fix(bridge): subprocess reaping + health probe + concurrent-stream safety

src/backends/claude.ts

@@ -36,6 +36,7 @@ import { hostSpawner } from '../executors/host.js'
 import type { Spawner } from '../executors/types.js'
 import { readProcessLines, waitForProcessClose } from './process-lines.js'
 import { writeStdinPayload } from './stdin-payload.js'
+import { killTree } from '../executors/process-tree.js'
 
 interface ClaudeStreamInit {
   type: 'system'
@@ -228,11 +229,10 @@ export class ClaudeBackend implements Backend {
     const earlySpawnError = spawned.spawnError?.()
     if (earlySpawnError) spawnErrorMessage = earlySpawnError.message
 
-    const timeoutHandle = setTimeout(() => {
-      child.kill('SIGTERM')
-    }, this.timeoutMs)
-
-    const onAbort = () => child.kill('SIGTERM')
+    // Tear down the whole process group (claude + every MCP/tool fork
+    // it owns). See backends/opencode.ts for rationale.
+    const timeoutHandle = setTimeout(() => { void killTree(child) }, this.timeoutMs)
+    const onAbort = (): void => { void killTree(child) }
     signal.addEventListener('abort', onAbort, { once: true })
 
     let emittedAnyToolCall = false
@@ -339,7 +339,10 @@ export class ClaudeBackend implements Backend {
     } finally {
       clearTimeout(timeoutHandle)
       signal.removeEventListener('abort', onAbort)
-      if (child.exitCode === null) child.kill('SIGTERM')
+      // Always tear down the whole subtree before releasing the slot.
+      // Reaps MCP servers and tool sub-processes claude spawned. Pre-fix
+      // this was `child.kill('SIGTERM')` which leaked grand-children.
+      await killTree(child)
       releaseSpawner()
       mcpMaterialised?.cleanup()
     }

src/backends/codex.ts

@@ -39,6 +39,7 @@ import { contentToText } from './content.js'
 import { hostSpawner } from '../executors/host.js'
 import type { Spawner } from '../executors/types.js'
 import { readProcessLines, waitForProcessClose } from './process-lines.js'
+import { killTree } from '../executors/process-tree.js'
 
 export interface CodexBackendOptions {
   bin: string
@@ -154,8 +155,9 @@ export class CodexBackend implements Backend {
     const earlySpawnError = spawned.spawnError?.()
     if (earlySpawnError) spawnErrorMessage = earlySpawnError.message
 
-    const timeoutHandle = setTimeout(() => child.kill('SIGTERM'), this.opts.timeoutMs)
-    const onAbort = () => child.kill('SIGTERM')
+    // Group-kill on timeout/abort — see backends/opencode.ts.
+    const timeoutHandle = setTimeout(() => { void killTree(child) }, this.opts.timeoutMs)
+    const onAbort = (): void => { void killTree(child) }
     signal.addEventListener('abort', onAbort, { once: true })
 
     let emittedToolCall = false
@@ -223,7 +225,9 @@ export class CodexBackend implements Backend {
     } finally {
       clearTimeout(timeoutHandle)
       signal.removeEventListener('abort', onAbort)
-      if (child.exitCode === null) child.kill('SIGTERM')
+      // Reap the whole subtree — codex spawns sub-processes for MCP
+      // tool calls, model HTTP I/O, etc. and we owe them a clean exit.
+      await killTree(child)
       releaseSpawner()
       codexHome?.cleanup()
     }

src/backends/kimi.ts

@@ -48,6 +48,7 @@ import { hostSpawner } from '../executors/host.js'
 import type { Spawner } from '../executors/types.js'
 import { readProcessLines, waitForProcessClose } from './process-lines.js'
 import { writeStdinPayload } from './stdin-payload.js'
+import { killTree } from '../executors/process-tree.js'
 
 export interface KimiBackendOptions {
   bin: string
@@ -166,8 +167,10 @@ export class KimiBackend implements Backend {
     const earlySpawnError = spawned.spawnError?.()
     if (earlySpawnError) spawnErrorMessage = earlySpawnError.message
 
