cc-session-control

tmux-first workbench (TUI + headless CLI) for the agent CLIs on one machine — Claude Code, Codex CLI, Kimi Code, and opencode: one session list across all four, plus a tmux-first project launcher.

CLI command: csctl

Features

• Sessions Tab — One machine-wide list of Claude Code, Codex, Kimi Code, and opencode sessions (CLI column: cc/cx/km/oc), discovered from each CLI's own on-disk state — not just sessions csctl started. Resume tmux-first (Enter resumes into a project-labelled window in the shared csctl tmux session via each CLI's native resume command; t bare-terminal fallback; R backgrounds into tmux; f forks where the CLI supports it; ⧉ marks tmux-resident sessions; Codex app-server-held rows show as hosted/read-only instead of dead), terminate, and delete; a cleanup submenu (c) prunes empty/short Claude sessions and sweeps orphan artifact directories, zombie session files, and aged global entries (cleanup models Claude state only)
• Projects Tab — The startup tab / launcher: member directories are discovered from evidence tiers (operator pins, any CLI's trust records, any CLI's session activity). Enter opens a CLI chooser (active providers only, claude focused first — so Enter-Enter starts claude) to open a new project-labelled window in the shared csctl tmux session; x/k/O jump straight to codex/kimi/opencode; p/h/H pin, hide, and reveal projects

Non-Claude liveness is deliberately conservative (ADR-0005). A real resume target in argv (codex resume <sid-or-unique-name> / kimi --session <sid> / opencode --session <sid>) has priority; csctl's own identity-checked tmux dispatch metadata is a supplement for dispatched windows, and is essential after Kimi rewrites its argv at runtime. A brand-new session started from the launcher (the Enter chooser or x/k/O) has no sid yet, so it stays unbound, as does any other bare-launched TUI; unbound processes are never stop/takeover targets. Codex app-server fd evidence is a separate hosted/read-only state: it never supplies a session pid, and csctl emits no resume/stop/fork/delete command for that row.

Kimi can close that gap opt-in via its official hooks: add this to ~/.kimi-code/config.toml and every kimi session — bare-launched ones included — self-reports its pid↔session binding (verified on Kimi Code 0.34.0, re-verified on 0.35.0; csctl re-verifies pid identity and process start time per entry, so stale or forged entries never bind):

[[hooks]]
event = "SessionStart"
command = "csctl _kimi-hook"
timeout = 5

[[hooks]]
event = "SessionEnd"
command = "csctl _kimi-hook"
timeout = 5

# Self-heal: re-registers the live session every 60 s, so a SessionStart
# kimi never delivered becomes a ≤60 s unbound window instead of a lifetime.
[[hooks]]
event = "SessionHeartbeat"
command = "csctl _kimi-hook"
timeout = 5

The hook fires when a session materializes, and the binding appears on csctl's next refresh.

A resumed session registers immediately; only a new one waits for its first prompt, because that is when kimi creates its id — which is also why csctl cannot learn a new session's id at spawn. SessionStart delivery is fire-and-forget and not guaranteed (a live 0.36.1 session went unbound for hours on 2026-08-16 with zero trace), so the SessionHeartbeat rule re-registers the session every 60 seconds and a missed start self-heals. A session already running when the rule is added keeps its old config — reopen it to bind. ~/.kimi-code/run/hook-errors.log says whether the hook ran at all — the endpoint records every run that did not register, including an unrecognized event name.

All agent sessions csctl dispatches — new, resumed, forked, or backgrounded — share the tmux session named csctl; their window names retain the project. Existing live sessions in older or user-created tmux sessions are entered in place and are never migrated automatically.

For phone terminals, the managed csctl tmux session gets a scoped second prefix when none is already configured: press Ctrl-A, then s, to open tmux's session tree from inside Claude, Codex, or Kimi. An existing prefix2 is preserved; csctl never changes the primary prefix or global tmux config.

Built with urwid.

UI language: Simplified Chinese (notifications and status text). CLI output is in English.

Requirements

• Python 3.12+
• At least one supported agent CLI installed and authenticated: Claude Code, Codex CLI, Kimi Code, and/or opencode — each is discovered automatically when its state home exists (~/.claude, ~/.codex, ~/.kimi-code, ~/.local/share/opencode; official relocation variables CODEX_HOME / KIMI_CODE_HOME are honored, and opencode's data home follows XDG_DATA_HOME). Several codex identities on one machine are supported by declaring their homes — see Configuration below
• tmux (the primary session-lifecycle carrier: launch, resume, background, and survive terminal/SSH disconnects)
• Linux / WSL. Other POSIX systems are not supported; without /proc, csctl shows degraded state and refuses destructive actions whose safety it cannot prove.

Installation

Install the latest published release:

uv tool install cc-session-control
# or
pipx install cc-session-control

Upgrade later with uv tool upgrade cc-session-control (or pipx upgrade cc-session-control).

Latest master build

To try the newest master before it is released, install from GitHub:

uv tool install --reinstall git+https://github.com/dzshzx/cc-session-control.git

csctl manages local state on the machine where it is installed: the active providers' ~/.claude, ~/.codex, ~/.kimi-code, and ~/.local/share/opencode homes, local tmux, and the project launcher entries recorded in ~/.claude.json. Install it separately on each machine whose sessions you want to manage. For working on the code instead of using it, see CONTRIBUTING.md.

Usage

The TUI is the primary surface — the project launcher and session cleanup live there (Projects tab and the Sessions cleanup submenu). The headless CLI keeps only the agent-facing command: resume.

# Open TUI
csctl

# Resume rescue (headless): list sessions of ALL providers across
# directories with ready-to-copy resume commands (native /resume only
# searches the cwd and hides sdk-ts/bridge sessions); non-Claude rows
# are tagged [codex]/[kimi]/[opencode]
csctl resume                 # Page 1, 20 per page
csctl resume mybug           # Keyword: sid/cwd/title, then transcript body
csctl resume --page 2        # Next page
csctl resume --all           # Everything, no paging

# Options
csctl --theme light            # Force the TUI palette (auto/dark/light)
csctl --version

Session rescue is built into csctl itself: use csctl resume or the TUI.

Configuration

Multiple codex identities

CODEX_HOME says which home one codex process uses, and it is inherited by everything that process spawns — including csctl. Reading it as the machine's inventory made csctl show only the identity it happened to be launched from. So when you run more than one codex identity, declare their homes in ~/.config/csctl/providers.json (XDG-respecting) instead:

{
  "codex_homes": [
    {"label": "cx",  "home": "~/.codex"},
    {"label": "cx2", "home": "~/.codex-eva02"}
  ]
}

That list is then the complete codex inventory and CODEX_HOME no longer takes part, so csctl shows the same machine state from any terminal. Each entry becomes its own provider — separate sessions, trust records, badges, and launcher entry — tagged in the CLI column by its label (ASCII alphanumeric, ≤3 characters). The first entry keeps the provider key codex; later ones become codex:<label>. Every command csctl synthesizes for a declared identity carries its CODEX_HOME explicitly. Without this file there is exactly one codex instance following CODEX_HOME/~/.codex, and a malformed file falls back to that same single instance with a visible reason. Changes take effect on the next csctl start. See ADR-0008.

License

MIT