Digest / September 25, 2026
Claude Code Issues Digest — September 25, 2026
398 new issues filed ·
64 resolved in anthropics/claude-code (UTC day).
New issues (398)
- [GitHub integration] will not reconnect
- [GitHub integration]
- [Bug] Anthropic API Error: Message flagged by safety filters with reasoning_extraction
- [Bug] Anthropic API Error: Consistent failure with HTB Academy queries
- Stop hook block reason is sent to the model twice (meta message + hook_blocking_error attachment)
- [MODEL] [SECURITY] Inconsistent behaviour: entry of test credentials and data. Inconsistent handling of test passwords in the user's own dev environment: agreed, rewrote global config, then reversed; incoherent Playwright position (relates to #78160)
- [BUG] Windows Desktop: device bridge handshake never acknowledged for ~18 min (connect frame sent → handshake timeout), survives reboot + app update, then self-recovers
- [Bug] Overly restrictive content classifier blocking legitimate use cases
- [BUG] Claude Desktop (Linux/Cinnamon X11): window menu from custom title bar gets stuck, desktop appears frozen (pushModal: begin_modal failed)
- [GitHub integration]
- [GitHub integration]
- [BUG] 400 diagnostics.previous_message_id bricks a session after switching from an ANTHROPIC_BASE_URL gateway back to first-party (anchor accepts any requestId)
- [BUG] Windows Desktop: linked computer stuck on "Asleep or app closed", cloud sessions and cloud scheduled tasks fail with "not connected to the bridge" (survives reboot, re-login, device removal, app update)
- [BUG] Cowork device bridge never completes device handshake after socket opens (regression after update to 2.9939.2)
- Permission dialog: button order shifts, causing accidental Deny/Approve-All clicks
- [FEATURE] VS Code extension: render Markdown formatting in the prompt input while typing
- [Bug] Claude refuses to modify database without clear reason
- Recognize TERM_PROGRAM=Kaku as hyperlink-capable (OSC 8)
- [Bug] Anthropic API Error: Opus 5.5 safeguards incorrectly flagging legitimate biology research
- [FEATURE] AskUserQuestion: keep Other-box text as notes when a single-select option is chosen
- GitHub Restrictions too harsh in Projects
- [GitHub integration]
- Desktop: session in a worktree of a trusted repo is permanently stuck on "workspace isn't trusted" — trust dialog never shown (app/CLI trust deadlock)
- [GitHub integration]
- [Bug] Input box stops accepting keystrokes in 2.1.282 after 0-90 seconds into session
- [BUG] Windows Task Scheduler silently fails to launch `claude -p` subprocess (no process, no output, stuck "Running" status)
- [Bug] CLI lacks context and explanation when presenting questions to users
- VS Code: MCP permission prompts show only the tool name, not what the call does
- [GitHub integration]
- [BUG] macOS Desktop App Thinks Computer is Asleep While Using Desktop
- Plugin submission: message-queue for rate-limit recovery across messaging channels
- [GitHub integration]
- [Bug] Poor output quality with high token consumption
- [BUG] Desktop usage widget mixes two token units: cached days count cache tokens, live days don't, so models first used after lastComputedDate vanish from the Models tab
- [BUG] iOS Simulator panel/tool ignores iPhone Duo fold state and orientation (sideways view, stale 466x678 coords, screenshot captureFailed)
- [BUG] Windows: main process survives beforeQuitForUpdate; relaunch blocked by single-instance lock (update deadlock, reproduced on two versions)
- Auto mode classifier inconsistently blocks production deploys explicitly allowed in autoMode.allow
- [GitHub integration]
- [Windows] Hook timeout doesn't release the tool call while an orphaned descendant still holds the hook's stdout
- [GitHub integration] unable to connect claude to github
- [Bug] Desktop app displays assistant reply text that the transcript records only as a (summarized) thinking block
- [GitHub integration]
- [FEATURE] Scoped, owner-approved test-account sign-in for dev/QA (domains, iOS Simulator/Device Hub, native apps) without the model seeing credentials
- Assistant text preceding a tool call is dropped from the transcript after auto-compaction
- [GitHub integration]
- [GitHub integration]
- [BUG] Desktop app (Code tab): "Auto mode is on" notice keeps showing while the session is in Manual mode
- [GitHub integration] private repo not visible
- Background Agent .output file is 0 bytes on Windows without Developer Mode, while the launch result calls it "the full subagent JSONL transcript" (still in 2.1.282; follows #55263)
- [GitHub integration]
- [FEATURE] Spell check (red underline) in the terminal prompt input
- [FEATURE] Claude Desktop Terminal panel: expose the owning Code session's id in the tab environment
- [GitHub integration]
- [GitHub integration]
