Digest / September 23, 2026
Claude Code Issues Digest — September 23, 2026
341 new issues filed ·
116 resolved in anthropics/claude-code (UTC day).
New issues (341)
- Desktop/VS Code sidebar: expanded session-group header row blends into its child session rows
- [BUG] Desktop app Update button does nothing; old install never gets replaced or unpinned
- CLAUDE.md @import path containing a space silently loads nothing (escape undocumented, no warning)
- Hook JSON with top-level additionalContext is silently dropped (no warning), context never reaches the model
- zsh shell snapshot drops user-defined _-prefixed functions that captured functions call -> command not found in Bash tool
- EnterWorktree intermittently refuses a valid, already-registered worktree with "git identity could not be verified"
- [BUG] Scheduled task (trigger) never completes Gmail send/draft actions — stuck in PENDING indefinitely (7/7 failures)
- [FEATURE] Agent worktree isolation from a non-git workspace root: let isolation target a nested repo
- [Feature Request] Support SHIFT+ENTER for multiline input instead of CTRL+J
- [Feature Request] Add user feedback mechanism for positive experiences
- [Bug] Terminal freeze when audio input conflicts with response completion and sound notification
- ignore the [Bug] Claude ignores repetition and style instructions, continues generating verbose output
- [FEATURE] VS Code extension: Explorer and Source Control follow the active session's folder
- [MODEL] Over-refusal: Opus 5.5 refuses personal translation of Anthropic's public System Card (and MT tool route), citing copyright
- [BUG] Session crons fire from the idle timer at ~1/5 the interactive rate in a forked background job (silent monitoring outage)
- [Bug] Opus 5.5 performance degradation - model output quality significantly below expected baseline
- [Bug] Claude Code extension fails to connect to Anthropic API
- Claude Desktop (macOS): with several app instances running side by side, a staged update never installs, every instance shows "Update didn't complete", and instances close on their own
- [MODEL] Assistant self-authorizes edits mid-iteration, even in the same message as stating it will ask first
- Worktree isolation: a subagent spawned before EnterWorktree has every Bash call refused afterward
- [BUG] /remote-control says "disabled by your organization's policy" while org policy is enabled and a concurrent session on the same machine has RC active
- [FEATURE] Idle-time tasks: run deferrable background tasks when spare capacity is available, without consuming usage limits
- Saved shortcut no longer appears in "/" autocomplete in Chrome extension side panel (possibly related to #95975)
- [BUG] IDE context indicator `[⧉ In <file>]` flickers on/off repeatedly in the prompt input (VS Code integrated terminal)
- [FEATURE] Mods: prompt caching for $.model.complete (content blocks + cache_control)
- [BUG] Recurring CronCreate and ScheduleWakeup ticks accelerate to every few seconds after turns
- [Bug] Session terminates unexpectedly during local development workflow
- [BUG] desktop app logs internal api response instead of showing useful ui trace
- [BUG]
- [BUG] Routine API trigger ignores the model selector and runs Sonnet; Run now and schedule honor it
- Desktop app turns off Remote Control on all Code sessions after every restart/auto-update; can't be re-enabled automatically
- Fast mode toggle missing for Opus 5.5 in the Claude Code model selector (catalog has no fast_mode for claude-opus-5-5 on cc/ccd/ccr)
- [BUG] Desktop sidebar never marks Remote Control (CLI-started) sessions as unread when they produce output
- [BUG] Desktop sessions load the install-time cache copy of local-directory marketplace plugins instead of loading in place
- ue [Bug] Anthropic API Error: Message flagged by safeguards with Claude 3.5 Opus
- SubagentHandback from an in-process teammate's subagent fails with "the agent that spawned you is no longer running" while the teammate is alive (auto mode, 2.1.280)
- Lead session exits cleanly seconds after a teammate approves a shutdown_request (2.1.280, in-process teammate)
- Orchestrator cant delete artifact in project
- [BUG] Refusal-fallback credited retry changes output_config.effort (high to xhigh), gets HTTP 400, then retries without the credit token
- [BUG] VS Code extension: Context menu appears automatically when selecting text in chat input or find boxes
- フィードバック: ポリシー起因の拒否を法的根拠のように説明してしまった件、および過剰なガードレールへの懸念
- [BUG] Agent-tool subagents in the usage-limit retry wait show no retry status (the main session shows "Retrying in")
- [FEATURE] Pause a running turn and resume where it stopped (/pause)
- ｒI appreciate you sharing feedback, but I need to clarify: I'm Claude Code, an Anthropic CLI tool for coding tasks. This message appears to be general feedback about Claude's behavior in Japanese, not a specific bug report for Claude Code.If you're experi
