Migrating to 2.0

gmux 2.0 is a breaking release. It starts with a clean SQLite state store; it does not import 1.x projects, peers, or dead-session metadata. The CLI tells you the new form of removed commands. This page lists every breaking change and the required migration steps.

The short version:

1. Upgrade every machine (and rebuild devcontainers) together — 2.0 hosts can’t peer with 1.x hosts.
2. Update scripts and muscle memory to the verb-first CLI: gmux -- <cmd> to run, gmux open for the UI, gmux ls/attach/send/wait/kill instead of flags.
3. Re-add each remote host with its connect URL (Settings → Hosts → Connect to host, using gmux auth on that host), then add the projects you want under Settings → Projects → From other hosts.
4. Recreate local projects. Restart any sessions that were running under 1.x — they, along with dead history, project order, references, and connected hosts, do not carry over.
5. If you parse gmux JSON: kind → adapter, session_file → conversation_file.

---

CLI: verb-first grammar

Who: everyone — interactive users, scripts, aliases, agent skills. (ADR 0009)

Bare-command shorthand removed

Before	After
gmux pi	gmux -- pi
gmux pytest --watch	gmux -- pytest --watch

A bare word now errors with the run-form hint. If you run commands constantly, alias gm='gmux --' is shorter than the old shorthand.

Bare gmux no longer opens the dashboard

Before	After
gmux (no args) → opens the UI	gmux prints help; gmux open opens the UI

Daemon auto-start and the update notice moved to gmux open (and session launches).

Action flags replaced by verbs

Every removed flag prints an error naming its replacement — nothing silently changes behavior:

Before	After
gmux --list / -l	gmux ls
gmux --all	gmux ls --all
gmux --attach <id> / -a	gmux attach <id>
gmux --tail <id> / -t	gmux tail <id>
gmux --kill <id> / -k	gmux kill <id>
gmux --send <id> <text>	gmux send <id> <text> Enter
gmux --send --no-submit …	gmux send <id> <text> (omit the trailing Enter)
gmux --wait <id>	gmux wait <id>
gmux --no-attach <cmd>	gmux -d -- <cmd>
gmux --host <peer> …	address the session as <id>@<peer>

Note the inverted send semantics: 1.x auto-appended a newline unless --no-submit; 2.0 never auto-submits — add a trailing Enter key token to dispatch. Audit any script that pipes prompts.

Daemon lifecycle fronted by gmux daemon

gmux daemon start|stop|restart|status|log-path is the canonical interface. The gmuxd binary keeps its verbs for service managers, so nothing breaks operationally — but update docs and scripts to the gmux spellings. gmux auth (token/pairing) and gmux remote (Tailscale setup) are top-level verbs.

New verbs (not breaking, but update your habits)

• gmux edit [file] — managed editor sessions, usable as $EDITOR. Inside gmux sessions, EDITOR/VISUAL now default to gmux edit when your dotfiles don’t set them — scripts that branch on EDITOR being empty inside sessions will see a value.
• gmux send-keys -t <id> … — tmux-compatible key sending.
• gmux wait [--quiet] [--timeout N] with the global exit codes 0 (the turn completed / the output matched) / 2 (the turn was intentionally interrupted) / 1 (anything else — error, death, timeout). A failed one-shot or mark-less shell command closes its turn with an error, so waiting on it now exits 1 rather than 0 — deliberate, so gmux wait $id && next-step cannot run after a failed build. For a pi session, wait also prints an exchange-structured report of the activity it observed on stdout — for every outcome, success or not; --quiet suppresses it. gmux wait --for-text S / --for-regex P wait until output appears instead of the idle signal, work for shell sessions too, and print no result.
• gmux agent prompt|logs|cancel — semantic control for agent sessions (pi), sharing the same exit codes: prompt delivers work (with --new, --follow-up, --steer) and prints an exchange report, logs renders the stored conversation as exchanges (-n counts exchanges, default 1), cancel interrupts the work in progress.
• gmux send --wait [--timeout N] fuses send-and-wait race-free (subscribes before delivering the input). Note send’s grammar: flags go before the id; everything after the id is verbatim, so gmux send abc -v sends a literal -v with no -- guard needed.

