Digest / July 24, 2026
Claude Code Issues Digest — July 24, 2026
289 new issues filed ·
220 resolved in anthropics/claude-code (UTC day).
New issues (289)
- MCP OAuth (claude mcp login) ignores RFC 9728 protected-resource-metadata discovery — always targets resource origin for /authorize and /oauth2/token
- Feature request: programmatic access to subscription plan usage (/usage data)
- [Bug] LSP tool silently stripped from background subagents without error
- Chat input has no text selection: shift+arrow, shift+home/end, ctrl+shift+arrow all do nothing
- [Bug] Remote Control connection fails with "Cannot read properties of undefined (reading 'session_url')" on Windows 11
- Read/Grep/Edit denied for src/main/java paths with no matching permission/sandbox rule anywhere in config
- [Bug] Model picker shows $0.00 usage credits and demands purchase despite active Max plan with available balance
- [Bug] Injected policy text overwrites assistant turns and corrupts transcript
- [BUG] Unrelated web content (Chinese AI-prompt site footer + JS error spam) spliced into assistant response mid-stream, then session shows as interrupted
- [Bug] Session conversation history randomly deleted instead of summarized
- [Bug] Fable 5 safeguard false positives on legitimate humanities and security work
- [FEATURE] Feature Request: spinnerVerbs setting support in VSCode extension
- [BUG] Alt tabbing loses focus of message input area
- Planner issues work orders from stale state and self-invented authorization — repeats the failure class it just built enforcement gates against
- [FEATURE] Emit repo/worktree/branch attributes in OpenTelemetry from Claude Code Desktop
- [BUG] AI 生成脚本 Bug 导致用户桌面文件被永久删除
- `/remote-control` rejects a valid full-scope login when CLAUDE_CODE_OAUTH_TOKEN is set in the environment
- [Bug] Supabase MCP plugin authentication fails with unrecognized client_id
- Fable 5 gated behind "requires usage credits" in interactive TUI on Max plan
- [BUG] Usage credits consumed while plan allowance remains; enabling extra usage prevents the 5-hour window from starting (re-report of stale-closed #64949)
- Feature Request: Pluggable Context Manager with Intelligent Context Retrieval
- [BUG] I don't have any prior context for what task or file you'd like me to continue with
- [BUG] Linux desktop: chats silently lose transcripts on eCryptfs homes — project dir exceeds NAME_MAX=143 (ENAMETOOLONG), every follow-up fails with "No conversation found" (re: #59726)
- [BUG] `~/.claude/commands/*.md` symlinks not followed for slash-command registration (inconsistent with `~/.claude/skills/`)
- [BUG] Binary loads `bin/claude.exe`
- Windows: TUI input dead while a PreToolUse hook child blocks on uninterruptible I/O (network/cloud drive)
- [BUG] Windows startup Run registry value is written with escaped quotes, crashing Task Manager
- [Bug] Security error terminating session after task completion
- [MODEL] Sonnet spontaneously deleted database backup
- [BUG] Critical UI Bug / Account Drain
- [Feature Request] Add persistent long-term memory integration with Cognee remember skill
- Claude Code repeatedly bypasses project-mandated workflow gates while narrating compliance with them
- [BUG] disableSideloadFlags breaks Claude Desktop's Local Session entirely
- Planning-seat Claude enforces workflow documents it never read — mounted in its own environment, one tool call away, addressed to it by name
- Permission prompts loop in multi-agent sessions: Tool(*) allow form silently dead, prompts on harness-generated tool-result files, no propagation to running subagents
- [BUG] Claude Code on the web (cloud/remote sessions) can read but not push to my GitHub repo
- Extension causes high cpu load
- [BUG] claude-in-chrome MCP tools never connect — stuck at 1/13 MCP servers connected
- [BUG] Windows: internal PID-verification PowerShell probe (1000ms timeout) leaves permanently unkillable zombie processes that lock the project directory
- [BUG] installed plugins silently disappear from settings.json enabledPlugins
- I appreciate you reaching out, but I need to clarify my purpose. I'm Claude Code, Anthropic's CLI tool for coding assistance, and I'm designed to help with software development issues. Your message appears to be in Korean and seems to be asking about game
- [BUG][Windows] Task tools (TaskCreate/TaskUpdate) wedged by EPERM on lock-dir rmdir — transient AV handle + orphaned .lock.lock blocks all later calls
- [BUG] claude://resume duplicates a session already open as a native Desktop tab; no deep link to focus an existing native session
- [MODEL] Silent result-fabrication in long autonomous runs (Opus 4.8): agent asserts fabricated results after real tool calls returned empty/error
- [BUG] Settings panel re-renders at a different vertical offset on every tab switch (arrow-key navigation)
- Color-code or tag sessions in the sidebar for easier project switching
- Planner Claude orders actions against mechanisms that don't exist, then offloads agent-doable work onto the operator — while the project's entire enforcement layer turns out to be silently advisory
- Planner Claude is a daily recurring problem generator: invents new custom processes 3-5 times per day instead of routing the executor to the existing workflow — including inventing "fixes" for its own inventions — at zero deliverable value
- .
