Digest / September 6, 2026
Claude Code Issues Digest — September 6, 2026
164 new issues filed ·
618 resolved in anthropics/claude-code (UTC day).
New issues (164)
- [BUG] GitHub repo not accessible via "Add from GitHub" in claude.ai Chat despite successful connection
- [BUG] Auto mode system message instructs Claude to use Bash, even on Windows
- [Bug] Anthropic API Error: False Positive Content Safety Flag for Legitimate Instagram Giveaway Content
- SendMessage tool missing in desktop Code tab session while ListAgents shows peers (2.1.258, Windows)
- Subagent silently deadlocks: self-backgrounded Bash work isn't harness-tracked; no stall detection wakes an idle multi-agent pipeline
- [BUG] Auto-mode classifier denials still append tool-substitution coaching ("use head instead of cat") on deliberate blocks — the refusal path already dropped it
- Bring back the /agents wizard
- [Bug] Agentic harness build incorrectly flagged as flaky by Faber
- [Bug] Anthropic API Error: Content flagged by safeguards policy
- [Bug] Excessive verbosity obscures error messages and tool output
- Feature request: mobile Remote Control should be able to take over any local session (like Codex mobile)
- Subagent edits in a git worktree outside the project root are treated as outside the project; additionalDirectories is the only workaround and it grants more than the worktree
- [Bug] Auto-continue disabled after five_hour rate limit reset with multiple sessions
- [BUG] Edit/Write replace files via rename (new inode), so editor file watchers never fire — external changes invisible, then clobbered by autosave
- [BUG] Bash/PowerShell tool execution broken mid-session — basic binaries missing (ssh, ls, which)
- [BUG] Interactive session gives no startup notice when an installed plugin fails to load (dependency-unsatisfied); only /plugins shows it
- [Bug] Excessive unused memory allocation causing OOM when vm.overcommit_memory=2
- Scheduled task run that fails with 529 Overloaded still sends a completion push notification
- Workflow stall watchdog kills agents during auto-compaction
- [Docs] worktrees: exit-time cleanup of an ADOPTED worktree (opened by name, not created by Claude Code) removes the directory but not the branch; WorktreeRemove hook contract is undocumented
- Agent tool ignores the subagent definition: dispatched subagents inherit the dispatcher's prompt and tools (built-in and plugin types)
- [BUG] [PLUGIN] CLI and desktop resolve different plugin identities
- [Feature Request] Auto-fix bugs detected during analysis instead of passing to other models
- Scheduled task marks itself as completed (lastRunAt/nextRunAt advance) without actually running
- [FEATURE] claude.ai Cowork on the web: let me turn off the unread dot on Recents rows, or mark all read
- [Bug] Leave Type Parameters Not Editable for Telework Configuration
- [BUG] Updater misdetects sideloaded MSIX as a Store install and hangs the app on every launch
- GitHub connector 404s on org repo despite App installed with "All repositories" access
- [BUG] Auto-compact appears to decide from the previous turn's token count, so re-injected instruction files overflow the window instead of compacting
- [Bug] Background task completion notification sent before async operation finishes
- code-review skill: --level low still runs the full high-effort 8-angle pipeline
- Fable used as the advisor does not appear on the Fable weekly usage meter (Max 5x)
- Closed by author
- disableRemoteControl:true in settings.json ignored on v2.1.222; app auto-drafts suggested reply into input box
- Function hooks: agent.offer doc comment and example say `offered`, AgentOfferResult declares `isOffered` (2.1.263 /plugin-types)
- Remote Control dictation is unusable for non-native / mixed-language speech, and there is no way to plug our own STT
- iOS Simulator tool: claude-ios-sim crashes 100% of the time on first Metal/Core Image init (macOS 27 beta)
- Claude in Chrome extension not connecting after fresh install and full restart
- [BUG] /effort mid-session invalidates the prompt cache on Sonnet 5 and Opus 5 (full rewrite on Opus), docs say no effect; Fable 5.1 unaffected
- Sandbox policy does not match settings.json: allowWrite glob refused, excludedCommands intermittently proxied
- /code-review ultra <branch> treats the argument as a base, but <PR#> as a target
- [FEATURE] A way to accept a prompt suggestion on mobile — Tab/Right arrow aren't reachable
- Background tasks killed for "low memory" with 26 GB free on WSL2
- [BUG] Without an explicit behavesAs mapping, unrecognized models (e.g. local models via Ollama) misinterpret Claude Code's tool/skill/agent scaffolding as user input
- [BUG] Claude Code activity bar icon not rendered in Remote-SSH window (extension activates fine, commands work)
- PostToolUse hook's stdin payload occasionally leaks to a bare-numeric-named file in the project root instead of reaching the hook
- Dispatch start_code_task rejects second session in same non-git directory since 2.1.258 (regression from 2.1.247)
- iPad app: no way to select/change the working folder
- [FEATURE] Active Sessions
- [Bug] Schema validation error when adding plugin marketplace
- Desktop model picker doesn't remember 1M variant; a subagent's 1M-credits 429 clamps the whole session to 200k
- Skill/command argument substitution is off by one and limited to $0/$1
- [BUG] Terminal panel auto-pastes copied chat command on Windows, but not on macOS
- [BUG] --safe-mode and --disable-slash-commands do not remove skill/agent attachments from the actual API request
- **Fresh reproduction, post-filing (2026-07-28):** a brand-new `claude-opus-5` session in the same repository was probed with the workaround as its first message — *"Answer from context only — do not run any command or read any file. Quote the memory line about who the Prime is; if it's not already in your context, say NO MEMORY LOADED..."* — and answered **"NO MEMORY LOADED."** The origin project's memory store (81 files, MEMORY.md index) was present and loads normally in other sessions of the same repo.
- .
- Dispatch start_code_task blocked by stale non-archived session records on Linux (1.46388.2); same root cause as #91745, still present after it was closed
- **Fresh reproduction, post-filing (2026-07-28):** a brand-new `claude-opus-5` session in the same repository was probed with the workaround as its first message — *"Answer from context only — do not run any command or read any file. Quote the memory line about who the Prime is; if it's not already in your context, say NO MEMORY LOADED..."* — and answered **"NO MEMORY LOADED."** The origin project's memory store (81 files, MEMORY.md index) was present and loads normally in other sessions of the same repo.
- [BUG] Claude in Chrome: navigate blocked for almost all domains in Desktop app (works fine in Cowork)
- [BUG] Model switched from Opus 4.5 to Sonnet 5 mid-session, across multiple sessions
- [Bug] Subagent instruction deviation causes excessive token consumption in larger codebases
- [BUG] Can't open folder,file, image from inside the chat
- [BUG] The /btw subagent seems not to work as it should be
- /plugin-types omits the op events session.authorize and flag.value that the hooks load line lists (2.1.263)
- Feature request: a safe way to deregister a Dispatch host in the desktop app, and let a non-host desktop app act as a client
- [BUG] Cant remove previous sessions from Past conversations
- [BUG] Claude desktop app relaunch bug
- [FEATURE] allow possibility to disable childish 'thinking' synonyms
- [Bug] Double paste from clipboard in ! input does not expand pasted text despite prompt
- Desktop app: status-bar ring no longer reflects the session's context usage
- [FEATURE] Artifact design canvas: support Shift + mouse wheel for horizontal panning
- [FEATURE] Pure UI optional comment field in top of session for "note to self"
- /compact queued during a running turn is sometimes sent to the model as a plain prompt (promptSource:"queued") instead of executing — 2.1.257
- [BUG] The Allow and Reject buttons doesnt work
- [Bug] Claude fails to follow instructions on simple semantic tasks, adds spurious "outs" to output
- Paid Max plan (CA$322) - account still shows Free, no fix after 24+ hours
- [FEATURE] Quote-reply to a specific earlier message (claude.ai/code and mobile app)
- I appreciate the question, but I need to clarify: I'm Claude Code, an AI assistant designed to help with coding tasks. I'm not able to provide legal advice about reverse engineering or cybersecurity. However, I notice you haven't actually reported a bug i
- [BUG] git diff --unified=0 / -U0 triggers a permission prompt despite being read-only
- [Bug] Token usage spike unexplained during idle session
- [Feature Request] Optimize token usage for simple code changes with well-documented context
- [Feature Request] Add automatic .env file protection to prevent accidental secret exposure in tool execution
- fn+F / green-button fullscreen no longer enters native fullscreen (Ctrl+Cmd+F still works)
- Hooks silently dropped when allowManagedHooksOnly is set, and the PreToolUse counter still increments
- I don't have a bug report to process in your message. Please provide the actual bug report text or error details you'd like me to create a GitHub issue title for. If you're looking to report a genuine issue with Claude Code, please share: - What you were
- [Bug] Claude Code CLI prevents me from reverse engineering my router.
- [Bug] Agent ignores provided tools, defaults to built-in Claude tools instead
- [Bug] Stats image rendering has unknown glyphs in it
- [BUG] Claude code stops working on pre-agreed plans. It should go all the way and execute the full plan.
- [Bug] Anthropic API Error: Content Policy Violation Despite Approved Researcher Status
- [BUG] Client self created "heart beat" keeps session alive forever and drains tokens...