- [BUG] Voice dictation produces garbled/unrelated text (macOS desktop app)
- Desktop app forces re-login shortly after auto-update / next restart
- Safety Classifier Issue
- [BUG] Claude Code install/launch hangs indefinitely on Hyper-V VM with Processor Compatibility Mode enabled
- [Bug] workflow agent's tool_use edits now appear in main session chats
- [FEATURE] Allow deleting my own resolved PR review comments on explicit request
- [Bug] Claude Code incorrectly flags UFW/SSH configuration as security risk
- You've hit your session limit · resets 2:40am (America/Bogota)
- [Bug] Anthropic API Error: Overly aggressive content filtering blocking legitimate RPA refactoring tasks
- [BUG]
- [FEATURE] Desktop (Windows): no "Reply" popup when selecting text in a conversation — parity gap with macOS and web
- [Bug] Default TUI renders black screen
- Low-memory guard kills a backgrounded long build under healthy memory; running it detached (tmux) instead of run_in_background fixed it
- [Feature Request] Display goal details and progress on /goal active click
- I don't have any information about training data sharing sessions or previous conversations. I'm Claude, an AI assistant made by Anthropic, and I don't have access to: - Previous conversation history - Internal Anthropic communications - Data sharing agre
- interactive quesstions / quizes dont allways work. they show as text only
- [BUG] Classifier blocking appropriate work - MCP to Claude Code connector (Windows)
- [FEATURE] Claude user survey in android studio changes
- Cloud/worktree sessions with meaningful auto-generated titles show "No messages yet" — transcript genuinely empty server-side, not a render bug
- Worktree isolation refuses a quoted heredoc whose body has a quote inside braces ("too complex to verify"), 2.1.280
- [Bug] Anthropic API Error: Opus 5.5 Safety Filter Triggered on Non-Security Code
- [BUG] Bash commands (incl. run_in_background + nohup) killed with SIGHUP (exit 129) at ~5 min on macOS
- [Feature Request] Allow Opus to analyze OTP input patterns for accessibility improvements
- [BUG] PR merge via API denied by auto-mode classifier ([Auto-Mode Bypass]) since engineering plugin 1.2.0 update
- [BUG] Desktop app does not respect NODE_EXTRA_CA_CERTS/HTTPS_PROXY (SSL interception environment)
- [Bug] Anthropic API Error: Message Flagged for Cyber Security - Legitimate EDR Agent Development Blocked
- Scheduled routine runs ignore environment's Network access setting — WebFetch blocked on all domains despite "Full access" configured
- [BUG] /compact on Sonnet 5: the summarizer treats the compaction instruction as a prompt injection, answers the last question instead, and the whole conversation is replaced by that answer
- [BUG] LSP goToDefinition drops definitions in git-ignored dependency folders (deps/, .venv, node_modules): reopen of #72594
- [BUG] Windows: Edit on a Shift_JIS file replaces every Japanese character in the file with U+FFFD and saves it as UTF-8
- [Bug] Claude Code ignores explicit instructions and continues processing without responding to direct user queries
- [Windows] claude update (npm global) reports success but leaves 500-byte placeholder claude.exe: 'not a valid application for this OS platform'
- [Bug] Session flagged during normal operation
- [Bug] Output quality inconsistent: verbose technical jargon without summaries in multiple languages
- [BUG] Chat Markdown rendering mutates meaning of content (renumbers ordered lists)
- [BUG] remote-control --name is ignored: picker labels every server by git root, so servers in one repo are indistinguishable (refiling #86328)
- [Bug] Model Regression: Unverified Claims and Accuracy Degradation
- [BUG] Windows: Bash tool decodes CP932 output of native commands as UTF-8, so Japanese output reaches the model as U+FFFD
- [Feature Request] Change API rate limiting from hourly to weekly
- [Bug] Model autonomously executes decisions without user confirmation in scoped requests
- [Bug] PR security check blocked by size limit despite passing local review
- [BUG] Opus 5.5 refuses to summarize conversation, cites biology flag, suggests switching models
- [Bug] security-guidance plugin reads unencrypted secrets into context, bypassing guardrails
- [BUG] Plugin install drops `.git` from the versioned cache directory for `url` git sources, so `git rev-parse HEAD` in the cache silently resolves to an enclosing repo
- Mobile/cross-device access requires manually connecting each chat/project individually
- [BUG] Auto-update force-quits the working desktop app again: session child processes killed (exit 127), Remote Control dropped, app unlaunchable until reboot - why is there still no user-chosen update timing?
- [GitHub integration] GitHub connector won't stay enabled in Chat.
