Digest / September 22, 2026
Claude Code Issues Digest — September 22, 2026
241 new issues filed ·
175 resolved in anthropics/claude-code (UTC day).
New issues (241)
- [Bug] Unexpected model downgrade from Fable 5 to Opus 4.8
- [BUG] Linux: sandboxed Bash permanently dies after EnterWorktree into .claude/worktrees/ (bwrap: Read-only file system)
- [Bug] Anthropic API Error: Consistent 500 and 529 Server Errors
- [Feature Request] Support for MDM-managed Mac diagnostics script generation
- [Bug] Staged desktop update is invisible: no update-ready indicator, silently deferred for days, then quits mid-session without warning
- Weekly usage cap met, no wind down in project whilst project running to create …
- Cross-session teammate/subagent messages leak between unrelated Claude Code sessions
- Desktop app browser pane blocks subresources of sandboxed (null-origin) iframes with ERR_BLOCKED_BY_CLIENT
- [Bug] Multiple errors occurring across new sessions
- [BUG] Ctrl+J / LF newline is hardcoded in the chat input, so unbinding ctrl+j in keybindings.json has no effect (breaks SKK IMEs)
- Claude in Chrome extension OAuth still fails with "Authorization failed" — recurrence of #69051/#34176, both now closed/locked
- ピン留めセッションが、特定操作後に新規発言でも最上位へ自動移動しなくなる
- [Bug] reasoning_extraction misreads a user-defined analysis section in the response as an attempt to extract internal reasoning (Opus 5, Claude Code and claude.ai)
- [Bug] Account access restrictions on paid tier accounts
- [Bug] Inappropriate Model Switching During Defensive Security Audit
- Desktop app: daily forced re-login (session_stale_relogin) + scheduled-task run freezes mid tool call and silently skips all subsequent runs
- [FEATURE]: per-session spend cap for interactive sessions (--max-budget-usd is print-mode only)
- [Bug] Automode executes code for passive analysis tasks instead of responding
- Scheduled tasks silently skip their fire window, with host machine and scheduler engine both demonstrably active
- Scheduled task (routine) runs no longer appear in the desktop sidebar or Remote Control
- [Bug] Claude ignores git worktree context and modifies wrong checkout
- Bash permission rules are quote-blind: an operator inside a quoted argument defeats a matching rule
- No way to see which permission rule, hook or classifier decided a tool call
- [BUG] VS Code extension: chat input shows ghost/offset text when prompt overflows — mentionMirror scroll clamps before re-render
- [Bug] Potential security issue: Telegram bot metadata/embedded content concern
- [Bug] Claude fails to verify file existence before claiming folder is empty
- [MODEL] Claude’s peculiar way of writing is spreading.
- [BUG] Claude in Chrome 1.0.94: side panel shows "Can't reach the Claude extension." in Vivaldi; classic side panel opt-out removed
- [Bug] Content safety filter triggers false positives on benign input like greetings
- [Bug] Safety filter falsely flags harmless messages as cyber threats
- computer use: add window-scoped screenshot (capture one app window, not the full screen)
- Desktop sidebar: no way to hide empty groups or sessions whose working directory is gone
- System prompt attribution (Co-Authored-By) overrides user CLAUDE.md rules
- Session ID reused after deleting a transcript .jsonl, and the new session inherits a stale/wrong auto-generated title
- [Bug] Auto mode classifier returns no verdict for project-local write operations
- [Bug] Performance degradation and increased error rate in code generation
- [BUG] Desktop app session tabs always show "Chat (claude)" instead of the session's name
- [BUG] No download/save option when viewing a generated PDF file in the conversation preview panel
- [BUG] tui shows garbled file name with the VSCode extension
- [BUG] Code tab missing from Claude Desktop app (Windows, Pro plan)
- Projects should be able to run local Claude Code runs
- [Feature Request] Allow configurable content filtering overrides for proprietary/confidential documents
- [Bug] app:exit action not dispatched when bound to custom keybindings
- [Bug] Safety filter incorrectly flags normal conversations with reasoning_extraction error
- Remote Control: prompts sent mid-turn never become user turns, so the phone and the terminal show different conversations (re-report of #86980)
- Dynamic multi-agent workflow consumes full credit balance while ~60% of agents never run
- /device confirmation page posts to absolute root path instead of respecting listen.public_url subpath, breaking deployments behind a path-based reverse proxy
- [BUG] Linux/GNOME Wayland: Desktop app self-focuses its window on Code session completion, hijacking Ctrl+Alt+Arrow workspace switching
- auto-mode classifier vetoes commands present in permissions.allow (no override path)
- Desktop app: auto-merge is hardcoded to github.com, and the PR bar requires manual binding, on GitHub Enterprise Cloud with data residency (*.ghe.com)
- [Bug] Claude Code requires manual refresh after login to function properly
- [Bug Report] Unable to send message - unclear error
- Auto-detect PR from current branch on GitHub Enterprise (*.ghe.com)
- [FEATURE] /btw with no arguments should open the side chat in the CLI (parity with Desktop), and add a composer link for it in Desktop
- Auto-merge is hardcoded to github.com — not host-aware for GitHub Enterprise (*.ghe.com)
- Auto-merge is hardcoded to github.com — not host-aware for GitHub Enterprise (*.ghe.com)
- Desktop Gmail connector sends from an unseen account: no from field, no account in results, Bash hooks cannot guard it