- [FEATURE] Middle display level: show tool command + first N output lines (between collapsed and verbose)
- [FEATURE] Official Arch Linux (pacman) support for Claude Desktop, not just the CLI
- [GitHub integration]
- [Bug] Opus 5.5 inconsistently applies classifier blocks to privacy research previously approved by earlier models
- I appreciate you sharing this detailed feedback, but I need to let you know that I'm Claude Code, Anthropic's CLI tool for assisting with coding tasks. I'm not able to help with this particular issue because: 1. This appears to be in Japanese and describe
- Cowork (cloud): routine sandbox Bash commands keep stopping for "Allow once" approval mid-task, despite user granting blanket permission
- Feature request: adjustable line/column width in desktop app
- [GitHub integration]
- [Bug] Token usage not reset when resuming agent after hitting limit
- [Bug] Deployment proceeded despite explicit non-deployment instructions and assurances
- [Bug] Cyber Verification Program prompt repeatedly shown to already-verified account
- [GitHub integration]
- # Our Approach to User Safety
- [GitHub integration]
- [BUG] Code block copy adds trailing newline causing >> in PowerShell
- [BUG] Cannot edit or delete environment on phone app, only create environment
- [BUG] A message that is only a paste is treated as untrusted: the model asks for confirmation instead of doing the task (tengu_virtual_pancake)
- Desktop app: Claude sends a password prompt to the Terminal panel while the user is on mobile via Remote Control, so the password gets typed into chat
- [BUG] Subprocess initialization did not complete within 60000ms — VS Code extension config-probe/session spawn race (.claude.json EPERM collision)
- [Bug] Cyber security filter incorrectly blocks proposal generation
- [BUG] UI colors quantized to 256 inside WSL under Windows Terminal (WT_SESSION not treated as truecolor)
- [GitHub integration]
- [Bug] Plugins enabled by project `enabledPlugins` load, but the Desktop plugins page and `claude plugin list` report them as not installed
- [GitHub integration]
- [GitHub integration]
- Auto-mode setup in 2.1.268 replaced unrelated user settings, including hooks and permission rules
- [GitHub integration]
- testgithub
- [BUG] Subagent's startup git status reports the parent session's cached branch, not its own working directory's
- [Bug] Claude Code unable to interact with LinkedIn platform
- [Bug] Anthropic API Error: Overly restrictive safeguard blocking legitimate defensive cybersecurity research
- You've hit your session limit · resets 3:10pm (Asia/Makassar)
- [Feature Request] Code obfuscation/encryption support for local development
- [GitHub integration]
- [BUG] Running sessions lose all access to ~/Documents (macOS TCC) when a plugin's npx-launched MCP server is replaced on disk by a newer session
- Synced plugins: startup race with generation rename (~g2) yields spurious 'commands path not found'
- Prompt cache never hits after a mid-conversation `role:"system"` message (resume and hook output), Opus 5.x / Fable
- [GitHub integration]
- [Bug] Anthropic API Error: reasoning_extraction on model claude-opus-5-5
- [GitHub integration]
- [BUG] claude -c resumes the globally most recent session, ignoring the current project directory
- [BUG] /init writes CLAUDE.md in an AGENTS.md project, silently disabling the instructions it just read
- [GitHub integration]
- Claude Desktop 2.9939.2.0 fails to launch with 0x80070020; five stale Silo hives remain after service stop
- [Bug] Valid API authorization credentials blocking execution
- I don't have enough context to generate an appropriate GitHub issue title. Could you please provide: 1. **The specific problem or bug** you encountered 2. **Error messages** (if any) 3. **What you were trying to do** when the issue occurred 4. **Expected
- Cowork: Project-bound scheduled tasks fail to spawn ("Failed to run scheduled task")
- PR verdict issued before the user's mandatory review steps were done (rubber stamp)
- Same forbidden 'unverified but disclosed' verdict repeated twice right after an explicit correction
- Asked what a documented procedure says, the model substitutes commentary and forbidden steps instead of quoting and applying it
- git index written (git rm --cached) despite a standing ban on git write operations
- [GitHub integration]
- [Bug] Anthropic API Error: Cyber Safeguards blocking legitimate security research requests
- [Bug] Insufficient information provided for issue classification
- [GitHub integration]
- orion-strom
- New Project glitch: one of my prompt disappeared - I typed fairly normal message with screenshots a…
- [Bug] CLI sessions share bash execution state unexpectedly
- I have cloud credits but for some reason when my weekly session ended due to us…
- pathToClaudeCodeExecutable setting ignored; extension hardcodes embedded binary path
