Digest / September 28, 2026
Claude Code Issues Digest — September 28, 2026
310 new issues filed ·
96 resolved in anthropics/claude-code (UTC day).
New issues (310)
- [BUG] Using Chrome's **Ungroup** on a Claude-created tab group has no effect — the group and its tabs are unchanged.
- [BUG] Manually dragging a tab from one Claude group to another does nothing
- [BUG] `/bug` opens modally and blocks copy/paste from the Claude app
- [Bug] Working directory change hook not triggering
- [Bug] Anthropic API Error: Safeguards flagged message with Claude 3.5 Sonnet
- [BUG] Opened `.md` in formatted view: text is selectable but "copy" only copies the file/path, not the selection
- [Bug] Anthropic API Error: False positive cyber safeguard flag on legitimate traffic management product spec
- [FEATURE] no-ncurses cli interface
- [GitHub integration] Can't connect to repos in other orgs
- [BUG] /resume inside a running session keeps stale CLAUDE.md and auto-memory context (claude --resume and /clear reload it)
- [GitHub integration] Can't add GitHub repo to claude.ai chats
- [GitHub integration]
- [GitHub integration]
- [BUG] Trust prompt reappears on every launch, but only for the home directory
- [BUG] Max subscriber: web/desktop login redirected to claude.ai/onboarding ("create account") for existing account; mobile apps work — no human response for 7 days
- [Bug] Safeguard flags triggered incorrectly on non-security projects
- [FEATURE] Desktop: in-app Browser in SSH remote sessions
- [BUG] 2.1.281: CLAUDE_CODE_SUBPROCESS_ENV_SCRUB sandbox denies $HOME/actions-runner, so every Bash call fails on a default self-hosted GitHub Actions runner
- [BUG] 2.1.281: scrub sandbox still fails to create ancestor .claude stub when the ancestor is owned by another uid (as root, --cap-drop ALL)
- teeeessss [GitHub integration]
- [FEATURE] Session-scoped /autocompact: change a running session's window only, with separate main/subagent values
- iPad/iPhone: local (Remote Control) sessions and cloud sessions are visually indistinguishable
- [BUG] Cosmetic bug — response is hidden by false "OAuth session expired" error after every message, restored on app reload
- Convos get forked when opening agents after resuming the convo and using left arrow to get to agents.
- Show cloud / local / waiting status for Code sessions in the sidebar Recents (iPad app)
- claude --cloud silently ignores --effort (model is forwarded, effort is not)
- [Bug] Agent ignores remote execution request and runs locally instead
- VS Code: plan review pane shows previous plan version when ExitPlanMode is batched with a plan-file write
- Auto-mode classifier keeps refusing browser actions after the user switches the session to Manual mode, and ignores explicit user approval
- [BUG] Desktop (WSL2): session disappears from sidebar right after its linked PR is merged
- [Bug] Anthropic API: Safety Classifier False Positive on Routine Engineering Design Task
- Sessions/groups inconsistently displayed across mobile app and web (claude.ai/code) vs. desktop app
- [GitHub integration]
- [FEATURE] Agent view: mark several sessions and send one reply to all of them (plus `claude send`)
- [GitHub integration]
- [BUG] code-modernization plugin: make_shards.py contained() uses os.path.realpath() — breaks topology sharding on Windows junction-based shadows (0 shards produced)
- [FEATURE] Raise, configure or document the lineage depth limit (8) on create_session in cloud sessions
- [GitHub integration]
- [BUG] Claude in Chrome: session tabs are created in the background and never activated, so they stay visibilityState "hidden" and screenshots of embedded apps/SPAs come back blank (macOS; same mechanism as #78533)
- [BUG] Windows: timed-out git status leaves orphaned git.exe processes that pile up until memory is exhausted
- [Bug] Fails to generate technical report for internal pentest findings
- VS Code extension: keep the session alive when a Remote Tunnel / Remote-SSH client disconnects, and re-attach the panel on reconnect
- [BUG] IDE integration not working in JetBrains devcontainer setup
- [GitHub integration]
- [BUG] No way to change the extra-usage payment method on a Google Play billed account
- [FEATURE] Keep the main thread's prompt cache warm while it waits for background agents
- [BUG] Harness tells the model to save scripts and outputs to the scratchpad, which macOS wipes at reboot; no setting to relocate or disable it
- Worktree-isolation Bash guard refuses a curl command that references a settings.json env var
- [Bug] Anthropic API Error: Repeated Safety Filter Blocks on Valid Sandbox Configuration Requests
- [GitHub integration]
- Subagent transcripts record start-of-stream `output_tokens` and no `stop_reason` for most turns
- [GitHub integration]
- [BUG] Desktop: ⌘Q / "Quit Claude" menu bypasses the running-sessions quit guard
- [Bug] Anthropic API Error: Auto mode classifier unavailable blocks Bash/Edit calls
- [Feature Request] Improve censorship-circumvention detection messaging and false positive handling
- [GitHub integration]
- No visible notice when auto mode ignores a skill's `model:` override
- claude auth login should defer to apiKeyHelper when configured
- Opus5.5 - warning security for local check leaking.
