-
-
Notifications
You must be signed in to change notification settings - Fork 1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Support command chaining #2222
Comments
Interesting idea! It'd help if someone explored the nuances of how click worked so we can build on their shoulders. What limitations does it have? How does it impact various UX features? What could be improved? I'm assuming any App that has this enabled would disallow sub-subcommands because that can be ambiguous with chaining. We do this with checks in We'd need to decide on how to expose this in
|
Copying over an example from another thread. This is a demonstration of how this could look using derive. Ideally this should be implemented for any type that is IntoIterator<Item = T>. #[derive(Parser)]
#[clap(rename_all = "snake_case")]
pub enum MyCli {
GetPrices {
#[command(subcommand)]
assets: Vec<AssetInfo>,
},
}
#[derive(Subcommand)]
#[clap(rename_all = "snake_case")]
pub enum AssetInfo {
Token {
/// <string>, "atom3h6lk23h6"
contract_addr: String
},
NativeToken {
/// <string>, "uatom"
denom: String
},
} This is a much simpler case then I have in reality, but useful for examining how this could work. An example run of the program could look like: my-cli get_prices token atom3h6lk23h6 native_token uatom When these example args are parsed it would result in: MyCli::GetPrices{
assets: vec![
AssetInfo::Token { contract_addr: atom3h6lk23h6 },
AssetInfo::NativeToken { denom: uatom }
]
} |
Why would we need to stop chaining when entering a subcommand? |
Allowing both chaining and sub-subcommands has the potential for ambiguity (in practice and from the users perspective) and implementation complexity. What will help is someone exploring what precedence there is in other implementations of chaining, exploring the various cases for how to avoid ambiguity, and exploring the use cases for why someone would need each kind of behavior we would implement for this. |
It should be possible to have some check to prevent ambiguities but yes I see the problem you're referring to. As far as use case goes for me it's as simple as trying to force clap to work as well as possible with any arbitrary enum structure using its derive functionality. |
FWIW, here is the relevant paragraph in Click's documentation. The limitations are:
|
I found this issue via #4424 and wanted to add my use-case. I'm not sure how this would work implementation-wise, but it'd be very useful: I'm building a fairly complex CLI that supports a dynamic number of duplicate configs. It's proxying traffic via different protocols, and adding filters on top of them. So for example, I might want to listen on tcp and udp, and parse tcp messages as strings but the udp messages as protobufs, and log the raw bytes first (just random simple examples). I can think of two logical approaches, and would be content if either were supported by clap; I'd just use whichever it supported. The first is
The second is more what this issue (and #4424) is discussing, if I understand correctly:
I personally find the first pattern more intuitive because of prior exposure to things like the
I've considered solving this short-term by requiring that you pass a number of --filter arguments equal to --uri arguments, but this starts breaking down with variable number of filters, or other arguments that should only apply to one of the listeners. It also requires doing a lot of custom parsing if you want to group multiple arguments (e.g. make --filter a If someone knows a way to achieve this already in clap, I'd love to hear it, and otherwise would love to see support get added. Clap has been such a fabulous and pleasant asset when working with rust CLIs. |
This comment was marked as off-topic.
This comment was marked as off-topic.
This comment was marked as off-topic.
This comment was marked as off-topic.
This comment was marked as off-topic.
This comment was marked as off-topic.
Wow, there's a lot happening there :-) I haven't digested all of it, but
at some point it seems like doing it that way is more trouble than just
manually parsing args from argv with something like nom (at least for my
use case). I'd _love_ to see repeated subcommands get added to clap.
It'd solve all my problems... I'd be happy to help contribute it, if we
came up with a design. I don't really know anything about clap
internals, and have only ever used it via the derive macros.
…On 3/27/23 13:28, Antoine Beyeler wrote:
@brownjohnf FYI, in [this crate](https://github.com/abey79/vsvg/tree/master/vsvg), I've pushed the "find" example to some extremities. I support repeated options and order, which the "find" example doesn't support IIRC. I also made some macros to reduce the boilerplate of defining commands. There are loads of limitations to it and, being new to Rust, this is likely messy and non-idiomatic on multiple counts, but it gets me far enough for this project.
