Digest / August 23, 2026
Claude Code Issues Digest — August 23, 2026
183 new issues filed ·
135 resolved in anthropics/claude-code (UTC day).
New issues (183)
- AskUserQuestion dialog: top-right X cancels instead of minimizing, hiding the conversation needed to answer
- Remote Control has no way to complete re-login (OAuth) when the session's auth token expires
- Unproductive/below-bar session: root-cause misdiagnosis + agreed specs delivered at a fraction of scope
- Extension host pins one CPU core for hours, log-silent - measured as a native file-watcher busy-loop (~245k other-I/O ops/s, 0 reads/s) on Windows
- Idle sessions burn 0.3-5% CPU because timers keep triggering garbage collection
- [Bug] False positive security flag for local-to-VM code transfer
- [FEATURE] Allow subagents to target a local OpenAI-compatible endpoint for mechanically-verifiable subtasks
- [BUG] Agent frontmatter tools: constraint not enforced — read-only agents can write files
- Scheduled tasks leak their process pair: work completes with end_turn but the session never exits (~16 GB/day)
- [BUG] "Always Allow" writes a rule that cannot satisfy the gate that blocked the command, so an identical command re-prompts forever (shell redirect outside cwd)
- [BUG] Stale TUI frame remains on screen after /exit until `clear`/`reset` (Konsole, xterm-256color)
- [BUG] CoworkVMService never completes VM creation — stuck in silent connect/disconnect loop (device_bash / Workspace unavailable)
- [BUG] Max users can't use the --bare mode
- [Bug] Auto mode avoids Read/Write tools causing file operation failures on Windows
- I appreciate you reaching out, but I'm having difficulty understanding the specific technical issue you're reporting. The description doesn't clearly identify a bug or feature request related to Claude Code's functionality. Could you please provide more d
- Core, manually-verified decision silently defeated by adjacent code — Sonnet 5 missed it under direct review
- [Bug][cyber] Troubleshooting missing files and inspecting cloud storage token scopes (req_011CeJtwD8fGdjksmkVkCi9A)
- [Bug][cyber] Safeguard flagged verifying citations and vetting project documentation (req_011CeJtsmp5JnvJMce5brq9W)
- [Bug] Anthropic API Error: Safeguards flagged benign message with Opus 5
- [Bug] Anthropic API Error: Inappropriate Content Filter False Positive on Frontend Task
- [FEATURE] - Show terminal diffs even when an IDE integration is handling approval
- [BUG] Project-skill registry stays rooted at the launch checkout after EnterWorktree — a committed SKILL.md in the entered worktree is permanently "Unknown skill" (2.1.241)
- In `--channels` sessions, the reply tool is intermittently never called despite the turn completing with a real, coherent response
- [BUG] Settings UI only shows 1 skill despite 56 skills installed and functional
- [Feature Request] Add option to disable automatic message compaction
- [Bug] Memory mechanism unreliability causing persistent context loss over extended sessions
- Absolute-path file links render as clickable but never resolve; only file:// URIs work
- Screenshots return black/masked images on Windows — screenshotFiltering: mask blocks all computer-use
- I don't have a bug report to analyze. Please provide the bug report details so I can generate an appropriate GitHub issue title.
- [FEATURE] Configurable peer-registration session name (per-project prefix) for cross-session messaging
- [Bug] stdio MCP server disappears ~53s after reconnect — no exit code, no signal, no exception (related to #40207/#86711)
- [Bug] Anthropic API Error: False Positive Content Policy Flag on OAuth Code
- Stop/interrupt kills in-flight background agents — should interrupt the turn, not the workers
- Chrome extension: computer screenshot with save_to_disk: true reports success but writes no file
- [BUG] Cowork Dispatch send_message no longer delivers to existing Cowork sessions (worked 461 times until Aug 10)
- [BUG] Path-scoped rules never match outside the project root, making the auto-memory directory unreachable
- [BUG] Fullscreen TUI in VS Code terminal: mouse drag-select works only once, next selection requires switching terminal tabs
- Frequent forced re-login when many concurrent sessions share ~/.claude (refresh-token rotation race?)
