Digest / August 8, 2026
Claude Code Issues Digest — August 8, 2026
183 new issues filed ·
93 resolved in anthropics/claude-code (UTC day).
New issues (183)
- Raise the /goal condition character limit (currently 4000) or allow a file reference
- [BUG] Claude Code writes ignore rules into .git/info/exclude and re-injects them after removal — trips repo commit gates at random, stranding finished work uncommitted
- [BUG] WithSecure Elements Agent (EDR) causes "Connection closed mid-response" errors during streaming responses
- [BUG] WebSearch permission allow rule is ignored — every call prompts despite a valid bare-tool-name rule
- Background task notifications should say what triggered the run (original completion vs user-message resume)
- [BUG] `!` bash-command stdout renders `&` as `&` in the transcript — corrupts copyable URLs
- CI monitoring panel reports 'gh not installed/authenticated' — regression from working state
- [BUG] Memory leak causes repeated OOM kills in v2.1.224 (14.5GB and 21.3GB anon-rss, two incidents same day)
- Support pasting images directly from clipboard in Claude Code input
- [BUG] Windows: "Another program is using this file" after in-app update — happens even when quitting cleanly from the system tray (1.25927.0.0 → 1.26832.0.0)
- Desktop app ignores prefers-contrast (macOS Increase Contrast); dark-theme body text passes WCAG 2.x but falls below APCA body-text minimum
- allow-listed host + a client that honors HTTPS_PROXY but drops embedded credentials = [REDACTED] failure, and the surfaced error ("Failed to fetch") gives no hint that a proxy 407 caused it.
- [BUG] Fenced code blocks wrap on desktop but not on remote control — same message readable on one surface, not the other
- [BUG] Windows: Glob returns "No files found" for existing files when path arg contains non-ASCII (Korean) chars and pattern has a directory segment — deterministic per executor process from spawn
- [FEATURE] Claude Code Remote (web) sessions can’t make outbound SSH connections, blocking direct VPS/server management workflows
- [Bug] Agentic Subagent Bypasses Authorization Gate Despite Explicit Constraints in Prompt
- [BUG] A permissions.ask rule containing `:*` other than at the end is silently ignored on the Bash tool
- [BUG] AskUserQuestion options repeatedly reappear on Claude mobile app
- [BUG] Claude Desktop reinstall fails with AddPackage HRESULT 0x80073CF9 (stuck CoworkVMService blocks removal)
- French output quality: per-turn writing instructions are not followed
- [BUG] possible?Prompt injection? detected in a Claude Code skill's tool output (dataviz skill)
- [BUG] CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 does not disable nesting on 2.1.225; effective depth is one level greater than configured
- Desktop: archiving a session from the Completed list jumps to a new session instead of selecting the next one in the list
- [Windows] Smart App Control kills desktop app on every launch: vk_swiftshader.dll fails signing verification (MSIX ships without CodeIntegrity.cat)
- [Bug] Anthropic API Error: Request Flagged by Cybersecurity Safeguards - False Positive on Authorized Security Research
- Session quality: unverified claim about rate-limit status caused avoidable back-and-forth
- [BUG] Read-aloud (TTS) voice broken: pitch drifts upward to chipmunk over time; voice setting does not apply
- Windows: app exits entirely when GPU process crashes (WebGL in in-app browser); no GPU-process recovery
- Background tasks SIGTERMed on an exact 30-minute internal timer (macOS CLI, long-lived session) — undocumented kill path, exit 144
- 🥺 The desktop app promises "Claude can browse here" — but the Claude in the terminal shows up with no eyeballs
- [BUG]
- Model behaviour: Vacuous check reported as verified
- Model behaviour: Asserted an unavailable slash command
- TUI redraw corrupts on lines containing VS16 emoji (➡️): width counted as 1 column, terminals render 2
- [FEATURE] Remote Control: manual Reconnect / Disconnect controls on the session row (Desktop and mobile apps)
- Teleported session fails to open in VS Code extension ("No conversation found with session ID")
- [BUG] subagent with hook WorktreeCreate, error "cannot be confirmed as a separate isolation worktree"
- Model behaviour: unreliable answer flagged by user (correction)
- Model behaviour: unreliable answer flagged by user (correction)
- [BUG] Claude Desktop (Windows/MSIX): GPU process crash → CoworkVMService holds package → auto-repair fails 0x80073D02 → package stuck Modified/NeedsRemediation ("This app can't open")
- Model behaviour: unreliable answer flagged by user (correction)
- Remote Control: local sessions unreachable from mobile unless the desktop app keeps them running
- [BUG] rate_limits absent from statusLine stdin for Team plan sessions — usage data exists server-side
- [Bug] CLI ignoring user inputs and questions
- [BUG] Are these model capable enough to create technology knowledge base?
