Roadmap

Where work stands today (May 2026) and what’s queued up next. Work is tracked as two parallel streams: the slice queue below (kernel/userspace work units, each one PR) and the GitHub issue tracker (everything else). This page is the map between them.

For the chronological log of what’s already shipped, see history.

Slice queue

Kernel and userspace work, numbered as focused mergeable units — usually one PR per slice.

Slice numbers were resequenced in May 2026 so the queue runs 1..N without gaps or duplicates. CLAUDE.md’s slice queue is the source of truth; this table mirrors it.

Slice	Title	Status	Issue
1	Reaper for dead-task user PDs	✅ shipped (`fcb8771`)	—
2	Per-task `task_t` plumbing	✅ shipped (`3a0ef78`)	—
3	Ring-3 lifecycle ktest	✅ shipped (`f48d730`)	—
4	100 Hz timer + `SYS_WRITE_SERIAL`	✅ shipped (`5e40001`)	—
5	Keyboard rewrite (layered decoder)	✅ shipped (#124)	—
6	Keyboard hardening	✅ shipped (#127)	—
7	Test-infra cleanup (ccache, fan-out CI)	✅ shipped (#125)	—
8	Per-task consumer migration (vtty/fd/cwd)	✅ complete	—
9	Linux-style signal subsystem	✅ shipped (#154)	#144
10	Preemption hardening	✅ shipped (#154)	#145
11	Per-TTY screen buffers + status bar + `/proc`	✅ shipped (#129)	—
12	Per-task FD table	✅ shipped (#134)	—
13	VFS `task->cwd` authoritative	✅ shipped (#135)	—
14	makbox multicall + `SYS_GETCWD` + exec race fix	✅ shipped (#137)	—
15	fork() + copy-on-write (15a-e)	✅ shipped (#166)	rolled into #121
16	execve + wait4 + `TASK_ZOMBIE` (16a-b)	✅ shipped (#166)	rolled into #121
17	UTF-8 terminal	⏭ deferred	#148
18	`ps`-style task listing (covered by `/proc/tasks`)	⏭ deferred	#147
19	VGA-text fallback per-TTY	⏭ queued	#146
20	Userland shell (`sh.elf`, multi-PR feature, 20a-f)	🟢 20a-c shipped, 20g (POSIX A1-A3 pipes/redir/list-ops) shipped via PR #181	—
21	COM2 serial-input mode for ui-test runner	⏭ planned	—
28	`SYS_PIPE` + `SYS_DUP2` + `FD_KIND_PIPE`	✅ shipped (PR #181)	refcounted `pipe_ring_t`, 4 KiB ring, blocking via `task_yield`. Single-arg `SYS_DUP` + FILE-kind open_file_t refcount still pending.

Bash-flavoured shell scripting

Shipped in #163 (May 2026): variables (NAME=value, $VAR, $?), if/elif/else/fi, while, for ... in, [ TEST ], sh script.sh, ./script.sh, makbox dispatch restriction, datetime/date/time builtins. Not on the slice queue (was a single PR that didn’t fit the slice cadence); covered in scripting.

Active themes

Higher-level groupings that cover several slices and/or issues.

Userspace + libc

Long-term goal: a self-hosting userspace. Get a real libc on Makar, then port dash (or extend the in-kernel shell), then bring up TCC for in-place compile-run. Detailed plan in Userland libc.

TCC port progress (May 2026, v0.9): all original phases shipped, and self-hosting now extends to the kernel itself. tcc.elf cross-builds against the libc shim and lives at /apps/tcc.elf on every ISO; userspace apps (hello, calc, sh, makbox) self-rebuild in-OS; the kernel rebuilds end-to-end with ./build-kernel-tcc.sh (host) or /apps/rebuild-kernel.sh (in-OS) — see rebuilding the kernel. Earlier groundwork: Phase 1 brought writable FD_KIND_FILE, O_CREAT/O_TRUNC/O_APPEND, SYS_STAT/FSTAT, SYS_READDIR, growable buffers (SYSCALL_FILE_MAX lifted to 16 MiB). Phase 2 brought the userspace libc shim: malloc/free/ realloc/calloc over SYS_BRK, <ctype.h>, strtol/atoi/strdup/ qsort/sscanf/getenv, setjmp/longjmp, a FILE* layer with snprintf family. Coverage: filetest.elf + alloctest.elf plus test_file_fd ktest suite. Phases 3+ shipped TCC itself + the self-rebuild milestones; v1.0 stays gated on the polish phase (full-green tests, hardened harness, “10× dev experience” tooling).

