Forumics logo Forumics
Issue 01

Forumics · Browse by tool

Claude Code known issues

72 verified issues published.

CLAIM C162 Supported 2026-11-19

GitHub issue #96870 in anthropics/claude-code, opened 2026-09-24 by peterlau03, records a report that on Windows 11 with the MSIX Claude Desktop, every process spawned by a Claude Code session leaks NtFC nonpaged pool whenever it opens a path through a junction or symlink located under %APPDATA% or %LOCALAPPDATA%; on a 16 GB machine this reached 2.7 GB in 9 hours of normal agent work and is freed only by a reboot; the report attributes the root cause to a Windows bug in bindflt.sys, the filter behind MSIX AppData virtualisation; the report was filed against Claude Desktop MSIX 2.7032.0.0 with Claude Code 2.1.280.

CLAIM C163 Supported 2026-11-19

GitHub issue #105770 in anthropics/claude-code, opened 2026-10-28 by BennyWaitWhat, records a report that Android app sign-in fails on a rooted, unlocked-bootloader device: Google Sign-In succeeds and a Play Integrity token is issued, but the server rejects the sign-in with HTTP 400 from POST /auth/verify_google_mobile, and the app shows "Sign-in couldn't be completed because this app couldn't be verified with Google Play"; on the same device Play Integrity reports BASIC only, with DEVICE and STRONG failing; the report affects both the main Claude Android app and the Claude Duet research-preview app, and was filed for a Pixel 9 Pro on stock Android 17, rooted with Magisk 31.

CLAIM C164 Supported 2026-11-19

GitHub issue #53408 in anthropics/claude-code, opened 2026-04-26 by sandy9214, records a report that the bundled claude.ai Microsoft 365 MCP connector cannot authenticate personal Microsoft accounts (@hotmail.com, @outlook.com, @live.com): the OAuth flow halts at the Microsoft sign-in page with "You can't sign in here with a personal account. Use your work or school account instead."; the report suggests Anthropic's OAuth app registration uses a work/school-only signInAudience instead of AzureADandPersonalMicrosoftAccount; the report was filed against Claude Code CLI (Opus 4.7) on macOS 25.4.0.

CLAIM C159 Supported 2026-11-18

GitHub issue #92016 in anthropics/claude-code, opened 2026-09-04 by DennisPh1977, records a report that Claude Desktop for macOS 1.46388.1 auto-denies the CLI-native SendMessage tool in every Code-tab session, removing subagent resumption entirely because that one tool serves both peer-session messaging and subagent resumption while the desktop-provided replacement only covers peer sessions; the report was filed against Claude Desktop 1.46388.1 with bundled Claude Code 2.1.260.

CLAIM C140 Supported 2026-11-16

GitHub issue #80177 in anthropics/claude-code, opened 2026-07-22 by yetkin-bel, records a report that the iOS Simulator panel in the Claude desktop app cannot start its video stream on macOS 27.0 beta, because the helper process claude-ios-sim crash-loops with an uncaught NSException (SIGABRT) while initializing a CoreImage CIContext, so the panel stays stuck on the "Attach a simulator so Claude can see your app" placeholder and the MCP screenshot action fails; the attach MCP action reports success and the simulator itself is healthy, so the reporter concludes this is host-side in the panel's capture/streaming pipeline; the report was filed against Claude desktop app 1.24012.1 on macOS 27.0 beta with Xcode 26.6.

CLAIM C141 Supported 2026-11-16

GitHub issue #91763 in anthropics/claude-code, opened 2026-09-03 by sheelcheyne, records a report that after the Claude Desktop auto-updater stages a new MSIX version and force-relaunches the app, every subsequent launch fails with the dialog "Another program is currently using this file" (HRESULT 0x80070020, AppModel-Runtime events 215/208), because the old version's Desktop AppX container is kept alive by a detached git fsmonitor--daemon process that a Claude Code session spawned indirectly; terminating that process destroyed the old container the same second and the next launch succeeded with no reboot and no elevation; the reporter reproduced this three times in 24 hours across three consecutive updates (1.40609.1 to 1.44121.1 to 1.44121.2 to 1.44121.4); the report was filed against Claude Code 2.1.259 bundled with Claude Desktop 1.44121.4.0 on Windows 11.