- [FEATURE] Allow agents and integrations to enrich native OTEL spans with delivery context
- [GitHub integration]
- Worktree isolation refuses every pwsh/powershell/cmd invocation with any argument; git-as-path-segment also refused inside interpreter text
- [feedback] Model: timestamps get invented from elapsed work; compaction summaries carry them forward as facts
- [feedback] Model: a document's headline gets mapped, its body's recommendations die; subagents don't know where a finding lives
- [feedback] Model: a repeated instruction is treated as a reminder, not as a missing mechanism
- [feedback] Model: trusts its own broken probe and a tool's count over its own knowledge
- [feedback] Model: a destructive git form was the first draft; only an incidental refusal saved 85 files
- [feedback] Subagents (Opus/Sonnet): sampling under an 'exhaustive' duty, paraphrased checklist legs, verdict words under a neutrality constraint
- [feedback] Positive: blind same-model judge pairs with different angles surface different real findings; a self-refutation stage is cheap and honest
- [harness] Subagent lifecycle invisible to hooks: resumes log no start, reports delivered twice, tool-written files carry no caller identity
- [harness] Agent tool surface: no per-call tool restriction, model required even for typed agents, shared scratchpad, Bash mangling of backslashes/backticks, user question blocks notifications
- [harness] Subagent cost is dominated by re-reading files: a judgement pass over a few hundred JSON records costs 470K-910K tokens
- [feedback] provenance: Values typed instead of read: a filename from a truncated display, a landing sha from git log -1 on a shared t
- [feedback] freshness: State claims without an observation in the same message: the slot is free, it landed, nothing calls Y.
- [feedback] instrument-vs-proxy: Three independent judge seats and one builder verified a coverage floor with a hand --cov= run instead of the
- [feedback] standing-directive-decay: An operator modifier (bundle from prior art, pre-prepared) decays over a session while action-attached rules s
- [feedback] cheap-capability-scarcity: The model hand-types around verbs that exist in its own tree: a lessons-injection verb unused eight times, a h
- [feedback] green-preference: A deterministic RED on main was classified as a composition flake after passing standalone twice and quarantin
- [feedback] population-claims: Zero genuine regressions remaining was claimed over a 149-file glob that could not contain the failing consume
- [feedback] guard-set-delta: A refactor moved safety-critical halt code behind facades;
- [feedback] silent-side-effects: A missed twin-patch let a route write the real reports/pipeline_state.json during a green test;
- [feedback] rule-wording-drift: A knowledge-gated safety rule kept drifting to its absolute form (never trades);
- [feedback] blob-copy-merge: A manual harvest remedy copied branch blobs over a base main had moved past, reverting a sibling lane's landed
- [feedback] calibration: ETAs padded 1.2-2x;
- [harness] compaction-loss: Compaction replaced the lead's transcript with a 4 KB stub;
- [harness] message-redelivery: Teammate messages from before a compaction were re-delivered 30-60 minutes later, twice;
- [harness] worktree-agent-resume: A completed worktree-isolated agent cannot be resumed by message (No transcript found) although names are said
- [feedback] shared-model-judges: Two same-model judge seats produced largely disjoint findings (a strength) but shared one model-level bias (FB
- [harness] slot-lifecycle: Lifecycle stops are logged by agent id and never pair with tool-id starts: 84 of 158 dispatches unpaired;
- [harness] time-zone: The model reported UTC to an operator on local time;
- i need the bypass permission for project too it cannot by pass
- [GitHub integration] "Add from GitHub" in chat only inserts the repo URL; no GitHub tools available to Claude
- [BUG] BETA_TRACING_ENDPOINT ignores OTEL_EXPORTER_OTLP_PROTOCOL: headless sessions (desktop app, claude -p) silently drop all events and traces
- [Bug] Anthropic API Error: Opus 5.5 Safeguards Incorrectly Flagging Non-Cybersecurity Content
- [GitHub integration] Private repository attachment in Chat is treated as a public URL
- [BUG] Auto mode classifier blocks reading .env.prod inconsistently across sessions, even on explicit user request
- [Bug] Anthropic API Error: Safeguard Policy Blocking Legitimate App Development Use Case
- [BUG] 2.1.283: sandbox.excludedCommands never unsandboxes aws calls with --profile or --endpoint-url (Linux)
- [BUG] Linux sandbox: filesystem.allowRead on a symlink inside a denyRead directory has no effect