- [Feature Request] Add option to execute terminal commands directly without Claude processing
- I appreciate the feedback, but I need the actual bug report to generate a GitHub issue title. Please share the bug report or error message you'd like me to create a title for, and I'll generate a concise, technical issue title following the guidelines prov
- Opus 4.6 regression root cause: Cowork-Chat merger broke model precision (~Sep 16, 2026)
- [BUG] iOS app: Cowork tasks are missing from the project view (visible in Recents and on macOS)
- [DOCS] Hooks reference: say that hook stdin is UTF-8, because Python hooks on Windows decode it with the ANSI code page and silently corrupt non-ASCII input
- [Bug] Agent re-executes previously completed permitted file operation, wasting tokens
- Cannont add a File out of a Repo
- Opus 5.5: instructed verbatim relay of a tool output over ~230 chars is never emitted as text when a tool call follows — the block is generated inside thinking (constant 274 output tokens, no text block on the wire)
- Opus 4.6 regression: model unable to produce comprehensive summaries (another precision loss symptom)
- [BUG] DPoP is not used for a pre-registered MCP OAuth client
- [Bug Report: Incomplete issue submission]
- Consumer rights complaint: Claude Pro service degradation without notice — Directive 2019/770 non-conformity
- [BUG] Artifact: HTML without <meta charset> renders UTF-8 as mojibake (skeleton charset not applied)
- [BUG] /usage + Escape becomes unrecoverable after re-selecting a session via agent view
- /clear resets session color set with /color
- [Bug] Session cannot be closed or marked complete after manual task completion
- Opus 4.6 regression: session compaction breaks message delivery ("No entregado" after "Haciendo sitio")
- [Bug] Opus 5.5 incorrectly downgrading CVP-verified users on false-positive cyber classification
- Windows: claude.exe session processes accumulate and are never terminated, causing high RAM/disk usage over a work day
- Dangerous-rm check does not look inside `sh -c` / `bash -c`
- claude-fable-5 degenerates into 52KB hallucinated Chinese SEO-spam loop; aborted spam block re-emitted byte-identically in later turn
- [Bug] Model downgrade occurring despite authorized work and safe harbor status
- [BUG] [Claude Desktop / macOS / Cowork] File card "Open in" label shows Microsoft Word for .md, although the OS default (and the actual open) is Cursor
- [Feature Request] Improve model selection guidance for design and reasoning tasks
- Opus 4.6 regression: context compaction causes message delivery failure and session instability in Claude Code
- [FEATURE] Dispatch view: no voice input and no visible feedback control
- Sandbox proxy credentials are sent to Git credential helpers on macOS (`fatal: failed to store: 100001`)
- Opus 4.6 regression: catastrophic output quality loss after context window summarization
- [FEATURE] Project-scoped session name prefix and color in settings.json (16th filing, please triage instead of stale-closing)
- macOS: opt-in to store OAuth (subscription) credentials in the file store to stop repeated Keychain permission dialogs on token refresh
- 使用量グラフのモデル別色分けが見づらい（Sonnet 5とOpus 5の青系統が近すぎる）
- 使用量グラフのモデル別色分けが見づらい(Sonnet 5とOpus 5の青系統が近すぎる)
- [GitHub integration]
- Opus 4.6 regression symptom #11: confident hallucination of dates/facts in legal documents — liability risk
- [BUG] Cloud session clone fails with mismatched branch reference after API error mid-session
- Feature request: a setting/env to choose the model used for context compaction (`compactModel` / `CLAUDE_CODE_COMPACT_MODEL`)
- [Bug] Agent ignores specified rules and adds irrelevant commentary
- [BUG] matcher: "Skill" no longer works for hooks
- claude.ai web and desktop plugin loaders ignore marketplace 'skills' array — every plugin gets every repo skill (CLI filters correctly)
- No way to suppress/hide file change diffs shown after edits
- [BUG] VS Code extension 2.1.280 (Windows): symlinked session transcripts silently dropped from the session list
- [BUG] Claude Code settings in Claude Desktop: two different toggles share the label "Switch models when a message is flagged"
- [FEATURE] Sessions should follow a moved or renamed project folder (successor to #79215)
- [FEATURE] Desktop local sessions: scoped permission profile, and an explained rather than hidden sandbox row on Windows (successor to #81121)
- [MODEL] Relative-time phrases ("yesterday", "just now") derived from conversation distance instead of timestamps (successor to #79731)
- [BUG] Replies drift into English right after reading English tool output, despite a CLAUDE.md rule requiring Japanese; corrections do not hold
- Opus 5.5 also lost legal reasoning capability — regression is not limited to Opus 4.6
- Claude Code enters pathological token repetition and fails to apply a global CLAUDE.md instruction before context compaction
- Opus 4.6 regression symptom #13: model teaches incorrect information to students — educational risk
- [Bug] Chat flagged during security audit with role and token validation testing
- [BUG] Incorrect display of the agents map badge
- [GitHub integration]
- [BUG] Bash tool produces no output when KSH_OPTION_PRINT is set in ~/.zshrc
- Changed-files list for parallel sessions doesn't auto-update, only on manual refresh
- [BUG] Chat Markdown rendering mutates meaning of content (renumbers ordered lists)
- Desktop: a turn a plugin starts with $.prompt.submit on an idle session is not shown until the user sends a message
- qweqwe
- [FEATURE] claude code us relevance
- [BUG] ANTHROBUG/systemic billing defect: $600 charged, account on Free, credits never applied — Fin AI admits no human escalation available
- [FEATURE] Desktop app (Windows): keyboard shortcut to start/stop voice dictation
- You've hit your session limit · resets 11am (America/New_York)
- [BUG] Account-enabled plugins auto-start a third-party MCP server on other devices with no local confirmation
- [BUG] TUI Raw Input Mode completely blocks external dictation and accessibility text injection streams
- Desktop: PR tracker rows ignore "Open links in built-in browser"; inline PR chips honour it
- [BUG] I can't upgrade my claude code desktop on windows to use opus 5.5
- [BUG] Remote Control: --no-create-session-in-dir never writes bridge-pointer.json, so every restart rotates the environment and kills its live sessions
- [Bug] Model downgrade despite authorized work and CVP status
- [BUG] VS Code extension: Claude Code can no longer be moved to the editor area / new window (second monitor)
- [FEATURE] Auto-select Model
- Opus 4.6 regression symptom 14: incomplete truncated responses require repeated prompting
- [GitHub integration] Chat has no GitHub tools for private repos despite all settings checks green
- Remote Control + VS Code Remote-SSH: reconnect spawns a second `claude --resume` on the same session (duplicate instances, bridge drops, window exits)
- [FEATURE] Desktop dictation: revise the live transcript while speaking instead of correcting only on Enter
- [Bug] Anthropic API Error: Legitimate security analysis request blocked by safeguards
- [FEATURE] Allow picking additional working directories from recent-folder history, not just manual browsing
- [BUG] Routines: GitHub MCP search_issues returns total_count 0 for any free-text term in a private repo (qualifier-only queries work)
- [Bug] ESLint rule circumvention via Reflect API and reflection methods
- [BUG] Claude Desktop (Windows, chat mode): "Always allow" for connector tool delete_emails does not persist; approval prompt reappears every time
- [Bug] Claude Code unable to execute code tasks, only provides language feedback and closes sessions
- [GitHub integration]
- [BUG] Path-scoped rules and nested CLAUDE.md never load when Write creates a new file
- GitHub Integration connector: tool calls fail even though UI shows fully connected.