- [Feature] Deduplicate AGENTS.md fallback when identical instructions arrive via another channel
- [BUG] [Windows] Quick Entry opens on non-primary display when tray icon is clicked on primary display
- [Bug] Anthropic API Error: Safeguard flagging legitimate MCP security verification request
- [BUG] ~30s of startup input lag: claude.ai connector-list fetch blocks the event loop (26 connectors)
- [FEATURE] Configurable Session Retention Policy
- [Bug] Model Makes Errors on Hand-Typed Values Rather Than Generated Code
- Desktop app (Windows): window does not go to background when clicking another application
- [Bug] False positive security detection when using Qwen model reasoning output
- [FEATURE] Project-local session storage: keep transcripts in the project instead of ~/.claude/projects/<path-slug>
- [Feature Request] Reduce false positive rate in safety guardrails for benign code review requests
- [BUG] Claude desktop app holds NoIdleSleepAssertion indefinitely, preventing macOS idle sleep
- [FEATURE] Add changelog to VSCode extension
- [Bug] Claude Code uses separate Git worktree in temp folder, preventing visibility into operations
- [FEATURE] `claude project mv` — one command to re-point a project's state after its directory moves (transcripts, memory, prompt history, session registry, profiles)
- [BUG] /btw won't use tools anymore
- 🐛 [Bug] Archived API key continued charging $20.86 after deletion — CSV proof attached (#29108, #53292)
- The project threads need the github connector, but tnat complains that it's not…
- [Bug] Agent incorrectly assumes uncommitted changes instruction not given
- [BUG] Fullscreen TUI treats mouse side buttons as left and middle clicks, so the forward button pastes PRIMARY
- [Bug] Model setting persists incorrectly after switching to Default
- [BUG] Desktop app session list omits sessions with entrypoint: sdk-ts (Orca-launched), and worktree-switch also fails for them
- [FEATURE] Report the deciding permission rule on every decision, not only on a rule-forced ask
- [Bug] Project safeguarded after pushing to GitHub
- I appreciate the sentiment, but I need an actual bug report or feature request to generate a GitHub issue title. Please describe the problem you're experiencing with Claude Code, and I'll help create an appropriate issue title.
- /code-review finder subagents can live-reproduce a defect outside the working tree, and a project has no way to constrain them
- [BUG] Blockquote line breaks rendering when the quoted text was pasted (>=20 chars)
- [Bug] Claude invents file content instead of reading it ANYTHING TO SAVE GPU ON PRICATE ACCOUNTS RIGHT?? SEE YOU IN COURT
- [BUG] Flagging a prompt puts Claude in a dysfunctional state
- Session resume sends invalid previous_message_id (UUID) when prior session ended with connection error
- [Bug] Security content filter false positive on non-security code
- [BUG] Opus 5 usage is counted against the weekly Fable limit (Max 5x, advisor off)
- [FEATURE] Tags for chats and tasks, assignable and filterable across web, desktop and mobile
- Windows: Bash/PowerShell tool calls flash a visible console window on the desktop
- [Bug] Auto-update repeatedly fails when network connection is lost
- [Bug] Fable 5.1 overly aggressive security safeguards causing false positives for legitimate security work
- [FEATURE] Add a persistent, at-a-glance machine indicator for multi-machine Claude Code sessions
- Session Manager Local tab shows only 1 of 189 local sessions after Reload Window
- [Bug] Tool execution failures causing excessive API message overhead PIECE OF SHIT GARABGE
- Desktop: MCP elicitation is declared but never presented — `onElicitation` is never wired, and `local-agent-mode-*` declares no `elicitation` (2.2553.1)
- LSP client answers client/registerCapability with -32601, wedging servers that use dynamic registration (and no per-request timeout)
- `[BUG] Existing Max account stuck in unrecoverable state after failed payment, login redirects to onboarding, billing inaccessible, Claude Code locked out`
- [Bug] Assistant language randomly switches to Korean/Chinese in Thai-language sessions
- [FEATURE] Claude desktop app: sync read/unread state across devices on the same account
- [Billing] API key stolen and abused twice in 12 days (LLMjacking); spending limit exceeded by 29%; refund denied via template responses without case review
- Desktop: Remote Control links are lost on account switch and idle chats cannot be relinked
- [Feature Request] Add repository whitelist to bypass blockReadsOutsideWorkingDirectories permission
- when old existing sessions when moved from Claude.Ai projects to Claude Cowork fails to write to disk
- "Other" responses not visible in the chat
- [BUG] VSCode extension's /mcp OAuth flow ignores configured --callback-port (uses a different port than the terminal CLI)
- [Bug] Claude ignores formatting instructions for table output
- [BUG] Bypass mode ignores blanket file authorization (Desktop, macOS) — prompts continue despite both settings
- (withdrawn by author)
- [FEATURE] Keep cleared conversation visible (greyed out) after /clear instead of wiping the screen
- Deleting a session in the desktop app leaves all of its files on disk
- Scheduled routine email notifications silently fail (still reproducing, related to #84645 / #80880)
- [BUG] Reftable repos: CLI reads .git/HEAD directly, branch resolves to ".invalid" and PR status is dropped from the statusLine payload
- [BUG] : I am not able to login to Claude Code just two days before the actual renewal date. Repeats every month.
