+        on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback))
+      },
+      stderr: {
+        on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback))
+      },
       stdin: { on: vi.fn(), end: vi.fn() },
       on: vi.fn((event, callback) => listeners.set(event, callback))
     }
@@ -159,7 +179,7 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
     ])
     expect(spawnMock.mock.calls[0]![2]).toMatchObject({
       windowsHide: true,
-      stdio: ['pipe', 'pipe', 'pipe']
+      stdio: ['pipe', 'pipe', 'pipe', ...(process.platform === 'win32' ? [] : ['ipc'])]
     })
     expect(child.stdin.end).toHaveBeenCalledWith(expect.stringContaining('"list_models"'))
   })
@@ -169,8 +189,12 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
     const child = {
       pid: 123,
       kill: vi.fn(),
-      stdout: { on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback)) },
-      stderr: { on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback)) },
+      stdout: {
+        on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback))
+      },
+      stderr: {
+        on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback))
+      },
       stdin: { on: vi.fn(), end: vi.fn() },
       on: vi.fn((event, callback) => listeners.set(event, callback))
     }
@@ -196,13 +220,20 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
 
   it('discovers dynamic models through the configured agent command override', async () => {
     // The host's own spawn shape: supervised on POSIX, direct on Windows.
-    Object.defineProperty(process, 'platform', { configurable: true, value: hostPlatform })
+    Object.defineProperty(process, 'platform', {
+      configurable: true,
+      value: hostPlatform
+    })
     const listeners = new Map<string, (value: unknown) => void>()
     const child = {
       pid: 123,
       kill: vi.fn(),
-      stdout: { on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback)) },
-      stderr: { on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback)) },
+      stdout: {
+        on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback))
+      },
+      stderr: {
+        on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback))
+      },
       stdin: { end: vi.fn() },
       on: vi.fn((event, callback) => listeners.set(event, callback))
     }
@@ -239,8 +270,12 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
       const child = {
         pid: 123,
         kill: vi.fn(),
-        stdout: { on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback)) },
-        stderr: { on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback)) },
+        stdout: {
+          on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback))
+        },
+        stderr: {
+          on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback))
+        },
         stdin: { end: vi.fn() },
         on: vi.fn((event, callback) => listeners.set(event, callback))
       }
@@ -283,8 +318,12 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
     const child = {
       pid: 123,
       kill: vi.fn(),
-      stdout: { on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback)) },
-      stderr: { on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback)) },
+      stdout: {
+        on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback))
+      },
+      stderr: {
+        on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback))
+      },
       stdin: { end: vi.fn() },
       on: vi.fn((event, callback) => listeners.set(event, callback))
     }
@@ -307,8 +346,12 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
     const child = {
       pid: 123,
       kill: vi.fn(),
-      stdout: { on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback)) },
-      stderr: { on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback)) },
+      stdout: {
+        on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback))
+      },
+      stderr: {
+        on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback))
+      },
       stdin: { end: vi.fn() },
       on: vi.fn((event, callback) => listeners.set(event, callback))
     }
@@ -331,8 +374,12 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
     const child = {
       pid: 123,
       kill: vi.fn(),
-      stdout: { on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback)) },
-      stderr: { on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback)) },
+      stdout: {
+        on: vi.fn((event, callback) => listeners.set(`stdout:${event}`, callback))
+      },
+      stderr: {
+        on: vi.fn((event, callback) => listeners.set(`stderr:${event}`, callback))
+      },
       stdin: { end: vi.fn() },
       on: vi.fn((event, callback) => listeners.set(event, callback))
     }
@@ -462,10 +509,17 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
       expect(spawnMock).toHaveBeenCalledTimes(2)
       secondChild.stdout.emit(
         'data',
-        Buffer.from(JSON.stringify({ models: [{ slug: 'gpt-5.5', display_name: 'GPT-5.5' }] }))
+        Buffer.from(
+          JSON.stringify({
+            models: [{ slug: 'gpt-5.5', display_name: 'GPT-5.5' }]
+          })
+        )
       )
       secondChild.emit('close', 0)
-      await expect(second).resolves.toMatchObject({ success: true, defaultModelId: 'gpt-5.5' })
+      await expect(second).resolves.toMatchObject({
+        success: true,
+        defaultModelId: 'gpt-5.5'
+      })
     } finally {
       vi.useRealTimers()
     }
@@ -494,10 +548,17 @@ describe('discoverCommitMessageModelsLocal on the Windows direct-child path', ()
 
