7 Open Issues in 5 Coding Agent Repos, and a Workaround for Each

• Book: Event-Driven Architecture Pocket Guide
• The series: Pocket Guides for Developers — 8 short books on agents, RAG, prompts, observability, events, databases and system design — the whole series
• My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools
• Me: xgabriel.com | GitHub

---

The Gemini CLI epic for parallel tool calling has zero thumbs-up. Two of its three sub-issues are already closed. The Codex bug where the VS Code extension swallows your queued messages has 3 thumbs-up and 28 comments, from people hitting it on Windows, macOS, Linux and WSL.

Vote counts are a rough map of what hurts. The seven open issues below each cost someone time this week, and each has a workaround you can use before a maintainer gets to it.

Counts are the 👍 reactions on each issue's opening post, as the GitHub API reported them on October 18, 2026.

For scale: what tops each tracker

None of the seven is its repo's most-upvoted open issue. Sorted by 👍, the leaders look like this:

Repo	Top open issue by 👍	👍
anthropics/claude-code	#45596 "Bring Back Buddy"	1,194
openai/codex	#8745 LSP integration for Codex CLI	503
anomalyco/opencode	#11176 official VS Code extension	162
google-gemini/gemini-cli	#19532 add gemini-3.1-pro-preview	110
github/copilot-cli	#223 "Copilot Requests" permission for org-owned tokens	80

Those are mostly feature requests. The seven below are mostly friction: a file you can't save, a message that never sends, a binary macOS won't open.

Claude Code and Codex

1. Claude can't write to Claude Project files (anthropics/claude-code #16550, 78 👍). It is filed in the Claude Code repo, but it's about Projects in the Claude app, labelled area:desktop. Claude reads the project's files and hands you a download when it changes one, so you delete and re-upload by hand. Open since January 7, 2026. Two commenters on October 10 and 13 said Claude had just edited their project files directly. No Anthropic reply is visible in the thread, and the issue is still open.

Workaround: move the files somewhere an agent is allowed to write. One commenter pointed out that in Claude Code, Claude keeps state files up to date on its own, and git can track the changes. Another moved the project into Cowork, which per Anthropic's help page "can only read and write files in folders you've connected," on paid plans in the desktop app. For the Claude Code route, copy your project files in and commit them first, so git diff shows only what Claude changed:

mkdir -p ~/code/project-notes && cd ~/code/project-notes
git init
cp ~/Downloads/checkpoint_log.md .
git add -A && git commit -m "baseline"
claude "Read checkpoint_log.md and add today's decisions"
git diff

2. The Codex VS Code extension drops queued messages (openai/codex #50403, 3 👍). You send a follow-up while Codex is working, and it never goes out, or it goes out an hour later. The Output log shows Failed to release queued message send lock with a JSON parse error. Commenters report stuck queues, vanished messages and a few prompts that ran twice. Open since October 2, labelled bug, extension, windows-os and app-server, with no staff reply in the thread as of October 18.

What people in the thread are doing:

• Command Palette → Developer: Reload Window. Several people say it sends the stuck prompt or unblocks the queue. It comes back.
• Update. One commenter on macOS says 26.930.21537 fixed it and kept GPT-6.1 Sol; nobody has confirmed that on Windows yet.
• Roll back to 26.917.62051. That worked for one commenter, but the build lacks GPT-6.1 Sol.
• Use the Codex CLI in a terminal. The bug is in the extension's queue, and one commenter used the bundled codex binary as a control that ran reliably.

code --install-extension openai.chatgpt@26.917.62051 --force

If you roll back, turn off Auto Update for that extension (Extensions view → gear icon), or VS Code will move you forward again.

Gemini CLI

The Gemini CLI README says that for unpaid-tier and Google One users, Gemini CLI "was replaced by Antigravity CLI on June 18th, 2026." These two issues matter if you still run Gemini CLI on a paid key or an enterprise plan.

3. Sub-agents, the "Post V1.0" epic (google-gemini/gemini-cli #3132, 50 👍, 98 ❤️). Opened July 3, 2025, asking for a reusable SubAgent component, so tools like the edit tool could run their own recovery loop without each one building it from scratch. It is the third most-upvoted open issue in the repo, labelled priority/p3. The feature it asked for shipped anyway: Google's April 15, 2026 post announced subagents, and the docs say "Subagents are enabled by default." The epic is a roadmap rollup marked maintainer only.

Workaround: use what's there. Call a built-in one with @, or drop a Markdown file in .gemini/agents/ (name and description are the required fields):

---
name: security-auditor
description: Finds security issues. Reports, never edits.
tools:
  - read_file
  - grep_search
---
You review code for vulnerabilities. Explain each one
and suggest a fix. Do not change any files.

Then @security-auditor review src/auth in the session.

4. Parallel tool calls (google-gemini/gemini-cli #17120, 0 👍). The epic has three sub-issues. Parallel read-only calls (#17117) closed as completed on March 10, 2026. A debugging mode for it (#17122) closed in May. Parallel writes "where safe" (#17121) is open. So reads fan out today, and edits run one at a time.