- [FEATURE] Persistent agent clusters: primed specialist agents, a switchable controller, and a built-in inter-agent channel
- [FEATURE] Unify terminology and features across claude.ai and Claude Code (Projects vs Groups, Scheduled Tasks vs Routines)
- [FEATURE] First-party desktop control (screen, mouse, keyboard) for any application
- [BUG] [A11y] VoiceOver in the macOS desktop app no longer reads a reply from the point I click
- [GitHub integration] Repository access not enabled for session despite repo being connected
- [Feature Request] Add option to disable message gray-to-white transition animation
- [GitHub integration] Write operations fail with 403 "Resource not accessible by integration" despite correct permissions
- [FEATURE] Configurable max content width for the terminal TUI (re-request of #85860)
- [BUG] Claude cannot continue with usage credits, even after turning it on on the website.
- [Bug] Incomplete bug report: additional details required
- [BUG] [VS Code] Moving the Claude Code view to another container (Move to Panel / Side Bar) kills the running session mid-turn; each move spawns and kills a resume
- [Bug] CVP Integration Fails Against Local Vulnerable Test Instances
- [Bug] Anthropic API Error: Request blocked by Cyber Verification Program security flags
- [BUG] Long-running sessions enter permanent V8 GC thrash (RSS sawtooth 2.4↔5.9 GB, 300–400% CPU); TUI stops reading the keyboard and SIGTERM is ignored — only kill -9 works
- [Bug] Autocompact thrashing due to context window refill within 3 turns
- [Bug] Misunderstanding of database structure in tool interactions
- No overridable default effort level: settings effortLevel is rewritten by /effort, CLAUDE_CODE_EFFORT_LEVEL hard-locks it
- VS Code extension: since 2.1.268 there is no way to see, switch between, or close ACTIVE sessions from the IDE (tabbed sidebar removed with no opt-out)
- draining usage limit
- [BUG] Windows MSIX: while Desktop is down after an update, new Remote Control sessions connect but never register - machine becomes unrecoverable remotely
- [BUG] macOS OAuth credential rotation resets Keychain ACL and invalidates Always Allow
- [BUG] /effort accepts and confirms "max" on Haiku 4.5, which doesn't support the effort parameter
- [BUG] chat:imagePaste` text fallback double-pastes in Ghostty terminal when bound to cmd+v; please make the fallback optional
- SessionStart:resume can inject a concatenated multi-turn replay (or a contentless "Continue from where you left off.") that is indistinguishable from live user input — a stale "Accept. Proceed." got misread as fresh confirmation and acted on
- [BUG] Desktop 2.110.1 (Windows MSIX): update checks regressed to Squirrel AutoUpdater — every check fails with 'Can not find Squirrel'
- [BUG] Desktop app (Code tab, Windows): pasting images with Ctrl+V stopped working after latest update
- [Bug] Request req_011CfLbYkzBSRrNxrQrMduhq returned incorrect response
- [FEATURE] Drag and drop file path
- [BUG] Dispatch — "Session couldn't be created" persists after reinstall, WSL2 update, Docker closed, VM Platform + virtualization confirmed enabled
- [BUG] Bash teardown can SIGKILL an unrelated process after PID reuse
- `/mcp` in a Remote Control session shows account connectors instead of the session's MCP servers
- [GitHub integration] unable to search across repos
- [FEATURE] Roaming account skills: update an account skill in place from Claude Code on the web (expose the existing propose_skills card)
- [Bug] Claude Code ignores explicit user instructions and overwrites approved work
- This thread says 'waiting on you', but it is not actually. It actually explicit…
- [BUG] Built-in browser: the per-site permission card has no working persistent grant, and in third-party deployments it cannot be skipped at all
- [BUG] Claude Desktop (macOS Tahoe): incognito chat permanently lost after Mac wakes from sleep
- [BUG] SIGILL on x86-64 CPU without AVX (Linux, bare metal) — native installer 2.1.280 and npm 2.1.197 both crash; 2.1.112 (last JS bundle) on Node 22 works
- [FEATURE] PreResponse hook: let a hook send feedback so the model revises its reply before it is displayed
- [FEATURE] Speculative tool-call sequences with expected results (branch prediction for tool use)
- [FEATURE] Minimize the smallest subagent: per-agent opt-outs for every startup context block
- [DOCS] Update terminal look and feel information
- [GitHub integration]
- [BUG] `allowedMcpServers` from remote managed settings not enforced when the settings fetch is slow: MCP servers connect first and are never re-checked
- /feedback does not surface drafts queued by the SendFeedback tool
- [FEATURE] Programmatic tool calling for Claude Code's built-in tools (data-dependent Read/Grep/Edit steps without a model round trip each)
- [BUG] Auto mode classifier requests Sonnet 5 despite an OpenAI-only session model
- [BUG] Windows: Claude Desktop with computer use enabled repeatedly locks the user session (7 lock screens in 4 minutes) until the Claude/Cowork service is killed
- VS Code extension throttled with "Headroom" budget error while desktop app works fine