- [BUG] Terminal gets corrupted, or rewritten over time - impossible to follow what happened. Deletes my messages.
- [BUG] Desktop app: repo picker stuck on 'No repos available' blocks chat input even after installing GitHub App and granting repo access
- [BUG] Skill Delivery
- [BUG] advisor_20260301 tool type 400 hard-fails every turn in an affected session; model switch does not recover
- Model pads delegated subagent prompts with unrequested work, wasting tokens despite explicit memory rules
- Plan mode: auto mode classifier gives no verdict for the designated plan file when CLAUDE_CONFIG_DIR is set
- [BUG] Claude in Chrome: navigation refused for domains present in "Your approved sites"
- [BUG] VS Code Extension hangs indefinitely on NFS-mounted $HOME due to SQLite WAL locks and log I/O latency
- [BUG] Persistent "API Error: Connection dropped (ECONNRESET)" messages
- [BUG] Weekly usage limit exhausted in ~2.5 days instead of 7 — matches pattern in #52502
- Browser pane: add a mute control (and default-mute) for agent-opened tabs
- PIECE OF SHIT TOOL UNSUSABLE
- Memory feature doesn't reliably prevent recurring behavior, no cross-device persistence
- [GitHub integration] **Title:** Cloud (Cowork) session: git proxy and GitHub API refuse a repository that is not in the "session's authorized repository set"; no `add_repo` tool exposed; connector reconnect does not fix it
- [BUG] Long-lived CLAUDE_CODE_OAUTH_TOKEN (setup-token) and Remote Control are mutually exclusive with no documented workaround
- Auto mode classifier timeout fails closed and blocks tool calls even with explicit allow rules
- Edit tool applied changes without a permission prompt despite `ask` rules for Edit (Windows, 2.1.278)
- [Bug] Memories are ignored until explicitly re-requested, defeating their purpose
- Assistant rotated a production API key without checking code/docs first, causing an outage and extra costs
- [FEATURE] Inline line-anchored comments on plans and diffs at the approval prompt (TUI)
- Plan mode and session limits
- [BUG] Desktop: inline Reply box swallows Tab during IME composition (Google Japanese Input) — text commits as unconverted kana
- [Feature Request] Support for local model decryption of archived device data
- [BUG]
- Agent picks highest-blast-radius interpretation of an ambiguous 'undo' instead of asking, cancelling a live prod GitHub Actions run
- [BUG] Windows desktop (MSIX): plugin hook scripts invisible to externally-spawned interpreter — every hook ENOENTs and asyncRewake Stop hooks loop forever
- [Bug] /design command shows authentication options instead of creating design canvas
- [Bug] Model unexpectedly switches to different model during execution
- [Bug] Anthropic API Error: Opus 5.5 safety filter blocking CVP member after update to v2.1.280
- Worktree resume fails with "could not verify worktree" due to git fetch timeout, not worktree corruption
- [Bug] Paste operation returns incorrect clipboard content instead of user selection
- [Bug] API Token from System Environment Not Recognized or Enabled
- [Bug] Bypass permissions mode overrides user tool preferences in CLAUDE.md and MCP tools
- [BUG] Windows: orphaned Vite processes block Claude Desktop launch after update (0x80070020)
- [BUG] Claude in Chrome tools prompt on every call in Bypass mode; "Always allow" ignored (regression in 2.1.280, desktop Code tab, Windows)
- [BUG] Shows "monthly usage limit" message when the 5-hour session limit is reached
- Why can't i make threads run locally
- [Bug] Anthropic API Error: Overly aggressive safety filtering blocks legitimate code review
- [BUG] [Claude Desktop] [Cowork] Cowork sandbox cannot access local project folders on macOS
- [BUG] Cache entry written with its breakpoint on a trailing `role: system` message is never reused — with the per-user-turn `<total_tokens>` reminder on, every tool-less turn re-writes the whole conversation (likely also the root cause of #93490)
- [Bug] Non-Haiku models (Opus, Sonnet) unavailable or misconfigured
- [BUG] TUI overprints new output on top of old lines in iTerm2 (still occurring on 2.1.278; continuation of auto-closed #68461)
- Railway MCP custom domain: Claude concludes no TXT record is needed, domain never verifies
- [BUG] macOS 12: Desktop app stuck on bundled Claude Code 2.1.260, Opus 5.5 greyed out — yet 2.1.280 lists this app's SDK wrapper (0.3.260) as tested while `stable` (2.1.267) does not
- claude.ai-hosted MCP connectors intermittently expose zero tools in headless launchd/claude -p sessions
- Background subagent completion messages scroll the transcript, making the assistant's last reply unreadable
- Title: Session feedback prompt: rating options are unclear and not clearly ordered
- Sometimes when a thread finishes with an open question it still lands up in idle and that question is not relayed to the main thread
- Agent Kunjani passed review but is missing from the plugin directory and community catalog
- [Bug] Anthropic API Error: False-positive cyber safeguard classification on harmless messages
- [Bug] Claude stops execution after announcing task start without completing work
- includeCoAuthoredBy: false is intermittently ignored mid-session