Workaround for writes: parallelize across checkouts, not inside one session. Google's own subagent post warns that parallel agents editing code at the same time can conflict and overwrite each other. Separate worktrees keep each agent out of the others' checkout:

git worktree add -b fix-auth ../app-fix-auth
git worktree add -b fix-billing ../app-fix-billing
# start one agent in each folder, merge when done

This is what Hermes IDE, which I build, does on ⌘N (Ctrl+Shift+N on Windows and Linux): each task starts on its own hermes/ branch in its own worktree, with the agent you pick for it.

Hermes 2.3.4: Codex and Claude Code side by side, each in its own hermes/ worktree, with status, context use, cost and branch in the sidebar.

OpenCode

5. Agent Plugins support (anomalyco/opencode #40993, 19 👍). Agent Plugins is a vendor-neutral format with a plugin.json manifest that bundles skills and MCP servers. Its steering committee lists maintainers from Amazon, Cursor, Microsoft, OpenAI and Vercel. The request, opened August 7, asks OpenCode to load those bundles. No labels and no maintainer reply yet.

Workaround: unpack the plugin by hand. OpenCode's skills docs list .agents/skills/<name>/SKILL.md as a project path it reads, along with .opencode/skills and .claude/skills. Put the skills there and the MCP server in opencode.json under mcp (format). The issue's author keeps one shared skills folder and symlinks it into each agent. Claude Code's skills docs say a skill entry in .claude/skills can be a symlink to a directory elsewhere:

mkdir -p .claude/skills
ln -s ../../.agents/skills/release-notes \
  .claude/skills/release-notes

6. V2 shipped without release notes (anomalyco/opencode #52184, 31 👍). Opened September 29. The public changelog stopped at v1.18.33, opencode.ai/v2/changelog returned 404, and the v2.0.20 GitHub release had a placeholder body. The reporter compared tags v2.0.19...v2.0.20 and found 31 commits across 179 files, including auth and permission changes.

Workaround, part one: on October 12 a commenter pointed to a CHANGELOG.md on the dev branch. It describes itself as "Release notes for OpenCode V2, newest first" and its top entry is v2.0.32, dated October 16. Part two, for any release it misses, do what the reporter did:

gh api repos/anomalyco/opencode/compare/v2.0.19...v2.0.20 \
  --jq '.commits[].commit.message | split("\n")[0]'

That prints the first line of every commit between the two tags. Swap in the versions you're moving between before you upgrade.

Copilot CLI

7. Gatekeeper blocks copilot after every Homebrew upgrade (github/copilot-cli #970, 23 👍). On a managed Mac with Gatekeeper set to App Store and known developers, each upgrade brings back "Apple could not verify 'copilot' is free of malware." Opened January 14, 2026, labelled area:installation, no GitHub staff reply in the thread. A July comment says it still happens on macOS Tahoe.

Workarounds from the thread, and talk to whoever runs your Mac fleet before using the second one:

• System Settings → Privacy & Security → Allow Anyway, then Open Anyway on the next run. Many companies don't allow this override.
• Remove the quarantine flag after each upgrade. The issue's author scripted it.
• On October 12 the author reported the real cause in their case: a corporate proxy presenting its own certificate broke Apple's check. Their fix was an SSL bypass rule for api.apple-cloudkit.com.

xattr -p com.apple.quarantine "$(which copilot)"
xattr -d com.apple.quarantine "$(which copilot)"

The first line shows whether the flag is there. The second removes it, which skips Gatekeeper's check for that binary. If your org forbids both, the honest workaround is a different agent for that task until GitHub ships a fix.

Where Hermes fits, and where it doesn't

I build Hermes IDE, so weigh this with that in mind. It fixes none of the seven. It doesn't sign Copilot's binary, touch the Codex extension or write to Claude Projects.

It helps with the pattern under items 2, 4 and 7: if one agent is stuck, move the task to another. Hermes runs Claude Code, Codex, Gemini CLI, Copilot CLI, OpenCode and others each in a real terminal, running the vendor's own CLI with your own sign-in (README). You pick the agent per task. When one hits its usage limit, you can continue the task in another, and every task gets its own branch and worktree, so two agents aren't editing the same checkout.

The ⌘N agent picker in Hermes 2.3.4: Claude Code and Codex installed, the other supported agents marked Not installed, plus Custom agent.

It's free, source-available under BSL 1.1, and runs on macOS, Windows and Linux. Download it or read the code on GitHub. It is not affiliated with Nous Research (or its Hermes Agent), Anthropic, OpenAI, Google, GitHub or the OpenCode team.

Which open issue in your agent's tracker costs you the most time, and what's your workaround? Leave the issue link in the comments.

---

If this was useful

Issue #50403 is a queue bug: a lock that throws on release, messages that vanish, and a few that run twice. Those are the same failures every message queue has to plan for, and the Event-Driven Architecture Pocket Guide covers them: at-least-once delivery, idempotent consumers, outbox and saga. It's one of 8 short books in Pocket Guides for Developers, on agents, RAG, prompts, observability, events, databases and system design.