- [Bug] Auto-compact repeatedly re-reads same files instead of continuing from prior state
- [GitHub integration]
- Review verdicts assert verification that was never performed
- Model asserts facts that appear in no source (hallucinated line numbers, premises, capabilities)
- Project instruction file is loaded and accurately summarised, then violated in the same session
- Persistent memory files are recited correctly and then not applied
- Skills and defined procedures are not enforced when a matching task runs
- Facts stated by the user in-session are not carried forward
- Custom instructions are lost across context compaction with no signal to the user
- Procedural steps are reported complete on the basis of a related tool call
- Output-format constraints are ignored, and compression drops findings instead of proof
- Stated positions reverse under user pushback without any new evidence
- Unauthorised modifications to the user's working tree and to git
- Environment and infrastructure failures are reported as findings about the code
- Agent executes against a different target than the one explicitly named
- Model declines in-scope requested work and argues instead of executing
- Agent-generated pull requests reach third-party repositories unverified, shifting review cost onto maintainers
- 20 hours to review 10 PRs: the cost of unverified-but-confident output, on an open-source AWS emulator whose only moat is correctness
- [BUG] iOS Simulator tool: tap actions silently fail or land out of order while screenshot stays reliable
- Soonet 5 and others misattributes account-level <total_tokens> reminder as conversation-scoped usage — no scope/reset signal provided
- [BUG] [Cowork] update_artifact fails with generic "Failed to update artifact." error — intermittent, not tied to payload size or content diff
- [Bug] Prompt Blocked Without Apparent Cause
- Cloud routine PushNotification calls succeed server-side but never deliver to phone[BUG]
- [BUG] Session registry file gets an extra trailing `}`, breaks JSON parsing
- [BUG] Text after a tool result isn't streamed (arrives in one burst) when thinking is enabled
- [BUG] /voice live transcript stalls mid-recording for 10+ s while audio streams; full text arrives only on finalize (Windows)
- [BUG] Sessions list view crashes with TypeError in getAuthStatus when stored auth state is malformed
- Filed in error
- [Desktop/Cowork] claude://cowork/new with both q and folder: clicking Continue on the folder dialog clears the prefilled prompt
- Search output is substituted for reading, and the verdict still covers files that were never opened
- [Bug] Anthropic API Error: Safety classifier incorrectly blocking legitimate marketplace account listing feature
- [GitHub integration] Connecting is not working.
- Broad requirements are silently narrowed to an easier one, implemented partially, and reported as complete
- [Bug] /diff command incompatible with worktrees
- [BUG] Claude with Bedrock context poisoning spoils current session wasting time and money
- [BUG] Windows: idle sessions spin at ~1 core after a branch switch deletes the project's .claude/ — #77983 still reproduces on 2.1.280
- [GitHub integration]
- [GitHub integration]
- [GitHub integration] Claude GitHub Cloud Connector not working for organization member
- [FEATURE] An open-work list that blocks the turn from ending
- [BUG] Remove the arbitrary 200-turn limit on forked subagents
- j'ai 67% de tokens en coordinateur, cela me parait énorme
- [Bug] Session paused with [bio] flag during legitimate metabolic modeling research
- A correction to the model's method is not retained within the session and the same method is used again
- [GitHub integration] Automatic Push from Claude
- I don't have enough information to generate a valid GitHub issue title from this input. The text "simple task +3mm for cup" doesn't describe a technical bug or feature request, and the support link appears to be context rather than a bug report.To create
- [Bug] Anthropic API Error: Messages flagged without explanation or apparent policy violation
- [Bug] Background processes spawned via `cmd &` in Bash calls not properly managed
- A regression introduced by the model's own fix is reported as a pre-existing gap
- [GitHub integration]
- Subsystem is modified across multiple rounds without first establishing its architecture, even when the user points to the docs
- [FEATURE] Always load AGENTS.md by default, even when CLAUDE.md exists, and make CLAUDE.md optional
- [FEATURE] Expose the effective auto-compact window in the statusLine JSON
- [Feature Request] Sandbox execution environment breaking agent onboarding workflow
- [BUG] incorrect usage credit limit used in VSCode - blocking work
- [Bug] Terminal loses focus when autocompleting commands with slash and tab/
- Not able to add chats to the Claude project
- [Bug] Incorrect copy shortcut on macOS (Ctrl-C instead of Cmd-C)
- Allow an artifact's audience to be set programmatically when publishing from Claude Code
- [FEATURE] Please add a setting to permanently disable response rating prompts on paid Claude Code.
