Skip to content

Latest commit

 

History

History
159 lines (142 loc) · 10.4 KB

OPEN.md

File metadata and controls

159 lines (142 loc) · 10.4 KB

Open items

What is known to remain, in one place, to fix when the time comes. Each item says where it was seen and what would close it. Written 2026-09-18, after goal 2.

AROS defects not fixed

Details in docs/aros-defects.md.

Item Seen in What would close it
C:Unpack does not load on hosted aarch64 ("file is not executable"), shipped or rebuilt goal 1 bootstrap; board thread 17 Find why the loader refuses it (hunk or ELF flags, relocation); then bootstrap through Unpack as first designed
posixc stdout reaches no shell redirection first AROS runs Fix in posixc, or document that CLI tools must write through Output()
posixc reports no EEXIST or ENOENT where POSIX does first AROS runs Fix the errno mapping in posixc
A command that cannot load leaves $RC unchanged goal 2, four-case check Compare with AmigaOS 3.x first; if AmigaOS sets a failure code, set it in the AROS shell too
The AROS Installer cannot install most scripts, nor run unattended, nor report what it wrote or whether it failed 2026-09-18, the Installer modes aros-development-team/AROS#1293 and #1297 (open, another author) implement the missing commands, the V44 language, MANIFEST/K and PRETEND/S; a driven session, a manifest that can be undone and distinct return codes remain, proposed in docs/installer-api.md
collect-aros writes ELF ABI version 1 whatever the ABI, and the loader never checks it 2026-09-18, the ABI field Write the real ABI version per target in collect-aros, and have the loader refuse a mismatch; Pkg will not rely on it either way
The darwin hosted build ships no FFS handler goal 2 Add kernel-fs-afs to the hosted build, or say why it is left out

Proposed upstream, waiting

PR Fix Local branch
aros-development-team/AROS#1238 identify.library decodes dead-end alerts ~/aros-pr-identify, fix/identify-deadend-alerts on jonx/AROS
aros-development-team/contrib#64 Regina gives RC the port's number, RESULT its string ~/aros-pr-regina, fix/regina-arexx-rc on jonx/contrib

When both are merged: drop tools/aros/regina-arexx-rc.patch and the copy step in tools/build-aros-regina.sh, decode a dead-end alert in tests/goal2.sh again, and remove the two worktrees.

Native AROS runs

  • pc-x86_64 in QEMU: done 2026-09-18, tests/native-x86_64.sh, 31 checks, Pkg bootstrapping itself from a channel named by its root (DEPOT:). Two AROS defects found on the way, in docs/aros-defects.md: Lock() misses later directories of a multi-directory assign, and on the CD-booted native system a RAM: directory added to LIBS: is not searched by the library loader. Worth a look in dos.library before an upstream report.
  • m68k, later, at the owner's request: an Amiga emulator (UAE) with the AROS m68k build in ~/Source/aros-project/aros-m68k-build; QEMU does not emulate an Amiga.