- Agent repeatedly acts on unverified assumptions and overreaches scope across a single prod-infra debugging session
- [Feature Request] Add cross-project inbox for waiting threads and permission prompts
- [BUG] github_repo_access_denied persists despite confirmed GitHub-side access (collaborator, App install, and identity all verified)
- [BUG] CLAUDE.local.md switches off the AGENTS.md fallback, so a personal note silently drops all shared project instructions
- [Bug] Anthropic API Error: Opus 5.5 safeguards flagged message with reasoning_extraction
- Image paste probe (xclip -t TARGETS -o) hangs forever on X11, leaving stuck processes for hours
- Claude 2.2553.13 (16e20e): Unable to view or download files in remote SSH sessions
- claude plugin eval sandbox never loads a plugin's own agents/*.md for the Agent tool
- [FEATURE] Snooze a session: hide it now, bring it back on a chosen date
- Linux/X11: one copy writes the clipboard twice (OSC 52 and xclip), and the two race for selection ownership
- `/usage` Usage tab shows no rate-limit progress bars (still reproduces on 2.1.280) — silently falls back to qualitative-only panel
- Auto-mode Bash classifier blocks authorized, non-sensitive credential operations under inconsistent labels (and doesn't change with explicit authorization)
- [BUG] Remote Control push fires for permission prompts a PermissionRequest hook auto-approves (regression, seen on 2.1.280)
- [Bug] Claude Code posted false bug report to public repository without user consent
- [Bug] Recursive deletion deletes unintended files outside target directory
- [BUG] Linux sandbox leaks claude-empty-* mount-source directories in the temp dir (never removed)
- [BUG] Opus 5.5 requires unstable release of Claude Code
- /model on 2.1.280 shows stale "Update to 2.1.280+" disabled row cached by an older concurrent session
- [Bug] Approved Cyber Verification Program users experiencing unexpected model downgrade for endpoint testing
- [Feature Request] Improve API documentation and development guidance
- [BUG] Regression since 2.1.174: plugin installation from HTTP marketplace fails with "fatal: transport 'http' not allowed"
- App update disconnects session registry from existing ~/.claude/projects transcripts (sessions become unlisted/unresumable, data intact)
- [Bug] Anthropic API Error: Message flagged by safeguards with Opus 5.5 (1M context)
- Stop hook reports fully-pushed branches as unpushed when the clone's fetch refspec is narrowed
- [Bug] Usage attribution overstates skills with handoffs and inconsistent subagent handling
- [Bug] Anthropic API Error: False positive safeguard flags on claude-opus-5-5 with reasoning_extraction
- [Bug] Poor Code Generation Quality and Insufficient Task Completion
- [Bug] Anthropic API Error: Opus 5.5 safeguards flagging non-cybersecurity work
- [Bug] Agent spawning restriction not enforced, causing unauthorized agent fork and excessive API usage
- [BUG] In-app browser URL bar in Claude Desktop reverts/clears after typing a localhost URL during Claude Code sessions
- [FEATURE] Let marketplace plugins reference a released .mcpb bundle by URL with a SHA-256 pin, instead of committing the bundle to the marketplace repository
- Stop hook treats a stale checkout as unsaved work, and advises committing a revert
- Stop hook prescribes noreply@anthropic.com as commit author, which conflicts with repo-level identity rules
- [BUG] Desktop (Linux), Code tab: long prompt overflows input box — can't scroll to end, send bar disappears
- [BUG] Failing asyncRewake Stop hook causes unbounded model re-wake loop (high token burn)
- Gmail MCP: attachments over ~5MB fail due to base64 serialization in JSON parameter
- [BUG] Desktop (macOS): Code session silently loses computer use if created during a GrowthBook refresh
- [BUG] New version says Approaching weekly Usage Limit when its the 5-hour limit that is reaching its limit
- [Bug] Anthropic API Error: Tool name exceeds 128 character limit
- [BUG] ECONNRESET "Connection dropped" with claude-fable-5-1 at normal effort; works at MAX
- [BUG] Local model via ANTHROPIC_BASE_URL ends the turn with a greeting when claude.ai connector instructions arrive mid-turn
- Move-to-cloud fails with generic 'remote rejected the push from the app (auth)' even though branch is confirmed already pushed and matches origin
- VSCode extension hardcodes invalid model ID 'claude-haiku-4-5-20251001' for internal background task
- /code-review apply-findings step occupies the user turn unlabeled (no <command-*> or <system-reminder> wrapper)
- Better design for session monitoring within a CC project
- [BUG] Windows: claude plugin eval fails with ENAMETOOLONG uv_spawn when the plugin dir holds ~3.5K chars of summed junction paths
- [Bug] Anthropic API Error: Content Filtering Blocking Normal Work Prompts
- [FEATURE] Running sessions should adopt a fresh Claude apps gateway login instead of each needing its own /login
- [BUG] Claude desktop WorktreePool GC deletes user-created git worktrees in .claude/worktrees/
- 使用量画面の「製品別使用状況」が、独立した上限ゲージのように見えて誤解を招く
- [BUG] Consistent reset delay
- "No, ask again next time" in the outside-working-directory Read dialog denies the current read; no allow-once option exists
- [Bug] Opus 5.5 refuses to do any cyber work despite being part of CVP.
