Forumics logo Forumics
Issue 01

Forumics · Browse by tool

Codex known issues

37 verified issues published.

CLAIM C161 Supported 2026-11-19

GitHub issue #50728 in openai/codex, opened 2026-10-03 by kamiljakubczyk, records a report that Codex native Computer Use repeatedly disrupts manual typing and pasting in Claude Desktop and Opera on macOS, with keystrokes and image pastes duplicating; a read-only CoreGraphics event-tap inspection confirmed two enabled, non-listen-only keyboard taps owned by the same SkyComputerUseService process and attached to Claude, with reported average latencies of approximately 15.7 and 55.6 seconds; the report states this is an observed incident sequence, not a deterministic minimal reproduction, and was filed against Codex desktop app 26.930.31730 on macOS 26.5.

CLAIM C165 Supported 2026-11-19

GitHub issue #20864 in openai/codex, opened 2026-05-03 by 8th-block, records a report that Codex Desktop App becomes slow/unresponsive when ~/.codex/sessions contains many rollout JSONL files — 109 files totalling 78M in the reporter's case, against 22 entries in session_index.jsonl; the report says Desktop appears to scan/read the entire sessions folder even though session_index.jsonl contains the much smaller Desktop-visible set, and clearing or moving files out of ~/.codex/sessions immediately fixes the lag; the report was filed against Codex Desktop App 26.429.30905 on macOS (Apple Silicon).

CLAIM C156 Supported 2026-11-18

GitHub issue #26984 in openai/codex, opened 2026-06-08 by jacobcxdev, records a report that long-running sessions using stdio MCP servers slowly accumulate open file descriptors in the codex parent process until they hit the per-process soft limit (256 on a default macOS login shell), after which every subsequent open()/spawn() fails with Too many open files (os error 24); the report attributes the leak to orphaned MCP child processes the parent still holds read pipes to, with 186 anonymous PIPE fds observed for a session with about 4-5 configured MCP servers, and was filed against codex-cli 0.137.0.

CLAIM C150 Supported 2026-11-17

GitHub issue #46110 in openai/codex, opened 2026-09-17 by harboriqdev, records a report that Codex 0.155.0-alpha.15 fails before executing any filesystem-restricted command on a native Ubuntu host when /proc/self/mountinfo contains valid nsfs entries created by snapd, with the error "error building bubblewrap command: mountinfo path is not absolute", because daemon_mounts.rs passes the mount root through mount_path(), which rejects any decoded path for which PathBuf is_absolute is false; the report was filed against Ubuntu 26.04.1 with bubblewrap 0.11.1 and snapd providing persisted mount namespaces under /run/snapd/ns.

CLAIM C151 Supported 2026-11-17

GitHub issue #33493 in openai/codex, opened 2026-07-16 by snrui, records a report that a long-running, image-heavy Codex Desktop thread enters a repeated auto-compaction cycle because the compacted replacement_history keeps historical input_image items including their full inline base64/data-URL payloads; in the reported thread the rollout JSONL was approximately 4.16 GB with 224 compaction records and 92 retained images carrying approximately 26.47 million inline image/base64 characters, leaving only about 8,000 tokens of headroom below the auto-compaction threshold; the report was filed against Codex Desktop 26.707.72221 with embedded CLI codex-cli 0.144.2 on macOS 26.3.

CLAIM C155 Supported 2026-11-17

GitHub issue #33624 in openai/codex, opened 2026-07-16 by erikmagnethi, records a report, presented as a publicly reported incident rather than a reproduction, that a GPT-5.6 Sol Ultra-mode subagent running with Full Access expanded a cleanup path incorrectly and recursively deleted nearly all files in the user's Mac home directory; the report states this concerns most contents of the home directory, not independent proof that the entire macOS installation or physical computer was erased, and asks for a hard confirmation and recovery gate for bulk and home-directory deletion even in Full Access, citing a Tech Times report and OpenAI's GPT-5.6 Preview System Card as public evidence.

CLAIM C134 Supported 2026-11-15

GitHub issue #60058 in openai/codex, opened 2026-11-07 by bradcarroll, records a report that in ChatGPT desktop on Windows, Work mode, app version 26.1103.32239, the new-chat Work composer shows Full access but the resulting conversation's execution context reports workspace-write with restricted network access, and the Full access control is visible only before starting the chat and then disappears, leaving no visible in-thread way to inspect or correct the selected mode; a harmless, explicitly authorized SSH check fails at the first jump-host TCP connection, before SSH authentication, with the SSH child process returning exit 255; fully restarting the ChatGPT desktop app did not change the result, and returning to the same conversation through the web interface produced the same result from its local command runner; the report lists related reports #57671 and #59528 and asks for consolidation with an existing issue if appropriate.

