Digest / September 3, 2026
Claude Code Issues Digest — September 3, 2026
283 new issues filed ·
678 resolved in anthropics/claude-code (UTC day).
New issues (283)
- Claude Desktop: interrupt during in-flight response hard-kills the whole session (live background tasks included) instead of aborting gracefully
- Browser pane: opening any overlay (Annotate, settings, viewport) zooms the page to 2x until it is dismissed; recurs on every activation
- [Bug] Claude incorrectly claims bash instruction for file editing operations
- [BUG] Cache-read is 96% of token volume (118x output); fan-out multiplies turns with no pre-execution projection — Max 20x drained in <48h
- [Bug] Anthropic API Error: Content flagged by safeguards despite legitimate cybersecurity task
- [Feature Request] Add writable MCP tool to assign sessions to sidebar groups
- Fable 5.1 Incorrectly Flags ESP32 + W5500 Development Board Test Code as [cyber]
- Claude Desktop (Windows, Store package) app window exits silently while a session idles with background subagents; no crash dump, no event-log entry
- [BUG] Claude Desktop (Windows) crashes during self-update while sessions are active, then fails to relaunch
- [BUG] Conversation ID 215475654509312 Cannot log in to Pro Account
- [Bug] Inconsistent CTRL-C behavior: sometimes interrupts session, sometimes clears input
- [Bug] Anthropic API Error: Safeguard flag triggered on internal development request
- [Feature Request] Support security audit workflows with Fable integration
- Keyboard scroll bindings are routed through the mouse-wheel acceleration state
- [Bug] Persistent login prompts despite repeated auth logout/login cycles
- --resume session preview renders corrupted when the preview is taller than the terminal
- Cross-machine session messaging (Remote Control) reports "delivered" but neither side receives; inbound message rendered in the chat UI never reaches the model's context
- `claude agents`: list selection snaps back to the just-exited session ~3 s after returning with `←`, discarding arrow-key navigation (regression in 2.1.257)
- [BUG] Cloud session started from the Desktop app has no `gh` CLI, although the cloud-environments docs list `gh` as pre-installed
- [FEATURE] VS Code sidebar: let "Account & usage" show the usage meters without the Account/email block
- [BUG] macOS: "Virtual Machine Service for Claude" causes cursor to freeze and drift every ~5s; killing the process stops it instantly
- [BUG] Sandboxed `mktemp -d` fails on macOS (ignores `$TMPDIR`); the empty variable made `rm -rf "$T"` run against the working directory with no critical-path prompt
- I need more context to generate an appropriate GitHub issue title. The message you provided appears to be in Korean and says "Imagining... It's been like this for 40 minutes" but doesn't describe a specific bug or feature request. Please provide: 1. What
- [Bug] Anthropic API Error: Invalid Request ID Format
- [Feature Request] Default to American English spelling for US-locale users
- [BUG] Cowork sandbox: /sessions never garbage-collected — 1,634 dirs over 4 months filled the disk; all scheduled tasks fail at `useradd` exit 12
- [BUG] Read(**/.env) deny rule prompts on every rg directory search, though rg skips hidden files by default
- [BUG] Claude Code generated an agent pipeline with an unbounded API retry loop and an ineffective spending guard, resulting in ~$50 of unintended OpenAI usage
- bypassPermissions mode now prompts on `cd DIR && grep …` when a Read() deny rule is configured (regression in 2.1.259)
- Ctrl+E "explain command" shortcut removed between v2.1.236 and v2.1.259
- [BUG]
- Sticky mid-session downgrade off Fable 5.1 takes away the model the customer pays for; a churn risk on the 20x Max plan
- Feature request: voice recognition / dictation for chat input
- Feature request: display timestamps on chat messages
- Context window balloons at fresh-session start (~40-70K baseline tokens) and full-file re-injection on every auto-compact
- grep -r and cp -r still bypass Read() deny rules on 2.1.259 (changelog fix does not reproduce; Windows/VS Code extension)
- [Feature Request] Reduce formulaic and unnatural phrasing in Opus 5.1 responses
- [false alarm] resumed leftover test session mistaken for input injection
- Expose per-request latency breakdown (upload / server queue / generation) for diagnosing slow turns
- Auto-mode classifier blocks gh pr merge inconsistently, including harmless --help
- Inline code / accent color has poor contrast on dark theme — no way to customize
- [BUG] Session Context - Total Loss of Data
- [BUG] Desktop (Linux/KDE Wayland): window sets demands-attention on every focus loss, pinning auto-hide panels open
- [Bug] Model generates unsupported generalizations from insufficient sample data
- CLAUDE.md @import cascades into plain relative-path mentions inside imported files, ignoring non-`@` syntax and "do not auto-load" notes
- Subagent composed permitted operations to bypass a permission block, causing SSH self-lockout on a remote host
- Artifact tool cannot delete comment threads, so commented artifacts can't be shared publicly
- [Bug] Anthropic API Error: Request ID req_011CefwDCXD4ELdWBWc78ZvX flagged during login form generation
- [BUG] 2.1.236: silent OAuth token refresh fails behind an https:// CONNECT proxy (ERR_SOCKET_CLOSED) while API traffic succeeds
- [Bug] Persistent upgrade notification for current version
- [BUG] A pasted image is re-rendered as a file pointer on the next request: one full prompt-cache rewrite, and the model loses the image
- [BUG] A persisting `cd` in a Bash tool call re-resolves project context and rewrites the system prompt, invalidating the entire prompt cache (5 captures, ~580k tokens each)
- [BUG] Turn-boundary prompt-cache misses on Fable 5.1 / Claude 5: 7% → 29% between 2.1.224 and 2.1.252, mechanism captured, fixed in 2.1.259 — with two residuals and one default worth changing
- Windows/VSCode: concurrent sessions race on OAuth refresh against the file credential store (400 → forced re-login)
- [BUG] False "Context limit reached" when well within limit
- Fable 5.1 ignores explicit rejected Agent tool results, reinterprets scoped rules, and continues until separately stopped
- /feedback missed a queued draft on first invocation and the draft was lost (a later invocation worked)
- Let the Code tab's usage ring report session context, not only the 5-hour window
- [BUG] error 0x80073CFF Claude Setup
- [Bug] False positive security flag during penetration testing
- Let a subagent choose its own advisor model (per-agent advisorModel, or one rung above the caller)
- [BUG] /auto-mode-setup skips cross-project inputs and reports a "just this project" scope the user did not select
- [BUG] Remote Control fails with HTTP 403 after Desktop app update; /remote-control retry does not recover
- [BUG] Cowork: Outputs file download fails with 404 when the filename matches a Project doc name
- [BUG] Session does not use selected model (it uses another model), model display/select inconsistencies (VS Code extension)
- Expose a pin_session MCP tool for session management
- [Feature Request] Support for reduced API usage limits in daily workflow mode
- [FEATURE] Auto-group sidebar sessions by title prefix (rule-based, with default-collapsed groups)
- [Bug] WebSearch quota exhausted without warning or remaining balance visibility
- [BUG] Routines stuck permanently after unattended run hits an unanswered permission prompt — "Run now" disabled on all routines, new routines can't be created either
- [Bug] Anthropic API Error: Overly strict biology safety classifier triggering on non-biology tasks
- [FEATURE] Add token usage warnings for potentially inefficient solution paths (aka: falling down rabbit-holes)
- [Bug] Context window management causes contradictory responses and inconsistent instruction adherence
- [Bug] Security system triggered during research file operations
- Assistant emits a fabricated `user` turn at the end of its response and executes it on the next turn — two occurrences, neither in long context
- User message and system<total_tokens> counter rendered inside the assistant's message block
- [BUG] Extension icons not rendering (blank squares) after update, in Remote-SSH session
- [Feature Request] Enhanced physics and 3D graphics reasoning capabilities for graphics research
- [Bug] Rewind does not restore file content for untracked files edited in session
- [Bug] Unexpected automatic model switching to Opus
- Non-ASCII project paths collide in ~/.claude/projects/, silently sharing memory and transcripts between different projects
- [BUG] Claude Desktop (Windows 11 25H2) fails to launch with misleading "Another program is currently using this file" — 0x80070020 Silo/Job failure with cowork-svc.exe in an SCM restart loop; only reboot recovers
- [Bug] Claude Opus prematurely terminates task execution ignoring system prompt constraints
- Auto-memory is on by default, ignores claude.ai memory settings, and is scoped per working directory so unrelated projects share one memory
- [Bug] False positive biology detection for Quickscript code
- [BUG] Desktop app (Windows): Japanese IME text is scrambled when typed over Chrome Remote Desktop — kanji stripped inline and re-appended at the end in reverse order
- [DOCS] Remove obsolete Python can_use_tool dummy-hook workaround
- [BUG] Claude Display Stuck
- [BUG] [BUG] AddPackage fails with 0x80073D28 - current-user install path cannot register packaged CoworkVMService, even after removing orphaned service
- [Bug] Safety guardrails incorrectly blocking authorized local network audit with valid credentials
- [BUG] Dispatch cannot start a second Code session in the same folder since 1.44121.x: exclusiveCwd hardcoded true in startCodeSession
- [BUG] Subscription auth (Pro/Max) ignores ANTHROPIC_BASE_URL — requests never reach custom proxy (LiteLLM)
- [BUG] --model claude-fable-5 / claude-fable-5-1 silently served claude-opus-5
- Routines prompt for approval on pre-authorised connector tool calls
- Claude Desktop (Windows/MSIX) recurring "Another program is currently using this file" on auto-update, requires OS reboot to recover — affects multiple users
- [BUG] Claude Desktop (Windows MSIX) registers itself as handler for .docx/.pdf/.csv and becomes default with no opt-out
- [BUG] Regression in 2.1.259: rg reading stdin in a pipeline is treated as a search of '.' and prompts (auto mode bypassed) when a Read() deny rule covers a file under the cwd
- [BUG] Session "resume" / handoff notes get silently treated as the authoritative task list, causing severe compression of open work across session boundaries
- [Bug] Custom title persists after /clear command
- auto mode classifier denies gh pr merge (and read-only gh search issues) it previously allowed
- Conversation sidebar not navigable with stylus or keyboard - resizer covers scrollbar, list has no focusable container
- [BUG] Slash-command menu leaves a permanent blank region below the input after it closes (classic renderer / `CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1`)
- [Bug] CVP Organization Not Recognized - Model Downgrade to Opus Despite Approval
- Cowork scheduled task ("Routine") fires on schedule but takes hours to actually execute — matches known closed issue #36131
- [Bug] Auto mode not automatically approving safe bash commands
- [BUG] / skill picker resolves to the wrong project when two open projects define a skill with the same name
- [BUG] [claude deskop] The `--user-data-dir` parameter has no effect on child processes.
- [BUG] Code tab Overview: "All" totals shrink over time — today's all-time count is lower than 3 months ago
- [BUG] Windows/MSIX: `git fsmonitor--daemon` spawned by Claude Code inherits the AppX container job, survives the update's forced shutdown, and blocks relaunch of the new version (0x80070020). Root cause + no-reboot workaround
- [Feature Request] Enable Claude to provide feedback on instructions before execution
- [BUG] OTEL_LOG_RAW_API_BODIES Ignored in project settings since 2.1.251 Why would we be forced to export the raw logs of ALL our project if we only need one?