       secondChild.stdout.emit(
         'data',
-        Buffer.from(JSON.stringify({ models: [{ slug: 'gpt-5.5', display_name: 'GPT-5.5' }] }))
+        Buffer.from(
+          JSON.stringify({
+            models: [{ slug: 'gpt-5.5', display_name: 'GPT-5.5' }]
+          })
+        )
       )
       secondChild.emit('close', 0)
-      await expect(second).resolves.toMatchObject({ success: true, defaultModelId: 'gpt-5.5' })
+      await expect(second).resolves.toMatchObject({
+        success: true,
+        defaultModelId: 'gpt-5.5'
+      })
     } finally {
       vi.useRealTimers()
     }

src/main/text-generation/source-control-agent-launch.test.ts

@@ -18,8 +18,12 @@ vi.mock('@orca/process-host', async (importOriginal) => {
   return { ...actual, spawnProcess: spawnProcessMock }
 })
 
-function fakeChild(): EventEmitter & { stdin: { on: () => void; end: () => void } } {
-  return Object.assign(new EventEmitter(), { stdin: { on: vi.fn(), end: vi.fn() } })
+function fakeChild(): EventEmitter & {
+  stdin: { on: () => void; end: () => void }
+} {
+  return Object.assign(new EventEmitter(), {
+    stdin: { on: vi.fn(), end: vi.fn() }
+  })
 }
 
 function lastSpawnSpec(): ProcessSpec {
@@ -61,10 +65,13 @@ describe('spawnSourceControlAgent', () => {
       args: ['-e', POSIX_PROVIDER_SUPERVISOR_SCRIPT, '--', '/opt/agent/claude', '-p', '--verbose'],
       cwd,
       detached: true,
-      stdio: ['pipe', 'pipe', 'pipe']
+      stdio: ['pipe', 'pipe', 'pipe', 'ipc']
     })
     expect(spec.env).toMatchObject({ PATH: '/usr/bin', AGENT_TOKEN: 'kept' })
-    expect(decodedSupervisorSpec(spec)).toMatchObject({ cwd, lifetime: 'one-shot' })
+    expect(decodedSupervisorSpec(spec)).toMatchObject({
+      cwd,
+      lifetime: 'one-shot'
+    })
   })
 
   it('spawns a Windows agent directly, without a supervisor', () => {
@@ -80,7 +87,10 @@ describe('spawnSourceControlAgent', () => {
     )
 
     expect(child.supervised).toBeUndefined()
-    expect(lastSpawnSpec()).toMatchObject({ program: 'C:/tools/agent.exe', args: ['--print'] })
+    expect(lastSpawnSpec()).toMatchObject({
+      program: 'C:/tools/agent.exe',
+      args: ['--print']
+    })
     expect(lastSpawnSpec().env).not.toHaveProperty('ORCA_PROVIDER_SUPERVISOR_SPEC')
   })
 
@@ -113,7 +123,11 @@ describe.runIf(process.platform !== 'win32')('supervised agent processes', () =>
 
     await expect(
       generateCommitMessageFromContext(
-        { branch: 'main', stagedSummary: 'M README.md', stagedPatch: '+test' },
+        {
+          branch: 'main',
+          stagedSummary: 'M README.md',
+          stagedPatch: '+test'
+        },
         { agentId: 'custom', model: '', customAgentCommand: `"${missing}"` },
         { kind: 'local', cwd: folder }
       )
@@ -129,7 +143,11 @@ describe.runIf(process.platform !== 'win32')('supervised agent processes', () =>
 
     await expect(
       generateCommitMessageFromContext(
-        { branch: 'main', stagedSummary: 'M README.md', stagedPatch: '+test' },
+        {
+          branch: 'main',
+          stagedSummary: 'M README.md',
+          stagedPatch: '+test'
+        },
         { agentId: 'custom', model: '', customAgentCommand: `"${agent}"` },
         { kind: 'local', cwd: folder }
       )
@@ -152,8 +170,16 @@ describe.runIf(process.platform !== 'win32')('supervised agent processes', () =>
 
     await expect(
       generateCommitMessageFromContext(
-        { branch: 'main', stagedSummary: 'M README.md', stagedPatch: '+test' },
-        { agentId: 'custom', model: '', customAgentCommand: `"${command}"` },
+        {
+          branch: 'main',
+          stagedSummary: 'M README.md',
+          stagedPatch: '+test'
+        },
+        {
+          agentId: 'custom',
+          model: '',
+          customAgentCommand: `"${command}"`
+        },
         { kind: 'local', cwd: folder }
       )
     ).resolves.toEqual({
@@ -168,7 +194,11 @@ describe.runIf(process.platform !== 'win32')('supervised agent processes', () =>
 