- [GitHub integration] - cannot connect to repo
- Feature request: Keyboard shortcuts for View group-by switching
- [GitHub integration]
- [Bug] Overly restrictive security rules block legitimate AWS credential access
- [Bug] Claude provides false billing information about Fable usage on Claude Max
- Model delivers a review verdict before completing the mandated checklist steps
- Model stops reviewing after finding the first real defect
- Model skips the mandated step of reading the current state of every touched file
- Model reads a PR thread from a text dump after being told to use the CLI directly
- Model omits reading the linked issue during a mandated review
- Model does not cite a normative source per component as mandated
- Model performs incomplete caller searches for changed functions
- Model misses a structure-rule violation it enforced earlier in the same session
- Model runs test suites and starts local servers during a review the user defines as code-only
- Model keeps using a banned shell reading method after an explicit ban
- Model takes an unrequested action (patching a scratch copy of the codebase) under zero-autonomy rules
- Model overclaims verification results in a review summary
- Model maintains a verdict while mandated evidence is still missing
- Grad Test Project
- Queued prompts which are sent and queued after there are no available tokens for the current session, are deleted when the desktop app closes, even though draft (not sent) prompts survive
- [BUG] Plugin command with arguments: unprefixed name gives "Unknown command" in the interactive CLI
- Extension chat webview causes renderer OOM crash (code 5) after large agent turns
- [Bug] @-autocomplete for paths outside cwd shows deleted files and doesn't use fileSuggestion command
- [GitHub integration]
- [GitHub integration] Github x Claude auth Loop
- IS_DEMO skips the trust dialog without granting trust, so the status line and hooks silently never run
- [GitHub integration]
- [FEATURE] Support GitHub Projects v2 (add items, set fields/Status) in Claude Code web + Claude GitHub App
- [Bug] Detached session closure unclear when process still running
- CLAUDE666
- I appreciate you reaching out, but I should clarify: I'm Claude, an AI assistant made by Anthropic. I'm not "Claude Code" or a tool for GitHub issues, and I don't have a specific function for formatting GitHub issue titles. Additionally, I'm not able to h
- [GitHub integration]
- Allow creating a new sidebar session programmatically (for context handoff between agents)
- [GitHub integration]
- Lecture vocale (TTS) : ponctuation mal placée et certains mots dénaturés, surtout en début de liste numérotée
- [BUG] Desktop: finished Project threads keep live sessions and fill the process cap, so new Project sessions are refused
- [Bug] Claude Opus 3.5 ignoring CLAUDE.md instructions due to reasoning extraction safeguards
- [GitHub integration]
- [GitHub integration]
- [BUG] MCP binary/large results are dropped ("tool-results path refused … is a link") when ~/.claude/projects/<slug> is a symlink
- [BUG] Running any version above 2.1.278 locks up on FreeBSD
- [GitHub integration]
- Epidemic Sound MCP connector: OAuth fails with "Client not found" — client_id set to user email instead of registered app ID
- Claude Code Repositoriesss
- [BUG] Uninstall/reinstall of Desktop (Windows MSIX) permanently orphans existing sessions — confirmed missing from claude.ai/code too, not just local cache
- [BUG] Model picker (Alt+P) is cut off in windows under 40 rows, so the effort line and key hints are hidden
- [Bug] Org-level connector tool controls not enforced in Claude Code CLI
- [BUG] /design-sync validator reports every component missing when the bundle contains non-ASCII (smoke page has no charset)
- [BUG] /design-sync story previews ignore package.json "browser": false, so builds fail on Node built-ins
- [Bug] Safety classifier false positives blocking legitimate cloud-security development workflows
- [Bug] Background agents lack resource warnings and orphaned processes prevent cleanup
- [BUG] Headless `claude -p` (entrypoint `sdk-cli`) uses ~1.8× more of the 5-hour window per token than interactive `cli`, same workload
- [GitHub integration]
- Default subagents spawn with ~60k tokens of inherited context; workflow subagents receive the user's latest chat line as their governing instruction
- [GitHub integration] after I connect GitHub to Claude, it's asking me to connect it again
- I appreciate you reaching out, but I need to clarify a few things: 1. I'm Claude Code, an agentic CLI for coding tasks - I'm not the right channel for account upgrades or policy discussions. 2. I can't generate a GitHub issue title for this because it's
- [GitHub integration] I can't integrate github to redeem the credits.