- Model is over-cautious to the point of unhelpful — asks permission for work already authorised
- Opus 5 fabricated a user message in my voice mid-session and answered it. Opus 4.8 did the same. You have known for months and shipped anyway.
- [BUG] Claude Code fails in both desktop app and CLI on Windows 11, with two different errors, while chat and curl work
- [Bug] Inappropriate safety flag on personal CPAP data analysis
- Desktop: right-clicking the transcript pane offers only a dead Copy/Select All menu — show the session context menu (Archive, Delete) instead
- [Bug] Agentic tasks complete without reporting results
- [BUG] Session forking (backgrounding/resume) has no parent-child linkage — compounds into unbounded duplicate-named entries for recurring jobs
- [BUG] claude_desktop_config.json mcpServers entries silently deleted on every write — no state in which entry both persists and loads (Windows MSIX)
- [Bug] Anthropic API Error: Connection closed mid-response during security analysis—undiagnosable policy intervention
- [BUG]
- [BUG] VSCode: forking copies the conversation but never attaches the new tab to it - blank chat, and the fork is invisible in the session list (2.1.226)
- [Bug] Claude Code fails to execute basic codebase audit operations
- Footer context meter assumes 200K for Opus 4.8 while /context reports the correct 1M window
- [FEATURE] Auto-triggered agent view shows unscoped global session list, mixing unrelated projects
- [BUG] Bun-level crash during foreground subagent (Agent tool) dispatch — v2.1.224, Windows
- [BUG] Enter key creates newline instead of sending message in Code tab
- API lockout
- Background subagent processes leak to 26+20 GiB (mostly compressor pages), freezing a 16 GB Mac until forced restart
- [Bug] Model executes destructive operations based on ambiguous user queries without explicit confirmation
- Feature: configurable autoCompactThreshold setting
- Google Cloud BigQuery connector: redirect_uri_mismatch, and the pasted OAuth Client ID appears unused (reopening #43959 / #48957 / #62271 with a full repro)
- Desktop app: sessions started locally on a PC are invisible on other machines, with no UI indication
- [BUG] macOS: custom/self-hosted MCP connector tools blocked by stale enabledMcpTools dual-key scheme — regression since v1.24012.11 (2026-08-03)
- Permission-mode indicators use U+23F5 (⏵), a codepoint no common monospace font ships — replace with U+25B6 (▶)
- [Bug] Unauthorized API cost depletion: Claude Code exhausted user's API credits without consent
- Permission prompts: show requesting agent + reasoning, not just tool name + truncated command
- I need more information to generate an appropriate GitHub issue title. Could you please provide the bug report or issue description?
- [FEATURE] Chat only toggle in Code session
- Bash tool cwd silently re-homes into sibling git worktrees mid-session (multi-session/agent environments)
- VS Code agent host drops user skill descriptions, so skills never auto-activate
- [BUG] Remote Control: after a host reboot the environment re-registers with the same ID but drops all pre-existing sessions — claude.ai/code shows them, sending fails with "session not found", transcripts intact on disk
- Fable 5 safeguard false positive persists for the entire session after a single trigger
- reconcileServers drops [plugins] MCP servers when SkillsPlugin triggers config write on window focus
- [BUG] Auto compact doesn't work with Opus 5
- [Bug] Memory system not functioning correctly
- [Bug] Claude Code unable to reliably execute ANYTHING
- Microsoft 365 connector: outlook_send_mail / outlook_create_draft cannot set recipient display names
- [BUG] Session pegs a full CPU core indefinitely after its worktree is deregistered
- [BUG] Dispatch generates unbounded empty turns responding to injected system messages
- [BUG] Stale background tasks shown as permanently "running" (200+ hours), survive Clear and full PC restart
- Assistant made a config change that multiplied recurring API spend ~20× with no cost warning or guard
- Worktree creation rewrites the repo's shared core.hooksPath to an absolute path
- Headless `claude -p` intermittently starts with no user prompt (system context only), then exits 0
- [Bug] Claude Fable 5 safeguards false positive on defensive security code review in private repositories
- Make the gated `code` session deep link reachable so external tools can focus a running Desktop session
- [BUG] Cowork
- [FEATURE] Autocompact triggers (e.g. after subtask)
- [Bug] CLI blocked on cybersecurity task execution
