|
root / docs / RELEASE_NOTES.md
RELEASE_NOTES.md markdown 229 lines 9.5 KB

zyginit Release Notes

0.2.7 — 2026-07-21

Bug-fix release. Fixes bug 004 — running the zyginit binary as a
non-PID-1 process (classically zyginit -v to check the version) fell through
into supervisor mode and clobbered the live PID-1 control socket, silently
breaking zygctl control of the running system until reboot. Three defenses:

  • Version flags exit safely. -v, -V, and --version print the version
    and exit — but only when not running as PID 1. The Hammerhead kernel passes
    -v (verbose) as a boot-arg to init, so that path must never short-circuit,
    or PID 1 would return from main() and be re-exec'd in a loop.
  • The control socket is guarded. create_socket now connect()-probes the
    path and refuses to unlink/rebind a socket a live server is already listening
    on; a non-PID-1 instance that cannot own the socket exits cleanly. PID 1
    keeps running without a control socket rather than halting.
  • Non-PID-1 supervisor mode is opt-in. A non-PID-1 run now requires
    --supervisor-test; a bare zyginit ... on a host where zyginit already
    owns PID 1 refuses and exits before any setup. The integration harnesses and
    the CLAUDE.md dev-run invocation pass the flag.

Builds clean and passes the full integration suite (104/104) on Reef 0.7.7;
no source changes were required for the 0.7.7 toolchain (its newly-rejected
program shapes are all Active-Object-related, which zyginit does not use).

0.2.6 — 2026-07-21

Bug-fix release. Fixes bug 003 — spurious exit=-1 service failures in
first-boot reports on Hammerhead. Root-cause analysis found exit=-1 was
never an exec-failure code (a real exec failure exits 127); it was the
"could not resolve an exit code" sentinel, produced by two distinct defects:

  • Exec robustness / no silent failure. Under first-boot I/O contention
    (devfsadm + ZFS pool-finalize) the child's exec of a valid start command
    could transiently fail, and the child then exited silently with an empty
    per-service log. The child now retries process_exec (5×, 100 ms backoff)
    and, on final failure, records the reason to both the per-service
    <name>.log (fd 2, survives a privilege drop) and a new central
    /var/log/zyginit/zyginit.log, then exits 127.
  • Reap-race reconciliation (the actual source of -1). The
    contract-empty event on the bundle fd can beat the child's exit-status
    posting; handle_contract_event used to commit the -1 sentinel and tear
    down tracking, after which the catch-all reaper discarded the real code. It
    now defers when there is no hint, no resolvable code yet, and a child is
    still tracked (rt.pid > 0), so the unconditional reap_children
    re-dispatches with the true code within ~1 s. This fixes both the 127→-1
    masking and the independent false-positive where an exit-0 oneshot that
    lost the race was reported FAILED.

New central operational log: /var/log/zyginit/zyginit.log (distinct from
the per-service logs). Adds integration Suite 9b (spawn-failure capture);
97/97 tests pass. The reap-race half rides the Hammerhead contract path
(not exercised by the Linux stubs) and awaits an hh-alpha first boot to
confirm in situ.

0.2.5 — 2026-07-17

Reef 0.7.6 migration; no init-engine behavior change. Migrated the suite
to Reef 0.7.6: Result/Option idioms, TOML inline arrays, and the
socketctlsocket module rename (avoids a clash with the stdlib
net.socket, which net.unix now imports). Internal/tooling release — no
source tarball was cut for 0.2.5.

0.2.4 — 2026-06-13

Service mutual exclusion. Adds the [dependencies].conflicts array:
mutually-exclusive services both fail if both are enabled, and a start is
refused if a conflicting peer is already running. Boot-time resolution fails
both sides and surfaces the reason in zygctl status. A one-sided
declaration is enforced symmetrically (both directions).

0.2.3 — 2026-06-02

Boot/shutdown progress gauge relocated into the header divider. The gauge
previously lived on the volatile last screen row, where interleaved
kernel/daemon writes would scroll it off. It now paints into the header's
horizontal rule. fmt_progress_bar and PROGRESS_WIDTH were removed and
fmt_divider_gauge added in src/ui.reef; two screen rows are reclaimed for
the rolling tape.

0.2.2 — 2026-05-21

Repo content reshape; no behavioral change. zyginit is now
strictly engine + examples + schema docs. The in-repo services/hammerhead/
tree (22 production service TOMLs) has been removed — distros own all
service-definition policy in their own overlay trees. The Hammerhead
team has absorbed the prior set into base/usr/src/zyginit/overrides/.