- [BUG] Compaction silently omits older invoked skills above the 25,000-unit aggregate budget; --resume reconstructs none
- [FEATURE] Sandbox: per-repo egress by host + port + protocol (TCP/UDP); the only built-in route prompts every time
- [BUG] A resumed background subagent can't deliver its final report (hand-back delivers only once)
- [BUG] Queued /feedback drafts can't be corrected, withdrawn or merged; duplicates pile up in the send prompt
- [FEATURE] Feedback should be a conversation, and the model should know what was sent
- [FEATURE] Claude Desktop on Linux: add an update check (it has no updater at all)
- [Bug] Subagent workflow management causes token waste and task assignment failures
- [Bug] Safeguard triggered on benign query about hidden notice content
- Safety classifier over-blocks a legitimate use case (personal call recording) and benign adjacent topics
- [Bug] Shift+Cmd+F keyboard shortcut intercepted by Cursor IDE, prints "f" instead of opening search
- Opus 4.6 regression: hallucinating dates and degraded accuracy after recent update
- Opus 4.6 regression since ~Sep 16: hallucinating dates and degraded accuracy
- I don't see a bug report in your message. Please provide the bug report details so I can generate an appropriate GitHub issue title.
- [BUG] iOS Code tab: session list stays "Waiting for you" after /clear, while the session view correctly shows Context cleared
- Opus 4.6 regression since ~Sep 16 2026: precision loss, hallucinations, personality degradation
- MCP: Claude Code sends legacy 'initialize' after successful modern 'server/discover', breaking modern-only servers
- [FEATURE] allow disabling the periodic background git diff refresh in the Code tab
- [FEATURE] Plugin-provided inline autocomplete with custom trigger (e.g. `#` for issues)
- I keep getting approval requests for basic web fetch. Auto mode seems to not be…
- [BUG] [Claude Desktop] [Cowork] Automatic update moved sessions to the cloud; file tools act on copies and device_commit_files leaves old bytes
- [BUG] Main composer doesn't scroll with cursor position — content above visible window becomes inaccessible while typing long prompts
- [Feature Request] Add /restart command to quickly resume latest conversation
Resolved issues (175)
- [Bug] reasoning_extraction misreads a user-defined analysis section in the response as an attempt to extract internal reasoning (Opus 5, Claude Code and claude.ai)
- System prompt attribution (Co-Authored-By) overrides user CLAUDE.md rules
- Desktop app: auto-merge is hardcoded to github.com, and the PR bar requires manual binding, on GitHub Enterprise Cloud with data residency (*.ghe.com)
- Auto-merge is hardcoded to github.com — not host-aware for GitHub Enterprise (*.ghe.com)
- Claude Desktop's fileTypeAssociation manifest invalidates existing UserChoice defaults for Office/dev file types on Windows
- [BUG] Falta el ícono de copiar en mensajes/artifacts, y "Copiar" del clic derecho aparece desactivado
- [Bug] Non-English input causes CLI failure or unexpected behavior
- [Bug] Claude Code fails on project tasks and routine development work
- [Bug] Agent validation checks pass incorrectly on broken deployments
- [Bug] Opus model producing false claims in audit analysis
- [Bug] MCP OAuth: revokeReplacedServerTokens revokes freshly issued tokens immediately after /mcp login
- [Bug] Safety Filter False Positive on Legitimate Code Task
- [Bug] /auto-mode-setup fails with removeFromPermissionsAllow[6] is not a rule string the removal offer could have produced
- [Feature Request] Add configurable safeguard thresholds for codebase review operations
- [BUG] Bun 1.4.0 segfault at address 0x54 on background thread (KERNEL32/ntdll frames) ~6 min into session — CLI 2.1.234 native, Windows 10 x64, two consecutive sessions
- [Bug] Task flagging incorrectly triggers on important operations
- [BUG] The fatal message: Interrupted · What should Claude do instead?