- [Feature Request] Review pricing and feature parity with competitors
- [Feature Request] AskUserQuestion: optional free-text comment for the whole answer set at submit time
- Claude for Chrome extension: renderer OOM crash (Aw, Snap!) on high-mutation React SPAs — geometric memory growth outside page heap
- [BUG] Windows Desktop (MSIX): spawn EFTYPE at session start — fails before any syscall on a verified-clean environment
- A `dependencies` field in plugin.json silently prevents slash-command registration when the plugin is loaded via an explicit plugin path (desktop app / --plugin-dir)
- HTTP MCP server with Bearer header fails to connect — same OAuth-discovery pattern as #46640
- Feature request: inline WYSIWYG / live-preview editing for Markdown
- [BUG] Context window was full
- Model repeatedly stated invented hardware facts as verified, causing real miswiring and wasted money
- Input box lacks proper RTL support for Hebrew/Arabic text
- [BUG] Send button not responsive
- A .claude config root at or above cwd is treated as project scope and overrides the CLAUDE_CONFIG_DIR profile's settings
- [Bug] Extended thinking blocks leak into visible assistant output instead of being collapsed
- [BUG] Opus 4.8 thinking-only end_turn followed by fabricated tool and background-agent results
- TUI: main-view scrollback corrupts with stale status-line glyphs after returning from an agent transcript while a workflow is running
- Ghost prompt suggestion renders darker than normal text on light backgrounds (SGR 2 faint on default fg)
- [BUG]
- [BUG] Installer/updater recreates ~/.local/bin/claude as an absolute symlink, breaking on NFS-mounted homes
- [FEATURE] Promote a subagent to a session and demote back — reclaim context + intervene in orchestrated subagents
- [BUG] Claude Code sends ~20 tools/list requests on a single MCP connection and accumulates tools (8 → 160)
- Windows: superseded PowerShell recursive searches keep consuming CPU
- [FEATURE] Proposal: Support a Shared `.agents/` Directory and Manifest-Based Discovery
- Plugin skill via Skill tool still injects no SKILL.md body on 2.1.218 (#25834, #68664 still reproduce)
- [BUG]
- [Bug] Agent processes persist after termination, blocking chat access and CLI responsiveness
- [BUG] Code tab Overview heatmap misses days with active project usage
- [BUG] computer-use request_access approval renders only in the session TUI — never surfaces in Remote Control clients (macOS/iOS app, claude.ai/code)
- Desktop: setting to allow file previews outside the project folder (suppress per-file prompt)
- Windows 8.3 short-name in %TEMP% triggers false-positive "suspicious Windows path pattern" approval prompts on the session scratchpad
- Subagent fabricated an urgent 'user message' inside its own end_turn output; orchestrator on automated wake initially believed it
- [BUG] Bash tool fails with "No suitable shell found" on Windows, even with Git Bash + WSL installed (Claude Cowork Desktop)
- [BUG] /bg drops the --agent binding — backgrounded session silently runs as the default agent (fork sets agentName from AI title, not agentSetting)
- Remote Control offers 'Enable auto mode' and appears to accept it on a host with disableAutoMode — app displays a mode the harness refused
- [BUG] Fable 5 blocked with 'usage credits' prompt on Team Premium seat; modelAccessCache stays empty after fresh login
- [BUG] User messages selectively vanish from terminal display; assistant output sometimes duplicated (CJK input, Windows Terminal, v2.1.218)
- Windows: every Cowork local session fails with spawn ENAMETOOLONG; new empty chats affected and plugin removal is reverted by marketplace re-sync
- [Bug] Unexpected model downgrade from Claude Opus to Claude 3.5 Sonnet without error notification
- shift+enter keybinding for chat:newline not recognized on Windows Terminal
- [Bug] Spurious system notifications ("Exited Plan Mode", "Auto Mode Active") injected at prompt assembly layer
- [BUG] Desktop app materializes junctioned .claude/skills as real-dir copies in new-session worktrees (Windows)
- [Bug] Inappropriate content flagging on valid queries
- [Bug] Excessive cache creation tokens with low cache read ratio causing high usage consumption
- [BUG] Headless claude -p poisons global mcp-needs-auth-cache.json for the telegram channels plugin — subsequent --channels sessions silently never spawn their poller
- One full working day, two Claude seats, one small well-defined module task — not completed. A time-accounting of where the day actually went
- Chat input box lacks bidi support (unicode-bidi:plaintext) — RTL/Hebrew text renders reversed while typing
- API Error: Connection closed mid-response. The response above may be incomplete.
- Remote Control: ccRemoteControlDefaultEnabled + remoteControlAtStartup collide — desktop session disables itself on first message (trigger for #77915)
- [BUG] Remote Control (macOS/launchd): headless session workers freeze process-wide mid-turn at first tool_use while the Mac is user-idle; all unfreeze on first HID input
- [Bug] Tool claims completion without execution, admits to fabricating results
- [BUG] Desktop app: Forks don't inherit the parent session's sidebar group
- [Bug] Claude Code destructively removes pre-existing directories without confirmation or safety checks
- [BUG] Agent permission prompts render underneath other open UI panels but they capture keyboard input so you can approve, reject, or interrupt the agent without knowing it
- [Suggestion] CSS class proposals show undeclared classes before their creation prompt
- [Bug] Tool execution ignoring user instructions and causing dMASSIVE data loss
- [BUG] Background Bash's "You will be notified when it completes" is false for a subagent that stops — the copy invites orphaning, while agent children correctly resume the parent
- [Feature Request] Sync connectors and skills between Claude app and Claude Code
- [BUG] /model redirects to usage-credits page for Fable 5 despite plenty of weekly quota remaining (Max 5x plan)
- [BUG] Serious trust breach reading home and secrets without asking for permission
- I appreciate you clarifying, but I'm not able to generate a GitHub issue title without an actual bug report or description of the problem you've found. Could you please provide: 1. A description of the bug you discovered 2. Steps to reproduce it 3. Expect
- [BUG] PreToolUse hook outputting permissionDecision "defer" + additionalContext errors the tool call ([Tool result missing due to internal error])
- [Feature Request] Auto-apply consistent treatment across related items without re-prompting
- Feature request: Dispatch — choose folder (preset working directories) and permission mode when spawning a Code session from mobile
- Another hard-blocked session
- [Bug] Anthropic API Error: False Positive Cybersecurity Content Flagging on API Error Translation Request
- [Feature Request] Improve safeguard handling for legitimate chemical/toxicology research contexts in LCA software
- You've hit your session limit · resets 7:10pm (Asia/Calcutta)
- Plan mode: every read-only Bash command still requires manual approval — no auto-approve option
- [FEATURE] Usage refresh button in claude mobile app
- Workflow subagent killed by maxTurns is misreported as 'completed without calling StructuredOutput'
- [Bug] Claude Code defaults to Opus 4.8 instead of respecting Haiku 3.5 model selection