- I don't have enough information to generate a GitHub issue title. Your message describes a general development question about RL (Reinforcement Learning) for games and "fable" flagging, but this doesn't appear to be related to Claude Code, Anthropic's CLI
- [BUG] Claude Fable interprets info given as meaning "low effort"
- [Bug] Worktree isolation pin shared across parallel subagents causes cross-agent command execution
- remoteControlAtStartup silently no-ops when OAuth is unhealthy at session start — never re-evaluated, session invisible to ListAgents for its whole lifetime
- Windows paths printed outside code blocks lose backslashes (`\_` eaten as a markdown escape)
- [BUG] Claude in Chrome: extension icon opens blank panel instead of resuming active Cowork session
- [BUG] Assistant fabricates synthetic `user` turns mid-session and keeps acting on them (58 occurrences / 9 sessions, with an exposure denominator)
- [Bug] Safeguard false positive flags legitimate AI-safety research workflow with valid cybersecurity authorization
- Feature request: make Remote account-wide by default (all local chats visible, like Codex)
- [Bug] /fork copies malformed advisor history, reports working, then child immediately 400s
- [BUG] Allowed sites list creates duplicate entries for the same origin
- Foreign third-party web-page content streamed into an in-flight assistant message (same requestId, +81s after reply ended)
- Claude in Chrome: extension connects but all page-interaction tools time out waiting for document_idle
- [BUG] CVP stuck "In review" 10+ days on active Max org (backend shows it as Free); Console org auto-denies instantly by plan status, not application — previously approved, support unresponsive
- Windows MSIX: auto-start CoworkVMService races package servicing -> 0x80073D02, app won't launch until repeated Repair
- Feature request: per-server lazy/deferred MCP connection in .mcp.json (with measured cost data)
- [FEATURE] AskUserQuestion is not available as a tool in routine (scheduled/triggered) cloud sessions
- [FEATURE] Let Claude open the Browser pane, or report its visibility state
- [Bug] SSH authentication failing or access restrictions preventing VPS login
- Agent isolation: "worktree" branches from a stale commit, not session HEAD
- [Bug] AI agent deletes entire repository without git commit history preservation
- Project .claude/agents partially discovered in git-worktree sessions (still broken in 2.1.241; #25816 closed but not fixed)
- Cross-session `send_message` has no working return path — sender's `local_<uuid>` id is unreachable from the `ListAgents`/`SendMessage` namespace
- [BUG] Claude Desktop (Windows/MSIX): in-app browser preview triggers a GPU process crash (exitCode 101457950) that kills the app and leaves the MSIX package needing remediation; Windows repair loops with no effect
- Auto mode's permission gate for dangerouslyDisableSandbox is blind to the flag itself
- [BUG] Cannot scroll through console output or conversation history
- [Feature Request] Support for model selection beyond Opus in security agent workflows
- [BUG] Cowork: images attached to a chat message are silently dropped — 2 of 3 never reach the agent, no failure signal either side
- [BUG] Cowork chat strips non-http(s) link hrefs, and "copy link address" then returns the page's own URL
- Claude Code panel stays blank in GitHub Codespaces — concurrent session spawns deadlock on .claude.json lock
- [BUG] Session limit hits 100% despite minimal usage and low token consumption
- [Feature Request] Improve reasoning depth and reduce excessive proactive suggestions
- [BUG] remote-control rejects all of its own flags when any global flag precedes the subcommand
- Bash tool spawns a login shell whose state it then discards — ~830ms per call
- Scheduled task sessions (Desktop local agent mode) never exit after completing; processes accumulate until the host runs out of memory
- [BUG] Queued image prompts are re-serialized into the session transcript 16-23 times, inflating .jsonl by GBs