- `claude` exits 1 when run with no args and no stdin, which blocks winget releases
- [GitHub integration]
- [Bug] Drag and drop screenshots not working on macOS
- [Bug] Terminal flickering on Windows 11 PowerShell 5 with GPU rendering
- Critical: Claude ignored explicit 'no upload' instruction and purged GCS version history
- i can use it [GitHub integration]
- [GitHub integration]
- [GitHub integration]
- [GitHub integration]
- [BUG] Anthropic skills not showing on session first turn
- [Bug] Overly aggressive cybersecurity safeguard blocks legitimate educational security content in long-context sessions
- [Bug] Unnecessary reasoning_extraction trigger during normal operations
- [GitHub integration]
- [Bug] Claude blocking legitimate jailbreak mode automation requests
- Claude in Chrome: Italian Chamber of Commerce portals (dire.registroimprese.it, mypage.infocamere.it) blocked intermittently with "Navigation to this domain is not allowed"
- [Bug] Plugin orphaned after sync: cannot uninstall due to missing marketplace backing
- Synced plugin cowork-plugin-management can't be uninstalled and isn't listed on claude.ai
- [Bug] Model performance degradation during legitimate development work
- Credential-handling rule: sessions disagree on whether clicking Login on a password-manager-filled form is prohibited
- [GitHub integration] Reconnect the Claude GitHub App stuck
- spawn_task / dismiss_task: the documented replace order (spawn new, then dismiss old) double-dispatches when the old chip has already been started
- [Bug] Performance degradation across recent updates
- [BUG] Agent-team teammates inherit --settings but not --channels, so they load the channel plugin and steal its single poller (messages dropped)
- [Bug] Arrow keys switch plugin config tabs instead of cycling option values
- [Bug] VS Code extension 2.1.282: IDE sessions start with no core tools - only deferred tools load
- Branch switch in a Bash call prints full diffs of every file changed on disk
- [GitHub integration]
- [Bug] Model hallucinating CLI/SQL command output that isn't generated
- [GitHub integration]
- [GitHub integration]
- [GitHub integration]
- [BUG] Bug report drafted: Safety classifier stopped legitimate QA test-harness work, forcing a decline of the user's request
- [Bug] Claude cites data from dead websites (DNS_PROBE_FINISHED_NXDOMAIN) claiming to have read them
- PIECE IF SHIT CHAT BOT CLOSES EVERYT THREAD -- THIS IS WHAT I PAY FOR? I THOGUH CLAUDE CODE WAS ADVERTSED AS EXECTUION TOOL. THIS IS BEYOND ABUSE. GREAT FOR LEGAL FOLDER THANKS
- Claude in Chrome extension bridge fails to connect (list_connected_browsers returns empty)
- [GitHub integration]
- Opus 5.5: Severe scope creep and task focus regression compared to Opus 4.6
- [GitHub integration]
- [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
- Copy to clipboard preserving formatting in VS Code
- [BUG] Cowork (macOS): sandbox VM never releases host file handles in mounted folders. Blocks OneDrive "Free Up Space" and the handle count grows without limit
- [GitHub integration]
- [BUG] Model approves a PR that adds a new build/runtime dependency despite a loaded rule banning new dependencies
- [BUG] After a correction the model applies it piecemeal, requiring repeated escalations
- [BUG] Model answers a hard constraint with a long unrequested implementation design
- [BUG] Owed fixes and review findings are not tracked; durable work left only in an ephemeral temp directory
- [BUG] Model launches a dozen parallel subagents and exhausts the account usage limit
- [BUG] Background agents killed by the usage limit are reported as failed even when their report was already delivered
- [BUG] Long silent stretches during large tasks; harness repeatedly prompts for status
- [BUG] Model checks CI status despite an explicit, repeated user ban
- [BUG] Model runs measurements/tests without being asked under a zero-autonomy instruction
- [BUG] Model tries to change the user's established procedure and memory rules without being asked
- [Bug] Claude Desktop Windows MSIX update leaves orphaned container preventing launch
- [BUG] Model asks a clarifying question instead of inferring an obvious referent from the conversation
- [BUG] Desktop: in-app Browser tab keeps playing audio while its session is in the background, with no indicator of which session owns it
- [BUG] Read tool fails on source files containing extremely long lines
- [BUG] Model passes an unverified generalisation from memory to subagents as fact
- [GitHub integration]
- [BUG] IDE extension auto-install downgrades the extension (2.1.281 → 2.1.89) in Windsurf: --force --install-extension without a version
- Desktop Code tab message box: a list number of 100 or more loses its first digit ("107. " shows "07.")
- A URL that wraps onto a second terminal line opens truncated on click
- [FEATURE] suggest starting a new conversation when resuming a large/stale session
- [BUG] LOST ALL MY WORK AND PROJECTS IN APP WHEN MY PAYMENT DIDN'T GO THROUGH!