|
For this to move forward, we need
|
OK, here is my take. Coming from Click, this paragraph describes pretty well what I would expect. For background, I created and maintain a Click-based CLI pipeline called vpype. I think it's an excellent example of how a chained multi-command pipeline CLI should behave (but I'm slightly biased). Note that I make this proposal with a very minimal knowledge of Clap's internals and a rather thin experience of actually using the library. So it may very well make little sense. In exchange, you get a fresh/candid feedback on one person's expectations for a CLI framework "with all of the bells and whistles". ;) Here is how I'd transform the builder tutorial to use command chaining (I marked my changes with use std::path::PathBuf;
use clap::{arg, command, value_parser, ArgAction, Command};
fn main() {
let matches = command!() // requires `cargo` feature
.chain(true) //NEW: enable chaining of sub commands
.arg(arg!([name] "Optional name to operate on"))
.arg(
arg!(
-c --config <FILE> "Sets a custom config file"
)
// We don't have syntax yet for optional options, so manually calling `required`
.required(false)
.value_parser(value_parser!(PathBuf)),
)
.arg(arg!(
-d --debug ... "Turn debugging information on"
))
.subcommand(
Command::new("test")
.about("does testing things")
.arg(arg!(-l --list "lists test values").action(ArgAction::SetTrue)),
)
.subcommand( //NEW: let's add another command for example purposes
Command::new("config")
.about("configures next test(s)")
.arg(arg!(-s --set <PARAM> <VALUE> "sets a config value")),
)
.get_matches();
// You can check the value provided by positional arguments, or option arguments
if let Some(name) = matches.get_one::<String>("name") {
println!("Value for name: {}", name);
}
if let Some(config_path) = matches.get_one::<PathBuf>("config") {
println!("Value for config: {}", config_path.display());
}
// You can see how many times a particular flag or argument occurred
// Note, only flags can have multiple occurrences
match matches
.get_one::<u8>("debug")
.expect("Count's are defaulted")
{
0 => println!("Debug mode is off"),
1 => println!("Debug mode is kind of on"),
2 => println!("Debug mode is on"),
_ => println!("Don't be crazy"),
}
//NEW: iterate over sub-commands
for subcmd in matches.subcommand_iter() {
match subcmd.name() {
"test" => {
if subcmd.get_flag("list") {
println!("Printing testing lists...");
} else {
println!("Not printing testing lists...");
}
}
"config" => {
println!("Configuring next test(s)...");
}
_ => unreachable!(),
}
}
// Continued program logic goes here...
} This would yield the following help string (directly inspired from Click):
And here is how I'd adapt the derive tutorial: use std::path::PathBuf;
use clap::{Parser, Subcommand};
#[derive(Parser)]
#[command(author, version, about, long_about = None, chain = true)] //NEW
struct Cli {
/// Optional name to operate on
name: Option<String>,
/// Sets a custom config file
#[arg(short, long, value_name = "FILE")]
config: Option<PathBuf>,
/// Turn debugging information on
#[arg(short, long, action = clap::ArgAction::Count)]
debug: u8,
#[command(subcommand)]
command: Option<Commands>,
}
#[derive(Subcommand)]
enum Commands {
/// does testing things
Test {
/// lists test values
#[arg(short, long)]
list: bool,
},
/// configures next test(s)
Config { //NEW
/// sets a config value
#[arg(short, long)]
set: Option<(String, String)>,
},
}
fn main() {
let cli = Cli::parse();
// You can check the value provided by positional arguments, or option arguments
if let Some(name) = cli.name.as_deref() {
println!("Value for name: {name}");
}
if let Some(config_path) = cli.config.as_deref() {
println!("Value for config: {}", config_path.display());
}
// You can see how many times a particular flag or argument occurred
// Note, only flags can have multiple occurrences
match cli.debug {
0 => println!("Debug mode is off"),
1 => println!("Debug mode is kind of on"),
2 => println!("Debug mode is on"),
_ => println!("Don't be crazy"),
}
//NEW: iterate over sub-commands
for subcmd in cli.command_iter() {
match subcmd {
Commands::Test { list } => {
if *list {
println!("Printing testing lists...");
} else {
println!("Not printing testing lists...");
}
}
Commands::Config { set } => {
if let Some((param, value)) = set {
println!("Setting config value for {param} to {value}");
}
}
}
}
// Continued program logic goes here...
} In most cases, some dispatch mechanism ( [(*) BTW, I'm slightly surprised it was marked as off-topic, as it provides a (somewhat) working of command (well... argument) chaining, which is a work-around for the present issue. Maybe I wasn't sufficiently clear about this in that comment?] |
Oh, that whole conversation was useful but since it ran its course, I was wanting to make it easier to find the relevant parts of the conversation for command chaining and to not get too off topic. I also feel like "argument chaining" is only a workaround in some limited cases and comment I did leave expanded will likely lead people to expand the follow ups. |
I assume that should be a |
Actually, since On the other hand, |
Could we be more explicit, rather than linking out to other documentation. If nothing else, it keeps the conversation in one place so we can make sure we are satisfying all of the requirements. For example
I assume this means that these commands effectively end the chain.