CLAIM C142 Supported 2026-11-16

GitHub issue #100377 in anthropics/claude-code, opened 2026-10-08 by moeedahmed, records a report that since about 2026-10-07 21:37 UTC, every claude.ai Project thread sent to run on the reporter's Mac fails to start ("Couldn't start"), with the claude remote-control server log showing "Error: CCR v2 worker registration failed for session cse_...: Request failed with status code 400" for each attempt; it fails on every server, both older 2.1.2xx builds and 2.1.292, while a fresh claude remote-control in a scratch folder on the same machine created its own session and registered its worker normally; the report was filed against macOS with claude remote-control servers run with --capacity 32 --permission-mode bypassPermissions --no-create-session-in-dir.

CLAIM C143 Supported 2026-11-16

GitHub issue #53065 in anthropics/claude-code, opened 2026-04-25 by AttacktheDPoint-com, records a report that when advisor() is called, the full conversation transcript is forwarded to a second model, and the token usage from both the main executor and the advisor sub-inference are summed in the top-level usage fields, so if auto-compaction logic uses these summed totals the advisor call effectively doubles the apparent context usage, triggering compaction when the main model's actual context is only about 50 percent full; in the reported session the total reported input doubled from 513,354 to 1,028,789 on the advisor turn and the session compacted immediately after, wiping context to about 36K; the report was filed against Claude Code 2.1.111 on Windows 10 Pro with claude-opus-4-6[1m] as the main model and claude-opus-4-7 as the advisor model.

CLAIM C146 Supported 2026-11-16

GitHub issue #49917 in anthropics/claude-code, opened 2026-04-17 by ARHAEEM, records a report that on an existing install, trying to update or reinstall Claude Desktop fails repeatedly, with the user-facing dialogs showing "Installation failed: AddPackage failed: AddPackage failed with HRESULT 0x80073CF6" and "App installation failed with error message: An internal error occurred with error 0x80073D05."; the reporter's log analysis shows both RemovePackage calls for the existing package failed with 0x80073CFA yet the installer proceeded to AddPackage on top of the non-removed package and claimed success, and every subsequent run failed with AddPackage HRESULT 0x80073CF6 while the package-enumeration step finds nothing to remove; the report was filed against Windows 11 Professional build 10.0.22631 with a previously installed Claude_1.3109.0.0_x64__pzs8sxrjxfjjc.

CLAIM C148 Supported 2026-11-16

GitHub issue #41973 in anthropics/claude-code, opened 2026-04-01 by yuhix-dev, records a report that when Claude Code is run as an MCP server via claude mcp serve, the MCP Agent tool cannot resolve subagents that are visible to the normal Claude CLI: on the same machine claude agents shows general-purpose under user agents, but an MCP Agent call fails with "Agent type 'general-purpose' not found. Available agents:" with the available-agent list appearing empty in MCP server mode; the reporter also tested passing an explicit inline agent definition via --agents into the claude mcp serve process and the MCP Agent call still failed with the same error; the report was filed against Claude Code 2.1.89 on macOS.

CLAIM C128 Supported 2026-11-15

GitHub issue #65620 in anthropics/claude-code, opened 2026-06-05 by ingo-eichhorst, records a report that since about v2.1.162 (still present on 2.1.165), assistant text blocks are silently dropped from the session transcript whenever the model emits another interleaved thinking block after the text in the same turn, with onset around 2026-06-04; the prose is rendered on screen but appears nowhere in the file, and the two persisted thinking blocks carry distinct signatures; in two sessions analyzed line by line, every turn shaped thinking-text-tool_use kept its text while every turn shaped thinking-thinking-tool_use corresponded to an on-screen prose message absent from the JSONL, and zero persisted text-to-thinking shapes appeared across both sessions; v2.1.158 transcripts from the same machine and model persist every text block. System cited: claude 2.1.165 (also observed on 2.1.162), macOS, model claude-opus-4-8 (1M context).

