Verification evidence

Run commands from the repository root after mise install and mise run setup. Provisioning uses the network; verification does not install dependencies.

mise run integration runs transport, core operations, control, decoded configuration and workspace groups in separate serialized child processes. Each retains the same 60-second cap and owned-process cleanup checks. ASDF test-op includes all five groups too.

Performance uses a separate tier: mise run benchmark compares identical native reads and measures graph acquisition separately. It is not run by CI or these test loops. The mid loop includes only the reporter’s four critical evidence checks. See the benchmark guide for sampling, owned cleanup, source identity, resource accounting and interpretation limits.

These commands exercise the implemented native core; they do not establish full-suite compatibility. New operation tests extend the integration suite as those APIs are implemented. check detects generated-document drift without rewriting files. docs deliberately rewrites generated pages and excerpts. Complete catalog reads exercise many native commands. Each of the five integration processes retains the same 60-second ceiling; the outer task runs every group and cleanup checks. No formatting tool is prescribed for Common Lisp; source checks and SBCL compilation catch whitespace and reader/compiler errors respectively. The testing guide describes FiveAM’s interactive workflow and the exact tmux build cache. Hosted results must identify the tested commit; local package or integration success does not establish a hosted CI result.

Recorded local compatibility¶

The native matrix record identifies the exact source revision, native source/binary identities, runtime and per-group results. On 2026-11-09, the existing native integration suite at 0e1a6f3 passed on all twelve supported releases: 3.3a, 3.4, 3.5, 3.5a, 3.6, 3.6a, 3.6b, 3.7, 3.7a, 3.7b, 3.7c and 3.8. The two-cell local sweep took 236.28 seconds, including startup. All cells checked owned-process cleanup. Eleven control cases are explicitly skipped below the optional feature’s 3.8 floor; older-daemon refusal is tested.

This record covers the implemented native integration surface at its pinned revision. It does not establish completion of the unfinished suite. CI continues to test only 3.3a and 3.8; source-archive consumers and complete examples have separate endpoint evidence. The record’s timings are verification costs, not performance measurements or latency guarantees.

Source archives and consumers¶

The package command writes dist/libtmux-source.tar.gz and dist/package-evidence.json. The consumer loads the extracted artifact from a temporary directory with an isolated ASDF source registry and output cache. It checks that query, regex, JSON and thread packages were not loaded by core, then loads libtmux/query separately without pulling those optional libraries. Separate consumers load and exercise libtmux/regex and libtmux/json through the same locked dependency setup. The artifact includes its setup pins, tools, guides and executable examples. package-examples checks its local links and runs its own copy of the example runner with an isolated core registry. dist/examples-evidence.json records the runtime versions, source hashes, tutorial block count and cleanup result. The installed workspace entry point reads the shipped JSON file, creates its declared live topology, and reports zero retained mutations on repeat apply. These checks qualify the implemented v1 behavior; full ownership metadata, drift inspection and explicit reconciliation remain separate unfinished work. The tutorial explains the recipes and expected values.

Select a tmux binary explicitly when exercising a compatibility cell:

$ LIBTMUX_TEST_TMUX=/path/to/tmux mise run integration

Support covers every tmux release from 3.3a through 3.8. CI uses the latest stable SBCL pin and tests only tmux 3.3a and 3.8 to reduce execution time. Intermediate releases remain supported; they are not additional CI jobs.

The fixture uses a unique libtmux-lisp- temporary directory, an explicit socket path and /dev/null configuration. Child environments omit TMUX and TMUX_PANE. A pidfd verifies termination before owned files are removed. Tests never address the default server. Attached-client tests launch ordinary PTY clients through SBCL’s fork path, without control mode. They join captured PIDs to retained child pidfds, observe native SIGTSTP stops, and use an owned bounded FIFO lock command. Request acknowledgements and frontend effects have separate assertions. Partial acquisition and nonlocal exits check cleanup. Session/server locking uses two ordinary frontends in separate sessions. Native suspended flags establish recipient scope, and independent FIFO readiness proves each requested lock program started before owned cleanup releases it. Wait-channel tests prove a waiter is queued through native listing before cancellation, then inspect the daemon’s remaining queue. Lost lock acknowledgements remain separate from queued-waiter disposal; neither closing nor cancellation automatically releases a granted lock. Message-printing tests distinguish literal and explicit format input, physical and link context, foreign authority, versioned intermediate limits and after-hook text in the ordinary printer stream. Key-binding tests send bytes through an owned ordinary PTY frontend. Native hook/channel acknowledgements prove stored command execution and order a second key after removal. Direct fixture reads verify literal values and parse-time environment effects retained after a failed binding. Printer profiles distinguish tmux 3.7 through 3.7c’s single-row status routing and missing-key success from normal multi-row output and other releases. Server-access fixtures use the existing nobody account and its primary group on private sockets, without changing OS identities or permissions. They prove entry lifecycle, literal names, group gating and tmux 3.8’s recomputation of an unrelated owner’s frontend flags. Cross-user authorization, connection and revocation remain unverified.

The integration runner is a dedicated child-process subreaper. It verifies that no test-owned worker or tmux daemon survives the test image, including an injected tutorial failure before its Lisp termination call. Production loading does not change the application’s process-reaping policy. Readiness FIFOs, semaphores and pidfds coordinate tests without polling sleeps.

Unit tests cover floor rejection without adding a third tmux build to CI. For an optional maintenance probe, use the exact 3.3 executable:

$ mise exec -- tools/lisp.sh tools/check-floor.lisp /path/to/tmux-3.3/tmux

This check asserts the executable’s release, expects !unsupported-tmux, and checks that opening the client created no daemon socket.

Broader stream/retained-memory qualification, source filtering, shared-site rendering, broader benchmarks and integrated acceptance of all required capabilities on both CI endpoints remain pending. MCP is excluded by policy. The capability table separates implemented behavior from those gaps. A passing core package or example does not cover them.