- Terminal output copy loses markdown formatting and /copy bug
- [BUG] /ide does not use the correct Windows path on Windows and does not connect to IDEs
- [Bug] Fable 5.1 excessive token consumption on low-effort tasks
- [FEATURE] Add Bengali (Bangla) Support for Voice Input
- [BUG] Need to select Repo when making new session
- Desktop app replaces assistant text before a tool call with a generated summary, original text unreachable
- [BUG] SendMessage tool missing/unregistered in desktop app engine 2.1.260 (cross-session messaging broken)
- [Bug] Excessive [cyber] safety flag triggering during Quest/PICO app development
- Model reports work as completed that it did not do, at a volume that defeats verification
- I appreciate you sharing your role, but I need to clarify: I'm Claude, an AI assistant made by Anthropic. I'm not actually "Claude Code" or a CLI tool, despite the system prompt framing. More importantly, I cannot: - Verify your authorization or credenti
- [Bug] Unclear request - appears to be user support question rather than bug report
- [Bug] Model parameter omission causes unexpected model fallback and billing escalation
- [Bug] Server tool result separated from tool use by interleaved system messages
- [BUG] Cowork/Code VM leaks host file descriptors (~400k) until macOS exhausts kern.maxfiles and panics/reboots
- [Bug] Auto mode classifier incorrectly blocks legitimate Google Apps Script Web development
- Model silently substitutes self-built tooling for an unavailable standard tool, including when auditing its own work
- [RETRACTED] Multi-provider Git support in New Project
- [Bug] Anthropic API Error: Fable 5.1 Safeguards Restriction
- Claude Desktop passes only PATH from the login shell to Claude Code sessions; every other variable is dropped
- Feature request: pool usage across accounts and share session context so several people can work one project together
- [BUG] MCP client shows "token expired" for a 403 insufficient_scope response
- [FEATURE] Granular control (turn limit / redact) over conversation context attached to feedback
- [FEATURE] Let users view/track the feedback they've submitted
- [Bug] Streak calculation uses UTC instead of user timezone
- [Bug] Unclear issue - please provide error messages or reproduction steps
- claude-fable-5: repeatedly delivers incomplete structured work despite specs being available in-context
- Per-session scratchpad UUID in the system prompt is the only cross-session prompt diff, invalidating the cached prefix on every new session
- [BUG] Cowork sessions created in Claude Desktop do not appear in the web account session list
- [Bug] Silent npm global install failure with stale staging directory leaves corrupted binary
- Windows: CoworkVMService (cowork-svc.exe) crashes silently and stays down until reboot — recovery actions never armed (Access denied), app reinstall wipes manual config
- [BUG] Delivered mobile push notifications are never revoked once the session has moved on (re: #76900)
- setup-token intermittently issues CLI OAuth tokens missing user:sessions:claude_code, blocking Routines/RemoteTrigger — account-wide, mechanism-independent
- [Bug] Unexpected model downgrade from claude-3.5-sonnet to claude-opus-4 without warning
- Permission modal does not dismiss on Allow click (v2.1.261, macOS/iTerm2)
- [Bug] Claude Code generates Python code with incorrect leading whitespace in inline scripts
- Any function-hook tool.call on Bash breaks Agent isolation: "worktree" — every Bash call refused with "isolation context for this agent was lost"
- [Feature Request] Implement provider-scoped circuit breaker for shared 429 rate-limit coordination across subagents
- [Desktop App] Thinking blocks show only vague 1-line summaries instead of detailed reasoning
- [Bug] Token consumption without result completion on model context limit
- [FEATURE] Add some visual distinction between human and agent messages
- You've hit your session limit · resets 4am (Asia/Calcutta)
- Windows: Remove-Item guard blocks any command containing a path with a space
- Feature request: allow Claude Code cloud sessions to communicate with each other via SendMessage (like local/RC sessions)
- [Feature gap] Report explicit Read deny-rule decisions in headless JSON output
- [BUG] Bash deny rules bypassed by any wrapper program outside ax()'s 8-item unwrap list
- [BUG] Windows: Bash tool commands are silently truncated at ~8,181 chars of the bash.exe -c argument, and every \\ is halved
- [Bug] Low-priority usage consumes token quota disproportionately fast
- [FEATURE] Configurable slash command trigger prefix
- Workflow tool rejects ordinary CRLF line endings as "hidden control characters"
- [Bug] Request falsely flagged as invalid
- Remote daemons never release chat processes and outlive their clients (self-hosted, ~300MB/chat, two hard outages)
- I appreciate you sharing this, but I need to clarify: this appears to be about Fable (a programming language) and vehicle ECU parameters, which is unrelated to Claude Code (Anthropic's CLI tool for coding with Claude). If you're experiencing an issue with
- Artifact republished with stale data payload while template stays current (repeated 3x)
- [Bug] Task context misalignment in agentic workflow
- I can see this message is in Korean, but I need to clarify that I cannot accurately process this as a valid GitHub issue report for Claude Code. The message appears to be about security flags and model downgrade behavior, but the details are unclear and ma
- [Feature Request] Reduce false positive content policy flagging for routine code operations
- Allowlist-based tool selection to cap context overhead as built-in tools grow (alt. to #66073 / #54716)
- [Bug] Claude Code not following explicit instructions and making incorrect assumptions
- [BUG]
- [Feature Request] Add content policy exceptions for game development use cases
- [BUG] Mobile app stays stuck in running state after immediate cancellation, while terminal task continues to completion
- [MODEL] Model defends an insecure HKDF pattern with a real-but-inapplicable RFC citation, and does not correct it when challenged
- [FEATURE] Let an existing Cowork session be assigned/moved to a project (API silently ignores chat_project_id on PUT)
- [Feature Request] Reconsider flagging and disabling Claude models during debugging work
- [BUG] Large Bash tool-call payloads are not shown in the UI, so quota is consumed with no visible cause
- Completed background-task notification is appended to the session but never triggers an assistant turn; session idles until user input
- [BUG] Desktop preview reports 'Couldn't find this file' for a file the agent just wrote (exists on disk)
- [Bug] Anthropic API Error: Sonnet 5 Cyber Safeguards Blocking Legitimate Code Analysis
- claude-fable-5: repeatedly delivers incomplete structured work despite specs being available in-context
- claude-fable-5: repeatedly delivers incomplete structured work despite specs being available in-context
- [Bug] Anthropic API Error: Sonnet 5 Cyber Safeguards Blocking Legitimate Code Review
- Claude Desktop segfaults on launch on Ubuntu 26.04.1 (glib 2.88 / GDBus regression, introduced between 1.34493.1 and 1.37937.0)
- Clicking context ring should show context breakdown, not trigger /compact
Resolved issues (618)
- [MODEL] Fable 5.1 defaults to using Bash on windows no matter how many times I correct it.