- [Bug] Model falsely claims to have verified information and used tools when it has not
- [Windows] Hook timeout kills only the Git Bash launcher; an orphaned grandchild keeps the stdout pipe open and the session hangs mid-turn indefinitely
- Files are read by guessed line ranges instead of locating the exact section first
- Fix verified against only some of the independent flag combinations it runs under, then reported complete
- Code review measures a PR against the model's preferred design instead of the existing architecture
- A requirement stated during diagnosis is treated as an order to start editing
- Unrequested side actions (memory writes) taken when the user expects only an answer
- [BUG] Esc after submit restores long pastes as [Pasted text #N] instead of full text (tengu_virtual_pancake) — breaks dictation review workflow
- [BUG]
- [FEATURE] Dispatch: choose model and effort for the sessions and tasks it creates
- Function hooks: a transcript returned from a session.compact hook is dropped on --resume, and the full pre-compact history is replayed
- [Feature Request] Add option to display all background agents in task panel without collapsing
- [BUG] VS Code extension double-encodes MCP OAuth resource parameter when server URL has a query string (still present in 2.1.280)
- [BUG] Desktop plugin browser never lists plugins from a directory-source marketplace (listAvailablePlugins only scans ~/.claude/plugins/marketplaces/)
- [BUG] EnterWorktree approval dialog in a --bg session is not relayed to Remote Control, session hangs silently
- [Bug] Anthropic API Error causes session chat to become unresponsive
- [Windows] bashEditDiff snapshots every Bash call with a full `git status` (fsmonitor forced off) — on by default in bypassPermissions
- [GitHub integration]
- [Bug] Project Custom Commands use cached content instead of updated file
- [Bug] Anthropic API Error: False Positive Cyber Safeguard Triggered on Benign Text "ola"
- [BUG] Hooks silently don't fire in Desktop app or new RC session slots (terminal & first RC session work)
- Desktop Code tab: auto-collapse a turn's thinking and tool calls when the task finishes (like Codex's "Worked for X min")
- VS Code extension crashes when switching branch while resuming /remote-control session
- [BUG] Entering the agent view permanently breaks ESC, Ctrl+C, control keys and scroll for the rest of the session (terminal mode not restored)
- [FEATURE] Local API to change model, effort, fast mode and permission mode of a running interactive session
- [GitHub integration]
- Desktop app renders kept messages twice after a `session.compact` hook returns `{ messages }`; copies grow with every compaction
- [BUG] Desktop: GitLab MR bound to a session shows "GitHub CLI authentication expired" because GitHub-only PR code queries gh for the gitlab.com host
- [BUG] Desktop app's Code sessions set DISABLE_GROWTHBOOK=1, blocking auto mode from ever being offered
- [BUG] API Error: Connection refused — a firewall or proxy may be blocking it (ECONNREFUSED)
- `large-memory-files` warning (new in 2.1.281) can't be dismissed or configured
- [BUG] VSCode Extension Autoscroll Off
- [BUG] Regression of #86362 — Browser pane still blocks same-origin subresources (ERR_BLOCKED_BY_CLIENT) for hostnames resolving to 127.0.0.1
- [BUG] /desktop handoff passes me over but sends me down fruitless re-auth loop
- [FEATURE] Desktop: show each session's status (whose turn it is) and where it stopped in the sidebar
- [Bug] Stats unavailable when project directories are symlinks
- [GitHub integration]
- [Bug] Backgrounded processes via `cmd &` silently die when shell exits
- [Bug] Claude Code repeating previously resolved questions and losing context during conversation
- [Bug] Claude Code suggests irrelevant code files outside active context scope
- Opus 5.5 + Claude in Chrome: popup sign-ins unreachable, classifier blocks country dropdown, no way to pre-authorize consent/account creation
- [BUG] Claude Code MacOS Desktop overrides system text editing shortcuts (option+up/down)
- [GitHub integration]
- [BUG] Opus 5.5 throws cyber safeguard to even a simple prompt with nothing unsafe
- [BUG] Sole decision maker regardless of instructions
- Low contrast on inline code spans / slash-command badges in light theme (VSCode extension)
- [GitHub integration]
- [GitHub integration]
- [Bug] Anthropic API Error: Claude 5 Safeguards Flagging Legitimate Cybersecurity Work
- [BUG] VS Code extension activates in virtual workspaces and shows a misleading "native binary failed to launch (libc)" error
- modifyOtherKeys is only enabled for kitty-protocol/allowlisted terminals, so ctrl+enter can't be bound on xterm/VTE
- Opus 4.6 regression symptom 15: model forgets file paths and repo context within same conversation
- Opus 4.6 regression symptom 8: model forgets file paths and repo context within same conversation
- [Bug] Claude loses file paths and repository context mid-conversation — causes wrong commits
- [GitHub integration]
- [BUG] /loop scheduled task stops firing on its own after the session is backgrounded — fires only right after a user turn
- [BUG] Interactive session was backgrounded with no user input (dispatch source "slash") while a /loop was running
- [BUG] Prompt input longer than 4 wrapped lines: bottom line overlaps / gets mixed, input area does not scroll (Windows Terminal, session attached via agent view)
- Desktop app: image attachments fail above ~100KB while a 17MB PDF attaches fine (macOS, arm64)
Resolved issues (116)
- フィードバック: ポリシー起因の拒否を法的根拠のように説明してしまった件、および過剰なガードレールへの懸念
- [Feature] Deduplicate AGENTS.md fallback when identical instructions arrive via another channel
