[BUG] Claude Code Desktop: Critical Knowledge Gaps Causing Token Waste

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

chrome-native-host.log

App Version: Claude Code Desktop 1.11187.4
Model: Claude Sonnet 4.6
OS: Windows 11
Date: 2026-06-07
Session IDs: d0bd13ec-263e-4204-b9e8-fb1ed17c9aa2, 128a6da9-be90-43f8-9329-ba58c0dc72ee
Logs in attach

I am an advanced Claude Code user and have experienced significant frustration and token waste due to Claude's lack of knowledge about its own product — Claude Code Desktop.

Specific failures observed:

Claude does not know the Desktop UI. It referenced menus and options by their English names while my locale is set to pt-BR. It did not know where settings, session history, or feedback options are located in the Desktop interface.
Claude suggested non-existent commands. It instructed me to type /p in the Desktop chat — a custom command that was never registered in the interface and therefore not available to type. Claude did not know the difference between CLI slash commands and what is actually available in the Desktop UI.
Claude does not understand how Desktop sessions work. It incorrectly assumed that copying .jsonl transcript files to the correct folder would make sessions appear in the Desktop sidebar. Sessions are account-bound and this approach did not work. Claude spent many turns on this dead-end.
Claude does not know how skills/commands scope works in Desktop. It created a custom /start command for a personal project, not knowing that global skills from another open project (la-positiva) would override it. The wrong skill executed, loading incorrect context and wasting tokens.
Massive token waste. Repeated failed PowerShell commands, wrong approaches, and incorrect assumptions across this session consumed a significant portion of the user's monthly token quota with no productive outcome for several issues.
Root cause:
Claude's built-in knowledge about Claude Code Desktop — its UI, session model, account binding, command availability, and skill scoping — is insufficient for users who rely on it daily. This gap causes Claude to confidently give wrong instructions, which is worse than admitting uncertainty.

Request:
Improve Claude's baseline knowledge of Claude Code Desktop, specifically: how sessions are account-bound, how custom commands are exposed (or not) in the Desktop UI, how skill scoping works across multiple open projects, and how to guide users through the Desktop interface in their configured locale.

What Should Happen?

Claude should have accurate knowledge of Claude Code Desktop behavior, including:

Custom slash commands created in .claude/commands/ are only available in the CLI, not typeable in the Desktop UI chat input
Sessions in Desktop are account-bound — copying .jsonl files to the project folder does NOT make them appear in the sidebar for a different account
Skills installed globally (e.g. in .agents/skills/) are loaded across all open projects in the Desktop simultaneously, causing skill name conflicts
Menu and UI references should be provided in the user's configured locale (in this case pt-BR, set in config.json)
Claude should say "I don't know" when uncertain about Desktop UI behavior instead of confidently giving wrong instructions

Error Messages/Logs

[warn] [PluginScan] Skipping legacy command "brand-voice:discover-brand" - name collides with skills/ entry
→ Confirma conflito de comandos entre projetos (o mesmo bug que causou o /start errado)

[error] [Chrome Extension MCP] Failed to copy native host binary: Error: EBUSY: resource busy or locked
→ Erro recorrente de MCP Chrome Extension

[warn] [CCD CycleHealth] api_error: API Error: Usage credits required for 1M context
→ Erros de consumo excessivo de tokens

[warn] [RemotePluginManager] Failed to fetch enabled state; defaulting to all-enabled
→ Plugin scan falhou, causando carregamento incorreto de skills

Total warnings/errors no log: 1930 lines
Full log attached.

Additional error: Log file path provided by Claude was incorrect.
Claude instructed user to attach logs from:
C:\Users\eBao01\AppData\Roaming\Claude\logs\main.log
Path does not exist on user's machine.
Claude read log content via PowerShell during the session but provided
the wrong path when instructing the user to locate the file manually.
Correct path unknown — Claude could not confirm it without running
another command.

Steps to Reproduce

to Reproduce

1.Open Claude Code Desktop (v1.11187.4) with two projects in the sidebar: Project A (with a global skill named /start) and Project B (with a local .claude/commands/start.md)
2Navigate to Project B and open a new session
3Type /start — observe that Project A's skill executes instead of Project B's local command
4Ask Claude to create a custom command (e.g. /p) by creating a .md file in .claude/commands/
5Try to type /p in the Desktop chat input — observe that the command does not appear and cannot be invoked
6Ask Claude to make sessions from another account appear in the Desktop sidebar by copying .jsonl files to C:\Users\<user>\.claude\projects\<project-key>\ — observe that sessions do not appear (they are account-bound)
7Note that throughout steps 1–6, Claude gives confident but incorrect instructions without flagging uncertainty

Claude Model

Sonnet (default)

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

1.11187.4 claude code desktop

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

chrome-native-host.log

App version: Claude Code Desktop 1.11187.4
OS: Windows 11
Locale configured: pt-BR (confirmed in C:\Users\eBao01\AppData\Roaming\Claude\config.json)
Session IDs affected: d0bd13ec-263e-4204-b9e8-fb1ed17c9aa2, 128a6da9-be90-43f8-9329-ba58c0dc72ee
Log file: C:\Users\eBao01\AppData\Roaming\Claude\logs\main.log
Impact: Significant token waste across multiple turns due to incorrect guidance. User is on a token-limited plan and repeated failed attempts consumed a meaningful portion of the monthly quota.
Suggestion: Add Claude Code Desktop behavior (session model, command availability, skill scoping, account binding) to Claude's system-level knowledge for sessions opened within the Desktop app.