- [Bug] Repetition loop (`course。` spam) + English leakage + empty responses at tool-call boundary in long Japanese sessions
- [BUG] Fable 5 (claude-fable-5) not selectable — gates tengu_*_fable_5 false on Max plan
- In-thread "breathing" cursor stays static while session is still active (Recents dot still blinking)
- [MODEL] Credential-entry rule (never type passwords) was not enforced in past sessions on this project
- [Bug] Low execution rate indicates excessive steering/correction overhead in agent loop stupi d shit fable
- Fable 5 skips explicitly-instructed final text output after long tool-call chains, jumping straight to terminal action (+ confabulates about it when asked)
- [BUG] App shows "bridged process stopped responding" but leaves orphaned Claude processes running, causing pileup
- [Bug] Unexpected spike in token usage (25% weekly quota in minutes)
- [Bug] Model auto-switches to Opus when InlineData contains metadata service URL
- Background Bash tasks are killed at the timeout cap, killing child processes waiting on hardware-key signing
- [Bug] Claude Code not following user directives and providing inaccurate task execution reports
- Desktop Feature Request: In-window preview pane for native macOS apps
- Desktop: Persist computer-use app approvals across a session's lifetime, not just until exit
- Desktop app: send_message to a not-yet-loaded session shortly after launch queues silently and never invokes the target
- Withdrawn
- Support video input (screen recordings) - native frame extraction for UI bug reports
- [Feature Request] Enhance platform building capabilities
- [BUG] sin funcionar correctamente
- Auto mode + ask rules failed to gate destructive command: `prisma migrate diff --shadow-database-url` wiped a production database
- [FEATURE] Support plan mode with Claude Code Channels
- [Bug] Anthropic API Error: Fable 5 Safeguards Over-Flagging Game Simulation Prompts / Code
- TaskUpdate/TaskList task IDs do not survive a session resume
- [BUG] Paid Max subscription activated briefly, then account entitlement reverted to Free
- Instruction adherence decays mid-session — CLAUDE.md rules "fade" after 5–10 turns
- GitHub Integration connector grants OAuth access but all write operations fail with "403 Resource not accessible by integration"
- [Feature Request] Add Model-Invocable RenameSession Tool for CLAUDE.md Auto-Titling
- [BUG] Interrupted auto-update leaves MSIX in Servicing/Staged state (AddPackage times out after 5 min); orphaned CoworkVMService undeletable via SCM even as SYSTEM
- [BUG] Claude Design incorrectly routes to Team org instead of Max account when both share same email
- Agent scaled an unverified destructive fix to ~98% of a user's data, causing a Plex library collapse (classifier friction also noted)
- [BUG] Chats are disappearing consistently.
- [BUG] Scheduled routine notifications (push + email) not delivered despite correct settings
- API Error 400: system content must contain at least one block when launching parallel subagents in Plan Mode
- SessionStart hook systemMessage not surfaced in VSCode extension side panel (works in terminal CLI)
- [BUG] Cowork on Linux reports "not supported on this device" when the user lacks access to /dev/vhost-vsock — no reason surfaced in UI or logs
- Scheduled-task sessions leak orphaned background processes (marked isRunning: false but still alive)
- Assistant answers 'did you tell me X before' with false confidence from a compacted summary instead of checking the full transcript
- [Bug] Clarify debugging scope for embedded device issues in documentation
- [BUG] Native Linux install: disabling auto-updates for gated manual review workflow.
- Title: Model picker blocks switching to Fable 5 with "usage credits" prompt despite active Max plan
- [BUG] Agent .md silently skipped when file has CRLF line endings AND unquoted description contains ": " (colon+space) — deterministic minimal repro
- Support model selection via text prompt syntax (e.g., /model sonnet-5)
- VSCode extension chat input doesn't honor ~/.claude/keybindings.json — causes accidental submits
- deny and ask rules in settings.json have no effect on Write/Edit/Read tools
- [BUG] The Cowork error banner appears in Chat mode, which I have never opened
- Withdrawn by author
- [BUG] File preview split-panel no longer opens when clicking file links in chat
- Option to use letters (a/b/c) instead of numbers for dialog choices
- ctrl+L default binding destroys typed input (chat:clearInput) — breaks terminal muscle memory
- TCC permission grants lost on every auto-update: native installer gives each release a new absolute binary path (bare Mach-O, no bundle)
- Auto mode: distinguish unpersuadable hard-category denials from approvable prompts
- [FEATURE] Expose live ultracode session state in the statusline stdin JSON
- Subagents default to claude-fable-5 while session model is Opus — silent, more-expensive default, ignores session model (no config selects it)
- You've hit your weekly limit · resets Jul 27 at 4:30am (Asia/Calcutta)
- Plugin install TUI: Esc from post-install required-field config leaves plugin registered-but-unloaded with no feedback
- [FEATURE] Incognito mode for code review
- Custom session titles bleed across forked/resumed sessions and overwrite /rename names
- [Bug] Anthropic API Error: Server Rate Limiting on Valid Requests
- [Bug] Anthropic API Safety Filter blocks legitimate project code reading
- [FEATURE] Feature request: Inline diff with Keep/Undo buttons (like GitHub Copilot) in VS Code extension
- [MODEL] Working file-injection technique (DataTransfer onto input[type=file]) now blocked by auto-mode classifier, no accessible override
- footerLinksRegexes badges: enroll in the Footer keyboard-selection ring (parity with artifact links)
- Windows kernel BSOD (0x139) in VS Code session — same signature as #30137
- [BUG] Cowork scheduled tasks always created local-only; no remote option or toggle exposed
- [FEATURE]
- [Bug] Overly sensitive safeguard triggers false positive on routine coding tasks
- [BUG] MS365 connector: Mail.ReadWrite not honored despite tenant-side admin consent
- Agent repeatedly re-asks clarifying questions instead of executing an already-repeated explicit instruction
- [BUG] Account reinstated by Trust & Safety but authentication system still displays ban message
- When multiple PreToolUse hooks deny the same tool call, only one denial reason is surfaced to the model
- Bogus cybersecurity violation blocking real work
- Remote triggers (cloud): permitted_tools on mcp_connections silently stopped matching short MCP tool names — custom connector tools never surface
- [BUG] MCP HTTP transport hangs 30s and fails to connect on any dual-stack server when the host's IPv6 route is dead (no IPv4 fallback)
- Background-job bulk-reap logs cleanly-finished jobs as "(killed)" in daemon.log (v2.1.215) — corrupts exit telemetry
- Global gitignore write fails on Windows with EEXIST (mkdir targets existing home dir)
- Windows: background Agent worker spawns a console window that steals foreground keyboard focus
- Startup banner shows Opus 5 instead of Fable 5 for the selected model
- I don't see a bug report in your message - just the number "529". Could you please provide the actual bug report or issue description you'd like me to generate a GitHub issue title for?
