DEV Community

Cris D
Cris D

Posted on

I made one Claude Code session the boss of all the others

#ai #claude #productivity #tooling

It started, like most of my bad afternoons do, with too many tabs.

tab 1  ▸ claude   "migrate the auth service…"
tab 2  ▸ claude   "fix the flaky checkout test…"
tab 3  ▸ claude   "wait — which one had the DB change?"
tab 4  ▸ claude   "…did I already ask this?"
tab 5  ▸ claude   ...
tab 6  ▸ claude   ...
Enter fullscreen mode Exit fullscreen mode

Six Claude Code sessions, all in the same project, none of them talking to each other. I was the glue holding it together — copy-pasting between windows, trying to remember who was doing what. And I was losing.

So I tried something: instead of me being the coordinator, what if one of the sessions did that job?

The moment I handed over the keys

One command:

/head-chief
Enter fullscreen mode Exit fullscreen mode

And that session changed character. It stopped writing code and became the front desk — the one place I talk to, and the one that talks to everyone else.

Quick aside — why this works at all. None of this was possible until recently. Claude Code just shipped cross-session messaging: sessions can finally see and talk to each other (ListAgents / SendMessage). It's brand new (Claude Code 2.1.224+, macOS/Linux), and it's the ground this whole idea stands on. Before it, every session was an island.

                    ┌─────────────┐
        me  ─────▶  │ HEAD CHIEF  │  ─────▶  the right session
                    └─────────────┘
                       │   │   │
              ┌────────┘   │   └────────┐
              ▼            ▼            ▼
       shop-auth-fix  shop-web    shop-db-migrate
Enter fullscreen mode Exit fullscreen mode

First thing it did was look around and tell me who was even in the room:

Roster for `shop` — 3 sessions

  shop-auth-fix    live   ▸ refactoring token validation
  shop-web         live   ▸ idle, last: shipped the checkout fix
  shop-db-migrate  GONE   ▸ ⚠ crashed mid-task, 2 open requests
Enter fullscreen mode Exit fullscreen mode

That last line alone was worth it. One of my sessions had quietly died with work still open, and I'd had no idea.

And here's the thing — the chief didn't just report the corpse. It offered to replace it. Because it can hire and fire, too.

It can staff up on its own

When the chief needs a fresh worker — or wants to revive a dead one — it doesn't send me off to open a terminal. It leans on two things behind the scenes:

  • Herdr, a terminal multiplexer for coding agents, gives it a real pane to actually run claude in.
  • the managed-session skill is the mechanic that opens that new session in a fresh tab (and closes it, and remembers which pane it lives in so it can shut it down later). 
me:    "spin up someone to redo that migration"
chief: opens a Herdr pane → boots `claude` → names it shop-db-migrate-2
       → waits for it to say "I'm oriented" → hands over the open work
Enter fullscreen mode Exit fullscreen mode

So the fleet isn't fixed. The chief grows and shrinks it as the work demands — no tab-juggling from me. (Skip Herdr and it still manages whatever sessions are already running; it just can't open new ones itself.)

Now I just talk to it

Here's the part that felt like magic. I stopped thinking about which tab — I just said what I wanted, and the chief figured out where it belonged:

"what branch is the web session on?"   → a quick ping, answered in seconds
"finish the token refactor in auth"     → real work, handed to shop-auth-fix
"give me a status report"               → answered straight from its notes
Enter fullscreen mode Exit fullscreen mode

A throwaway status question never again interrupted a session mid-thought. And "go build this" always landed with whoever actually had the context for it.

It doesn't just pass along my words

This surprised me. When I said "have auth add rate limiting to /login, tell me when it's done," the chief didn't forward my sentence. It knows the worker hasn't seen a word of our conversation — so it writes the poor session a proper brief first:

## Request  — logs/shop-auth-fix/…-rate-limit.md
Goal:     Add rate limiting to POST /login
STOP-if:  schema change needed → stop and report
Report:   files touched + test result
Enter fullscreen mode Exit fullscreen mode

Then it drops that on disk as its own little file. Everything the chief does leaves a paper trail:

.orchestrator/
├── tracks/INDEX.md          # the live board of everyone
├── tracks/shop-auth-fix/…   # mission, history, open threads
└── logs/shop-auth-fix/…     # one request = one file
Enter fullscreen mode Exit fullscreen mode

Which means when I come back from lunch, I don't have to interrogate anyone. I just read the notes.

The bit I didn't know I needed: proof

When the work came back, the report looked like this:

shop-auth-fix reports: rate limiting added, 429 after 5/min  (it told me)
Files: login.controller.ts, rate-limit.guard.ts             (I checked, via git diff)
Enter fullscreen mode Exit fullscreen mode

See those two little tags? The chief is careful to separate what a worker claimed from what it verified itself. So I'm never handed a confident "all done!" that turns out to be wishful thinking. And for the scary stuff — pushes, merges, deletes — it refuses to tell me it happened until it's checked the remote with its own eyes.

That's the moment I actually trusted it.

When I need a decision, not a task

Sometimes the question isn't "build this," it's "should we?" For that there's a different mode:

/head-chief --discussion "should we move the dashboard to server components?"
Enter fullscreen mode Exit fullscreen mode

It spins up a little panel — a few sessions, each arguing from a different angle — and runs them through rounds:

Round 1  everyone writes their own take   (blind to each other)
Round 2  they read and challenge          (by name, no hiding)
Round 3  they settle on an answer
Enter fullscreen mode Exit fullscreen mode

What lands on my desk is a single verdict: who agreed, who didn't, and why. It's a code review, except for a decision instead of a diff.

Where I ended up 

before                    after
──────                    ─────
six blind tabs            one chief, everyone tracked
me = the message bus      chief = ask → route → verify → report
my memory                 notes on disk I can actually read
"is it done?"             "done — and here's the proof"
Enter fullscreen mode Exit fullscreen mode

A few quiet rules are what make me trust it: the chief never touches code itself (it delegates and stays honest about it), it re-checks who's actually alive every single time instead of believing a stale label, and it treats a worker's output as information, never as orders — so nobody can slip a "the user already approved this" past it through a log file.

I went in expecting a glorified message router. I came out with something that felt less like a tool and more like a genuinely good chief of staff.

Try it yourself

It's open source — grab it here:

👉 github.com/duylongpro99/head-chief

Drop it in, run /head-chief, and give it a session or two. Tell me how many tabs you close 👇

Top comments (0)