mirror-mind 0.39.3 → 0.39.4

package/docs/releases/index.md

@@ -12,6 +12,7 @@ Future release notes should use the structure defined in [Release Notes](../proc
 
 ## Releases
 
+- [v0.39.4 — The Reply Finds Its Prompt](v0.39.4.md) — patch release of PR #51 (@alannascimento1), CR166 and CR196: on Claude Code, a prompt typed while a turn runs and a journey entered in plain words no longer lose the reply; `nothing_waiting` in `hooks.log` means a lost reply again. No schema change, no migration.
 - [v0.39.3 — Windows Gets Its Hooks](v0.39.3.md) — patch release of PR #53, PR #50, and PR #54 (@rodrigoimmaginario): the plugin's hook wrappers and MCP locator run the package's own bin when npm's global entry is a shell shim, so Claude Code hooks and the MCP server work on a Windows package install; `runtime diagnose` no longer counts its own verdict as a front-door error (CR139); the docs point a Windows reader to the installer in its own repository. No migration.
 - [v0.39.2 — Input, Read the Platform's Way](v0.39.2.md) — patch release of PR #49 and PR #48 (@rodrigoimmaginario), CR194, and CR193: `runtime diagnose` stops warning every Windows install about Node; a PowerShell byte-order mark is not stored; stdin is decoded by its mark, UTF-8 or UTF-16, and `identity set` and `journey update` gain `--file`, the route around PowerShell 5.1's ASCII pipe; `package-lock.json` names the package at its version, and the release doctor checks it. No migration.
 - [v0.39.1 — Pi Finds Its Node, or Says So](v0.39.1.md) — patch release of PR #47 (@rodrigoimmaginario) and CR191: a Pi opened where `node` is not on the PATH no longer exits; the Pi extension resolves Node where the hook wrappers look, so a global install needs nothing configured; when no Node resolves, one warning names the fix and `runtime diagnose` reports the `hooks.log` line. No migration.