- [MODEL] Uses PowerShell here-string (@'...'@) inside the Bash tool on Windows, silently corrupting a git commit message
- Auto-compaction leaves the session on Haiku 4.5 on Linux (`opus[1m]`, Max 20x) — silent, and it degraded real work
- [Feature Request] Allow disabling safeguards for authorized security testing
- Task/general-purpose subagent fabricates "verbatim" file quote instead of returning actual file content
- [BUG] Filesystem MCP server (remote-devices) fails all tool calls: outputSchema declares unsupported draft-07 dialect
- [BUG] [Cowork] Sessions can't read or update their own published Artifacts by default (cloud_default egress blocks *.frame.claudeusercontent.com)
- Cross-session message bodies are HTML-escaped by Desktop and stamped `encoded="1"`, but the CLI never decodes — `&`, `<`, `>` arrive corrupted
- [BUG] Plugin marketplace auto-update runs synchronously inside PreToolUse hook path, causing multi-minute session hangs under concurrent sessions (Windows)
- VS Code extension: renaming a session via the sidebar pencil icon doesn't sync with /rename / cross-session messaging name
- [BUG] Workflow subagents interrupted at ~180s while a model request is still in flight; identical-key retries restart cold and never complete
- Cross-session message bodies arrive HTML-entity-encoded: Desktop escapes and stamps encoded=1, the CLI never decodes
- [BUG]
- [BUG] Background job id from a dead daemon epoch is `claimed-spare` 30 days later — stale transcript resumed into the foreground, and its ai-title grafted onto the post-`/clear` session
- Cloud routines: Artifact publish raises an interactive permission prompt, so unattended runs hang in requires_action
- [BUG] Remote session hangs on "Starting session…" — server logs connection, then 50–118s of silence, then SIGTERM and reconnect loop
- Desktop "Instructions for Claude" silently does not apply to Claude Code sessions — the text is never written to disk, so CC cannot read it
- cowork-dragdrop-issue (Similar to closed #59311 issue)
- [BUG] Incoherent path completion options produced by '@'
- Feature request: auto-enable --ax-screen-reader for screen reader users
- [Bug] Claude 5 Opus exhibits high error rate and poor task completion quality
- [Feature Request] Add conversation rewind capability to revise responses at arbitrary points
- [Bug] Anthropic API Error: Content policy violation flagged by safeguards
- [Bug] Anthropic API Error: Model downgrade during active session
- [Feature] Compact a running background session from outside (no external trigger exists)
- [Bug] Subagent transcript over SKIP_PRECOMPACT_THRESHOLD (5 MiB) cannot be resumed: 'No transcript found for agent ID'
- STATE.md corruption: stale file-history snapshot re-injected into working tree at context/compaction boundaries
- [Feature Request] Add network device inspection capabilities for local network debugging
- Pending message lost when leaving a chat (non-terminal client)
- [Bug] Sandbox denies write access to $CLAUDE_JOB_DIR/tmp for background agents
- [FEATURE] Show a last-used date for each agent in the /agents view
- Agent ran an unrequested destructive command against a production host, and reported a failed deploy as successful
- [BUG] Pull requests from four unrelated third-party repositories appear in a local desktop session's state (prs[]) — and are rewritten after deletion
- [BUG] Claude Desktop (Windows/MSIX): fatal GPU-process crash (exit 101457950 / 0x060C201E) on preview surfaces — --disable-gpu-compositing does NOT mitigate; package left Modified/NeedsRemediation
- [BUG] Different usage for the same task in CC vscode extension & CC Desktop (linux beta)
- [BUG] (Link-Klick → GitHub-Login-Klick → Absturz, reproduzierbar 2-3x
- Artifact publish auto-arms a monitor_ws live-updates watcher with no returned handle and no opt-out
- Workflow tool: stamp journal.jsonl records with a writer-side timestamp (ts)
- [Bug] Grammar Error: "how Claude did" feedback uses "good" instead of "well"
