Digest / April 20, 2026
Codex Issues Digest — April 20, 2026
92 new issues filed ·
40 resolved in openai/codex (UTC day).
New issues (92)
- Github Plugin not showing organization
- Codex exec does not fire Stop / SessionStop / PostToolUse hooks (only UserPromptSubmit fires)
- Allow slash commands to be searchable in English regardless of the selected UI language
- Improve Plan mode decision prompts with richer context and a standard more-context path
- Codex closeout introduced ungrounded “user-directed exception” framing in audit narration, plus compact/link follow-up failure
- Feature request: add documented ChatGPT device-code auth support to codex app-server
- Mac Codex Desktop misclassifies sessions from different projects under the same remote worktree as a single project
- Sandboxed shell commands fail on Windows with `CreateProcessWithLogonW failed: 1326` / `1909`
- CLI output & thinking order error, out of sequence & disordered
- First click to display menu does nothing
- Windows: "Open with VS Code" fails even when VS Code works normally outside Codex
- Unable to change reasoning effort
- Desktop threads can be poisoned by inline base64 tool images, leading to `{"detail":"Bad Request"}` on resume and likely inflated token usage
- Support file mentions in `/review` custom review instructions like normal chat.
- Ability to reliably generate transparent images
- Improve resume hint rendering
- Codex App cannot switch accounts, and the logout button is unresponsive
- Codex Desktop can sometimes save and display a garbled thread title instead of a normal auto-generated title.
- threads / projects in sidebar missing
- Codex scans filesystem for Git repositories without permission, triggers runaway under certain conditions
- [Bug] Mini window: Permission request cards duplicated and no visual feedback after approval
- [Redacted by author]
- Image paste fails in some Codex CLI sessions (Ctrl+V and Right-click Paste both fail)
- Windows desktop app intermittently fails streamed responses behind Clash Verge Rev proxy (TLS EOF / response body decode error)
- Windows: npm installs 0.121.0 but resolves to 0.119.0-win32-x64 binary
- Add Meeting recorder
- Codex CLI incorrectly treats Plus account as Free plan
- Support for running multiple simultaneous profiles in the app
- Stdio MCP servers can trigger terminal job-control suspension during startup on macOS
- Cannot scroll up/down to read full plan output in Codex CLI in integrated terminal in VS Code
- VS Code extension spawns runaway R child processes (R --file=- / R -e) that pin CPU on macOS
- Codex-cli lag on every prompt when "background" apps/plugins discovery requests not disabled
- `file_change` fails under `workspace-write` when `sandbox_workspace_write.network_access = true`
- Quota fully consumed after a single message
- Windows desktop shell exposes bundled rg.exe, but executing it fails with Access is denied
- Plus Codex Limits Are Too Low for Real Development Work
- Bug: in apply-patch silently swallows mismatched replacement ranges
- Bug: rejects update_plan in Plan mode — contradictory error message
- Bug: Race condition in memory Phase 2 agent status subscription
- Bug: Flaky tests left ignored without resolution
- Doc: SDK README uses incorrect case for zod-to-json-schema target value
- Bug: Review turns lack TurnStarted event, inconsistent with other task types
- Security: init_firewall.sh does not re-resolve DNS, allowing DNS rebinding bypass
- Windows Codex app: npm/network DNS fails in sandbox despite network enabled; cannot reach local proxy
- Resume session: stream disconnected before completion
- Feature request: make request compression explicit instead of conditional behind enable_request_compression
- Codex App: make the file tree/sidebar resizable or show full filenames
- No usage after buying plus plan
- can’t see my Codex usage, and the dictate feature also isn’t working
- 403 forbidden error.
- We're currently experiencing high demand, which may cause temporary errors.
- Feature request: batch-spawn forked subagents from a shared preloaded context
- User turns can continue in completed subagent thread after parent spawned it, leaving subagent role active
- Codex cannot connect
- 403 Forbidden error in the middle of normal usage
- Error running remote compact task: unexpected status 403 Forbidden: Unknown error, url: https://chatgpt.com/backend-api/codex/responses/compact, cf-ray: 9ef4e4f49ed1fdb1-PDX Version 26.415.40636 (1799)
- macOS: Cloudflare 403 blocks all chatgpt.com/backend-api/codex/responses requests — CLI unusable while browser works on same network
- Codex suddenly stopped working for me
- rollout file not written until first API turn completes, no way to get thread_id at spawn time for embedding tools
- I suggest OpenAI fully acquire Godot and Blender and launch AI collaboration skills.
- GPT-5.4 Fast mode often feels no faster than Standard, but still consumes credits at 2x
- Desktop performance collapses in profiles with a few very large local conversation histories (typing, scrolling, thread list, random exits)
- Treat git global pager options as unsafe for auto-approval
- Usage fully consumed after a single message only For 2 weeks
- Codex Desktop: Enabling Memories causes disproportionate quota usage
- MCP Chained Tool Routing Failure
- Headed screenshot capture on Windows is not deterministic enough for bounded UI proof passes on my display setup.