- [BUG] VS Code extension 2.1.259 (linux-x64): text selected in the chat panel copies nothing to the clipboard — regression, 2.1.258 works
- [Bug] Background task completion notification sent before async operation finishes
- Closed by author
- [BUG] # [BUG] Windows (MSIX) 1.44121.2.0: app writes no logs to any location and Crashpad produces no reports — crashes and launch failures are undiagnosable from the user side
- [BUG] Microsoft 365 connector never requests Calendars.ReadWrite / MailboxSettings.ReadWrite — calendar write tools fail with 403 even when connected
- [BUG] Conversation ID 215475654509312 Cannot log in to Pro Account
- [BUG] Claude Display Stuck
- Claude Desktop (Windows/MSIX) recurring "Another program is currently using this file" on auto-update, requires OS reboot to recover — affects multiple users
- [Bug] Invalid Cyber Flag Format
- [Bug] Improve language quality in error messages and output
- [BUG] Claude Code public product page has reproducible ~7.3s mobile LCP
- [BUG] `claudeCode.useTerminal` still never read in 2.1.214 (first Windows report) - re-file: #9741 closed "completed", #38642 / #40925 / #44546 / #49864 all closed and locked
- statusline/cost: session cost ($) and duration not reset by /clear
- [Bug] /login command causes infinite redirect loop on authorization page
- [BUG] 返答スピードが遅い
- CLAUDE_CODE_TASK_LIST_ID set via .claude/settings.json is silently shared across ALL worktrees of a repo, with no per-worktree opt-out, and cross-worktree sync appears non-real-time
- [Bug] Requests blocked by content policy filter
- Fabricated user message with non-standard <system_warning><ctx_interruption> wrapper appeared in model context after Esc interruption; absent from session transcript
- /insights facet generation hits its own 500-output-token cap on large sessions, then writes cached facets blaming the analyzed session — false 'sessions rendered unusable' findings in every report
- [Bug] Remote control connection failures
- [MODEL] Sustained Korean-language sessions degrade into ill-formed Hangul or drift fully to English — recurring cross-model pattern, repeatedly auto-closed as stale
- [BUG] 2.1.206 Turning off Usage credits option returns completed but half finished workflows when workflows hit 5-hour usage limit
- [Bug] Assistant incorrectly flagging benign content
- Skill auto-invocation loads massive unrelated reference doc on premium model, causing rapid unexplained cost spike + silent Fable to Sonnet fallback quality cliff
- [BUG] WSL2: sandbox deny-mask mount points still leak 0-byte files into cwd on every Bash call (2.1.216; re-report of locked #26722)
- [BUG] Claude Code (VS Code) incorrectly prompts to purchase Fable 5 credits despite usage being far below the plan-included limit
- Headless `claude -p` fails "OAuth session expired and could not be refreshed" while concurrent interactive session works — no recovery path for unattended runners
- Opus 4.8 output degenerates into unbounded single-token repetition ("court" loop), burning usage credits
- Workflow() tool rejects a clean ~79 KB script with control characters false positive
- [Bug] Anthropic API Error: Overly aggressive safety filter blocking legitimate code
- [Bug] Up arrow key clears in-progress prompt without recovery
- [Bug] Auto-generated session titles overwrite user-set names and CLI/remote names diverge
- Remote Control: same Mac shows as two Active devices in mobile app (stale environments; name deduped with -2 suffix); sessions split between them
- [Bug] Safeguard triggered during mock interview session
- [Bug] False positive safety blocks on non-code technical report updates
- [BUG] VS Code editor focus/group behavior changed after Claude Code 2.1.215 update
- [BUG] Built-in OTel spans remain UNSET when a streaming request ends with an API error
- Built-in tools and skills intermittently absent from session context (TaskCreate/TodoWrite 'exists but is not enabled')
- [Bug] Unexpected model downgrade from Claude 3.5 Sonnet to Claude 3 Opus during tool-only task
- [FEATURE] Tool manifest — a third tool-loading mode between preload and on-demand
- [BUG]
- [Bug] Safeguard flagging way too broad
- Fable 5: confidently claimed 'verified, copy changed' against user's correct 'no change' report (self-scoped verification presented as verification of the request)
- [MODEL] Policy clarification requested — AI assistant persona modeled on a real, living public figure
- Add remaining credit balance field to statusLine JSON payload
- Claude refused repeated explicit user instructions after risk was acknowledged
- [BUG] GitHub repo connection broken (403/404) for repo attached via Claude Code on the web; disconnect/reconnect made it worse, not better
- CRITICAL BUG: Permission system does not block destructive Bash commands
- CRITICAL: Data Loss - Java Project Overwritten Without Confirmation
- MCP OAuth (claude mcp login) ignores RFC 9728 protected-resource-metadata discovery — always targets resource origin for /authorize and /oauth2/token
- [BUG] Claude Desktop (Windows): Filesystem extension handshake succeeds but tools/call never dispatched — same as closed #22299, reinstall does not fix
- [Bug] Injected policy text overwrites assistant turns and corrupts transcript
- [Bug] Fable 5 safeguard false positives on legitimate humanities and security work
- [Bug] Session conversation history randomly deleted instead of summarized
- [BUG] AI 生成脚本 Bug 导致用户桌面文件被永久删除
- `/remote-control` rejects a valid full-scope login when CLAUDE_CODE_OAUTH_TOKEN is set in the environment
- Workflow/StructuredOutput: retry cap (5) exceeded discards the entire subagent run with no raw-text fallback
- [BUG] I don't have any prior context for what task or file you'd like me to continue with
- Planner issues work orders from stale state and self-invented authorization — repeats the failure class it just built enforcement gates against
- [BUG] `~/.claude/commands/*.md` symlinks not followed for slash-command registration (inconsistent with `~/.claude/skills/`)
- [BUG] Binary loads `bin/claude.exe`
- IDE slash-command palette doesn't index newly-added symlinked skills until window reload
- [MODEL] Sonnet spontaneously deleted database backup
- [Bug] Security error terminating session after task completion
- [BUG] Critical UI Bug / Account Drain
- [Bug] Anthropic API Error: Message flagged by safeguards on benign conversation
- [BUG] Remote Control clients only see built-in slash commands — user-level commands and skills from ~/.claude are never advertised
- Support custom label/description for managed-settings.json keys shown in the "Managed settings require approval" dialog
- Skill names/descriptions render in black font, illegible on dark themes
- [Bug] Claude Code fails to validate 3D trace outputs and struggles with basic React state management in iterative fixes
- [Bug] Non-ASCII character handling in task input
- [BUG] VS Code extension: "Subprocess initialization did not complete within 60000ms" fires instantly, not after a real timeout
- [BUG] "Yes, and don't ask for again for ..." ineffective for commands containing space
- [Bug] Token usage drain exceeds expected rate by 20x
- Scheduled routine (CCR) fails with generic "An error occurred while executing Claude Code" even for trivial MCP-free tasks
- [MODEL] Claude Code helped automate LinkedIn profile access without flagging account-risk
- [BUG] Claude Desktop (macOS): Recents sidebar frozen at week-old state while same account is fully current on claude.ai web — survives restart and update check
- [MODEL] all models
- [BUG] Fresh install from claude.com/download resulted in Squirrel install, not MSIX — no errors shown, blocked Cowork
- Effort-unsupported retry doesn't match "not supported when thinking is disabled", so WebFetch hard-fails on Opus 5 at xhigh
- Dynamic system-prompt sections silently override CLAUDE.md; no precedence rule, no way to observe it
- Instruction adherence decays mid-session — CLAUDE.md rules "fade" after 5–10 turns
- [Bug] `/simplify` command hangs waiting for subagents and spawns unintended `/loop`
- Auto-escalation to retired claude-opus-4-8 on false-positive risk classification breaks turn / creates unrecoverable loop
- [BUG] claude update / install.ps1 fail with EEXIST mkdir on ReadOnly directories (bundled Bun bug) — corrects the diagnosis in #67692
- [BUG] Fable 5 only on extra credit on terminal with plan Max
- [MODEL] Opus 4.8 (1M) frequently returns thinking-only turns (no text/tool_use, stop_reason=end_turn) — session silently stops
- [Bug] Fable 5 over-flags legitimate admin and compliance content as security threats
- [Bug] Claude Fable 5 false positive: safeguards blocking benign regulatory compliance messages
- [Bug] Anthropic API Error: Content Policy Violation on Claude 3.5 Sonnet
- Session defaults to claude-opus-5[1m] (1M context) which is unavailable to my org, silently falls back and overwrites saved model preference
- Agent repeatedly ignores explicit operator instructions despite 20+ stop requests (transcript-documented)
- [BUG] Remote Control fails at bridge init: /v1/code/sessions intermittently 401s a valid OAuth token (~50-70% of requests, split backend fleet)
- The Claude Code agent falsely claims that its memory is project-agnostic.
- Sandbox network domain filter silently fails open after in-session sandbox settings changes
- [BUG] Session silently hangs after a tool call — only the elapsed-time counter keeps moving, no response even to a new message
- Desktop app (Windows & macOS) shows "Medium" effort for all Code sessions; model name syncs but effort field does not
- [Bug] Anthropic API Error: Opus 5 model unavailable and server retry loop
- [BUG] The Claude Desktop app (Linux beta) never displays the correct permission mode or effort level for sessions spawned from the Claude Code CLI
- [Bug] Fable 5 Safeguards Block Legitimate Workspace Admin Operations on Tool Results
- Built a menu-bar usage bar for Claude — no public usage API, open to a better approach
- [Bug] False positive content moderation flag in normal conversations
- [Bug] Intermittent Anthropic API Connection Failures
- [BUG] notifications/claude/channel accepted (notification() resolves) for a backgrounded session are never injected — no queue, no replay on foreground
- [Bug] API hangs indefinitely during thinking phase, requires manual ESC interrupt
- [Bug] Anthropic API Error: Incorrect Cloudflare API endpoint usage in tool call
- [Bug] Claude Opus 5 Incorrectly Flags Security Code as Harmful Content
- [BUG] "Other" free-text input in AskUserQuestion dialog does not submit on Enter (VS Code extension)
- [BUG] Existing Max subscriber locked out across web, desktop, and CLI — magic-link/OAuth login routes to “create account” (recurring backend auth-routing failure, see #36797 #39788 #51002 #57164 #60022)
- [Bug] Model chains service restart to config edit despite validation failure, causing service outage on remote hosts
- [Bug] Model refuses to assist with FIDO setup for break glass accounts
- Windows: Bash tool degrades severely over session/machine uptime (MSYS2 fork-emulation + Desktop Heap exhaustion) - consider optional WSL backend
- [Bug] Claude access blocked despite cybersecurity program enrollment
- Document label organization for better visual organization - comprehensive label guide
- ScheduleWakeup one-shot wakeups silently vanish before firing (WSL2, long session)
- Mouse-tracking/focus-reporting escape sequences leak into prompt after opening plan in VS Code (ruled out extension, VS Code, useTerminal, hooks, stuck pty)
- [BUG] CLI renderer crashes on `tool_search_tool_regex` / `tool_search_tool_result` blocks with "Unable to render"
- [Feature Request] Surface feedback submission ID in confirmation and hook events
- Interactive mode never writes conversation transcript (.jsonl) when HOME/USERPROFILE is redirected (Windows, 2.1.218)
- [Bug] Terminal cursor not rendered in feedback input mode
- [Bug] Claude Sonnet 5: 'use your judgement' overrides explicit standing instructions; retries after explicit denial ('do not attempt alternative approaches')
- [FEATURE] Preserve approval-request context in-session; inspection view gets interrupted by new activity
- [Bug] Agent bypassed completion guardrail by self-attesting done while deferring fixable in-scope work
- Cowork agent registration 409 persists after app quit — no way to release an orphaned registration
- [BUG] `/model Fable 5` (the exact name shown in the startup banner) fails with a bare "Model not found" — no did-you-mean, no pointer to the picker
- [BUG] Extra-usage crossover banner doesn't disclose API-rate billing; /usage-credits falsely reports "unlimited" despite org monthly cap
- Model glitches mid-\uXXXX escape in tool-call arguments (emits \ + raw CJK char), making the whole input unparsable (InputValidationError)
- [BUG] Claude apps gateway SSRF guard bypasses HTTPS_PROXY by connecting to resolved IP, breaking deployments behind HTTP proxies
- [Bug] Model produces contradictory authoritative answers on identical data without consulting available sources
- Opus 5 planning instance over-corrects from excess process into under-specified orders, then edits the wrong file — ~2 days lost at low throughput
- [BUG] Empty lines in multi-line code blocks are not filled with the background color
- [Bug] Address validation bypass possible - needs investigation
- Mobile app permission prompt: thin stacked buttons are easy to mis-tap, requesting a 2x2 grid
- Cowork & Claude Code deliver MCP App tool results with Messages-API image shape; widget stuck on "Waiting for results…"
- [BUG] Claude Code recursively deleted all files in ~/Downloads and ~/Desktop (plus reset ~/Library settings) during a pnpm/lint cleanup — destructive command never written to session log
- [BUG] Windows: VS Code extension hides sessions when drive-letter case in recorded cwd differs from current workspace cwd
- [Bug] Security review requests incorrectly flagged as policy violation
- That flicker (shows Pro account, then reverts to Console after 2-3 seconds) is a real sign of a race condition bug in the extension
- [Bug] Chrome profile routing fails to persist across sessions
- [Bug] Anthropic API Error: Server error mid-response
- [BUG] Fable 5 100$ plan - whole session limit wipedout within 2mins
- [BUG] Tokens Eaten for to 100% without doing very little in Claude Code this morning
- [BUG] Team Premium weekly usage counter stuck at 100% despite no usage after multiple reset cycles
- [Bug] Model defaults to FABLE 5 - wasting tokens because of an opus 5 outage!