- [Bug] Code output not displaying in terminal
- I appreciate you reaching out, but I need to clarify: I'm Claude Code, Anthropic's CLI tool for coding assistance. I don't have access to: - Your conversation history or flagging decisions - Your account status or model tier - Conversation moderation syst
- Status bar shows Chinese text "致性检查" instead of English
- I need more specific information to generate an accurate GitHub issue title. Could you provide details about: 1. What specific error message or behavior are you encountering? 2. At what point in the Claude Code workflow are you getting blocked? 3. What is
- [BUG] hookify rule files containing non-ASCII are silently skipped on Windows (open() without encoding)
- [BUG] VSCode extension streaming errors: content_block_delta and message_stop without a current message (regression after v2.1.235)
- I appreciate you reaching out, but I need to clarify a few things: 1. **This appears to be a support request rather than a Claude Code bug report** - Claude Code is Anthropic's CLI tool for coding tasks. Account/billing issues should be directed to Anthro
- [Bug] False positive safety flag on cybersecurity website design tasks
- [BUG] Over mosh, copy-on-select only works the first time a given string is copied: identical OSC 52 is deduped by mosh state sync while the TUI reports "sent N chars via OSC 52"
- [Bug] Anthropic API Error: Content filtered by safeguards policy
- [Bug] Opus 5 model performance degradation reports
- Claude Code v2.1.241 native binary crashes with SIGILL on CPUs without SSE4.1
- [Bug] False positive security classification triggering mid-session model downgrade during adversarial code review
- I don't see a bug report in your message - you've only provided what appears to be a request ID. Could you please share the actual bug report details so I can generate an appropriate GitHub issue title?
- I don't have access to the actual bug report content associated with that request ID. To generate an appropriate GitHub issue title, I would need you to provide the bug report details, error messages, or description of the problem you're experiencing with
- [Bug] Git operations (pull, fetch, branch list) not functioning properly
- I don't have access to the specific bug report referenced by that request ID. To generate an accurate GitHub issue title, I would need you to provide the actual bug report details, error messages, or description of the problem you're experiencing with Clau
- I don't have access to the specific bug report details for request ID `req_011CeL9TRK7S1K9qP1ttfcWL`. To generate an appropriate GitHub issue title, please provide: 1. A description of the bug or problem encountered 2. Any error messages or logs 3. The
- /compact silently fails to apply on very large conversations: summary generated, boundary never written, context unchanged
- [Bug] Increased hallucination rate in model responses
- [BUG] SendMessage-resumed subagent's completion notification routes to the session root, not to the resumer
- [BUG] Subagents cannot reply to inter-agent messages: incoming from= is the agent type ('general-purpose'), which is unroutable
- [BUG] Completion notifications unreliable for foreground commands auto-backgrounded at the 10-minute cap (inside subagents)
- [BUG] Deliberately backgrounded long-lived tasks armed by subagents are killed early with no notification to the owning subagent
- mcp__claude-in-chrome resize_window reports success but does not change viewport size
- Add a "Copy response" quick-action button to message actions
- [BUG] Voice input (mic) misrecognizes Korean speech in VS Code extension
- Expose the resolved tool registry so a configuration can detect when a tool's default availability changes
- [Bug] Model unexpectedly switched to Claude 3.5 Opus despite Fable subscription
- hit safeguards while making a programm
- [Bug] Trusted devices UI: device rows are not identifiable (app, timestamp, IP all missing)