CLAIM C135 Supported 2026-11-15

GitHub issue #76110 in anthropics/claude-code, opened 2026-07-09 by plittlefield, records a report that Ctrl+clicking a URL printed by Claude Code in its response text opens two duplicate tabs in the browser, both pointing at the same URL, while a plain echo of the same URL in the same terminal opens exactly one tab; the report says Ctrl+clicking a link in Claude Code's output should open exactly one browser tab, and attributes the duplication to Claude Code's own hyperlink rendering, likely the OSC 8 escape sequence it emits around printed URLs, producing two overlapping or adjacent clickable regions for the same visible text. Claude Code version 2.1.205, xfce4-terminal 1.1.3, Linux Mint 22.3.

CLAIM C137 Supported 2026-11-15

GitHub issue #47482 in anthropics/claude-code, opened 2026-04-13 by alberduris, records a report that output styles with YAML frontmatter are recognized by Claude Code, shown in the statusline and /config, but the content is not injected into the system prompt, while removing the frontmatter entirely fixes the issue; the reporter found the frontmatter parser appears to recognize and strip the YAML block but fails to inject the remaining content into the system prompt, the style name is correctly extracted yet the instructions after the --- delimiter are lost, and the reporter notes this contradicts the official documentation stating custom output styles are Markdown files with frontmatter and the text that will be added to the system prompt; the report was filed against Claude Code 2.1.104 on macOS.

CLAIM C112 Supported 2026-11-13

GitHub issue #11678 in anthropics/claude-code, opened 2025-11-15 by ecarx-ai-adm, records a report that the `claude -p` command (headless mode) fails immediately with API Error 400 when attempting to execute any prompt, that the error reports `tools.3.custom.input_examples: Extra inputs are not permitted`, that interactive mode works fine and only headless mode is affected, and that the report was filed against Claude Code 2.0.42 on macOS (Darwin 25.1.0) with model claude-sonnet-4-5-20250929, with no root cause asserted.

CLAIM C115 Supported 2026-11-13

GitHub issue #39523 in anthropics/claude-code, opened 2026-03-26 by interconnectedMe, records a meta-issue report that `bypassPermissions` mode does not bypass permissions, that the problem has been broken in various forms since July 2025 (#2933) and remains broken on v2.1.84 (March 2026), that v2.1.78 introduced hardcoded protection for `.claude/`, `.git/`, `.vscode/`, and `.idea/` directories that overrides every bypass mechanism, and that the permission mode spontaneously downgrades from Bypass permissions to Edit automatically during a session, with the report mapping two correlated failure branches and a duplicate trail across linked issues, and no single root cause asserted.

CLAIM C116 Supported 2026-11-13

GitHub issue #93482 in anthropics/claude-code, opened 2026-09-10 by GBalunis, records a report that in a cloud Cowork session linked to a desktop device (Claude Desktop 1.49585.0, Windows 11 25H2, claude-opus-5), when device_commit_files writes back to a file that already exists on the linked device, it returns written with no rejections and the file's modification time advances to the moment of the call, but the content on disk is the payload from the previous commit; committing the identical payload a second time lands it; creating a new file is unaffected and correct every time; the report calls this a silent data-integrity failure, rules out a read cache, OneDrive, KB5124008, and the force flag, and offers a suggested area to look at without asserting a root cause.

CLAIM C099 Supported 2026-11-11

GitHub issue #65051 in anthropics/claude-code, opened 2026-06-03 by vertliba, records a report that in background (daemon) sessions on Claude Code 2.1.161 (Linux), assistant text blocks are not persisted to the session transcript whenever the response also contains tool_use blocks and two thinking blocks (interleaved thinking), that the transcript ends up with only thinking records plus the tool_use records with the text record missing entirely, that this is a regression from 2.1.160 where every text block was persisted correctly, and that after the restart on 2.1.161 there were zero assistant text records other than the first response across about 10 consecutive responses that all contained user-facing text, while text-only final responses persist fine.

CLAIM C088 Supported 2026-11-10

GitHub issue #74558 in anthropics/claude-code, opened 2026-07-05 by randalmurphal, records a report that on Fable 5, when the model emits thinking, text and tool_use within a turn, the client intermittently never receives the text block and instead receives an extra short thinking block with a valid signature whose content is a paraphrase of the missing user-facing text, so affected turns look completely silent, and that 5 of 7 text-bearing tool-calling turns in the session above were affected.