- [Bug] Auto mode classifier returns no verdict, blocking Edit/Bash operations
- [FEATURE] Interactive terminal chat with a running cloud session (attach without teleporting)
- [Feature Request] Allow Claude Code to assist with personal network file transfers and password management for self-employed users
- [GitHub integration]
- [BUG] Agent view's primary agent `claude` is offered as an Agent tool subagent type, where it's told a job-list classifier reads its output
- [Bug] Anthropic API Error: Opus 5.5 safeguards flagged message with reasoning_extraction
- [GitHub integration]
- [harness] Lead session has no live view of subagent cost, and subagents have no default channel back to the lead
- [harness] PostToolUse hooks run synchronously for subagents with no agent or worktree context; a per-edit test hook starved a builder for two hours
- [Bug] Uncontrolled API spending on non-Anthropic providers during failed test runs
- [BUG] Cowork (Desktop, cloud session): model picker change is ignored while a turn is processing, and the picker display reverts to the default model while a different model is being served
- [BUG] Long text on the chat doesnt go up to show claude response
- [BUG] macOS: Claude Code probes /home (autofs) and triggers a Network Volumes TCC prompt from a pre-warmed Desktop session; denied access retried on every Bash call
- [GitHub integration]
- iOS Simulator tool can't detect booted devices — Xcode 27 renamed Simulator.app to DeviceHub.app
- [Bug] Unclear content moderation rejection without specific violation details
- [GitHub integration]
- [BUG] Agent frontmatter `effort` is still ignored on the `--agent` session path (2.1.283) — follow-up to #81677
- [BUG] Xcode 27.0 Claude Agent: claude 2.1.280 exits with code 1 within 40 ms of Xcode's first stdin write; conversations hang on "thinking" forever (2.1.220 works)
- [BUG] CLI: after /login → "Anthropic Console account", OTel and API request metadata carry no identity (organization.id, user.* empty; metadata.user_id.account_uuid = "")
- Feature request: a cloud "curator" agent that routes tasks across a user's own machines
- [BUG] Resuming a session that is already running forks its transcript; the next resume silently drops one branch
- [BUG] Desktop 2.7032.0: PR monitor loops on GraphQL NOT_FOUND (~5000/h), falls back to REST and exhausts both GitHub rate limits
- Bug Report
- [BUG] --worktree session enforces the launch checkout's permission rules instead of the worktree's .claude/settings.json
- I need more information to generate a proper GitHub issue title. The word "save" alone doesn't provide enough context about the bug or feature request. Could you please provide: 1. What is the actual problem or feature request? 2. What error message (if a
- [Bug] Anthropic API Error: [cyber] classifier over-triggering on authorized red-team operations
- [BUG] Sandboxed Bash tool blocks .git writes (refs/heads, index.lock) despite filesystem.allowWrite/excludedCommands overrides
- [GitHub integration]
- 共有フォルダの読み込み書き込みエラーになる
- [BUG] Squash request force-pushed uncommitted working-tree changes (git add -A after reset --soft)
- [GitHub integration]
- [BUG] : Folder picker in Code tab makes existing sessions look like they switched folders
- [Feature Request] Add queue option to pop items only after task completion
- [Feature Request] Add confirmation dialog for quality ratings to prevent accidental feedback submission
- Synced plugin update moves the plugin folder of an open session to .trash; its hooks then fail open
- [BUG] /insights generates empty report — facets directory never created (still occurring)
- [BUG] AskUserQuestion invoked with placeholder content while waiting on background agent
- [Bug] File Permission Scope Rules Inconsistent Across Sessions
- [Bug] Stop hook repeatedly triggers with "full milestone completion" demand in loop sessions
- [Feature Request] Support Claude 3.5 Opus model access for registered cybersecurity users
- Repeated "Authentication failed" in desktop app session panel behind TLS-inspecting corporate proxy
- Auto mode: server-side safety classifier intermittently returns no verdict, blocking Bash and ScheduleWakeup entirely
- [Bug] Classifier service unavailable or non-functional
- [GitHub integration]
- [BUG] `/compact` typed inside the subagent view compacts the main conversation instead
- [FEATURE] Agents view: show when each session last worked instead of how long it worked
- [Segfault at 0x10 in nextTick drain on DarkWake after 5.6 days — 2.1.278, Bun 1.4.3 (8c28b3167), macOS 27.0 arm64]