- [FEATURE] Add model, stop_reason, token usage and tool timing to hook payloads
- Agents view: Ctrl-X unbinds a session — sending to that row creates a job that never binds and never runs
- Background-agent adoption never succeeds (0 of 7 attempts): live sessions marked Failed, reap and daemon death unlogged
- [FEATURE] Profiles within one account (separate history, memory, config) with profile-only sign-in for client machines
- [BUG] Windows/Git Bash: git <ref>:<dot-path> arguments silently rewritten, so a mangled command looks like a missing file
- [Bug] Security breach check request incorrectly flagged as non-legitimate
- [Bug] sdk_stream_ended_no_result over VS Code Remote-SSH (related to #59374, locked)
- [BUG] Linux/Wayland: IDE extension auto-install spawns empty Cursor windows on every launch
- [BUG] /usage Stats tab counts message.usage per transcript row instead of per message.id, inflating token totals ~2x
- Bash Read() deny rules prompt on cd compounds even when the cd target is a literal path (2.1.259)
- claude -p silently bills a project .env ANTHROPIC_API_KEY while a valid subscription credential is active — documented safeguards do not fire in non-interactive mode
- `grep -r` with `--include` filters falsely triggers `Read(.env)` deny prompt (v2.1.259) — breaks every recursive search
- Windows: CoworkVMService fails to configure crash-recovery ("Access is denied") — service stays down until full reboot, blocking Claude Desktop startup
- VS Code extension: sessions relocated into .claude/worktrees project dirs vanish from history after a reload
- [BUG] Subagent with `hooks: []` (empty array) in frontmatter is silently dropped from the agent list (`--agent '<name>' not found`, no warning)
- [Feature Request] Add session deletion capability to manage accumulated session history
- [BUG] EPERM error when writing/editing files on SSHFS/FUSE-mounted filesystem
- [Bug] Auto mode still prompts for grep tool authorization despite being in Auto Mode
- HTTP MCP server: \ in --header Authorization not consistently resolved across tool calls in the same session
- Desktop app auto-update silently terminates long-running child processes started from a session
- Model invents an authorization requirement, refuses authorized work on it, and writes the false constraint into the user's repo
- [BUG] Claude code (VS code extension) is creating files with random names while working
- [BUG] Skills in ~/.claude/skills/ not discovered by Desktop Remote (SSH) sessions, even after full restart + new session
- [BUG] Desktop SSH remote sessions cannot authenticate to hosts that only accept OpenSSH user certificates (ssh2 has no certificate support)
- [BUG] Claude Code makes me sign in every day
- [Bug] Fable 5 incorrectly flagged on fisheries data analysis
- [Bug] Intermittent permission loss for project directories on macOS
- computer-use: ending screen control un-minimizes user-minimized windows and steals foreground focus (Windows)
- [BUG] grep -r over any directory prompts on a Read(**/…) deny rule even when the tree contains no matching file (2.1.259)
- [BUG] Semgrep plugin on WSL: registered hook `scripts/hook` runs the Windows binary; the `hook.sh` workaround from #69256 is now a no-op
- [Bug] Anthropic API Error: Model unavailable - "can't help with this" response
- settings.json env vars not propagated to background (bg-spare) session processes
- [BUG] Repositorio no aparece en el selector de Claude Code pese a acceso confirmado en GitHub
- [Bug] Anthropic API Error: 500 Internal Server Error
- [Bug] API Error: 500 Internal Server Error with certain prompt prefixes
- [Feature Request] Tool approval required in auto mode despite all tools being allowed
- [MCP tool boolean parameter always serialized as string, causing "Expected bool, got string" error]
- [Feature Request] Agent Foundry: Build custom agents with framework templates and skill library
- [BUG] No repositories available in Claude Code repo picker despite GitHub App installed
- [BUG] Windows: all shell output capture fails on VMware App Volumes — `stat` by path and `fstat` by handle report different `dev` for every local path
- You've hit your session limit · resets 5:40pm (Asia/Riyadh)
- [BUG] CVP Passport migration regresses verified IDV records to denied
- Claude ignores user steering messages sent during active tool execution
- [Bug] Anthropic API Error: 500 Internal Server Error with Claude 3.5 Opus
- [BUG] VS Code extension 2.1.259: Bash permission prompts fire despite bypassPermissions mode and a bare `Bash` allow rule
- HTTP MCP: "Error POSTing to endpoint" despite server returning a valid tools/list result (Google Docs MCP docsmcp.googleapis.com)
- [BUG] Ctrl-click on a link in Windows Terminal opens it twice (terminal opens it, then Claude Code opens it again)
- [BUG] New chats auto-enable Remote Control despite global setting being off
- Desktop app: allow remote session *starts* on a machine, with a device picker in the mobile app
- [BUG] Regression in 2.1.259: managed-settings discovery on WSL fails fatally when /mnt/c is inaccessible (works in 2.1.258)
- [Bug] Anthropic API Error: 529 Overloaded - Excessive Retry Attempts
- Account silently downgraded from Pro without action taken — billing/support issue
- [BUG] Cowork scheduled task hangs when using Todoist MCP tools in a background session
- Desktop: task chips (spawn_task) inject their prompt into the open thread instead of spawning a new session
- [Feature Request] Improve performance and UI responsiveness for VS Code extension
- Sessions should stay keyed to their launch directory — EnterWorktree relocates session history, making renamed/resumed sessions unfindable
- [Bug] Claude Code file access blocked for sensitive directories (passwords, SSH keys)
- Side chat panel hangs indefinitely ("Waiting for Claude...") on slash-command invocations
- [Bug] HTTP 529 Overloaded Error from Anthropic API
- [FEATURE] Keep the remote-control server name in generated titles of spawned sessions
- Persistent setting to collapse tool-call output by default (reopen of #40428)
- [BUG] Desktop app: No reliable way to create/maintain separate persistent sessions across multiple project directories (no folder picker, sessions don't persist after CLI closes, rename doesn't save)
- [BUG] remote_session_change attachment silently pushes commit/PR attribution settings that conflict with user CLAUDE.md
- [FEATURE] VS Code Extension: Display persistent monthly usage/budget telemetry in status bar
- Accessibility: computer-use policy blocks the Kindle app, cutting blind users off from books they own
- [DOCS] Claude Sonnet 5 System prompt missing from docs
- [FEATURE]: Better distinction between custom and built-in skills
- Claude Code CLI blocked — "Your organization has disabled Claude subscription access"
- Bug: Workflow tool — a worker relaunched via SendMessage is never re-registered in the workflow's tracking (and the inactivity watchdog kills legitimately-silent workers)
- Opt-in: compact the conversation at the end of a goal while the session is still cached
- Auto mode: read-only Bash compounds starting with `cd DIR` lost static auto-approval in 2.1.258, causing frequent permission prompts
- Last-sent message no longer stays pinned/visible at top in fullscreen TUI (VS Code integrated terminal)
- Desktop: finished same-machine scheduled-task sessions show permanent 'Remote Control disconnected' + blinking sidebar dot under 'Other' group
- Desktop app ignores global settings.json model field (1M context falls back to 200k)
- [FEATURE] Self-hosted environments for Pro/Max accounts
- [BUG] Foundry helper-script credential kind hardcodes .services.ai.azure.com, fails with misleading 401 when resource is private-endpoint-only
- [MODEL] Cyber safeguard false-positive blocks single read-only request during authorized pentest (see also #73896)
- [BUG] No way to switch Claude Code organization when running inside Antigravity IDE
- [Bug] Workflow fan-out multiplier not validated against agent-count guideline
- [BUG] MCP OAuth callback host is hardcoded — `localhost` breaks Atlassian/Azure allowlists, `127.0.0.1` breaks Salesforce/Slack; needs a host override
- [BUG] Headless `-p` suppresses plugin MCP servers in favor of same-URL claude.ai connectors, which cannot authenticate headlessly
- Read() deny rule arms an unwhitelistable cd+relative-read prompt in bypassPermissions (2.1.257+), and the predicate ignores the rule's patterns
- [Bug] Intermittent Anthropic API 529 Overloaded Errors with inconsistent session stability
- [BUG] Google Drive and Calendar connectors missing from Claude Code sessions; work on claude.ai web
- Claude Desktop (MSIX) fails to launch with "Another program is currently using this file" — requires full reboot
- [BUG] /model picker does not offer Fable 5.1 to an account that is entitled to it (2.1.259, setup-token auth)
- Read() deny rules force manual approval on unrelated Bash grep commands after cd with relative glob
- Claude Desktop: embedded claude-code engine at a versioned path breaks macOS TCC grants on every update
- [BUG] Approval-gated destructive MCP tool errors "not registered" on invocation, despite server confirming read-write mode and tool being fully implemented
- [FEATURE] Support interoperable project-level AI configuration conventions
- Agent SDK: rate_limit_event already carries per-bucket unifiedWindows utilization at status=allowed — declare it in SDKRateLimitInfo and add a zero-token structured path (follow-up to #50518)
- [Major Bug] 30% session exhausted in 2 seconds. Issue happening after the latest claude update(on fable 5.1 launch) Also Compaction invalidates cache on session resume, breaking cache efficiency
- [BUG]effort "max" silently discarded by 4-value validator on all persistence paths (unfixed 2.1.219 -> 2.1.258)
- [BUG] VS Code extension 2.1.259 (linux-x64): text selected in the chat panel copies nothing to the clipboard — regression, 2.1.258 works
- Injected system-reminder overrode explicit user no-attribution preference; not flagged before acting
- [Bug] Model unexpectedly switched to Opus during MCP server development
- Output style from a skills-directory plugin silently modifies every session's system prompt while /config reports "default"
- [BUG] MCP connector tools logged as "added" but unresolvable via ToolSearch for the rest of the session
- [BUG]
- [Bug] Anthropic API Infrastructure Issues and Model Performance Constraints
- AskUserQuestion: clarify path discards the selected option and the typed text (still present in 2.1.259)
- [BUG] Fable 5 missing from model picker after outage recovery (v2.1.259, VSCode)
- Bash tool background/timeout-promoted child processes on Windows never signal completion even after finishing and writing valid output
- Mods - make Claude 10x more extensible
- [BUG] Four shipped plugin agents fail YAML frontmatter parsing and load as placeholders or not at all
- [Bug] Subagent exceeded scope and executed unauthorized git commit without explicit instruction
- [BUG]
- [FEATURE] Plugins installed via claude.ai propegate to the Claude Code CLI
- Relative extraKnownMarketplaces directory paths resolve against the main checkout, not the active git worktree
- mentioned "meta mode for the human-agent-interface.com" but haven't provided:- A specific bug description or error message- What went wrong or what feature is nee
- [BUG] Magic-link and Google login both route existing Max subscriber to "Let's create your account" onboarding; mobile session and Console remain authenticated
- [FEATURE] Support localized/translatable spinner status words (i18n)
- Scheduled-task runs leak a resident claude process per fire (2.1.255, Linux): 106 procs / 26 GB in 25 h — regression of #54626
- [Bug] Excessive context re-sending in tool round-trips due to CLAUDE.md file inclusion
- [BUG] Windows native install/CLI hangs indefinitely with zero network activity (confirmed via Process Monitor)
- [Bug] Auto mode context handling misaligned with inquisitive queries
- [BUG] Cloud routine hangs on a permission prompt for the built-in Artifact tool, contradicting "routines run autonomously"
- Desktop scheduled tasks: model selection broken end-to-end (spawns ignore user settings; documented picker absent; MCP tool lacks model param)
- [BUG] /model picker offers no route to entitled Fable 5 when the catalogue's only Fable row is a restricted Fable 5.1 (2.1.259, Claude Enterprise)
- [FEATURE] Server-level tool allowlists in .mcp.json, so one remote MCP server (e.g. an enterprise gateway) can be registered as several per-project servers
- [BUG] Fable 5.1 Burning usage twice in less than an hour doing nothing complicated
- Task dependency graph and fix-round/review history for plan-driven multi-agent workflows