     await expect(
       generateCommitMessageFromContext(
-        { branch: 'main', stagedSummary: 'M README.md', stagedPatch: '+test' },
+        {
+          branch: 'main',
+          stagedSummary: 'M README.md',
+          stagedPatch: '+test'
+        },
         { agentId: 'custom', model: '', customAgentCommand: `"${missing}"` },
         {
           kind: 'local',
@@ -194,12 +224,22 @@ process.stdin.on('end', () => setTimeout(() => console.log(prompt.includes('READ
 
     await expect(
       generateCommitMessageFromContext(
-        { branch: 'main', stagedSummary: 'M README.md', stagedPatch: '+test' },
-        { agentId: 'custom', model: '', customAgentCommand: `"${process.execPath}" "${agent}"` },
+        {
+          branch: 'main',
+          stagedSummary: 'M README.md',
+          stagedPatch: '+test'
+        },
+        {
+          agentId: 'custom',
+          model: '',
+          customAgentCommand: `"${process.execPath}" "${agent}"`
+        },
         { kind: 'local', cwd: folder }
       )
     ).resolves.toMatchObject({ success: true, message: 'Update README' })
     expect(lastSpawnSpec().program).toBe(process.execPath)
-    expect(decodedSupervisorSpec(lastSpawnSpec())).toMatchObject({ lifetime: 'one-shot' })
+    expect(decodedSupervisorSpec(lastSpawnSpec())).toMatchObject({
+      lifetime: 'one-shot'
+    })
   })
 })

src/main/text-generation/source-control-agent-launch.ts

@@ -2,6 +2,8 @@ import { spawnProcess } from '@orca/process-host'
 import { withCliRuntimeOnPath } from '../../shared/node-cli-command-resolution'
 import { resolveCliCommand } from '../codex-cli/command'
 import { createProviderSpawnSpec } from '../provider-process/provider-process-supervisor'
+import { guardSupervisedProviderGroup } from '../provider-process/provider-group-orphan-guard'
+import { SOURCE_CONTROL_KILL_SITE } from './source-control-local-process'
 import { wslAwareSpawn } from '../git/runner'
 import type {
   SourceControlAgentSpawnInput,
@@ -107,7 +109,11 @@ function spawnSupervisedAgent(
     env: spec.env,
     cwd: spec.cwd,
     detached: spec.detached,
-    stdio: ['pipe', 'pipe', 'pipe']
+    stdio: ['pipe', 'pipe', 'pipe', ...spec.groupReportStdio]
+  })
+  guardSupervisedProviderGroup(child, {
+    site: SOURCE_CONTROL_KILL_SITE,
+    supervised: spec.supervised
   })
   return Object.assign(child, { supervised: spec.supervised })
 }

src/main/text-generation/source-control-local-process.ts

@@ -24,7 +24,7 @@ import type {
   TextGenerationOperation
 } from './source-control-text-generation-types'
 
-const SOURCE_CONTROL_KILL_SITE = 'source-control-text-generation'
+export const SOURCE_CONTROL_KILL_SITE = 'source-control-text-generation'
 
 export async function killSourceControlAgentProcess(
   child: SpawnedSourceControlAgentProcess
@@ -126,7 +126,10 @@ export function runLocalSourceControlPlan(input: {
     } catch (error) {
       markProcessClosed()
       if (error instanceof WindowsCmdLineBreakError) {
-        resolve({ success: false, error: userFacingUnsafeWindowsBatchArgs(plan.label) })
+        resolve({
+          success: false,
+          error: userFacingUnsafeWindowsBatchArgs(plan.label)
+        })
         return
       }
       console.error('[commit-message] Failed to spawn local generator:', error)
@@ -170,7 +173,11 @@ export function runLocalSourceControlPlan(input: {
     const cancel = (): void => {
       canceledByUser = true
       startTermination()
-      finalize({ success: false, error: 'Generation canceled.', canceled: true })
+      finalize({
+        success: false,
+        error: 'Generation canceled.',
+        canceled: true
+      })
     }
     setLocalGenerationCancelToken(laneKey, cancel)
     timer = setTimeout(() => {
@@ -219,7 +226,11 @@ export function runLocalSourceControlPlan(input: {
     const onClose = (code: number | null): void => {
       markClosedAfterTermination()
       if (canceledByUser) {
-        finalize({ success: false, error: 'Generation canceled.', canceled: true })
+        finalize({
+          success: false,
+          error: 'Generation canceled.',
+          canceled: true
+        })
         return
       }
       if (outputLimitExceeded) {