- Codex Desktop: BrowserOS MCP "Always allow" does not persist and permission prompts keep reappearing
- codex hallucinates by ending the turn responding to previous turn instructions
- Allow edit on any message, not just the last one
- We have to open Codex App to load chats in IDE extension
- `is_likely_sandbox_denied` false positive when command output contains "sandbox" in file paths
- Codex Desktop ignores `plan_mode_reasoning_effort` when entering Plan mode, while CLI applies it correctly
- this is just an example why CODEX Is better, codex would ask, what do you mean by refactour this codebase
- /copy always copies most recent message, even after rewinding
- Codex auto-compact loses critical task details and causes wrong execution, even on small fixes
- Cannot use Codex due hitting the limits, though I didn't reach them
- Codex CLI sessions become unrecoverable after ~37-40 minutes on Windows 11 after websocket disconnects
- shortcut to open browser with side panel all at once
- Desktop app does not show older local CLI/TUI sessions; pinned chats still open, but unpinned CLI sessions never appear in app
- Memory apparently ignored by Codex until specifically asked to read it
- Codex should provide meaningful OSC 0/2 for Terminal Title
- /status shows stale usage limits on first run, requires a second /status to display correct values
- Bug: Codex macOS app ENOENT when viewing remote repo over SSH
- memory creation refusing /// not adding new memorise up to a point
- macOS packaging error: SkyComputerUseClient built for macOS 15.0 crashes on macOS 14.x
- Terminal combined view
- We're currently experiencing high demand, which may cause temporary errors.
- Switching models mid-session does not update provider/endpoint — Ollama endpoint persists after switching to gpt-5.4
- We're currently experiencing high demand, which may cause temporary errors.
- session starts up with an inscrutable MCP nastygram
- Codex treats skill injection via $ command as a request to read the skill.md file.
Resolved issues (40)
- Codex exec does not fire Stop / SessionStop / PostToolUse hooks (only UserPromptSubmit fires)
- Github Plugin not showing organization
- Codex closeout introduced ungrounded “user-directed exception” framing in audit narration, plus compact/link follow-up failure
- Fast Mode has disappeared from Codex App
- Codex App is misreporting the state of /fast mode
- Codex App: Active conversation in the sidebar has insufficient contrast
- Unable to change reasoning effort
- write_stdin dispatch is not gated on exec session stdin liveness
- Codex resume does not show the id in 0.121
- `file_change` fails under `workspace-write` when `sandbox_workspace_write.network_access = true`
- Shell snapshot logs false ENOENT on temp cleanup after thread/start
- No usage after buying plus plan
- Codex Desktop can sometimes save and display a garbled thread title instead of a normal auto-generated title.
- Codex cannot connect
- Windows: npm installs 0.121.0 but resolves to 0.119.0-win32-x64 binary
- Quota fully consumed after a single message
- Plus Codex Limits Are Too Low for Real Development Work
- Bug: rejects update_plan in Plan mode — contradictory error message
- Bug: Race condition in memory Phase 2 agent status subscription
- Bug: Flaky tests left ignored without resolution
- Bug: Review turns lack TurnStarted event, inconsistent with other task types
- /mcp output regressed starting in codex-cli 0.116.0: missing Status/Resources/Resource templates
- Codex suddenly stopped working for me
- can’t see my Codex usage, and the dictate feature also isn’t working
- We're currently experiencing high demand, which may cause temporary errors.
- rollout file not written until first API turn completes, no way to get thread_id at spawn time for embedding tools
- Usage fully consumed after a single message only For 2 weeks
- Codex Desktop: Enabling Memories causes disproportionate quota usage
- codex hallucinates by ending the turn responding to previous turn instructions
- Clear context before implementing plan
- this is just an example why CODEX Is better, codex would ask, what do you mean by refactour this codebase
- Sandboxed apply_patch falls back to vendored bwrap because exec-server fs helper clears PATH
- Desktop app does not show older local CLI/TUI sessions; pinned chats still open, but unpinned CLI sessions never appear in app
- Codex Desktop on WSL: apply_patch always requests approval in writable workspace, while normal shell writes do not
- [sandbox_workspace_write].network_access = true causes fs sandbox helper write failures on Linux
- TUI exits to shell on subagent/thread switch when startup skills/list fails; parent rollout stays empty
- /copy always copies most recent message, even after rewinding
- We're currently experiencing high demand, which may cause temporary errors.
- Switching models mid-session does not update provider/endpoint — Ollama endpoint persists after switching to gpt-5.4
- tui onboarding: warn when trusting git subdirectories