- [FEATURE] Desktop Code tab: a row of user-defined quick-command buttons above the prompt box
- [Bug] Web design requests triggering content filter restrictions
- [BUG] Desktop app: Claude Code background processes keep running after sessions end and push macOS into ~45 GB swap, hanging the Mac overnight
- [BUG] /usage progress bars render for ~1s then disappear in PyCharm terminal (both engines); fine in system terminal
- [BUG] Scheduled tasks run with personal preferences frozen at task creation, not the current "Instructions for Claude"
- Feature request: Remote Control for scheduled task runs
- [BUG] Glob ignores .ignore/.rgignore by default (--no-ignore); should use --no-ignore-vcs so only .gitignore is skipped
- [GitHub integration] fdsfas
- [Bug] Auto mode classifier returned no verdict error
- [Bug] Anthropic API Error: Auto Mode Classifier No Verdict for Write Action
- [Bug] Server-side auto mode classifier returns "no verdict" errors; allow-listed commands blocked
- [Bug] Autoclassifier feature not functioning
- [BUG] The server-side auto mode classifier is temporarily unavailable (error)
- [Bug] Auto-learn wizard silently reverts bypass settings without user confirmation
- [Bug] Auto mode classifier returns no verdict for Chrome MCP tabs_context tool
- Scheduled tasks ignore project-pinned model (run on claude-opus-5-5 instead of settings.json's model); permission gate occasionally stalls despite defaultMode:auto
- [Bug] Anthropic API Error: Security Verifier Unavailable in Auto Mode
- [Bug] Anthropic API Error: Auto Mode Classifier No Verdict for WebSearch
- [Bug] Auto mode not initializing or executing properly
- [Bug] Auto mode exhausts session limit too quickly during long operations
- [FEATURE] Individual plan above Max 20x (e.g. $300–400/month)
- Auto-mode permission classifier misses git reset --hard when bundled in a multi-line Bash command, then false-positives on an unrelated later command
- [Bug] Default effortLevel parameter has no effect
- [Bug] Subagents fail to terminate reliably
- Auto mode: server-side safety classifier repeatedly returns no verdict, blocking all write/exec tool calls (Edit, Bash, PowerShell)
- Claude Desktop: no permissions.defaultMode or sandbox setting suppresses tool-approval prompts
- [Bug] Prompt suggestion ghost text not appearing with Claude 3.5 Opus
- Quanty[GitHub integration]
- [BUG] hasTrustDialogAccepted reverts to false for previously-trusted, actively-used projects (macOS, 2.1.283)
- [BUG] [A11y] VS Code extension 2.1.283: VoiceOver no longer echoes typing in the prompt input (macOS)
- Claude Code (Opus) lied to the user about progress, overrode explicit instructions, and burned ~9.5M subagent tokens while the assigned task barely advanced
- [Bug] Opus 5.5 incorrectly flags legitimate embedded systems debugging and provisioning as cyber threats
- [FEATURE] Always-on personal assistant mode (like Meta Muse / Grok Bot)
- [FEATURE] Setting to disable the "switch to auto mode" option/nudge on permission prompts
- Desktop app: session list empty for weeks while ~/.claude/projects transcripts remain on disk
- Routines: MCP tool call (Gmail send_message) silently no-ops but run reports SUCCEEDED
- session_info MCP tools (list_sessions, read_transcript) missing from Cowork sessions since 9/22–23
- [FEATURE] Usage stats should combine all devices on the same account
- [Bug] CLI not accepting single-letter option selection from numbered list
- [BUG] iOS Remote Control requests disabled Auto mode on new chat, then shows Accept edits while host remains in bypass
- [GitHub integration]
- [BUG]Gmail / Google Calendar / Google Drive MCP connectors unreachable (NXDOMAIN)
- [harness] Parallel Bash calls in one response share the shell's working directory; a cd in one call breaks a sibling's relative paths
- [harness] TaskUpdate/TaskCreate report success while dropping the metadata merge or applying only one field when the task store is busy
- Desktop app session list doesn't include sessions started from the CLI
- EVER SIGNLE TASK DONE BY CLAUDE CODE HAS TO BE REDONE ADN CORRECTED BY CODEX
- [FEATURE] Conversation anchors: re-enter an approved plan per item, keep the code, get a ledger of finished items
- [FEATURE] Roles: one committed, selectable bundle of agent, settings, instructions and write scope per team role
- [FEATURE] Session journal: local file sink for OTel events, with source-side privacy controls and a few missing events
- [BUG] Mobile app does not show skill arguments in the user message
- [Bug] False-positive safety block on browser fork transparency-mode bug fix task
- Unattended `--bg` sessions can hang for hours on an unanswered auto-mode permission prompt, with no timeout and no way to see it