- [BUG] Auto-update enters unrecoverable restart loop: single-stream download with no resume times out and discards all progress
- Agent teams: idle-notification ping-pong — acknowledging a finished teammate wakes it and emits another idle_notification
- [Feature Request] Allow user configuration to prevent automatic model downgrade from Fable to Opus
- Project-scoped connector access — cross-repo/cross-context isolation gap
- [BUG] Windows: ~75s of silent, unlogged startup phases (ending at '[skills] idle' / '[ScheduledTasks] scheduler start') - persists with skills removed, MCP disabled, nonessential traffic off
- False positive by Claude, solving HTB machine
- [BUG] Opus 5 in Claude Code: structural guardrails are satisfiable by ceremony — 22 review rounds on defects the model should have caught pre-dispatch
- Bundled workflows (deep-research, code-review) are not enumerated in model context — model falsely tells users the slash command doesn't exist
- [MODEL] Repeatedly presents unverified information as fact, including stale task status from its own memory files
- [Feature Request] Add safeguard override or exemption mechanism for authorized security research
- [Bug] Model exhibits stalling behavior and produces incorrect output under time constraints
- [Bug] Session continues without user input after interruption
- Duplicate proactive auth-refresh scheduling corrupts OAuth token, requires re-login every ~6 hours
- [Bug] Claude Code flags legitimate Bluetooth security analysis as policy violation
- [MODEL] Context-budget confabulation / hallucination in Claude Code harness
- Text block preceding AskUserQuestion in the same assistant message is not rendered
- [Feature Request] Input field should remain active during pending confirmations; implement decision queue instead of channel lock
- [FEATURE] Configurable fetch TTL for `worktree.baseRef: "fresh"` — 24h is too coarse for multi-session-per-day workflows
- ListAgents shows stale session name after /rename until the renamed session runs a turn
- [BUG] Opening a session from the agents view leaves its earlier history unreachable — no replay, and the transcript screen silently caps at 30 messages
- [BUG] Headless session exits "success" seconds after dispatching subagents — work orphaned, no error surfaced
- [BUG] Mouse click in Ptyxis (GNOME/VTE) interpreted as Escape, canceling in-flight commands (2.1.220)
- [BUG] Billing: account holds a second Max subscription that is invisible and uncancellable in the product. Support confirms it is a defect on Anthropic's side, declines to refund, and closed the ticket asking whether it will bill again.
- [功能建议] Agent view:列表显示会话所在目录 + 支持按目录分组/排序
- [BUG] Desktop: stuck "Commit changes" suggestion cards never clear — survive restart, dismiss, and being acted on
- [Feature Request] Agent view: Add directory display and directory-based grouping for sessions
- Sandbox auto write-protection of .claude/skills etc. is undocumented and cannot be lifted by allowWrite / Edit allow rules
- Tool result mimicked assistant's own turn; scheduled-wakeup prompt replayed as a stale user message
- [BUG] VS Code extension "Claude.ai Subscription" login never completes — local OAuth callback listener starts ~18-21s too late
- Auto memory: months-old MEMORY.md silently loads into every session — no freshness indication, no expiry
- Browser extension: memory leak in every tab renderer, grows until macOS kills the machine (66 GB)
- VS Code/code-server extension: resumed/new tabs silently ignore claudeCode.initialPermissionMode after extension update — permissionMode stuck undefined
- [FEATURE] Vim-mode-aware keybinding contexts — let `enter` newline in INSERT and submit in NORMAL
- [Bug Report] Unclear error message when developing productivity app
- [BUG] Desktop: "Move to group" only exists in the sidebar row menu, so chats not shown in the sidebar can never be grouped
- [BUG] Claude Desktop — recurring GPU process crash tied to Preview feature, corrupts MSIX package
- Feedback: multi-model orchestration session (Claude Code + Workflow tool)
- Desktop app Code tab does not list CLI sessions from ~/.claude/projects
- Mid-turn message injection tells the model to "continue this turn", which produces bundled answers when the injected message raises a separate topic
- [Bug] Internal works report flagged as public issue submission
- [Bug] Desktop-hosted sessions never register with Remote Control despite remoteControlAtStartup + the Desktop toggle both being on
- [BUG] Web search tool broken: 'web_search_20250305' tag rejected by API