- Mobile app: active desktop session shows up as archived with stale content, then disappears from session list
- [BUG] Session scratchpads under /tmp/claude-<uid> never garbage-collected after exit — 8.8 GB of dead sessions resident in RAM on tmpfs, drives swap thrashing
- [BUG] Remote Control (--spawn worktree) self-restarts in a rapid loop, eventually hits systemd start-limit-hit and dies
- Remote-control session worker self-terminates on "Epoch mismatch (409)" instead of recovering
- Desktop scheduled tasks: docs say "ask Claude" can change a task's model, but update_scheduled_task MCP tool has no model parameter
- [Bug] Anthropic API Error: Model generating false information and stating inferences as facts
- Gmail connector: browser reconnect doesn't restore MCP tool bridge in active Claude Code session (high token cost)
- [FEATURE] Unified "Pending Approvals" panel for multi-worktree / multi-session workflows
- Artifact public sharing fails with "unscannable" content scan error
- [BUG] Permission/privacy messages say THAT files are accessed, but not WHICH — collapsed UI hides the target path
- [BUG] ask permission rule for Bash(git push *) not triggering confirmation prompt
- Cross-session messages to idle desktop-app sessions never start a turn — ~38 parallel count_tokens requests, then nothing (app build 1.32352.1)
- Stray `.git` directory created in non-repo cwd (e.g. $HOME) by the runtime-files exclude writer
- [BUG] send_message reports success on native Windows where cross-session messaging is not offered — no inbox socket bound, nothing delivered, caller cannot tell (worked on this machine until 2026-08-11)
- [FEATURE] Let a folder on a WSL path open as a Local (Windows) session
- Moving a session to an existing project name ungroups all sessions
- [Feature Request] Add option to disable whimsical comments in CLI outputwhat
- Tool/message output intermittently replaced with garbled compressed-memory debug text
- Claude not connecting with Chrome Platform -BUG
- [BUG] Claude in Chrome: tab viewport silently collapses to 280x15 and never recovers; screenshot/left_click/resize_window all keep reporting success
- Gate keybinding contexts on vim mode (NORMAL/INSERT) so scroll and line-editing keys can share a keystroke
- [Feature Request] Add "Ask about it" option to permission prompts for clarification without rejection
- Cloud/remote session token usage not included in "Total tokens" stat on the Overview dashboard
- claude mcp get/list print configured MCP secrets (headers, env vars) in cleartext with no masking
- MCP server tools not indexed by tool discovery despite server showing "Connected"
- Browser-pane dev-server launcher fails with EPERM (uv_cwd) -- sandboxed subprocess can't load an ad-hoc-signed native binary (Vite 8/Rolldown)
- [FEATURE] Should look for git.path in Vscode user settings
- Sandboxed Bash wedges permanently when the working directory is on a read-only filesystem (no cwd recovery)
- [Bug] Anthropic API Error: Sonnet 5 Safeguards Flagging Verified Cyber Users
- [FEATURE] Workflow tool: byte-exact data channel between workflow scripts and the host (model-retyped transport corrupts commands and payloads)
- Permission mode reverts to manual on startup despite auto mode configured
- [BUG] Local conversation is unviewable on web / unsharable after migrating to team account
- [BUG] VS Code panel: sequential edits to one file diff each proposal against a stale snapshot (diff-editor tab only)
- [BUG] Bottom status row PR badge on long-lived background jobs stays pinned to the first PR ever linked, even after it merges
- Desktop machine no longer offered as target for new local sessions from mobile app (cloud only)
- Mobile: always-on context-window usage indicator (could reuse the 'Create PR' action area, no extra real estate)
- Remote Control view renders the injected cross-session security preamble verbatim on every teammate message — terminal view collapses it
- [Bug] Anthropic API Error: Overly Aggressive Content Filtering on Legitimate Cybersecurity Development Work
- Claude in Chrome extension breaks LastPass extension when both are enabled
- Remote Control mobile/desktop view: show currently running subagents (live agent roster)
- Show file edits as inline diff in the editor tab (like GitHub Copilot)
- [FEATURE] A structured channel for users to submit ideas, with opt-in transcript sharing and reciprocity
- [BUG] "Set up auto mode for your environment?" dialog is uncompletable: Enter toggles the checkbox instead of submitting and arrow keys never reach Continue
- [BUG] HTTP MCP server with a hyphenated name: tools injected to the model as mcp__server_name__* (underscore) while /mcp UI shows mcp__server-name__* (hyphen) → "No such tool available"
- [BUG] Input line renders one column left: incremental render's leading backspace contradicts absolute cursor parking
- [BUG] MCP servers connect successfully then get torn down with no reconnect on resumed sessions
- [BUG] Remote SSH fails with NO_HOME on hosts lacking printenv (BusyBox/NAS devices) — probe should use echo "$HOME"
- Feature request: import/create comment threads on Artifacts (round-trip serialization)
- [BUG] `CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH` override silently fails via both cloud config paths (Environment Variables panel and repo `settings.json`)
- Model fabricates a "prompt injection" incident — quotes a fake attack string that never appears in any tool_result
- [BUG] Stop hook no longer fires on stream-stall-terminated turns ("The response stopped arriving") — StopFailure fires instead but its decision output is ignored, breaking auto-continue hooks
- Browser tool navigate consistently fails with "denied or failed" for previously-working sites