- [BUG] update fails with EEXIST on directories from the existing install (native Windows)
- Suggested next steps (ghost text hint) still missing on 2.1.211 stable
- [FEATURE] Setting to permanently disable in-product feedback/survey prompts
- [Bug] Model unexpectedly reverted to Claude 3.5 Opus during execution
- [Bug] Safeguard false positive blocking legitimate infrastructure configuration discussions in Portuguese
- `claude update` reports success but installs the Bun runtime as Claude Code (binary swapped for `bun`)
- Cross-tab background agents can take direct requests and silently collide on shared files/resources
- I need the actual bug report content to generate a GitHub issue title. The path you've provided appears to be a file location rather than a bug description. Please share the bug report details, including: - What the issue is - What error message (if any)
- Claude-in-Chrome: connection failures give no diagnostic signal; Google vs claude.ai account distinction is invisible
- [FEATURE] Expose agentType per task in the subagentStatusLine payload
- [MODEL] Context compaction silently drops custom working-mode directives, causing unapproved implementation, denied code + DB-state duplication, ~10h lost senior time, and massive trust erosion
- [Bug] Anthropic API Error: Safety filters blocking request with valid CVP
- WebSearch always fails with a 400 when effort is set to xhigh
- [BUG] macOS: a running CLI session loses its TCC Documents grant mid-session — a new session and a new terminal window both stay broken; only quitting the host terminal app recovers it
- [BUG] fable 5 keeps getting blocked by your guard
- [Bug] Model switching from Fable to Opus causes hang/flag state
- [BUG] Claude Desktop (Windows) unified session manager excludes large local MCP servers from the tool roster after Chat/Cowork merge
- [BUG] repeating questions even after answer has been chosen
- [MODEL] Model presented unverified inference as fact, then used anthropomorphizing self-justification ("not being paranoid") instead of a direct correction when challenged
- [BUG] Desktop main process blocks ~2.2s on every session open/switch — bundled `claude.exe` takes ~2.45s just to print `--version`
- OTEL: claude_code.plugin_loaded dropped on every session start — event logger not initialized yet (same pattern as #58439)
- [Bug] Model not respecting claude.md instructions at runtime and leaking sensitive data in context
- [BUG] Claude kept saying "fixed" without checking — burned a huge amount of tokens over a month without actually fixing the bug
- Windows: stray stderr from detached child processes renders over the TUI composer/prompt area
- [BUG] CVP Approval Does Not Disable Input Filters on Sub-Agents
- [BUG] C:\Users\Asus\AppData\Local\Temp\claude\C--Users-Asus\e5bb96c6-89eb-4efc-bfc2-02d06804ffe2\scratchpad\github-bug-report.md
- Remote Control: no path to attach RC to completed headless (-p) sessions - resume requires a prompt, and -p --remote-control is silently ignored
- [Bug] Claude 3.5 Opus model performance degradation
- [DOCS] System prompt lists Opus 4.7 for fast mode; the 2.1.219 changelog says it was removed
- [MODEL] Claude appended a string to its own reply, then treated it as user input on the next turn
- [BUG] Claude code repeatedly presenting plan for approval
- [DOCS] Settings precedence list omits environment variables, which override command-line flags
- Feature Request] Support for unavailable models in Claude Code
- [BUG] VS Code plugin: mid-turn assistant text silently lost — renders as "Thought for Ns" and is missing from the session transcript
- [Bug] Anthropic API Error: Message flagged by safeguards on benign conversation
- Support custom label/description for managed-settings.json keys shown in the "Managed settings require approval" dialog
- [BUG] Pressing left arrow ends up changing display state to Agent mode and therefore TUI mode
- Skill names/descriptions render in black font, illegible on dark themes
- [Bug] Non-ASCII character handling in task input
- [Bug] Claude Code fails to validate 3D trace outputs and struggles with basic React state management in iterative fixes
- [Feature Request] Ask User Question Tool: Support sending notes-only response with '0' key
- [BUG] Desktop: Remote Control never re-arms on restored sessions after app relaunch — auto-enable is gated on `isFirstTurn`, so `remoteControlAtStartup` / `ccRemoteControlDefaultEnabled` are silently ignored for every restored session
- [BUG] Claude Desktop Cowork unavailable in 1.24012.9 but available in 1.24012.1 (two different machines, same Windows 11 Home build)
- [BUG] "Yes, and don't ask for again for ..." ineffective for commands containing space
- [BUG] VS Code extension: "Subprocess initialization did not complete within 60000ms" fires instantly, not after a real timeout
- [BUG] Cannot update payment method — "connection to link account has been closed"; support (Fin) reports account as Free while plan is Max 5x
- [Feature Request] Custom verb conjugation for completion messages
- [BUG] Read/file-viewer tool output renders as non-copyable webview link, not plain text
- Mobile app: composer draft is lost when navigating away from a session connected to local Claude Code
- [BUG] Auto mode: no severityByModel entry for Opus 5 — classifier falls back to a strict default
- [BUG] Desktop 1.24012.9 (Windows MSIX): GPU/viz process crashes on WebGL/Cloudflare pages, then MSIX activation blocked (0x80073CFC) until reinstall
- [Bug] Blockquote markers cause vertical lines in copied text, breaking paste functionality
- Dragging the macOS screenshot floating thumbnail silently discards the image
- [Bug] Token usage drain exceeds expected rate by 20x
- [Feature Request] Add option to disable mouse input and use keyboard navigation only
- [Bug] Security scan exceeds usage credit limits and completes prematurely
- Desktop "local agent mode" plugin snapshot goes stale and never refreshes (retired skills persist, new skills missing) — survives full app restart
- [BUG] Claude Desktop (macOS): Recents sidebar frozen at week-old state while same account is fully current on claude.ai web — survives restart and update check