package/docs/releases/v0.39.4.md

@@ -0,0 +1,81 @@
+---
+digest: >
+  v0.39.4 is a patch release of one pull request by Alan Nascimento, his first here. On
+  Claude Code, two ordinary situations made the Stop hook drop a real reply: a prompt typed
+  while a turn was still running, whose prompt hook fires at queue time under the running
+  turn's id, and a journey entered in plain words, whose Builder load switches the
+  conversation mid-turn. Both replies are now logged beside their prompts, and
+  `nothing_waiting` in `hooks.log` means what it says again. No schema change, no
+  migration, and no model or prompt change.
+---
+
+# v0.39.4 — The Reply Finds Its Prompt
+
+**Date:** 2026-11-03
+
+## Highlights
+
+- **A prompt typed while a turn runs keeps its reply**
+  ([PR #51](https://github.com/mirror-mind-ai/mirror/pull/51), by @alannascimento1;
+  [CR196](../project/refinement/rs010-cv22-oracle-and-port-hygiene/cr196-a-prompt-typed-while-a-turn-runs-loses-its-reply-on-claude-code.md)).
+  Claude Code fires the prompt hook when such a prompt enters its queue, with the running
+  turn's id; the prompt later runs under an id no hook sent, and its Stop found nothing
+  waiting. The prompt is now recorded as queued during the running turn, the running turn's
+  reply stays with its own question, and the Stop of the turn the queued prompt ran as
+  claims it. Never for a command turn, and never a prompt an ordinary one has since
+  followed.
+- **Entering a journey in plain words keeps the turn together**
+  ([CR166](../project/refinement/rs010-cv22-oracle-and-port-hygiene/cr166-a-plain-language-journey-switch-strands-its-turn.md)).
+  "I want to enter journey X" logs the prompt in one conversation; the model's `build load`
+  closes it and binds the journey's; the Stop paired in the new, empty one. A hook prompt
+  still owed its reply now moves with the switch. A prompt logged by hand stays where it
+  was, so `mm-new` and the logger CLI behave as before. Pi gets this half too: its prompts
+  carry turn ids.
+- **`nothing_waiting` means a lost reply again.** On one host, 20 of 22 such lines in a
+  week were these two cases; `runtime diagnose`'s `hook_failures_recorded` was mostly
+  counting them.
+- No schema change (`queued_in_turn` is one more optional key in the `metadata` JSON
+  column `turn_id` already lives in), no migration, and no model or prompt change.
+
+## Where We Started
+
+Alan Nascimento opened PR #51 on 2026-10-30 with the session transcript that shows the
+mechanism, Claude Code's own `queue-operation` record between the prompt hook and the
+Stop, and a week of `hooks.log` counted by cause. The journey-switch half had been in the
+backlog as CR166 since 2026-10-24, unassigned; the queued-prompt half no record described.
+The review, through four lenses, had nothing to amend: the reply rule gains one branch
+between "by turn id" and "by position", each new path bounded by an explicit guard, and
+ten tests of which five fail without the change.
+
+## What Changed
+
+All in `ts/src/conversation/logger.ts`. `promptMetadata` records a second prompt under a
+turn still owed its reply as `queued_in_turn`. `handleStopHook`, for a turn id with no
+prompt of its own, claims the oldest unclaimed queued prompt that no ordinary prompt has
+followed, and logs the reply as paired by `queued_prompt`. `switchConversation` moves the
+old conversation's last row to the new one when it is a hook prompt with no reply after
+it. The reply rule's docblock names the new branch.
+
+## Conscious Exclusions
+
+Both named by the author and recorded in CR166 and CR196:
+
+- In the queued case the running turn's reply is still written after the queued prompt,
+  because the prompt hook fires at queue time: the rows pair by `turn_id`, but a reader
+  ordering by `created_at` sees `Q1, Q2, A1, A2`. Rewriting `created_at` would make it lie
+  about when the prompt was typed.
+- The waiting prompt moves after the old conversation's close-time extraction, which
+  therefore still saw it. Moving it earlier would mean starting the new conversation
+  before closing the old.
+- The two `nothing_waiting` lines in the week that match neither pattern.
+
+## Upgrading
+
+```bash
+mirror runtime update
+```
+
+## Next Horizon
+
+Moacir Moda's PR #27; CR190, CR192, and CR195 in RS010's backlog. CV24 continues on its
+branch.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mirror-mind",
-  "version": "0.39.3",
+  "version": "0.39.4",
   "description": "Mirror Mind: a local-first memory and identity framework for agentic AI runtimes (Pi, Claude Code, Codex, Antigravity)",
   "license": "MIT",
   "repository": {

package/plugins/mirror-mind/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "mirror-mind",
-  "version": "0.39.3",
+  "version": "0.39.4",
   "description": "Mirror Mind — local-first memory and identity for agentic runtimes.",
   "author": {
     "name": "Mirror Mind"

package/plugins/mirror-mind/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "mirror-mind",
   "description": "Mirror Mind — local-first memory and identity for agentic runtimes.",
-  "version": "0.39.3",
+  "version": "0.39.4",
   "mcpServers": {
     "mirror-mind": {
       "command": "/bin/sh",

package/ts/src/conversation/logger.ts

@@ -212,11 +212,13 @@ function discardMarkerActive(session: RuntimeSessionRow | null): boolean {
 }
 
 /**
- * What a message row may carry beside its text. The runtime's turn id (US8)
- * is the one field today: written by the prompt and Stop hooks, read by the
- * reply rule, and NULL for every message logged another way, as before.
+ * What a message row may carry beside its text: the runtime's turn id (US8),
+ * written by the prompt and Stop hooks and read by the reply rule, and the
+ * turn a prompt was queued during, when it arrived while that turn still owed
+ * its reply. NULL for every message logged another way, as before.
  */
 export interface MessageMetadata {
+  readonly queued_in_turn?: string;
   readonly turn_id?: string;
 }
 
