Flight Deck: a worktree-per-agent control panel for Claude Code

Customer

Solo software engineers and small-team devs who already use Claude Code or Codex daily and want to run 3-6 agent sessions at once instead of babysitting one chat window

Problem

Once a developer tries to go past a single agent session, nothing manages the mechanics: each agent needs its own isolated checkout to avoid clobbering files, each blocks silently waiting for approval or a clarifying question, and there is no single place to see which agent is working, stuck, or finished. So people improvise with a couple of terminal tabs, lose track of who is waiting on them, and fall back to one ticket at a time \u2014 the parallel throughput never materialises.

Proposed product

A local terminal UI (single binary, `flightdeck`) that turns each task into an isolated git worktree with its own agent session and shows all of them in one status board. You paste in a task (or point it at an issue), it creates a branch + worktree, launches the agent there, and streams a one-line status per agent: RUNNING / NEEDS YOU / DONE / FAILED. Agents that stop for a question or a permission prompt are surfaced at the top of the board with a desktop notification; you hit a key to jump into that session, answer, and drop back to the board. When an agent finishes, Flight Deck shows the diff and offers a one-key push/PR, then tears down the worktree.

Why now

An open issue on the official Claude Code repo says that when you run several sessions at once there is no way to tell which one is blocked waiting on you — you must check each tab or window by hand — and asks for native OS notifications. A second feature request shows why that gap persists: the existing idle_prompt hook fires after every response, so it cannot distinguish 'I'm done' from 'I need you', and a related issue (#36850) asks for a bell specifically on tool approval. A third-party orchestrator, cmux, has an issue saying its 'Claude waiting' notice is too generic to tell which workspace needs attention, and asks for reason classification. So attention routing across parallel agent sessions is a real, repeatedly reported and still-unsolved need. But the surrounding ground has shifted: several write-ups report Claude Code now ships its own `claude --worktree` flag, and tools such as Claude Squad (reported ~6.8k stars), Conductor, Vibe Kanban, Crystal and Nimbalyst already do worktree-per-agent with diff review and merge. The opening, if there is one, is narrower than the original pitch: accurate per-session state classification and routing, not worktree creation.

Smallest useful version

A CLI + TUI for Claude Code only, on one repo, that does four things: (1) `flightdeck run \"<task>\"` creates a worktree/branch and starts an agent there; (2) a board listing every live session with detected state by parsing agent output for idle/prompt patterns; (3) notification + hotkey attach when a session is waiting on the human; (4) `d` to view diff, `p` to open a PR, `x` to clean up the worktree. No inter-agent delegation, no cloud, no multi-repo, no scheduling \u2014 just isolation plus visibility plus fast context-switching.

Potential ways to charge

• Free open-source core with a paid licence for teams (per-seat, roughly the price of one lunch per developer per month) covering multi-repo and shared boards — noting no purchase signal was found for any tool in this space
• One-off paid desktop app for the macOS/Windows notification layer, priced like a small indie utility, with the CLI free
• Sponsorship or paid support tier if the project gains stars first, following the open-source-tooling pattern the competing tools appear to use
• Consulting or bespoke setup for small teams wanting agent parallelism configured, with the tool as a lead-in

Routes to early customers

• Comment on and subscribe to the exact Claude Code issues requesting native notifications and a tool-approval bell — those subscribers are self-identified sufferers of this problem
• Approach maintainers and users of cmux, which has an open issue asking precisely for per-session reason classification, and offer the classifier as a companion
• Post in the same roundups and newsletters that currently list Claude Squad, Conductor, Vibe Kanban, Crystal and Nimbalyst, positioning against them on attention routing rather than worktrees
• Direct outreach to practitioners who have blogged about losing track between terminal tabs or building their own session monitors

Main risks

• Differentiation is actively weakened by the evidence: the described feature set overlaps almost completely with shipped tools (Claude Squad, Conductor, Vibe Kanban, Crystal, Nimbalyst), and reported built-in `claude --worktree` support absorbs the worktree half outright.
• No willingness-to-pay signal was found at all — no job posting, freelance listing, keyword data or paid review — and the nearest alternatives appear to be free and open source.
• The core technical value depends on classifying state from a noisy upstream signal; the evidence shows the idle_prompt hook fires after every response and that a competitor's notifications are too generic to be useful. If Anthropic fixes this upstream, the product's remaining surface disappears.
• The original riskiest assumption is untested by the evidence: nothing here shows a developer will actually keep 4+ sessions productive once visibility improves, rather than reverting to one task at a time because of review load.
• Parsing agent output for idle/prompt patterns is brittle across Claude Code versions and will break silently, which is exactly the failure mode that destroys trust in a notifier.

Questions to test first

• Build only the classifier: instrument a handful of real sessions and measure how often NEEDS YOU is correctly distinguished from DONE. If you cannot beat the noisy idle_prompt behaviour described in the issues, stop.
• Ask the subscribers on the Claude Code notification issues and the cmux classification issue to try the classifier and report false positives — cheapest available panel of confirmed sufferers.
• Install Claude Squad and Conductor and document precisely where their attention routing fails, to establish whether there is any gap left after the platform's worktree flag.
• Verify the reported `claude --worktree` flag against Anthropic's own documentation, since the evidence treats it as reported-by-source only.
• Run a diary test with five developers using the board for a week: log how many sessions they actually keep live and whether it rises, to test the parallel-throughput assumption directly.