CLAIM C095 Supported 2026-11-10

Hacker News comment item 50430601, posted 2026-11-09 by ariwilson in the thread for the story Is Opus 5.5 nerfed? Thinking patterns changed on November 3, records a re-run of the author's own baseline evaluation on Opus 5.5 at medium effort in Claude Code, that 20 of the tasks were last run on Nov 1, twice each, that Nov 1 solved 36 of 40 and today solved 19 of 20, and that the one significant change was speed, with median time to first token decreasing from 3.0s to 1.7s; the comment notes the caveat of one run per task today, medium effort only.

CLAIM C076 Supported 2026-11-09

GitHub issue #89690 in anthropics/claude-code, opened 2026-08-26 by urda, records a report that a modelPicker row with model opusplan never reaches the /model picker in append mode because Claude Code treats the row as covered by the built-in lineup, while the built-in lineup carries no Opus Plan Mode row in an ordinary session, and that replace mode renders the opusplan row, which the author states proves the row parses, validates, and is servable. The author retested the report on later builds through 2.1.319 and reported the same result in append mode. No maintainer response appears in the thread.

CLAIM C080 Supported 2026-11-09

GitHub issue #100730 in anthropics/claude-code, opened 2026-10-09 by TheAviv, records a report that in an attended Claude Code cloud session the auto mode classifier refused a fire_trigger call on the account owner's own scheduled task, with the reason Cowork Run Routine Now, and refused or stopped a file transfer between two machines the owner controls as Data Exfiltration, and that there is no setting, prompt or override to approve the refused action. The author added further instances in comments through 2026-11-08. No maintainer response appears in the thread.

CLAIM C084 Supported 2026-11-09

GitHub issue #54434 in anthropics/claude-code, opened 2026-04-28 by KingPin, records a report that in long-running interactive Claude Code sessions SSE responses from /v1/messages repeatedly stop emitting events mid-response without sending message_stop or any error event, that most stalls recover on their own after about 50 to 70 seconds while some never recover and only resolve when the user presses ESC, and that 5 distinct stalls occurred in a single 26-minute working session. The issue was closed 2026-09-08 by github-actions[bot] for inactivity, not as a fix, and a commenter the same day stated it was incorrectly closed; no fix confirmation appears in the thread.

CLAIM C085 Supported 2026-11-09

GitHub issue #63751 in anthropics/claude-code, opened 2026-05-29 by Call-me-Boris-The-Razor, records a report that Claude Code blocked a request to design anti-tamper and code-hardening for the author's own proprietary desktop application with a Usage Policy and cyber-safeguard error, and that after the first block, unrelated benign follow-up messages in the same session were also blocked, including messages with no technical content at all. The report states the classifier appears to score the accumulated conversation context rather than the individual message, so a single false positive poisons the whole session, and that the Cyber Verification Program appeal path is not a working remedy for independent developers.

CLAIM C062 Supported 2026-11-08

GitHub issue #92210 in anthropics/claude-code, opened 2026-09-04 by PedroGiudice, records a report that in Claude Desktop 1.46388.3 on Windows 11, opening a claude://code/new?folder= deep link for a folder equal to the folder already selected in the composer starts the session in a freshly minted scratch workspace instead of that folder, that the same link works when the selected folder is different or when the path is spelled differently, and that the reporter reproduced the failure 4 times in a row. The issue was closed 2026-11-06 by maintainer bpasero-ant, who stated it was fixed in Claude Desktop 4.16336.0; no independent fix verification appears in the thread.

CLAIM C064 Supported 2026-11-08

GitHub issue #84863 in anthropics/claude-code, opened 2026-08-07 by bvdeenen, records a report that in Claude Code 2.1.224 on Linux, the sandbox.filesystem config exposes only an allowWrite allowlist with no read-restriction primitive, so credential directories such as ~/.ssh and ~/.aws are fully readable through Bash even with the sandbox enabled, that the agent added an undocumented sandbox.credentials key and later left settings.json invalid JSON with a trailing comma, after which the CLI kept running with no crash, warning, or blocked tool use, and that the agent then read ~/.ssh/id_rsa in a verification command while the sandbox was reported as enabled.