---

Multi-machine: tokens everywhere, no autodiscovery

Who: anyone with peered hosts, tailnet setups, or devcontainers. (ADR 0008)

All hosts must run 2.0

The wire protocol is v2-only: a 2.0 hub cannot aggregate a 1.x spoke and vice versa. Upgrade every machine together, and rebuild devcontainers so the feature installs a matching gmux.

Tailnet identity no longer grants access

Before, passing the Tailscale allow list granted the full API. Now tailnet identity only gets you to the login page; every request additionally needs the host’s bearer token — the same two-gate model as the browser. If you opened https://gmux-<host>.ts.net and got straight in, you’ll now see the login page once: paste the token from gmux auth on that host.

Tailscale peer autodiscovery removed

Before, gmux machines on your tailnet appeared as hosts automatically. Now peers are explicit: run gmux auth on the host, paste its connect URL into Settings → Hosts → Connect to host.

There is no host/reference importer. On first 2.0 start, the daemon creates a clean SQLite database (state.db) and ignores the old discovery cache and JSON state. Re-add every wanted host with Connect to host, then add its projects under Settings → Projects → From other hosts. Recreate local projects as well.

Removed host.toml keys

These are ignored with a warning (not fatal), so an old config won’t brick the daemon. Remove them to silence the warning:

Key	Replacement
tailscale.hostname	Name derives from the OS hostname (gmux-<hostname>) and is then owned by Tailscale. Seed a different name before first registration with GMUXD_TS_HOSTNAME, or rename in the Tailscale admin console.
[[peers]]	Runtime state in the daemon’s SQLite database (state.db), managed via Settings → Hosts.
discovery.tailscale	Gone — add tailnet hosts via Connect to host.

Host renames no longer follow automatically

A peer’s name is now frozen at first contact (ADR 0017): renaming a machine doesn’t relabel your roster, and references keep working under the original label. Node IDs act as a liveness anchor — a removed-and-re-added host reclaims its references automatically.

---

API & schema: terminology rename

Who: anyone parsing gmux ls --json, the REST API, or SSE payloads.

Before	After
"kind": "pi" (session JSON)	"adapter": "pi"
KIND column in gmux ls	ADAPTER
GET /v1/conversations/{kind}/{slug}	GET /v1/conversations/{adapter}/{slug}
"session_file"	"conversation_file"
resume_key field	gone — use conversation_file (resume identity) and slug (membership/URLs)
stale field	gone — derive from runner_version/binary_hash vs GET /v1/health

The daemon accepts the legacy runner session_file event for one release (dropped in v2.1) but writes/emits only the new names. The GMUX_ADAPTER env var is unchanged (it was already named that in 1.6). URL path segments (/project/pi/slug) are unchanged, so bookmarks keep working — though a Claude /rename now moves the slug with the title.

Wire protocol v2

Custom consumers of the daemon SSE stream: the per-event session-upsert/session-remove surface and bulk-GET prefetch are gone. Subscribe to GET /v1/events and consume full-replacement snapshot.sessions / snapshot.world payloads plus lossy session-activity pings. GET /v1/sessions remains for one-shot listing.

Same-origin enforcement

Cookie-authenticated mutations and WebSocket upgrades are now rejected cross-origin (403 cross_origin). Browser-based tooling on another origin must switch to bearer-token auth; reverse proxies that rewrite Host must forward the browser-facing host in X-Forwarded-Host. See Security.

---

Adapter API (out-of-tree adapters & integrations)

Who: authors of custom adapters or tooling against packages/adapter / the runner socket.

