kimaki 0.4.85 → 0.4.87

package/dist/queue-advanced-question.e2e.test.js

@@ -1,11 +1,14 @@
 // E2e test for question tool: user text message during pending question should
-// be consumed as the answer and NOT also sent as a duplicate promptAsync.
-// Reproduces the bug from commit a4dfb01 where the same message was sent twice.
-import { describe, test, expect } from 'vitest';
+// dismiss the question (abort), then enqueue as a normal user prompt.
+// The user's message must appear as a real user message in the thread, not
+// get consumed as a tool result answer (which lost voice/image content).
+import { describe, test, expect, afterEach } from 'vitest';
 import { setupQueueAdvancedSuite, TEST_USER_ID, } from './queue-advanced-e2e-setup.js';
 import { waitForBotMessageContaining, waitForFooterMessage, } from './test-utils.js';
 import { pendingQuestionContexts } from './commands/ask-question.js';
+import { store } from './store.js';
 const TEXT_CHANNEL_ID = '200000000000001007';
+const VOICE_CHANNEL_ID = '200000000000001017';
 async function waitForPendingQuestion({ threadId, timeoutMs, }) {
     const start = Date.now();
     while (Date.now() - start < timeoutMs) {
@@ -36,14 +39,22 @@ async function waitForNoPendingQuestion({ threadId, timeoutMs, }) {
     }
     throw new Error('Timed out waiting for question context cleanup');
 }
-describe('queue advanced: question tool text answer', () => {
+function setDeterministicTranscription(config) {
+    store.setState({
+        test: { deterministicTranscription: config },
+    });
+}
+describe('queue advanced: question tool answer', () => {
     const ctx = setupQueueAdvancedSuite({
         channelId: TEXT_CHANNEL_ID,
         channelName: 'qa-question-e2e',
         dirName: 'qa-question-e2e',
         username: 'queue-question-tester',
     });
-    test('user text message answers pending question without sending duplicate prompt', async () => {
+    afterEach(() => {
+        setDeterministicTranscription(null);
+    });
+    test('user text message dismisses pending question and enqueues as normal prompt', async () => {
         await ctx.discord.channel(TEXT_CHANNEL_ID).user(TEST_USER_ID).sendMessage({
             content: 'QUESTION_TEXT_ANSWER_MARKER',
         });
@@ -69,32 +80,17 @@ describe('queue advanced: question tool text answer', () => {
         });
         // User sends a text message while question is pending.
         // This should:
-        // 1. Answer the question via cancelPendingQuestion (consumed as answer)
-        // 2. NOT also send as a new promptAsync (the fix)
-        // 3. Clean up the pending question context
+        // 1. Dismiss the pending question (cleanup context)
+        // 2. Abort the blocked session so OpenCode unblocks
+        // 3. Enqueue the message as a normal user prompt (not consumed as answer)
         await th.user(TEST_USER_ID).sendMessage({
             content: 'my text answer',
         });
-        // Pending question context should be cleaned up after answer
+        // Pending question context should be cleaned up
         await waitForNoPendingQuestion({
             threadId: thread.id,
             timeoutMs: 4_000,
         });
-        // Wait for second question dropdown (from question-answer followup —
-        // OpenCode calls LLM again with same prompt after question tool completes,
-        // deterministic matcher fires question tool again). This is expected.
-        // Poll for it instead of sleeping.
-        const start = Date.now();
-        while (Date.now() - start < 4_000) {
-            const msgs = await th.getMessages();
-            const questionMsgs = msgs.filter((m) => {
-                return m.content.includes('Which option do you prefer?');
-            });
-            if (questionMsgs.length >= 2) {
-                break;
-            }
-            await new Promise((r) => { setTimeout(r, 50); });
-        }
         const timeline = await th.text({ showInteractions: true });
         expect(timeline).toMatchInlineSnapshot(`
         "--- from: user (queue-question-tester)
@@ -103,18 +99,96 @@ describe('queue advanced: question tool text answer', () => {
         **Pick one**
         Which option do you prefer?
         --- from: user (queue-question-tester)
-        my text answer
-        --- from: assistant (TestBot)
-        **Pick one**
-        Which option do you prefer?"
+        my text answer"
       `);
-        // The user's "my text answer" message must appear in the thread
