huntedbytheirs/spectral
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

readme, roadmap: say what is true now

Browse Source 
The status block no longer says scaffold, "The loop" no longer says none of it
runs yet, and the table of what each verb actually runs names format-patch
rather than a bare git diff.

Configuration is real: the file and its keys, the two environment variables,
init, and a doctor transcript showing what got resolved and from where. What is
left is Later and Far, plus three gaps named rather than papered over: real
mail, a full kernel build, and the fix round trip against a real kernel patch.

The tests section describes the fixture harness and the three tests that want
the host's own tools.

ROADMAP's Now and Next record what landed and what each item does now. The
series paragraph explains why the sidecar records a whole range rather than a
base: a re-roll that regenerated base..HEAD would pick up every commit that
landed on the branch since create and mail it with the set.
This commit is contained in:
huntedbytheirs
2026-09-16 21:52:05 -04:00
parent 1c668c3c1e
commit c19cf75b2f
2 changed files with 206 additions and 119 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+157 -84
View File
@@ -2,11 +2,13 @@
A monolithic kernel work wrapper, to make it easy.
> **Status: scaffold.** The command surface parses, every verb is wired
> through `main` to its module, and the paths resolve. Every body is a
> `todo!()` stub. Nothing builds a kernel, boots qemu, scrapes bugzilla, or
> sends mail yet. What exists is the shape of the tool and the seams to fill
> in, and *Where to pitch in* below is the list.
> **Status: the loop closes.** Every verb below runs. The chain from a random
> open bug to a patch that `git send-email` accepts has been walked end to end,
> including the v2 re-roll threaded off the original. `kernel test` builds a
> tree and boots it under qemu, `doctor` says what is missing before you start,
> and `init` writes the config file. `Next` in `ROADMAP.md` has landed; what is
> left is `Later` and `Far`, plus the three gaps named under *What is not
> proven* below.
## Why
@@ -32,10 +34,12 @@ What each verb actually ends up running:
|---|---|
| `patch check` / `patch format` | `scripts/checkpatch.pl` **from your tree** |
| `patch commit` | `git commit` |
| `patch create` | `git diff <base>` |
| `patch create` | `git format-patch` |
| `patch update` | a rename, or `git format-patch --reroll-count` for a series |
| `patch submit` | `scripts/get_maintainer.pl` **from your tree**, then `git send-email` |
| `kernel test` | `make`, then `qemu-system-x86_64` |
| `kernel quest` | HTTP against `bugzilla.kernel.org` |
| `kernel quest` | the bugzilla.kernel.org REST API |
| `doctor` | git, the tree's own scripts, qemu, and the paths |
checkpatch.pl and get_maintainer.pl are worth calling out. They come from
the tree you are working in, so spectral cannot drift out of date with the
@@ -49,17 +53,19 @@ the command yourself.
## The loop
What the finished tool should feel like. None of this runs yet:
What the finished tool feels like. Every line below has been run:
```console
$ spectral kernel quest
12345 usb: xhci: device does not enumerate after resume
component: Drivers/USB
https://bugzilla.kernel.org/show_bug.cgi?id=12345
# ... go and fix it, in the tree ...
$ spectral patch check
$ spectral patch format # checkpatch --fix, then check again
$ git add -A # `patch commit` commits what you staged
$ spectral patch commit "usb: xhci: re-arm the port after resume" --signoff
$ spectral kernel test # build, then boot it under qemu
$ spectral patch create 000-xhci-port-rearm
@@ -75,11 +81,35 @@ $ spectral patch update 000-xhci-port-rearm.patch -v 2
$ spectral patch submit v2-000-xhci-port-rearm.patch --in-reply-to '<msg-id-of-v1>'
```
Staging is yours on purpose. `patch commit` runs `git commit`, so what goes in
is what you put there; spectral will not sweep a dirty tree into your patch,
because a stray file in a kernel patch is a review comment you did not want.
For a set rather than a single patch:
```console
$ spectral patch create xhci-port-rearm --range master..HEAD \
--cover-letter "Re-arm the port after resume"
~/.spectral/patches/xhci-port-rearm
$ spectral patch submit xhci-port-rearm --dry-run
$ spectral patch update xhci-port-rearm # -v is inferred; every file moves
```
The series directory holds the patches plus a `series.toml` recording the range
it was generated for (both ends as hashes), the revision, the file list and the
cover letter, which is what lets a re-roll happen without retyping the range.
Re-rolling regenerates exactly that range: a commit that landed on the branch
since `create` does not join the set, because a re-roll should not mail commits
nobody asked to send. The blurb's first line becomes the cover letter's subject.
Leave `--cover-letter` off and the set is written with git's template subject in
it, which `git send-email` refuses: an unsendable series is better than a cover
letter nobody wrote.
One deviation from the obvious design: `patch update` is a single subcommand
taking `-v N` rather than a generated `update-v2` / `update-v3`. clap
subcommands are static, and a flag beats a subcommand you cannot tab-complete.