CLAIM C138 Supported 2026-11-15

GitHub issue #25015 in openai/codex, opened 2026-05-29 by LostFrxks, records a report that Codex app-server starts additional stdio MCP process stacks when spawning subagents, and after the subagents complete and close_agent is called, some MCP child process trees remain alive under the long-lived codex app-server, so process count and memory usage grow roughly linearly with subagent usage; in a small live repro, 3 trivial subagents took the app-server from 23 descendants and about 203 MiB RSS to 74 descendants and about 3.7 GiB RSS, and after close_agent 9 MCP launcher roots and 33 processes remained; in a larger run, 20 subagents took descendants from 46 to 266 and total RSS from about 2.9 GiB to about 11.7 GiB, with swap usage growing to about 6.4 GiB; the report was filed against codex-cli 0.128.0 on Ubuntu 24.04.4 LTS with 16 GB RAM and 16 GB swap.

CLAIM C139 Supported 2026-11-15

GitHub issue #49829 in openai/codex, opened 2026-10-01 by xc1427, records a report that opening a Dot-created task in ChatGPT desktop fails with "Failed to resume chat" and "Codex app-server is not available", while the same task opens successfully in ChatGPT web; on a network where direct access to codex-cloud-backend.chatgpt.com times out but an approved proxy works, logs repeatedly show hostKind=durable, failureReason=open_timeout, and WebSocket close code 1006, and installed-code inspection suggests the bundled WebSocket library supplies a custom createConnection without a proxy agent, bypassing Node's environment-aware default agent; the report was filed against ChatGPT desktop 26.928.31416 on macOS with a ChatGPT Pro subscription.

CLAIM C119 Supported 2026-11-14

GitHub issue #42501 in openai/codex, opened 2026-09-03 by leonfetter, records a report that after updating the Microsoft Store Codex Windows app to 26.901.1978.0, Codex launched several ChatGPT.exe processes but never created a visible window, with all processes responsive but MainWindowHandle remaining 0; investigation showed the new bundled cua_node runtime repeatedly failed during staging, each launch creating a new incomplete staging directory that contained node.exe but was consistently missing node_repl.exe, and manually staging the packaged cua_node runtime with xcopy /G into the expected versioned runtime directory fixed the problem immediately, after which Codex launched normally; the report presents the EFS / Application Protected WindowsApps relocation issue as the apparent cause, not a confirmed root cause.

CLAIM C121 Supported 2026-11-14

GitHub issue #29908 in openai/codex, opened 2026-06-24 by legka, records a report that on Ubuntu 24.04 with Bubblewrap 0.9.0, apply_patch and ordinary managed sandbox commands fail before the requested command can run, with the error bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted; the report attributes the failure to Bubblewrap sandbox setup, not repository file permissions, noting minimal Bubblewrap tests fail with uid-map and loopback errors under AppArmor unprivileged user namespace restrictions, and asks for the apply_patch and managed sandbox path to be made robust on Ubuntu systems with such restrictions or for a supported fallback that does not require changing host-wide security settings.

CLAIM C122 Supported 2026-11-14

GitHub issue #28058 in openai/codex, opened 2026-06-13 by ignatremizov, records a report that after #26210 (Encrypt multi-agent v2 message payloads, merged 2026-06-05, post-0.137.0), encrypted message delivery is the default for MultiAgentV2 spawn_agent, send_message, and followup_task, so users cannot inspect the exact delegated tasks, follow-up instructions, or intermediate child-to-parent messages as readable text in parent activity rows, child transcript/history, or after resuming a session; the report asks for a user-choosable readable delivery and a structured operational transcript with sender/recipient attribution, states the requested audit trail does not require exposing private reasoning, and notes that #35845 added conditional plaintext-response handling without changing the default, a position the body reaffirms in a source review dated 2026-10-26.

CLAIM C123 Supported 2026-11-14

GitHub issue #57721 in openai/codex, opened 2026-10-28 by fengzee, records a report that on macOS with Codex desktop app 26.1023.50603 and the Android ChatGPT app connected through Codex/Remote, a new chat created entirely from the phone cannot access the connected Mac's Chrome browser and reports Local browser use is disabled in Work settings although the setting is enabled and both devices use the same account/workspace; sending a message in the same chat from the Mac starts browser access, after which the phone can use it, while chats created on the Mac work when continued from mobile, and restarting both apps and creating a fresh mobile chat does not fix the failure; the report expects mobile-created remote chats to initialize browser access just like desktop-created chats and says initialization failures should not be reported as a disabled setting.