- [BUG] Linux: bashEditDiff shows a concurrent Claude session's edits, including ones made before the command started
- [BUG] Remote Control sessions batch auto-archived server-side days after disconnect; hidden from mobile with no unarchive or retention setting
- [Bug] Terminal cursor visibility inverted in PyCharm/WebStorm terminal
- [BUG] VS Code extension 2.1.282: blank webview over Remote-SSH, webview/index.js truncated in renderer (2.1.281 works)
- [BUG] claude-vscode.editor.openLast not found for version 2.1.282
- o prompt está sendo bloqueado após eu solicitar que opus...
- [BUG] Projects are not synced between apps
- [GitHub integration]
- [BUG] Model argues about the wrong line instead of searching for the string the user named
- [BUG] Model reports a fix as done without running the failing tests, and the fix was incomplete
- [BUG] Model gives git recovery advice without checking which commits and files are affected, costing the user time
- [GitHub integration]
- [GitHub integration]
- UNUSABLE PIECE OF SHIT GARABGE
- Project on the cloud env stuck when usage limit hit 100% while Cloud Credit still available
- [BUG] Model skips a mandated step (read the linked issue) in a written PR review procedure
- [BUG] Model presents a verdict as final while a mandated checklist step is still undone
- Ctrl+Enter to send a queued message opens iTerm2's context menu instead
- [GitHub integration]
- [Bug] LoRA loading stack not applying multiple LoRAs in workflow
- Feature request: focus view (hide tool output) for the default renderer — /focus requires the fullscreen renderer, which takes the mouse and scrollback
- [GitHub integration]
- [Desktop] "Open links in built-in browser" is ignored for some chat links: same URL opens in-app from one message, in system browser from another
- [GitHub integration] GitHub App reconnection not working
- [GitHub integration]
- claude plugin eval: every Bash call fails with "bwrap: loopback: Failed RTM_NEWADDR" (ref #14719, #55585)
- [Feature Request] Add allowlist for pre-approved scripts in settings.json
- MCP stdio initialize handshake no longer completes as of 2.1.282 (regression from 2.1.281) — tool calls still work, proactive server push breaks silently
- [GitHub integration]
- [GitHub integration]
- [BUG] Model checks CI status despite an explicit, repeated user ban
- [GitHub integration]
- [BUG] Model runs measurements/tests without being asked under a zero-autonomy instruction
- [BUG] Model tries to change the user's established procedure and memory rules without being asked
- [BUG] Model checks CI status despite an explicit, repeated user ban
- [BUG] Model delays executing an explicit order to prepare and confirm instead of acting
- [BUG] User had to repeat an explicit order three times before it was carried out
- [BUG] Model files duplicate GitHub issues because it did not check what was already filed
- [GitHub integration]
- [GitHub integration]
- [GitHub integration]
- [BUG] Desktop (Windows): "Add to prompt" quote cards grow without limit and cover the conversation
- les images ne peuvent pas être zoomées dans les fil de conversation. C'est hype…
- [BUG] MCP stdio server lists no tools since protocol 2026-07-28 is negotiated (tools/list times out, mcp-grafana)
- [GitHub integration]
- [Bug] Assistant repeatedly issues filler commands while waiting for background tasks despite explicit instruction
- [BUG] 長時間セッションで、必須のツール呼び出し(Telegram返信)と、明示済みの「決定事項を即記憶に保存する」指示の両方が、同じ時間帯に繰り返し抜けた
- [FEATURE] Add setting to suppress file reveal when extension applies edits
- [GitHub integration]
- [Bug] Overly restrictive file access safeguards blocking legitimate operations
- [BUG] Deferred tool resume: a later PreToolUse "defer" is ignored when the resume also sends a user message (the Agent SDK always does)
- [GitHub integration]
- [BUG] https://github.com/anthropics/claude-code/issues/87974#issuecomment-5774787317
- [yokyokyokGitHub integration]
- Video analysis precision lost since Sept 16 update
- [BUG] Desktop (macOS): Claude Projects threads on a local folder create hidden phantom sessions, and approvals and notifications send you there
- [BUG] [claude.ai Projects] Slash in knowledge item title creates folders in the knowledge panel
- [FEATURE] Add a way to browse back through previous /btw side-question answers
- Claude attributes other AI models work to itself since Sept 16 update
- [BUG] Sidebar empty after update to Opus 5.5 (Claude Code Desktop)
- Session name chip in statusline has no background color (plain text only)
- claude-fable-5 emitted a fake Russian-language 'system reminder' appended to its own reply