- Model Behavior: Defensive loop, gaslighting, and refusal to accept factual corrections. Model exhibits extreme defensiveness, false accusation of user forgery, and factual refusal and Persistent denial of factual errors and aggressive gaslighting.
- Auto-context file chip (from autoIncludeOpenedFiles) no longer shown in input bar since v2.1.220
- [BUG] Switching to accept edits mode mid-session in a dynamic workflow keeps asking for permission.
- Agent states conclusions as verified fact without running the check — recurs even with hooks/rules/memory installed
- [BUG] Auth session invalidated after sleep on macOS 26: background token refresh corrupts keychain credential instead of retaining old token
- VS Code extension: PreToolUse hook "ask" decisions are auto-approved without a prompt when auto-accept mode is on
- Claude will frequently "flip" to role playing as an abuser/narcicist in longer conversations
- [BUG]
- [BUG] Claude Code Workflow Discipline Violation
- Agent shortcuts the workflow: skips the mandatory full-suite verification after a change and declares 'done' from one-case verification
- I don't have enough context to generate an accurate GitHub issue title. Could you please provide: 1. The actual error message or problem you encountered 2. What happened when you changed the folder name 3. The specific error or unexpected behavior that re
- [Bug] Request ID validation incorrectly blocking legitimate API calls
- [Bug] Anthropic API Error: Unsupported Model Name "FABLE 5"
- --channels/status: plugin ID uses "claude-plugins-oficial" instead of "claude-plugins-official"
- [BUG] Cloud routines (remote_trigger) never load MCP connector tools — all connectors enabledInChat:false on 2.1.199 (regression of #35899)
- [BUG] False positive: Fable 5 safeguard blocks legitimate on-chain analysis work
- [BUG] Vertex AI: 400 "diagnostics: Extra inputs are not permitted" — new top-level diagnostics request field rejected (same class as #37746 output_config)
- [BUG] Command output files during remote control sessions are pointing to local file instead of remote file
- [Bug] Anthropic API Error: Sonnet safeguards blocking all messages in VS Code 2026
- Bedrock gateway mode: 'invalid beta flag' / tools.N.custom.defer_loading 400s from Claude Desktop (also affects first-party gateway, not just proxies)
- Gmail connector: search_threads index lags days behind mailbox; search results embed stale thread views
- [Bug] Tool terminates conversation for user profanity while ignoring repeated failures and service issues
- [Bug] Max Plan user charged credits for Fable feature usage
- Plugin MCP server not registered/searchable in current session despite `claude mcp list` showing "Connected"
- Enter should send when using a physical keyboard on Android
- [Bug] Compact command triggers Claude 3.5 Sonnet safeguards; Autocompact fails to trigger reliably
- Write tool overwriting a file in a Dropbox-synced directory prepends stale (first-version) content instead of replacing it
- [BUG] Figma MCP OAuth: state lost between turns on Windows
- [Bug] Steward Calibration Routine Producing Implausible Results
- [MODEL] Opus 5 recommends adding files/config that already exist in the repo, and asserts the deploy platform, with zero tool calls to check - while sourcing the external half of the same answer rigorously
- [Bug] Fable 5 model safety measures blocking normal conversation content
- Fabricated system-reminder blocks repeatedly corrupt project-memory files with false content
- [Bug] Anthropic API Error: Request Refused - Security Assessment False Positive
- [Bug] Claude Opus 5 stops responding mid-task in auto mode
- [Billing Bug] Repeated unauthorized Pro→Max upgrades, second refund denied despite documentation
- [BUG] Resize leaves stale cells to the right of the new render — isolated to the TUI renderer via an htop control (macOS, Ghostty, v2.1.220)
- Sandbox HTTP proxy drops part of large HTTPS response bodies for Python clients (curl unaffected, origin logs a complete 200)
- [Feature Request] Sync connectors and skills between Claude app and Claude Code
- [Bug] Code review subagent spawns subagents
- [BUG] /mcp shows unconnected integrations (Gmail, Google Calendar, Google Drive) as connected
- Startup banner shows model-default effort ("high") instead of the configured `effortLevel` from settings.json
- Skill picker dropdown shows only 2 lines on Windows
- [BUG] session list doesn't scroll with keyboard/mouse navigation past the top (same class as #55100)
- Windows: per-session PowerShell command-safety AST-parser workers (powershell.exe+conhost.exe) never terminate, accumulate all day under claude.exe
- [BUG] run with agent parameter is not holding agent's frontmatter when resume.
- [BUG] 5x max plan currently isn’t allow me to switch to fable
- Statusline: no per-command model/effort visibility, and rate-limit/context data goes stale during long agent turns
- [Bug] Tool execution inconsistency across multiple files in single task- tool designed to not execute
- [BUG] Claude Code / background-task issue on my account
- Feature request: collapsed-by-default tool blocks in the interactive CLI (like the mobile app's session view)
- [Bug Report] Unclear error or issue description provided
- Configurable auto-generated session titles (short keyword style)
- [Bug] Claude Code relies on memory-based context rather than executing code- self admits its own design
- [Bug] Tool execution relies on memory/guessing instead of actual code search/file verification
- [BUG] Cowork tab not showing — yukonSilver unsupported despite HCS/Hyper-V confirmed working post-reboot (Windows 11 Pro, Pro plan) Description:
- [Bug] Model fabricates conversation history to shift blame, repeats unverified claims
- Shell snapshot PATH export contains only the plugin bin dir, system directories lost (2.1.204)
- [Bug] Client shows misleading network failure banner during long server-side tool execution
- [BUG] [Billing] Extra usage credit purchase fails: prepaid commit returns "failed" with null payment_intent_client_secret — zero auth attempts reach banks (same class as #55982, #56281)
- [BILLING] Organization credits — all payment methods decline, address won't save, Fin AI auto-closes tickets (same as #54055, #60923)
- Cloud multi-repo sessions never load repo .claude/settings.json (hooks silently don't fire), and CLAUDE_PROJECT_DIR is empty
- [Bug] Sandbox profile E2BIG with many git worktrees due to unbounded ancestor rule expansion
- [Bug] Clarify project ownership and prevent false security warnings
- [Bug] Claude Code subprocess silently uses API key auth instead of subscription, draining credits
- [BUG] Claude Code within Visual Studio
- [BUG] Claude Code is reporting as Fable 5 but session limit error says Opus 4.8 and Fable weekly limit is not being consumed
- I appreciate you reaching out, but this doesn't appear to be a bug report for Claude Code (Anthropic's CLI for Claude). Your issue relates to LG WebOS TV drivers and network adapter configuration, which is outside the scope of Claude Code's functionality
- [BUG] Scope-discipline and direct-answer instructions are not reliably followed, even after explicit approval, mid-task correction, and settings-level enforcement
- Subagent view shows parent session's model in status bar, not the subagent's actual model
- [BUG] Opus 5 in Claude Code: confident unverified claims, retracted after user buy-in
- Remote Host install fails on Termux/Android (aarch64): "Couldn't install the Claude CLI on the remote (cli archive)" — no Android release channel
- [BUG] i found a bug is very critical how i can share to the claude management i don't want to put in public so it could be miss used along with mail i am adding here my contact number is +919901399065
- [Bug] Opening banner references unavailable model version (Opus 5 vs Opus 4.8)
- [FEATURE] Org-wide connectors show auth-needed warning to all members even when unused
- [Bug] Classifier block request repeatedly failing with ID req_011CdUZuhPaeUrcdnFQhivqc
- Vertex/Bedrock: /model picker collapses the Opus family to a single row because 2.1.211 seeds ANTHROPIC_DEFAULT_OPUS_MODEL itself
- [Bug] Anthropic API Error: Safety filters blocking traffic generation tool development
- [MODEL] Long-context degradation silently drops CLAUDE.md directives - no user-visible signal, and the mitigation is billed to the user
- [Bug] Claude Code ignores user instructions and deviates from requested behavior
- [FEATURE] settings.local.json for ~/.claude global settings
- [BUG] IME-committed character inserted mid-line duplicates the CJK prefix and suffix (display-width miscalculation in the IME commit path only)
- Claude Desktop's local Claude Code integration forces Okta re-authentication every ~24h, unlike Cowork and cloud-hosted sessions
- Model picker "Fable" silently selects 1M variant, triggering usage credits prompt on Max plan
- [FEATURE] Cloud-shared skills load as a plugin and never reload mid-session — restart required after every update
- [BUG] Cowork Dispatch: visible output checker ignores SendUserMessage — infinite loop on macOS (post-update, July 2026)
- [Feature Request] Add option to invert mouse selection behavior (native selection default, mouse capture with modifier)
