Claude Code Haha — Renderer Crash + Sidecar BOM Bug Report

• Date: 2026-09-29 (CST)
• Reporter: kizemo (via MiniMax Code diagnostic session)
• App version: claude-code-desktop/0.6.7, Electron/42.7.0
• Sidecar binary: claude-sidecar-x86_64-pc-windows-msvc.exe (Tauri + Bun, ~115 MB)
• Affected repo: github.com/NanmiCoder/cc-haha

TL;DR

cc-haha shows a blank "出现异常。错误已记录到诊断日志。" error page on every launch when the user has at least one session whose transcript contains a malformed AskUserQuestion tool_use whose options[*].description is an object instead of a string. The root cause is a missing defensive render in the renderer that propagates ErrorBoundary to the whole tree; a side-channel bug (scheduled_tasks.json corrupted to NUL/BOM) is unrelated to the user-visible crash but produces noise in logs.

Two distinct bugs are reported in this document:

1. Renderer crash (primary, user-visible) — React error 调用工具有问题？ #31, {path, $text} hast object rendered as a child.
2. Sidecar JSON parsing (secondary, log noise) — CronScheduler fails to read scheduled_tasks.json when the file is empty / NUL-filled / BOM-prefixed.

1. Environment

2. User-Visible Symptoms

1. App launched; sidecar started; HTTP server bound to 0.0.0.0:64467.
2. WebSocket connected to session 91f665d5....
3. UI rendered most of the message history normally, then showed:
   • "出现异常。"
   • "错误已记录到诊断日志。"
   • "重试" button (no effect on click)
   • Doctor card + "重置安全 UI 状态" button (resets UI fields but does not clear transcript)
4. Renderer process repeatedly restarted: [renderer] [process-gone] reason=killed exitCode=1 → [renderer] [recovery-started] trigger=process-gone:killed → recovery completes → same error again.

3. Diagnostic Logs

3.1 ~/.claude/cc-haha/diagnostics/runtime-errors.log

[2026-09-28T23:49:24.266Z] ERROR client_react_error_boundary
summary: Minified React error #31; visit https://reactjs.org/docs/error-decoder.html
         ?invariant=31&args[]=object with keys {path, $text}
componentStack:
  at p / div / div / button / div / div / div
  at pU (App-DWep8kRD.js:345:388)
  at App-DWep8kRD.js:346:97765
  ...
  at u (ErrorBoundary-BnweEjNj.js:1:292)

(5 occurrences total, all with the same {path, $text} signature; timestamps cluster at 2026-09-28T23:49:24–23:49:42.)

3.2 ~/.claude/cc-haha/diagnostics/electron-host.log

[claude-server] [stderr] [CronScheduler] Error during tick:
  ApiError: Failed to read scheduled tasks:
    JSON Parse error: Unrecognized token '<NUL>'  (or '\uFEFF' after attempted repair)
  at readTasksFile (claude-sidecar.js:11102:18809)
  at async listTasks   (claude-sidecar.js:11102:17406)
  at async tick        (claude-sidecar.js:11231:3122)

[claude-server] [stderr] [CronScheduler] Error cleaning up stale runs:
  SyntaxError: JSON Parse error: Unrecognized token '...'
  at Fe5 (claude-sidecar.js:11230:1203)
  at e_8 (claude-sidecar.js:11230:1325)
  at async cleanupStaleRuns (claude-sidecar.js:11232:3904)

[renderer] [process-gone] reason=killed exitCode=1
[renderer] [recovery-started] trigger=process-gone:killed

3.3 ~/.claude/scheduled_tasks.json (corrupted)

Original (before any repair):

size: 713 bytes
hex: 00 00 00 00 00 00 ... 00 (all NUL)
JSON parse: ✗ Unrecognized token '\u0000'
last write: 2026-09-26 12:30:31

4. Root Cause Analysis

Bug A — Renderer crash on {path, $text} (PRIMARY)

A.1 Trigger

A model output produced a malformed AskUserQuestion tool_use where one option's description field is an object instead of a string. Sidecar's runtime validation rejects the call and emits InputValidationError:

<tool_use_error>InputValidationError: AskUserQuestion failed due to the following issue:
The parameter `questions[0].options[0].description` type is expected as `string`
but provided as `object`</tool_use_error>

A.2 Transcript evidence

Path: ~/.claude/projects/F--soft-00selfmade-filemanager/<session>.jsonl, line 515 (the offending assistant tool_use):

