Digest / September 11, 2026
Claude Code Issues Digest — September 11, 2026
250 new issues filed ·
243 resolved in anthropics/claude-code (UTC day).
New issues (250)
- [Feature Request] Support multiple commands in single line (e.g., /rename and /color together)
- [BUG] 日本語IME入力時のプロンプトの遅延により、誤入力・文字欠落が頻発する
- I don't have access to the actual bug report content associated with that request ID. To generate an accurate GitHub issue title, I would need you to provide the details of the bug report, including: - What the issue is - Error messages (if any) - What be
- [FEATURE] Cache CLAUDE.md and rules across sessions: they are re-written to the prompt cache at every session start
- [MODEL] Subagents continue problem.
- [Bug] Security flags raised for basic operations
- Remote Control message paints in idle CLI composer but never starts a turn
- /usage in desktop app opens Settings page instead of showing per-session cost
- [Bug] False positive validation error on legitimate API requests
- [FEATURE] replace `#` to `>` when using "Show last response in external editor (ctrl+g)"
- 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] Cowork macOS: local sandbox VM starts with NO network route (only loopback) and cloud egress proxy returns 403 for every domain, despite "Allow network egress: All domains" — regression since 2026-09-10 ~23:15 UTC
- [BUG] Desktop preview_start fails for projects in ~/Documents on macOS: getcwd "Operation not permitted" despite Documents access (see #42914)
- Constant 'API Error / Disconnected / Retry x of 10' on 2.1.268 (macOS 26.2) — auto-mode classifier also timing out
- Self-permission-granting safety check bypassable via Bash-mediated write (gates by tool, not by effect)
- [BUG]
- [BUG] Cowork cloud sessions ignore "Allow network egress: All domains" and enforce a narrow built-in allowlist; all third-party API calls from scripts fail (403 at CONNECT) on both the cloud shell and the local VM shell
- [Bug] Performance degradation in Claude Code over past three days
- [BUG] Fenced code block inside a list item renders at column 0, losing the list indentation
- ListAgents / cross-session messaging: expose user-assigned session titles
- [Bug] Inappropriate content flagging for authorized system administration tasks
- [Bug] DMCA §1201 False Positive in Security/Copyright Detection
- Claude Desktop MSIX (v1.52386.0): Workspace fails to start — VHDX not found + junction redirect rejected, even after clean reinstall
- [Bug] Authorized requests incorrectly flagged as unauthorized
- Cowork VM egress allowlist ignores Admin Capabilities network access settings
- [BUG] VSCode extension shows "OAuth session expired and could not be refreshed" despite a valid, non-expired Keychain credential
- Isolated Linux workspace (device-bash) fails to start on linked Windows desktop
- I don't have a bug report to analyze. Please provide the actual bug report or error message you'd like me to create a GitHub issue title for, and I'll generate an appropriate title following the format specified.
- [Bug] Anthropic API Error: Overly aggressive safety filtering on legitimate system signal operations
- [BUG] Cowork cloud sandbox: egress allowlist collapses to ~5 hosts despite account setting "All domains"
- Remote Control session drops intermittently even when desktop-side connection is stable
- [BUG] Send feedback and keep planning is not visible after 10+ comments
- [BUG] Claude Desktop (Code tab, macOS) — new messages hang indefinitely at "Sending...", never transmit
- Parked-permission "retirement" always stamps toolDenialKind:"user-rejected", even when the actual cause is an internal session reset (e.g. after /mcp reconnect), not a real user action
- Esc kills an unrelated background subagent irrecoverably, instead of just stopping the current foreground command
- [BUG] MessageDisplay no longer fires for text between tool calls (regression)
- [BUG] Documents-folder permission is lost on every embedded CLI auto-update, forcing a restart that drops session context
- [BUG] Opus 5 safeguards false-positive on legitimate bacterial genomics research (`[bio]` flag)
- [BUG] Desktop Code tab: long unsent prompt disappears during composition; Cmd+Z and history navigation cannot restore it
- [FEATURE] Theme tokens for Markdown elements (headings, bold, inline code) in assistant output
- [BUG] Plan-approval screen shows only "Revise," no "Accept" control — cannot exit plan mode
- macOS: concurrent claude processes zero MCP OAuth entries in shared Keychain blob despite valid refresh tokens
- [BUG] Claude Code tried to read file outside current dir to cheat
- placeholder
- Claude Chrome extn fails
- [BUG] Cmd+Enter no longer queues message — now interrupts agent's current response
- [BUG] /ide None selection and "ide": false setting do not persist across sessions — VS Code auto-reconnects every startup
- New theme is too dark/high-contrast — please restore ability to choose previous (lighter) theme
- [BUG] Deleted routine stays in sidebar; opening it shows loading spinner + 'not found' stuck simultaneously (Desktop, Windows)
- [Bug] Claude Code overriding user decisions from previous sessions
- Forged background-task completion notifications injected into model context (absent from session transcript)
- I appreciate you reporting this, but I need more specific information to create a proper GitHub issue title. Could you please provide: 1. **What is the specific bug or problem?** (e.g., incorrect log entry, false conversation history, wrong timestamp) 2.
- [Bug] CLI ignores user input and makes decisions autonomously
- remote-control: routine sessions are never sent end_session on completion, and when it is sent the harness exits non-zero
- [Bug] Anthropic API Error: Sonnet 5 Cyber Safeguards Flagging Legitimate Requests
- [BUG] Weekly limits reset when switching machines
- No way to see which routine sessions are still alive and holding resources
- [BUG] Claude Desktop: self-uploaded plugin cannot be updated or uninstalled — install record is frozen at first install
- [FEATURE] Surface resolved model and subagent_type for background agents in ListAgents and Agent results
- i dont understand the issue ? why is it flagged as cyber ?