We enforce invariants as
Could we make this like the |
Additional variants are just additional subcommands at the same level. We still need a way to store (and access) multiple instances of the enum |
If I understand you correctly (which I'm not entirely sure of), you mean that the enum in the example represent "level 1" sub-commands, which are by definition exclusive in the current (chaining-less) model, so adding variants and expecting that multiple instances exist is in contradiction with the current implementation. I guess it's both a technical (on which I can't comment much) and API question. On the API side, I think I would be happier with What's for sure is that enforcing as single level of subcommand when Edit: I'll try to answer your other message later tonight. |
Not quite. Take prog test and that would map to Cli {
name: None,
config: None,
debug: 0,
command: Some(Commands::Test {
list: false,
})
} With that said, what would the following map to? prog test test test config |
(sorry, re-posting from the right account :-)) From an API perspective, I'd like to see something like this: #[derive(Parser)]
struct Cli {
#[clap(subcommand)]
subcmd: Vec<Subcmd>,
}
#[derive(Parser)]
enum Subcmd {
Listen {
#[clap(short, long)]
port: u16,
#[clap(short, long = "proto")]
protocol: String, // Should be an enum, but this is an example
// etc....
}
}
fn main() {
let cli = Cli::parse();
for sub in cli.subcmd {
println!("{:#?}", sub);
}
} And then used like this: $ mycli listen -p 2345 listen -p 1234 # etc...
Listen {
port: 2345,
}
Listen {
port: 1234,
} While it would be wonderful to get "arg chaining" (I guess you'd call it?) a la In terms of algorithm/parsing, as a developer I'd expect clap to:
So, if you have top-level flags or args and multi-subcommands, the subcommands would parse "greedily" until they fail, and then start all over at the beginning. |
I've managed to confuse myself by hastily answering on the move with my previous messages—sorry for that. The
Right, here is (a starting point for) the spec of sub-command chaining. Sub-command chaining can be enabled on a given command (defaults to disabled), and affect the way its subcommand are handled.Builder API: let matches = command!()
.chain(true)
.arg(...)
.subcommand(...)
.subcommand(...)
.... Derive API: #[derive(Parser)]
#[command(chain = true, ...)]
struct Command {
#[clap(subcommand)]
subcmd: Vec<Subcmd>,
...
} A command with chaining enabled accepts more than one sub-command.That is kind of the point of this whole issue :) A given chained sub-command may be repeated multiple times and order of invocation is maintainedThis means that Sub-command chaining may be enabled on an arbitrarily nested commandThis is slightly more general than my initial take. Taking git as a well known example CLI:
Here I hypothetically enabled sub-command chaining for This seems implied by the Click documentation, although I haven't tested that it actually works. Chained sub-command may not have nested sub-command(s)Taking the same example as before, this is not allowed:
where Likewise, since
In other word, chained sub-commands must be leaves in the command tree. Chained sub-commands must parse greedilyAs @brownjohnf, when it appears, a chained sub-command should eagerly parse all arguments until failure, at which point the parser must expect the next sub-command. This implies that Click's Technically, it think this would mean that the "chained sub-commands must be leaves" rule could be lifted as well. Both of the counter-example above could work. In the first, I'm not sure of the technical implication of that though, and it would certainly make for a potentially confusing UX. Notes/TBD:
|
I was looking at the parser which reminded me of some of the complexities involved, including
I'm starting to suspect that the cheapest / easiest way to implement this will be for the parent command to copy its subcommands to the command being parsed.
This would cause the
Except even that comes with its own challenges
|
The way Click does it is Real world example (abbreviated):
Some additional spec/remarks come to mind. The
For this to work, nothing must happen besides showing help when Another aspect is sub-command grouping in the top-level help, as shown above. All of these commands are "sister" sub-command of the top-level CLI—no nesting or anything. Yet they appear thematically grouped. Click doesn't support it out-of-the-box and a stack overflow recipe is needed. It's simple enough, but it would be nice to have built-in support. Finally, in my previous message I mentioned variable length arguments. To clarify, tt would be really nice to have both variable length options and positional arguments. The |
#### Chained sub-command may not have nested sub-command(s)
In other word, chained sub-commands must be leaves in the command tree.
