# Support and remaining gaps Source: https://libtmux.org/en/erlang/latest/guides/compatibility/ > Support and remaining gaps: libtmux_erlang documentation. Support covers tmux **3.3a–3.8**, including intermediate releases, on Linux x86_64 and arm64. The [workflow](https://github.com/libtmux/libtmux-erlang/blob/d1b6308440cf9eb8f6647e7ae434bce60e849368/.github/workflows/ci.yml) tests the oldest supported and newest stable tmux releases on both architectures. OTP 29.1.2 is the only supported Erlang baseline; older-OTP compatibility branches are not maintained. Native records remain an experimental language dependency in OTP 29. Do not persist their internals or assume compatibility with another OTP line. | Platform | tmux releases in the required matrix | | --- | --- | | Ubuntu 24.04 x86_64 | 3.3a, 3.8 | | Ubuntu 24.04 arm64 | 3.3a, 3.8 | CI uses only the latest stable Erlang/OTP release. Testing the endpoints keeps the routine matrix small; it does not exclude releases between them from support. Earlier intermediate-version results retain their recorded scope. Inspect the individual jobs at the commit you will consume. A minimum-version check does not prove runtime support on every newer release. Tracking builds, macOS, native Windows and psmux are not advertised as qualified platforms. Linux must provide readable `/proc/self/task//children`, enabled by CONFIG_PROC_CHILDREN. The helper checks this interface before launching a child. An unavailable interface returns `unsupported_capability` with `capability => proc_children`, the OS `errno` and `outcome => not_started`. See the [helper protocol](https://github.com/libtmux/libtmux-erlang/blob/d1b6308440cf9eb8f6647e7ae434bce60e849368/apps/libtmux_erlang/c_src/README.md) for cleanup evidence and the limits of native process ownership. ## Retained rows and upgrades Native model version 2 adds a private captured target association to each observed row. Actions check the entity, stable ID, origin and snapshot ID; pane associations also retain the observed window and index, and window-link associations retain their session, index and expected physical window. Actions reject edited target fields before dispatch. Valid-shaped binding mismatches return `observation_mismatch`; detached and malformed rows retain their validation errors. [`tmux_server:bind/2`]() reads the requested target and returns a new actionable row. Ordinary record matching and detached constructors remain available. The private value retains identifiers, not the snapshot graph. It prevents accidental retargeting through public row updates; it is not an authorization boundary against trusted code running inside the VM. The internal construction functions remain implementation details. Discard version-1 rows and snapshots when upgrading to model version 2. Stop consumers that retain them, clear their process/ETS state and freshly observe the server after restart. This release supplies no `appup`/`relup` conversion. Old or unknown model versions fail row validation. New code must validate a retained value before accessing fields added by another model version. Every exported-field addition, rename, removal or default change requires a model compatibility review independently of the OTP version. Its changelog must state whether retained values remain usable, need an explicit validated conversion, or must be discarded. Do not assume hot loading changes fields in existing native values. [`tmux_model:to_map/1`]() extracts the documented domain fields from a valid row. Validate those fields with the destination model's constructor when migrating data. The map loses actionable provenance; bind it through a freshly constructed row before an action. It is not a versioned storage or cross-port wire format: absence atoms, binary bytes and field inventories still require an application-defined schema. ## Implemented surfaces | Surface | Current behavior | | --- | --- | | Model | Exported native session/window/window-link/pane records; runtime validation | | Capture | Immutable complete graph, coverage, ordering, conflict detection and read intervals | | Local query | Map predicates, Boolean callbacks, quantifiers, cardinality and bounded JSON | | Source query | Explicit plans/fetch, conservative candidates, full residuals and final limits | | Operations | Bounded one-shot submission, waits, cancellation, transfer and cleanup | | Control attachment | Explicit session, same-connection identity, no-output/ignore-size, skipped environment update and separate native cleanup | | Daemon ownership | Empty private startup, separate ownership token, caller lifetime, explicit stop and identity-checked namespace cleanup | | Domain commands | Initial create/read/rename/kill calls, explicit client/session selection, checked window memberships, pane split/break/join/move/swap/resize, pane/window respawn, window layouts, literal/key input and capture | | Options | Explicit local/global tables, local or inherited reads, indexed writes/unsets and generated source metadata; hooks remain outside the option API | | Hooks | Versioned builtin arrays, indexed command values, inheritance and explicit execution contexts | | Buffers | Named or automatic binary load, exact show, metadata lists, delete/rename and paste; native version differences are explicit | | Server utilities | Command values for literal source-file/stdin configuration and native [wait-for](https://libtmux.org/en/tmux/latest/manual/wait-for/) signals, waits and explicit advisory locks | | Pane pipe | Literal shell command, output/input/both directions, native toggle and explicit stop; tmux-owned process and delivery checks | The [examples](https://github.com/libtmux/libtmux-erlang/blob/d1b6308440cf9eb8f6647e7ae434bce60e849368/examples) exercise the current public API. Real integration checks use private sockets and verify owned process retirement. A captured row can disappear before an action; stable IDs are not permission to reinterpret a missing target positionally. ## Not yet shipped Control command submission and groups and output subscriptions/streams remain implementation work. Workspace planning/apply and the CLI use the one-shot path. The [workspace guide](https://libtmux.org/en/erlang/latest/guides/workspace-json/) includes executable loader and apply examples. SIGTERM exercises explicit cancellation and returns JSON. Installed-consumer checks also cover terminal Ctrl-C and SIGKILL: both exit by signal without a receipt, retire observed helper/client processes, and preserve the borrowed server and existing panes. Remote steps can remain applied and temporary files can remain; see [CLI interruption](https://libtmux.org/en/erlang/latest/guides/workspace-json/#interrupt-the-cli). Complete performance qualification remains open. MCP packages, servers and SDK integration are excluded by [project policy](https://libtmux.org/en/erlang/latest/topics/project-policy/). Typed command coverage, generated capability inventory, registry delivery, public cross-port filter conformance, published native API pages, shared-site documentation and independent implementation reviews remain open. Current test success does not pass those unfinished surfaces. No registry or documentation publication is claimed. ## Source installation The [source-package check](https://github.com/libtmux/libtmux-erlang/blob/d1b6308440cf9eb8f6647e7ae434bce60e849368/tools/check_source_package.py) builds each application separately with pinned `rebar3_hex` tooling. It serves the exact tarballs through a signed local registry, then creates a fresh Rebar3 consumer whose only dependency is workspace. Core must resolve transitively and compile its helper from packaged C source. Missing compiler support fails installation. The installed examples, native-record imports, model-upgrade check and owned transport checks run against that consumer. CI runs this gate on both supported architectures with tmux 3.3a and 3.8. Inspect its receipt for the commit being consumed. These local source artifacts retain the development version; the check neither publishes to Hex nor proves an independently retrieved registry release. The older compiled-archive check remains separate evidence for its own payload. ## Performance evidence The current suite checks behavior and resource limits. It does not establish a control/grouping/streaming comparison because those execution lanes are not implemented. Any published benchmark must compare equal workloads and values, state the source/tool/platform identities, and report errors and cleanup as well as time. Test-loop durations are not throughput benchmarks.