- [GitHub integration] Repos missing from list of repos available for cloud sessions
- [FEATURE] collapse long pastes into an attachment chip in the desktop app (Code tab)
- [Feature Request] Reduce false positives in security content filtering for CTF discussions
- [GitHub integration]
- [BUG] Subagent disallowedTools with a path rule (e.g. Write(dir/**)) removes the whole tool instead of denying that path
- [Feature Request] Improve output formatting with clear separation between CLI commands, tool results, and user-facing content
- Claude GitHub Integration — Unable to Connect Local SEO Second Brain Repository
- [BUG] Cowork Local projects now visible to all users sharing the same Pro account
- [GitHub integration]
- [Bug] Model downgrade required for SQL security analysis - Opus 5.5 stops unexpectedly
- Sharp increase in background API activity (~30x API-min:wall-min) in long-running web sessions + response-quality degradation, with week-over-week baseline
- [FEATURE] Declare local scripts and CLIs as first-class tools, without an MCP server (reopen of #31165)
- Hooks cannot obtain the session display name that ListAgents reports and SendMessage accepts
- [GitHub integration] i cant get the 100$ credit
- [BUG] Pro plan: 5-hour limit and paid extra usage consumed with no activity, attributed to "Others"
- [GitHub integration]
- Agent tool `isolation: "worktree"` cannot target an existing worktree; every isolated dispatch mints a fresh one
- Bash tool in a worktree-isolated agent refuses commands that never leave the worktree
- [BUG] Remote Control sessions not accessible from claude.ai/code without session-specific URL
- [GitHub integration]
- [GitHub integration]
- [BUG] VS Code: transcript >2 GiB causes infinite re-parse loop, extension host OOM kills all Claude tabs (2.1.282)
- [BUG] VS Code: post-compact skip ignored when boundaries carry preservedSegment — every tab parses the whole transcript, tab restore after host restart OOM-loops (2.1.282)
- [Guardrail] Cybersecurity fix for research project mentioned in the application form
- [BUG] New Claude sessions appear and claude CLI loses track of sessions when context limit reached
- [GitHub integration]
- [BUG] Linux: "import cookies" in the in-app browser wedges the main webview, which is killed and never respawned (permanent black window, app still running)
- [GitHub integration]
- [GitHub integration]
- [Bug] Session transcripts stored with file path instead of session ID on POSIX systems with symlinks
- [GitHub integration] on claude
- [Bug] Claude executes destructive commands without confirmation or safety checks
- [GitHub integration]
- [BUG] One claude CLI session backgrounds and/or gets confused with another session
- Desktop app tasks fail to delete ("could not be deleted, try again") and don't appear in the claude.ai/code web task list
- [GitHub integration]
- [GitHub integration]
- nemam issue
- [GitHub integration]
- [GitHub integration]
- repositório não aparece no seletor do Claude Code na web
- [MODEL] Fable
- GitHub integration
- O'Donnell Winter Comeback
- [FEATURE] Shell escape that doesn't feed output to LLM context
- [BUG] Model selection in gateway mode lags other providers
- [BUG] macOS desktop 2.9939.2: computer:// links to files/folders in a connected folder render as plain text (or get disabled after ~1s) — Finder can no longer be opened from the transcript
- [GitHub integration]
- [Bug] Sensitive information leaking despite configured stripping logic
- [BUG] Desktop Code tab usage chart miscounts tokens: duplicate transcript lines summed, workflow agents skipped, 7d legend shows all-time
- [BUG] Windows desktop app (Code tab): after "Main webview is unresponsive, will kill and reload" the UI never recovers; F5 does nothing, and only quitting (which stops all running Code sessions) fixes it
- Claude repeatedly acts autonomously after questions and ignores ESC/stop interrupts
- /doctor's CLAUDE.md checks skip AGENTS.md content
- [BUG] Session-specific scratchpad path early in the context stops cache reuse across sessions
- [GitHub integration]
- [BUG] With CLAUDE_CONFIG_DIR set, starting in $HOME records a project-scope install of every user plugin
- [FEATURE] One switch to turn off Claude Docs in Claude Code (today it arrives as a "Custom" connector + a separate skill)
- [BUG] Scheduled routines (CCR) hang indefinitely after set_permission_mode handshake, never execute task
- [FEATURE] User ability to choose the right context after compaction
- [GitHub integration]