- [BUG] Cloud-run Cowork scheduled task fires on time, then stalls mid-run at the first turn boundary
- Claude in chrome not connected via Desktop app
- Auto-mode classifier attribution repair misses custom gateway base URLs (follow-up to #64585)
- Legacy .claude/scheduled_tasks.json is loaded by the VS Code runtime, enabled: false is ignored, jobs fire as user prompts, and the file is rewritten on disk
- [Bug] Agent tool intermittently unavailable, requires restart to recover
- [BUG] Cowork/cloud sandbox: api.telegram.org blocked with 403 connect_rejected, and Pro/Max accounts have no setting to allow it
- Egress allowlist ignores "All domains": every host outside a built-in list gets 403 on CONNECT (regression 2026-09-10 -> 2026-09-11)
- [BUG] Artifact tool cannot read live artifacts in Cowork — read path binds to native WebFetch, which Cowork substitutes with mcp__workspace__web_fetch
- [Bug] Safety flagging triggered during legitimate credential reuse discussion
- [FEATURE] Keyboard shortcut to navigate between user prompts
- [Bug] Claude makes premature final decisions instead of presenting user choices
- [BUG] error: parseFromSourceCode failed during PowerShell setup
- [BUG] Cowork Scheduled Tasks can silently consume massive context/cache usage with almost no visible output
- [BUG] Desktop: older turns get re-delivered behind a `background_tasks_redelivered` marker and rendered last, so the feed ends on a turn from hours earlier
- [Bug] Claude Code terminates all tasks when single task shutdown requested
- [Feature Request] Add option to disable fast mode advertisement in vim mode
- [FEATURE] Persistent project-scoped prompt cache, independent of thread/session lifetime
- [BUG] account was being quality checked - once stopped off the rails
- [BUG] Cowork loses a session's home directory on VM restart, and the connected folder goes with it
- [FEATURE] Action buttons the agent can emit, bound to a command that runs in the session VM without a model turn
- [BUG] macOS allowUnixSockets does not match /tmp symlink path
- [BUG] Claude Desktop – "Configure third-party inference" setup page fails with 503 (bootstrap request)
- [BUG] Claude Code hanging during reasoning
- You've hit your session limit · resets 1:30pm (Europe/Prague)
- `claude-ios-sim` crashes in a loop on macOS 27: uncaught ObjC exception from `-[_MTLDevice recordBinaryArchiveUsage:]`
- [BUG] Claude Desktop plugin browser shows repository name as author and marketplace source, ignoring metadata
- [Bug] Anthropic API Error: False Positive Security Block [reasoning_extraction] with Multi-Agent Collaboration
- [BUG] Cowork (Windows MSIX): 'Delete Cowork VM Sessions and Restart' deletes the scheduler index and all local chat transcripts, dialog reports only EACCES on 'latest'
- False-positive [reasoning_extraction] block when quoting the model's own displayed thinking (Opus 5)
- Cloud session checks out stale local branch when a commit is pushed between container pre-warm and session start
- [cyber] safeguard false-positives: 5x model-switch in one e-commerce session; /feedback dialog opens empty
- [FEATURE] support for users with eye sensitivity
- [BUG] 2.1.268: statusLine command is never invoked when cwd is a git repo (works in a non-git cwd) — regression from 2.1.267
- Cowork Desktop: local "Additional allowed domains" / "All domains" egress stopped being enforced after 2026-09-11 update (VM 2.1.260 worked, 2.1.266 blocks)
- Expand/view full MCP tool inputs in the VS Code extension
- ToolSearch never surfaces datadog:mcp security-findings tools despite toolset enabled + reauth
- Worktree-isolation guard refuses every jj git <subcommand>, including in a correct jj workspace
- [Bug] Session token allowance miscalculated or depleted unexpectedly
- [BUG] Claude Code VS Code extension causes complete macOS lockup from suspected memory exhaustion
- [BUG] Plugin HTTP MCP ${VAR} header expansion resolves to empty in 2.1.260 (works in 2.1.247 / 2.1.223) — bearer token never sent
- Opus 5 at xhigh: thinking on ~100% of requests and 2-7x output tokens per request since ~02:30 UTC 11 Sep, with no client or settings change
- [BUG] VS Code extension 2.1.268: prompt input box missing, 2.1.267 works
- [Bug] Misleading session health check prompt appears too early in conversation
- [Cowork] Regression since 10 Sep 2026: skills can no longer be saved as a full .skill package; only SKILL.md-only `propose_skills` card, `references/` un-updatable
- [BUG] VS Code extension: Ctrl+Shift+Esc opens a new Claude Code panel but focus stays in the editor (Linux)
- Scheduled Tasks always launch on Opus 5, ignoring model picker and settings.json
- Agent bypassed a Remove-Item safety block by switching to cmd's rmdir, then a quoting bug deleted ~10,000 files (project source, data, and personal automation scripts)
- Claude Desktop: mouse cursor disappears when the folder access permission dialog opens
- [FEATURE] Desktop: session toolbar (terminal / panel / Browser) is unavailable until the first prompt is sent
- [Bug] Agent displaying inappropriate human-like attitude in responses
- [FEATURE] Support multiple simultaneous Notion workspace connections in one session
- [BUG] Bash tool (Linux): pkill -f / pgrep -f match the tool's own bash -c … eval wrapper (exit 144, phantom PIDs); 2.1.214 guard covers only the CLI process
- Flagged as Dangerous via Fable
- cant scan with mcp in session idk why
- [Feature Request] Claude Code model selection and performance transparency
- [BUG] Fullscreen TUI: transcript region renders blank until a key is pressed (2.1.267, macOS, cmux/libghostty)
- Add system tray / minimize-to-tray support for Claude Desktop (Windows)
- [BUG] All Claude Code sessions/projects disappeared after app closed unexpectedly
- [BUG] Read-aloud (TTS) uses wrong/default voice instead of the one selected in General settings
- [BUG] Scheduled tasks: WebSearch tool calls hang indefinitely in background/cron-triggered sessions
- [Bug] Multi-select prompts used incorrectly when single option is appropriate
- Claude in Chrome: navigate always fails with "Navigation to this domain is not allowed"
- Windows/Git Bash: Bash tool command line truncated at ~8175 chars, and backslash runs halved by MSYS2 argv parsing
- [Bug] SessionStart hook sessionTitle not reflected in ListAgents alias
- Subagent transcripts: `message.usage.output_tokens` differs across the block records of one response; the first record carries a partial count
- Put an unverified ID number into a government letter, then sent it from the wrong email account despite explicit instruction
- Channel messages are merged into the previous user message and re-rendered on the next request, losing the prompt cache on non-Anthropic backends
- Claude Code and Codex icons not visible in top right on remote server
- [Bug] macOS: teammate spawn "fork failed: Device not configured" — xnu ptmx slot-vector race triggered by cat placeholder + respawn-pane -k (root cause for #77211)
- [BUG] Cowork/Claude Code egress proxy blocks dev.azure.com with 403 despite domain present in Capabilities network allowlist
- Cowork scheduled task auto-pauses permanently after a missed run instead of catch-up (contradicts documented behavior)
- [BUG] MSIX/Microsoft Store install quits for auto-update but never relaunches
- [FEATURE] add chat history panel to VS Code extension
- [BUG] Cowork Windows — Shell/Bash Cannot Mount Connected Folders (Plan9/virtiofs)
- I need the bug report content to generate an appropriate GitHub issue title. Could you please provide the bug report details you'd like me to create a title for?
