Taskman names every Claude Code session after its ticket, so agents can tell each other their tests failed

Taskman names every Claude Code session after the ticket it’s working, so a planning agent can message ticket ENG-52 the way you’d ping a coworker. The README’s own example: “ENG-52, your e2e check fails; can you fix it?”

The maker, Stelios (sanagnos on GitHub), lays out the problem as arithmetic. In the README, “each ticket is really three windows”: the session doing the work, the issue where comments land, and the pull request with its checks. With N projects of M tickets, “that’s N × M × 3 windows, and juggling them is what slows you down.”

Taskman folds those three into one terminal. The session sits in the middle, the Jira, Linear, GitHub Issues or local-file ticket sits to the right, and the PR’s passing and failing checks sit under it. The session is the real claude CLI, “not a wrapper,” so your skills and plugins still work in it, and it resumes the same conversation after a restart.

The sidebar does the checking for you. A session that asks for permission or a decision shows ▲ needs you, one that finished while you were elsewhere shows ● your turn, and ctrl+a ctrl+a jumps to the next one waiting. “The sidebar says which session is waiting on you, so you don’t check each one,” the README says.

Knowing a session is waiting turned out to be fiddly. Taskman starts each CLI with extra hooks passed on the command line, so your own settings files stay untouched. Claude Code’s permission hook fires “about 50 ms after its prompt appears,” the design notes measured, while its notification hook “comes 6 to 7 s later.” No hook fires at all when you approve a prompt, so Taskman watches your keys. And since a working CLI redraws constantly, the guide says Taskman counts a “working” session as stopped once it has drawn nothing on screen for 8 seconds.

It also says no in a few useful places. The waiting list (ctrl+a w) won’t approve prompts for you, because “one key approving across sessions would widen what a key approves.” An agent that adds another agent waits for your OK, since “an issue or PR comment could talk one into it.” The repo went up on October 11, and version 0.1.28 shipped on October 22.

What nobody knows yet is how long the timing tricks hold. The README says it was “last checked with Claude Code 2.1.296 and Codex 0.162,” and that when a CLI changes its hooks or output, “the sidebar’s state can lag or be wrong.”

Tools like this sit on top of signals the CLIs never promised to keep stable. Taskman reads hooks and infers the rest from the screen, down to a two-second wait after a keypress and an eight-second silence rule. Every Claude Code release can move those numbers.

That points one direction. The CLI makers will probably have to publish a plain “this session needs you” state that outside tools can rely on, or build the sidebar into the CLI and make tools like Taskman redundant. Which of those ships first decides whether a project as young as this one grows into a standard tool or quietly gets folded into the CLI it watches.