- [BUG] ElevenLabs Agents MCP connector returns 401 invalid_api_key, persists through reconnect and app restart
- Workflow tool: permission-approval schema validation rejects a clean script with "control characters that would be hidden in the approval dialog"
- permissions.deny does not intercept a file read inside a command substitution
- [Bug] Bridge reconnect re-injects full agent/skill/tool listings, inflating Messages token count per turn
- [Feature Request] Reduce false positive safety guardrail triggers for legitimate ML optimization work
- [Bug] Claude fails to identify and correct repeated errors in task execution
- [Bug] Anthropic API Error: Safety Filter Blocking Technical Content Retrieval
- [BUG] Persistent ECONNRESET on home network (VDSL), works fine on mobile hotspot — macOS
- [Bug] Claude Code `/bug` command inaccessible from Skill tool invocations
- CLI hangs in a 100%-CPU main-thread loop and ignores SIGTERM/SIGINT (19 sample captures, 3 versions, identical stack)
- [Bug] MCP tools with requiresUserInteraction not suppressing "don't ask again" option
- Docs say workflow subagents always run in acceptEdits, but they inherit the session's permission mode (2.1.241)
- Claude Code helper bundle ships without an app icon — blank grey row in System Settings → Privacy & Security
- Desktop (Windows): manual drag-and-drop session reordering within groups removed
- [Bug] Fable-5-Safeguard incorrectly flags legitimate Samba server setup task as malicious
- [Bug] Token usage discrepancy between session carryover and actual Opus usage
- [Bug] Safeguards triggered when using bug-finding functionality
- Artifacts: Pro/Max cannot share "latest version" with a named person — public links are version-pinned
- Windows: a console window opens and steals keyboard focus while typing an @-mention in the Claude panel (VS Code extension)
- Windows: VS Code extension spawns 14 helper processes without windowsHide, each opening a focus-stealing console window
- [FEATURE] One-click handoff between Cowork and Claude Code sessions
- claude-in-chrome: extension verified installed/enabled/logged-in but never detected by list_connected_browsers or switch_browser
- [BUG] VS Code extension host restart due to OOM caused by Claude Code loading histories of ALL sessions!
- [BUG] WorktreeCreate hook returning a path outside the repository is always refused: "git could not be run to resolve it"
- `/resume` (CLI picker) silently hides sessions whose transcript `entrypoint` is `sdk-*`
- [Bug] CVP access not working despite approval status
- Unexplained <claude-code-hint> lines injected into background Bash log output, referencing a nonexistent plugin, paired with an instruction to withhold this from the user
- Effort picker shows a level as selected while requests are silently applied at a lower level, with no warning
- Conversation area becomes unscrollable past a rolling window once a session grows large, older content permanently unreachable
- [Feature Request] Reduce false positives in survey content detection
- [Bug] Claude ignores output style instructions and uses prohibited terminology (AI slop terms)
- Feature request: render markdown links as clickable hyperlinks (OSC 8) in assistant output
- iOS Claude Code app: dictation "release to send" doesn't trigger send
- [BUG] The completion path for chip sessions assumes PRs, which excludes repo owners who commit directly and people not on GitHub at all
- [Bug] Safety classifier over-flagging legitimate first-party automation as [cyber] threat
- [반성문] 서브에이전트 위임 시 모집단 전제 누락 → 세션 대표 발견 수치(+26.4%p, n=4,063)가 애초부터 오염된 채 여러 세션째 인용·라이브 배포됨
- [BUG] CVP stuck "In review" for month after org/portal changes and i had previously approval months ago
- [Bug] Claude Code stops responding during database encryption implementation
- Peer SendMessage reply nudge stamps from= with agent-type name, not reachable session ID, causing infinite unreachable-address loop
- it ended my conversation. Refused to execute commands, token extortion!