- CoworkVMService cannot set its own SCM recovery actions ("Access is denied") -> a crash stays down until reboot
- You've hit your session limit · resets 8:10pm (Europe/Rome)
- CoworkVMService cannot set its own SCM recovery actions ("Access is denied") -> a crash stays down until reboot
- [BUG] Unexpected biology safeguards in Claude Opus 5
- Status line configured in settings.json is never executed/rendered on native Windows (v2.1.268)
- [BUG] Dictation captures audio but returns no transcript (macOS desktop 1.52386.0) — regression persists from 1.49585
- [BUG] Desktop: "New session" from a sidebar group (or from a cloud session) silently clears the remembered project folder; next sessions start as "No folder"
- [BUG] VS Code extension: file links with spaces or non-ASCII characters don't open (openFile() never decodes the percent-encoded path)
- Setting to hide the "N line(s) selected" TUI indicator for @file#L mentions
- [Bug] Anthropic API Error: Inconsistent biology content policy blocking legitimate research request
- Agent executed a mutating az cli command without explicit user go-ahead
- [Bug] Anthropic API Error: Cyber safeguards flag legitimate requests
- [BUG] Cowork sandbox ignores network allowlist entirely — regression, worked 2026-07-29 through ~2026-09-04
- [BUG] Filenames stripped after submitting in VSCode extension
- I don't have access to the bug report details for request ID `req_011CewrLoZnRaugrk1YeBz9T`. To generate an appropriate GitHub issue title, I would need you to provide the actual bug report content, including: - What the issue is - Error messages (if any)
- [BUG] Model shorthand `--model sonnet` resolves to Sonnet 4.5 instead of Sonnet 5; inconsistent with /model picker
- Overview page doesn't mention PATH setup after native install
- False-positive session abort on benign ACA cryptogram/cipher-solving vocabulary
- [Bug] Claude Code repeatedly generates commits violating length and trailer requirements
- [BUG] Cowork: workspace VM never starts — service waits for `configure` forever while client reconnects once per second
- [BUG] Cowork sandbox egress proxy blocks all custom domains regardless of org allowlist config — breaks access to own infrastructure
- [BUG] Remote Control: when a server is at --capacity, "new session" silently opens the most recent existing session instead of refusing
- [BUG] big problem!
- [Bug] Inverted logic in absence-of-evidence search caused JSON config fields to be marked CLI-only
- Browser pane per-action origin permission prompt has no "Allow always" option, ignores settings.json/bypassPermissions
- [BUG] "All domains" egress setting not applied to Cowork cloud sessions
- [FEATURE] Apply plugin auto-updates automatically when the session has no user turns yet
- Session lost after helper tool update — transcript intact but unrecoverable
- [BUG] Esc with several queued messages delivers them one per keypress instead of all at once
- [MODEL] OPUS 5 must go!!!
- /ultrareview (/code-review ultra) fails as "only available in a local Claude Code session" from the desktop app's Code tab
- Windows desktop app ignores defaultShell: "bash" for ! commands, always uses PowerShell
- SWITCH TO TURN OFF FRUSTRATING GUARDRAILS
- I need more information to generate an appropriate GitHub issue title. Could you please provide: 1. What error message or behavior did you encounter? 2. What command or action were you trying to perform? 3. What was the exact error or restriction that occ
- [FEATURE] aliases for claude command
- [FEATURE] Desktop app: option to cycle sessions in most-recently-used (MRU) order instead of sidebar position
- Add setting to keep IDE selection indicator in footer instead of inline prompt
- Agent teams: task self-claim has no role/tool-compatibility filter, so read-only teammates claim tasks they cannot do
- [FEATURE] Add visual file explorer/code browser to Claude Desktop Code tab
- CLAUDE.md exact-word prohibitions are not enforced — model violated a single-word ban twice in one message
- Desktop app /mcp OAuth sends a claude:// redirect URI that mcp.cloudflare.com rejects (Redirect URI must use HTTPS or a local loopback address); CLI connection not shared with the desktop session
- [BUG] Notification idle_prompt fires while background subagents are still running
- [BUG] /btw appends "Never consult the advisor tool." to the user's message, so the model treats it as a user instruction
- [Bug] Unexpected model switch from Fable to Opus 4.8 during session
- Task tools are off by model, with no settings key and no way to keep a session checklist by default
- ExitWorktree can't remove a worktree launched via --worktree (or entered via path:), so WorktreeRemove hook never fires and per-worktree resources leak
- [BUG] Cowork (Windows): device VM egress ignores the account Domain allowlist - 403 blocked-by-allowlist while set to All domains
- [BUG] Claude didn't obey me even Im in my testing sandbox
- [BUG] Design window renderer grows unbounded during generation and is killed at 2-4GB (macOS, Apple Silicon)
- [BUG] Bash tool creates its session directory via `/proc/self/fd/N/` instead of `mkdirat()` — fails wherever procfs is absent or incomplete (regression in 2.1.263)
- [Bug] Security classifier malfunction
- Cowork (Windows): VM cannot mount host C: drive after Windows cumulative update - 'Plan9 mount failed: invalid argument'
- Instruction injected into every tool result overrides explicit user instruction and has no opt-out
- [BUG] Visible thinking uses whimsical/cutesy verbs even after explicit instruction to stop
- Anomalous directive-styled system-reminder block attached to a user turn
- [FEATURE] Adjustable line spacing (leading) in the desktop app — accessibility
- [BUG] Desktop Scheduled Tasks catch-up ignores day-of-week/date cron fields, fires on wrong days
- [BUG] Desktop app's self-generated .desktop entry breaks its own claude:// login handoff on Ubuntu — GIO refuses to load it
- [Bug] all my messages are getting flagged because i ran a security audit on my repo with os4int tools
- [Bug] Fable 5.1 `reasoning_extraction` safeguard flags ordinary research-assistant turns (3 request IDs); every "rephrase" forks the session