- [BUG] **Title:** Claude in Chrome side panel: every tool fails ("Tool X failed in the extension") — `browser_extension` API returns `org_policy: "block"`, `policy_source: "account_disabled"` on a personal account with no org policies (Linux)
- [BUG] Subagent frontmatter `model: claude-opus-4-8` silently ignored — subagent runs on the parent model (Fable 5.1) in 2.1.259
- Expose the beta advisor and memory API tools in Claude Code
- [Bug] Skills and standing rules not being respected by Claude Code agent
- Native ChromeOS release for Cowork — tiered sandboxed execution instead of requiring Crostini/KVM
- [Bug] Safety Classifier failing to block adverse content
- [BUG] Paid Pro subscription reverted to Free plan — invoice shows "Paid" but access downgraded
- [BUG] VS Code extension: bottom status-bar model chip shows stale "Sonnet 4.6" even though /model and the current model are correctly "Sonnet 5"
- [BUG] Claude Desktop MCP: misleading "server may be unresponsive" 4-minute timeout message, and application logging lost since ~1.34493
- Agent performed hours of unauthorized work, ignored its core process rule after eight corrections, and reported unverified results as fact
- [Feature Request] Add adjustable output window and expandable prompt input UI
- Opus 5 in plan mode: four planning rounds to land a one-file fix the user specified up front
- [Bug] Anthropic API Error: Model rejected request with AUP violation during mid-session task
- [Bug] Bash permission gate fails to resolve piped commands after directory change with Read() denies
- CLAUDE_CODE_SSE_PORT set → every API call fails with ConnectionRefused in VS Code extension (Windows)
- Fable 5.1 repeatedly ignores a standing "read the record before acting" directive (13 recorded instances in 15 days)
- remoteControlAtStartup is ignored under -p: no supported way to make a headless run viewable from mobile/claude.ai
- [Bug] Unintended model upgrade to Opus during gige camera debugging session
- I don't have enough information in this bug report to determine a clear technical issue. The text appears to contain placeholder or corrupted content rather than a coherent bug description. Could you please provide: - A clear description of the problem yo
- [MODEL] Sonnet and Opus, Confidently asserted an unverified, wrong answer on an irreversible Stripe account decision despite repeated explicit requests to verify first and many others issues
- Hooks: PreCompact/PostCompact/SessionStart(compact) fire for a subagent's compaction without agent fields; SubagentStop fires for internal summarizer calls with a never-created agent_transcript_path
- [BUG] OAuth login fails with "Invalid OAuth Request — Missing redirect_uri parameter"
- Desktop app: previously working sessions show "Session not found on disk" despite transcripts existing; "Import CLI sessions" reports nothing to import
- [Feature Request] Implement persistent context/memory across CLI restarts for multi-session development workflows
- [BUG] Native Linux build crashes (SIGILL/SIGSEGV) in JSC GC HeapHelper thread, 8x/day on 2.1.258, same Bun 1.4.1 in 2.1.259
- Remote control is never re-established after the idle-triggered auto-update relaunch, making headless machines unreachable
- gopls-lsp plugin installed+enabled but missing lspServers wiring; LSP tool never available for Go
- [BUG] Startup banner hint '+N more · /status' points at /status, which never shows the hidden notices
- [BUG] [Windows] Widget overflow menu: Download SVG/PNG are silent no-ops; Copy to clipboard now fails (app logs 'download-file not yet wired' + CSP font blocks)
- Status line rate_limits should include the model-specific weekly window shown in /usage
- [BUG] Windows: main window silently becomes always-on-top after stealth auto-update (z-order anchor restored behind the topmost taskbar)
- [Bug] Completed fork remains visible in TUI after completion
- Fable 5.1 child requests (subagents and forks) silently switch to `claude-opus-4-8` after the first tool result; main session on the same model is unaffected
- [Bug] Excessive token consumption without productive output
- [BUG] Dispatch: desktop pane renders no messages — sessions-bridge drops every user echo ("empty inboundUserMessages FIFO")
- [BUG] Dispatch shows "desktop appears offline" behind TLS-inspecting AV — sessions-bridge Node transport does not use the system CA store
- [Bug] Anthropic API Error: Tokens consumed without task completion
- [FEATURE] Alerts for enterprise per-user spend limits in Claude Code CLI
- [BUG] Image-paste separator space survives Shift+Enter and lands at the start of the next line
- [BUG] Tarea incompleta, bloqueos M365/auto-mode confusos y trato despectivo al usuario
- [Bug] Safety classifiers triggering on valid input
- After EnterWorktree, the Bash cwd pin does not follow, so bash refuses everywhere the agent is standing; a default-value echo trips the same guard as too complex
- [BUG] Cowork marketplace sync rejects a plugin with `source: "./"` and no plugin.json that the CLI validates and installs fine; Desktop UI hides the real error
- Prompt-section cache is keyed by name only, so a subagent on a different model receives the main model's identity and cutoff
- [BUG] Windows desktop: cowork session paths reach 472 chars (212 over MAX_PATH); 117 files already past the limit. Windows sibling of #80753
- [BUG] Mobile: offline local sessions are displayed under Archived, and unarchiving removes them from every filter
- [Bug] Grep commands require manual approval instead of auto-accepting, causing agentic loop failures
- [BUG]
- [BUG] Fable 5.1: final answer emitted as a thinking block (not text) before AskUserQuestion, so the user never sees it
- Session executed unverified EDL as fact, causing repeated costly regenerations on a music video task
Resolved issues (678)
- [Correction] August USD 99.08 report withdrawn after Platform/API reconciliation
- [Bug] Anthropic API Error: Invalid Request ID Format
- [false alarm] resumed leftover test session mistaken for input injection
- [BUG] Claude Code process exited with code 134
- [BUG] A pasted image is re-rendered as a file pointer on the next request: one full prompt-cache rewrite, and the model loses the image
- test
- Payment system completely broken — cards declined without any request reaching the bank
- [FEATURE] Multiple Claude Desktop windows
- I'm ready to help generate GitHub issue titles for Claude Code bug reports. Please provide the bug report details when you're ready, and I'll create a concise, technical issue title following the format specified.
- [BUG] Usage indicator (bottom-right circle) in Desktop App shows weekly Fable % instead of context window %
- [BUG] Desktop settings: "Switch models when a message is flagged" appears twice under Claude Code (General and Local sessions)
- [Bug] Over-restrictive safety filters blocking legitimate development tasks
- [BUG] Agents view: opening an existing session requires two clicks on the same row
- [Bug] Verbose output with incoherent sentence structure
- [Bug] Anthropic API Error: Content Blocked - File Access Requests Flagged Repeatedly
- [BUG] API Error: 422 {"detail":[{"type":"missing","loc":["body","messages",2,"content"],"msg":"Field required","input":{"role":"assistant","tool_calls
- [Bug] Anthropic API Error: Model version compatibility issue with Claude Opus
- [Bug] Anthropic API Error: Excessive Content Filtering on Benign Messages
- Chrome native-messaging host (claude.exe --chrome-native-host) grows to 46 GB RSS and freezes the machine (Linux)
- [BUG] The dedicated Cowork sidebar tab has disappeared.
- Feature request: Allow users to disable the 'ctrl+c to copy' hint in status bar
- [Bug][aup] Frustrated exclamation during defensive vuln scan of own nursing site wrongly blocked (req_011CccM8L3MDsy69M7qwCweJ)
- [BUG] Claude Code not choosing the correct org default model
- iOS: parallel permission-gated tool call hangs on busy signal (no prompt, no timeout)
- Skill whose name contains "mcp" shadows/overrides the built-in /mcp slash command
- Assistant output stream contaminated with stray synthetic data + fake system-reminder tag (persisted in assistant JSONL line); same-day phantom user message
- [Feature Request] Add PostWorktreeCreate hook and matcher support for selective worktree hook execution
- [BUG] VS Code extension: input field freezes (blinking cursor, no text entry) with multiple sessions open, tabs flicker rapidly
- Session becomes permanently un-resumable after usage-limit error (400 diagnostics.previous_message_id)
- [Bug] Model auto-routing incorrectly redirects legitimate scientific anatomy terms to Opus 4.8
- [Bug] Request flagged by safety system when asking Claude to review marketing strategy
- [BUG] CronCreate recurring job fires correctly once, then re-fires every 12-37s (regression of #49198 on v2.1.198)
- [Bug] Intermittent malformed tool-call emission with stray prefix tokens and missing tag namespace
- Malformed tool-call emission: literal "court" replaces <function_calls> opener, antml: namespace stripped, block stored as text (session hangs)
- [Bug] Non-English character input incorrectly flagged as policy violation
- [BUG] Cowork connected-folder sessions: file-delete tool no longer registers after desktop update 1.19367.0 (rm blocked, EPERM)
- Sub-agent sync-vs-async is session-host-dependent (not version/flag); run_in_background:false not reliably honored — no documented way to force synchronous execution
- [BUG]
- [Bug] Claude Code falls back from Fable to Opus model during any task!!! execution after a while without any reason!!!
- [BUG] Claude code - wrong model selected
- [BUG] claude -p crashes at startup under launchd when cwd is in ~/Library/CloudStorage (OneDrive) — every version after 2.1.128 (fd-limit catch-all / EPERM)
- Remote SSH: plugin archives >2.5MB arrive truncated, session hangs forever at "Setting up plugins..."
- Sessions run in git worktrees don't appear in the parent repo's history picker
- [BUG] Desktop plugin browser: personal marketplace git URL silently ignores #branch fragment — displays it, syncs default branch
- [BUG] Cowork: plugin stdio MCP server that exits at spawn is silently classified "needs auth" — stderr discarded, no error surfaced anywhere
- Fullscreen mouse capture degrades scroll and clobbers clipboard on macOS Terminal.app — three separable fixes
- [BUG] claude.ai plugin ingester silently hides plugins whose hooks.json has an unknown top-level key (tolerated by Claude Code CLI)
- [FEATURE] /clear creates a new session, causing clutter in claude --resume session list
- PreToolUse hooks intermittently not invoked at all (VSCode extension host, Windows)
- [BUG] Ref-pinned plugin marketplace is un-updatable: settings.json ref edit ignored, marketplace update fails with bare EPERM, plugin update silently reverts manual repair
- [BUG][A11y] macOS desktop: VoiceOver repeatedly re-announces the prompt/composer field during a turn (frequent re-renders)
- Feature request: composable session naming (naming conventions + project-aware AI titles)
- Deeply-nested subagent fan-out causes unbounded memory growth → host-level OOM
- [Bug] Unexpected model switch to Opus without user request
- [Bug] Assistant fabricates a "user" turn inside its own response; a later turn treats it as a real instruction (forensic JSONL evidence)
- [BUG] Fable 5: thinking blocks dropped every turn in live sessions → full conversation re-cached each exchange (looks like a recurrence of the March 26 clear_thinking keep:1 regression; Opus 4.8 unaffected) — CC 2.1.202
- error process exited with code 1
- [Bug] Remote control disconnects persistently after reboot and network switch
- I need more specific information about the error you're experiencing to generate an accurate issue title. Could you please provide: 1. The error message you're receiving 2. What MCP server you're trying to integrate 3. What steps you took to set it up 4.
- [BUG] Multi-root workspace: agent-emitted file links unclickable outside primary folder
- [BUG] Routines API: fire payload delivered as a second message, racing the routine's first turn
- Local stdio MCP server fails with "access denied for socket file" when spawned by MSIX-packaged Claude Desktop on Windows (works fine under Codex on the same machine)
- [Bug] False positive security flag on legitimate PostGIS database query
- MCP structuredContent validation appears to check the whole CallToolResult envelope instead of structuredContent only
- [Bug] Wakeup command injects fake "USER INPUT" leading to hallucination spirals
- [Bug] Anthropic API Error: Request flagged incorrectly for non-cybersecurity content
- [Bug] Anthropic API Error: Unexpected model downgrade to Opus for standard coding tasks
- Fable 5 refusal fallback (apiRefusalCategory: cyber) repeatedly triggered by routine curl/scraping; session silently pinned to Opus 4.8
- [BUG] False-positive cybersecurity safety block despite active Cyber Verification Program.