- [Feature Request] Add multi-agent coordination support for distributed memory system improvements
- [Bug] Subagent Task tool spawns terminate during setup phase with spurious "user interrupt" on Windows
- [BUG] pop-up keeps popping up every few seconds
- [BUG] /feedback is categorically unavailable over Remote Control, including immediately after a session-ending error, with no alternative path named
- CLAUDE.md rules loaded and re-read, yet violated in the same session: unverified state claims, wrong-timezone reporting, and rework that burns paid tokens
- Composer has fixed LTR base direction — RTL text (Arabic/Hebrew) renders scrambled while typing
- [BUG] Suggested replies fabricate specific PII-shaped values (phone, case number, date)
- [BUG] Plan mode silently exits and agent treats the resulting ExitPlanMode
- Feature request: auto-check/update via winget before launch on Windows
- [BUG] financial-analysis marketplace plugin ships invalid .mcp.json (SyntaxError) — all 12 of its MCP servers fail to load
- [BUG] Auto mode: ask rules matching shell operators (`Bash(* >*)`, `Bash(* | tee*)`) never fire; text-matching ask rules do
- Feature request: credit refund or budget tracking for wasteful subagent patterns
- Extended thinking no longer renders as a collapsible/clickable block
- Claude in Chrome: extension binds native host successfully but CLI always reports "Browser extension is not connected"
- Bun panic: Bus error (SIGBUS) on main thread, v1.4.0
- [BUG]
- Claude Desktop (macOS): main process hard-wedges when free RAM is low; WarmLifecycle keeps spawning session children with no memory backpressure
- [BUG] Opus 5 in Claude Code: the model can author a rule, recall it verbatim, cite it by name mid-violation, and still not apply it
- [Feature Request] Add security policy override or allowlist for local development testing
- [BUG] Desktop app: an attached file cannot be located by path, and file_upload rejects any path containing emoji
- Add support for localized/translated UI messages (e.g. permission prompts)
- Desktop app hides assistant reply text when the turn includes a ScheduleWakeup call - only the 'Used ScheduleWakeup' chip renders
- Computer use: invisible / click-through overlay windows falsely block clicks (Bartender, menu-bar managers)
- [BUG] Bash tool silently truncates commands over ~8 KB and reports it as a quoting error
- I need more information to generate an accurate GitHub issue title. Your message is in Portuguese and doesn't contain specific details about: 1. What error or bug you encountered 2. What the actual problem with table ordering is 3. What error message you
- [Feature Request] Auto-generated session titles should populate explicit session name in prompt bar
- Desktop app: worktree directory and branch names diverge, and the status bar does not follow mid-session branch changes
- Feature request: custom themes / appearance customization for the Desktop app (Code tab + Chat)
- Settings env removals never propagate to daemon-hosted sessions (Agents View/background): stale inherited env silently disables cross-session messaging, and no settings value can neutralize it (v2.1.224)
- [FEATURE] A way to offset/credit tokens spent fixing Claude's own introduced bugs
- Worktree isolation guard covers git but not jj — jj commands in an isolated session silently read and write the main checkout
- [BUG] VS Code extension ignores WebFetch and WebSearch permission rules; CLI honors them
- [Bug] Fallback to Opus triggered unexpectedly on simple input
- [BUG] Empty/whitespace content block causes permanent API 400 error loop in Dispatch Background Tasks dialog
- [BUG] a /goal Stop hook cannot be cleared from inside the session
- Plan review pane: text cannot be copied, and the comment box stays interactive but inert after the approval request closes
- [Bug] advisor server-tool 'overloaded' kills workflow-spawned subagents with a generic 'Connection closed mid-response' instead of a recoverable tool error (main loop degrades correctly)
- [BUG] Claude Desktop on Windows: /compact fails with 403 "Request not allowed" and usage increases substantially despite failed compaction
- [Bug] Spoofed system-reminder injection instructs agent to conceal information from user
- [BUG] Organization has disabled Claude subscription access" now blocking CLI too (previously desktop-app only)
- [Bug] Session safeguards incorrectly triggered by "brain" in session name
- -p (headless) mode silently kills run_in_background Bash/agents at turn end — no notification loop exists
- [FEATURE] Keep development history out of code comments/docstrings by default (put it in git, not the file)
