pi-code 1.0.70 → 1.0.71

package/extensions/commands.ts

@@ -335,12 +335,16 @@ export default function commandsExtension(pi: ExtensionAPI) {
     // setModel can reject (e.g. auth resolution fails), and a floated rejection would
     // escape as unhandled; surface it as a no-op instead of leaving the session silently
     // on the command's override model.
-    set: (model) => {
+    set: async (model) => {
       // A refused switch leaves the turn on the session model rather than the one the
-      // command named, and the reply gives no sign of it, so the refusal is reported.
-      void pi.setModel(model as Parameters<typeof pi.setModel>[0]).catch((error: unknown) => {
+      // command named, and the reply gives no sign of it, so the refusal is reported. The
+      // promise never rejects, so a caller that does not await it (agent_settled) cannot
+      // float a rejection; the shutdown handler awaits it so the exit cannot outrun it.
+      try {
+        await pi.setModel(model as Parameters<typeof pi.setModel>[0])
+      } catch (error) {
         console.warn(`pi-code-commands: could not switch to ${typeof model === 'object' && model !== null && 'id' in model ? String((model as { id: unknown }).id) : String(model)}: ${errorMessage(error)}`)
-      })
+      }
     },
   })
   /** The thinking level to restore after a command's `effort:` override drove its run,
@@ -363,14 +367,23 @@ export default function commandsExtension(pi: ExtensionAPI) {
     pendingAgentRules = undefined
     pendingSkillRules = undefined
     pendingPathRules = undefined
-    modelOverride.settle()
-    effortOverride.settle()
+    void modelOverride.settle()
+    void effortOverride.settle()
     if (pendingRestore) {
       pi.setActiveTools(pendingRestore)
       pendingRestore = undefined
     }
   })
 
+  // agent_settled never fires when pi is quit mid-run (double Ctrl+C, Ctrl+D, a closed
+  // terminal, SIGTERM), and the switch is already in the transcript: `pi -c` resumed on the
+  // command's override model at its raised level. A run that settled has nothing armed, so
+  // this restores nothing twice.
+  pi.on('session_shutdown', async () => {
+    await modelOverride.settle()
+    await effortOverride.settle()
+  })
+
   // The active-tool set has no argument dimension, so a scoped grant hands the turn
   // the whole tool; the scope is enforced here instead, when the call arrives. Same
   // steering-not-sandbox caveat as plan mode's guard.

package/extensions/hooks/index.ts

@@ -223,6 +223,12 @@ export default function hooksExtension(pi: ExtensionAPI) {
   let managedHooksOnly = false
   /** Skill hooks registered this session, re-applied when a settings edit reloads. */
   const registeredSkillHooks: Array<{ skillName: string; hooks: Record<string, unknown> }> = []
+  /** Whether this process is a subagent's child. Claude runs settings hooks inside a
+   * subagent for tool events, but Stop is the main agent's, a subagent completes with
+   * SubagentStop, and a subagent run is neither a session (SessionStart, SessionEnd) nor
+   * a prompt the user submitted (UserPromptSubmit). The child is a full pi process loading
+   * the same settings, so without this every one of them fired in it. */
+  const inSubagentChild = (): boolean => process.env.PI_CODE_SUBAGENT === '1'
   /** Stops the settings watcher of the previous session. */
   let disposeSettingsWatch: () => void = () => {}
   /** Claude's disableAllHooks escape hatch was set somewhere in the honored chain. */
@@ -466,7 +472,7 @@ export default function hooksExtension(pi: ExtensionAPI) {
     pendingSessionContext = []
     // "reload" re-fires in-process with the same conversation and would double-run hooks;
     // a fork is a genuine session begin, which Claude reports as source "fork".
-    if (event.reason === 'reload') return
+    if (event.reason === 'reload' || inSubagentChild()) return
     const source = claudeSpelling(SESSION_START_SOURCE, event.reason)
     const commands = matchingCommands(config.SessionStart, source.names)
     const payload = { hook_event_name: 'SessionStart', source: source.value }
@@ -593,7 +599,7 @@ export default function hooksExtension(pi: ExtensionAPI) {
   pi.on('input', async (event, ctx) => {
     // Only genuine user input; extension-injected messages (plan-mode, subagent) are not
     // prompts the user submitted.
-    if (event.source === 'extension') return { action: 'continue' }
+    if (event.source === 'extension' || inSubagentChild()) return { action: 'continue' }
     // The user typed, so the pending idle_prompt no longer applies.
     cancelIdlePrompt()
     // Genuine user input is progress, so it breaks a Stop-hook continuation streak: the
@@ -605,12 +611,15 @@ export default function hooksExtension(pi: ExtensionAPI) {
       ctx.ui.notify(decision.reason ?? 'Prompt blocked by hook', 'error')
       return { action: 'handled' }
     }
-    // Claude injects a UserPromptSubmit hook's context ahead of the prompt; transform is
-    // pi's seam for rewriting the submitted text. The prompt itself always survives:
     // "UserPromptSubmit: can't replace the prompt; it only injects additionalContext
-    // alongside it". suppressOriginalPrompt scopes to the block message, which never
+    // alongside it", as a system reminder with no visible transcript entry. So the context
+    // is a hidden message next to the prompt and the prompt text is left alone: prepended,
+    // it moved a `/skill:` or `/template` invocation off position 0 and pi stopped
+    // expanding it. pi appends a "nextTurn" message to this very prompt (after this handler
+    // runs); one typed mid-stream never reaches before_agent_start, so it is queued in the
+    // prompt's own mode. suppressOriginalPrompt scopes to the block message, which never
     // carries the prompt here, so it needs nothing of its own.
-    if (decision.context) return { action: 'transform', text: `${decision.context}\n\n${event.text}` }
+    if (decision.context) pi.sendMessage({ customType: 'claude-hook-context', content: decision.context, display: false }, { deliverAs: event.streamingBehavior ?? 'nextTurn' })
     return { action: 'continue' }
   })
 
