Releases

62 commits since v0.5.3. Three of them break existing behaviour — read those first if you are upgrading.

Breaking changes

The session detach chord moves to `ctrl-]` then `d`

Was ctrl-w. That key is termios-special (VWERASE): the line discipline consumed it before the application ever saw the byte. The daemon no longer intercepts ctrl-w or its CSI encodings anywhere — it is forwarded to the shell like any other key.

The leader is now configurable per client through a [session-keys] section and negotiated per attach channel, so two clients attached to the same session can each use their own chord. Pressing the leader twice forwards it verbatim, which is the primitive for nested sessions. (#1202)

Exec-channel requests are tagged, not string-sniffed

The daemon previously decided what an SSH exec request meant by checking whether the string began with min , conflating three intents in one untagged field. Both failure modes were real: an argv could not be sent, because sh -c 'echo A B C' arrived reshelled and A became $0; and a session's own min --version was unreachable, because the daemon claimed the prefix and refused it.

Requests now carry an explicit tag from a vocabulary both sides share — min://shell, min://argv, min://task/run, min://package/build, min://check. Argv words travel as JSON and are exec'd with no shell in between. Anything untagged now belongs to the session, which is the rule that makes the hijack impossible. Adds min session run <session> <task>. (#1308)

Task `env_vars` resolve on the client

A task declaring env_vars.X = { inherit = true } resolved against the daemon's own environment — on macOS, not even on the same side of the VM boundary as the shell that typed the command. Your export could never be visible to it, so the task failed at spawn with environment variable not found regardless. [session.vars] used identical syntax and worked, because activation resolves it client-side.

Values are now resolved where they exist and carried over the same channel env mechanism that already carries the session id. A variable declared inherit but unset now fails on the client, naming the task and the variable, instead of surfacing later as a spawn failure. (#1312)

Added

Diagnostic bundles capture the shutdown cause, host memory, and macOS panic reports. A "my Mac rebooted under Minimal" report could not previously be answered from a bundle: the power collector filtered out exactly the lines that date a restart, nothing recorded host memory to judge a VM allocation against, and the panic report was never read. (#1263)
Bundles record the terminal they were collected on — per-stream tty status, window size, and tty device. Garbled-TUI, wrong-column-wrap, and resize-does-nothing reports are claims about a terminal the bundle never described. (#1264)
Bundles are scoped to the project they came from. (#1265)
New sessions start fish when it is the configured shell and is actually installed, rather than falling back to bash. (#1320)
Loadouts gain a patch source anchor for the loadout root. (#1297)
Patch files keep their permission bits when copied into a session. (#1302)
Nickel parse logic for container specs in decode.

Fixed

Sessions and the daemon

A version-skewed daemon is refused instead of destroying sessions. A CLI upgraded past its minimald failed mid-activation, not at connect: FinalizeSession succeeded on the daemon, its reply would not decode against the newer wire types, and cleanup then tore down the session it had just created. Every session appeared to be created and vanish at once. (#1260)
A stalled host no longer wedges every session RPC. A daemon in the field served GetVersion in 54µs while ListSessions, GetSessionRecord and CreateSession hung for fifteen minutes. The split was exact — the RPCs that hang are the ones that round-trip through the sessions Manager actor, which awaited get_attrs inline. (#1281)
The terminal is unwound when a session process exits, SIGKILL included. Exit notices and the shell-exit prompt were rendered into whatever input modes the dead application had left set. (#1261)
Package-declared ~/ paths expand explicitly, and credentials are dropped. Inside the guest minimald is pid 1 with HOME=/, so ~/.claude.json resolved to /.claude.json and min task run died with EROFS on the read-only rootfs, naming a path nobody had written down. An unset HOME panicked outright. (#1262)
exec returns when the child exits, not when its pipes close. (#1342)
The default mfile is scaffolded before composing. (#1339)
Lifecycle hook stdout and stderr are surfaced during session activation. (#1296)
Activating a path that already has a session now warns. (#1316)
A session that cannot be composed names the directory in the error. (#1340)
The client is told when session hostnames will not route. (#1310)
The proxy bridge exits when the daemon socket closes. (#1282)

Command-line output

mip trace logs go to stderr, not stdout, where they corrupted piped output. (#1315)
min add renders download progress. (#1213)
min init writes its result line to stdout. (#1285)
min add confirmation drops the array brackets. (#1292)
Host-side min add --session output is qualified so it no longer implies a live install. (#1309)
Generated minimal.toml headers name the binary min. (#1284)
min stop --force help matches the live-shell guard it actually applies. (#1294)
The removed-profiles warning names the migration path. (#1300)
--loadout is validated before the no-config banner prints. (#1153)
Sync runs when a minimal.toml is present, before the VCS check. (#1321)
Shells are deduplicated before the completions install loop. (#1283)
command_capture output is bounded by size as well as by time. (#1318)

Build graph and determinism

An unparseable layer-cache entry is a miss, not a fatal error. LayerCacheDir::get deleted the file it could not deserialize and propagated the error anyway, so one poisoned entry failed the whole graph load. A session came up with no package environment at all — no base, git, go, make or vim on PATH — and nothing to say why. (#1259)
min init output is deterministic. Stack tables order by name (#1332), the mfile task table is ordered and a redundant sort dropped (#1334), OutputKind::OciImage vars are ordered (#1333), and package-derived setup has a fixed order (#1319).
The R2 read default survives Config::into_builder(), which rebuilt remote_cache_url unconditionally from a non-optional field and so treated the default GCS bucket as an explicit user choice. (#1291)
decode returns errors instead of panicking.

Installer and dependencies

install.sh wraps its body in main(), so a short read of the piped script cannot execute a truncated program. (#1276)
h2 0.4.16 → 0.4.18. The 0.4.16 bump taken for RUSTSEC-2026-0258 shipped an empty-DATA-frame flood guard that misfires on legitimate clients multiplexing many streams over one connection. (#1255)
wide 1.6.0 → 1.6.1; 1.6.0 was yanked. (#1314)

Documentation

New build reproducibility reference page. (#1254)
The architecture.md crate map covers minimal-client, minimal-tui, and mlog. (#1336)
Documented where { inherit = true } diverges by surface. (#1343)
The min loadout list example matches what the command prints. (#1280)

Fourteen further commits in this range are CI regeneration, dependency-group bumps, test coverage, and a README change, and are omitted deliberately. Two decode commits landed without a pull-request reference and so appear unlinked.

Full changelog: v0.5.3...v0.5.4

minimal 0.5.5-dev.49.gcb384e31

Uh oh!

minimal 0.5.5-dev.33.g24dbbfe2

Uh oh!

minimal 0.5.5-dev.15.gc6cc5fe5

Uh oh!

minimal 0.5.5-dev.14.ga514ac9e

Uh oh!

minimal 0.5.5-dev.10.g8e7e72c2

Uh oh!

minimal 0.5.5-dev.9.g323ae081

Uh oh!

minimal 0.5.4

Breaking changes

The session detach chord moves to ctrl-] then d

Exec-channel requests are tagged, not string-sniffed

Task env_vars resolve on the client

Added

Fixed

Sessions and the daemon

Command-line output

Build graph and determinism

Installer and dependencies

Documentation

Uh oh!

minimal 0.5.3

Uh oh!

minimal 0.5.2

Uh oh!

minimal 0.5.1

Uh oh!

The session detach chord moves to `ctrl-]` then `d`

Task `env_vars` resolve on the client