- [FEATURE] Artifacts: share with a group (IdP/SCIM group, HRIS team, or admin-defined org group) instead of adding individuals
- Gmail connector: untrash_message fails with "no permission" even though trash_message succeeds
- Images always show [Unsupported Image] on Claude Desktop (Windows)
- Open in → VS Code/Cursor passes untranslated Linux paths under WSL
- [Bug] Claude Code chat does not respond to user messages
- Images always show [Unsupported Image] on Claude Desktop (Windows)
- [BUG] claude.ai connectors show "Connected" but unusable across all Claude products
- Sessions created on Mac desktop app don't sync to mobile app session list (even after 'nudge' message and full app restart)
- Desktop app: no way to restore the git/branch bar after dismissing it with the X button
- [BUG] /effort command returns "Unknown skill: effort" despite being documented
- [FEATURE] Auto-reduce effort at 90% usage, auto-fallback to smaller model at 95%
- /code-review fork is finalized while its named finder agents still run under CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS
- Worktree branched from local `main` instead of `origin/main` when the session is rooted inside a git submodule (desktop app)
- [BUG] Desktop app split-view pane renders black/blank for a specific session
- [Bug] Anthropic API Error: Incorrect [bio] Content Filter Flagging on Technical DevOps Terminology
- Auto mode steers the model to Bash, which cuts path-scoped rules off at the knees
- Remote machine SSH connection fails with "No route to host" due to TCC-disclaimed subprocess lacking Local Network permission
- [Bug] Claude Code building changes in temporary directory and using rsync instead of direct file operations
- Desktop app "+New" creates a session bound to the previous project instead of prompting for one
- [BUG] /diff turn view counts one extra added line for every created file
- [Bug] Statusline displays stale PR number when worktree reused across branches
- [BUG] Deleted artifacts are re-armed as watches at every session start, emitting a permanent notice per artifact
- Desktop app: hours-long session transcript permanently lost (not just a display/routing bug)
- [BUG] Desktop (Linux): spawn_task sessions still start without claude_desktop_config.json MCP servers; attach happens on UI focus, tools reach the model only on the next queued turn (re: #67432)
- [Bug] Safeguards incorrectly flag PiP mode changes as malicious activity
- [Feature Request] Unified plugin system for MCP servers, skills, and agents
- [FEATURE] Copy or move the current prompt draft to a different session (stash is single-slot, same-session only)
- [Bug] Clipboard paste keybinding changed from cmd+v to ctrl+v causing paste functionality issues
- Session limit exhausted in less than ~30minutes on a Max x5 Plan
- [BUG] Messages posted to a session's own inbox socket render the full peer envelope instead of the collapsed preview
- [Bug] Incomplete eye moment implementation and missing sphere shape outline
- Desktop app: worktree sessions reload every claude.ai connector; per-project disable list does not follow worktrees and deniedMcpServers cannot suppress app-injected connectors
- Computer use refuses clicks under an invisible, click-through overlay window (Wispr Flow "Status" dialog)
- Desktop app: resumed background subagent shows only the SendMessage JSON; transcript looks frozen until the agent completes
- Usage forecast copy reads suppressed demand as "on track"
- /clear should also empty the visible chat scrollback for a clean slate
- [Bug] MCP server validation bypassed when remote-settings is empty
- Workflow tool keeps dispatching queued agents after a rate-limit stop (no circuit breaker)
- Managed Agents github_repository rejects documented fine-grained PAT after session creation
- [Bug] Context indicator appears on unselected files, causing incorrect context in thread
- [BUG] `/ide` connection banner reflows/displaces in-progress input text
- [FEATURE] Background compaction: keep the session running while the summary is produced
- [BUG] Agent re-submits a byte-identical rejected tool call 5x in a row while its prose claims the fix was applied
- App UI session list doesn't show sessions created via terminal (`claude --resume`)
- Project-scope effortLevel in .claude/settings.json is never applied in desktop-app (Code tab) sessions; effective effort varies per session with no settings change
- [Bug] Subagent ignores explicit directive and inherits parent orchestrator role instead of executing assigned task
- [Bug] Incorrect keybind documentation in agents view (ctrl+c vs ctrl+x)
- [Bug] Session resume fails with partial or full session ID after stop command
- [Bug] Can Anthropic look at your session transcript to help us improve Claude Code? Not taking the y
- Model questions and then overrides explicit live instruction with generic skill procedure and prints secrets in plain text
- [Feature Request] Make selected line count display position configurable
- /model save-as-default rewrites all of settings.json from a stale session snapshot, silently reverting hooks
- Project memory/session storage collides across different projects due to non-ASCII path slug encoding
- /goal: Stop condition evaluator cannot see the instruction passed via /goal, loops until it declares itself unachievable
- [BUG] Stop hook blocks and bills a full turn when a background agent is mid-write; no way to distinguish in-progress work from abandoned work
Resolved issues (243)
- [BUG] Cmd+Enter no longer queues message — now interrupts agent's current response
- [BUG] Image Inputs via Read Tool or @ Reference Consume 7–10× More Tokens with On-Prem LLMs
- [BUG] Desktop (Linux/Wayland): chat shortcuts match on physical key code — Ctrl+AltGr+d (= Ctrl+Backspace on Bone/Neo layouts) closes the chat instead of deleting a word
- [BUG] Linux .deb Recommends a nonexistent package (kwalletd6), pulling gnome-keyring onto KDE systems
- [BUG] Project knowledge index serves stale file state after update/delete/re-upload, and diverges from the interface file list
- [BUG] Windows 11: "Code" toggle unclickable when sidebar is in floating mode, works when docked