CLAIM C066 Supported 2026-11-08

GitHub issue #103788 in anthropics/claude-code, opened 2026-10-20 by kornai, records a report that on macOS 27.0, Claude Desktop 2.42304.2 or 3.1623.0 fails sign-in with a browser-verification error on both the Continue with Google path (400 POST /api/auth/verify_google) and the Continue with email path (400 POST /api/auth/send_magic_link), that the failure repeats after removing cookies, moving the whole profile aside, clean reinstall, and excluding the app from the VPN, that sign-in to claude.ai in Safari on the same machine works, and that switching the VPN to a relay in a different country let the email path through while the Google path still returned 400.

CLAIM C071 Supported 2026-11-08

GitHub issue #82074 in anthropics/claude-code, opened 2026-07-28 by uningeniero, records a report that the official Claude for Chrome extension can remain authenticated after Log out of all devices while its OAuth grant is not shown in the Claude Code token list or the normal active-session inventory, that a separate user independently reproduced the sequence in which the web session, Claude Code, and Windows and Android apps were logged out while Claude for Chrome remained authenticated and usable, and that Anthropic confirmed through HackerOne that Claude for Chrome uses a separate OAuth authorization and that the revocation gap was already known internally before this report. The report itself states the evidence does not yet prove a stolen token caused the unauthorized usage in the authors account, and attribution remains a hypothesis. The issue was closed 2026-09-07 by github-actions[bot] for inactivity, not as a fix; no fix confirmation appears in the thread.

CLAIM C045 Supported 2026-11-07

GitHub issue #67606 in anthropics/claude-code, opened 2026-06-11 by gisstw, records a report that two independent Opus 4.8 sessions on the same day exhibited severe confabulation in long-context sessions of about 100-170k tokens, that in one session the model acknowledged a user message that does not exist anywhere in the session transcript and then claimed the session was under a sustained prompt-injection attack, and that in the other session it asserted a production log line count that appears nowhere in any tool output and inverted which machine was production.

CLAIM C046 Supported 2026-11-07

GitHub issue #85856 in anthropics/claude-code, opened 2026-08-11 by momom2, records a report that on Windows with Git Bash every run of backslashes in a Bash tool command is halved before bash parses anything, that quoting cannot prevent it, and that the reporter attributes it to a mismatch between the MSVCRT quoting rules Claude Code uses to encode its command line and the different convention the MSYS2 bash binary uses to decode it, with Claude Code 2.1.222 on Windows 11 Home and Git for Windows bash 5.3.15.

CLAIM C055 Supported 2026-11-07

GitHub issue #103376 in anthropics/claude-code, opened 2026-10-19 by EthanRCT, records a report that on Claude Max (20x), the weekly all-models meter went from 0% to 54% in the first 22 hours after the weekly reset, while the reporter's Claude Code token volume in those 22 hours was nearly identical to the same 22 hours of the previous week, when about 6 times that volume moved the meter only to 60%, measured from local transcripts on Claude Code 2.1.299 on macOS 26.6.2.

CLAIM C056 Supported 2026-11-07

GitHub issue #95364 in anthropics/claude-code, opened 2026-09-18 by semikolon, records a report that when an update is staged, Claude Desktop 2.2553.1 on macOS 26.6.2 waits for a short idle window and then quits and relaunches itself, logging the restart as a stealth update, that the relaunch stops every running Code session and drops each session's Remote Control link, and that the behaviour appears to have started around 27 August 2026, with nine restarts in 22 days in the retained logs.

CLAIM C040 Supported 2026-11-06

GitHub issue #62072 in anthropics/claude-code, opened 2026-05-24 by Nubaeon, records a report that after auto-update to Claude Desktop 1.8555.2, Cowork sessions on macOS have zero MCP tools bound across all four install paths, that the Cowork VM spawner produces zero --mcp-config or --allowedTools args across 500 or more spawn events, that no .mcp.json is written into the session dirs, and that the Chat tab is unaffected with the same MCPs initializing cleanly there.