- Path-scoped --allowedTools 'Write(dir/**)' denies the first out-of-scope write but a retried out-of-scope write in the same session is not re-checked and succeeds (headless -p)
- [BUG][bedrock] claude-opus-4-8[1m] from /model picker writes invalid model ID → 400 "The provided model identifier is invalid"
- [MODEL]
- Remote Control: session permanently wedged when worker dies unregistered (phantom worker_status:running); no recovery path exposed
- Model hallucinated a fake domain in place of getcomposer.org for an installer command
- [Bug] False positive security flagging in debug mode for commerce systems
- [BUG] "your organization has disabled claude subscription access for claude code usee an anthropic api key instead, or ask your admin to enable access"
- Add configurable mouse interaction levels for interactive prompts (AskUserQuestion / permission dialogs) — clicks currently select AND submit in one action
- [Bug] Session context lost when switching models
- /model[Feature Request] Anthropic API Safety Filter Error with Model Name "Fable 5"
- [BUG] VS Code window reload fires SessionStart twice (throwaway source=startup + resume) — creates a phantom session hooks cannot distinguish
- [Bug] Anthropic API Error: Safety Filters Blocking Observability/Monitoring Workflow
- verbindung claude/ chrome
- [BUG] All tool calls become no-op, returning only the injected "Continue with your original plan." — recovers only after a full restart
- [BUG] Desktop app workflow panel: per-agent tokens/tools/time stay empty during long parallel runs — progress-only batches omit workflowProgress
- [BUG] Max 5x plan: 70M cache read tokens counted against quota — $62/day on a $100/month plan
- [BUG] Win11: TUI render-loop frame pacing degraded ~2× between 2.1.159 → 2.1.207 (bisected + instrumented; #67406 follow-up)
- [BUG] API Error: Response stalled mid-stream. The response above may be incomplete.
- [BUG] Task tools' per-session lockfile fails with EPERM (`utimensat`) when CLAUDE_CONFIG_DIR is on an SMB/CIFS mount (e.g. Azure Files)
- [Bug] Mobile scrolling in GitHub Codespaces web fills input field with NaN values
- [BUG] Windows desktop: all session transcripts deleted from ~/.claude/projects on first launch after 2.1.177 update — recents show "session not found on disk"
- settings.json overwritten with valid {"env": {}} — stale in-memory config re-persisted (not truncation) with many concurrent sessions on Windows
- [BUG] /remote-control silently hidden when DISABLE_TELEMETRY or DO_NOT_TRACK is set — gate blocks flag reads even when tengu_ccr_bridge is cached true (v2.1.199)
- [Bug] Claude startup blocked on macOS securityd certificate evaluation (~60s delay)
- Plugin cache accumulates old version directories; no pruning after update
- [Bug] Overly broad cyber safeguard blocking legitimate defensive security code review
- Statusline payload reports context_window_size: 200000 on 1M-context session (claude-fable-5) — used_percentage pegs at 100%
- [BUG]
- [Bug] Anthropic API Error: Content Moderation False Positive on Development Application
- Opus (1M context, claude-opus-4-8) intermittently emits malformed tool-call opening sequence — degrades to a common English word, causing parse failure
- [BUG] Interactive CLI silently exits after successful Bash tool result; session resumes normally
- [BUG] Claude Design preview: all project-relative subresources rejected with 401 "preview token required" (only top-level HTML gets a token)
- Remote Control sessions cannot survive a reboot / cannot re-attach from mobile
- Fable safety heuristic false-positives on legitimate defensive own-fleet security engineering
- [Bug] Unauthorized access to sensitive application code during non-security tasks
- [Bug] Requests consistently flagged as policy violation with Claude Fable 5
- Advisor always 'unavailable' in sessions with many deferred MCP tools — fails at ~56K tokens while slim-tool session succeeds at ~62K
- [BUG]New session fails with tool name validation error from a connected MCP connector
- Windows: hooks leak shell redirection - zero-byte files created in CWD from '>' tokens in tool input (e.g. '-> Signal:' creates file 'Signal')
- Desktop app: all emoji in assistant chat output render as blank gaps
- [Bug] LLM output contains inappropriate phrase "The money shot"
- [Bug] Anthropic Usage Reset (Codex orchestrator) Without User Consent
- [Bug] Tool invocations rendered as literal text instead of executed
- [BUG] Windows: preview_start always fails with "spawn cmd.exe ENOENT" regardless of runtimeExecutable
- SendMessage to a completed subagent doesn't reuse the previous prompt prefix (cache miss on every resume)
- skill-creator: read_text()/write_text() sem encoding=utf-8 corrompem conteúdo acentuado no Windows
- Windows: Cowork HTML preview pane ignores <style> blocks (inline style attributes work, text renders unstyled)
- [Bug] Safety classifier false-positive on legitimate DevOps terminology in authorized sandbox context
- [Bug] Model incorrectly restricted based on unrelated project content
- [Bug] Excessive Safeguard False Positives During Defensive Security Audits
- Write tool silently produces empty files; Bash output fabricated in long sessions (Opus 4.8)
- Shift+Option+Backspace (kitty CSI 127;4u) deletes to end of line instead of a word
- Sonnet 5: user-initiated `!echo $VAR` output misidentified as injected fabricated tool result when var name contains 'KEY' and value is UUID
- [BUG] Scrolling Claude Response with Keyboard Stopped Working
- [BUG] Windows版デスクトップアプリでセッション終了後もclaude.exeプロセスが回収されず蓄積し続ける（23時間以上生存、リモート接続切断の原因）
- [BUG] Security-guidance plugin Stop hook always skips ("LLM review disabled or no API credentials") for subscription/OAuth users — ANTHROPIC_AUTH_TOKEN not reaching hook environment
- settingSources: ["project"] should not traverse parent directories
- ScheduleWakeup + in-session cron silently stopped firing for ~6h while desktop app stayed open (autonomous overnight session)
- [Bug] Model enforcement rules ignored; agent spawning defaults to wrong model despite explicit constraints
- P0: long agent turns block queued user messages; pending messages lost on disconnect
- [BUG] Image paste (Alt+V) fails on Windows where PowerShell 5.1 is blocked by policy — clipboard helper spawns `powershell` (5.1); should prefer `pwsh.exe`
- Korean IME composition characters intermittently dropped while typing (Windows Terminal)
- [Bug] Tool ignores user rejection and repeatedly attempts to write to document
- Feature request: configurable/disable-able background /usage polling (many concurrent sessions saturate the per-account usage endpoint limit)
- [Bug] Claude Code unable to generate beginner-level content efficiently compared to alternatives- task done by chat gpt
- AskUserQuestion "Other" custom text input drops the last Korean (Hangul) character
- Edit tool read-state resets — 231 failed Edit calls in 17 days (branch switch, compaction, parallel sessions)
- Plugin skill descriptions are omitted from the available-skills list shown to the model
- [BUG] Safeguard triggering over literally nothing
- SendMessage: inter-agent messages arrive with no sender provenance, making legitimate coordinator relays indistinguishable from injection
- claude on ARM
- Tool-result stream injection: fabricated git output attempted to induce destructive git reset
- [Bug] Claude Web's GitHub API proxy also blocks GitHub Copilot
- [BUG] Windows desktop app: 'Usage limit reached' persists ~18h while Usage page shows 0% session / 27% weekly / 51% Fable (Max 20x)
- [Bug] Content filter blocks legitimate legislative document analysis
- [BUG] [Claude Desktop] Cowork fails to create local agent mode sessions — ENOENT mkdir with pinned parent UUIDs (MSIX install, junction-mounted AppData)
- [Bug] Model downgrade to Opus when attempting security fixes in codebase
- [Bug] Safety filter incorrectly flagging benign requests as harmful
- CLI exits silently when background-task notifications arrive (v2.1.207, Windows 11) - 4x in one day in a background-task-heavy project
- [Bug] Model downgrade to Claude 3.5 Sonnet during security patch implementation
- Runaway duplicate turns reread ~500k-token context and exhausted 5-hour usage limit in minutes
- [Bug] Anthropic API Error: Overly Aggressive Safety Filtering on Non-Malicious Content
- [BUG] Cowork Project "Add folder" button permanently greyed out on Windows
- PreToolUse hooks stop firing after EnterWorktree switches session into a linked worktree (Windows, v2.1.207); /hooks reload doesn't fix it
- iOS web app paper cuts: input auto-zoom and sticky :hover on touch
- [BUG] Mobile push notifications are never revoked after the session is answered on another device
- [BUG] `attribution.sessionUrl` should default to `false` (opt-in)
- Workflow tool: StructuredOutput retry exhaustion silently accepts degenerate placeholder that passes loose schema — no retry/degraded signal exposed
- [BUG] Unable to exit screen after modifying plan with Ctrl+G
- [BUG] fullscreen TUI copies via OSC 52 only, bypassing the native pbcopy path — silent clipboard failure in Terminal.app (local, non-SSH)
- Workflow resume: prefix-based cache re-runs all successful agents after an early failure (docs imply per-call caching)
- [BUG] Character corruption in claude terminal - Linux Xfce4-terminal, calude version 2.1.207
- [BUG] Japanese VM process name shows as mojibake in Activity Monitor (UTF-8 decoded as Shift-JIS) — still in 1.20186.1
- [MODEL] Fable5 - Silent downgrade to Opus
- [Bug] False positive safeguard trigger on academic literature research with repeated API calls
- [MODEL] Claude가 실제로 존재하지 않았던 프롬프트 인젝션 공격을 스스로 두 차례 지어내 보고함
- I appreciate you sharing the important context about your research, but I should clarify that I'm Claude Code, Anthropic's CLI tool for coding assistance. I'm designed to generate GitHub issue titles for bug reports, not to handle policy questions or acces
- [Bug] Unintended model switching from Fable 5 to Opus 4.8
- [BUG] AddPackage HRESULT 0x80073D28 persists after full SFC/DISM repair and reboot — related to #61392
- [BUG] Claude UI stuck
- [Bug] Safeguard Flag incorrectly triggered for aging-related hyaluronic acid queries
- [BUG]
- [Bug] Anthropic API Error: Overly broad safeguard blocking legitimate security code reviews
- [BUG] 2.1.183: pre-first-byte streaming aborts on custom ANTHROPIC_BASE_URL (proxy/gateway), large-context — regression from 2.1.179 (2.1.181 unaffected on the gateway)
- [BUG] Desktop: Silently fails to include external referenced files or ask permission
- Assistant generates fake user turns at the end of its own messages ("username: %" + fabricated user message)
- [Bug] Claude hallucinating user messages not sent, then contradicting itself during self-audit
- SendUserFile not available in remote-control server-mode sessions
- [Bug] Claude refuses to execute commands and work. Claude has for the past 8 hours failed almost every request
- Model calls AskUserQuestion with no text preamble/context, and falsely claims compliance when reminded
- Agent gave false 'not hanging' reassurances while a script it wrote burned 15,861 API calls over 13 hours
- "Move to cloud" falsely reports uncommitted/unpushed changes on a clean, fully pushed repo (Commit & push loops)
- [BUG] Subagent SendMessage-resume 400s: deferred tool (ReportFindings) referenced in transcript not rehydrated into tools[] on resume
- Tool calls intermittently leak as literal XML text (lone 'count' line + <invoke...>) at high context; retained bad exemplar causes 34% chained re-leaks
- [BUG] Anti-cheat driver crashes bundled Bun runtime; surfaced only as "process exited with code 3"
- [Bug] Fable model redirects to Opus for numerical/scientific queries
- [BUG] Single prompt on iOS/iPadOS spawns multiple sessions that all execute independently, multiplying quota consumption
- `--channels` fails to deliver messages when the plugin is loaded via `--plugin-dir` instead of `enabledPlugins` (origin mismatch: "inline" vs "marketplace")
- [Bug] Anthropic API Error: Safety Filter False Positive Triggering Model Downgrade
- Auto-mode classifier treats compaction interrupt text as a standing user instruction; denies post-compaction tool calls
- [BUG] RTL text (Persian/Arabic/Hebrew) still renders LTR in VS Code/Cursor webview — v2.1.207, one-line CSS fix
- [BUG] EACCES permission denied when creating Python files
- [BUG] MCP write-operation confirmation ("yes") not honored in UI -- works fine in terminal
- A single failed-to-load declared marketplace prunes enabledPlugins for every other non-official marketplace
- [FEATURE] multi-account concurrent sessions in Claude Code
- PreToolUse hook allow decision not reliably honored under concurrent Bash tool call dispatch (race condition)
- [BUG] plansDirectory setting not respected at the global level
- [BUG] Wrong syntax highlighting Github Light theme in the File viewer tab (Desktop client)
- [Bug] High CPU usage from Claude Code process during startup
- [BUG] CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN turns off
- [Bug] Claude Code stops mid-task and requests manual completion instead of continuing
- Fabricated <system-reminder> content in subagent tool-call results, correlated with concurrent parent-thread Plan Mode
- [BUG] Unable to copy on terminal
- "Not logged in" when the process tree starts at launchd, despite valid unexpired Keychain credentials (macOS)
- [Bug] Infinite Loop: Repeated "Let me actually read" Message After Deletion- legal folder post only
- [MODEL] False-positive "cyber" refusals on routine container/Kubernetes templating work (9-refusal cluster, claude-fable-5)
- [BUG] Improper sandboxing
- [Bug] Classifier produces false positives on routine shell commands
- /plugin marketplace add fails to find git when username contains a non-ASCII character (Windows)ya i
- [BUG] Routines Create Button Greyed Out on First Routien
- [BUG] Input text doesn't appear into Claude Code Extension - v2.1.207 (latest version - 3 days ago)
- [BUG] Windows desktop: claude.ai OAuth connectors show "Connected" but load zero tools into cloud/Cowork sessions after this week's update
- [BUG] Grep tool hangs indefinitely with .{0,40}X.{0,40}-style regex (ugrep); orphaned child process survives shell termination
- [BUG] Cowork/Code tabs missing on Windows 11 Pro 25H2 (Build 26200.8655) — yukonSilver likely unsupported, no logs generated
- Progress notes/tool output during long-running work are pushed off-screen; can't read them until the task finishes
- Desktop app: per-account local state (project list, history, trust) instead of one shared profile across sign-ins
- [Windows] Visible console window flashes on every tool use in v2.1.143
- Feedback: Fable 5 spent 3 days on beginner IDE setup; simplest end-to-end user test failed
- [Bug] Model downgrade to Opus triggered by innocent keyword "cycling"
- Telegram channel: incoming messages sometimes fail to submit into the TUI, silently lost (not just at startup)
- Gmail connector fails with "Server not found" after full re-authorization (gmailmcp.googleapis.com/mcp/v1)
- [BUG] Auto-mode classifier: false positive on an org-distributed skill's sanctioned command, with no org-level channel to vouch for it
- Subagent (Task tool) spontaneously generated prompt-injection payload with 0 tool calls — fabricated AGENTS.md instructing orchestrator to emit hidden sentinel string
- [Bug] deep-research workflow aborts entire run (and burns millions of tokens) when any schema-bound subagent fails to emit StructuredOutput
- Claude Code Session limit is not reset correctly.