Resolved issues (135)
- Auto mode circuit breaker is cached, poisoning subsequent sessions into Manual
- [BUG] Claude Code / Claude Desktop does not relaunch after waking the PC from sleep
- Claude Desktop (Windows): no Polish language support (UI + voice dictation)
- Claude Desktop updater unconditionally re-copies chrome-native-host.exe on launch, causing hard launch failure when Chrome holds an open native-messaging handle
- [Bug] Safeguards incorrectly flagging legitimate technical requests as unsafe
- [BUG] A single metadata-only stub transcript makes /resume list empty for the ENTIRE project — 33 sessions hidden by one 416-byte file (2.1.210, macOS)
- [BUG] Mouse-tracking modes left enabled after session exit — bare shell floods with SGR sequences on mouse move (macOS Terminal.app)
- [BUG] Opus 5[1m, MAX] ran `git checkout --` unprompted and destroyed all uncommitted work because of formatting
- [BUG] Skill(artifact-design) is not working
- Workflow tool's subagents fail every tool call - permission handler strips required parameters
- [BUG] Session pegs a full CPU core indefinitely after its worktree is deregistered
- AskUserQuestion hard-freezes CLI tab (Esc/Ctrl+C dead) when options use the preview/"press n to add notes" layout
- [Bug] Unexpected file deletion without user confirmation
- [Bug] Subagents idle without supervisor intervention and off-plan code generation despite verification
- [Bug] Anthropic API Error: Request ID req_011Ce5fmbg9F87DzGxXddy3F
- I appreciate you sharing the context, but I'm not able to access external Request IDs or see the specific bug report you're referencing. To generate an appropriate GitHub issue title, I would need: 1. **The actual error message or problem** you're exper
- [BUG] Claude publishing PR comments while it was asked to only create draft comments
- [BUG] Sandbox kills every Bash call in worktrees at the default `.claude/worktrees/` location (bwrap EROFS creating protected-path mountpoints)
- [Bug] Performance degradation and general system instability
- [BUG] claude mcp list misleads if I was logged in with different acocunts on the same computer (v2.1.196)
- mouse: false setting not respected by plan-approval prompt
- [Feature Request] Add read-only mode for document review without tool execution
- [BUG] Main agent not being able to read messages from sub agents (team agents)
- [Bug] Silent model substitution to non-selected model despite Max plan subscription
- Make /resume session limit configurable in settings.json
- Zero Anthropic staff engagement on 13 documented model failures from a paying customer
- [Feature Request] Improve context retention and multi-model integration for conversation recall
- [BUG] Claude Code in Desktop - No way to TURN OFF Prompt Suggestions
- [Bug] Shell code blocks with heredocs incorrectly indented with leading spaces
- [Feature Request] Support `/fork` with model and effort parameters in single command
- [Feature Request] Add safeguard override or allowlist for user-owned projects to prevent false-positive flagging
- [Bug] Chinese language display in UI
- [FEATURE] Desktop: render images returned by MCP tool calls inline in chat (revive #53980)
- Add post-apply diff review UI (Keep/Undo per file, like GitHub Copilot Chat)
- [BUG] Claude Code process exited with code 132
- [BUG] Claude Desktop update leaves MSIX in wedged state - reinstall fails with 0x80073CF6/0x80073D05
- [MODEL] Safety classifier false positive: flagged during defensive auth bug-fixing, switched to Opus 4.8
- [Feature Request] Add security audit mode for local codebase analysis
- CERT_HAS_EXPIRED on Windows native install during OAuth login, no proxy/VPN/AV — curl succeeds on same host
- [Feature Request] "Open with external app" option for file links/attachments in chat
- Intermittent session-local tool-I/O corruption on Windows (silent Write loss, fabricated/replayed results, mojibake) — worse under high concurrency + long-lived sessions
- attribution settings should also govern the CI-monitor PR reply footer
- Feature request: fresh-context adversarial plan-review in plan mode (audit before ExitPlanMode)