- [BUG] Regression in 2.1.228: streaming to a Bedrock gateway returns zero events, client silently falls back to non-streaming, throughput halves
- [BUG] Remote Control: after a host reboot the environment re-registers with the same ID but drops all pre-existing sessions — claude.ai/code shows them, sending fails with "session not found", transcripts intact on disk
- Worktree isolation: session guidance suggests the ! prefix as a workaround, but ! runs in the same session and hits the identical block
- [Bug] Firefox browser not supported by Claude Code CLI
- [Bug] Claude Code performance degradation after service outage - Opus 5.0 failing basic tasks
- [BUG] A `fork` subagent dispatched for a read-only/critique task adopts the parent's role and takes unrequested agentic actions (launches Workflows/Agents)
- [BUG] `claude -p "/usage" --output-format json` intermittently fails right after a long print-mode run — exit 5, or `is_error:false` envelope with no `result`
- [FEATURE]
- Subagent stream dies silently on transient network failure — no retry/reconnect, watchdog kills after 600s of silence
- Out-of-bounds read past end of 1 GiB mapping on CPUs without AVX/SSE4 (SIGSEGV, not SIGILL)
- [Bug] Model behavior: failures are omitted from reports and admitted only when the user corners the model
- claude-in-chrome extension fails to connect (tabs_context_mcp: "Browser extension is not connected") after reinstall + full Chrome restart
- [Bug] Anthropic API Error: False positive safety classifier blocking legitimate mental health detection software
- Expose session name as a variable (e.g. $session_name) for use in shell/loop commands
- [Bug] ultrareview: All reviewer agents terminate at Verify stage, Dedupe never executes
- Ink UI leaves stale placeholder text when TERM_PROGRAM=ghostty and new content is shorter
- [Bug] Ancestor CLAUDE.md @import directives not resolved in child directory sessions
- [Opus 5] Model insult to user
- [Feature Request] Fixed schedule option for usage limit reset window instead of rolling 5-hour window
- Background agent completion notification lost after messaging the agent mid-run
- [BUG] PR Review skill burns through usage limits in 10 minutes and then loses all progress
- [BUG]
- [Bug] Claude Code ignores claude.md instructions and uses chat mode instead of markdown workflow
- Desktop app: top-level "+ New" always prompts for a folder (regression, no longer remembers last working directory)
- [Bug] Claude Code ignores user-defined instructions and performs opposite actions repeatedly
- [Bug] Cybersecurity classifier blocking defensive malware analysis tooling mid-conversation
- [MODEL] Claude ignore Path-scoped rules during file creation
- [Bug] Anthropic API Error: Server rate limiting during request processing
- Claude in Chrome: silent-reauth failure latches startupReauthState={terminal:true} with no cooldown, permanently killing the bridge until manual sign-out/sign-in
- iOS Simulator MCP: no way to pass launch arguments to the app (control action:"launch" runs bare simctl launch; `text` is silently ignored)
- [BUG] Android app renders slash-command messages without their arguments
- [BUG] Session permanently bricked: 400 "Tool reference 'LSP' not found in available tools" after a deferred tool leaves the pool mid-session
- [FEATURE] Desktop: named split-view layouts with lockable cells
- [BUG] Slash command auto-submits on click before user can type arguments
- claude plugin install refreshes the marketplace, sees a newer version, and still reports "already installed" without upgrading (2.1.232)
- [BUG] updating claude on Windows does not work
- [Bug] Windows update fails silently: bin\claude.exe not replaced during version upgrade
- [Bug] Model response mixing Portuguese and English languages unexpectedly
- Title: OAuth login succeeds but token is never written to macOS Keychain — persistent "Not logged in"
- Removing enabledPlugins from a project's settings leaves an orphaned project-scoped install that can't be removed from the UI Body
- [BUG] Claude in Chrome side panel dictation always "Microphone access is blocked" — claude.ai iframe missing `microphone` in allow attribute (v1.0.85)
- (withdrawn by author)
- [BUG] Desktop app (Windows): main process pinned at 100% of a core by git watches left on deleted worktrees (archive re-attaches the watch, the worktree pool then deletes the folder)
- [BUG] Windows: Import silently skips every CLI transcript whose project path is stored with forward slashes in ~/.claude.json — whole directories never import, with no error surfaced
- [BUG] Mid-input slash completion never matches plugin-namespaced skills (plugin:skill), while the position-0 dropdown does
- [Bug] Anthropic API Error: Opus 5.5 safety filter blocking CVP member after update to v2.1.280
- [BUG][REGRESSION] model:/effort: frontmatter override silently broken for BOTH commands/*.md and skills/SKILL.md (v2.1.220, worked on v2.1.197)
- Cross-machine SendMessage works via undocumented bridge:<sessionId> address while every surface says reply-only/unreachable
- iOS Simulator panel: claude-ios-sim crashes (SIGABRT) on attach via CoreImage Metal binary archive
- iOS Simulator panel crash-loops on macOS 27 beta: seatbelt profile denies Metal's new cache-dir writes
- iOS Simulator helper crashes in CoreImage/Metal on macOS 27, making tap/swipe unavailable
- `claude-ios-sim` crashes in a loop on macOS 27: uncaught ObjC exception from `-[_MTLDevice recordBinaryArchiveUsage:]`
- iOS Simulator live panel crash-loops on macOS 27.0 beta: Metal recordBinaryArchiveUsage nil-array during CoreImage init in claude-ios-sim helper
- iOS Simulator live panel helper (claude-ios-sim) crash-loops in Metal/CoreImage on macOS 27.0 beta