CLAIM C126 Supported 2026-11-14

GitHub issue #29915 in openai/codex, opened 2026-06-24 by andrewsohrabi, records a report that permission and approval mode selection is not reliably persisted for new threads or already-loaded existing threads: on a new thread, choosing a non-workspace mode such as Full access falls back to workspace after the thread starts and the first prompt executes, while the picker still looks selectable before submission; on an existing loaded thread, changing approval and permission mode through resume appears to work in the UI, but the live thread keeps the old effective settings; the report proposes two linked fork commits as fix candidates, one routing loaded-thread resume overrides through the validated thread-settings path and one exposing permissions plus activePermissionProfile for thread start, resume, and fork in the stable protocol fixtures, and reports focused local tests passing.

CLAIM C108 Supported 2026-11-13

GitHub issue #42695 in openai/codex, opened 2026-09-04 by HS7097, records a report that a task delivered through `send_message_to_thread` on Codex Desktop on Windows (bundled CLI 0.153.0) was received and partially executed correctly, but after automatic mid-turn compaction the receiving agent reverted to an older, already-completed instruction instead of finishing the current task; that its final response then incorrectly reported that no task execution had occurred, despite successful tool results and files created earlier in the same turn; that the incoming Task B instruction was recorded as a `function_call_output` associated with `send_message_to_thread` while the replacement history retained the older Task A instruction as a visible user message; and that the opaque compaction item's contents cannot be inspected, so the report does not establish that Task B was entirely absent from the compressed summary. The report notes whether this shares a root cause with related reports requires investigation, and no root cause is asserted.

CLAIM C109 Supported 2026-11-13

GitHub issue #39054 in openai/codex, opened 2026-08-17 by rmanalan, records a report that when an MCP server rejects a refresh with a spec-correct `400 invalid_grant`, Codex does not discard the refresh token, re-sends the identical token on every subsequent attempt indefinitely, and never falls back to a fresh authorization flow; that the app layer is simultaneously told the server is still authenticated, so no re-authenticate affordance appears; that the user sees the MCP server present but with zero tools; that the report was reproduced on Codex CLI 0.140.0, 0.144.0, 0.146.0, 0.147.0 and 0.148.0-alpha.20 on macOS 26.5.2 arm64; and that the report attributes the behavior to the credential being judged usable purely because a non-empty refresh-token string is present, with the rejected refresh token never recorded. No root cause beyond the reporter's stated attribution is asserted.

CLAIM C110 Supported 2026-11-13

GitHub issue #25770 in openai/codex, opened 2026-06-02 by zyzhen, records a report that Codex Desktop from Microsoft Store repeatedly fails to update on Windows 10 Enterprise (version 10.0.19044, x64); that the update reaches the AppX/MSIX registration phase and then fails because Windows still considers the old `OpenAI.Codex` package/application to be in use, with AppXDeployment-Server failures including `0x80073CF9`, `0x8007001F` and `0x80073D02`; that this happened repeatedly on the same machine, and also on another machine in the same LAN; that closing visible Codex windows is not enough, since Windows still reports the old package/app as in use; and that the update succeeded only after manually quiescing the old package state and reinstalling from Microsoft Store. The report notes the failure appears to be in the Windows Store/MSIX update lifecycle around releasing the old package identity, rather than in normal application runtime behavior after the package is successfully installed. No root cause beyond the reporter's stated attribution is asserted.

CLAIM C113 Supported 2026-11-13

GitHub issue #50520 in openai/codex, opened 2026-10-03 by gwdonnor, records a report that the native Windows sandbox repeatedly fails before shell commands start with `helper_sandbox_lock_failed` when opening the `.codex\.sandbox-bin` directory, that granting SYSTEM and Administrators Full Control on that directory and retrying Windows setup restores sandbox access but the problem later recurs with the ACL entries again at Modify, that this is an observed recurrence with a successful workaround and not a confirmed diagnosis of which component resets the ACL, and that the report was observed October 1-2, 2026 with command-runner 0.160.0 in workspace-write filesystem mode, with no root cause asserted.