The examples/ directory is slimmed to four heavily-commented
canonical examples:

  • daemon-with-restart/ — long-running daemon with [restart] policy
  • oneshot-setup/ — type="oneshot" with [dependencies].requires
  • task-with-condition/ — type="task" with [condition].exists_file
  • with-stop-script/[exec].stop + companion stop.sh

scripts/install-to-be.sh was deleted; its "deploy our services to
a BE" purpose evaporates in an engine-only world. (Distros can ship
their own equivalent.)

docs/SMF_MIGRATION.md has been rewritten as a methodology guide
(how to convert an SMF manifest to a zyginit TOML) rather than a
catalog of our specific conversions.

Tarball: ~22 KB smaller than 0.2.1. No clang / reefc / API changes.

0.2.1 — 2026-05-15

Patch release with two operational improvements:

  • type = "transient" in 0.1.x-style TOMLs no longer silently demotes
    to daemon. The parser now treats it as task (the renamed type) and
    emits a deprecation warning. This prevents a service that should run
    as a task from hanging in STATE_STARTING after an incomplete upgrade.
  • Missing /etc/zyginit/services/ directory at boot now produces a
    distinct, actionable log message instead of silently entering the
    idle loop. The operator is told to run the migration tool from a
    recovery medium.

0.2.0 — 2026-05-15

Breaking change. The filesystem layout under /etc/zyginit/ is
restructured. Operators upgrading from 0.1.x MUST run the migration
tool before installing 0.2.0 binaries:

/path/to/scripts/migrate-layout.sh --dry-run    # preview
/path/to/scripts/migrate-layout.sh --apply      # apply

The tool is idempotent (safe to re-run) and refuses to proceed if
both old and new layouts coexist for the same service.

New: per-service directory layout

Each service is self-contained:

/etc/zyginit/services/<name>/
├── service.toml
├── start.sh          (optional)
└── stop.sh           (optional)

enabled.d/<name> symlinks target ../services/<name>/ (the
directory), not ../<name>.toml (the file).

New: type = "task" service type

For services that may run for a while but are expected to eventually
exit cleanly — platform probes, one-time setup work, etc. Replaces
the previously-unused type = "transient".

[service]
name = "acpihpd"
type = "task"

Semantics:

  • Exit 0 → STATE_STOPPED (counts as online in the boot banner)
  • Exit ≠ 0 → STATE_FAILED (counts as failed)
  • Never restarts
  • Skips the daemonize handshake (tasks don't double-fork)

New: [condition] block

Declarative pre-flight gating. Services whose conditions fail at
startup go into a new SKIPPED state — they don't fork, don't fail,
don't loop. Supported keys:

[condition]
exists_file     = "/dev/acpihp"
exists_file_any = ["/dev/acpihp", "/dev/acpi"]
command         = "/usr/sbin/acpi_probe"     # 5-second timeout

All specified keys must pass (AND). Absent keys are no-ops. The
command form is bounded by a 5-second timeout.

New: SKIPPED state in zygctl status

Skipped services render with the · (dim) glyph and the skip reason
in the NOTES column. The boot banner and final card include a
"skipped" counter alongside online/failed.

Hammerhead vendoring

The Hammerhead team runs the migration tool against their override
tree in base/usr/src/zyginit/overrides/ as part of consuming this
source drop. Follow the same dry-run → apply pattern.

Other changes

  • Renamed SERVICE_TYPE_TRANSIENTSERVICE_TYPE_TASK in the API.
  • type = "transient" in TOML produces a deprecation warning and is treated as type = "task" (see 0.2.1).

0.1.8 — 2026-05-14

Late-failure race fix. A 1-second settle period after the last tier
catches daemons that fork successfully but die immediately (e.g.,
acpihpd on Hammerhead without ACPI hardware).

0.1.7 — 2026-05-14

Boot screen no longer scrolls the [Z] banner off the top of the
framebuffer. Drop the trailing newline that was advancing the cursor
past the last row.

0.1.6 — 2026-05-14

illumos build fix: <sys/termios.h> now explicitly included for
struct winsize and TIOCGWINSZ. Linux's <sys/ioctl.h> pulls
those in transitively; illumos doesn't.

0.1.5 — 2026-05-14

(Broken on illumos: missing <sys/termios.h> for TIOCGWINSZ — use
0.1.6+.) Spinner-in-tape redesign (option B): each tape row tracks
its own active flag, lifecycle procs update rows in place,
parallel-startup ready. TIOCGWINSZ for dynamic screen size. Oneshot
single-count fix. Dynamic SERVICE column in zygctl status.

0.1.4 — 2026-05-14

Framebuffer-first mode detection (PID 1 + isatty + no TERM defaults
to truecolor). Sets TERM=sun-color in zyginit env so children
inherit it. Plain-mode emit no longer mislabels event types.

0.1.3 — 2026-05-12

Initial visual pass: [Z] sigil, palette, rolling tape, progress
bar, spinner, final cards, mirrored shutdown UI, zygctl status
banner+table.