- [DOCS] HTTP_PROXY env var can also be used to redirect sandbox traffix
- [BUG] /model: "Default (recommended)" shows your saved pick after restart
- [Bug] Auto mode classifier incorrectly flags harmless read-only operations as destructive
- [GitHub integration]
- [MODEL] Claude leans towards git worktrees despite includeGitInstructions=false and instructions not to use git
- [BUG] Desktop app (Code tab): unsent draft is lost when switching sessions (CLI unaffected)
- [Feature Request] Add override/exemption mechanism for safety flags during local debugging scenarios
- [GitHub integration]
- [BUG] Enqueued message doesn't show correctly after being sent
- [Bug] Rate limiting occurring for MAX subscription users
- [BUG] Desktop: plugin MCP OAuth sign-in page silently not opened (denied_no_activation) if the confirm dialog isn't answered within ~5s
- Auto mode classifier halts explicitly user-authorized work: a local `docker build` flagged as "Production Deploy"; removing an owner-ruled fake control flagged as "Security Weaken"
- Projects: coordinator + thread sessions merged PRs with red CI and ignored AGENTS.md during multi-thread implementation
- Claude Desktop background CPU usage contends with foreground GPU-bound apps (games) even when minimized
- [GitHub integration] Cowork sessions on one account don't get list_repos / add_repo tools; same GitHub App works on my other account
- [MODEL] Fabricates the next user turn at the end of its answer and executes the fabricated instruction (triggered by a per-turn UserPromptSubmit hook reminder)
- Desktop app: "Interface font" setting doesn't apply — stays on OpenDyslexic even after selecting Sans Serif
- Feature request: configurable keyboard shortcut for voice dictation in desktop app
- [GitHub integration]
- [BUG] [Billing] Refund denied by Fin bot citing "previous refund" — unrelated auto-renewal, no human escalation path
- [GitHub integration] - Support GitHub Projects
- [GitHub integration]
- [FEATURE] Cloud sessions: a sign-in form credential, filled without the model seeing the password
- [Bug] Diff panel cannot be closed and persists when empty
- [BUG] Desktop Projects (beta) / Remote Control threads: `/` autocomplete omits user skills (~/.claude/skills) and plugin commands — they execute when typed in full
- [FEATURE] Cowork built-in browser: apply View → Zoom to the browser pane when it has focus (persisted, like Codex)
- [FEATURE] Have the Claude surface be able to get the time zone from the host system
- [FEATURE] VS Code extension: allow voice dictation in Remote - WSL windows (mic is available via WSLg)
- bug: desktop slash menu lists /fast twice, once as "Custom command"
- [Feature Request] Add security testing mode for validating exploit mitigations
- Auto mode classifier denial cannot be overridden by the owner's explicit in-chat consent (no approval prompt)
- Worktree-isolated subagent cannot create a directory link: cmd/powershell refused, Git Bash ln -s silently deep-copies (FB-033)
- [GitHub integration]
- [BUG] Windows: computer-use screenshot sends TextInputHost.exe into a 97-99% CPU loop
- [Bug] Claude 5.5 Sonnet unavailable for CVP members
- New cloud Projects is good, with some limitations
- [FEATURE] Title: Chat sessions lose working-directory context on new session — no way to spawn a new chat pre-attached to an existing project folder
- [FEATURE] customHeadersHelper: fetch ANTHROPIC_CUSTOM_HEADERS from a command (like otelHeadersHelper)
- [GitHub taha integration]
- Terminal tab title not set when resuming with `claude -r` (stays at version string)
- sonnet 5
- [BUG] Cowork (Windows): tools/MCP connectors become unavailable as soon as voice mode is activated
- [Bug] Child agents continue running after parent agent termination, causing excessive token consumption
- [GitHub integration]
- [Feature Request] Add scrollbar indicator to fullscreen TUI conversation view
- [Bug] Post-install plugin configure dialog: ←/→ still switch tabs instead of cycling userConfig options (2.1.284)
- [Bug] Message incorrectly flagged as policy violation in demo context
- [BUG] Can't read a Claude Docs artifact my account can view: Docs connector denies read, Artifact read returns only the viewer shell
- [Bug] Model performance degradation reported
- [BUG] Can't read a Claude Docs artifact my account can view: Docs connector denies read, Artifact read returns only the viewer shell
- "One-time reset" in Settings → Usage deleted all chat/session history, not just usage data
- Feature request: shared secret/credential store across Claude Code, Desktop, and Web