-    const timeoutHandle = setTimeout(() => child.kill('SIGTERM'), this.opts.timeoutMs)
-    const onAbort = () => child.kill('SIGTERM')
+    // Tear down the whole process group (kimi + every tool/MCP subprocess
+    // it forks). See backends/opencode.ts for the rationale.
+    const timeoutHandle = setTimeout(() => { void killTree(child) }, this.opts.timeoutMs)
+    const onAbort = (): void => { void killTree(child) }
     signal.addEventListener('abort', onAbort, { once: true })
 
     try {
@@ -332,7 +335,9 @@ export class KimiBackend implements Backend {
     } finally {
       clearTimeout(timeoutHandle)
       signal.removeEventListener('abort', onAbort)
-      if (child.exitCode === null) child.kill('SIGTERM')
+      // Always tear down the whole subtree (kimi + any MCP/tool forks)
+      // before releasing the slot. Idempotent; waits for actual exit.
+      await killTree(child)
       if (configFile) await cleanupConfigFile(configFile)
       mcpMaterialised?.cleanup()
       releaseSpawner()

src/backends/opencode.ts

@@ -27,6 +27,7 @@ import { hostSpawner } from '../executors/host.js'
 import type { Spawner } from '../executors/types.js'
 import { readProcessLines, waitForProcessClose } from './process-lines.js'
 import { writeStdinPayload } from './stdin-payload.js'
+import { killTree } from '../executors/process-tree.js'
 
 export interface OpencodeBackendOptions {
   bin: string
@@ -140,8 +141,12 @@ export class OpencodeBackend implements Backend {
     const earlySpawnError = spawned.spawnError?.()
     if (earlySpawnError) spawnErrorMessage = earlySpawnError.message
 
-    const timeoutHandle = setTimeout(() => child.kill('SIGTERM'), this.opts.timeoutMs)
-    const onAbort = () => child.kill('SIGTERM')
+    // killTree kicks off the SIGTERM→grace→SIGKILL ladder against the
+    // ENTIRE process group (opencode + everything it forked). We fire
+    // and forget here — the actual await happens in the outer finally
+    // so the generator can still emit a clean final delta.
+    const timeoutHandle = setTimeout(() => { void killTree(child) }, this.opts.timeoutMs)
+    const onAbort = (): void => { void killTree(child) }
     signal.addEventListener('abort', onAbort, { once: true })
 
     try {
@@ -251,7 +256,12 @@ export class OpencodeBackend implements Backend {
     } finally {
       clearTimeout(timeoutHandle)
       signal.removeEventListener('abort', onAbort)
-      if (child.exitCode === null) child.kill('SIGTERM')
+      // Always tear down the whole subtree before releasing the slot.
+      // killTree is idempotent and waits up to gracefulMs+500 for the
+      // process to actually exit, so by the time we hit releaseSpawner
+      // there's no orphan to leak. Pre-fix this was `child.kill('SIGTERM')`
+      // which left opencode's HTTP-client + MCP children alive.
+      await killTree(child)
       releaseSpawner()
       mcpMaterialised?.cleanup()
     }

src/backends/pi.ts

@@ -43,6 +43,7 @@ import { contentToText } from './content.js'
 import { hostSpawner } from '../executors/host.js'
 import type { Spawner } from '../executors/types.js'
 import { readProcessLines, waitForProcessClose } from './process-lines.js'
+import { killTree } from '../executors/process-tree.js'
 
 export interface PiBackendOptions {
   bin: string
@@ -162,8 +163,9 @@ export class PiBackend implements Backend {
     const earlySpawnError = spawned.spawnError?.()
     if (earlySpawnError) spawnErrorMessage = earlySpawnError.message
 
-    const timeoutHandle = setTimeout(() => child.kill('SIGTERM'), this.opts.timeoutMs)
-    const onAbort = (): void => { child.kill('SIGTERM') }
+    // Group-kill on timeout/abort — see backends/opencode.ts.
+    const timeoutHandle = setTimeout(() => { void killTree(child) }, this.opts.timeoutMs)
+    const onAbort = (): void => { void killTree(child) }
     signal.addEventListener('abort', onAbort, { once: true })
 
     try {
@@ -293,8 +295,9 @@ export class PiBackend implements Backend {
     } finally {
       clearTimeout(timeoutHandle)
       signal.removeEventListener('abort', onAbort)
+      // Reap the whole subtree before releasing the slot.
+      await killTree(child)
       try { releaseSpawner() } catch { /* best effort */ }
-      if (!child.killed) child.kill('SIGTERM')
     }
   }