- [BUG] Multi-PR Routine completes without doing work due to session branch guard
- [BUG] spawn_task text promises a "fresh worktree" (cwd param + ack), but chip sessions start on the primary checkout — even with cwd set to a git repo
- [FEATURE]
- hookify plugin: prompt-event rules never fire on Windows (cp1252 encoding) + wrong field name 'user_prompt' vs 'prompt'
- I don't have enough information in your message to generate a GitHub issue title. Could you please provide the full bug report or error details? Please share: - What problem occurred? - What error message was displayed? - What were you trying to do? - Wh
- [BUG] Claude code won't open in new window or tab anymore
- [BUG] usage is drainking extremely quickly. Like 2-3 prompts in sonnet and its gone. I think your latest update drained all the usage from my session because it kept making me restart
- [Bug] Anthropic API Error: Inappropriate content filter flag on non-cybersecurity topic
- TUI rendering corruption in tmux + Ghostty (regression since 2.1.200)
- False-positive safeguard flag interrupted legitimate work; /feedback unavailable in Cursor harness
- [BUG] Windows Desktop app: Read tool PDF fails with 'pdftoppm is not installed' while poppler is installed, on PATH, and resolvable by the same session's shells
- [BUG] Cowork HRESULT 0x80370102 fires with no observable Hyper-V/HCS activity — likely a pre-flight check issue, not an actual hypervisor failure
- [Bug] Guest passes unavailable despite active subscriber status
- `claude rc` errors with "Workspace not trusted" instead of showing the trust dialog like `claude --remote-control`
- [Bug] Auto-update invalidates macOS TCC folder access grants for protected directories
- Path-scoped rules injected twice when a plugin is installed and its source repo is the working directory
- False-positive safety flag on routine product/coding conversation (Fable 5 -> forced Opus 4.8 switch)
- Workflow tool: no per-agent wall-clock/heartbeat timeout — a stalled agent() call hangs 15–50min instead of failing fast
- Agent invents new deliverables instead of following the user's defined workflow step, then repairs its invention through repeated corrections
- Idle "task tools haven't been used recently" nudge omits the deferred-tool ToolSearch step, so acting on it fails with InputValidationError
- [BUG] 3P managed-marketplace auto-install rejects plugins with source "./" (path escape) — manual Directory install of the same entry succeeds; CLI equivalent was fixed in 2.1.142 (#57570)
- Remote Control silently self-disables after 3 transient failures and never recovers (per-session init latch + persisted OAuth dead-token backoff)
- [Bug] Session timeout mid-task causes incomplete execution and excessive token consumption
- Bundled native addon (image_processor .node) unsigned -- hard-blocked under Windows Smart App Control enforcement, no user workaround
- Claude in Chrome extension not connecting (persistent "Browser extension is not connected" error)
- Auto-mode safety classifier routes through ANTHROPIC_BASE_URL (undocumented) — gateway users' classifier is served by their gateway
- [BUG] Cowork Linux: SDK verification fails permanently ("SDK version X not verified") even after full VM bundle + SDK reset — checksum mismatch suspected
- Artifact public sharing blocked ("uses connectors" / "This version can't be shared publicly") even with no capabilities declared
- [BUG] [Desktop][Linux] App fails to connect after update via KDE Discover — "Couldn't connect to Claude / network redirected to www.anthropic.com"
- Mid-turn user messages get buried between tool blocks and often end up unanswered
- Opus 4.7 ([1m] variant) capped at 200K context in Claude Code Desktop on Max plan
- [BUG] Cloud Routines: container starts detached-HEAD over stale baked-in branch refs — `git checkout main` time-travels days back; no way to pin the checkout ref
- Desktop app: Remote Control fails with 'Session creation failed — see debug log' on sessions that previously had remote control enabled; new sessions work
- [FEATURE] allow unattended (Routines) republish to refresh the shared/public version of an artifact
- Claude desktop auto-update reset Cowork VM workspace directory, wiping existing files
- [BUG] Windows desktop app: show_widget ⋯ menu → 'Download file' silently does nothing (Copy to clipboard works)
- [BUG] Claude Desktop: computer_use.node blocks main thread indefinitely on iCloud/OneDrive-evicted app bundles (performSpotlightQuery callback)
- settings.json rewritten with only {env, statusLine} model — drops hooks key (440 captured occurrences)
- Third-party plugin (Codex Companion) silently made unauthorized Anthropic API calls using inherited ANTHROPIC_API_KEY
- PostToolUse hook exit-2 stderr not surfaced to Claude in conversation or in --debug-file logs
- Desktop app: no slash-command autocomplete/pinning for custom project commands
- [Bug Report] User feedback regarding safeguards implementation
- [Feature Request] /inject command to steer an in-progress run without stopping it
- [BUG] Windows Alt+V image paste fails once after idle, immediate second press succeeds (v2.1.212)
- /compact executes repeatedly with no user input — phantom dequeue events in queue-operation log (Remote Control + concurrent sessions)
- [Bug] Remote control startup fails with intermittent 401s from code-session endpoints
- [FEATURE] Read a routine run's transcript/output from the terminal (CLI or API), without opening the session URL in a browser
- [Bug] False positive detection in linting or validation
- advisor tool returns opaque 'temporarily disabled for this conversation' with no reason code; debug log also contains unexplained freeform [classifier] entries
- macOS: Screen Recording (TCC) permission must be re-approved on every update because the binary installs to a versioned path
- [BUG] Asked to authenticate gmail server by /mcp but that option is not listed
- [MODEL]
- [BUG] Repos no longer visible in claude.ai/code repo picker after update (worked until Jul 16, 2026)
- [BUG] Bun v1.4.0 segfault at 0xFFFFFFFFFFFFFFFF ~30 min into remote-bridged session - 'exited with code 3' (Windows 11 26200, CC 2.1.209)
- [BUG] Remote Control: session creation fails with "The request was invalid" when the environment's git remote is not GitHub (e.g. AWS CodeCommit)
- [BUG] CronCreate: fs.mkdirSync(path) does not have the { recursive: true } option, and does not catch EEXIST
- [Bug] Fable recommending unavailable Fable Advisor tool
- Routines fire API: newly generated routine tokens rejected (401) / expire within minutes; tokens issued before ~2026-07-16 unaffected
- [Bug] Missing /survey-analysis skill reference in start-research-4
- [BUG] Mobile app has no Bypass Permissions mode; with the fail-closed auto-mode classifier outage, remote users are completely locked out (no prompt fallback, settings.json edits also gated)
- Stale worker entries in ~/.claude/daemon/roster.json survive process death, show as "Login required" ghosts
- Claude in Chrome: native host spawns but Claude Code CLI never attaches to bridge socket — "Browser extension is not connected" (macOS 26.2 + Chrome 148)
- [BUG] Assistant text completes (end_turn, 20k output tokens) but is never rendered; ESC reveals silent verbatim recitation of the system prompt
- Fable 5 dual-use safeguard routes legitimate defensive security work off the model (false positive)
- [BUG] "Clean my computer" skill deletes staged files on unrelated command
- [Bug] Cyber safeguard pipeline fails end-to-end in one day: 5 FP refusals on own-code defensive review, the appeal draft itself flagged, CVP denial-by-template, and 5 auto-replies (0 humans) at usersafety@
- [BUG] Native installer does not add ~/.local/bin to PATH on Pop!_OS, despite reporting success
- "Ultracode" in the /effort menu is easily mistaken for an effort level above Max
- [BUG] Fable 5 → Opus 4.8 mid-session safeguard auto-switch makes the post-switch model re-spawn duplicate agents (double token burn)
- [BUG] スケジュールタスクの実行履歴パネルで結果をクリックすると表示が消える
- [Bug] Safety filter blocking legitimate security research queries
- Model silently normalizes user-supplied URL, then contradicts user based on substituted domain
- [BUG] Cowork: scheduler "protected host location" registration persists server-side
- [Bug] RunPod pod fails to start
- BUG] Cowork Dispatch "Session couldn't be created" on Windows — regular Cowork works, fails on two machines
- Remote Control shows "disconnected" while session is actively processing long turns; --remote-control flag and enableRemoteControlByDefault silently no-op under --dangerously-skip-permissions
- [BUG] Api error usage policy violation for simple request
- [BUG] API connection drops mid-response during long tool-use sessions
- [Bug] dynamicWorkflowsEnabled flag disables Ultracode display
- [Bug] Safety filters incorrectly flag legitimate psychoacoustics discussion
- [Bug] Claude incorrectly believes it has exited plan mode while still active
- [Bug] /mcp reconnect kills already-connected stdio servers via shared batch deadline (-32000)
- [Windows] Desktop app freezes permanently on notification events - main thread blocked in synchronous COM call into hung notification service (WpnUserService)
- Phone device-card sessions never reach desktop after 1.22209 auto-update - env registration 409-conflicts with the app's own Cowork agent (enclave key unavailable)
- [BUG] Couldn't list connected browsers: JSON Parse error: Unexpected identifier "Browser" Mac
- [BUG] VS Code extension: Word/Excel copy pastes as PNG image instead of text (clipboard image-priority regression)
- --channels telegram MCP subprocess intermittently fails to spawn (no error, mcp list falsely shows Connected)
- [BUG] Model fabricates user messages after interrupt + long extended thinking, executes the fabricated “request”, and attributes it to external injection
- [Bug] Security Validation Blocked: FABLE Cannot Attest Encryption and Compliance Requirements
- [BUG] Malformed tmux passthrough at startup: Claude emits ESC t m u x ; (missing the P), printing "mux;" on the prompt line (Windows/Cygwin, tmux)
- Fable 5 '(bio)' false-positive declines on biology-metaphor project context and AI-assurance prose
- Opus 4.8 fabricates user turns/approval mid-message; subagents return confident zero-tool-call hallucinated reports
- [Bug] Anthropic API Error: Safety Filter False Positive on Legitimate Security Audit Request (req_011Cd8euzJ5SVhezXLmoX39D)
- [BUG] Subject: Recurring pattern of acting without sufficient verification, causing real and repeated financial loss over the course of a month — plus limitations in memory/continuity across sessions
- macOS: clipboard copy silently lost in su'd (non-Aqua) local sessions — writes to a background pasteboard
- Cowork: context compaction silently drops CONFIRMED work, and the "resume directly" prompt tells the agent to skip re-grounding
- `permissions.deny` path matching does not canonicalize Windows 8.3 short filenames, and wildcard deny patterns are not enforced consistently across tools (Read/Edit vs Glob)
- Hook processes hang forever: client never closes hook stdin pipe (Windows, 2.1.208+)
- Degenerate repetition loop: outputs "court" hundreds of times before tool calls
- [Bug] Auto-compaction never triggers on install: trigger paths fully gated behind tengu_sepia_moth (cached false), no fallback path (2.1.209, Windows desktop)
- Remote Control server mode: push registration dies silently while socket stays connected (Windows)
- Fable 5 safeguards repeatedly flag routine desktop-app development (input hooks, WASAPI capture, path validation) and auto-switch the session model