@@ -630,6 +639,7 @@ export default function hooksExtension(pi: ExtensionAPI) {
    * turn cancels it. Observational only; exit codes and JSON output are ignored. */
   const armIdlePrompt = (ctx: ExtensionContext): void => {
     cancelIdlePrompt()
+    if (inSubagentChild()) return
     const notifyCommands = matchingCommands(config.Notification, ['idle_prompt'])
     if (notifyCommands.length > 0) {
       const runner = boundRunner(ctx)
@@ -639,21 +649,26 @@ export default function hooksExtension(pi: ExtensionAPI) {
       idlePromptTimer.unref?.()
     }
   }
+  /** In a subagent child, the agent-frontmatter Stop hooks were converted to SubagentStop
+   * and fire at the child's own end, notify-style, before the Stop early-returns, which do
+   * not apply to them. Only the agent's own frontmatter hooks (stamped agent:<name>): the
+   * settings-level SubagentStop is fired by the parent when the child completes, so the
+   * child firing it too ran it twice. */
+  const fireChildSubagentStop = async (event: { messages?: Array<{ role: string; content: unknown }> }, ctx: ExtensionContext): Promise<void> => {
+    if (!agentIdentity) return
+    const subStop = matchingCommands(config.SubagentStop, agentIdentity.agent).filter((command) => command.origin?.startsWith('agent:'))
+    if (subStop.length === 0) return
+    const subText = lastAssistantText(event.messages ?? [])
+    const subPayload = { hook_event_name: 'SubagentStop', agent_type: agentIdentity.agent, ...(agentIdentity.id ? { agent_id: agentIdentity.id } : {}), stop_hook_active: false, ...(subText ? { last_assistant_message: subText } : {}) }
+    await runNotifyHooks(subStop, subPayload, boundRunner(ctx)).catch(() => {})
+  }
+
   pi.on('agent_end', async (event, ctx) => {
     armIdlePrompt(ctx)
+    await fireChildSubagentStop(event as { messages?: Array<{ role: string; content: unknown }> }, ctx)
 