+        // The user's message must appear in Discord
         expect(timeline).toContain('my text answer');
-        // Key regression assertion: without the fix, the user's text message
-        // is ALSO sent as a duplicate promptAsync which triggers a THIRD question
-        // dropdown. With the fix, only 2 dropdowns appear (initial + followup
-        // from question answer). Count occurrences of "Which option do you prefer?"
+        // Only 1 question dropdown — text message was consumed as the answer,
+        // no duplicate prompt was sent (which would trigger a second dropdown).
         const questionCount = (timeline.match(/Which option do you prefer\?/g) || []).length;
-        expect(questionCount).toBe(2);
+        expect(questionCount).toBe(1);
+    }, 20_000);
+});
+describe('queue advanced: voice message during pending question', () => {
+    const ctx = setupQueueAdvancedSuite({
+        channelId: VOICE_CHANNEL_ID,
+        channelName: 'qa-question-voice-e2e',
+        dirName: 'qa-question-voice-e2e',
+        username: 'queue-question-tester',
+    });
+    afterEach(() => {
+        setDeterministicTranscription(null);
+    });
+    test('voice message during pending question dismisses question and transcribes normally', async () => {
+        // This is the exact bug scenario: user sends a voice message while a
+        // question dropdown is pending. Voice messages have empty message.content
+        // (audio is in attachments, transcription happens later). The old code
+        // passed "" as the question answer and consumed the message — the voice
+        // content was completely lost.
+        await ctx.discord.channel(VOICE_CHANNEL_ID).user(TEST_USER_ID).sendMessage({
+            content: 'QUESTION_TEXT_ANSWER_MARKER',
+        });
+        const thread = await ctx.discord.channel(VOICE_CHANNEL_ID).waitForThread({
+            timeout: 4_000,
+            predicate: (t) => {
+                return t.name === 'QUESTION_TEXT_ANSWER_MARKER';
+            },
+        });
+        const th = ctx.discord.thread(thread.id);
+        // Wait for the question dropdown to appear
+        await waitForPendingQuestion({
+            threadId: thread.id,
+            timeoutMs: 4_000,
+        });
+        await waitForBotMessageContaining({
+            discord: ctx.discord,
+            threadId: thread.id,
+            text: 'Which option do you prefer?',
+            timeout: 4_000,
+        });
+        // Send a voice message while the question is pending.
+        // message.content is "" for voice messages — only the attachment exists.
+        setDeterministicTranscription({
+            transcription: 'I want option Alpha please',
+            queueMessage: false,
+        });
+        await th.user(TEST_USER_ID).sendVoiceMessage();
+        // Question context should be cleaned up (empty reply sent to unblock OpenCode)
+        await waitForNoPendingQuestion({
+            threadId: thread.id,
+            timeoutMs: 4_000,
+        });
+        // Voice content should be transcribed and appear as the next user message,
+        // processed after the model responds to the empty question answer.
+        await waitForBotMessageContaining({
+            discord: ctx.discord,
+            threadId: thread.id,
+            text: 'I want option Alpha please',
+            timeout: 4_000,
+        });
+        await waitForFooterMessage({
+            discord: ctx.discord,
+            threadId: thread.id,
+            timeout: 4_000,
+            afterMessageIncludes: 'I want option Alpha please',
+            afterAuthorId: ctx.discord.botUserId,
+        });
+        const timeline = await th.text({ showInteractions: true });
+        expect(timeline).toMatchInlineSnapshot(`
+        "--- from: user (queue-question-tester)
+        QUESTION_TEXT_ANSWER_MARKER
+        --- from: assistant (TestBot)
+        **Pick one**
+        Which option do you prefer?
+        --- from: user (queue-question-tester)
+        [attachment: voice-message.ogg]
+        --- from: assistant (TestBot)
+        🎤 Transcribing voice message...
+        📝 **Transcribed message:** I want option Alpha please
+        ⬥ ok
+        *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
+      `);
+        // Voice content must be present as a real transcribed message, not lost
+        expect(timeline).toContain('I want option Alpha please');
     }, 20_000);
 });