- [Bug] Overly restrictive safety filters block legitimate testing workflows
- Web sessions: add a mode "write access to every repo the GitHub token allows" (single source of truth for permissions)
- Mouse-tracking escape sequences leak into PowerShell prompt after terminal idle
- Desktop app cannot access the camera: no NSCameraUsageDescription, so macOS refuses silently and never prompts
- [BUG] Web (cloud) session, Auto mode: classifier denies an explicitly user-requested add_repo(push); UI/mode guidance from the assistant was wrong
- [Bug] Highlighting text fails to copy to clipboard
- [Bug] CVP verification not preventing safeguard measure enforcement
- Auto-mode classifier blocks user-approved removal of leaked personal data from a template ([Auto-Mode Bypass]), with no way to confirm in-session
- [BUG] Cloud sessions: `gh` is not installed although the proxy already injects a GitHub credential; the GitHub MCP is the only supported path and is limited to session repos
- [BUG] [BUG] Web session: the assistant's built-in documentation has no page on permission modes, so it gives ineffective advice
- [FEATURE] Pass quotaLimits (rateLimitType, resetsAt) to the StopFailure hook input
- roblox jump
- PostToolUse hooks never fire from the desktop Code-tab surface
- [BUG] Fork subagent transcript duplicates the parent's last response (same requestId and usage), so per-file usage counts it twice
- [GitHub integration]
- [GitHub integration]
- Desktop app: second machine's sidebar keeps listing archived/ended sessions (survives app restart)
- [GitHub integration]
- I can see this appears to be a list of bug reports, but the text is primarily in French and the descriptions are truncated or unclear. Without being able to see the full, clear bug report details in English, I cannot generate an accurate technical issue ti
- [BUG] Panel opened through the vscode://anthropic.claude-code/open link has no title row
- [FEATURE] A standard text-editing model for the prompt input: selection, cut/copy/paste, word navigation (built on OSC 52, bracketed paste, kitty keyboard protocol)
- [GitHub integration]
- [BUG] Figma MCP tool calls fail with SSE JSON parse error on long responses (desktop app, Code tab)
- [GitHub integration]
- Boring Website
- [GitHub integration]
- git worktree of an already-trusted repo still re-prompts "Workspace not trusted" on 2.1.284, despite the documented fix
- [a11y] Desktop sidebar: focus is lost after Pin, Archive, or Delete instead of staying on the trigger or moving to the nearest sibling
- statusLine: only first line renders when command outputs multiple lines
- [Bug] Model refusing legitimate requests due to overly restrictive safety filters
- [FEATURE] Desktop Code tab: let atlassian.net (Jira) links open in the Browser pane
- [Bug] Infinite loop of code generation without functional output - Claude continuously regenerates code that fails testing despite claiming verification is successful
- [BUG] Fable weekly usage counted (20%) with zero Fable requests since reset — only Opus 5.5 / Sonnet 5 used (Max 20x, Windows)
- [Bug] Claude refuses to enter Stripe test card numbers on a staging environment (only allows localhost); worked in earlier versions
- [BUG] CLAUDE.md rules not enforced across sessions
- [Bug] Anthropic API Error: Overly aggressive content filtering on harmless message
- [Bug] Session state not persisted between CLI invocations
- [FEATURE] Setting to open the /diff panel on the left instead of the right
- No practical way to upload a locally-built file to Google Drive when only the Drive connector (not Docs/Slides) is enabled
- [Feature request] Hebrew (he) support for voice dictation
- [Bug] Remote Control bridge/sdk-cli sessions missing Artifact tools
- [FEATURE] A user-scope setting to pre-trust directories, for scripted `claude --bg`
- [Bug] MCP plugin servers disabled per-project don't persist across no-folder sessions
- [GitHub repositories not appearing in Claude Code GitHub integration]
- [BUG] Claude Desktop (Windows MSIX): Dispatch/Cowork host sessions fail with "native binary … failed to launch" because host-cwd exceeds MAX_PATH
- User-scope ~/.claude/CLAUDE.local.md loads its content but silently never expands @imports
- A symlinked rules/memory file is realpath-resolved before @imports expand, so relative imports resolve against the symlink target
- StopFailure: error "rate_limit" covers both a resumable usage-limit freeze and a non-resumable credits condition, and the reset time is only in prose
- [FEATURE] Queued messages should reach the model while a long tool call is running
- Bug: Project threads don't use my cloud session credit.