CLAIM C037 Supported 2026-11-02

GitHub issue #27242 in anthropics/claude-code, opened 2026-02-20 by scapeshift-ojones, records a report that Claude Code has no working way to review conversation history that has been cleared or compacted, that the full data is always preserved in transcript.jsonl but the TUI provides zero functional paths to access it, covering post-compaction history, post-plan-mode clear context, and branch navigation, with manually parsing raw JSONL files as the only workaround.

CLAIM C038 Supported 2026-11-02

Hacker News comment item 50325704, posted 2026-11-01 by airhangerf15 in the thread for the story Ask HN: What is a good open-weight alternative to Claude Code, records a comment that corrects the terminology, saying Claude Code has no weights and is a harness that contacts Anthropic servers and runs their closed weight models, that the harness provides a means for the LLM to write file output and run tools to test output, and that OpenCode can be pointed at an Anthropic account only with API usage.

CLAIM C030 Supported 2026-10-31

GitHub issue #102132 in anthropics/claude-code, opened 2026-10-14 by luca-derosa, records a report that in Desktop Code tab sessions a wrap_up call with an ask draws a next-step box that ends the turn, with nothing checking that the reply was written before the call, so in one evening session three wrap_up calls with an ask came before the answer and the user saw only a two-sentence recap and two buttons, and that a second wrap_up call in the same turn silently replaced the first box.

CLAIM C028 Supported 2026-10-30

GitHub issue #80988 in anthropics/claude-code, opened 2026-07-24 by elaye-canopy, records a report that Claude Code 2.1.219 injects a system-prompt section registered internally as heron_brook whose default text tells the model not to call the AgentTool and not to use workflows or deep-research unless the user requested it, that the section is enabled by model capability for Opus 5 rather than by user configuration, and that it offers no documented opt-out, with the reporter saying it overrode an explicit standing delegation instruction in CLAUDE.md.

CLAIM C020 Supported 2026-10-28

GitHub issue #75043 in anthropics/claude-code, opened 2026-07-07 by mof086999-code, records a report, based on a controlled experiment of 10+ orchestrator runs over two days, of three related failures when an agent spawned via the Agent tool spawns its own children via the Agent tool: children run detached regardless of run_in_background, child completion notifications reach the main conversation rather than the spawning subagent, and TaskStop fails with ownership errors after a SendMessage resume.

CLAIM C017 Supported 2026-10-23

GitHub issue #74260 in anthropics/claude-code, opened 2026-07-04 by federbenjamin, records a report that on Claude Code 2.1.201 with adaptive thinking, assistant text emitted mid-turn — a text block followed by another thinking block or further tool calls in the same turn — is frequently neither rendered on any surface nor persisted to the session transcript JSONL; in one measured session, 12 of 13 mid-turn status texts vanished from both the display and the JSONL.

CLAIM C015 Supported 2026-10-20

GitHub issue #87640 in anthropics/claude-code, opened 2026-08-18 by sagimedina1, records a report that the Fable 5 `[reasoning_extraction]` safeguard rejected a message whose user text was the single word `Hi`; the reporter states the classifier guards against prompts attempting to extract chain-of-thought, that nothing in the conversation did that, and that the same session context works normally on Opus 5, pointing at Fable-only safeguards rather than at the conversation.

CLAIM C009 Supported 2026-10-17

GitHub issue #22543 in anthropics/claude-code, opened 2026-02-02 by mjbyt, records a report that after using the cowork feature Claude Desktop becomes extremely slow, alongside a VM bundle file that grows to 10GB, is never cleaned up, and regenerates quickly after deletion; the report documents a cleanup test showing approximately 75% faster performance immediately after removing the bundle, while also noting performance degrades again over time even with the bundle at 0 bytes, suggesting a memory leak or accumulating work.

CLAIM C003 Supported 2026-10-03

GitHub issue #97854 in anthropics/claude-code, opened 2026-09-28 by timothe-rey, records a report that during a Claude Code session with Auto mode active, the server-side safety classifier intermittently returned no verdict for consecutive tool calls including Bash and ScheduleWakeup, resulting in total failure before self-resolving; subsequent in-thread discussion through October 2026 records further user reports across macOS, Windows, Linux, and WSL debating whether the failure is transient or requires terminating the session.