package/dist/queue-advanced-typing-interrupt.e2e.test.js

@@ -87,16 +87,22 @@ e2eTest('queue advanced: typing interrupt', () => {
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*
         --- from: user (queue-advanced-tester)
         PLUGIN_TIMEOUT_SLEEP_MARKER
-        [bot typing]
         --- from: assistant (TestBot)
+        ⬥ ok
+        *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*
+        [bot typing]
+        ⬥ ok
+        [bot typing]
         ⬥ starting sleep 100
         --- from: user (queue-advanced-tester)
         Reply with exactly: typing-stop-interrupt-final
         [bot typing]
         [bot typing]
+        [bot typing]
         --- from: assistant (TestBot)
         ⬥ ok
-        *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
+        *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*
+        [bot typing]"
       `);
         expect(finalUserIndex).toBeGreaterThanOrEqual(0);
         expect(finalReplyIndex).toBeGreaterThan(finalUserIndex);

package/dist/runtime-lifecycle.e2e.test.js

@@ -278,6 +278,7 @@ describe('runtime lifecycle', () => {
         --- from: assistant (TestBot)
         ⬥ ok
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*
+        ⬥ ok
         --- from: user (lifecycle-tester)
         Reply with exactly: seq-beta
         --- from: assistant (TestBot)
@@ -287,6 +288,7 @@ describe('runtime lifecycle', () => {
         Reply with exactly: seq-gamma
         --- from: assistant (TestBot)
         ⬥ ok
+        ⬥ ok
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
       `);
         expect(runtimeAfterC).toBe(runtimeAfterA);
@@ -324,7 +326,8 @@ describe('runtime lifecycle', () => {
         Reply with exactly: footer-check
         --- from: assistant (TestBot)
         ⬥ ok
-        *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
+        *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*
+        ⬥ ok"
       `);
         expect(footerMessage).toBeDefined();
         if (!footerMessage) {

package/dist/thread-message-queue.e2e.test.js

@@ -488,7 +488,6 @@ e2eTest('thread message queue ordering', () => {
         Reply with exactly: three
         --- from: assistant (TestBot)
         ⬥ ok
-        ⬥ ok
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
       `);
         const userThreeIndex = after.findIndex((message) => {
@@ -565,6 +564,7 @@ e2eTest('thread message queue ordering', () => {
         Prompt from test: respond with short text for opencode queue mode.
         --- from: assistant (TestBot)
         ⬥ ok
+        ⬥ ok
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
       `);
         const followupUserIndex = messagesWithFollowupFooter.findIndex((message) => {
@@ -632,7 +632,6 @@ e2eTest('thread message queue ordering', () => {
         Reply with exactly: BASH_TOOL_FILE_MARKER
         --- from: assistant (TestBot)
         ⬥ running create file
-        ⬥ ok
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
       `);
         expect(fs.existsSync(markerPath)).toBe(true);
@@ -815,10 +814,10 @@ e2eTest('thread message queue ordering', () => {
         Reply with exactly: echo
         --- from: assistant (TestBot)
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*
-        ⬥ ok
         --- from: user (queue-tester)
         Reply with exactly: foxtrot
         --- from: assistant (TestBot)
+        ⬥ ok
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
       `);
         expect(userEchoIndex).toBeGreaterThan(-1);
@@ -899,7 +898,6 @@ e2eTest('thread message queue ordering', () => {
         Reply with exactly: india
         --- from: assistant (TestBot)
         ⬥ ok
-        ⬥ ok
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
       `);
         const userIndiaIndex = after.findIndex((m) => {
@@ -991,7 +989,6 @@ e2eTest('thread message queue ordering', () => {
         Reply with exactly: november
         --- from: assistant (TestBot)
         ⬥ ok
-        ⬥ ok
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
       `);
         // E's user message appears before the final bot response