Leave `-v` off and it infers the next revision from the file's current `vN-`
prefix.
prefix, or from the sidecar for a series.
## Install
@@ -108,26 +138,58 @@ You also need:
## Configuration
There is no config file yet. Two paths matter, and both come from the
environment with `$HOME`-relative defaults:
Two paths and a couple of settings, resolved in this order: the environment,
then `~/.config/spectral/config.toml`, then the built-in default.
| what | resolution |
|---|---|
| kernel tree | `$SPECTRAL_KERNEL`, else `~/.spectral/linux` |
| patch dir | `~/.spectral/patches` |
| what | environment | file key | default |
|---|---|---|---|
| kernel tree | `$SPECTRAL_KERNEL` | `kernel_tree` | `~/.spectral/linux` |
| patch dir | `$SPECTRAL_PATCH_DIR` | `patch_dir` | `~/.spectral/patches` |
| always CC | | `always_cc` | none |
| send-email identity | | `send_email.from` | git's own `user.email` |
```console
$ export SPECTRAL_KERNEL=$HOME/src/linux
$ spectral init --tree $HOME/src/linux # or --clone <url> [--depth N]
wrote /home/you/.config/spectral/config.toml
tree: /home/you/src/linux
```
`init` verifies the tree before it writes anything, and refuses to overwrite a
config that points somewhere else unless you pass `--force`. Run it twice with
the same tree and the second run says so and changes nothing.
A tree is accepted only if it has `scripts/checkpatch.pl`. Otherwise you get
`not a kernel source tree` rather than a confusing failure three steps later.
A missing tree reports the path and the variable that would have set it. The
tree is not needed for `kernel quest`, which only wants the network.
`~/.config/spectral/config.toml` and a `spectral init` to clone the tree are
planned, not built. `src/config.rs` is the only file that has to change to
add either.
A path in the config file has to be absolute, and an unknown key is an error
rather than a shrug: a mistyped setting that silently does nothing is worse
than one that stops the run and says which line is wrong.
Run `spectral doctor` to see what got resolved and from where:
```console
$ spectral doctor
ok kernel tree /home/you/src/linux (from the config file)
ok kernel tree shape scripts/checkpatch.pl is there
ok checkpatch.pl scripts/checkpatch.pl (the only thing that decides kernel style)
ok get_maintainer.pl scripts/get_maintainer.pl (the only thing that knows who to mail)
ok MAINTAINERS the maintainer map the script reads
ok git git version 2.55.0
ok git send-email runs
ok sender identity [email protected]
warn mail route no sendemail.* config: --dry-run works, a real send will not
-> configure `git send-email` before dropping --dry-run
ok qemu QEMU emulator version 11.1.1
ok patch directory /home/you/.spectral/patches (exists, the built-in default)
ok config file /home/you/.config/spectral/config.toml
doctor: nothing failed, 1 warned
```
A warning keeps the exit code at 0; a failed check exits 1 and the summary names
the first one, because that is the one to fix first.
## Commands
@@ -140,75 +202,81 @@ Usage: spectral <COMMAND>
Commands:
kernel Find work, build it, boot it
patch Carry a change from working tree to mailing list
doctor Check the tree, the scripts, the mail setup and the paths
init Point the config at a kernel tree, verifying or cloning one
help Print this message or the help of the given subcommand(s)
```
`spectral kernel quest` · `kernel test`
`spectral kernel quest`, `kernel test`
`spectral patch check` · `format` · `commit` · `create` · `submit` · `update`
`spectral patch check`, `format`, `commit`, `create`, `submit`, `update`
Every one of them has `--help` that says more than this README does, and for
now a body that panics with a description of what it is supposed to do.
`spectral doctor`, `init`
Every one of them has `--help` that says more than this README does.
## What is not proven
Named rather than papered over, because a thin wrapper is only worth trusting
where it says where it stops:
- **Real mail is unverified.** No patch from this project has been sent to a
mailing list by spectral. The acceptance evidence is `git send-email
--dry-run` reaching `Dry-OK` with the recipients spectral chose.
- **A full kernel build is not part of the test suite.** `kernel test`'s build
step is exercised against a Makefile that writes the image; the boot is
exercised against a real qemu with a real `bzImage`, but nothing here builds
a kernel from source, because that is hours of work that proves nothing about
this crate.
- **`patch format`'s fix round trip is proven against a made-up patch**, not
against a real kernel posting whose damage checkpatch can repair.
## Tests
```console
$ cargo test --locked
```
No test needs a kernel tree or a network, and none needs `git send-email`
configured. Each one builds a throwaway kernel-shaped git repository in a temp
directory, points `HOME`, `SPECTRAL_KERNEL` and git's global config at it, and
drives the built binary through `CARGO_BIN_EXE_spectral`. The fixture's `PATH`
and git config are the test's, not the machine's, which is how a missing qemu, a
missing sender identity or a missing `git send-email` is produced deliberately
instead of inherited.
One host tool is assumed: git with its `send-email` support. The `doctor` tests
