fleetops

An ops board for every Claude Code session running on your machine.

Each Claude Code session is a ship; fleetops is the bridge's ops board — one terminal view of the whole fleet: which sessions are working, which are idle, and which are blocked waiting on you, plus tokens spent, context-window fill, and the wezterm pane each one lives in (jump to any of them with a keypress).

Features

• Session discovery — finds every live Claude Code session from ~/.claude/sessions/<pid>.json, confirming liveness against /proc (PID-reuse-safe: the file's recorded start time must match /proc/<pid>/stat). Stale files for dead PIDs are counted, never shown as live.
• At-a-glance status — a pure fold over each session's native status, pending-question flag, and transcript activity yields one of: working, idle, needs answer (a pending AskUserQuestion), waiting (blocked on input the transcript can't show, e.g. a permission prompt), stalled? (busy but the transcript stopped growing), shell, or unknown (a native status this build doesn't recognize — a drift signal, never hidden).
• Tokens & context % — reads the transcript tail for the last assistant usage line and renders a context-window gauge (out of 200k, or 1M once a session exceeds 200k) plus a compact token count. Approximate, never a bill.
• wezterm pane mapping & jump — matches each session to its wezterm pane (exact WEZTERM_PANE identity when forwarded, else title/cwd) and jumps to it on Enter (activate-tab then activate-pane). Discovers every live wezterm instance so a board running on one monitor can still drive panes in another.
• Codex lane — Codex CLI sessions (which keep no per-pid session file) are discovered from /proc + their rollout transcript and folded onto the same board.
• Read-only over the fleet — the only actions that change anything are focusing a pane (the jump) and an optional brief highlight of the jumped-to pane (disable with --no-highlight). fleetops never writes into any Claude config or session directory.
• doctor and snapshot subcommands — fleet doctor prints a read-only drift report (are the undocumented sources still parseable?); fleet snapshot emits one JSON object of exactly what the board would render, for dashboards and scripts.

Platform: WSL2 + macOS

fleetops targets WSL2 (Linux) and macOS; native Windows is not supported (the board comes up empty there and fleet doctor says why).

• WSL2 — session discovery reads /proc; pane control shells out to the Windows wezterm.exe across the WSL interop boundary. The Windows-side wezterm socket directory is auto-detected (from $WEZTERM_UNIX_SOCKET, or by scanning /mnt/c/Users/*/.local/share/wezterm) — no per-machine username configuration.
• macOS — session discovery uses libproc + one sysctl (no /proc needed); liveness compares the session file's procStart (a UTC ctime string there) against the kernel start time, exactly. Pane control drives the native wezterm CLI; sockets are discovered from $WEZTERM_UNIX_SOCKET plus ~/.local/share/wezterm, filtered to live wezterm-gui processes owned by the invoking user. Mux domains with a custom socket_path are not discovered (documented limitation). No wezterm installed → the board still works; only the jump/highlight lane sits empty.

Install

cargo build --release
# binary at target/release/fleet

Requires a recent stable Rust toolchain (see rust-toolchain.toml) and, for the pane lane, wezterm installed on the Windows side (wezterm.exe reachable from WSL).

Using Claude Code? Run claude in the repo and type /setup — a checked-in command (.claude/commands/setup.md) walks Claude through checking your platform, building, installing onto PATH, detecting which lanes your machine supports, and interpreting fleet doctor.

Quick start

fleet              # launch the board
fleet --no-highlight   # board, without the jump-target pane highlight
fleet doctor       # read-only diagnostics / drift report
fleet snapshot     # one-shot JSON of the current board, to stdout

Keys: j/k or ↑/↓ move the selection · Enter jumps to the selected session's wezterm pane · h pins the selected row to the top, or unpins it if it's already pinned (u also unpins) · J/K move a pinned row down/up · r refreshes · q/Esc quits.

Pinned rows are marked ▪ and stay at the top across refreshes, in the order you arranged them with J/K. There is no mode: j/k always move the cursor, so pinning one session never stops you navigating to another. The manual order survives a degraded sweep too — a pin whose session is missing goes dormant rather than being dropped, and wakes up if the session comes back. The # column keeps each agent's number regardless of where you park its row, so it still matches fleet snapshot.

What it reads & privacy

Everything fleetops reads is local, and nothing is ever transmitted off the machine. It reads no credentials — not tokens, not API keys. Specifically, per live session it reads:

• ~/.claude/sessions/<pid>.json — pid, session id, cwd, native status, name, version.
• /proc/<pid>/stat — process start time, for the liveness / PID-reuse check.
• /proc/<pid>/environ — only two variables are kept: CLAUDE_ACCOUNT (account label) and WEZTERM_PANE (exact pane identity). Everything else in the environment is ignored.
• /proc/<pid>/fd/1 — the session's controlling pty, kept only when it resolves under /dev/pts/ (the target for the optional pane highlight).
• Transcript tail (~/.claude/projects/<slug>/<uuid>.jsonl, last 256 KiB) — only token counts, the ai-title, and a pending-question flag are extracted. Message text is never read into state, logged, or stored.
• wezterm pane list — wezterm.exe cli list --format json, for pane titles/cwd and jump targets.

No data leaves your machine; fleetops makes no network requests.

Unofficial

fleetops is an independent, unofficial tool. It is not affiliated with, endorsed by, or supported by Anthropic. "Claude" is a trademark of Anthropic. It relies on undocumented, internal file formats that can change at any time — fleet doctor exists to surface exactly that kind of drift.

Maintenance

Passively maintained. Issues and PRs are welcome, but responses may be slow and features are added only as they earn their keep.

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual-licensed as above, without any additional terms or conditions.