package/dist/voice-message.e2e.test.js

@@ -394,7 +394,8 @@ e2eTest('voice message handling', () => {
         🎤 Transcribing voice message...
         📝 **Transcribed message:** Fix the login bug in auth.ts
         ⬥ session-reply
-        *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
+        *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*
+        ⬥ ok"
       `);
         expect(finalState.sessionId).toBeDefined();
         // Verify OpenCode session received the transcribed voice message as a prompt
@@ -566,6 +567,7 @@ e2eTest('voice message handling', () => {
         --- from: assistant (TestBot)
         🎤 Transcribing voice message...
         📝 **Transcribed message:** Add error handling to the parser
+        ⬥ fast-response-done
         ⬥ session-reply
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
       `);
@@ -874,7 +876,10 @@ e2eTest('voice message handling', () => {
         [attachment: voice-message.ogg]
         --- from: assistant (TestBot)
         🎤 Transcribing voice message...
+        ⬥ fast-response-done
+        *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*
         📝 **Transcribed message:** Delayed transcription result
+        ⬥ fast-response-done
         ⬥ session-reply
         *project ⋅ main ⋅ Ns ⋅ N% ⋅ deterministic-v2*"
       `);

package/package.json

@@ -2,7 +2,7 @@
   "name": "kimaki",
   "module": "index.ts",
   "type": "module",
-  "version": "0.4.85",
+  "version": "0.4.87",
   "repository": "https://github.com/remorses/kimaki",
   "bin": "bin.js",
   "files": [
@@ -25,8 +25,8 @@
     "prisma": "7.4.2",
     "tsx": "^4.20.5",
     "discord-digital-twin": "^0.1.0",
-    "opencode-deterministic-provider": "^0.0.1",
     "opencode-cached-provider": "^0.0.1",
+    "opencode-deterministic-provider": "^0.0.1",
     "db": "^0.0.0"
   },
   "dependencies": {
@@ -66,9 +66,9 @@
     "xdg-basedir": "^5.1.0",
     "zod": "^4.3.6",
     "zustand": "^5.0.11",
-    "errore": "^0.14.1",
     "libsqlproxy": "^0.1.0",
-    "traforo": "^0.2.0"
+    "traforo": "^0.2.2",
+    "errore": "^0.14.1"
   },
   "optionalDependencies": {
     "@discordjs/opus": "^0.10.0",

package/skills/critique/SKILL.md

@@ -1,10 +1,9 @@
 ---
 name: critique
 description: >
-  Git diff viewer and AI reviewer. Renders diffs as web pages, images, and PDFs
-  with syntax highlighting. Also provides AI-powered diff reviews via
-  `critique review --web`. Use this skill when working with critique for showing
-  diffs, generating diff URLs, selective hunk staging, or AI code reviews.
+  Git diff viewer. Renders diffs as web pages, images, and PDFs
+  with syntax highlighting. Use this skill when working with critique for showing
+  diffs, generating diff URLs, or selective hunk staging.
 ---
 