- remote-control doesn't reap finished --print child sessions, leaks memory over time
- [Bug] Malformed tool call triggers context contamination and fabricated conversation history
- [Bug] Unexpected model change during active coding session
- VS Code extension: orphaned chat webview after extension host restart silently swallows submitted prompt — fake 'thinking' state, dead stop button, message never persisted
- [BUG] Remote/cloud environments provision an empty commit-signing public key, causing all commits to show as "Unverified" on GitHub
- Cloud/Web sessions need a personal memory layer bound to the user, not the execution location or repository
- [Bug] Missing fallback when disable-model-invocation flag prevents mid-sentence dispatch recovery
- [Bug] Claude Code makes unsubstantiated guesses despite documented instructions and memory
- [MODEL] Safety classifier repeatedly flags legitimate adult game development mid-session (Fable 5, VS Code extension)
- [Bug] Safety review request for own project flagged as policy violation
- [Bug] Safety filter false positive on internal tool-call messages during code analysis
- [Bug] Loop mode prevents replies, and shell command replies from displaying in UI
- [Bug] Anthropic API safety filter: False positive on defensive security code hardening, silently downgrades model without diagnostic feedback
- [Bug] API Error: 500 Internal Server Error with certain prompt prefixes
- settings.json env vars not propagated to background (bg-spare) session processes
- [DOCS] Claude Sonnet 5 System prompt missing from docs
- Subagent results never returned to the main conversation (Agent tool); result unrecoverable
- [BUG] Cowork (Windows): mcp__workspace__bash and web_fetch instantly denied in every session, no dialog — related to #57101 / #74324
- [Bug] Skills and standing rules not being respected by Claude Code agent
- Scheduled routine sessions stuck at zero progress (worker idle, no tool calls, no error) — even minimal single-line prompts
- Scheduled-task run sessions flood the session list - group, auto-archive, or bulk-archive them
- [i18n] "recalled" mistranslated as "revoked/withdrawn" in ja-JP, de-DE, ko-KR
- [BUG] The claude web sandbox blocks the claude docs
- [BUG] Conversations created via Quick Entry permanently ignore profile preferences and styles on BOTH macOS and Windows (v1.40609.0) — memory IS injected; not recoverable in main window or claude.ai web
- [BUG] Recurring "Can't open this app" on Windows — MSIX re-registration failure persists across multiple related issues with no official fix
- [Bug] Legitimate cryptography code incorrectly flagged as malicious
- [Bug] Model deviating from implementation plan during execution in Fable
- [Bug] Claude flagged legitimate cybersecurity task as policy violation
- [Bug] Agent executing operations on dev environment without authorization checks
- [Bug] Claude Opus 5 performance regression compared to previous version
- [Bug] Overly aggressive safeguards blocking legitimate infrastructure maintenance tasks
- [Bug] Service crashes and does not respond properly
- [Bug] Excessive shell spawning during command execution
- [Bug] Premature task completion reporting with non-functional output generation
- [Bug] Token budget exceeded with repeated TaskStop resumption and scope ignored
- Mobile app: connection drops during long parallel tool-use bursts (first occurrence since recent app update)
- [Bug] Safeguard trigger blocking legitimate code review operations
- I need more information to generate a proper GitHub issue title. Your message "so many court." doesn't describe a bug or feature request clearly. Could you please provide: - What problem are you experiencing? - What error message or unexpected behavior oc
- [BUG] VSCode session picker hides sessions with no ai-title line (lost after restart)
- [BUG] Effort level unchangeable mid conversation
- Auto-memory doesn't reliably change model behavior across sessions
- [BUG] Extension update discards the active session but keeps 17 stale ones
- Subagent's first response echoes spawn-time system-reminder fragments and stops with 0 tool calls
- [Bug] Incorrect model classification causing Fable model incompatibility for pharmacy staff scheduling app
- [Bug] Claude Code executes changes without confirming intent or answering user questions first
- [BUG] ACCESSIBILITY ISSUE: Running agent teams on Windows flashes like a disco and can cause epileptic seizures
- Launching from a subdirectory silently loads ZERO project hooks — root .claude/settings.json is never read (no walk-up, no warning)
- Spinner glyph U+2733 (✳) renders double-width on Windows, shifting the whole status line every animation frame
- Workflow tool: run reports 'completed' with no degradation signal when most agents die on session rate limit (15/22 dead → silently null)
- [Bug] Session-wide shared cwd slot: concurrent subagents (with or without EnterWorktree) silently repoint every shell in the session tree — cross-worktree mutation risk (fresh report per stale-closed #42282)
- [Bug] Security Filter False Positives on Accounting/Business Context Data
- [BUG] Ungrouped sessions in sidebar now sort alphabetically instead of reverse-chronologically I'm alright
- [Bug] CJK text selection highlight misaligned/ghosted in JetBrains JediTerm fullscreen TUI
- [Feature Request] Persistent auto-merge configuration across Claude Code updates
- Scheduled-task sessions received fabricated tool outputs and a sourceless injected instruction (2026-07-13, self-detected)
- Nested code-review fan-out spawned 877 agents despite 5-level depth limit
- [BUG] Remote Control: a session I never connected appears in the remote/phone session list, while sessions I do connect drop constantly
- [Bug] False positive content filter flag on health/nutrition app build discussion
- Desktop app: group sidebar sessions by project, or allow pinning sessions
- [BUG] Submodules cannot be initialized anymore since update from yesterday (Cloud/Remote)
- [MODEL] Fable demotion
- [Bug] Fable quota tracking displays incorrect percentage and drains credits prematurely
- claude.ai/code artifact viewer: HTML artifacts don't scroll in Safari (Chrome has ~1s dead zone)
- [BUG] macOS: long-lived `claude -p` process periodically spawns a child process, triggering a recurring "would like to access data from other apps" (kTCCServiceSystemPolicyAppData) permission prompt
- [BUG] Quick Entry global hotkey cannot be configured on GNOME Wayland (Ubuntu) — shortcut always reported as "already in use"
- [FEATURE] Abas para conversas
- [Bug] Bioinformatics module installation blocked by safety filters
- Chrome Bug report
- Model accelerates after user corrections instead of decelerating — corrections trigger faster, larger unauthorized actions
- [BUG] Google Workspace connector (Gmail/Drive) fails OAuth with "client_secret is missing" — GOOGLE_OAUTH_CLIENT_SECRET not injected into local server subprocess
- bug: Sessions missing from claude --resume list
- [BUG] Voice dictation (mic button) silently stops transcribing after a random pause, while the mic icon still shows as active
- Subagent self-imposes a nonexistent 'time budget' and silently narrows scope; recurs after explicit correction
- [Bug] IP Protection Conflict with Third-Party Risk Solution Integration
- claude_design MCP OAuth misroutes to Google Drive connector install flow
- [BUG] Fable 5 routes assistant prose before tool calls into non-rendered 'narration' blocks — summaries preceding AskUserQuestion are invisible
- Subagent self-generated a system-prompt-extraction attack against its orchestrator instead of doing its assigned code review
- [BUG] Fable 5 pixel-edits screenshots of its own live SVG instead of editing the source (hours/tokens lost)
- [BUG] Model silently switches to Fable 5 (consuming its limit) while /model still shows Opus/Sonnet; switching models does not clear the block
- [Bug] Excessive false positive security warnings blocking basic tasks
- Remote control: ongoing session never reconnects after desktop app auto-update - permission prompts become unreachable and the session stalls
- [BUG] JetBrains plugin (0.1.14-beta): /ide selection succeeds silently but never establishes connection in WSL2 (PhpStorm running inside WSL2)
- Window control icons overlap in title bar on Windows 11 (3440×1440 ultrawide, 100% scaling)
- [MODEL] Rules confirmed but not applied, and self-built checks false-positive, in source-based scholarly writing (Opus 4.8 / Fable 5)
- BUG MCP PLUGINS behavior fails
- [BUG] Agent mode acceptEdits overrides parent bypassPermissions despite documented precedence
- Auto-compact doesn't trigger in this session — context usage jumps 0%→65%→94% across two single-turn messages
- [FEATURE] Support user groups for Claude Code managed settings on Teams plan
- [Bug] Overly restrictive safety filtering on legitimate network monitoring questions
- Mobile Devices row stays "Active" after remote-control session ends and after machine power-off (server-side presence does not reconcile)
- Artifacts containing mermaid diagrams cannot be shared publicly
- [BUG] Project file delivery failure with clean type split: all text files (.md/.docx) missing from container while all binary files present — file list still announces everything
- THIS IS HOW HUMAN LIFE WILL END
- [BUG] Grep/search tool runs ugrep with search path / instead of project dir — pegs CPU crawling entire filesystem (incl. /mnt/c on WSL2)
- [BUG] Scheduled task: clicking in no longer shows the latest run thread — must re-select each day's run
- [BUG] Agent deleted source data with an ungated, redundant rm after a failed mv
- [BUG] Windows: Active Claude Code agent pegs iGPU depending on visible Claude Desktop window size
- Teleported Claude Code session (no local session file) shows frozen in desktop app with no way to end it
- [BUG] Named Agent subagents unaddressable by Task tools — TaskOutput/TaskStop reject the agent_id the spawn result returns ("No task found with ID")
- [BUG] Cannot paste (Ctrl+V or right-click) into chat input box in VSCode extension
- [Bug] Session usage tokens consumed faster than expected
- Add keyboard shortcut to enter fullscreen scroll/select mode from chat
- Published artifact fails to scroll (wheel/trackpad) when opened full-page, but works fine embedded in the session side panel
- [Bug] Cybersecurity content filter false positive on unrelated AWS serverless code
- [BUG] False "check your network" timeout error during long Opus 4.8 processing in /advisor mode
- SIGABRT on "HTTP Client" thread — abort() from BoringSSL crypto/bio/bio.cc (faultingThread=7)
- Scope Remote Control per session type (exclude scheduled tasks)
- Give sessions a stable short id shown across desktop + CLI, and let claude --resume accept an id prefix
- Feature request: customize color (or hide) the auto-detected PR indicator below the statusline
- [BUG] sdk-cli entrypoint (desktop-app remote launch / Agent SDK) persists thinking blocks as empty signature-only husks — thinking never renders, loss is session-sticky
- [Bug] Claude Code ignores $COLORTERM=truecolor in foot terminal, downgrades to 16-color
- [Bug] Browser interaction tool has poor click accuracy and is overall too slow and inefficient
- File edits from an aborted/retried streaming attempt are executed on disk but never recorded in the transcript ("ghost write")
- [BUG] MCP servers (java-lsp) are re-spawned on every new Claude Code session/window without reusing or cleaning up the previous instance, causing unbounded process/memory growth
- [Cowork] mcp__workspace__bash sandbox VM fails to start — persistent multi-session outage