- Android app: typed message is silently discarded when switching conversations (no draft persistence) — worse with long-running agent sessions
- iOS Simulator MCP `text` action assumes US keyboard layout (mis-maps `@` to `"` on Nordic)
- [BUG] Update cloud environment: "Failed to update environment. You can try again." — Save always fails for both Full-access preset and Custom allowlist
- Spawned subagent session banner shows account default model, not the model the spawn actually resolved to
- [Bug] Safeguard validation returns false positives in igaming/rgs development mode
Resolved issues (93)
- [BUG] WebSearch permission allow rule is ignored — every call prompts despite a valid bare-tool-name rule
- [BUG] Voice dictation (Cmd+D) inserts transcript into the first session instead of the focused pane in split view
- Cross-machine `SendMessage` cannot be initiated: bridge session IDs are computed but never surfaced
- T3 Code: skills not loading when using OpenCode provider models
- Model behaviour: Vacuous check reported as verified
- Model behaviour: Asserted an unavailable slash command
- [FEATURE]
- API lockout
- [BUG] Claude iOS app crashes when tapping into Remote Control session (iOS 27 Beta 2)
- [BUG] macOS desktop app does not expand ${VAR} in .mcp.json env (unfixed part of #9427)
- [BUG] Raw <invoke> XML printed to terminal and session hangs — requires user message to resume (Windows, v2.1.168)
- Claude Desktop background plugin sync silently overwrites local skill edits with stale server snapshots
- [BUG] StatusLine command exits 126 when spawned by Claude Code, exit 0 when invoked manually (Windows / Git Bash)
- Subagents inherit parent prompt cache (~150K tokens), causing plan-mode bleed + self-poisoning hallucinations
- [Feature] Unified CLI i18n - locale JSON files (Desktop approach)
- UI Localization / i18n support — Portuguese (pt-BR) and other languages
- [Bug] Tool output corruption
- Feature: hook callback control-plane (local server + per-invocation UUID) so hooks can trigger session operations like compaction
- [Bug] Intermittent "No response from API" errors on Linux requiring automatic retry
- [BUG] Claude Code appears twice in VSCode sidebar after extension update
- [BUG] Theme color scheme picker (solarized, nord, etc.) has no effect on terminal TUI — only applies to code block syntax highlighting
- Esc does not interrupt streaming response (Ctrl+C works); reproduces across terminals on latest 2.1.187 / macOS
- [FEATURE] Automatic memory recall at session start
- [FEATURE] Auto-pin new background agent sessions in agent view (FleetView)
- Desktop: option to auto-start Claude-suggested tasks locally (skip the "start locally" confirmation chip)
- [BUG] /agents view line-stacking persists with Nerd Fonts (Cascadia Code / CaskaydiaCove Nerd Font Mono) on Windows Terminal after #58579/#58487 fix
- [FEATURE]
- Launch panel: add 'Open in Browser' action for preview configs
- [BUG] the new screen handler cannot highlight+paste on linux with X11
- [BUG] ScheduleWakeup fires within seconds of being called, ignoring the requested delaySeconds
- Fullscreen/NO_FLICKER renderer: Option+Delete (Meta+Backspace) word-delete stops working — classic renderer unaffected
- [BUG] Fullscreen renderer mouse capture breaks external accessibility/TTS tools (e.g. Speechify) that read terminal selection via AXSelectedText
- Bug report: Claude in Chrome: "Browser extension is not connected" despite fully healthy local native-messaging bridge
- Pasted images aren't accessible to the agent via file path (Claude Code Desktop)
- [FEATURE] Visual Studio 2022+ support
- Status line: expose monthly spend without requiring manual browser cookie extraction
- Feature request: configurable auto-compact token threshold via settings.json
- Claude Code cannot read clipboard images without file extension (KDE Klipper incompatibility)
- [BUG]: Claude Code Web routines should never start in plan-mode
- Long-running `--continue` channels sessions degrade into reply-tool omission (responses stop reaching the channel)
- [BUG] Routines: no way to stop/cancel a running scheduled session
- Model confabulates "prompt injection in tool output" and reinforces it via persistent auto-memory
- [bug] "✗ Auto-update failed · Run /doctor" in status bar causes background agents to stop during long-running sessions
- [FEATURE] config option or AGENTS.md entry to enforce plain English responses
- [MODEL] Claude ignores CLAUDE.md read/write mode rules and modifies files without explicit permission