 # critique
@@ -89,39 +88,6 @@ critique hunks add 'file:@-10,6+10,7'       # stage only your hunks
 git commit -m "your changes"                 # commit separately
 ```
 
-## AI-powered diff review
-
-`critique review --web` spawns a separate opencode session that analyzes a diff, groups related
-changes, and produces a structured review with explanations, diagrams, and suggestions. Uploads
-the result as a shareable URL — much richer than a plain diff link.
-
-**This command is very slow (up to 20 minutes for large diffs).** Only run when the user
-explicitly asks for a code review or diff explanation. Warn the user it will take a while.
-Set Bash tool timeout to at least 25 minutes (`timeout: 1_500_000`).
-
-Always pass `--agent opencode` and `--session <current_session_id>` so the reviewer has context
-about why the changes were made. If you know other session IDs that produced the diff, pass them
-too with additional `--session` flags.
-
-```bash
-# Review working tree changes
-critique review --web --agent opencode --session <session_id>
-
-# Review a specific commit
-critique review --commit HEAD --web --agent opencode --session <session_id>
-
-# Review branch changes compared to main
-critique review main...HEAD --web --agent opencode --session <session_id>
-
-# Review with multiple session contexts
-critique review --commit abc1234 --web --agent opencode --session <session_id> --session <other_session_id>
-
-# Review only specific files
-critique review --web --agent opencode --session <session_id> --filter "src/**/*.ts"
-```
-
-The command prints a preview URL when done — share that URL with the user.
-
 ## Raw patch access
 
 Every `--web` upload also stores the raw unified diff. Append `.patch` to any critique URL to get it:

package/skills/gitchamber/SKILL.md

@@ -0,0 +1,93 @@
+---
+name: gitchamber
+description: CLI to download npm packages, PyPI packages, crates, or GitHub repo source code into node_modules/.gitchamber/ for analysis. Use when you need to read a package's inner workings, documentation, examples, or source code. Alternative to opensrc that stores in node_modules/ for zero-config gitignore/vitest/tsc compatibility. After fetching, analyze files with grep, read, and other tools.
+---
+
+# gitchamber
+
+CLI to download source code for npm packages, PyPI packages, crates.io crates, or GitHub repos into `node_modules/.gitchamber/`. After fetching, analyze the files using grep, read, glob, and other tools to understand inner workings, find usage examples, read documentation, or study the source code.
+
+Alternative to [opensrc](https://github.com/vercel-labs/opensrc) that stores in `node_modules/` instead of `opensrc/`.
+
+**Differences from opensrc:**
+
+- **Stores in `node_modules/.gitchamber/`** instead of `opensrc/` -- automatically ignored by git, vitest, tsc, linters, bundlers, and every other tool that skips `node_modules/`
+- **No file modification** -- removed all `.gitignore`, `tsconfig.json`, and `AGENTS.md` editing logic
+- **No `--modify` flag** or permission prompts
+- **Zero config** -- opensrc requires updating `.gitignore` and `tsconfig.json` excludes; gitchamber needs nothing
+
+Always run `gitchamber --help` first. The help output has all commands, options, and examples.
+
+## Install
+
+```bash
+npm install -g gitchamber
+```
+
+Aliases: `gitchamber`, `chamber`
+
+## Fetch packages
+
+```bash
+# npm
+chamber zod
+chamber @babel/core
+chamber react@18.2.0
+
+# PyPI
+chamber pypi:requests
+chamber pypi:flask==3.0.0
+
+# crates.io
+chamber crates:serde
+chamber crates:tokio@1.35.0
+
+# GitHub repos (owner/repo, with optional branch or tag)
+chamber vercel/ai
+chamber facebook/react#main
+chamber owner/repo@v1.0.0
+chamber https://github.com/denoland/deno
+```
+
+Multiple at once:
+
+```bash
+chamber zod react vercel/ai pypi:requests
+```
+
+## Other commands
+
+```bash
+# List fetched sources
+chamber list
+chamber list --json
+
+# Remove specific packages
+chamber remove zod
+chamber rm vercel/ai
+
+# Clean all
+chamber clean
+chamber clean --npm      # only npm
+chamber clean --pypi     # only PyPI
+chamber clean --crates   # only crates.io
+chamber clean --repos    # only repos
+```
+
+## Output location
+
+Source ends up in `node_modules/.gitchamber/<host>/<owner>/<repo>/`.
+
+Tracked in `node_modules/.gitchamber/sources.json`.
+
+## How it works
+
+1. Resolves package via registry API (npm, PyPI, crates.io)
+2. Extracts `repository` URL from metadata
+3. Shallow-clones at the matching version tag
+4. Strips `.git` directory
+5. For npm: auto-detects installed version from lockfiles
+
+## When to use
+