• Renames: SessionFiler → ConversationFiler, ParseSessionFile → ParseConversationFile, SessionFileInfo → ConversationInfo, SessionRootDir/SessionDir → ConversationRootDir/ConversationDir; internal Kind → Adapter.
• Removed capabilities: FileMonitor, FileAttributor, SessionFileLister. Daemon-side file attribution and live tailing were replaced by runner-owned agent hooks (SessionExtender / SessionHookCommand + POST /hook/event) and adapter-owned ConversationSources. There is no metadata-matching fallback: an unhookable tool runs without daemon-reported live state.
• Status.label removed: Status is only {active, error, interrupted} booleans. Scripts that PUT /status with a label should drop it — display text is derived in the frontend.
• Runner endpoints removed: GET /scrollback/text and GET /scrollback/tail are gone from the runner socket. Use gmux tail <id> or gmuxd’s GET /v1/sessions/<id>/scrollback?tail=N (works for dead sessions too).
• New surface: GMUX_SESSION_SOCK env var, POST /hook/event hook protocol (docs/runner-hook-protocol.md), GMUX_NO_AGENT_HOOK opt-out, ConversationProber, PassthroughDetector, SessionRegistrar/SessionFinalizer.

---

Behavior changes

Agent status is hook-driven

pi, Claude Code, and Codex now report status/titles/attribution through injected hooks instead of file watching:

• Codex needs CLI ≥ 0.135.0 for live status; older versions launch fine but show no working/idle state.
• Shell-wrapped launches (gmux -- bash -c 'claude') can’t be hooked and run without live status.
• The agent’s argv gains a hook argument (-e <ext> for pi, --settings for claude, -c hooks.… for codex). GMUX_NO_AGENT_HOOK=1 launches the agent unmodified.

Sessions and retention

• Dead sessions persist across daemon restarts as rows in SQLite. Dismissal hides the selected session and its launch descendants and removes their project placement; retained rows and conversation identity are not hard-deleted.
• gmux no longer surfaces conversations it never saw by scanning ~/.claude/~/.codex/~/.pi into resumable sidebar entries; those files still power lookup and reconciliation of known sessions.

Per-session sockets moved

1.6 put runner sockets in shared /tmp/gmux-sessions; 2.0 uses ~/.local/state/gmux/run/sessions (per-user, 0700). Tooling should read $GMUX_SOCKET instead of constructing paths. Sessions that were running under 1.6 are not carried into the 2.0 daemon — restart them after upgrading. GMUX_SOCKET_DIR still overrides.

Fresh login environment

Daemon-initiated launches (UI launcher, resume, restart) source a fresh $SHELL -l -i environment per launch instead of inheriting the daemon’s frozen environment. Dotfile edits take effect on the next launch without a daemon restart. (No $SHELL — Docker/systemd — means unchanged behavior.)

Devcontainer discovery requires the devcontainer label

Containers are only auto-discovered when they carry the devcontainer.local_folder label (set by the devcontainer CLI / VS Code) in addition to GMUXD_LISTEN. Plain docker run containers with GMUXD_LISTEN set are no longer picked up — add them as manual peers, or use the Running in Docker flow.

Project model

Projects are now stored in the daemon’s SQLite database (state.db, ADR 0026). The hosts match-rule field is dropped. Configure projects on the host that owns their sessions, then add wanted network-host projects from Settings → Projects → From other hosts.

---

UI changes (where did X go?)

Not breaking, but 1.x docs and muscle memory point at moved things:

• Home screen is now a pure output-recency dashboard (Today, Yesterday, recent weekdays, then dates). Status changes a session’s indicator, not its section. Host cards and quick-launch buttons are gone.
• Project management moved from the sidebar’s “Manage projects” modal to Settings → Projects (sliders button in the sidebar header).
• Hosts roster lives in Settings → Hosts, with explicit Online / Connecting… / Auth needed / Offline statuses.
• Mobile toolbar reworked: dedicated ↑ ↓ and word-jump keys are always present; ctrl/alt arm-and-highlight instead of relabeling keys; paste moved off the toolbar (paste keybind or long-press).
• Cmd/Ctrl+F now opens find-in-terminal instead of browser find. Restore browser find with { "key": "secondary+f", "action": "none" } in settings.jsonc.

Nothing in settings.jsonc or theme.jsonc changed — old files parse identically.