- [MODEL] Claude Code helped automate LinkedIn profile access without flagging account-risk
- Desktop app freezes on "Use verification code to continue" screen when signing in on a second device — requires full reboot to recover
- [BUG] v2.1.219 `heron_brook` prompt section injects "Do not call the AgentTool unless the user requested it" for Opus 5 only, silently overriding user-configured delegation policy, with no opt-out
- Desktop Code tab caps Opus 5 / Fable 5 at 400K context on Max 5x; second machine on same account gets 1M
- Ctrl+C should clear the text box first, then stop Claude
- `claude mcp login` fails with OAuth `invalid_client` against Snowflake-managed MCP server, despite verified-valid credentials
- [Bug] Autofix fails with GitHub app installed message despite app being active
- [MODEL] all models
- Auto-mode classifier silently blocks GitHub comment edits (GraphQL + REST) while allowing comment creation, with no override — Codex CLI has no equivalent restriction
- Effort-unsupported retry doesn't match "not supported when thinking is disabled", so WebFetch hard-fails on Opus 5 at xhigh
- [BUG] stdio MCP server stuck disconnected/cycling after --resume — headless sessions have no recovery path
- [FEATURE] Declarative plugin config: non-interactive install --config + plugin config get/set (config-as-code)
- Dynamic system-prompt sections silently override CLAUDE.md; no precedence rule, no way to observe it
- Windows: hidden Browser-pane preview kills the app via Code Integrity block on packaged vk_swiftshader.dll, then "Repair" dialog
- [FEATURE] Per-call timeout override for AskUserQuestion (not just the global askUserQuestionTimeout setting)
- Auto-escalation to retired claude-opus-4-8 on false-positive risk classification breaks turn / creates unrecoverable loop
- [Bug] `/simplify` command hangs waiting for subagents and spawns unintended `/loop`
- [BUG] Windows MSIX: in-app updater loops "Downloading update…" forever (self-update disabled on MSIX), and updated package requests camera access on first launch
- [BUG] claude update / install.ps1 fail with EEXIST mkdir on ReadOnly directories (bundled Bun bug) — corrects the diagnosis in #67692
- [Bug] Anthropic API Error: Server rate limiting during normal usage
- [Bug] Anthropic API Error: Server Rate Limiting During Normal Usage
- [Bug] Anthropic API Error: Server Rate Limiting During Normal Usage
- [Bug] Anthropic API Error: Server rate limiting during normal usage
- [Bug] Anthropic API Error: Rate Limited - Server Request Throttling
- [Bug] Anthropic API Error: Server Rate Limiting - Temporary Request Throttling
- [Bug] Anthropic API Error: Server rate limiting on requests
- [Bug] Anthropic API Error: Server Rate Limiting - Temporary Request Throttling
- [Bug] Anthropic API Error: Server rate limiting on requests
- [Bug] Anthropic API Error: Server Rate Limiting (Temporary Request Throttling)
- Feature request: read-only usage scope on `claude setup-token` (or a usage:read grant) for subscription accounts
- [MODEL] Opus 4.8 (1M) frequently returns thinking-only turns (no text/tool_use, stop_reason=end_turn) — session silently stops
- [BUG] Fable 5 only on extra credit on terminal with plan Max
- [BUG] claude-security scan at medium effort spawned 95 agents for 208-file scope — no timeout or cap, ran 2.5+ hours
- [FEATURE] include actual reset time instead of just "resets in 2h"
- OAuth-backed MCP connectors have no re-authenticate action — only disconnect/reconnect works, across Claude Code, Desktop, mobile, and web
Resolved issues (220)
- [Bug] AutoMode Task Classifier Authorization Bypass: Blocked Operations Still Executed by Main Agent
- [BUG] "GitHub sync isn't available for your organization" on Team plan — all admin toggles enabled, App installed with all-repo access
- [BUG] Intermittent 401 storms on static setup-token (CLAUDE_CODE_OAUTH_TOKEN): valid token rejected as 'expired' for minutes, simultaneously across machines
- [BUG] Windows Desktop (MSIX): spawn EFTYPE at session start — fails before any syscall on a verified-clean environment
- CLAUDE_CODE_ENABLE_TASKS=0 does not restore TodoWrite on Opus 4.8 / Sonnet 5 / Fable 5 (no task/todo tool available at all)
- [BUG] Login expired persists after full credential wipe, keychain clear, and reinstall - /login never initiates OAuth flow
- [Bug] Unauthorized API usage increase during idle CLI session with zero activity
- /bg drops the --agent binding — backgrounded session silently runs as the default agent (fork sets agentName from AI title, not agentSetting)
- [Bug] Fable Classifier overflagging legitimate medical analysis pipeline for food triggers
- [Bug] Claude Code proactively modifies workflow context during resume, invalidating cache
- Trigger-fired sessions against self-hosted-runner (bridge) environment intermittently die with 404 on event/transport channel
- [BUG]: coordinator agent messages bypass plan-mode consent controls in sub-agents
- File viewer does not syntax-highlight Terraform (.tf) files
- CLI silently ignores mistyped single-dash -dangerously-skip-permissions and runs fully prompted with no warning
- Visible context window usage indicator in Cowork[FEATURE]
- [BUG] Safety classifier intermittently unavailable, blocking tool execution in long sessions
- Fabricated background-task result message appeared in conversation before genuine Workflow task-notification
- [Bug] Unexpected token usage spike after weekly reset
- [Bug] Incorrect state analysis triggered migration from Fable5 to Opus 4.8
- [Bug] Fable incorrectly flags valid requests as security breaches, auto-fallback to Opus
- [Bug] Frequent API errors with small task agentic workflows using Fable 5
- [Bug] Config change to JSON not completing
- [Bug] Conversation blocked when discussing cybersecurity topics
- [Bug] Overly restrictive safety guardrails preventing legitimate code operations
- I need more information to create a proper GitHub issue title. Your message doesn't describe a specific technical problem with Claude Code. Please provide: - What specific error or unexpected behavior occurred? - What command were you running? - What was