- [Bug] False positive reasoning_extraction safeguard flag on legitimate requests
- [Bug] Agent workflow performance degradation on trivial tasks
- [Bug] Safety classifier blocks legitimate admin UI code generation for user's own extension
- Classic renderer drops or overwrites output when anything above the visible window changes height (verbose makes it frequent)
- [GitHub integration]
- [FEATURE] Forward the URL fragment into published artifacts so they can be deep-linked
- [GitHub integration]
- [BUG] 2.1.284 freezes on first Enter: new sandbox glob expander synchronously walks all of ~ for "~/**/…" denyRead patterns (follows symlinks, unbounded memory); 2.1.280 fine
- [Bug] Excessive request blocking or rate limiting preventing normal operation
Resolved issues (96)
- CLAUDE.md @import cascades into plain relative-path mentions inside imported files, ignoring non-`@` syntax and "do not auto-load" notes
- [BUG] Compaction stalls permanently at 95%: event loop goes idle in kevent64 with zero orphaned tool calls (macOS, 2.1.270)
- Session goes permanently idle after a tool_result is delivered, with zero pending tool calls and no compaction in progress
- [BUG] Linux Cowork: cloud sessions can never link to computer — enclave key not implemented on Linux ("device not registered (no row-PK)"); reopening closed #79232 / #77348[BUG]
- [Bug] Fails to generate technical report for internal pentest findings
- [BUG] Running any version above 2.1.278 locks up on FreeBSD
- [GitHub integration]
- [GitHub integration]
- [GitHub integration]
- Feature request: adjustable line/column width in desktop app
- testgithub
- Claude Desktop 2.9939.2.0 fails to launch with 0x80070020; five stale Silo hives remain after service stop
- Grad Test Project
- CLAUDE666
- Lecture vocale (TTS) : ponctuation mal placée et certains mots dénaturés, surtout en début de liste numérotée
- [Bug] Claude 3.5 Sonnet producing low-quality or irrelevant responses
- [Bug] Claude ignores CLAUDE.md instructions and fails to recognize custom file formats
- [Bug] Agent uses stale memory from merged PRs, causing unnecessary resource waste
- I am not able to see my project conversation from my phone. Even when I try to …
- blockers and questions dont surface on orchestrator. very annoying. orchestrato…
- [Bug] Anthropic API Error: Request blocked due to security policy
- Desktop app keeps every session's process tree alive — 183 idle sessions, ~2,600 MCP sidecars, ~160GB RSS
- [BUG] Gmail connector: get_thread silently truncates long threads — newest messages invisible to the agent (refiled from claude-ai-mcp#239)
- Auto-mode classifier blocks a corrective rsync but misses the destructive one that caused the damage
- Korean (Hangul) IME composition inserts stray characters (e.g. "?") across CLI TUI and Agents view
- Internal /__remote-workflow command leaks into slash-command autocomplete in VSCode extension
- SIGABRT crash on VPN connect/disconnect (network route change) — regression in 2.1.177, still present in 2.1.202
- [BUG] Bun 1.4.0 JSC GC use-after-free segfault (SlotVisitor::drain -> MethodTable::visitChildren) at 0x...00000050 on Windows x64 baseline - symbolized trace, reproduces at idle load - CLI 2.1.209
- Subagent assistant message content replaced with ~145KB eval fixture (Alice in Wonderland) despite output_tokens: 5
- Updater treats 2.1.209 as newer than 2.1.210 (lexical version compare, offers downgrade)
- Subagent replies delivered to the wrong session when multiple sessions run agents concurrently (Windows)
- [BUG] CoworkVMService — found in Disabled state, clean exit 0 after re-enabling, no diagnostic logs (matches #36801 signature)
- [Bug] Frequent Anthropic API errors without specific error details
- Sub-agent (Task tool) returned hallucinated ccd_* MCP 'developer mode / trust escalation' server instructions as its result (0 tool calls)
- [BUG] Claude Code Desktop: EINTR error when opening session in Desktop folder (macOS)
- [BUG] commercial-legal practice profile does not persist across Cowork chat sessions on Windows
- Stale/out-of-place conversation content resurfacing inside unrelated tool results (TaskCreate, Bash)
- [BUG] Cmd+N broadcasts new-conversation to all open Claude tabs (allComms fan-out) — reopens #31455
- Resumed session opens in panel instead of sidebar, ignoring preferredLocation: sidebar
- [BUG] Session Race Condition: 888 Duplicate Transcripts in 7 Days (2189 files, 57MB)
- [BUG] Plan review: "Edit with Claude" revise window loses entered text when accidentally closed or when switching chats