Slices 15+16 closed the fork/execve/wait4 half of #121; the remaining libc/musl/dash half is still open under the same issue. Slice 20 (paused) carries the userland-shell follow-on.

Networking

NIC driver → lwIP → DHCP/DNS → simple HTTP/Telnet daemons.

Open:

#122 — RTL8139 + lwIP + DHCP/DNS — full stack epic.
#141 — IP/TCP over NE2000 or virtio-net — alternative driver target (NE2000/virtio) for the same stack.

Makar × Medli interop

Sharing data + binaries with the Medli project.

Open:

Background: Makar × Medli.

Kernel hardening

Things the kernel needs to be properly preemptive + signal-aware.

Shipped (PR #154):

#144 — Linux-style signal subsystem (slice 8) — per-task handler table, scheduler-driven default-terminate delivery, SYS_KILL(37) / SYS_SIGNAL(48) / SYS_SIGRETURN(119), ring-3 trampoline so user-installed handlers actually run, Ctrl+C → SIGINT migration (removed g_sigint / keyboard_sigint_consume), task_t.unkillable flag protecting idle + shell tasks, maktop.elf htop-style task viewer with F9 signal picker.
#145 — Preemption hardening (slice 9) — in_schedule re-entrancy guard cleared before task_switch, irq_save_disable / irq_restore wrapping the critical section, per-task kticks accounting in /proc/tasks, runtime-tunable g_sched_quantum (sched_quantum shell builtin), test_preempt busy-loop ktest.

Filesystem & devices

The on-disk FAT32 layout adopts Medli’s canonical structure (/Users/, /Apps/x86/, /System/Data|Logs|Libraries|Modules/, /Temp/ — see Medli’s Paths.cs and Makar × Medli). Alongside it, the kernel exposes Unix-style synthetic mounts so userspace tools have a uniform handle on drivers and devices.

Open:

#151 — Adopt Medli-compatible /Users/ /Apps/ /System/ layout (migrate today’s flat /apps/ to Medli’s canonical FAT32 structure from Paths.cs so user data + binaries move between Makar and Medli without translation)
#152 — Proper HDD install with GRUB2 (Ubuntu/Arch-style) (today’s in-kernel install is fragile: whole-disk format, copied core.img, no grub.cfg generation, no module verification. Fix it into a real installer — depends on #149/#150/#151)

Already shipped:

/proc/ synthetic filesystem with cpuinfo/meminfo/tasks/uname/rtc (#129)
FAT32 R/W; disk filesystems under /mnt (/mnt/root, /mnt/boot, /mnt/cdrom); mount /dev/hdaN /mnt/<name>; flush + unmount on shutdown/reboot
#150 — /dev/ synthetic VFS: raw block devices /dev/hda[N] + /dev/cdrom (devfs); byte-addressed sector I/O
#149 — fdisk: fdisk.elf MBR editor reads/writes the partition table through /dev/hda

Display + terminal

Open:

#146 — VGA-text fallback per-TTY backing buffers (slice 19)
#148 — UTF-8 terminal with ASCII fallback (slice 17, deferred)

Shell

Open:

#147 — ps-style task listing (slice 18; today’s cat /proc/tasks covers this)
Slice 20 — userland sh.elf (multi-PR feature, paused)

(serial shell over COM1 lives under the Makar×Medli theme as #139.)

What’s already shipped

See history for the changelog. Highlights:

Kernel: paging, heap, scheduler (preemptive 100 Hz), per-task page directories, ring-3 lifecycle proven by ktest.
Storage: FAT32 R/W, ISO 9660, IDE PIO, VFS with auto-mount, synthetic /proc.
Userspace: ELF loader, ring-3 iret, full POSIX fork + execve + wait4 (slices 15-16 with copy-on-write page-table clone), syscall surface (Linux i386 subset + Makar extensions 200–218), makbox busybox-style multicall, apps (calc, vix, fdisk, basic, kbtester, forktest, execvetest, …).
Display: VESA framebuffer (720p default), VGA 80×50 fallback, per-TTY backing grids, makmux status bar, four independent TTYs with Alt+F1–F4.
Tooling: single-entrypoint run.sh, ccache toolchain image, build-once fan-out CI (4 parallel jobs including a UI-test job).

Tracking conventions

Slices are kernel/userspace work units numbered in the table above. Each slice should land as one focused PR.
Issues use feat(<scope>): <topic> titles for consistency with commit messages. Scopes seen so far: kernel, shell, display, networking, makbox.
PRs reference the issue they close in the body (Closes #N or Fixes #N). History lives in docs/history.md, not in PR descriptions — once the PR is merged, the changelog entry is the canonical record.