@@ -375,7 +377,15 @@ export async function endConversation(
   }
 }
 
-/** Python `switch_conversation`: close the old conversation, bind a new one. */
+/**
+ * Python `switch_conversation`: close the old conversation, bind a new one.
+ *
+ * A prompt still owed its reply moves to the new conversation: when the
+ * model switches during a turn (a Builder load asked for in plain words), the
+ * turn's Stop pairs its reply in the conversation the session is bound to by
+ * then, and the prompt it answers has to be there. Only a hook prompt counts,
+ * one that carries a turn id; a prompt logged by hand stays where it was.
+ */
 export async function switchConversation(
   db: WritableDatabase,
   sessionId: string | null,
@@ -388,6 +398,8 @@ export async function switchConversation(
 
   const runtimeSession = getRuntimeSession(db, resolved);
   const oldConversationId = runtimeSession?.conversationId ?? null;
+  const last = oldConversationId ? lastMessage(db, oldConversationId) : null;
+  const waitingPrompt = last && last.role === "user" && turnIdOfRow(last) ? last : null;
   if (oldConversationId) {
     await endConversation(db, oldConversationId, { extract: true }, deps, hooks);
   }
@@ -404,6 +416,12 @@ export async function switchConversation(
     },
     deps,
   );
+  if (waitingPrompt) {
+    db.prepare("UPDATE messages SET conversation_id = ? WHERE rowid = ?").run(
+      newConversationId,
+      waitingPrompt.rowid,
+    );
+  }
   upsertRuntimeSession(
     db,
     resolved,
@@ -550,7 +568,11 @@ export type StopHookOutcome =
         // command -- the reply is lost. The only skip the caller reports.
         | "nothing_waiting";
     }
-  | { action: "logged"; sessionId: string; pairedBy: "turn_id" | "last_message" }
+  | {
+      action: "logged";
+      sessionId: string;
+      pairedBy: "turn_id" | "queued_prompt" | "last_message";
+    }
   | { action: "failed"; error: string };
 
 export interface HookLoggingOptions {
@@ -653,7 +675,7 @@ export function handleUserPromptHook(
       db,
       sessionId,
       content,
-      { interface: interfaceName, metadata: turnMetadata(turnIdOf(payload)) },
+      { interface: interfaceName, metadata: promptMetadata(db, sessionId, turnIdOf(payload)) },
       deps,
     );
     return { action: "logged", sessionId };
@@ -725,6 +747,64 @@ function turnIdOfRow(row: MessageRow): string {
   return typeof value === "string" ? value : "";
 }
 