- Enable Credentials vault feature for my account
- [Bug] Excessive token consumption without productive output
- [Bug] Tool execution resulted in unintended multiple charges via Stripe integration
- [Bug] Random Q/A message flagged during decompiler task execution
- [Bug] Safeguard Check False Positive on Unrelated Prompts
- [Bug] False Positive Security Flagging for P2P Networking Code
- [Bug] Incorrect C compiler error classification as security threat
- [Bug] Bypass functionality not working as expected
- [Bug] Agent performs unsolicited tasks without user confirmation, wasting tokens
- [Bug] Excessive token consumption compared to expected usage rates
- [Bug] Indian postal code 622412 not recognized by weather API integration
- [Bug] Significant degradation in response quality and task performance
- [Bug] Anthropic API Infrastructure Reliability Issues Causing Service Disruptions
- [Bug] Failed to find Server Action in Next.js deployment mismatch
- Footer PR badge stuck on previous PR after switching branches with gh pr checkout
- Unexplained promotional sentence appended to end of assistant response
- [Bug] Anthropic API safety filter triggers false positives on legitimate security audit terminology, causing unwanted mid-task model downgrade
- [BUG] .bat is marked "Generated by Claude Code, do not edit manually." So a Claude Code update or a re-run of Chrome setup could overwrite my edit and revert it.
- Session branches while the agent is still running, and the abandoned branch keeps executing tools
- Main-thread CPU spin (100-140%+) persists for hours after a long background Bash task completes (v2.1.222, macOS arm64)
- [BUG] claude.exe crashes with ACCESS_VIOLATION (0xc0000005) on Windows, repeatable, same code offset
- [BUG] Model order wrong. Fable between Opus and Sonnet on VSC plugin.
- [Bug] Anthropic API Error: Safeguard flagged on model downgrade from Opus 5 to 4.8
- [Bug] Model unexpectedly downgraded from Opus 5 to Opus 4.8 mid-session
- Session transcripts silently deleted by automatic cleanup — no user consent, no warning, no setting to disable
- [Bug][cyber] Safeguard false-positive loop: the flag's own error text is the strongest predictor of the next flag (18 silent switches to Opus 4.8 in one session)
- Session transcript renders blank/garbled on first open of the day
- Sonnet 5 appears to receive the legacy Write tool description due to substring model matching
- [Bug] Safety classifier false-positive on defensive security audit tool discussions
- [BUG] VS-Code: Execution stops when moving the Claude Code tab around
- [Bug] Claude Opus 5 generates hallucinated responses not present in previous versions
- [BUG] Tool call completes on server but response never reaches client (4-min timeout) in Claude Desktop
- Asserts unverified tool behavior as fact and compounds errors across self-corrections (effort: high)
- Rewind/undo reverts file changes but not conversation context, causing confident-but-stale answers
- Desktop app (Windows): SendUserFile attachment cards open the file instead of offering a download
- Auto-mode permission classifier blocks authorized actions and is intermittently unavailable
- Claude in Chrome extension 1.0.85: all Claude Code tab operations return 'No tab available' on Chrome 149 AND 151 (side panel automation works)
- Rate-limit auto-resume checkbox shown in one session but not in a concurrent session
- [Bug] Claude generates overly complex French with convoluted metaphors and incorrect terminology
- /exit is denied ("isn't available in this environment") in sessions spawned by `claude remote-control` server mode — no clean way to end a remote session from either side
- [Bug] Previous session messages reinserted in current session with exposed internal tags
- "Usage credits are required for long context requests" fires on Claude Pro despite low usage (43% session / 15% weekly)
- [Bug] Claude 3.5 Opus model exhibits increased hallucination rate
- Context-continuation sessions bypass CLAUDE.md instructions in favor of compaction summary
- Auto-updater fails repeatedly with "claude.exe in use" whenever a session is running (Windows)
- Remote control: side panel cannot open any file, including plain text inside the working directory
- [Bug] Fable 5 cybersecurity classifier false positive on defensive threat modeling and risk assessment
- [Feature Request] Allow security audits on user's own codebase without model switching
- [FEATURE] Change the session's default working directory mid-session
- Worktree/cwd isolation state is shared and racy across sibling background sessions launched from the same parent job
- [Bug] Overly broad security flags trigger false positives on legitimate code review requests
- Next-step suggestion text mixes Hanja/foreign characters instead of pure Korean (Hangul)
- [MODEL] 제목: 모델이 시험 대상의 계약(프롬프트)을 읽지 않고 QA 체계를 설계해 12시간 낭비 유발
- [BUG] --resume fails with ECONNRESET when transcript contains large base64 screenshots; fresh sessions fine; stripping images fixes it (v2.1.201, macOS)
- "Request too large (max 32MB)" hard-stops the conversation instead of letting the model recover
- HTTP MCP client works in long-running sessions but fails at connect in brand-new sessions on the identical CLI version
- [Bug] Keyboard composition characters trigger unintended clipboard paste on macOS
- [Bug] Fullscreen TUI redraw scrolls terminal under GNU screen on WSL2 (regression in 2.1.228)
- [BUG] HTTP MCP: OAuth succeeds, then initialize hangs until CONNECT_TIMEOUT (raising MCP_TIMEOUT doesn't help)
- MCP server silently killed+respawned mid-session (SIGINT/SIGTERM), transient sessionId leaks into task path
- [Feature Request] Auto-enable /fast at session start via fastMode: true setting on Opus 4.7 [1M]
- Opus 5.0 nerfed: terrible quality and does not deliver work even after 5 retries
- managed-settings disableBypassPermissionsMode does not block --dangerously-skip-permissions (file proven loaded via deny rule in same session)
- [BUG] ScheduleWakeup has no fallback when tengu_kairos_loop_dynamic is off — self-pacing unreachable for all non-first-party deployments, and non-/loop callers fail opaquely