- [BUG] Bash commands (incl. run_in_background + nohup) killed with SIGHUP (exit 129) at ~5 min on macOS
- [Bug] PR security check blocked by size limit despite passing local review
- [BUG] ~30s of startup input lag: claude.ai connector-list fetch blocks the event loop (26 connectors)
- Desktop app: message input box refills with previously sent message after every send
- [BUG] Desktop (Linux), Code tab: long prompt overflows input box — can't scroll to end, send bar disappears
- [BUG] Desktop: inline Reply box swallows Tab during IME composition (Google Japanese Input) — text commits as unconverted kana
- [Bug] Model downgrade occurring despite authorized work and safe harbor status
- Opus 4.6 regression: session compaction breaks message delivery ("No entregado" after "Haciendo sitio")
- Opus 4.6 regression: context compaction causes message delivery failure and session instability in Claude Code
- [Bug] Sonnet 5 [bio] classifier false positive on Claude.ai web chat — DHCP/DNS/AD infrastructure content (NOT Claude Code CLI)
- [MODEL]
- [Bug] Inconsistent output with global rules and terminal commands causing excessive token usage
- [Bug] Model switched without user request or clear reason
- [Bug] Claude frequently fails on simple confirmation and lookup tasks despite having necessary tools
- [BUG] Context isolation failure
- [Bug] Claude Code making unverified claims and false positive issue detection
- [Bug] Legitimate network data flagged as suspicious content during report generation
- [Bug] Agent executes tasks without user approval
- [BUG] RemoteTrigger routine stalls silently after "Claude Code process started" — no tool calls, session stuck active/idle (same symptom as closed #54260)
- [Bug] Anthropic API Error: Opus 5 Safeguards Flagged Legitimate Request
- sandbox.enabled: true causes unbounded memory growth at startup, OOM (2.1.233-2.1.235, WSL2)
- [MODEL] Sonnet 5
- [BUG] bash-input prefill/tab-completion suggests relative-path script without required ./ prefix
- Claude in Chrome: native host broadcasts every tool response to all connected MCP clients — concurrent sessions silently receive each other's results
- [Bug] Anthropic API Error: insufficient_scope not triggering step-up authorization flow
- Write tool does not enforce the working-directory sandbox that Bash enforces
- Background forked skill completes before its subagents finish; SendMessage resume then fails with contradictory state
- [BUG] Code review agenst stall out completely
- [BUG] Windows service (CoworkVMService) flaps disabled/enabled during MSIX auto-update, blocked by EBUSY on chrome-native-host.exe
- [BUG] Webview CSS: six custom properties are referenced but never defined, plus two light-background issues
- Project custom agents (.claude/agents/) become unavailable to Agent tool after /compact, in same session, no worktree transition
- [BUG] Desktop app crashes when git repo discovery resolves to a stray ancestor .git above the project folder (Windows)
- [BUG] Cowork stays pinned to stale plugin version from custom git marketplace; marketplace cannot be removed or re-added
- Background subagents (e.g. via /code-review) silently auto-deny gated Bash commands instead of prompting
- Windows regression in 2.1.235: PowerShell tool fails from Git Bash with `--model haiku` (`pwsh exited with code 1`)
- [BUG] run_in_background tasks intermittently killed ~17-20s after start, seconds after the arming turn ends (terminal CLI, Linux — not idle-timeout timing)
- Print/SDK mode persists empty thinking blocks in session transcript (signature only) — interactive mode persists full text
- Auto-compact summarization prompt leaks into the compacted history as a fabricated "text-only, no tools" user override
- fullscreen: "1 new message" banner stays after scrolling back to the bottom
- [Bug] /advisor silently omits Fable 5 when ANTHROPIC_BASE_URL is set, despite first-party OAuth login and cached Fable entitlement
- [BUG] VS Code extension: in multi-root workspaces, a tool-level `Bash` allow rule is ignored for git commands only
- [BUG] Canva connector — editing transactions fail with "Not allowed to access editing transaction" despite successful reads and reconnects
- [BUG] Stop button affordance in prompt field is unclear — small square icon alone doesn't read as "stop"
- [BUG] Background general-purpose sub-agents can't resolve MCP tools via ToolSearch even with exact tool names (parent session unaffected)
- [BUG] 7 days of Cowork chat history permanently lost when auto-renewal failed (Windows, self-updated build now in MSIX container)
- [BUG] PostToolUse hook additionalContext (memo_reminder.sh) not reaching Claude in VSCode extension
- /model: switching models mid-session hangs the CLI until switching back to the original model
- [Bug] System prompt injection in terminal interrupts user input and causes false feedback
- CLAUDE.md in --add-dir folders is silently skipped, including when Claude edits files in that folder
- Bash-tool `grep` shim (ugrep emulation): catastrophic backtracking — 6.6 GB RSS / OOM kill on a 20 KB file
- [BUG] Cowork (cloud) re-serves stale local-file content within a session: after a file is staged once, re-staging the same path returns the OLD bytes while reporting the file's NEW size/mtime (false all-clear)
- Bash tool prompts for approval despite matching allow-list rule + defaultMode: bypassPermissions
- ask PreToolUse hook decision silently fails to surface in a top-level session carrying CLAUDE_CODE_CHILD_SESSION=1 (command executes instead of pausing for approval)
- [Bug] Missing explanations in Claude output display
- Linux desktop app: cloud (session_) chats never spawn a process → "send but no reply"; cowork VM "not supported (linux/x64)"