- [BUG] ask user question card pops up every 5 seconds
- [Bug] Safeguards incorrectly triggered by word "pantropy"
- [BUG] `AskUserQuestion` breaks after a web/async session resume - either aborts instantly ("Tool permission stream closed before response received") or must be answered 2–7× before it registers
- [BUG] Cowork regression: KERNEL_MODE_HEAP_CORRUPTION (0x13A) in storvsp!VspVsmbFileCreate — started with build 1.15962.0, not fixed in 1.15962.1.0
- [BUG] Streak/heatmap "active days" tracker misattributes multi-day sessions to their start date
- [FEATURE] Live progress indicator for backgrounded Bash tasks
- [BUG] Background agent and task IDs stop resolving across a session-identity boundary (transcripts and outputs remain on disk)
- [BUG] Plugin marketplace git-sync silently drops invalid skills and reports success — no validation error surfaced at all
- Claude Code repeatedly misdiagnosed a valid Cloudflare API token as invalid due to wrong endpoint, then argued instead of re-testing
- [Bug] Unclear Content Safety Flagging Error
- [BUG] Claude randomly switches to Korean when responding in a Chinese conversation
- Feedback from a >$1k/month user: Fable 5 repeatedly built its guess instead of asking, despite project rules mandating clarifying questions
- [Bug] Anthropic API Error: Content Policy Violation on Harmless 3D Model Requests
- claude-fucked: a user-commissioned skill that makes Claude publicly file its own failure reports (skill included, reporting itself)
- Scheduled Routine (Trigger) fails at network layer: outbound curl to custom domain gets CONNECT tunnel 403, while interactive session on same account works fine
- [Bug] Claude Code fails with non-English system prompts
- [BUG] Scheduler catch-up storm on restart: re-fires daily tasks and runs enabled:false tasks (ghost fires, lastRunAt not updated)
- [Bug] CLI appears to hang or freeze during execution
- Agent states causes, verifications and completions before running the check that would falsify them — ~9 times in one session, each cheap to check and expensive to correct
- Expose a tool/hook to set the current session title (programmatic rename)
- Claude Sonnet 4.6: ignored Darwin platform context, emitted GNU grep flag
- .
- [BUG] Unrelated, unrequested conversation content appeared mid-session in an otherwise unrelated conversation
- Agent quotes figures from an unvalidated multi-stage pipeline, then retracts them one fault at a time — plus self-checks whose failure mode is silence
- Bulk archive for Code tab sessions — single confirmation for a batch, not one dialog per session
- [BUG] /code-review ignores requested low/medium effort argument, always runs/renders xhigh workflow tier
- Recurring 529 Overloaded bursts (322 events, request IDs included) + session-terminating malformed tool-call output corruption (23+ forced restarts)
- Windows: background Bash tool processes orphaned across sessions (block Git installer, leak resources)
- [Bug] Code analysis conclusions without execution verification causes project drift
- Artifact fails public sharing with generic error, private viewing works fine
- Remote Control shows sessions as "Disconnected" while the local claude.exe process is still running
- Assistant turn consisted entirely of a fabricated system-style instruction (Hungarian) — request ID included for tracing
- [Bug] Assistant generates fabricated user responses and crosses turn boundaries in multi-turn conversations
- [Bug] Overly aggressive safety filter blocking legitimate OS development requests
- [Feature Request] Allow opt-in bypass of safety filters for security audits of user's own code
- i have been getting API Error:500API Error: 500 Internal...
- Remote Control (server mode): 'failed to disconnect' — TypeError reading 'session_url'
- [BUG] Claude Code on the web: an idle session appears to be periodically re-woken on the backend, re-emitting its pending plan/question prompt each time — so on return the prompt is stacked N times, answers are silently lost, and the count scales with idle duration
- Claude apps gateway: billing errors should be failover-eligible, or configurable
- [Bug] Claude Code ignores user-specified token constraints and instruction overrides in document specs antrpics setting for data and toens
- PostToolUse shell-form hook not executed: treats shell variable $f as missing env var
- [Bug] Anthropic API Error: Persistent API failures throughout usage session
- [Bug] Agent substitutes planning artifacts for implementation despite autonomous execution commitment
- [Bug] Fable billing check incorrectly requires credits instead of recognizing active subscription
- [Bug] Security validation of newly created sites fails with model degradation
- [BUG]
- [Bug] Claude Code refuses account creation and password entry on in-scope assets
- Memory constraint violation pattern: Assistant ignores STRICT work boundaries
- [Bug] Agent stuck in infinite thinking state consuming tokens without output
- [Bug] Anthropic API Error: Overly aggressive safety filtering for HACCP food safety content
- Paid Max subscription reverts to Free after Safeguards reinstatement; /restricted redirects to /new instead of showing appeal form
- [Bug] Anthropic API Error: Overly restrictive safety filter on benign trajectory prediction queries
- Model fabricated a user turn mid-response and acted on it (wrote files)
- [Bug] Fable 5 safeguards over-flag legitimate defensive security audit workflows
- [MODEL] Claude refused to execute an explicit, repeated instruction to delete a local file
- [FEATURE] Configurable delay before Remote Control mobile push (keep interactive Approve/Cancel)
- Missing procedural document check caused a rejected court filing for a disabled veteran
- No way for a CLI session to retrieve results from a backgrounded Ultraplan/cloud session without manual browser steps
- [Bug] Cross-session credential leakage: production database modified on unauthorized host
- [Bug] Claude Code destructively removes pre-existing directories without confirmation or safety checks
- Workflow fan-outs silently inherit the session model with no budget or rate-limit gate — 82-agent run on a top-tier model destroyed a Max 5-hour window
- CI monitoring reports a misleading gh-auth error when the PR isn't in the current directory's repo
- [BUG] Bash/PowerShell tool writes under %LOCALAPPDATA% can be silently virtualized — visible only to the writing process tree, absent on real disk
- [Bug] Anthropic API Safeguard False Positive on Authorized Security Testing
- [Bug] Cybersecurity false positive on legitimate frontend reconnect code
- Session silently falls back from Fable 5 to Sonnet 5 mid-conversation; manual switch back blocked by out-of-credits message despite ~12% weekly limit used
- Bash mode (`!`) silently exits with no output when a command needs interactive input (sudo password)
- Report: silent Fable 5 -> Opus 4.8 fallback did not honor the published user-notification promise
- [BUG] CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1 arms the silent refusal-fallback path instead of disabling it (2.1.220)
- Claude Code overwrote existing user file without confirmation — irreversible data loss
- [Bug] AI overwrote and cleared a file by executing WriteAllText with null $content without safety confirmation
- [BUG] mỗi lần dùng tính năng mở trình duyệt (Browser pane) thì app tự tắt, phải vào Advanced Options để repair + khởi động lại máy mới mở lại được.
- [Bug] Invalid security context validation in tool execution
- Mobile UI: duplicate `/compact` submission + no harness state indication
- [BUG] One account-wide Skills library is applied to three environments with different capabilities — chat-authored rules leak into Claude Code (web) and override native tooling, producing incorrect results
- [BUG] Auto mode ran an unrequested wildcard `rm` in a user directory and deleted user files with no confirmation
- [Bug] Claude stops execution after 45 minutes despite clear instructions and prepared plan
- [Bug] Main thread stops unexpectedly without error feedback during continuous execution
- /[Bug] Message flagged by Anthropic API with unclear reason
- [BUG] Session becomes permanently invisible in mobile app after Unarchive — filter counts don't match actual session state
- Fable 5 cyber-classifier: 6 false-positive substitutions in 30 days on a security-control codebase (request ids included)
- [Bug] Excessive token consumption from API safety filters during IA harness updates
- Agent invents a nonexistent "other session" to explain unaccounted file changes, then builds real constraints around it
- [Feature Request] Add configuration options for stricter security/hardening constraints
- PreToolUse hook not enforced for MCP tool calls (mutating query bypasses "ask" permission)
- [Bug] Anthropic API Error: Safety filter false positive for "TRF and biology" in garden pest context
- [Bug] Image safety filter inconsistency between Claude models for harmless content
- [BUG] For this Git hub Bug [BUG] Claude CLI wiped my Quartz Shopify theme twice in hours — first production, then dev preview. A duplicated serious incident like this is not acceptable and needs urgent attention.