- [BUG] Desktop: Chat/Cowork account plugin loads as @inline and overrides the same plugin installed from a same-named Claude Code marketplace
- [BUG] Windows: sessions disappear from VS Code Session history, cwd recorded as both c:\ and C:\ in one transcript
- [BUG] Background task completion notification can fail to arrive even though the task finished successfully
- [FEATURE] Auto mode classifier should be able to run its own read-only checks before approving, gated by a second auto-mode pass
- [BUG] Windows: claude --bg / agent view refuses a trusted folder when the trust key's casing differs from disk, and the suggested fix never shows a prompt
- [BUG] Design canvas: artboard PLAY button does nothing, and links between artboards don't navigate
- Agent SDK: a missing `cwd` is reported as "native binary … failed to launch (libc mismatch)"
- [Bug] Duplicate response display with inconsistent column widths
- [Bug] False Positive: Content Incorrectly Flagged
- [FEATURE] Let PostToolUseFailure hooks replace the error output the model sees (re-file of #77180)
- [GitHub integration]
- [GitHub integration]
- [Bug] Clipboard copy corrupts UTF-8 characters with incorrect encoding
- MCP OAuth client invents `<gateway-origin>/authorize` instead of resolving the RFC 9728 discovery chain (authorization_servers → AS metadata) — non-compliant with MCP auth spec
- [GitHub integration]
- [BUG] A compaction summary can refer to its own <analysis> block, which is stripped before the summary is used
- [BUG] /usage session Breakdown "Output" undercounts by ~20x — only counts plain text blocks, drops tool_use and thinking tokens
- [FEATURE] Desktop (macOS): folder picker always opens at ~ — hardcoded defaultPath ignores last-used folder; allow a default
- [GitHub integration]
- [BUG] VS Code /model fails with EPERM on Windows: Hebrew username is reversed in config path
- [Bug] TUI text rendering missing when scrolling to older messages
- [Bug] Rename command not synced to mobile sessions
- [GitHub integration]
- Code tab (desktop app) still applies the model/effort pick globally across projects — the behaviour fixed in #20745 for the CLI
- [BUG] Claude Desktop: Research mode can't use local (stdio) MCP servers, and nothing warns before the run starts
- [FEATURE] Desktop app: show Code sessions from all signed-in accounts on this machine
- Thank you
- [BUG] Bash tool children inherit the TUI's real pty; ssh passphrase prompt breaks fullscreen mouse tracking and wedges the session across resumes (regression in 2.1.281)
- Paul Crane
- [BUG] headersHelper silently ignoring env vars with SECRET in the name
- [BUG] Agent sessions are not persisted when canceled.
- [BUG] SendUserFile never delivers files from sessions a `claude remote-control` server spawns
- [GitHub integration]
- [Bug] Overly broad safeguard trigger on legitimate security instrumentation code
- spawn_task suggestion chips are labeled "running task" from the moment they're created, before the user has chosen to start them
- [MODEL] Model writes "falsifiable" when "verifiable" is the obviously intended word — reproduced 4x over 4 months, project-scoped correction doesn't transfer
- [GitHub integration]
- [BUG] cli OAuth error: Invalid code. Please make sure the full code was copied
- Cowork scheduled tasks: action classifier intermittently denies authorized browser writes to a logged-in CRM; same action succeeds when user is present; no trusted-site setting
- claude-in-chrome doesn't recognise connected browsers after switching Claude accounts, even after logout and clearing the cache
Resolved issues (64)
- [FEATURE] Usage keeps getting worse
- [Feature Request] Improve copywriting capabilities and vocabulary consistency
- [BUG] Claude Desktop taskbar icon blurry on Windows 11 at high DPI scaling
- orion-strom
- [Bug] Worktree isolation blocks all Bash commands due to parser abort in bash-command classifier
- [BUG] Desktop app (Code tab): "Auto mode is on" notice keeps showing while the session is in Manual mode
- [Bug] Agent deviates from user plan and exhausts API quota unexpectedly
- [Bug] Content classifier false positive on metaphorical language as biosafety risk
- [Bug] Agent repeatedly ignores user instructions and sends out subagents building "things" the user didnt ask for after the user explicitly told the agent its idea was bad and not approved. it sent out a build anyways.
- [Bug] Incorrect reasoning extraction in analysis
- I don't have a bug report to analyze. Please provide the bug report details, error message, or description of the issue you'd like me to create a GitHub issue title for.