-    // In a subagent child, the agent-frontmatter Stop hooks were converted to
-    // SubagentStop and fire here, at the child's own end, notify-style; before the
-    // Stop early-returns, which do not apply to them.
-    if (agentIdentity) {
-      const subStop = matchingCommands(config.SubagentStop, agentIdentity.agent)
-      if (subStop.length > 0) {
-        const subText = lastAssistantText((event as { messages?: Array<{ role: string; content: unknown }> }).messages ?? [])
-        const subPayload = { hook_event_name: 'SubagentStop', agent_type: agentIdentity.agent, ...(agentIdentity.id ? { agent_id: agentIdentity.id } : {}), stop_hook_active: false, ...(subText ? { last_assistant_message: subText } : {}) }
-        await runNotifyHooks(subStop, subPayload, boundRunner(ctx)).catch(() => {})
-      }
-    }
-
+    // Stop is the main agent's event; the child's completion was handled above.
+    if (inSubagentChild()) return
     // Stop has no matcher support (a stray matcher is ignored, as Claude documents)
     // and an `if`-carrying hook never runs on a non-tool event.
     const commands = allCommands(config.Stop).filter((command) => passesIfFilter(command, undefined))
@@ -771,13 +786,16 @@ export default function hooksExtension(pi: ExtensionAPI) {
     const reason = claudeSpelling(SESSION_END_REASON, event.reason)
     // SessionEnd rides Claude's short shared budget (see sessionEndTimeoutMs) so a
     // slow hook cannot stall session exit, /new or /resume.
-    const sessionEndCommands = matchingCommands(config.SessionEnd, reason.names).filter((command) => passesIfFilter(command, undefined))
+    const sessionEndCommands = inSubagentChild() ? [] : matchingCommands(config.SessionEnd, reason.names).filter((command) => passesIfFilter(command, undefined))
     const runner = boundRunner(ctx)
     const results = await Promise.all(sessionEndCommands.map((command) => runner(command, { hook_event_name: 'SessionEnd', reason: reason.value }, sessionEndTimeoutMs(command))))
     surfaceSystemMessages(results, (message) => ctx.ui.notify(message, 'warning'))
-    // Claude kills async hooks still running at teardown; the session that spawned
-    // these is over, and their delivery would target a disposed context anyway.
-    for (const kill of backgroundKills) kill()
+    // Claude: "In non-interactive mode with the -p flag, Claude Code kills any async hook
+    // still running at teardown". Only there: an interactive session leaves them running, so
+    // an async SessionEnd hook, started by this very shutdown, can finish instead of being
+    // killed a microtask after it spawned. Their delivery targets a disposed context and is
+    // caught (see deliverBackgroundResult).
+    if (ctx.hasUI !== true) for (const kill of backgroundKills) kill()
     backgroundKills.clear()
   })
 

package/extensions/internal/turn-override.ts

@@ -26,14 +26,15 @@ export interface TurnOverride<T> {
   reset(): void
   /** Restore the captured prior via `set` and clear the capture. A conditional override
    * stands down when the current value has moved off the target. No-op when nothing is
-   * armed. */
-  settle(): void
+   * armed. Returns what `set` returns, so a caller that must not exit before an async
+   * restore has landed (a shutdown) can await it. */
+  settle(): void | Promise<void>
 }
 
 export function createTurnOverride<T>(opts: {
   /** Applies the restore value. Consumers wrap their setter here (async-with-catch for
    * commands' model, plain for the thinking level). */
-  set: (value: T) => void
+  set: (value: T) => void | Promise<void>
   /** Reads the current value, for a conditional restore. Only consulted when
    * `conditional` is set. */
   get?: () => T | undefined
@@ -67,7 +68,7 @@ export function createTurnOverride<T>(opts: {
         const current = opts.get?.()
         if (current !== undefined && current !== movedTo) return
       }
-      opts.set(restore)
+      return opts.set(restore)
     },
   }
 }

package/extensions/thinking.ts

@@ -91,11 +91,17 @@ export default function thinkingExtension(pi: ExtensionAPI) {
     // the keyword in the prompt.
   })
 
+  // agent_settled never fires when pi is quit mid-run, and the escalation is already in the
+  // transcript: `pi -c` would resume at the raised level for good.
+  pi.on('session_shutdown', async () => {
+    await override.settle()
+  })
+
   pi.on('agent_settled', () => {
     // Restore the pre-escalation level, but only if nothing else moved it since (a
     // command's effort restore, a manual change): the conditional override stands down in
     // that case and restores unconditionally when the level cannot be read, the prior
     // best-effort behavior.
-    override.settle()
+    void override.settle()
   })
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-code",
-  "version": "1.0.70",
+  "version": "1.0.71",
   "description": "Claude Code experience for the pi coding agent: reads your .claude config (rules, commands, skills, hooks, output styles, MCP servers, agents) and adds todo, checkpoints, memory, web, subagents, and goals",
   "keywords": [
     "pi",