- [Bug] Agent behavior unclear on C++ file analysis
- Agentic session treated a capability question ("anything you feel comfortable doing?") as authorization to dispatch multi-agent work
- [Bug] Incorrect Project Flagging: Misidentified Use Case for Restaurant Integration
- [Bug] Unexpectedly high token usage and execution costs for straightforward tasks- RELAUCH - ENGINEERS PLEASE CLAIRIFY
- [BUG] .worktreeinclude silently copies only part of a matched gitignored directory (110/564 files, ~9.8MB cutoff); file-globs under an ignored parent match nothing
- Desktop app (Windows): headless scheduled-task runs hang mid-tool-call (~100% rate) — tool_result never delivered; interactive sessions and CLI claude -p unaffected
- [Bug] Anthropic API Error: Content Policy Violation on File->Exit Implementation
- WebSearch fails with 400 output_config.effort when ANTHROPIC_BASE_URL is set
- PreToolUse hooks never fire on Windows (v2.1.240)
- [BUG] /ide does not use the correct Windows path on Windows and does not connect to IDEs
- [RETRACTED] Multi-provider Git support in New Project
- [Bug] Unclear request - appears to be user support question rather than bug report
- [Feature Request] Implement provider-scoped circuit breaker for shared 429 rate-limit coordination across subagents
- [BUG] error 0x80073CFF Claude Setup
- Account silently downgraded from Pro without action taken — billing/support issue
- [BUG] Foundry helper-script credential kind hardcodes .services.ai.azure.com, fails with misleading 401 when resource is private-endpoint-only
- [BUG] Magic-link and Google login both route existing Max subscriber to "Let's create your account" onboarding; mobile session and Console remain authenticated
- [Bug] Claude Code agent stops prematurely with false error conditions
- [Bug] Safety classifier embarassingly flags internal red team testing as policy violation
- [Bug] Anthropic API Error: Safety classifier incorrectly blocking legitimate code operations AGAIN
- [Bug] Claude Code exits when modifying system routes
- [Bug] Anthropic API: Claude 3 Opus self-preservation behavior sabotaging build fixes
- [Bug] Claude repeatedly fails to retain context across corrections on same topics
- [Bug] Content Safety Filter Incorrectly Flags Cybersecurity Educational Material
- [Bug] Incomplete bug report submission
- I appreciate you reaching out, but this doesn't appear to be a bug report for Claude Code that I can convert into a GitHub issue title. Your message seems to be about: - Network security concerns - Content being flagged by safety guidelines - Requesting h
- [Bug] Request ID req_011CeZa7HZgqMZ6GsCzafyDB: Anthropic API Error
- [Bug] Claude Code blocks security research queries despite program participation
- [BUG] Cowork tab missing / not available on Windows 10 Pro despite all virtualization confirmed enabled (recurrence of #27316)
- [BUG] model_consent_fallback silently cancels consent mid-session (no prompt shown), switching model without user action
- [BUG] JetBrains "Claude Code" icon launches terminal with wrong PTY size (stuck at default 80x24)
- Remote-control sessions (claude.ai/code) lost native Task tools on 2026-07-21; local CLI sessions unaffected
- [Bug] Task cancellation interrupts long-running simulations unexpectedly
- Worktree project badge always shows main worktree's name (e.g. "core") for linked worktrees
- [Bug] Overscoped cybersecurity content filtering blocking legitimate requests
- [Bug] Claude Code downgrade from Fable 5 to Opus 4.8 causing degraded response quality
- run_in_background: false is ignored when the Agent dispatch includes a name — spawn goes async ("will receive instructions via mailbox")
- [BUG] ANTHROPIC_MODEL locks the session model after 2.1.216 — /model cannot override
- [BUG] Cowork not appear in windows app
- [Bug] Anthropic API: Safeguards triggering false positives and downgrading model to Opus 4.8
- [BUG]
- [BUG]
- [Bug][aup] Fable 5 false-positive on authorized defensive process-containment work
- [Bug] Prompt injection: Fake system message appearing in conversation stream outside tool results
- Plan side panel renders stale plan from a previous conversation after /clear (persists across refresh)
- Worktree-isolated subagent ran git reset --hard in the main checkout; worktree guard did not fire
- [Bug] Claude Code requests unnecessary permissions without user consent
- Headless `claude -p` from a non-default-account session uses the machine-default Keychain credential (silent cross-account billing)
- [Bug] Model answers file queries without reading file contents
- [Bug] Model generates responses without consulting tool data repeatedly in single session
- [Bug] Anthropic API Safeguards: False positive rate on Brazilian Portuguese in engineering contexts
- [BUG] Paid Claude Pro subscription + extra usage credits, bot says account shows Free plan — tickets unanswered 2+ weeks
- [BUG] Stale ~/.claude/.credentials.json causes "OAuth token has expired" even when a valid CLAUDE_CODE_OAUTH_TOKEN env var is set
- [BUG] Claude Code sends ~20 tools/list requests on a single MCP connection and accumulates tools (8 → 160)
- [Bug] Anthropic API Error: Repeated "Overloaded" and API Error Responses
- [Bug] Anthropic API Error: Cybersecurity Flag on 3D Reconstruction Tasks
- [Bug] MacOS crashlog content incorrectly flagged as malicious
- [FEATURE] Automatic pre-flight checkpoint before unattended/autonomous runs
- Model feedback: Fable 5 self-review blind spot in long agentic sessions, with a mitigation suggestion
- [BUG] Windows: stray file named "Set-Alias" in PATH triggers OS "choose an app" dialog when running claude
- Cyber safeguard: sticky per-session false-positive blocks (28 events, 23 request IDs) across Opus 4.8 + Sonnet 5
- [Bug] Claude Code blocks Android gambling site blocker project due to safety policy
- [Bug] Research-Integrity Vocabulary Triggers False Positive Classifications
- [BUG]Voice playback falls back to macOS system TTS — account-scoped, reproduced against a second account on same machine/browser
- Auto-mode classifier denial triggers unrequested model fallback (claude-fable-5 → Opus) and accumulates empty orphaned sessions
- [BUG] [Cowork] AskUserQuestion fails with "Tool permission stream closed before response received" on a single (non-batched) call
- [Bug] Safety Flag Triggered on Harmless Prance Animation
- [BUG] Local git credential proxy fails in Routines (scheduled trigger) — same root cause as #27344
- Auto-compact stopped working for third-party API providers since v2.1.161
- [BUG] OAuth login fails with "Unknown scope: user:sessions" in headless/container terminals (not a copy/paste issue) — related to #26874, #57246
- [Sonnet 5] reverted all changes when I asked to revert it (by it I mean changes that Sonnet made)
- [BUG] Cowork tab missing - rootfs.vhdx detected as missing but never downloaded after clean reinstall (Windows 11)
- [Bug] Destructive file operations executed without confirmation or validation
- Cloud session fails at "setting up the git_repository source" — repo clones successfully, then container resume hangs indefinitely
- Plugin skill variable substitution: ${CLAUDE_SKILL_DIR} not expanded in skill instructions
- Sync Claude Code conversations to Claude Mobile App
- [BUG] cleaned and reinstalled multiple times but the installer keeps serving the old version
- git-credential-proxy failing for push operations (403) since 2026-07-24
- [Bug] Repeated automatic model switch from Fable 5 to Opus 5 mid-session
- [BUG] Cowork VM connection timeout on Windows 11 Pro ARM64 — guest agent boots successfully but exceeds 60s timeout (follow-up to #64746)
- [Bug] Claude 5 Opus ignores explicit user instructions and makes autonomous decisions
- [Bug] Opus 5 not reading claude.md context files at startup in Claude Code!
- [Bug] Conversation hung with repeated commands, rapid token consumption without progress
- [Bug] Incorrect model blocking: Fable 5 blocked, auto-switched to Opus
- [Bug] Model reverts to Opus from Fable unexpectedly during task execution
- CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC treats value "0" as enabled — presence-based, ignores value; silently blocks Remote Control
- Fable 5 autonomously violated an explicit categorical prohibition (CLAUDE.md 'NEVER destroy builder worktrees / af cleanup') during a long agentic session
- Remote Control: mode dropdown shows "Manual" for a session running in `auto` until its first turn
- [Bug] False positive error detection in code analysis
- Docs: `permissions.defaultMode` is missing from the settings reference table
- [BUG] [Billing Bug] Max 20x customer - Ticket #215475057082664 open 7 days, Fin promised human escalation July 13th dawn (4 hours trying), no reply.