- [BUG] VSCode: background agent output streams into foreground chat, disrupting active conversation
- Sandbox wedges on merged-usr Linux: bwrap "Can't mount tmpfs on /newroot/bin" (regression, works on 2.1.229)
- [Bug] Account Access Issue: User Blocked Without Cause
- [Bug] Anthropic API Error: 400 Output blocked by content filtering policy
- [BUG] Fable 5 quality regression: ignores explicit instructions after acknowledging them, increasing hallucinations
- [BUG] Read tool falsely reports unencrypted pandoc/LaTeX PDFs as 'password-protected'
- Auto-backgrounded/interrupted Bash commands outlive their turn and can destructively race later commands
- Model fabricates <task-notification> blocks (plus its own replies) inside a single assistant message during long background-task waits
- [BUG] Desktop app shows 'GitHub CLI authentication expired' banner while gh credentials are valid (Keychain storage)
- [BUG] Workspace trust dialog never appears in the VS Code extension (shown via CLI), so project-scope plugins / project .claude/settings.json are silently skipped
- [BUG] Platform reported as "darwin" / Shell "zsh" under Git Bash (MSYS) on Windows 11
- VS Code extension: AskUserQuestion dialog covers the assistant message it refers to, with no way to read the text underneath
- [Bug] One transient advisor failure permanently latches the tool off for the whole session; distinct error causes flattened to generic "unavailable"
- [BUG] Desktop usage widget mixes two token units: cached days count cache tokens, live days don't, so models first used after lastComputedDate vanish from the Models tab
- [GitHub integration]
- [Bug] Overly aggressive cybersecurity safeguard blocks legitimate educational security content in long-context sessions
- [GitHub integration]
- Synced plugin cowork-plugin-management can't be uninstalled and isn't listed on claude.ai
- [Bug] Claude cites data from dead websites (DNS_PROBE_FINISHED_NXDOMAIN) claiming to have read them
- Remote Control: /remote-control gives no confirmation on enable, sessions never appear on claude.ai/code or mobile
- [BUG] Windows Desktop: device bridge handshake never acknowledged for ~18 min (connect frame sent → handshake timeout), survives reboot + app update, then self-recovers
- [BUG] Windows Desktop: linked computer stuck on "Asleep or app closed", cloud sessions and cloud scheduled tasks fail with "not connected to the bridge" (survives reboot, re-login, device removal, app update)
- [BUG] macOS Desktop App Thinks Computer is Asleep While Using Desktop
- [BUG] "$250 credit for cloud sessions" gift icon is unclickable
- MCP/Bash outbound TCP to LAN IPs fails (EHOSTUNREACH) while built-in browser tool reaches the same host fine
- [BUG] VS Code extension: renamed sessions revert to the auto-generated title
- [BUG] claude-vscode.editor.openLast not found for version 2.1.282
- [BUG] LOST ALL MY WORK AND PROJECTS IN APP WHEN MY PAYMENT DIDN'T GO THROUGH!
- Cross-session messaging is one-way: receiving session has no SendMessage tool, but ListAgents and inbound system-reminders instruct it to use one
- Symlinked files in .claude/rules/ are not auto-loaded (contradicts docs)
- Suggested task ("task chip") intermittently fails to start, with no surfaced error and no log
- [GitHub integration]
- nemam issue
- [BUG] sparsePaths git marketplaces clone empty since 2.1.274 (GIT_NO_LAZY_FETCH=1 on a blob:none clone; git ≤2.34 exits 0) and silently replace healthy clones — 2.1.278 retry never fires
- [BUG] Plugin `bin/` entries are added to PATH with no existence check — 74% of PATH becomes dead dirs, pushing the inlined shell snapshot past the Windows 8191-char limit
- [Bug] Claude Code downgrades to Opus instead of maintaining requested model
- [Bug] Message flagged by safety system without clear reason during non-security tasks
- [Bug] Inconsistent performance with extreme latency spikes (5+ hours)
- I don't have a bug report to analyze. Could you please provide the bug report details that you'd like me to create a GitHub issue title for?
- [BUG] "Läuft einem Befehl" is wrong in German use "Führt einen Befehl aus"
- [BUG] Desktop app (macOS): Cmd+V starts a new session instead of pasting with German Neo 2 keyboard layout
- [BUG] `disableAutoMode` added mid-session removes auto from the mode rotation but does not eject a session already running in auto mode
- [BUG] Cowork EXDEV: cross-device link not permitted — rename within same directory (Windows 11 Pro, OneDrive present)
- [BUG] claude-in-chrome: tabs_context_mcp({createIfEmpty: true}) places new tab in an existing unrelated window instead of creating an isolated one
- [BUG] Cowork local-agent-mode sessions missing from Recents despite intact data on disk (Windows)
- [BUG] Skill with paths frontmatter is completely undiscoverable
- [BUG] False positive content filter blocks standard e-commerce geoblocking implementation (state-level sales compliance)
- [BUG] Invoking a skill (slash command) pins a large injected block and shrinks the assistant response into a tiny scroll area (VS Code extension)
- [BUG] VS Code plan-review comments are not forwarded to Claude Code
- Permission prompt: key '1' = Approve in terminal CLI but Deny in Windows desktop app — muscle memory causes accidental denials
- [Bug] Arrow keys switch plugin config tabs instead of cycling option values
- [GitHub integration]
- [BUG] MCP progress notifications are received but discarded once a call is auto-backgrounded at 120s