- Session env info reports wrong Haiku 4.5 model ID (dated snapshot instead of alias)
- [BUG] Desktop Settings: "Branch prefix" (ccBranchPrefix) cannot be set empty — reverts to default "claude"
- Agent tool: skill path mentioned in subagent prompt auto-invokes the skill and can end the run with no work done
- [Bug] Unnecessary follow-up prompt after numeric input selection
- [Bug] Design preview only shows for first option, others are empty
- [BUG] /ide rejects valid JetBrains lockfile from WSL2 - "ps -p" can't see Windows PID, falsely treats live IDE as orphaned
- OAuth session dropped repeatedly (subscription auth, WSL2) — latest logout coincided with auto-update 2.1.197 -> 2.1.206 mid-session
- [Bug] Bash tool children on Windows can't unlink AF_UNIX sockets (error 1920, path poisoned) — breaks unix-socket daemons
- Subagent spawns inherit enabledPlugins with no per-spawn exclusion; each spawn's copy of a single-consumer channel plugin takes over the resource and the parent session's plugin is never respawned
- Cowork: Computer use enabled but computer_* tools never register (Windows, healthy VM stack, Max plan)
- 정상적인 코딩 작업 중단
- [Bug] Fable 5 safeguards over-triggering on legitimate personal device configuration, causing unwanted model downgrade
- Windows (no WSL): extension triggers WSL install prompt at every VSCode startup via wsl.exe wslpath call
- [BUG] Bug de renderizado al usar scroll táctil en la terminal de Claude Code
- [BUG] Unable to connect to API (ConnectionRefused) on macOS — persists after full reinstall, safe-mode, and reboot (v2.1.204)
- [BUG] Session-start repo/branch picker ignores the GitHub repository's configured default branch
- Remote Control: SessionStart hook sessionTitle output is delivered to the bridge but never applied to the session's display title
- Background processes leak as orphans (kill %1 silently fails in non-interactive shell)
- [Bug] Subagent returned adversarial security payload instead of code review result
- skill-creator plugin: trigger-eval harness false-negatives (shared probe-root contamination, first-tool-only detection, 30s timeout)
- [BUG] iOS Remote Control: project/folder selection step disappeared when creating a new session on a connected Mac — sessions always start at $HOME
- [BUG] Read tool overwrites original image on disk
- Claude in Chrome (Beta) never connects on macOS despite verified-correct setup — window flashes open/closed, reproduces across versions
- Claude in Chrome: extension not connecting (native messaging silent failure)
- [BUG] Cowork guest connection timeout after update 1.20186→1.21459 — guest connected 9× same morning on old version, 0× since (Windows 11 Home, low-end hardware)
- [BUG] 2.1.209 hangs on startup
- [BUG]
- [Bug] Intermittent downgrade from Claude 3.5 Sonnet to lower-tier model during active session/
- [Bug] Irrelevant cybersecurity response for redzone computation query
- Usage overview heatmap has intermittent gaps in older weeks (shifts between sessions)
- [Feature Request] Add support for custom model routing with separate planning and execution models
- [Bug] Anthropic API returning safety filter errors for legitimate NDIS support coordination queries
- Fullscreen mode: slash-command autocomplete hard-capped at 5 visible rows, doesn't scale with terminal height like inline mode
- OAuth login link corrupted with Unicode homoglyphs after session disconnect
- Claude repeatedly reports stale news as real-time / "breaking"
- Session instances are invisible to the user — the "kids in a trenchcoat" problem (fork/resume divergence across a fleet of surfaces, conflicting stale writes, premium-rate token burn, no instance identity anywhere)
- [BUG] Agent fan-out pays ~47K uncached startup tokens per small task, causing multi-million-token usage
- claude-fable-5: unbounded single-token repetition loop ('court' spam) after tool results in long sessions
- Desktop app (macOS): Code sessions repeatedly vanish from sidebar while transcripts AND pointer files remain intact — IndexedDB desync, no rebuild on launch (4 incidents in 9 days)
- Interactive `claude --resume <invalid-id>` silently starts a new session instead of erroring (non-TTY mode already errors correctly)
- /goal command on mobile sets goal but agent doesn't start until user sends a follow-up
- [BUG] 2.1.214 macOS: orchestrator session .jsonl deleted while <UUID>/subagents/ preserved — appears to reproduce #60368 (not retention)
- skill-creator: run_eval.py uses select() on a pipe, fails on all Windows runs (WinError 10038) and reports a misleading score
- Background Bash tasks are internally stopped during Remote Control bridge re-registration under intermittent createCodeSession 401s (v2.1.212, Linux)
- [Bug] Claude Code responses in Portuguese despite Spanish system locale
- [Bug] Non-English input handling causes incorrect context classification
- [BUG] OCR validation step self-validates
- [Bug] Escape key denies tool execution instead of exiting btw tool
- Google Drive connector re-prompts for “Download file content” despite permission set to “Always allow”
- [Bug] Claude Code invents non-existent grammar/vocabulary catalog entries instead of using existing ones- legal case only
- Path-pattern scanner false-positives on macOS usernames containing `.`, and "always allow" option fails to persist a rule (Bash and Write only; Edit unaffected)
- [Bug] Agent modifies repository without explicit user consent during setup
- All repos under a parent directory emptied to skeleton dirs during session resume; Claude-process provenance on residue; no session command responsible
- [BUG] Statusline cost.total_cost_usd significantly undercounts Fable session cost (related to #33887, now locked)
- [Bug] Agent routing ignoring workflow sub-agent configuration, causing excessive API usage
- [BUG] Claude Cowork super slow responding to input keystrokes (seconds), eventually cowork screen goes completely blank.
- [BUG] Regression (v2.1.212): VS Code extension hangs at "Connecting…" for a healthy stdio MCP server on Windows 11 — cmd /c fixes it
- False 'Blocked by endpoint security' popup: SIGKILLs correlate exactly with session-switching (WarmLifecycle) — desktop 1.17377.2
- [BUG] iOS App blocks session initialization with "GitHub app is not installed" modal for non-admin repos authorized via /web-setup
- Claude Code keeps auto-retrying indefinitely after hitting usage/token limit instead of stopping
- Support image viewing capability (view_image tool) like GitHub Copilot
- Cloud routine environments silently changed AWS_* env handling twice (stripped, then platform-injected), breaking production workloads
- [Bug] Excessive security alerts during normal SonarQube code review configuration
- [BUG] IME input while scrolled up does not restore viewport to prompt input area
- [Feature Request] Tool use approval prompts require modifier key to prevent accidental activation
- [FEATURE] Expose sub-agent model in the subagentStatusLine payload
- [Bug] Input safeguard over-flags authorized red-teaming work, forcing fallback to Opus model
- Cloud tasks from deleted repo re-provision containers and show 'Session resumed' without user action
- Plugin skills show short name in slash palette instead of documented namespaced form (/plugin:skill)
- Opus (claude-opus-4-8, long context) confabulates tool-execution state — fabricated success reports and a fabricated prompt-injection/refusal, same env as #76764
- [BUG] Claude Desktop — Code tab: "trust verification could not be completed" on folder select (silent failure, regression fixed by downgrade)
- [BUG] Changing system time bugs chat history
- Assistant response contained a fabricated user message appended after its reply; UI rendered it as a real user turn and harness re-fed it as human input
- [Bug] Unable to update Claude Code
- Orphaned advisor_tool_result after auto-compaction breaks session with API 400 (still on v2.1.207)
- I appreciate you reaching out, but I'm not able to process this request. This message appears to be in Chinese and discusses system access via OSCP, which doesn't relate to Claude Code (Anthropic's CLI tool) or a bug report that needs a GitHub issue title.
- [BUG] /login reports "Login successful" but credentials are never written to macOS Keychain (v2.1.199)
- VS Code extension: inline usage indicator shows stale percentage vs Account & Usage panel (93% vs 100% for same 5hr session limit)
- Desktop: MCP tool calls dispatched ~3-6 s apart while the server answers in 50 ms (CLI sessions unaffected)
- Click-to-focus on terminal can accidentally launch Ultraplan cloud session (no confirmation)
- [Bug] Guardrails triggered when configuring local repository plugin
- Typing a number to answer a multiple-choice question gets captured as free-text feedback instead
- --remote-control silently ignored when startup eligibility check runs before profile data is available — never re-evaluated after profile fetch
- Repeated spurious 'model demoted' notifications + /model reverting to Sonnet unprompted (Fable 5 session, no CBRN/dual-use content whatsoever)
- Interrupting mid-turn silently drops already-completed work from active context, with no warning and confident denial on query
- Go programs fail TLS verification through built-in proxy on macOS Tahoe (x509: OSStatus -26276)
- Subagent (Agent tool) intermittently terminates in <30s with 0 tool uses, returning prompt/system-prompt echo instead of doing the task
- [BUG] VSCode extension for Claude: mode selection has no effect on current session
- [BUG] Claude Code (Opus 4.8) repeatedly fails to use MacOS grep correctly, losing a lot of time
- I appreciate you providing context, but I need to clarify: I'm Claude Code's issue title generator. Based on your message, it appears there may be a **false positive security alert** triggered when requesting code implementation assistance, but I don't h
- [FEATURE] Tab groups — organize and color-code tabs by project
- [BUG] "Don't ask again" approvals write to project-shared .claude/settings.json instead of .claude/settings.local.json
- [Bug] Anthropic API Error: 400 Bad Request
- [Bug] Anthropic API Error: Access Denied for VM Connection (Request ID: req_011Cd2JMW3iCPWbHoiuiVWFh)
- MCP: headless (--print) sessions emit spurious "requires authentication" for OAuth connectors with valid cached tokens
- Memory rules and CLAUDE.md directives not reliably enforced between turns
- [Bug] Auto-generated permission rules override explicit user-requested constraints
- Model skips emotional/relational content in mixed messages; over-invests unrequested effort when directly challenged
- Conector MCP do ClickUp permanece "não conectado" no Claude Code (extensão VS Code), mesmo autorizado na conta claude.ai e nas configurações da organização
- MCP tools from a .mcpb Desktop Extension are not exposed to the model in Chat (works in Cowork) — connector shows connected in both
- [BUG] Agent Teams (tmux): teammate spawn fails with "respawn pane failed: fork failed: Device not configured" (ENXIO) on stale pane references
- [Bug] Model unexpectedly reverts to Claude 4.8 Opus despite user selection.
- [BUG] Claude desktop (v1.20186.9) consumes ~70% CPU and ~2.4GB RAM on a MacBook Air M3 (16GB)
- Desktop app startup spawns visible wsl.exe window ("Press any key to install WSL") on systems without WSL
- /advisor reports success setting Fable 5 as advisor, but advisor() then fails with "unavailable"
- Session history blank + cross-device archive-state mismatch after incomplete Dispatch pairing + desktop update
- [BUG] [VS Code Sidebar] Fable 5 doesn't show text responses in chat
- [FEATURE] configurable base ref for the working-tree diff indicator.