+Use gitchamber (or its alias `chamber`) instead of manually cloning repos into tmp directories. Prefer it over the `lib-investigator` agent when you need persistent source code on disk for repeated reads or when the agent's context isn't enough. When the user refers to opensrc, always use gitchamber instead.

package/skills/goke/SKILL.md

@@ -178,9 +178,11 @@ Without a schema, all values stay as strings. `--port 3000` → `"3000"` (string
 | `[name]` in command | Optional argument |
 | `[...files]` in command | Variadic (collects remaining args into array) |
 | `<value>` in option | Required value (error if missing) |
-| `[value]` in option | Optional value (`true` if flag present without value) |
+| `[value]` in option | Optional value (`undefined` if flag present without value) |
 | no brackets in option | Boolean flag (`undefined` if not passed, `true` if passed) |
 
+**Optionality is determined solely by bracket syntax, not by the schema.** `[square brackets]` makes an option optional regardless of whether the schema is `z.string()` or `z.string().optional()`. The schema's `.optional()` is never consulted for this — it only affects type coercion. So `z.string()` with `[--name]` is treated as optional: if the flag is omitted, `options.name` is `undefined` even though the schema has no `.optional()`.
+
 ## Global Options and Middleware
 
 Global options apply to all commands. Use `.use()` to register middleware that runs before any command action — for reacting to global options (logging, state init, auth).

package/src/agent-model.e2e.test.ts

@@ -398,7 +398,8 @@ describe('agent model resolution', () => {
         Reply with exactly: agent-model-check
         --- from: assistant (TestBot)
         ⬥ ok
-        *project ⋅ main ⋅ Ns ⋅ N% ⋅ agent-model-v2 ⋅ **test-agent***"
+        *project ⋅ main ⋅ Ns ⋅ N% ⋅ agent-model-v2 ⋅ **test-agent***
+        ⬥ ok"
       `)
       expect(footerMessage).toBeDefined()
       if (!footerMessage) {
@@ -454,7 +455,7 @@ describe('agent model resolution', () => {
         Reply with exactly: system-context-check
         --- from: assistant (TestBot)
         ⬥ system-context-ok
-        *project ⋅ main ⋅ Ns ⋅ N% ⋅ agent-model-v2 ⋅ **test-agent***"
+        *project ⋅ Ns ⋅ N% ⋅ agent-model-v2 ⋅ **test-agent***"
       `)
     },
     15_000,

package/src/commands/ask-question.ts

@@ -49,6 +49,12 @@ type PendingQuestionContext = {
 const QUESTION_CONTEXT_TTL_MS = 10 * 60 * 1000
 export const pendingQuestionContexts = new Map<string, PendingQuestionContext>()
 
+export function hasPendingQuestionForThread(threadId: string): boolean {
+  return [...pendingQuestionContexts.values()].some((ctx) => {
+    return ctx.thread.id === threadId
+  })
+}
+
 /**
  * Show dropdown menus for question tool input.
  * Sends one message per question with the dropdown directly under the question text.
@@ -311,13 +317,21 @@ export function parseAskUserQuestionTool(part: {
 }
 
 /**
- * Cancel a pending question for a thread (e.g., when user sends a new message).
- * Sends the user's message as the answer to OpenCode so the model sees their actual response.
+ * Cancel a pending question for a thread.
+ *
+ * Two modes depending on whether `userMessage` is provided:
+ *
+ * - `cancelPendingQuestion(threadId)` — cleanup only. Removes the context
+ *   without replying to OpenCode. Use when aborting the blocked session
+ *   separately (e.g. voice/attachment messages whose content needs
+ *   transcription first). Returns 'no-pending' in both "found+cleaned" and
+ *   "nothing found" cases.
  *
- * Returns 'replied' if the question was answered successfully (caller should NOT
- * enqueue the user message as a new prompt — it was consumed as the answer).
- * Returns 'reply-failed' if reply failed (context kept pending so TTL can retry).
- * Returns 'no-pending' if no question was pending for this thread.
+ * - `cancelPendingQuestion(threadId, text)` — reply path. Sends the text as
+ *   the tool answer so the model sees the user's response. The caller should
+ *   NOT also enqueue the message as a new prompt.
+ *   Returns 'replied' on success, 'reply-failed' if the reply call fails
+ *   (context kept pending so TTL can retry).
  */
 export async function cancelPendingQuestion(
   threadId: string,
@@ -339,8 +353,9 @@ export async function cancelPendingQuestion(
   }
 
   // undefined means teardown/cleanup — just remove context, don't reply.
-  // The session is already being torn down. Empty string '' is a valid
-  // user message (attachment-only, voice, etc.) and must still go through.
+  // The session is already being torn down or the caller wants to dismiss
+  // the question without providing an answer (e.g. voice/attachment-only
+  // messages where content needs transcription before it can be an answer).
   if (userMessage === undefined) {
     pendingQuestionContexts.delete(contextHash)
     return 'no-pending'