I kept losing track of which Claude Code session was waiting on me, so I built Paneherd.
Every agent in its own pane with live status, your 5h and weekly limits, a worktree per agent that says when it's ready to merge, and a phone remote.
Free for VS Code and Cursor.
@iamleannmuller /context is great, I just forget to run it. Claude Code also passes context_window.used_percentage to the status line after every reply, so I keep it next to the 5h %. Watching it climb does the reminding for me.
Claude's 5-hour limit hit 95% mid-task.
My VS Code extension asks once: hand this pane to another agent? Pick Codex and a new pane opens in the same folder with a brief: branch, uncommitted files, Claude's last message, and "run git diff first".
Free, no Pro needed.
@LucaCaponeX 23% over four days is about 6% a day. What helps me most is seeing the cost per task: Claude Code passes your weekly % to the status line after every reply, so put it there and you notice right away when one run eats 4%.
@geezerrrr_r Review before merge is the part I'd keep too. What still bit me with one worktree per agent: two of them change the same file and nobody notices until the merge. Does cc-huddle's review see the other workers' diffs, or only its own branch?
@BabaDev_ The 6 AM message stings less when you see it coming. Claude Code passes your 5h and weekly % to the status line after every reply, so you can keep it in your prompt or tmux bar and stop starting big runs at 90%.
@marcthecreatorr On a subscription the $ figure matters less than what each run did to the 5h window. Did you catch the % before and after? Claude Code puts it in the status line JSON after every reply and Codex shows it in /status, so it's easy to log next time.
@techNmak This gets more valuable with a few agents running in parallel. Two plans that both touch the same file are cheap to spot at plan time. I only catch that at edit time right now, and by then one of the agents has to redo work.
@alexmichael The hooks-that-never-match line is the one I'll use. My VS Code extension shows each session's status from Claude Code hook events, and a hook that silently stops matching is exactly how a pane ends up showing the wrong state.
@ClaudeDevs One catch with several agents merging: a green run only counts for the exact commit it ran on. My merge queue fast-forwards only that commit, so one agent's merge can't quietly invalidate another's passing check.
@IntuitMachine Past 4 or so agents, my cap stopped being verification and became attention: which session is stuck on a permission prompt right now. I ended up reading Claude Code's hook events to show that per pane instead of watching terminals.
@yoavgo One thing I hit running 4 agents on my own VS Code extension: an agent can edit the test it's judged by. So my merge queue flags any branch that touched its own tests and leaves it for me to pick by hand.
I kept losing track of which Claude Code session was waiting on me, so I built Paneherd.
Every agent in its own pane with live status, your 5h and weekly limits, a worktree per agent that says when it's ready to merge, and a phone remote.
Free for VS Code and Cursor.
I run 3-4 Claude Code and Codex sessions at once. Two things kept biting me: missing which one waited on me, and two editing the same file.
I built Paneherd, a grid of agent terminals in VS Code. It's on Product Hunt today.
What does your multi-agent setup still not tell you?
@thehenrylabs Matches what I hit running 4 agents. In my VS Code extension the merge queue fast-forwards only the exact commit whose check passed, and a branch that touched its own tests gets flagged and starts unpicked. Haven't solved the judge-on-base-branch part yet.
@myalya Held that job for a while. What fixed most of it: one worktree per agent, then land them one at a time: rebase onto main, run the tests, fast-forward only if they pass, send conflicts back to the agent that made them. I ended up building that queue into my VS Code extension.
@sattyyouneed I switch to Codex for the small stuff. In the VS Code extension I'm building, once Claude's 5-hour window is 90% used, new light worker tasks run on Codex instead, and the bigger prompts sit in a queue that starts them when the window resets.
@NabhaniMehdi Same, that one's why I built this. Claude Code's hooks make it easy: Notification fires on a permission prompt, Stop when it's your turn again. Paneherd listens to both, marks the waiting pane and pops a notification. That part is free.
Running 3+ Claude Code or Codex sessions at once, what bites you most?
1. Not noticing which one is waiting on you
2. Two sessions editing the same file
3. Hitting the 5-hour limit mid-refactor
4. Reviewing and merging all those branches after