wrap it so a machine without that support still passes, and the send-email flow
tests use the real one, because that is the tool under test.
Three tests do want the host's own tools, and they are ignored unless asked for:
```console
$ SPECTRAL_REAL_CHECKPATCH=1 SPECTRAL_REAL_TREE=/usr/src/linux \
cargo test --locked -- --ignored real_checkpatch
$ SPECTRAL_REAL_GETMAINTAINER=1 SPECTRAL_REAL_TREE=/usr/src/linux \
cargo test --locked -- --ignored real_get_maintainer
$ SPECTRAL_LIVE=1 cargo test --locked -- --ignored live_bugzilla
```
They are where a kernel-version change to checkpatch's summary line, or to
get_maintainer's role words, or to bugzilla's REST answer, shows up as a test
failure rather than inside someone's patch workflow.
## Where to pitch in
Every stub is a small function that already has its signature, its doc
comment, and its caller in `main`. Pick one and you only need to read the
file it lives in.
The verbs are written and the loop closes. What is left is in `ROADMAP.md`:
`Later` holds more quest sources, more checks, review tracking and packaging;
`Far` holds the TUI and `spectral auto`.
Roughly in order of how much they unblock.
Two of those are more useful than they look. `kernel/quest.rs` is behind the
`QuestSource` trait, so a syzbot source is one impl and no change to the
command. `patch/series.rs` holds the sidecar a series carries, and review
tracking is the feature that will want to add to it.
### patch submit
The most self-contained win, and the only one that needs no network. It does
need a kernel tree and a working `git send-email`.
- `patch/maintainers.rs` · `lookup` runs `get_maintainer.pl --git` over the
patch, splits the output into `To:` and `Cc:`, and collects the files the
patch touches.
- `patch/maintainers.rs` · `add_cc` folds `--cc` flags in without duplicates.
- `patch/mod.rs` · `submit` ties those two together with `git send-email`,
with `--dry-run` stopping one step short.
### patch check, format, commit, create
The everyday verbs. `patch/checkpatch.rs` · `run` already has the `Target`
enum (working tree, a revision, or a patch file) and `Report` with
`is_clean`. What is missing is the process call and pulling the error and
warning counts out of checkpatch's output.
### patch update
Small, and worth writing a test for. `patch/mod.rs` · `reroll_path` holds the
naming rule in one pure function: strip one leading `vN-`, then prepend
`v<revision>-`. Running it twice at the same revision should do nothing,
which is easier to assert than to explain.
### kernel quest
The fun one if you like HTML. `kernel/quest.rs` · `Bugzilla::fetch_open` is
the only place `reqwest` and `scraper` earn their place in `Cargo.toml`, and
`run` then applies `--filter` and picks one issue. The `QuestSource` trait is
where you would add syzbot or a lore.kernel.org thread later.
### kernel test
Needs a machine you are willing to boot kernels on. `kernel/qemu.rs` ·
`build` then `boot`. Streaming serial output as it arrives beats buffering it
until qemu exits, and it is the difference between a boot hang you can
diagnose and one you cannot.
### git plumbing
`git.rs` · `current_branch`, `diff_against`, `commit`, `rev_parse`. Each is a
few lines over the `run` that is already written, which makes these the
easiest place to start.
There are no tests yet. The naming rule in `reroll_path` and the recipient
split in `lookup` are the two that most want one.
Longer-term work, and the things spectral deliberately will not do, are in
`ROADMAP.md`.
The three gaps at the end of *What is not proven* are honest small tasks: a
real send, a real kernel build under `kernel test`, and the fix round trip
against a real kernel patch.
## Development
@@ -216,9 +284,10 @@ Longer-term work, and the things spectral deliberately will not do, are in
$ cargo fmt --all
$ cargo clippy --all-targets -- -D warnings
$ cargo test
$ cargo doc --no-deps
```
All three are clean on `main` and are the bar for a change.
All four are clean on `main` and are the bar for a change.
A few conventions:
@@ -228,12 +297,16 @@ A few conventions:
- Command modules return what they produced and `main` prints it, so output
formatting lives in one place.
- Errors go through the one `Error` enum and `?`. Nothing in the command
paths calls `unwrap`.
- A new stub gets a `todo!()` naming the command it will end up running. A
module that is unreachable until its caller exists carries a single
`#[allow(dead_code)]` with the reason attached. There are no crate-wide
allows, so the warnings come back as the stubs get filled in.
- No `unsafe` and no FFI.
paths calls `unwrap`. An external command that fails reports what it said,
from stdout as well as stderr, because git puts some refusals on stdout.
- Every module that shells out has a comment saying why that tool and not
another API: `checkpatch.pl` and `get_maintainer.pl` come from the tree, and
`git format-patch` does the series naming because that is already its job.
- No `unsafe` and no FFI. `std::env::set_var` is `unsafe` under edition 2024,
which is why environment-dependent behaviour is tested through the binary
rather than in a unit test.
- No `todo!()` is left in the command path. If you add one, say in the test
suite what will prove it gone.
## License