{
  "type": "tool_use",
  "name": "AskUserQuestion",
  "input": {
    "questions": [
      {
        "question": "上述 5 个文件确认删除吗？",
        "header": "删除确认",
        "options": [
          {
            "label": "确认删除全部 5 个",
            "description": {                            ← BUG: should be string
              "path": "`（一次一个，不用 wildcard）。CLAUDE.md 禁止 wildcard 删多个文件。",
              "$text": "我会逐个执行 `rm "
            }
          },
          {
            "label": "先别删，再想想",
            "description": "你会修改选择，或先备份再说。"   ← correct: string
          }
        ],
        "multiSelect": false
      }
    ]
  }
}

The description value is a hast (hypertext AST) node — {path, $text} is the diagnostic React error #31 gives when an mdast/hast object is mistakenly passed as a React child. The model appears to have wrapped part of its prose in an AST literal mid-stream instead of producing a string.

A.3 Renderer behavior

App-DWep8kRD.js:345:388 (pU) and App-DWep8kRD.js:419:4400 (mge) consume tool_use.input.questions[].options[].description and pass it directly to a markdown renderer. The renderer returns a hast tree; when this tree is placed inside JSX as a child, React throws #31: Objects are not valid as a React child (found: object with keys {path, $text}).

The top-level ErrorBoundary (ErrorBoundary-BnweEjNj.js:1:292) catches the throw and replaces the entire UI with the screenshot's "出现异常" page — including the unrelated Doctor card — making recovery without manually clearing transcript impossible from the UI.

A.4 Why it persists across restarts

desktop-ui.json (in ~/.claude/cc-haha/) stores pet.lastSessionId = "91f665d5-...". Every launch re-opens this session and re-renders the same line, triggering the same crash. The "重置安全 UI 状态" button only clears UI-state fields, not transcript content (per the error page's own copy: "聊天历史、模型配置、Skills、MCP、IM 和 OAuth 始终受保护").

A.5 Side note — type=queue-operation

The transcript contains {"type":"queue-operation","operation":"remove","timestamp":"...22:47:32.351Z"} immediately after the error result. This indicates the runtime flagged the tool_use as failed and removed it from the execution queue, but the transcript (and thus the renderer replay) still contains it.

Bug B — Sidecar CronScheduler JSON parse (SECONDARY)

B.1 Trigger

scheduled_tasks.json becomes an invalid JSON file (NUL-filled or BOM-prefixed). Each CronScheduler tick (every 60 s) tries to read it via:

async readTasksFile() {
  try {
    let q = await ud.readFile(this.getTasksFilePath(), "utf-8"),
        K = JSON.parse(q);
    if (!Array.isArray(K.tasks)) return {tasks: []};
    return K
  } catch (q) {
    if (q.code === "ENOENT") return {tasks: []};
    throw B6.internal(`Failed to read scheduled tasks: ${q.message}`);
  }
}

Path resolution:

function getTasksFilePath() {
  let q = process.env.CLAUDE_CONFIG_DIR || $e5.homedir() + "/.claude";
  return $e5.join(q, "scheduled_tasks.json");
}

With CLAUDE_CONFIG_DIR unset and USERPROFILE=C:\Users\Duanyi, this resolves to C:\Users\Duanyi\.claude\scheduled_tasks.json — which is mirrored/identical to the one under E:\Users\Duanyi\.claude\.

B.2 Reproduction

1. Overwrite scheduled_tasks.json with 713 NUL bytes (any cause — power loss during write, antivirus quarantine, fs corruption, etc.).
2. Launch app.
3. Observe CronScheduler log spam every 60 s.

B.3 BOM regression (during this user's attempted repair)

PowerShell 5.1's Set-Content -Encoding UTF8 writes a BOM (EF BB BF) by default. The user repaired the file with PowerShell; Bun's JSON.parse is strict and rejects BOM. The sidecar then logs Unrecognized token '\uFEFF' instead of '\u0000'. A re-write with System.Text.UTF8Encoding(false) (no BOM) clears the symptom. This is a Windows-onboarder footgun; not the user's fault.