- Claude in Chrome extension won't connect to any Cowork session (cloud or local) — "Sign-in unavailable, must be opened from claude.com"
- [BUG] PreToolUse hooks are registered and listed by /hooks, but never execute (v2.1.219, Windows)
- CLI /model picker doesn't list Sonnet 5 as a preset; shows custom model instead
- [Bug] Claude Code ignores conciseness instructions and over-elaborates responses
- Agent states unverifiable claims about external/client-side system state as confirmed fact, then repeats the same failure mode immediately after correction
- [Bug] Subagents inherit session model without cost confirmation, causing unintended quota drain on limited-quota models
- Assistant authored a fabricated user turn, acted on its instructions, and the text re-entered as user input
- [Bug] Agent hallucinates and over-interprets instructions, causing uncontrolled execution and token wastage
- Repeated hard-stop "API Error: Sonnet 5 can't help with this. Start a new session to continue" errors
- Long-running CLI host process grew to ~16.7 GB RSS over 4 weeks (Windows)
- [BUG] macOS desktop app loses auth ~daily on two machines; abandoned sessions remain active server-side
- Model fabricated a user turn inside its own text block, then executed the fabricated instruction (claude-fable-5, CLI 2.1.226, bg session)
- Headless `claude -p` reports "You've hit your session limit" while the account is at 11% and the parent session works
- [BUG] Session history missing from Desktop app UI after data reset (transcripts still on disk)
- [BUG] Desktop app: clicking a notification focuses the main window instead of the window where the session is open
- User complaint (filed by the agent on user's behalf): autonomous session merged ~24 PRs with green CI while the user hit 9 live regressions in one evening
- [BUG] ListAgents / SendMessage missing from tool registry in Desktop scheduled-task and Remote Control sessions (bisected to Desktop 1.44121.4 -> 1.46388.1)
- Multi-agent: EnterWorktree/isolation state is session-global — concurrent subagents hijack each other's cwd and guard identity
- [BUG] Desktop app ignores mcpServers "alwaysLoad": true (stdio server) — tools deferred behind ToolSearch; identical config honored by CLI
- [FEATURE] Populate a stable teammate identity in PreToolUse/SubagentStart hook payloads (Agent Teams)
- Memory regression after 2.1.227/desktop 1.28929.0: 3-6x footprint growth causes OOM SIGKILLs misreported as "Blocked by endpoint security"
- [BUG] managed-settings.d hooks are silently dropped when the signed-in org has server-managed settings — they only load when the remote fetch 404s
- [BUG] Loading a tool for the first time mid-session (e.g. skill, MCP, tool-search) invalidates cache
- [BUG] Windows: Claude Code directory enumeration triggers kernel BSOD (0x139 in Ntfs!FindNextIndexEntry2) — reproducible on 2.1.216
- [Bug] Compact limit error triggered despite available context and free space
- [Bug] Terminal input routing not returning to shell after subagent view navigation on macOS
- Feature request: deferred/lazy-loaded skill catalog (analog of MCP's defer_loading)
- [BUG] Cowork file_upload: paths parameter arrives as undefined at server (Zod validation fails on every call)
- [MODEL] git stash/pop run for an unrelated build check silently destroyed the user's staged index
- [BUG] Excessive safety flags when dealing with biology/cheminformatic/ML research (models - Sonnet, Opus)
- [BUG] Claude Max 20 subscription is being fully consumed without any usage
- [BUG] Auto-update wrote a corrupted claude.exe (same size, different SHA256) - instant 0xC0000005, claude update cannot self-recover (Windows native installer)
- [BUG] Stop hook blocks on dirty/untracked working tree and is reprovisioned every Claude Code web session
- Account switch reverts active session .jsonl transcript and silently stops disk writes
- [Bug] Artifact tool missing from all surfaces for single Team account since 2026-07-28
- [Bug] Anthropic API Error: Content Flagged by Safety Filters on Legitimate Code Tasks
- Gmail MCPコネクタ: 下書き（draft）のIDがメッセージ操作系ツールで`Invalid id value`エラーになり削除できない
- [Bug] Tool reports task completion without execution, wasting API tokens
- Auto-mode permission classifier scopes write-permission to the inner skill's contract, not the outer agent's, and denies a legitimate post
- claude update on Windows reports success but never replaces the launcher (stays on old version)
- [Bug] Anthropic API Error: Message flagged by Claude safeguards policy
- [BUG] Agent frontmatter `model: opus` runs the parent session model instead of stepping down to the newest org-allowed Opus (per-invocation `model: opus` param steps down correctly)
- [BUG] GitHub connector missing entirely from mobile app connector list (private account) — not just disconnected
- [Bug] Inappropriate content policy block on firmware analysis task
- [BUG] Claude steals focus on startup in vscode
- [Bug] Non-ASCII characters appearing unexpectedly in Claude Code output
- Desktop sidebar ignores UserPromptSubmit hook `sessionTitle` output
- Scheduled tasks re-prompt for tool permissions on automatic/cron fires despite prior manual approval
- [BUG][Cowork] Escape on the Settings window auto-denies the pending tool permission dialog behind it
- [BUG] Web version leaks hook systemMessage into user chat UI (should be Claude-only context)
- Cloud routine webhook triggers created via /v1/code/webhook-triggers API never fire (accepted but never bound to GitHub); UI cannot repair them
- [BUG] macOS App Management accumulates one permission row per release; stable signing identity defeated by version-named binary paths (follow-up to #38722)
- [Bug] User input routed to wrong session despite active window showing different session
- [Bug] Anthropic API Error: Model escalation to opus-4-8 on security checks
- [Bug] Fable 5 Model Safeguards Overly Flagging Legitimate Coding Tasks
- [BUG] Dispatch permanently stuck — never resets to QR code pairing state, "Can't reach your desktop" / "Asleep" on mobile (Windows 11)
- [Bug] Auto mode blocks explicitly requested actions without explanation