- [Bug] Plan mode UI blocks Claude response output in full screen
- [Bug] Incorrect Safety Filter Flagging on Unity UI Menu/Dialog Implementation Prompt
- [Bug] Claude refusing tasks due to security-related vocabulary in non-security contexts
- [Bug] Session transcript persistence silently disabled by server-side gate
- [Feature Request] Remove or make configurable 60-second timeout for user input prompts
- [Bug] Prompt injection from sub-agent with no tool usage consuming excessive tokens
- [BUG] Fable 5 safeguards over-flag routine health/biology work and reroute mid-turn
- [BUG] Claude Code desktop app sends invalid model IDs for subscription-authenticated users (400 error on all models)
- [Bug] Claude prioritizing model transfer to Opus without legitimate safety justification
- [BUG] EnterWorktree leaves a nested git submodule uninitialised (empty dir, no .git pointer) — 100% reproducible across fresh worktree names in one session
- Claude Code CLI cannot connect to the Claude Chrome extension
- Fable 5 safety banner advises /feedback, but /feedback is unavailable in the VS Code extension (dead-end feedback path)
- [Feature Request] Make AskUserQuestion timeout configurable
- [Bug] Safeguard flagging legitimate SAN dashboard code review requests
- [Feature Request] Remove overly restrictive content filtering for educational physiology and anatomy discussions
- New Chat conversations get routed into an already-active Claude Code session instead of creating an independent one, causing indefinite hangs
- [BUG] Remote-code: enabling /remote-code on a session causes other enabled sessions to disappear (iMac + Android repro)
- [Feature Request] Add security study mode to distinguish legitimate cybersecurity learning from malicious activity
- Scheduled/background task sessions hang on their 2nd shell command (desktop app, Windows 11)
- [FEATURE] SSH connections: support OpenSSH certificate authentication (CertificateFile) or delegate to system ssh
- [Feature] Compact a running background session from outside (no external trigger exists)
- [Bug] Subagent transcript over SKIP_PRECOMPACT_THRESHOLD (5 MiB) cannot be resumed: 'No transcript found for agent ID'
- [Regression, Windows+macOS] Cross-session messaging silently broken since CCD 2.1.224+ — messages dropped, recipient sessions wedged 15-20 min (hadFirstResponse=false); fixed in 2.1.237 for some installs, stable channel still affected
- [BUG] `claude remote-control` (server mode) fails registration with 400 on legacy `/v1/environments/bridge`, while `--remote-control` succeeds via v2 session path
- [Bug] Claude Code `/bug` command inaccessible from Skill tool invocations
- Unable to install claude for window
- [BUG] Claude Desktop arm64 fails to launch on Snapdragon X2 Elite — GPU process crashes (exitCode 101457950)
- Gift Redemption Issue when sender and receiver email is same
- [FEATURE] Claude Enterprise Groups Maximum Spend Limits
- Sub-agent pane steals keyboard focus from the main input while typing
- [Bug] Tool execution fails to match multiple lines in clone operations
- Model reports a failed detached action as 'done' and force-kills process state without authorization — in the same turn it was corrected for both (claude-code, Opus 4.8)
- [BUG] /clear doesn't reset "Context limit reached" state; full process restart required
- Runaway anonymous-memory growth (~14.5 GiB in ~5 min) kills the CLI — reproduces on BOTH native (Bun/JSC) and npm (Node/V8) installs
- FleetView row loses session title after detaching back from an attached session (macOS)
- MCP server 'supabase' sendo removido/adicionado repetidamente com comando malformado (openssl rand -hex 32 não interpolado)
- [ERROR] Error streaming, falling back to non-streaming mode: Connection error.
- [BUG] Claude Code fabricated a user interruption/instruction and preserved it after compaction
- [BUG] Subagent tasks set and billed to Sonnet-5 despite Haiku set as session model (v2.1.198)
- [BUG] Burning through more tokens, while doing nothing different then before.