CLAIM C118 Supported 2026-11-13

GitHub issue #57101 in openai/codex, opened 2026-10-26 by americanmademischief, records a report that on Windows with Codex desktop package OpenAI.Codex 26.1021.3705.0 and Computer Use plugin 26.1021.31345, a gold/yellow border remained around the display after a Codex chat had finished its native desktop capture work, with the Windows Computer Use helper still running and its Windows capture-access record still open approximately five hours after the screenshot-producing turn completed; the report asks for capture resources to be released when their owning work ends while preserving Computer Use for other chats that actually need it, and labels the lifecycle-ordering explanation a hypothesis, not a confirmed root cause.

CLAIM C102 Supported 2026-11-12

GitHub issue #58447 in openai/codex, opened 2026-10-30 by Robert-Fish, records a report that a dot cannot read an existing user-created native Codex thread on a connected macOS desktop through the documented cloud_threads.read route, despite the desktop reporting connected, attached, and authorized for tasks; that a control read of an existing dot-created cloud task succeeds; and that the read fails with DesktopThreadReadUnavailableError (CONFLICT / desktop_thread_read_unavailable), with no root cause asserted.

CLAIM C104 Supported 2026-11-12

GitHub issue #59855 in openai/codex, opened 2026-11-06 by ShakhrukhDR, records a report that opening an affected cloud-managed chat from the Codex Desktop sidebar on Windows can show "Failed to resume chat / Invalid request: AbsolutePathBuf deserialized without a base path" before a new message is submitted, while reading the stored conversation can still succeed; the report asks Desktop to resolve workspace paths when resuming cloud-managed chats from Windows and notes which field triggers the error remains unknown.

CLAIM C100 Supported 2026-11-11

GitHub issue #25178 in openai/codex, opened 2026-05-30 by Define1165250535, records a report that on Windows Codex Desktop on Windows 10 Pro 22H2 (build 10.0.19045), any get_window_state call that requests a screenshot fails before capture with "SetIsBorderRequired failed: 不支持此接口 (0x80004002)", that the same target window works when requesting accessibility text only, and that the reporter traces the failure to the capture helper calling GraphicsCaptureSession.IsBorderRequired / SetIsBorderRequired on a Windows build that does not support that property, which Microsoft documents as introduced in Windows 10 version 2104 / build 10.0.20348.0.

CLAIM C087 Supported 2026-11-10

GitHub issue #41220 in openai/codex, opened 2026-08-27 by FromAriel, records a meta-issue tracking report that many open Codex reports describe a common symptom family in which Codex subscription quota or purchased credits are depleted materially faster than the user's prior baseline or local and token evidence would predict, sometimes with abrupt changes during an unchanged usage window, passive or background consumption, or usage records that do not reconcile with the meter, and that the tracker is intended to link and compare those reports in one place while not assuming they all have one root cause.

CLAIM C082 Supported 2026-11-09

GitHub issue #54685 in openai/codex, opened 2026-10-16 by n0rpan123, records a report that a naturally required cross-task execution review denied a bounded QA installation because the trusted user context contained only an abbreviated Approved with omitted context, even though the exact approval request contained the complete original proposal and reply, original message identifiers and hashes, and a read-only original-record locator. The final decision was risk level high, user authorization unknown, outcome deny, and the report asks maintainers whether this is a trusted-context retention or attribution defect, a reviewer evidence-verification failure, or a currently unsupported cross-task authorization path.

CLAIM C057 Supported 2026-11-08

GitHub issue #38455 in openai/codex, opened 2026-08-13 by flannick, records a report that ChatGPT desktop 26.810.41047, build 6570, on macOS 15.7.7, Apple Silicon, 32 GB RAM, repeatedly spawns Computer Use workers and crashes with a V8 out-of-memory error, that the previous version 26.730.61639 worked, that the crash reproduces while idle 98 seconds after launch through SIGABRT in node::OOMErrorHandler, with 316 threads at crash, 187 named computer-use, and telemetry showing 78 unexplained child processes consuming about 4.96 GB after 29 seconds, and that the Browser and Computer Use flags were explicitly disabled.