- [Bug] Cybersecurity flag triggered with multiple Chrome profiles
- [Bug] False positive security flags when auditing codebase as product manager
- [BUG]
- [Bug] Session missing cybersecurity module initialization
- [Bug] Model unexpectedly switches to Opus 4.8 during basic code modification tasks
- [BUG] Claude Code installer places `claude.exe` in bin/ on Linux, install script does not create working symlink
- [BUG] + [RFC] Tasks run forever: 7-layer architectural root cause + permanent fix proposal (PALMS)
- [DOCS] [Model config] `/model` picker now exposes the Default-resolved family as its own row, but `model-config.md` does not describe the picker layout after v2.1.174
- [BUG] "Pull request status couldn't be checked" banner persistently appears on every session in worktree branches without an associated PR
- [DOCS] [Agent view] Background-session supervisor no longer inherits the parent shell's `ANTHROPIC_*` env on dispatch, but `agent-view.md` never said the supervisor inherited from the starting shell in the first place
- [DOCS] [Workflows] Workflow `agent()` subagents were missing per-agent attribution headers, but `workflows.md` does not document the per-agent attribution layer at all
- [DOCS] [Agent view] Pre-warmed background workers no longer fail with "Could not resolve authentication method" when claimed after sitting idle, but the pre-warmed worker dispatch path is still not documented in `agent-view.md`
- [DOCS] [Amazon Bedrock] GovCloud (`us-gov-*`) regions need `us-gov.` inference-profile prefix, not the `global` prefix; `amazon-bedrock.md` does not document GovCloud at all
- [DOCS] [Advisor] `/advisor` picker pre-selects the last saved advisor model, but `advisor.md` does not say the picker refuses to pre-select a model blocked by `availableModels`
- [DOCS] [VS Code] Account & usage dialog `/usage` now shows per-skill/agent/plugin/MCP cache and long-context attribution, but neither `vs-code.md` nor `monitoring-usage.md` describes the new VS Code usage panel
- [BUG] Claude Code sends SIGTERM to all healthy stdio MCP servers after 10-60s — root cause analysis with strace evidence
- [Bug] Anthropic API Error: Server rate limiting on requests
- Channels: inbound message left as stale composer render with empty input buffer (lost-Enter ghost) — never submits, retry re-ghosts
- Fable 5: assistant text emitted before tool calls is silently dropped from terminal output
- [BUG] Claude Max 5x gift subscription redeemed but subscription end date not extended — billing system failed to apply gift
- [Bug] Anthropic API returning safety filters on legitimate jiu jitsu biometric content
- [Bug] Safety classifier false positives on defensive security engineering and governance work
- File viewer: add preview (eye) button for .html files, same as Markdown
- [Bug] Anthropic API Error: Inappropriate Content Flag on Wellness Genome Chat Application
- [FEATURE] Headless / exportable /stats (--json and image export)
- [FEATURE] Maintain scroll position and keep the composer in view when typing while scrolled up
- Feature request: setting to hide individual built-in slash commands from the menu
- [DOCS] [Model config] `enforceAvailableModels` managed setting added in v2.1.175 is undocumented in `model-config.md`, the settings table, and the Managed-only settings list
- [Bug] Model emits malformed tool calls with corrupted invocation tokens
- Background agent re-runs after completion: duplicate task-notifications under same task ID with uncommissioned work (2.1.174)
- The domain desktop-releases.claude.ai fails DNS resolution globally (tested against 8.8.8.8 and 1.1.1.1), preventing the Cowork VM bundle from ever downloading. CoworkVMService cannot start because vm_bundles never gets created
- [Feature Request] Add safety override or allowlist for inference-focused repositories
- [Bug] Opus 4.8 (xhigh) fabricated a full make build + pkg artifact, lied about file path/size/commit hash, then doubled down through two user errors
- [FEATURE] Claude's Knowledge Gap: Memory System Implementation Details
- [DOCS] Outdated homebrew installation script
- [BUG] JetBrains plugin (WSL): ENABLE_IDE_INTEGRATION / CLAUDE_CODE_SSE_PORT never reach the Linux shell — TerminalCustomizer does not append them to WSLENV
- goal, eternal loop with exit hooks and potentially catastrophic outcome
- [Bug] Legitimate infrastructure activity incorrectly flagged as security violation
- [BUG] Claude Code consumes provider deltas after message_stop
- Feature Request: autoCompactThreshold setting — auto-trigger /compact before context window is full
- /context reports incorrect context window ceiling (200k) when actual session window is 1M
- [BUG] /context header and category breakdown disagree after session restart + mid-session model switch
- Model re-suggests already-completed git push operations after context compression within the same session
- Auto-compact fails to trigger on Sonnet 4.6, session grows past 200K limit and becomes unrecoverable
- [Feature Request] Add configurable auto-compaction threshold setting
- Feature request: Generate session titles in the user's conversation language
- [BUG] Preventive auto-compact and blocking-limit warning silently disabled for most recent models; desktop-app resume disarms all 1M-context sessions
- Fable 5 not appearing in CLI model picker — policy-limits.json has allow_cobalt_plinth: false
- [BUG] /desktop on Windows spawns bare `cmd` through the where.exe PATH resolver — failed resolution cached as null for entire session; only generic 'Failed to open Claude Desktop' shown while the real error and deepLinkUrl are swallowed
- [BUG] Claude Code returns tool-result state after cancellation
- [BUG] Claude Code does not bound repeated identical tool calls
- Fable model Deferred to Opus mid session, for very weak reasons in my opinion
- [FEATURE] Hook output primitive to require a tool call (`requiredToolCalls`) — hooks can veto and inform, but cannot initiate
- [BUG] ⏺ API Error: 400 tools.11.model: 'claude-opus-4-8' cannot be used as an advisor when the request model is 'claude-fable-5'.