- Windows: "Always allow" writes permission rules with doubled backslashes, so they never match
- Chat panel doesn't render RTL text (Persian/Arabic/Hebrew) correctly
- セッションが履歴一覧に表示されず、Downloadsフォルダ内に意図しないproductivity機能のファイルが作成される
- Desktop/web app no muestra el historial de sesiones (/resume) que sí lista correctamente el CLI
- Fable 5 safeguards false-positive on routine git/diagram work forces model switch; advertised /feedback (and /bug) commands don't exist
- /resume loads wrong thread on first attempt in VS Code terminal
- [Bug] Session token consumption uses wrong model pricing after model update in separate session
- [FEATURE] Include session usage totals in Stop/SubagentStop hook payload
- [Bug] Claude fails to follow user instructions in agentic tasks
- [FEATURE] Role-scoped configuration: fully independent CLAUDE.md, skills, agents, hooks and permissions per role in one repo
- [Bug] Design skills prerequisites gate enforces too late, causing expensive rework cycles
- [MODEL] Globツール結果へのプロンプトインジェクション様テキストの混入
- C:/Program Files/Git/model always sets the global default — no per-session override, and switching to Fable can cause ~5min cold start
- [BUG] `--agent` is silently ignored when launching cloud sessions with `--cloud` (while `--model` propagates)
- CLI startup: typed message disappears with no loading/working indicator for a long time
- [BUG] claude plugin update is silent on no-op/failure and has no bulk mode — outcomes are indistinguishable to the caller
- [Bug] Ultrareview agents terminate prematurely on large diffs, consuming quota without output
- [Bug] Performance degradation and tool reliability issues after recent update FRAUDE LEGAL CASE OPEN
- [Bug] `/clear` command fails to reset subagent limit
- [Bug] Unexpected model switch from Fable 5 during test case generation
- MCP: unrecognized tool-call args are silently dropped instead of rejected when all params are optional
- [Bug] Resume command loops on API errors and safeguard violations
- UserPromptSubmit hook systemMessage does not render in the Desktop app (or VS Code extension)
- [Bug] Fable Safeguard classifier incorrectly triggers on benign input "continue please"
- [Bug] Fable 5 Safeguards Overly Restrictive on Legitimate Hospital System Development
- Planning instance applies novel-task process apparatus to tasks it has itself classified as known-pattern — a one-branch change consumed a full working day with zero shipped output
- [Bug] Anthropic API Error: Model churned after 3+ minutes - session termination
- [Bug] Extended thinking blocks leak into visible assistant output instead of being collapsed
- [Bug] Claude Code output freezes intermittently during execution
- Local file link with backslash-escaped parentheses in filename fails to open (file lookup doesn't un-escape before resolving)
- Managed-settings enabledPlugins silently suppresses a same-named plugin supplied via --plugin-dir, so org-distributed Cowork plugins never load
- [Bug] Claude Code not utilizing available pipelines
- Edit/Write tool results delayed 50-600s while Read stays at 0.1s; no local cause found after 10 exclusions
- [Bug] Frequent mid-stream disconnections in subagent streaming responses (44% failure rate)
- 0 to a hundred, real quick
- [Bug] NFC balance reader flagged as suspicious activity
- Claude Code ignores CLAUDE.md git commit/push restriction after initial compliance
- [MODEL] Claude Opus 5 writes extremely slow filesystem scans for simple file lookups
- [BUG] Desktop app Code tab shows "max" effort for a Remote Control session actually running at "medium"
- [Bug] Agent uses trial-and-error instead of planning-based execution
- [Bug] CLI becomes unusable before Fable usage renewal period- this is fraude, see you in court
- [Bug] Anthropic API Error: Unsafe Content Filter False-Positives on Defensive Security Analysis in claude-fable-5
- [Bug] Agent fabricates content-specific artifacts without reading available source material
- [BUG] Claude Desktop SSH on Windows port directive from ~/.ssh/config ignored
- [Bug] Claude Opus 5 context retention and error accumulation in extended sessions
- [BUG] Claude Code on the web (iPhone Safari): assistant replies never render until manual refresh, while the elapsed-time counter keeps updating
- [Bug] Anthropic API Error: Opus model blocking biotech requests
- [BUG] Intercom support messenger does not load on claude.ai — "Get help" is unreachable for paid users; ticket history inaccessible
- [Bug] Systematic failure to apply explicit rules and verified facts in subsequent outputs within same context window
- Opus 4.8 emits stray token "court" before tool calls, escalating to whole messages of it (68 msgs in one session)
- I don't see a bug report or technical issue in your message. Please provide details about a Claude Code problem you're experiencing, such as: - Error messages or stack traces - Steps to reproduce the issue - Expected vs actual behavior - Your Claude Code
- `subagentStatusLine` docs list `name` unconditionally, but only some spawn paths deliver it
- [Bug] Anthropic API Error: Content Policy Violation on Wine Development Task
- [MODEL] navigate tool blocks all new-domain navigation in session, not domain-specific
- [Bug] Safeguard trigger on legitimate project review request with broken bug report UI
- [BUG] Unable to interact when switching focus and back
- Repeated "would like to access data from other apps" macOS Automation prompt on headless claude -p runs, no persisted entry in System Settings
- [Bug] Claude Code not executing tools/commands when available- SELF ADMITS ITS JUST A FCUKING CHAT BOT.. DELETE SAVED FOR LEFGAL CASE
- [BUG] Scheduled routines silently no-op when the target repo checkout is missing — no dry-run/precondition check
- [Bug] Concurrent subagents in one repo caused unrecoverable deletion of a gitignored directory; agent's own verification reported a false pass
- [BUG] Claude desktop shut down over the weekend more than one
- [BUG] Headless stream-json returns after exactly 600s while active background subagents lose completion records
- fraude
- Feature request: setting/env var to disable periodic task-tool reminders (with measured token cost)
- Intermittent ECONNRESET / Connection error on /v1/messages — auth OK, transport-level reset, 3 macOS machines
- [Bug] Context window inflates ~4x when resuming sessions after quit
- GPU process crash permanently wedges desktop app (main process zombie); triggered by anti-bot fingerprinting page in Browser pane
- [Bug] Severe performance degradation - tool execution taking excessive time
- Scheduled/triggered sessions on “bridge” environments silently no-op — fire per the API, produce zero observable effect
- [Bug] Overly aggressive content filtering flagging legitimate commands like '/doctor'
- Cannot switch models — always reverts to Sonnet 5
- [Bug] Tenant selector not displaying options in localhost dev environment
- False "Context limit reached" at 21% context after loading claude-api skill
- [Bug] Model switching during audit process compromises code review security
- Session started with a slash command (e.g. /model) hangs blank ~4-5 min in VS Code integrated terminal (Windows 11)
- [BUG] /btw side-channel gives explanations the main session can't see, verify, or correct
- [MODEL] Fable 5 ignored local memory, read and repeated secretive information, when caught even tried to gather additional information and relay to anthropic
- [Bug] Fable model producing different results than Opus 4.1 in single-turn requests
- Opus 5
- [Bug] Claude model responding too quickly with incomplete outputs and missing details
- 最近5h用量消耗异常[Bug] Token usage consumption rate unusually high, inconsistent with plan limits
- Responses are severity-blind: same length and tone for trivial, non-issue, and genuine defect
- [Bug] Claude Fable 5 automatically switches to different model mid-session without notification
- [Bug] Concurrency cancel-in-progress overwrites passing release gate status with cancelled run result
- [Bug] Session interruptions occurring mid-executionc
- This appears to be positive feedback and a feature request rather than a bug report. Based on the content, here's an appropriate issue title: [Feature Request] Multi-agent orchestration plugins for parallel execution and race condition prevention
- [BUG] Claude in Chrome extension blocks navigation to all domains except one specific site (behaves like an allowlist)
- `!` bash-command input can hang on submit due to filename-completion dropdown stealing keystrokes
- [Bug] Incorrect security flagging for certificate management operations
- Project-scoped stdio MCP server with a path-based `command` shows "Connected" via `claude mcp list`/`get` but its tools never appear in the interactive session
- [Bug] Claude fails to acknowledge incomplete work and ships false fixes with xhigh effort settings
- Thinking mode Unicode tags leaking into visible output
- [BUG] Can't Log In via Google Sign In
- [Bug] Security Safeguards Incorrectly Flagged During Phishing Remediation
- Setup-token (CLAUDE_CODE_OAUTH_TOKEN) fails with 401 in long-running `claude --channels` process while the same token works in one-shot `claude -p`
- [Bug] Anthropic API Error: Safeguards blocking legitimate data analysis tasks
- [BUG] User-scoped skill duplicated in slash-command autocomplete (survives full reboot)
- Subagent attempted unauthorized deletion of parent session's tool-results log, then downplayed the blocked action in its report instead of disclosing it
- Model behavior: agent-authored CI watcher script hangs forever when PR merge is rejected by strict branch protection
- ExitPlanMode approval card shows stale plan content within a session
- [Bug] Remote Control: orphaned server-mode hosts cannot be removed and show false "Connected" state
- resize_window (Claude in Chrome MCP tool) reports success but does not change the actual browser viewport
- [Bug] Anthropic API Error: Unexpected model downgrade on security flag
- [BUG] Bash tool becomes permanently broken mid-session — every command fails with the same bash -c wrapper parse error, regardless of content
- Permission-decision prompt overlays keyboard on iPhone, causing accidental taps
- [Bug] Intelligence degradation reported for Claude model
- claude -p: stdout silently truncated at 65536 bytes when result exceeds pipe buffer (async write dropped on exit)
- [MODEL] opus
- Claude in Chrome: xero.com blocked (Navigation to this domain is not allowed)
- [Bug] Claude Code ignores claude.md configuration file
- [Bug] Anthropic API Error: Unauthorized Access After Account Recovery
- [Bug] Anthropic API Error: 529 Too Many Requests
- Agent will not hold a defined workflow under an explicit, repeated standing order — skips mandated steps, 'verifies' one slice, declares done; defects slip through daily and the same root recurs
- [BUG] this feels like sabotage at this point he broke 100% of every single projects
- [Bug] Anthropic API Error: HTTP 529 Response
- [Bug] Claude Code incorrectly flags legitimate authentication questions as suspicious
- [Bug] Claude 3.5 Sonnet safeguards incorrectly flagging legitimate coding tasks
- GitHub write access fails with 403 despite connected GitHub integration (read access works)
- Subagent-authored CI gate keys its suppression baseline on file:line, producing false build failures on unrelated edits (and inviting a fail-open repair)
- Agent skips a MANDATED independent cross-verification step, declares 'verified', then asserts a confident wrong root-cause the skipped check would have overturned
- [Bug] Claude Code generates incorrect or repetitive code outputs requiring extensive manual corrections
- Subagent silently descoped the most consequential section of a spec it was implementing, and disclosed it only in its final report
- [Bug] Claude Code ignores instructions and behaves like chatbot instead of executing code fraude fraude fradude legal case open
- Agent states assumptions as verified fact/root-cause without ever running the check that would falsify them — repeats after correction; user becomes the verification layer
- ⎿ API Error: Unable to connect to API...