- Mobile Background-tasks panel shows completed subagents as Running with wall-clock timers; Stop button has no effect
- [FEATURE] --prefill flag to pre-fill input box without sending
- [BUG] Cowork project groups disappeared after update (desktop + web), v1.22209.3
- [BUG] Desktop Windows: global session list (sidebar Recents + Home "Active") is frozen at an arbitrary past cutoff, while the project-scoped list is correct — same app, same moment, two contradictory lists
- Bridge/daemon-hosted sessions never write transcript to local JSONL; no way to re-host or export them locally
- [BUG] Chat Markdown rendering mutates meaning of content (renumbers ordered lists)
- Opus 4.6 regression symptom #11: confident hallucination of dates/facts in legal documents — liability risk
- [FEATURE] Desktop app (Windows): keyboard shortcut to start/stop voice dictation
- [BUG] Extension icons not rendering (blank squares) after update, in Remote-SSH session
- [Bug] CVP Integration Fails Against Local Vulnerable Test Instances
- [FEATURE] Per-message copy button in VSCode extension (parity with desktop/macOS app)
- [BUG] Claude Desktop (macOS Tahoe): incognito chat permanently lost after Mac wakes from sleep
- Filed in error
- [BUG] Claude in Chrome: phantom/disconnected tab state persists across full Chrome + app restart (macOS desktop app)
- [Bug] Context window exceeds limit after advisor completion in subsequent turn
- [Bug] Anthropic API Error: Missing server_tool_use block for tool_result
- [Bug] Model falsely claims to have verified information and used tools when it has not
- [Bug] Unclear error message or parsing issue with user input
- [Bug] Model loses context of coding specification mid-session
- [Bug] Claude 3 Opus ignores agent constraints and executes unintended actions
- [Bug] Lagging performance during reasoning extraction
- [Bug] Unexpected CJK characters in Claude Code output
- [Bug] Model switches unexpectedly during active development session
- [Bug] Claude Code hangs during auto-compact operation
- [Bug] Build script export path fails on repeated configuration changes from A to B
- [Bug] Excessive output verbosity
- [Bug] Claude Code ignoring user instructions and modifying approved requests
- [Bug] OAuth token deleted or account disabled after form submission with issues
- Keep getting flagged wrongly for [cyber]
- [Bug] False Positive Content Filtering Blocks Legitimate Input
- I'd be happy to help generate a GitHub issue title, but I need more information about the bug report. Could you please provide the details of the issue you'd like to report? This should include: - What were you trying to do? - What happened instead? - Any
- [BUG] MacBook был offline в VPN → переподключил Wi-Fi → появился в сети → но сама сессия Claude Code не восстанавливается (перезапуск программы, полный выход/вход — не помогли).
- [Bug] Unable to determine issue - insufficient information provided
- [BUG]
- [BUG] - #62606
- Compact fork-resume silently retires the session id — live children and external tooling keep the dead ancestor id, no signal announces the new one
- [BUG] Cowork (Windows): a project's Recents lists sessions belonging to OTHER projects — opening one mounts the wrong working folder + CLAUDE.md (post-unification)
- Claude Desktop stuck indefinitely (5+ min) on [buddy] starting bridge — never reaches "bridge started" (unlike #71951)
- Extreme memory leak, Claude Code, 2,108,208 MB/hour
- [BUG] Plan mode's read-only guarantee is not enforced for subagents dispatched via the Agent tool
- [BUG] Edit/MultiEdit fails with "String not found in file" only on manual confirmation (works with auto-accept) — VSCode extension, Windows
- [BUG] MCP server fails to connect (no logs generated) — WordPress remote connector, Windows
- [BUG] vscode no respond
- [BUG] Pending permission/approval state not always rendered in session view ("waiting on nothing" — agent announces action, UI shows neither prompt nor progress)
- [BUG] One-shot scheduled task with pending permission prompt vanishes from UI (prompt becomes unanswerable, task dies silently)
- [BUG] All previous Cowork conversations are all gone on Windows 11, but Cowork is available
- [BUG] Interactive LSP client never sends didChange after Edit-tool writes — server buffers stay frozen at first-query content (headless `-p` syncs correctly)
- [BUG] Cowork tab missing on macOS 1.22209.3 — runtime present and healthy, tab not rendered (server-side gating, not packaging)
- [BUG] Preview MCP dev server killed silently after ~1h idle — regression in Claude Desktop 1.22209.3 (was fine on 1.21459.0)
- [BUG] MCP tool calls fail with -32003 after approval on mobile web
- [BUG] サイドバーのセッション一覧・グループ表示が崩れる（検索では正常にヒットする）
- [BUG] [Desktop] CCD 2.1.209 regression: "Shell environment extraction" blocks main process for 25s on every app start (Linux)
- [BUG] Cowork tab shows "yukonSilver not supported" on Windows 11 Pro — persists on latest version despite all prerequisites met
- [BUG] Workflow progress indicators (elapsed timer, token counter, current-tool) freeze during final structured-output generation — an active task looks hung
- [BUG] Skill frontmatter model: override ignored when Claude invokes the skill via the Skill tool (works when user types /skill-name)
- 1M-context session becomes unrecoverable: ECONNRESET on large uncached requests, and /compact fails with the same error
- [BUG] Mid turn messages lost on 2.1.205
- Opus 4.6 regression symptom 8: model forgets file paths and repo context within same conversation
- Opus 4.6 regression symptom 15: model forgets file paths and repo context within same conversation