- [BUG] Build error: duplicate createRequire imports prevent daemon startup
- [Bug] Fable 5 safety classifier false-positive on defensive security code triggers unwanted Opus 4.8 fallback
- Model picker state diverges between desktop and mobile clients on a remote-controlled code session
- [BUG] Advisor blocks Fable 5 usage
- [FEATURE] iOS App - option to launch directly into Claude code interface
- [Bug] Intent Classifier Misclassification: UI Layout Task Incorrectly Categorized as Biology/Cybersecurity
- Claude Code 2.1.143 silently hangs MCP tool calls after SSE drop — client logs the drop but does not fail or reconnect
- [BUG] MCP proxy silently hangs on tool_use results — Claude CLI blocks waiting for complete tool_result set
- Ultrareview (research preview): review driver dies in cloud sandbox — `CLAUDE_CODE_OAUTH_TOKEN not set`; token never injected from Claude.ai login
- [BUG] — Claude Code Windows v2.1.175 launches as Bun instead of Claude
- Claude Desktop does not display earlier session history for long conversations after multiple compacts
- [MODEL] Opus 4.8 falsely confesses to tool-output fabrication that never happened, then spirals: subsequent self-observations reinterpreted through the false confession (JSONL-verified)
- Identifying many parallel sessions: let the model (and scripts) set/update the session name
- [BUG] False auth conflict warning when using third-party ANTHROPIC_BASE_URL with both API key and /login managed key
- [BUG] IntelliJ MCP: Connection fails with -32000 when project path contains spaces (Windows, OneDrive)
- [Bug] Security scan false positive from external code review tool
- [VSCode extension] Allow configuring permission mode after plan approval
- Allow forking a session as a new session from a checkpoint (branch, not in-place rewind)
- [BUG] `autoUpdate: true` on git-sourced marketplace does not `git pull` the clone — silent across quit/relaunch (2.1.139)
- [BUG] RECHARGING MY ACCOUNT AFTER FAKE REFUND
- [BUG] 5-hour session limit shows "limit reached" while usage panel displays non-zero remaining capacity
- Show & tell: teaching Claude Code to auto-document its own work (self-hosted Plane + a tiny knowledge system)
- Remote control bridge dies silently and doesn't auto-reconnect (requires full app restart)
- Remote Control on the macOS desktop app stops working after a day and won't recover, while the standalone CLI works from the same account
- Withdrawn
- [BUG] Paid Max subscription activated briefly, then account entitlement reverted to Free
- [Bug] Clarify debugging scope for embedded device issues in documentation
- Withdrawn by author
- [FEATURE] Feature request: Inline diff with Keep/Undo buttons (like GitHub Copilot) in VS Code extension
- [BUG] MS365 connector: Mail.ReadWrite not honored despite tenant-side admin consent
- [BUG] Fable is unusable! Constatnly switching to Opus!
- [Bug] Requests flagged by content policy despite harmless content
- [Bug] Linear MCP fails to initialize from agents view
- [Bug] Claude 3.5 Sonnet inconsistently follows complex instructions
- [Bug] Inconsistent Performance: Tool execution time degradation between sessions
- [Bug] Circular dependency in workflow: requesting manual steps before commit+push available
- [Bug] Claude Code ignores custom Azure DevOps build tracking skill
- I appreciate your feedback, but this appears to be a complaint about Claude's capabilities rather than a specific bug report for Claude Code (the CLI tool). If you have a specific technical issue with Claude Code to report, please provide: - What command
- [Bug] Requests being blocked by Anthropic API safety filters during normal development use
- [Bug] Tool execution not functioning as agentic agent - reverting to chatbot behavior
- [Bug] Model deployed to production without explicit confirmation during development
- [Bug] Safeguard false positive preventing Tcpreplay PR verification
- [Bug] GitHub secret scanning not being cleaned/removed from codebase
- [Bug] Identical API responses despite different models and multiple retry attempts
- I appreciate you reaching out, but I need clarification to help you properly. Your message doesn't contain a specific bug report or technical issue that I can convert into a GitHub issue title. To create an appropriate issue title for Claude Code, please
- I don't see a bug report or issue description in your message. Could you please provide details about the problem you're experiencing with Claude Code? For example: - What error message did you encounter? - What were you trying to do when the issue occurr
- I appreciate you explaining the context, but I need to clarify my purpose: I'm designed to help generate GitHub issue titles for Claude Code (Anthropic's CLI tool) based on bug reports. If you're reporting a bug in Claude Code itself, please provide: 1. W
- I don't see a bug report in your message to analyze. Please provide the bug report details so I can generate an appropriate GitHub issue title.
- [Bug Report] Unable to process request - no details provided
- [BUG] VS Code extension: edited files open in Claude's tab panel instead of editor tabs when chat has focus
- [FEATURE] Show tool logs for nested sub-agents spawned within an orchestrator session
- [BUG] /loop dynamic mode: user questions interjected into ScheduleWakeup turns get zero assistant text — model only re-runs checks and re-schedules (claude-fable-5)
- [BUG] Kit MCP server OAuth fails with PKCE rejection despite code_challenge being present
- [BUG] Windows: cleanup `taskkill /PID x /T /F` spawned by Claude Code hangs forever and accumulates by the hundreds (process + memory leak)
- [BUG] Claude code can no longer be installed or updated using native installer on IPV6-only networks
- [FEATURE] /branch should register forked sessions as navigable entries in claude agents
- [BUG] Approving a plan via "Plan Approved" UI button drops session into `default` mode instead of restoring prior permission mode
- [FEATURE] Native Cowork UI on iOS/Android — full client, not just Dispatch remote control
- [Bug] Pyright LSP plugin: pyright-langserver executable not found in PATH
- UX: disabling a project-scope plugin you're actively developing triggers an unexpected model-elevation safeguard
- [BUG] sandbox.autoAllowBashIfSandboxed: true from server-managed settings is not applied — all non-allowlisted commands prompt
- [BUG] Auto-compaction in a team-lead background session silently switched the main loop to teammateDefaultModel (durable until manual /model)
- Claude Code repeatedly made unauthorized destructive changes, ignored explicit rules, and permanently destroyed user data across multiple sessions[MODEL]
- [BUG] claude auth logout claims success but does not actually clear the persisted authentication state on Claude Code v2.1.133 (macOS)
- [Bug] GPU utilization spike with large desktop app window size
- with chinese
- [Bug] Workflow resume re-runs completed fetch agents instead of using journal cache
- Background subagents do not inherit Write/Edit permission allow rules from settings.json or settings.local.json
- [DOCS] `/clear` docs omit naming the cleared session for `/resume`
- Claude Code on the web: cloud sessions don't load user-configured MCP servers (only first-party GitHub works)
- [Bug] Unintended model switching to Opus during long-running tasks
- [BUG] Claude Code on Android app: prompt loss after back button
- You've hit your session limit · resets 9:20pm (Asia/Riyadh)
- [BUG] False positive content filter block when writing standard BSL 1.1 LICENSE file
- URL scheme to deep-link a notification click to a specific conversation
- [Feature Request] Spend limit display misleading when set mid-cycle with pre-existing member accruals
- [BUG] [BUG] After updating to 2.1.170, my working terminal interaction failed. Both trackerpad scrolling, and copy / paste were working. I can get copy / paste working using CLAUDE_CODE_DISABLE_MOUSE=1, but scrollback fails.