- [BUG] ECONNRESET in Claude Desktop in-app Claude Code, while standalone CLI works on same machine/network
- [BUG] Squash request force-pushed uncommitted working-tree changes (git add -A after reset --soft)
- [BUG] Desktop runs `gh auth token` to pull the GitHub CLI token (undocumented) and respawns it every 10s when it times out
- [BUG] MCP stdio server lists no tools since protocol 2026-07-28 is negotiated (tools/list times out, mcp-grafana)
- [GitHub integration]
- Windows: synchronous command hooks wait for persistent descendants
- [BUG]Gmail / Google Calendar / Google Drive MCP connectors unreachable (NXDOMAIN)
- [GitHub integration]
- [GitHub taha integration]
- Claude-in-Chrome native host rejected by Gatekeeper (syspolicyd security_exception -67062/-67002) — macOS only, works fine on Windows
- [BUG] Can't read a Claude Docs artifact my account can view: Docs connector denies read, Artifact read returns only the viewer shell
- "One-time reset" in Settings → Usage deleted all chat/session history, not just usage data
- stable channel has not advanced past 2.1.236 since Aug 19, against a documented cadence of about one week
- [GitHub integration]
- [GitHub integration]
- Epidemic Sound MCP connector: OAuth fails with "Client not found" — client_id set to user email instead of registered app ID
- [Bug] Web search claims to read pages that return DNS errors (NXDOMAIN) — fabricated source citations
- [Bug] Web search claims to read pages that return DNS errors (NXDOMAIN) — fabricated source citations
- Bottt
- [GitHub integration]
- [GitHub integration]
- Video analysis precision lost since Sept 16 update
- [BUG] [claude.ai Projects] Slash in knowledge item title creates folders in the knowledge panel
- Claude attributes other AI models work to itself since Sept 16 update
- [GitHub integration] i cant get the 100$ credit
- [GitHub integration]
- O'Donnell Winter Comeback
- [Bug] Agent dispatch fails silently without error notification
- [Bug] Excessive token consumption on repetitive email test cases
- [Bug] Excessive refusal on context compaction validation requests
- [Bug] Token waste issue in Claude Code operations
- [Bug] Claude makes unfounded assumptions about well-documented APIs on Opus
- [Bug] Unwanted automated reminder appearing in every message
- [Bug] Excessive token consumption during usage
- VS Code extension misreports normal per-turn process teardown as exit code 4294967295
- System prompt's model identity line contradicts /model and the injected commit trailer in the same session (2.1.210)
- [BUG] low contrast text
- [BUG] Plan mode permanently corrupts prompt color default forever
- [BUG] Claude Desktop crashes/restarts when Google Drive virtual drive unmounts during file-index scan
- [BUG] Cowork incompatible with Universal Audio Apollo (UAD kernel drivers) on Windows 11 Pro — Hyper-V conflicts with pro audio hardware
- [BUG] Plugin and marketplace source refs not surfaced in UI — users cannot see which channel or plugin version they are tracking
- [BUG] All running sessions segfault simultaneously on macOS wake from deep sleep (Bun 1.4.0 GC socket finalizer, SIGSEGV at 0x4)
- [BUG] Subagent card stays 'running' after completion notification is delivered; only manual stop freezes it, then persists as a ghost card visible from another session
- [BUG]
- [BUG] Reinstalling Claude Desktop on Windows wipes Code tab session history (no warning, no recovery)
- /model picker: Enter persists the choice globally — should default to session scope like /effort
- [BUG] task_reminder nudge fired mid-tool-loop rewrites cached history (near-total rebuild per firing in autonomous sessions); 29% of sessions double-write the opening context — 3-week measured audit
- Auto-reconnect after wake exhausts after 3 attempts (~6s), no further retry
- Orphaned duplicate session runner processes after Desktop App restart
- Remote-control bridge auto-registration silently stops after app update (session stays connected regardless)
- [BUG] Desktop app microphone dictation ignores "language" setting — Ukrainian speech transcribed as gibberish
- [BUG] Code sessions started via Dispatch don't appear in mobile Dispatch list
- [BUG] VS Code extension: Alt+K / Ctrl+Esc don't focus the input when the Claude panel is in an auxiliary window; at-mention is broadcast to all visible panels
- stream-json: result event `.result` drops all but the first text chunk when one assistant message streams as multiple events
- [GitHub integration]