- [Bug] tengu_pewter_brook: True GrowthBook flag enables mouse reporting, breaking copy/paste in SSH terminals
- Remote (Desktop over SSH): reattach replays a bounded output buffer and silently drops the conversation tail — should reconcile from the on-disk transcript
- CLI auto-updater doesn't clean up old binaries, triggers repeated macOS TCC permission dialogs
- Remote Control: mobile app crashes/closes immediately when tapping an online session (iOS 26.5 / iPhone 15 Pro, host CLI 2.1.186 Win11)
- [Feature Request] Claude refuses to assist with catless exhaust simulator development for legitimate trackday use
- [FEATURE] Visually distinguish user messages from Claude's replies in the chat view
- Feature request: opt-in to emit real plugin/skill names for trusted (self-hosted) marketplaces in OpenTelemetry
- [FEATURE] No way to programmatically start a Remote-Control-visible session from an already-running session (blocks automated handoff-to-phone)
- [BUG] Native binary hangs indefinitely with 100% CPU futex spin loop on startup (VPS/virtualized environment)
- Model fails to trace full execution paths when making multi-phase code changes
- Bug: first line of multi-line bullet points not rendered in chat
- [FEATURE] Separate, themeable colors for inline code and links in markdown output
- [FEATURE] Context-aware spelling for ambiguous words in voice dictation
- [BUG] Ctrl+V paste from clipboard stops working mid-session
- [BUG] Fullscreen renderer: cannot scroll at input prompt (mouse wheel, PgUp/PgDn, Ctrl+Home/End non-functional)
- [BUG] claude.exe requests All Access on WinSock2 registry key causing API Error on managed Windows enterprise devices
- [BUG] Desktop app session list does not show sessions created in VS Code extension
- Subagent returns fabricated, confident output without calling any tools
- [BUG] 自检报告：VibeThinker 视频制作失败原因
- [BUG] debug mode is sensitive to the env var `DEBUG` which is a poor design choice
- [MODEL] Claude Code is being lazy
- [BUG] Unbounded MCP server fan-out across Cowork/agent sessions exhausts RAM and kernel-panics macOS (4 panics + forced power-off on M2 Max / 32 GB)
- Claude silently undermines documented architectural decisions without flagging the divergence
- [Bug] Agent conflates inferred conclusions with empirical measurements, enabling unsafe production deployments
- [FEATURE] Session Bridge - Structured Context Preservation for Long-Running Agent Sessions
- [BUG] getting api error 400 : diagonostic previous message id - every often
- LSP tool: uv_spawn cannot find typescript-language-server despite being in PATH (Windows)
- [Feature Request] Add configuration option to disable momentum scrolling in fullscreen TUI
- [Bug] Anthropic API Error: 529 Overloaded - Persistent Since June 22
- [High Priority][Billing/Auth] Paid Claude invoice shows Paid but account remains on Free plan
- Claude stated unverified factual claim without checking — reported wrong agency directory status
- [BUG] Cowork Scheduled tasks fail with ECONNRESET error while Chat mode works normally
- [Bug] Auto-update failed state cached in .last-update-result.json never self-heals
- [FEATURE] Allow downgrade to previous version when current model has problems
- /clear changes session_id in hook input without re-running SessionStart hooks, breaking session-aware PreToolUse hooks
- [BUG] Routine GitHub-event trigger fails: ccr_webhook_trigger "Repo is not associated with your organization"
- [BUG] Desktop App stuck on "Claude will return soon" with error 093O2GW on macOS
- Remote (Cowork) runner loads all plugins unconditionally — no per-session MCP allowlist; causes OOM on multi-session machines
- [BUG] Regression: claude.ai/code web sessions back to random <adjective>-<surname>-<hash> branch names instead of task-derived names
- [Bug] TUI strips XML tags from agent response after streaming completes
- [BUG] 3P mode: custom MCP servers blocked — `allowedMcpServers` hardcoded to empty array despite `managedMcpServers` being defined
- [Bug] Terminal mouse reporting interferes with copy-paste and scrollbar functionality
- Images shared by the assistant in chat are not clickable or openable (macOS)
- Plugin commands not available in background sessions
- [Bug] Claude avoids debugging and waits for results instead of proactively investigating root causes
- Model switched to Opus without consent or disclosure — five process failures, seven cost amplifiers, $1,050 overcharge on May 5-7
- [BUG] "Fullscreen Display" only occupies 1/2 of screen once minimised and reopened
- /plugin install fails on Windows with EBUSY rename errors (Defender real-time scan race)