5. Evidence Files (paths only, not contents)

~/.claude/cc-haha/diagnostics/runtime-errors.log           (timestamps: 2026-09-28T23:23..23:49)
~/.claude/cc-haha/diagnostics/electron-host.log            (process-gone, sidecar exit/entry)
~/.claude/cc-haha/diagnostics/diagnostics.jsonl            (consolidated structured events)
~/.claude/projects/F--soft-00selfmade-filemanager/<sid>.jsonl   (offending transcript, 521 lines)
~/.claude/scheduled_tasks.json.bak.20260929_073521          (713 NUL bytes, pre-repair snapshot)
~/.claude/scheduled_tasks_log.json.bak.20260929_073521      (20810 bytes, pre-repair snapshot)
~/.claude/projects/.../<sid>.jsonl.bak.20260929_084704      (transcript snapshot before description fix)

6. Suggested Fixes (upstream)

Fix A — Defensive rendering (PRIMARY)

In App-DWep8kRD.js around line 345 (pU) and 419 (mge), before passing option.description to any markdown / hast pipeline:

const desc = option.description;
if (typeof desc !== 'string') {
  // Render label only, or a sanitized placeholder, or the error reason.
  // Never forward a non-string into the markdown pipeline.
  return <p className="opt-desc-broken">{option.label} <em>(description unavailable)</em></p>;
}

Or, more conservatively: skip the option entirely from the rendered UI but keep the label, since the upstream model already received the validation error and won't act on it.

Also: do not let the ErrorBoundary replace the whole tree on a single tool_use render failure. Either scope it to the message bubble, or auto-recover by skipping the offending option in a componentDidCatch handler.

Fix B — Sidecar JSON hardening

1. Read scheduled_tasks.json with explicit BOM stripping:

   const text = await Bun.file(path).text(); // Bun auto-strips BOM
   // or, in Node-style: text = Buffer.from(buf).toString('utf8').replace(/^\uFEFF/, '');

2. On JSON.parse failure, treat as {tasks: []} rather than throwing — but log the bad payload to corruption-evidence.json (the existing diagnostics/corruption-evidence.json is already wired for this; just call it on parse error too).
3. On empty file (0 bytes), do NOT re-throw. Re-write a fresh {"tasks":[]} to disk and return that, instead of looping forever on parse errors.
4. Document that PowerShell Set-Content -Encoding UTF8 writes a BOM; ship a tiny cli-diagnostics note warning Windows users.

Fix C — Transcript schema validation at write time

When persisting an assistant tool_use into *.jsonl, validate that the shape matches the tool's JSON schema (Zod or hand-written). If a model produces description: {…} for AskUserQuestion, rewrite it to description: "" before persistence, so the transcript never carries a poison record.

Fix D — desktop-ui.json recovery

Add a checkbox to the Doctor card: "Skip last session on next launch" — sets a flag the renderer honors to land the user on the project list instead of auto-opening the crashing session. This unblocks users without requiring manual desktop-ui.json edits.

7. User Workarounds (until upstream fix)

1. Clear the offending tool_use from transcript: locate the <session>.jsonl line containing "description": { followed by "path" and "$text", replace it with "description": "..." (string) or delete the whole tool_use entry. Restart the app.
2. Edit ~/.claude/cc-haha/desktop-ui.json: change or remove pet.lastSessionId. Restart the app; it will open the project list instead of the crashing session.
3. Repair scheduled_tasks.json if it gets corrupted again: write {"tasks":[]} without a BOM. On Windows, use System.IO.File]::WriteAllText($path, $json, (New-Object System.Text.UTF8Encoding $false)) from PowerShell — not Set-Content -Encoding UTF8.

8. Environment Snapshot at Time of Diagnosis

electron host version: claude-code-desktop/0.6.7
chrome: 148.0.7778.280
electron: 42.7.0
node_pty: 1.23.251008001
sidecar stack: B:/~BUN/root/claude-sidecar.js (Bun runtime, source maps present in asar)
sidecar binary: resources/app.asar.unpacked/src-tauri/binaries/claude-sidecar-x86_64-pc-windows-msvc.exe (115,294,208 bytes)
db files:
  index-v1.sqlite       124,239,872 bytes
  search-index-v1.sqlite 188,358,656 bytes
  trace-index-v1.sqlite  136,228,864 bytes

9. Minimal Reproduction for Upstream

1. Add to a new session, ask the model to use AskUserQuestion with a Chinese-language description that contains backticks and code fragments.
2. Confirm model emits description: {path: ..., $text: ...} (object) instead of description: "..." (string).
3. Sidecar returns InputValidationError, model retries.
4. Renderer renders the failed tool_use from transcript on session reopen → React error 调用工具有问题？ #31 → ErrorBoundary → "出现异常".

A unit test could simply feed { description: {path: "x", $text: "y"} } into the pU component and assert no crash.

Prepared for upstream issue tracker. All sensitive content (auth tokens, user-specific paths) redacted.