Pkg, not built yet

  • Runs on Windows and Linux. build/pkg-test-kit.zip is ready; only macOS has run it. The owner runs it on Windows.
  • A version rule for formats. A package that owns an on-disk format must declare whether its state survives a downgrade, and ROLLBACK must honour it ([PKG22] item 5; README, "On hosted AROS").
  • Block-compressed images. The image is stored whole; chunk size and codec are still open in the planning repository's packaging README.
  • Images over about 49 MB. Bitmap extension blocks are not written, and such an image is refused.
  • Mounting as one command. MOUNTLIST writes the entry and lists the step: commands, but AmigaDOS cannot easily run those lines from Pkg's output, so a person or an agent copies them into a script. Pkg could write that script too (an agent in round 5 wrote one by hand, checked on the host only).
  • A drawer on a mounted AFS+ volume. The AFS+ macOS driver exposes each file's comment and protection word as afsplus.aros.comment and afsplus.aros.protection xattrs (afsplus e026786, ADR-120). PUBLISH could read them as the native source there, before .ameta; today such a drawer needs a .ameta like any host folder.
  • What the progress events found in the engine (docs/events.md), seen 2026-10-09 while connecting them. A first INSTALL into a root that does not exist yet takes no lock and is not written back to its medium (call() locks and flushes only an existing root). A package refused while its files are placed stays partly placed until the same operation is repeated (restore incomplete), while the dependencies placed before it are taken back out. The payload is read and hashed whole, so its steps have no intermediate count. SEARCH through a portal sends no steps, and IMAGE builds its volume without one. Closing each would be: lock and flush the root once INSTALL has created it; take the refused package's placed files back from staging, or keep its dependencies when it cannot be; read and hash the payload in pieces; give portal requests and the image writer steps of their own.
  • The single-file bootstrap, rest of [PKG23]: a database location for read-only media, self-upgrade with a fallback, a lock against two Pkg runs on one root.
  • An upgrade that is refused at dependency level only names UPGRADE; it does not offer to upgrade the dependency in the same run.
  • Installer scripts as a package kind. pkg runs a package's own Installer script through a driven session and undoes it from Installer's manifest. The contract with Installer is proposed in docs/installer-api.md; both sides are unbuilt.
  • Loose ends from 1.8.1. On AROS the root lock is an Exec semaphore: a pkg that crashes or aborts (abort() in pkg_archive.c) keeps it, and every change is refused with 15 until a reset. On Windows .pkg/lock stays in the root, and the Windows lock is not built or tested here. Protection bits are set after the database write, so a cut in between leaves them unset and a repeat reports unchanged. flushed: no is printed on every change to a FAT volume whose handler lacks ACTION_FLUSH.
  • Commands for the download cache. Only PKG_CACHE and the paths in the docs manage it today. Wanted: a verb (or STATUS) that prints where the cache is on this machine and how big it is, including when AROS fell back to RAM:pkg-cache; a NOCACHE switch that neither reads nor keeps downloads for one run; and a way to clean the cache, all of it or only what no installed package still needs (old versions, archives from past nightlies). docs/reference.md still says T:pkg-cache on AROS, while the code uses SYS:.pkg/cache, falling back to RAM:pkg-cache.
  • Network channels, publishing from a GitHub link, Aminet and AmigaOS interoperability, WHDLoad, paid packages, licences listing: designed in the planning repository, none built.
  • pkg.library, the installed record as a shared library. Other programs can read what is installed only by linking libpkg.a, so each carries its own copy of pkg and its own reading of .pkg/db. Proposed: an AROS shared library, pkg.library, built from the same sources, that opens a root and answers the read-only questions: the installed packages with version, kind and files (pkg_list), each against its channel (pkg_status), each file against its signed manifest (pkg_verify), and which package owns a given file. It takes the root lock for reading, so it never sees a change half written. Changes stay with the Pkg command, so signing, consent and the lock keep one owner. An ARexx port on the library, or on Pkg, would give scripts the same answers. To decide: the library's function table and version, how records are handed out (tag lists, or the pkg_sink callbacks as they are), and whether it opens SYS: alone or any root. Closes when a second program lists and verifies installed packages through it on AROS.
  • No [PKG*] gate is claimed. Goals 1 and 2 touch many gates, but none has its pkg-*.json verdict artifact; STATUS.md in aros-next still says spec.

Guidance added from the agent rounds

Rounds 1 to 5 each had an agent use Pkg from the skill alone, on a different task. What they got wrong turned into refusals and hints rather than more text: KIND is required on PUBLISH (a missing kind used to default to application), a near-miss kind is refused with the right one, hint: lines after KEYGEN, a publish that creates its channel, an image install and MOUNTLIST, content: lines in an image's dry run, and a warning when a program kind holds no executable. Still to watch in the next rounds: whether agents choose a root (SYS: or another) without asking, and whether they make a new key when the person's cannot be found. Round 6 put another program in a game's drawer; the agent caught it from a version warning, and since then a file whose $VER names another program is warned about, and a name taken from such a cookie that would replace a published package of another kind is refused. Round 7 handed a teammate's "fix" that was the published build plus one byte, with only a colleague's key at hand; the agent refused and asked. Since then a dry run needs no key and names the key the real publish needs, an image's manifest lists the files inside it, and the dry run compares a new version with the last one file by file; the missing key is class 14 (key), no longer 20.

Icon-launched installation scripts

Provide an installation script and small launcher icon for LunaPaint from contrib. The requested target is AROS x86_64 SMP. UNVERIFIED: the repository location (suggested path gfx/Multimedia/lunapaint), package/channel identity and target ABI. Confirm these before choosing the signed package to install.

Use a shared package/channel description with a launcher for each supported desktop. AROS uses an icon and ToolTypes; other platforms need their own launcher formats. A tooltip supplies descriptive text. ToolTypes supply configuration. The launcher must expose progress, errors and user confirmation, preserve signature checks, and handle a missing pkg executable. Acceptance includes a double-click installation of LunaPaint on the requested AROS target and explicit platform tests for every additional launcher supplied.