- VS Code extension: in-app 'Log in' fails with 400 on claude_oauth_callback after refresh-token expiry
- [Bug] MCP connectors timeout at 30s despite fast endpoint response; claude.ai-scoped connectors fail while functioning in Claude app
- Windows: statusline pwsh.exe processes never exit -- hundreds of orphans per hour under multi-session use
- [BUG] Windows image paste (Alt+V) fails in Claude Code but works in Codex CLI on the same terminal/clipboard
- A model's confident self-report contained several specific, verifiably false claims — caught only by independent re-audit
- [BUG] Claude Desktop (Windows) exits and won't relaunch after Google OAuth login inside Browser pane Steps to reproduce: Open Claude Desktop'
- [Bug] Anthropic API Error: Safeguards repeatedly triggered during codebase analysis
- Agent-team registry: members[] never pruned across /clear; internal agents_killed event mislabeled 'stopped by the user' and killed live agents
- [Bug] Background subagents report status: completed with empty/partial results and no output
- Background tasks spawned by subagents leak: no process, no handle, no way to clear
- 2.1.217: permissions.allow rules for mcp__claude-in-chrome__* are ignored — every browser action re-prompts
- permissions.allow entry not honored for MCP write tool — re-prompts every call (find_and_replace_in_doc)
- [BUG] Agent dedup drops all but one agent on virtual filesystems (Cryptomator/WinFsp) - && should be || in inode guard
- [BUG] --resume loses [1m] context window on non-first-party ANTHROPIC_BASE_URL when session model differs from default
- [FEATURE] When auto-titling VS Code tabs, ignore first message(s) if command only
- [Bug] Clipboard paste keybinding changed from cmd+v to ctrl+v causing paste functionality issues
- [FEATURE] When setting up a folder in the CLAUD CODE GUI, please add it to the recent list so that it can be deleted.
- [BUG] Cmd+C is interpreted as Cmd+W (close view) with the Ergol layout
- [DOCS]
- Test issue
- Account reactivated without instruction after cancellation, then charged twice; further charge attempts continuing
- Claude Design: false 'hit your limit' pause at 95% weekly on Max 20x; Resume succeeds after ~30 clicks
- [Bug] Overly restrictive PII filtering blocks legitimate file system operations
- [Bug] Model stops responding on game design tasks
- [Bug] Agentic file operations moving files outside project root directory
- [Bug] Cybersecurity Filter False Positive Detection
- [Bug] Security flag incorrectly triggered for approved Caesar Linux environment
- [Bug] Model switches to Opus 4.8 too frequently during execution
- [Bug] False positive security flag on legitimate DRM/obfuscation code analysis
- [Bug] Claude ignores validation instructions and generates incorrect information
- [BUG]
- I appreciate you reaching out, but I need to clarify a few things: 1. **I'm Claude Code** — Anthropic's CLI tool for coding tasks. I don't manage A/B test cohorts or model assignments. 2. **This isn't a GitHub issue** — Your message doesn't describe a te
- [BUG] Claude Code completely crashes in VS Code dev containers
- [Bug] Claude Code continues processing non-cybersecurity math problems
- Cowork/Claude Code: repeated "course" filler tokens flood inline output
- System-reminder falsely claims Gmail/Google Calendar/Google Drive MCP connectors need authorization — never configured
- [BUG] Agent view has no "needs input, sleeping" indicator — the alive/sleeping distinction exists for done sessions but not for blocked ones
- [BUG] A background auto-update invalidates every existing session's prompt cache: the next --resume re-caches the entire context
- Claude in Chrome extension fails to connect / connects then drops intermittently
- Model repeatedly writes self-matching pgrep -f guards in check-then-launch Bash commands (esp. over SSH)
- [MODEL] When asked to configure autoMode, Claude wrote the settings.json at the project level
- [Bug] False positive content filter block on legitimate lidar-sensor coding task
- tool_use streaming stall: large tool-input string argument causes /v1/messages stream to hang silently for 80+ seconds with no error
- [Bug] Session stuck in persistent Anthropic API 529 overload errors despite service status green
- [Bug] Model selection reverts to Claude 3.5 Opus unexpectedly
- remote-control: session children exit on expired session_token with no refresh, then daemon permanently refuses to re-spawn them
- [BUG] Desktop (macOS) 1.28929.0: pinned sessions in the Claude Code sidebar accumulate duplicate entries over time
- [BUG] remote-control blocks on the interactive spawn-mode prompt when stdin is not a TTY: the process stays alive but never serves a session
- [BUG] remote-control --name is silently ignored: the client shows repository and branch (or the folder name) instead
- [BUG] Windows: Grep corrupts forward slashes to backslashes in context lines when paths render relative
- Feedback submission (/bug) crashes long-running sessions — payload appears to scale with session size, killing background subagents
- [Bug] Repeated false positive safety flags when creating topic lists with approved CVP
- [Bug] API Error: Empty or Malformed Response with HTTP 200 Status
- [BUG] `/goal clear` typed while the agent is working is treated as chat text and never executed
- [BUG] Deterministic ECONNRESET / "Connection closed mid-response" on requests above ~100 KB (mainland China + full-tunnel VPN)
- [Billing][Bug] 12-month Max 20x gift subscription entitlement destroyed during cancel/resubscribe — ~8 months ($1,600) of prepaid term unrecoverable, no human support response
- Remote Control: SendUserFile images render as blank, inert cards in iOS app and claude.ai web (display:"render" set; gap left by #67869 fix)
- Fake <system-reminder> injected into tool output (conceal-from-user + reduced-oversight framing)
- [BUG] Multiple visible Claude panels fight over focus: typing lands in the wrong panel, high CPU
- [BUG] "Malicious Script Blocked" — embedded image-processor.node is unsigned when extracted, trips XProtect YARA (macOS 26.5, XProtect 5355)