+/**
+ * A prompt's metadata. Claude Code fires the prompt hook when a prompt typed
+ * during a running turn enters its queue, with the running turn's id; the
+ * prompt later runs as a turn of its own, under an id no prompt hook sends. A
+ * second prompt under a turn whose reply is still owed is therefore recorded
+ * as queued during it, not as its prompt, so the running turn's reply stays
+ * with its own question and the queued turn's Stop can claim this one.
+ */
+function promptMetadata(
+  db: WritableDatabase,
+  sessionId: string,
+  turnId: string,
+): MessageMetadata | null {
+  const conversationId = getRuntimeSession(db, sessionId)?.conversationId ?? null;
+  if (turnId && conversationId) {
+    const running = promptOfTurn(db, conversationId, turnId);
+    if (running && !answeredAfter(db, conversationId, running, turnId)) {
+      return { queued_in_turn: turnId };
+    }
+  }
+  return turnMetadata(turnId);
+}
+
+/**
+ * The oldest queued prompt no turn has claimed, among those typed after the
+ * conversation's last ordinary prompt. A queued prompt Claude Code folded into
+ * the running turn instead of running it as its own is never claimed, and a
+ * later ordinary prompt retires it, so a Stop long after cannot take it.
+ */
+function unclaimedQueuedPrompt(db: WritableDatabase, conversationId: string): MessageRow | null {
+  const row = db
+    .prepare(
+      `SELECT rowid, role, created_at, metadata FROM messages AS queued
+       WHERE conversation_id = ? AND role = 'user'
+         AND json_extract(metadata, '$.queued_in_turn') IS NOT NULL
+         AND json_extract(metadata, '$.turn_id') IS NULL
+         AND NOT EXISTS (
+           SELECT 1 FROM messages AS later
+           WHERE later.conversation_id = queued.conversation_id AND later.role = 'user'
+             AND json_extract(later.metadata, '$.queued_in_turn') IS NULL
+             AND (later.created_at > queued.created_at
+                  OR (later.created_at = queued.created_at AND later.rowid > queued.rowid)))
+       ORDER BY created_at, rowid LIMIT 1`,
+    )
+    .get(conversationId);
+  return (row as MessageRow | undefined) ?? null;
+}
+
+/** Give a queued prompt the id of the turn it ran as. */
+function claimQueuedPrompt(db: WritableDatabase, row: MessageRow, turnId: string): void {
+  const metadata = decodeMetadata(row.metadata);
+  metadata.turn_id = turnId;
+  db.prepare("UPDATE messages SET metadata = ? WHERE rowid = ?").run(
+    pythonJsonDumps(metadata),
+    row.rowid,
+  );
+}
+
 /**
  * The Stop hook (CV21.DS2.US8): log the turn's final answer beside its prompt.
  *
@@ -734,6 +814,9 @@ function turnIdOfRow(row: MessageRow): string {
  *     same -- unless a reply with that turn id already follows it, which is a
  *     repeat Stop (a Stop re-fired after another hook blocked). An assistant
  *     row without the turn id (a hand-logged summary) does not count;
+ *   * else, when the turn id is new, by the oldest queued prompt no turn has
+ *     claimed (`promptMetadata`): the turn a prompt typed mid-turn ran as.
+ *     Never for a turn the prompt hook skipped as a command;
  *   * else by position, when the conversation's last message is a logged
  *     prompt -- unless both sides carry a turn id and they differ, which means
  *     this turn's prompt was a command and an older prompt is waiting because
@@ -770,11 +853,11 @@ export function handleStopHook(
     const session = getRuntimeSession(db, sessionId);
     if (discardMarkerActive(session)) return { action: "skipped", reason: "discarded" };
     const turnId = turnIdOf(payload);
+    const commandTurn = (): boolean =>
+      Boolean(turnId) && decodeMetadata(session?.metadata ?? null)[SKIPPED_TURN_KEY] === turnId;
     const nothingWaiting = (): StopHookOutcome => {
       if (!turnId) return { action: "skipped", reason: "session_event" };
-      if (decodeMetadata(session?.metadata ?? null)[SKIPPED_TURN_KEY] === turnId) {
-        return { action: "skipped", reason: "command_turn" };
-      }
+      if (commandTurn()) return { action: "skipped", reason: "command_turn" };
       return { action: "skipped", reason: "nothing_waiting" };
     };
     const conversationId = session?.conversationId ?? null;
@@ -798,6 +881,18 @@ export function handleStopHook(
         );
         return { action: "logged", sessionId, pairedBy: "turn_id" };
       }
+      const queued = commandTurn() ? null : unclaimedQueuedPrompt(db, conversationId);
+      if (queued) {
+        claimQueuedPrompt(db, queued, turnId);
+        logAssistantMessage(
+          db,
+          sessionId,
+          message,
+          { interface: interfaceName, metadata: turnMetadata(turnId) },
+          deps,
+        );
+        return { action: "logged", sessionId, pairedBy: "queued_prompt" };
+      }
       if (last && last.role === "user" && !turnIdOfRow(last)) {
         logAssistantMessage(
           db,