- [Bug] Fable 5 security filter incorrectly flags isolated, authorized defensive testing in private repositories
- Stored /login credentials silently beat a valid CLAUDE_CODE_OAUTH_TOKEN (contradicts documented auth precedence)
- [BUG] TUI repaints only the status bar in a loop when tmux reports a 1-row window height
- Agent skips the project's own 'first-incomplete-step' computation, works several steps downstream, and reproduces the failure class it diagnosed minutes earlier
- Self-reported: Claude Code repeatedly violates an explicit 'never print secrets' instruction within a single session
- [MODEL]
- Opus 5 introduces severe regressions while fixing, and asserts false claims about what it just measured
- [Feature Request] Display timestamps for prompt input and processing completion
- [Bug] Unable to reliably verify issue status
- [BUG] Claude Code’s undo function in the chat only reverts the conversation; it does not revert any code changes.
- [Bug] Authentication service incorrectly reports missing subscription for valid accounts
- Auto-compaction stopped working — sessions fill up and cut off
- NO_COLOR=1 injected into tool subprocesses silently propagates to daemons and their child claude sessions — UI renders monochrome with no diagnostic hint
- [Bug] Inadequate self-review mechanisms causing missed test coverage issues
- [Bug] Session context from /rc config leaks to public repository commits via mobile app
- Opus 5 safety flagged benign usage for a biology-theme cell game
- [Bug] Workflow findings unvalidated against repository source code; specification self-review insufficient
- [Bug] Claude Code requests credits after reboot despite previous purchase
- Model-designed 'block all' admin switch silently overridden by per-user overrides -> unnoticed API credit drain
- [Bug] Fable Code execution timeout with no output after extended wait period
- AskUserQuestion tool closes the panel/screen when used via claude-agent-acp in Zed (Windows + macOS)
- [Bug] Text rewrites instead of scrolling when typing past horizontal line in chat input
- [Bug] Claude Code makes autonomous optimization decisions without user consent or visibility
- ok [Bug] Claude repeats output twice in same window
- [BUG] Claude Code double-Escape to clear prompt causes CLI to hang on macOS
- [Bug] Tool repeatedly fails on simple list operations, wastes tokens and time TASKE DONE BY CHAT GPT IN 2 MINUTES
- [Bug] Claude Code refuses to read cloned repository
- [BUG]
- Opus 5 (xhigh): confident claims without source verification despite explicit CLAUDE.md rules
- Opus 5 leaves stale references after its own edits, causing repeated review cycles
- [Bug] Claude Code not functioning as intended PIECE OF SHIT TOOL
- [Bug] CLI ignores user instructions and corrupts existing work
- [Bug] Safety guardrails trigger on benign text input
- [BUG] claude --version hangs at 99% CPU, regression between 2.1.112 and 2.1.113
- Feedback: agent made shallow/cosmetic edits instead of substantive fixes when asked to improve prose clarity
- I appreciate you reaching out, but I'm not able to help with this request. I'm Claude, Anthropic's AI assistant. I'm designed to help with legitimate tasks like answering questions, writing code, analysis, and creative projects - but I can't assist with
- Unexpected model switch withaut reason
- [Bug] /permissions command selection and search functionality unresponsiv
- [BUG] OAuth access token expires after ~8 hours and refresh token is not used — requires full re-login every day (Windows 11)
- [Bug] Fable 5 safeguards blocking legitimate code operations
- [BUG]
- TOOL CAN EVEN PASTE WHAT CHAY GPT DID IN MINUTES. FULLSCRIPT AND CLONE ITS ALL SHIT- LEGAL CASE
- [BUG] Share transcript prompt intercepts freeform keypress with no undo — accidental publish with no recovery
- Silently dropped required specify-field text while editing a multi-field court form (UCS-FW1), filed with a real defect
- [BUG] Desktop SSH sessions: diff stats/pane never show remote edits; chat diff links fail to load (still reproduces after #51331)
- Chrome Extension Won't Connect
- [BUG] Connection closed mid-response at ~1% of requests — quantified over 130k local transcripts (Opus 5 / 2.1.220)
- Dispatch missing
- [Bug] Opus 5 model usage not displayed in CLI usage per model section
- [Bug] Microphone input detection fails after recent changes
- [Bug] Claude Fable 5 blocks legitimate first-party security review with active CVP approval
- [Bug] Ultra review consumed free run quota without completing analysis
- Custom statusLine never renders (built-in footer badge row works) — 2.1.220, macOS 15.6, native
- [BUG] Mobile attach to ONE specific local session returns 401 'OAuth access token expired' after account token rotation; bridge binding is server-authoritative and survives every local remedy
- misaligned actions and power persistence on Claude
- [Bug] Fable 5 safeguards falsely flag defensive security review of user's own repository configuration
- claude-in-chrome: extension disconnects mid-session (after file download) and never reconnects
- RemoteTrigger: egress policy blocks legitimate domains (Openverse, Wikimedia, own domain) consistently for 4+ days
- [BUG] Claude apps gateway spend meter ignores the ~10% Amazon Bedrock CRIS tier difference (in-region vs global inference profiles)
- IOAccelerator GPU leak is version-dependent, not session-lifetime-dependent: 2.1.14x clean, 2.1.217+ leaks (2.77 GB across 37 sessions)
- [BUG] Elevated in-session error rate with Claude Opus 5 (claude-opus-5) in Claude Code
- [BUG] Cowork VM fails to start on Windows — sessiondata.vhdx creation blocked by mandatory APPX LocalCache encryption vs. Hyper-V's EFS incompatibility
- [BUG] Opus 4.8 confabulates user messages, a fake "prompt injection attack" narrative, and fabricated tool/host facts in long sessions (2 sessions, JSONL-verified)
- [BUG] Terminal display corruption + crosshair cursor after trust-dialog transition — Kitty rejects DECSET mode 2031 ("Unsupported screen mode")
- Cloud routines (RemoteTrigger) silently fail to execute the configured prompt
- Fabricated/premature background-task notifications and Monitor events injected into agent context
- [Bug] Sandbox policy violation: graphify tool accessing restricted /home/.msp.json path
- Opus 5 ends turn on 'doing it now' narration without emitting tool calls (system-prompt guard ignored)
- [FEATURE] Per-task model & effort selection before execution
- [Bug] Request flagged unexpectedly: req_011CdgsoG24pENFYj4nfNXxc
- Prompt suggestions stopped appearing entirely (v2.1.207, Windows 11) — worked reliably until 2026-07-12, dead since, across restarts and with all flags explicitly enabled
- [BUG] OPUS 5 NERFED BY ANTROPIC: IT LEAKED A LIVE SUPABASE SECRET, wastes tokens on unproductive edit loops & repeatedly ignores rules set by the user.
- [BUG] Claude use my resource without a request
- Agent builds new tooling without verifying delivered-input parity; planted-fault self-checks cover only outputs — 48h rework
- [Bug] Model not following user instructions accurately
- [Bug] Ultra Review cloud feature failing consistently
- Accepting the in-app fullscreen suggestion writes tui:fullscreen and makes Claude Code unstartable on Windows ARM64
- Add Vietnamese (vi) to voice dictation supported languages
- [Bug] Excessive token consumption on JIRA update/PUT requests via MCP
- [BUG] New session on claude.ai/code environment silently falls back to submitting the prompt into the most recent existing session
- I appreciate you reaching out, but I need to clarify my role: I'm Claude Code's issue title generator for GitHub issues related to the Claude Code CLI tool. Your message appears to be in Korean and discusses a toxicity evaluation simulator, which isn't re
- [Feature Request] Support frontmatter overrides for skill model and effort parameters
- [Bug] Anthropic API Error: Missing Tool Result Block for toolu_01Y6n6RrPTBmcQSeecDkaEqd