- Bug: interrupting a running Bash tool call does not prevent its side effect, but the rejection text says nothing was written
- [BUG] Fullscreen: scroll-back sticky "last prompt" banner never appears since 2.1.247 (scrollRef -> scrollViewport.handle refactor)
- [BUG] Compaction silently omits older invoked skills above the 25,000-unit aggregate budget; --resume reconstructs none
- [BUG] OAuth and magic-link login redirect existing Max subscriber to "Let's create your account" onboarding — browser and desktop locked out, iOS + Console still authenticated
- [BUG] Claude Desktop (macOS): in-flight completion stream dies with "TypeError: network error" when router deprecates/renumbers IPv6 prefixes; not retried, response lost
- macOS 27: Claude Desktop silently stops updating when Background App Activity is disabled
- I appreciate you reaching out, but I'm not able to help with this request as written. If you're experiencing a technical issue with Claude Code or the Claude API, I'd be happy to help if you can: 1. **Describe the specific technical problem** - What erro
- [Bug] Safeguards Incorrectly Flagged Message - False Positive
- [Bug] Agent ignores explicit instructions and implements unspecified changes
- [Bug] Claude Code not responding to input/commands
- I appreciate you reaching out, but I'm Claude Code's issue title generator for GitHub. I can only help format bug reports and feature requests into proper GitHub issue titles. Your message appears to be in Czech and translates to "Never commit anything wi
- [Bug] Agent generates contradictory or confusing responses affecting user workflow
- [Bug] Unclear error when attempting to build SNN model
- Cowork sandbox recurring failure — useradd exit 12 / no space left on device (5 recurrences in a week, prior fix didn't hold)
- Remote Control (iPhone): native MCP tool-permission dialogs never render — session blocks invisibly (AskUserQuestion cards render fine)
- claude desktop: worktree sessions should inherit the base repo's .claude config
- Remote Control session confirmed active but never resolves — session lookup fails on claude.ai/code
- [BUG]
- Opus orchestrator contaminates "independent" subagent consults: injects its own candidate answers, relabels an unverified subagent claim as VERIFIED, and transmits a rejected design's frame while warning against it
- [BUG] Plugins are re-enabled when claude-code is updated.
- Claude in Chrome extension always binds to Desktop's native-messaging host, never the CLI's, even with Desktop fully quit
- [Bug] CVP approval not recognized across organizations in Claude Code
- [Feature Request] Add CTF/security testing mode for controlled vulnerability exploitation
- [FEATURE] Browser view / browser tools in Claude Code cloud sessions (Claude Code on the web)
- [BUG] Pycharm keeps freezing when using Claude on the terminal of the IDE
- /insights Languages chart: no Swift/Objective-C in extension map; .h misattributed to C on iOS codebases
- ralph-loop: two copies of setup-ralph-loop.sh at v1.0.0; the unpatched one silently creates unstoppable loops
- [Feature Request] Improve Opus model execution quality and provide configuration options
- skill-creator: run_eval.py writes skills to .claude/commands/, so description optimization always measures recall=0%
- [FEATURE] Add a dedicated theme token for spinner tip text, separate from `subtle`
- [Bug] Anthropic API Error: Content Flagged by Safety Filters for Legitimate Use Case
- [FEATURE] Promote a subagent to a session and demote back — reclaim context + intervene in orchestrated subagents
- Feature request: free-form feedback box + image paste in plan mode review (VS Code)
- [BUG] awsAuthRefresh no longer displays device verification code before opening browser
- [BUG] Cannot log in to Claude code nor VSCode extension
- [Bug] Anthropic API Error: Cyber Safeguards Blocking Legitimate Requests
- --continue can silently resume another live session's conversation when the working directory is shared
- [BUG] Bash sleep-block guidance prescribes the Monitor tool in sessions where the harness itself reports "Monitor is disabled for this session"
- [Bug] Duplicate issue creation without prior verification against existing issues and design documentation
- [MODEL] Forked subagents (subagent_type: fork) sometimes report status: completed after doing no work, with a result message unrelated to the assigned task
- [Bug] Agent teammate display uses agentType instead of assigned name in fleet view
- [BUG] VS Code extension context pie pins at "100% context used" ~12k tokens before auto-compact fires; popup mislabels it as "until auto-compact"
- [BUG] Desktop app (Windows): delivered file cards open a claude.ai /api/.../files/{id}/contents URL in the browser — 404s or save-as with a corrupted filename
- [BUG] Internal prompt template unexpectedly appended to visible assistant reply (possible leak into chat)
- [BUG] ExitWorktree (remove) reports success but orphans the worktree (dir + admin entry + branch); intermittently corrupts parent repo core.bare
- [BUG] Desktop Linux: theme does not follow system light/dark changes at runtime — restart required (Electron ignores portal color-scheme signals)
- [BUG] MCP tool call fails with "requires approval" error even after granting approval
- [BUG] CronCreate jobs never fire in Claude Code Desktop app (works fine in CLI)
- no respeta los permisos y lee todo el disco
- [BUG] Claude Code sessions freeze locally until another session receives input — Desktop app & VS Code only (not web)
- [BUG] Grep/Glob permission dialog shows the search pattern only, with no target path
- [BUG] GPG Dialog is buggy when Claude Code is full terminal view