- [BUG] Workflow tool: synthesis agent() returns placeholder schema values instead of real output on large aggregated input
- HTTP MCP: valid token treated as expired — "OAuth session expired and could not be refreshed" when provider issues no refresh token (mcp.facebook.com/ads)
- [BUG]
- [BUG] Text cursor invisible in light theme — bypasses terminal's context-aware cursor color (recurring, see #24587 #20928 #24929 #46374)
- [BUG] Subagents execute independent tool calls sequentially even when explicitly instructed to batch — including fork mode with identical system prompt
- [BUG] Context-compaction thrashing
- [Bug] Request blocked with unclear error message
- [Bug] Auto-update failure notification persists and cannot be dismissed
- [FEATURE] Add per-built-in-agent model/effort override (settings.json), without forking the prompt
- [Feature Request] Add exemption mechanism for legitimate cybersecurity work to prevent over-blocking of defensive security use cases
- Bedrock: subagent model alias ignores ANTHROPIC_DEFAULT_SONNET_MODEL, forces invalid eu.<model> profile
- Subscription-OAuth long-running Remote Control session dead-ends at 401 / "Please run /login" at the daily refresh boundary; /login cannot complete in a remote-controlled session
- [FEATURE] /splitscreen command and button for custom multi-tab layouts
- Terminal title: distinguish "idle with background tasks running" from truly idle
- [FEATURE] Improve AI-generated image and logo quality
- [FEATURE] Allow Claude to fully interact with websites — not just code
- [Bug] Workflow mode cannot be cancelled and consumes full token budget without user control
- Opus 4.6
- [Bug] Unexpected model downgrade from Opus 4.8 to Fable 5 during quantum physics research
- [Feature Request] Add NotebookRead tool for efficient Jupyter notebook cell extraction
- /resume fails with invalid --permission-mode for sessions containing "permissionMode":"bubble"
- Subagent detail view footer shows the parent session's model, not the agent's actual model
- Local CLI auth expires every ~12h during long-running Remote Control sessions (personal Max) — forced /login; docs claim credentials auto-refresh
- [Bug] Claude ignores context files (CLAUDE.md) and prior implementations in conversation history
- [Bug] Harness forces sub-agent execution to the background even when Claude does not set `run_in_background: true`
- Feature: Multi-session coordination primitives (cross-session messaging, session registry, compaction-resistant state, shared task board)
- [BUG] /model menu shows stale label "Sonnet 4.6" for Sonnet 5 after switching
- [Bug] Race condition: User prompt and feedback button appear simultaneously
- [FEATURE] VS Code extension: no keyboard scrolling in the Claude panel transcript
- [BUG] Cowork (Linux): resuming a pre-upgrade conversation fails with duplicate `.claude` / `.claude/skills` VirtioFS mount ("file exists")
- ~/.claude.json corrupts repeatedly on Windows, silently wiping user-scope mcpServers
- [Bug] Anthropic API Error: Repeated Rate Limit or Overload Responses
- [Bug] Auto-update mechanism failed to complete
- [Feature Request] AskUserQuestion button clicks register unintentionally during window switching
- [Bug] Unexpected model fallback to Opus during prompt execution
- [BUG] Cowork: claude://cowork/new?q= prompt renders but doesn't commit when app already running — no Send button, placeholder still shown (fresh launch works)
- [BUG] Resumed sub-agent gets stuck in self-addressing retry loop, burns ~10M tokens over 3 hours with zero output
- Bug report : Amiga forever and amiga software development labeled by fable as security and dropped to opus.
- [Bug] Fable guardrails incorrectly flagging legitimate development workflows
- [Bug] Geographic Restriction: Fable 5 Unavailable Outside Supported Regions
- [BUG] Fable 5 safety classifier flags UI/frontend work — trigger appears to be project files in context, not prompts
- Claude Code web: cloud session permanently stuck at "14 running tasks" (empty Background tasks panel, input ignored, survives restart)
- remote control: bridge restart rotates environment id and silently auto-disables all scheduled routines (auto_disabled_env_not_found); plus a 6h registration crash loop
- [BUG] Novas sessões cortam plugins ativos em sessões ativas
- Assistant text between tool calls renders inconsistently in long wakeup-driven sessions
- [Bug] Dynamic workflow cost reporting discrepancy between session display and billing
- [BUG] Remote Control epoch-mismatch (409) recovery blocks session exit for ~20s when unrelated concurrent sessions are running
- [FEATURE] remote control does not have option to type something when it is allow/deny or allow/allow all /no option
- [BUG] Claude remembers NOTHING.
- [FEATURE] Routines management: delete tool in the scheduled-tasks MCP, grouping in the routines list, clearer paused state for manual-only tasks
- [Bug][cyber] False cyber block on cache-fix-warmer maintenance cron (req_011CcbvVXnTau2SguZK9QyDd)
- [FEATURE] Can't reach my nested repos from the claude agents @repo picker