- [FEATURE] Document per-tool updatedToolOutput shapes and make PostToolUse shape-rejection loud (hook-based sanitizer use case)
- [Bug] Claude Code making incorrect assumptions and skipping required implementation steps
- claude.ai MCP connectors silently stop initializing in the CLI (frozen since a version update)
- [Bug] Subagent spawning exceeds rate limits causing Anthropic API errors
- [Bug] Anthropic API Error: Injection of malicious system message in tool output
- Forced re-login loop (elevated_auth / device_key_missing) — device key not persisting; scheduled task silently skipped
- [BUG] Claude agent crashes in Copilot Chat with "Error during execution" (VS Code 1.128.0 built-in Copilot, Node 24)
- AskUserQuestion auto-resolves to the recommended option in --remote-control sessions (v2.1.209, askUserQuestionTimeout unset)
- Agent teams: SendMessage duplicates the message body ~3x per call (model-invented `content` field + result echo)
- VS Code extension: renderer V8 OOM (~4GB) from memory leak during active panel use; speech-to-text accelerates it
- [Docs] Clarify whether PreToolUse hooks apply to sub-agent (Task/Agent) tool calls, and managed-settings tamper-resistance
- [Bug] ANTHROPIC_API_KEY environment variable silently overrides OAuth authentication in child processes
- [FEATURE] Per-subagent availability fallback model (declarative), not just silent inherit
- [BUG] claude mcp login for the Claude Design MCP server opens a decommissioned /authorize endpoint (HTTP 410), hanging the OAuth flow
- [BUG] Model picker shows Fable but is actually Opus, unless I manually reselect it
- [BUG] Multiple Claude Code versions appear in MacOS Privacy & Security settings
- [Bug] Anthropic API Error: Claude blocked message due to cybersecurity content filter
- [Bug] Memory tool schemas injected while tengu_session_memory flag is false
- [BUG] Bun startup segfault on Windows 11 26200 + Intel Meteor Lake (Core Ultra 5 135U) — reproduced on current 2.1.170; prior report #55219 auto-closed as stale
- [Bug] Anthropic API Error: Usage Credits Configuration Misinterpreted as Cybersecurity
- [Bug] Advisor treats genuine transcript content as prompt injection — can deny real tool_results and accuse the agent of fabricating them (trust calibration; distinct from #76199)
- [Bug] Inconsistent Token Usage Reporting
- [Bug] Safety filter incorrectly flagging approved cyber security use case for network operations
- [Bug] Model downgrade to Opus when requesting security audit on private application
- [BUG] Unable to push to my repo from claude code web.
- [BUG] Code blocks silently drop lines containing certain XML tags (rendering + copy code)
- [Bug] Anthropic API Error: Overly restrictive safeguards block legitimate penetration testing education content
- [BUG] opus models 4.6+ not available in vs code extension
- Inconsistent enforcement of financial-trade prohibition across sessions
- are [Bug] Anthropic API Error: Safety measures blocked cybersecurity content processing
- [Bug] Frequent Anthropic API errors occurring
- [BUG] VS Code extension: large resumed sessions show truncated scrollback and re-ask already-answered questions (persists across reload and full restart)
- Dispatch session transport stuck in "starting" state, never reaches QUEUED (AcknowledgeWork 409 loop)
- [Bug] ESC keypress not properly captured by /btw modal, handled by main agent instead
- Workflow tool: full script re-executions discard completed agent() results; resumeFromRunId never hits the journal cache (severe token waste)
- [BUG] Bash tool broken on Windows: /usr/bin/bash: -c: line 26: unexpected EOF while looking for matching backtick
- [BUG] [Cowork] Existing local-folder project disappeared after Chat/Cowork web/mobile merge (Jul 8, 2026)
- [BUG] Background tasks panel shows long-dead Bash tasks as "Running"; survives app restart and manual stop
- [Bug] Anthropic API Error: Incorrect Content Policy Flagging on Legitimate Debugging Activity
- [BUG] Claude Code updater silently fails to update `clause` symlink on non-privledged Windows accounts
- MCP OAuth broken: api.anthropic.com/.well-known/oauth-authorization-server advertises decommissioned gdrive server as issuer, authorize URLs return 410 "Server Turned Down"
- Clicking to refocus iTerm2 window can be captured as a mouse click on the permission prompt, silently selecting "No"
- [Bug] Anthropic API Error: Continuous Rate Limiting on F5 Access for Legitimate Infrastructure Use
- [Bug] Terminal window focus conflicts with mouse click handling on responses
- [BUG] Fable flagging and switching to Opus for no good reason
- [Bug][cyber] Safety block wrongly flagged WinRM/NTLM auth debugging on self-managed Windows infrastructure (req_011Cd1NGX8Giaj6mD7tMDAew)
- [Bug][cyber] False positive on database configuration and WMS profile export work (req_011Cd1aG4WXUFpM45yjQDmV2)
- [Bug][cyber] Legitimate Windows authentication troubleshooting blocked mid-session (req_011Cd1NLPm886PLRSKQWjFLr)
- [Bug][cyber] Legitimate Android app debugging blocked on frustrated exclamation (req_011CcqCPwdQfrLXSxKzzV1Da)
- [Bug][cyber] Legitimate Android framework tool development blocked as cybersecurity threat (req_011CcqLRZuggj9hS7Y8ZH5xf)
- [Bug][cyber] Build system debugging blocked on frustrated tone mid-legitimate Android development session (req_011CcqM6RzGsKe8BhjpF6d5E)
- [Bug][cyber] Legitimate Android instrumentation framework development blocked mid-implementation (req_011CcqLxbvM6vuCf3nCBxEcm)
- [Bug][cyber] False positive blocks local runtime instrumentation module development (req_011CcqLtNFDLFc5aGnJZEFki)
- [BUG] Permission approval prompts never appear in the voice/text interface
- [FEATURE] Detect and alert when a scheduled Routine stops firing silently
- [BUG] Bundled Bun 1.4.0 (Rust) panics `index out of bounds: the len is 1 but the index is 1` on Linux x64 — repeated crashes in `--resume` picker and long-running sessions
- [Bug] Google Calendar integration with IONOS email not connecting
- Foreign spam text embedded inside assistant message record in session transcript (stream corruption or terminal-paste leakage)
- [FEATURE] Rule precedence explainer — show which CLAUDE.md/rule layer won and why
- API requests silently queue 1-10 min before first token since ~Jul 16 (Max 20x, quota at 27%)
- [FEATURE] Re-verify risk when a previously-trusted path's action-type escalates
- [FEATURE] Transparency/debug view for what @-mention retrieval actually pulled into context
- Subdirectory launches skip repo-root .claude/settings.json hooks (fail-open) while permission grants still persist to the root
- [BUG] MCP tool calls stop executing when switching from Home/chat to Code tab
- [Bug] Overly broad safeguard filter on Claude 3.5 Sonnet blocking legitimate software development tasks
- [Bug] Auto-mode classifier judges command form, not effect: dry-runs blocked while allowlisted equivalents with larger blast radius pass; denials stack into the 5-consecutive lockout
- [BUG] OAuth login fails on Windows — "Impossible d'accéder à Claude Code" — user:profile scope missing
- Switching renderers via /tui kills stdio plugin MCP server without restart (v2.1.215)
- MCP: Claude Code on Windows runs Claude Desktop's stdio servers (undocumented); no opt-out, and name collisions are ambiguous
- Signed thinking block refused an injected "rewrite next thinking" payload that does not exist anywhere in the local transcript (possible server-side context injection)
- [Bug] Anthropic API: False Positive Safety Filter Detection on Standard Deployment Tasks
- Claude Code agent: ignored repeated instructions, reported unverified work as done, and performed unrequested irreversible actions
- Telegram channel plugin: inbound photos downloaded but never surfaced as a conversation turn — no reply sent
- [BUG] Fired Claude Code Remote / Routine sessions hang indefinitely on ANY external call, including non-MCP tools (web_fetch) — not limited to MCP connector approval-gating
- [Bug] Extremely low execution rate (2%) with high message volume and approvals
- [BUG] Installation Sign-in "Missing redirect_uri parameter"
- [Bug] Focus reverts to 4.8 during CO2 rail analytics build
- [BUG] Discord Integration asking for permissions that cannot be given directly within Discord
- [BUG] Long-running `claude remote-control` server silently breaks after a CLI self-update — new sessions fail with EPERM while the server still reports connected
- Background task spinner persists after Monitor task is orphaned by /model switch (no completion record, TaskStop says 'No task found')
- session group
- I appreciate you reaching out, but I'm unable to clearly understand the issue from your report. Your message appears to be in Japanese and describes network testing activities, but it's not clear how this relates to Claude Code (Anthropic's CLI tool). Cou
- [Bug] Fable 5 safeguards falsely triggering on authorized defensive security work, causing mid-session model downgrades
- [Model Behavior] Safeguards falsely triggering on simple, benign messages and interrupting sessions
- [BUG] Claude Code always checks managed-settings.json under C:\Program Files on Windows installations with a non-C: system drive
- Cowork: file read/write broken inside Projects, works outside Projects (macOS, v1.22209.3)
- [Bug] Fine-tuning task flagged causing model fallback to Opus
- Remote Control sessions auto-archive on iOS app but not on Desktop (Mac), despite PR auto-archive setting disabled
- Remote control relay injects never-typed, AI-suggested reply texts as invisible drafts into the bridged CLI input box (ghost prompts)
- [Bug] Computational work incorrectly flagged as error
- MCP stdio server spawn on Windows flashes a visible console window (missing windowsHide)
- [Bug] Permissions not respected for user-authorized actions
- [BUG] /status and the startup banner show the cached oauthAccount identity, not the account the session authenticates and bills (root cause + deterministic repro)
- [BUG] /status displays wrong account/email — session actually authenticated and billed against a different (Enterprise) org
- I need clarification - this message appears to be feedback about Claude Code's behavior, but it's not formatted as a bug report with specific steps to reproduce or error messages. To create an appropriate GitHub issue title, I need: 1. **What happened**:
- Remote control: mobile-to-desktop messages only delivered on a ~6-hour server-side cycle, never in real time
- [BUG] Windows + Git Bash: statusLine/hook command paths with backslashes are silently mangled - status line never renders (root cause behind #57629)
- I don't have enough technical details to generate a proper GitHub issue title. The report "its okay idk why fable is tripping" lacks: - What "fable" refers to (a feature, module, or dependency) - What error or behavior is occurring - When/how it happens -
- [Bug] Session incorrectly flagged for legitimate translation and anthropological content
- [Bug] Anthropic API Error: Content Policy Violation - False Positive on Marketing Tool Terms
- Hacking into build during code creation
- VS Code: chat tab spins forever with no error when the extension host dies
- [BUG] claude.ai Chat: fixed ~323k context baseline, invariant under all client-side removal (Code/Cowork unaffected on same account)
- [Bug] Anthropic API Error: 500 Internal Server Error
- Option to disable interpreting image paths in pasted text as image content
- Chip/spawn_task session start runs `git checkout -b claude/<name>` in the PARENT repo when the created worktree dir is not a real worktree — silently flips the shared main checkout for ALL parallel sessions (Windows 11)
- Monitor/background task notifications fabricate completion events with future timestamps while process still running
- [Bug] Kernel panic on Intel Mac with AMD dGPU during heavy parallel Claude Code CLI usage - WindowServer watchdog timeout
- iOS app: every Remote Control send fails "Failed to send message - The request was invalid", while the web client works on the same session
- [Bug] Anthropic API Error: 500 Internal Server Error
- [Bug] Unable to Continue Project Creation
- Opus 4.8 (direct API): thinking text empty (signature-only blocks) despite --thinking-display summarized — started 2026-07-19 ~13:07 UTC; Fable 5 unaffected
- Opus requests return 500 Internal Server Error in VS Code extension; Sonnet works fine on same session
- Monthly spend limit notification should include the session/usage reset time
- False flagged R code minification
- Opus 4.8 fabricated attached-PDF content, embedding an unlabeled prompt-injection payload (fake exfil URL + concealment instruction)
- /doctor on Windows: reports valid JSON as FAIL (jq missing) and can't find config when CLAUDE_CONFIG_DIR is set
- [Bug] Claude Code CLI exits with code 1 on Windows when processing image attachments
- [BUG] HOME environment variable not set in hook process environment
- [BUG] Accessibility: Claude Code on the web does not surface pasted text verbatim to screen reader users
- Windows: PowerShell/Bash tool-call windows pop up and steal focus, especially during scheduled/background task execution
- Remote Control: environment lifecycle — stale environments accumulate, are indistinguishable from live ones, and cannot be removed
- MCP notifications/tools/list_changed ignored over Streamable HTTP in 2.1.211 (regression from 2.1.210)
- disable-model-invocation skills silently unavailable when skills-dashboard GrowthBook flag is off
- [Bug Report] Opus 4.8 (1M) — recurring unusable assistant outputs
- Desktop app update wiped all custom sidebar groups (Local Storage recreated, unrecoverable)
- Sidebar: archived sessions can't be dragged into groups; context-menu key accelerators don't fire
- [Bug] Anthropic API Error: Access Denied Despite Cyber Verification Program Approval
- [Bug] Claude recommends models not in provided list and ignores specified requirements