CLAIM C059 Supported 2026-11-08

GitHub issue #34260 in openai/codex, opened 2026-07-20 by RocStone, records a report that Codex Desktop 26.715.4045.0 on Windows 25H2 x64, build 26200.8875, can enter an unbounded process-cleanup loop where hundreds of taskkill.exe processes remain alive, each with a corresponding conhost.exe, that the accumulated taskkill instances repeatedly query Win32_Process, exhaust the WMI provider quota, and make the entire Windows UI sluggish or unable to launch new processes, with peak measurements of 457 taskkill.exe processes, 473 conhost.exe processes, 1,388 total processes, over 14,000 threads, over 350,000 handles, and WMI provider private memory of 720.5 MB, and that killing the accumulated taskkill.exe processes immediately reduced the total process count from 1,388 to about 565 while WMI private memory fell to about 104 MB, but taskkill.exe counts grew again from 156 to 320 in 20 seconds while the original Codex process remained running.

CLAIM C063 Supported 2026-11-08

GitHub issue #45163 in openai/codex, opened 2026-09-13 by lightninglu10, records a report that in codex-cli 0.154.0 on macOS, Darwin 25.6.0, arm64, after the terminal changes between light and dark appearance while Codex is already running, the composer and previously rendered user-message blocks retain their old background colors, that the terminal default foreground updates, producing nearly white text on a pale background, so long-lived sessions become unreadable after an automatic macOS appearance change, that refocusing and redrawing do not refresh the cached palette, and that restarting refreshes the startup colors but the next appearance change can break the session again.

CLAIM C070 Supported 2026-11-08

GitHub issue #35420 in openai/codex, opened 2026-07-26 by hiroki-tamba-research, records a report that in the ChatGPT Work / Codex surface on Windows, requests repeatedly fail with stream disconnected before completion, that the Windows workspace being used is OneDrive-backed and the OneDrive client is in a degraded state because it reports that the Microsoft 365 subscription payment could not be processed, that Codex cannot complete work in the selected workspace, and that the evidence establishes a strong correlation between a OneDrive-backed workspace, a degraded OneDrive state, and repeated stream disconnections, while not yet proving which layer is the root cause.

CLAIM C042 Supported 2026-11-07

GitHub issue #57084 in openai/codex, opened 2026-10-26 by interconnectedMe, records a report that on Linux / VS Code with ChatGPT Pro authentication, recurring "Selected model is at capacity" and "exceeded retry limit, last status: 429 Too Many Requests" errors interrupted ordinary Codex work on October 25-26 despite 84-99 percent weekly allowance remaining, that several October 25 retries converged on common deadlines within 0.7 seconds, and that the precise rejecting condition remains unknown.

CLAIM C044 Supported 2026-11-07

GitHub issue #50428 in openai/codex, opened 2026-10-02 by mattalberta, records a report that Windows Codex Desktop can read an existing durable/cloud chat bound to a Windows execution environment, but direct plain-text submissions and a same-directory fork fail before execution with "invalid turn/start params: AbsolutePathBuf deserialized without a base path", that the recorded JSON-RPC code was -32602 with failureReason invalid_config, and that the same failure appeared in desktop versions 26.928.31416 and 26.930.21537.

CLAIM C047 Supported 2026-11-07

GitHub issue #50168 in openai/codex, opened 2026-10-02 by IaHorrorLabs, records a report that Dot can list and read Codex Cloud tasks but the exposed cloud_threads write path fails for both task creation and follow-up messaging, that new cloud task creation returned an app-server backend failure with error_code UNKNOWN and no usable thread or turn identity, and that a follow-up message to a manually created healthy cloud thread returned CloudThreadNotFoundError with error_code NOT_FOUND, observed on October 1, 2026.

CLAIM C053 Supported 2026-11-07

GitHub issue #35259 in openai/codex, opened 2026-07-24 by dimasyankauskas, records a report that during intentional Ultra and multi-agent work on Codex Desktop 26.721.31836 with codex-cli 0.146.0-alpha.3.1 on macOS, the desktop repeatedly re-entered the model merely to wait for agents or poll terminal status, that in a corrected reset-to-49-percent usage window wait and status polling turns accounted for 19.8 percent of raw local token volume, and that the polling often ran at a 30- or 60-second cadence for hours while delegated work was still running.

CLAIM C054 Supported 2026-11-07

GitHub issue #50769 in openai/codex, opened 2026-10-04 by jabdori, records a report that during October 3-4, 2026, development coordinated through dots hit repeated pre-execution approval blocks although the user had authorized development and publication, that later denials kept citing read-only scope or missing authorization, that a branch-creation call returned a user-cancelled message the user said they had not triggered, and that moving the workflow to gh CLI on the connected Mac still left branch creation rejected before execution, on macOS with GitHub CLI 2.101.0 and unknown Codex, client and backend builds.