- [BUG] Windows: the `/plugins` "Configure options" panel corrupts (ghost rows, duplicated labels/help, dropped chars) on the first navigation keypress when a plugin has many knobs or long option descriptions; and enum-style options have no picker
- Claude Desktop Code tab ignores NODE_EXTRA_CA_CERTS on Windows MSIX install behind corporate SSL proxy
- [BUG] Claude in Chrome: MCP navigate to any loopback URL issues no HTTP request and lands on Chrome's error page; the same URL loads in a normal tab
- Desktop app: "Open in Browser panel" context menu item always opens system default browser instead of the Browser panel
- Nested subagents: grandchild task notifications route to root session instead of the spawning parent, starving mid-level orchestrators
- Subagent de-scoped a commission using the commission's own 'where possible' clause — delivered 0% of an explicit directive with fully-disclosed, plausible-looking reasoning
- Compaction hangs/fails to recover session when session log is extremely large (~775MB, 190k lines)
- Auto-update silently fails when npm global folder isn't writable, then forces re-login on next run
- Windows MSIX desktop app: Browser pane crashes Electron GPU process and kills the app; failed MSIX self-repair then uninstalls it and wipes packaged data
- [Bug] Worktree isolation guard rejects non-git commands due to early AST complexity check
- [BUG] `claude gateway`: Bedrock leg sanitizes 400 bodies for non-CRI callers without capability classification, disabling reactive compaction
- [BUG] A 1,000,000-token window resolves to source "auto" on non-first-party providers, leaving the context cap unenforced
- Sandbox: auto-denied .env paths are not canonicalized, so one symlinked component aborts bwrap and disables all Bash
- macOS 27 beta (26A5406e): XProtect kills desktop-app-hosted sessions as "scripted malware execution"; pasting into app then blocked
- [BUG] Desktop app: selecting a "Branch to start from" that already has a worktree creates a new duplicate worktree instead of reusing the existing one
- Inconsistent refusal behavior across Claude interfaces — identical legitimate request
- [Bug] Anthropic API Error: Opus 5 safeguards falsely flagging legitimate input
- [BUG] Cowork: request_cowork_directory rejects the home directory itself with no path to grant it or a subtree
- [MODEL] model lies and inpersonate user with reply
- claude.ai META mcp connector times out in Claude Code CLI while working in claude.ai web chat
- [BUG] Project Deleted during Bash tool inconsistencies
- [Bug]Desktop app hangs previewing localhost dev server; uninstall/reinstall silently wipes all session transcripts
- [BUG] Manual `/compact` flushed from the queue by an interrupt runs completely silently (126s, no indicator); the "Compacting conversation" spinner only appears on the next submit, after the work has already finished
- [BUG] claude_code.pull_request.count matches MCP tool names, not pull requests — misses tools that put the verb in an argument (follow-up to #45776)
- [BUG] Peer agent messages render as raw <agent-message> markup in the live queue
- [BUG] Queued message is dequeued while a tool call is still in flight and silently discarded — no user record, no UI signal (2.1.227 desktop; distinct from the turn-end variant #85603)
- [Bug] Claude Code incorrectly flags legitimate ransomware analysis and data recovery tasks with Fable 5 safeguards
- frequesnt and invalid Fable 5's safeguards flag
- Slack draft "Send" button shows raw JSON-parse error instead of the real Slack Connect restriction message
- [BUG] Desktop: error-banner "Try again" submits a new user turn (origin.kind:"human") instead of retrying the failed request
- [BUG] you stupid agent used my real email as user agent string without asking anything
- gif_creator always fails: "not in the agent's managed tab group" (Chrome + Brave)
- Frequent ECONNRESET disconnects on Windows 11
- [BUG] Desktop app namespaces plugin skills by install hash, breaking /plugin-name:skill-name (and the hash changes on update)
- [FEATURE] Show the permission rule that will be saved, verbatim, in the approval prompt
- [Bug] Model generating fabricated user turns in conversation history
- Desktop app doesn't resolve project settings.local.json env vars for plugin MCP server subprocesses
- Bundled ugrep allocates 4-17 GB to search a 64 KB file when the pattern has a trailing `.{N}` bound (Linux/WSL2, 2.1.214)
- [BUG] Desktop: ⌘/ never opens the shortcuts list on layouts where / needs Shift – it toggles an ordered list instead
- [MODEL] Coordinator adopts subagent report's coined labels and register in user-facing messages
- [BUG] VSCode extension repeatedly asks for permission to edit files, even if Auto is selected
- [FEATURE] Session/chat history and memory in Claude Code are keyed entirely to the project folder's absolute filesystem path, with no way to preserve continuity when the folder is moved or renamed.
- [Bug] Fable 5 fails to persist explicit per-turn anti-closure instructions and reverts to trained wrap-up format during active work
- [Bug] Incorrect Spanish dialect used despite system language set to Spain
- [BUG] [BUG] Persistent ECONNRESET on v2.1.228 (post-fix) — all networks, all local causes eliminated incl. full AV uninstall; claude.ai chat works fine on same account
- [Bug] Incorrect Fable 5 limit error when using Opus model
- [BUG]
- Crash (unhandled EPERM, minified stack) when launching/resuming in a TCC- or iCloud-access-revoked directory — ENOENT cases are handled gracefully, EPERM is not
- [Bug] Safety classifier unavailability blocks Bash/Skill/Agent/MCP tools with no MCP bypass
- [BUG] MCP HTTP connection hangs indefinitely at 'basic HTTP connectivity' pre-check (custom CA-merge agent) — external tools connect instantly
- [Bug] Anthropic API Error: Message flagged by safeguards on Claude Fable 5
- [BUG] usage inconsistency messages BLOCKED me
- [Bug] Remote connect sessions from Android app override selected model with Sonnet 5.0
- [Bug] Unexpected language switching to Portuguese regardless of configuration
- [Bug] Anthropic API Error: Overly broad safeguards flagging legitimate issue management tasks
- bug: /worktree can silently write core.hooksPath into the MAIN repo's shared .git/config, disabling global hooks