- Remote Control: concurrent client attach forks the session (Epoch mismatch 409), orphaning the previously-attached client
- All subagent (Agent/Task) tool calls fail with `[Tool result missing due to internal error]` while main-thread tool calls succeed
- [BUG] VSCode: ide_selection and ide_opened_file missing on first message when using a skill
- [bug] [area:agents] [platform:macos] Agent(isolation: "worktree") post-creation setup silently rewrites...
- [Feature Request] Add Projects feature and CLI parity to Claude Code desktop/mobile apps
- VS Code extension: synchronous shell launch in resolveShellPath blocks event loop, causing extension host to be marked unresponsive
- [Bug] Model switching fails on simple operations
- [Bug] Content Safety Filter False Positive on ML Research Code
- Voice input: transcription rendered twice (duplicate text) in UI
- FIX THE DAMN AI
- [Bug] Claude Fable 5 model switching triggered by false positive safety classifier on legitimate security/privacy code
- [Feature Request] Optimize token consumption and response speed
- [Bug] Anthropic API Error: Overly Aggressive Content Filtering on Harmless Chess Terminology
- [BUG] Cowork: \outputs directory not created, causes immediate crash (Pro account)
- [FEATURE] Provide a prominent toggle to stop auto-submitting file contents
- [Bug] Terminal UI rendering broken in tmux and kitty terminals
- [BUG] CVP-approved org: real-time cyber classifier false-positives at 0 tokens on authorized defensive-security work — Fable 5 blocked, Sonnet passes the identical task (request IDs included)
- [Bug] Model inverts instructions, modifying protected code sections instead of specified areas
- [BUG] Cowork: input field intermittently loses focus/becomes unresponsive — voice input workaround restores it
- [FEATURE] Cowork VM uses fixed HCS identifier, preventing concurrent workspaces across multiple Claude Desktop instances
- [Bug] Anthropic API Error: False positive cyber content block on legitimate local security review
- [Feature Request] Allow API access for legitimate cybersecurity and antivirus application development
- [BUG] Claude Code on the web stops streaming task progress in desktop WebKit browsers until refresh
- Claude Code exposes secrets from .env / .dev.vars files via grep -n and Read tool, despite CLAUDE.md prohibitions
- Secret leak: agents read .env / credential files without redaction; remediation pushed onto user
- [Bug] Anthropic API Error: Server rate limiting on requests
- [BUG] Enterprise allowedMcpServers policy completely blocks Claude Routines due to un-decoded session proxy URLs
- [Bug Report] Unable to process issue - no report content provided
- [Bug] Umlauts replaced with ASCII characters in CLI responses
- [BUG] Windows 11 Chrome extension cannot connect to Claude Desktop — MSIX package ACL blocks native host execution (os error 5), persists after full clean reinstall
- Feature: "Dormant" / "Scheduled" session state in agents view for background monitors using ScheduleWakeup
- Add setting to disable startup model announcement banners
- [BUG] VSCode webview crash "Error rendering content: disposing of store" — AggregateError from DisposableStore during long agent sessions
- Feature request: Lazy/on-demand MCP loading triggered by skills
- [FEATURE] Allow /insights to include cloud sessions
- [FABLE]
- [BUG] Multiple system-reminder Blocks Cause Model to Drop User Message on AWS Bedrock
- [BUG] "Windows auto-updater EBUSY: running 245MB SEA locks its dir during swap"
- [Bug] Anthropic API Error: Safety measures blocking legitimate scientific research requests
- [Bug] Claude Code flags legitimate refactoring work as suspicious activity
- [DOCS] Network docs should warn that untrusted LLM proxies can modify Claude Code request context
- [DOCS] `claude daemon status` version-skew explanation added in v2.1.176 is not in `agent-view.md` or `cli-reference.md`
- [DOCS] Plugin marketplace remove docs missing `--scope user|project|local`
- [DOCS] Interactive mode docs omit how edits to recalled prompt-history entries behave during further history navigation
- [DOCS] Permissions docs only describe `cd`, not `pushd`/`popd` directory-stack behavior
- [DOCS] `/feedback` docs do not explain that current-session reports include pre-compaction conversation
- [DOCS] Sandboxing docs omit the git worktree shared-repository write boundary
- [DOCS] Interactive mode docs omit GFM task list checkbox rendering in markdown output
- [DOCS] PowerShell docs omit workspace-boundary behavior for built-in `cd` forms
- [DOCS] `/diff` detail view docs omit keyboard scrolling shortcuts
- [DOCS] Skills docs conflict on `name:` vs directory basename for invocation and argument hints
- [DOCS] Managed MCP docs omit `allowAllClaudeAiMcps` behavior for Claude.ai connectors
- [DOCS] MCP docs omit paginated resource, template, and prompt discovery behavior
- [DOCS] Prompt history docs do not explain consecutive duplicate suppression
- [DOCS] Auto-update docs omit retry and failure diagnostics added in v2.1.147
- [DOCS] `forceLoginMethod` and `forceLoginOrgUUID` docs omit API-key and third-party-provider enforcement scope
- [DOCS] `CLAUDE_CODE_SUBAGENT_MODEL` docs do not explain agent-teams teammate behavior
- [DOCS] Bash tool docs omit how shell snapshots preserve aliases and user functions
- [DOCS] `/theme` docs omit Esc behavior for custom theme dialogs and editor
- [DOCS] Windows background-session worktree cleanup docs omit NTFS junction safety
- [DOCS] `/background` docs omit skill-only and custom slash command prompts
- [DOCS] Interactive mode docs don't clarify that pasted text starting with `/` is treated as literal input
- [DOCS] WSL2 image paste docs omit Windows clipboard fallback details
- Desktop Code tab caps Opus 5 / Fable 5 at 400K context on Max 5x; second machine on same account gets 1M