I'm not sure why we'd add this restriction? It seems harder than just
blindly following recursion as far as it goes? Once all input has been
parsed, or no further progress can be made with remaining arguments, we
return.
This might be a bit of a hot take, but as a developer I'm fine with
allowing people to create confusingly unusable interfaces. I've gotten
myself into some corners just using the existing quite rich capabilities
in clap as it is. As an end user, I see it as either a) I'm using a
power tool (like find) that just requires reading the manual to
understand what it's going to do or b) the CLI author failed to design a
reasonable interface. Neither case seems like it's clap's fault.
As @brownjohnf, when it appears, a chained sub-command should eagerly
parse all arguments until failure, at which point the parser must expect
the next sub-command.
I'd loosen this up a bit. A sub-command should eagerly parse all
arguments until failure, at which the parser should pop up the stack and
resume parsing the _parent_ (sub)command until failure, etc.
This allows arbitrarily nesting (sub)commands, and provided we're
explicit about "the parser is greedy" I think it makes things easy to
reason about, even if it does theoretically allow people to make very
confusing CLIs. Here's an example though of something that I think makes
sense to me (contrived, and indented for clarity):
```shell
$ mycli \
listen -p tcp -h localhost \
listen -p udp \
proxy -p / --proto tcp -h 123.456.0.1 -vvv \
proxy -p /foobar -h 145.123.2.3 --log-level ERROR \
-v
# Interpreted as:
# With default log level of INFO, listen on tcp://localhost, and do
# nothing. listen on udp:// using default host and proxy all traffic on
# / to tcp://123.456.0.1, and enable TRACE logging. Proxy a subset of
# traffic on /foobar to upstream host 145.123.2.3 using the default
# protocol, but only log at an ERROR level.
```
That seems like a greedy recursive parser would have no issue with that
(I realize in practice, implementing this in clap is more complex).
…On 3/30/23 01:54, Antoine Beyeler wrote:
I've managed to confuse myself by hastily answering on the move with my previous messages—sorry for that. The `command` field would indeed need to be a vec. Also, I agree with everything that @brownjohnf just said.
> Could we be more explicit, rather than linking out to other documentation.
Right, here is (a starting point for) the spec of sub-command chaining.
#### Sub-command chaining can be enabled on a given command (defaults to disabled), and affect the way its subcommand are handled.
Builder API:
```rust
let matches = command!()
.chain(true)
.arg(...)
.subcommand(...)
.subcommand(...)
....
```
Derive API:
```rust
#[derive(Parser)]
#[command(chain = true, ...)]
struct Command {
#[clap(subcommand)]
subcmd: Vec<Subcmd>,
...
}
```
#### A command with chaining enabled accepts more than one sub-command.
That is kind of the point of this whole issue :)
#### A given chained sub-command may be repeated multiple times and order of invocation is maintained
This means that `cli op1 op2 op1` is allowed and is not equivalent to `cli op1 op1 op2`. Clap's API must reflect the repetition and the order.
#### Sub-command chaining may be enabled on an arbitrarily nested command
This is slightly more general than my initial take. Taking git as a well known example CLI:
```
git submodule foreach op1 --arg op2 -a op3 op1
```
Here I hypothetically enabled sub-command chaining for `foreach`—itself a twice-nested subcommand—such that it accept a sequence of subcommands (`op1`, `op2`, `op3`).
This seems implied by the Click documentation, although I haven't tested that it actually works.
#### Chained sub-command may not have nested sub-command(s)
Taking the same example as before, this is not allowed:
```
git submodule foreach op1 --arg op2 -a subop op3 op1
```
where `subop` would be `op2`'s sub-command.
Likewise, since `submodule` may not accept chaining sub-command as per this rule, and assuming `summary` is a sub-command of `submodule` (as is the case in the real-world git), this would not be allowed either:
```
git submodule foreach op1 --arg op2 -a op3 op1 summary
```
In other word, chained sub-commands must be leaves in the command tree.
#### Chained sub-commands must parse greedily
As @brownjohnf, when it appears, a chained sub-command should eagerly parse all arguments until failure, at which point the parser must expect the next sub-command.
This implies that Click's `nargs=-1` restriction does not apply. Any chained sub-command may have any number of variable lengths arguments. (FWIW, I have actually suffered from this restriction of Click before, locally forcing me into a lesser UX.)
Technically, it think this would mean that the "chained sub-commands must be leaves" rule could be lifted as well. Both of the counter-example above could work. In the first, `subop` is recognised as `op2` sub-command so parsing continues one level down. On the second example, `summary` is not recognised as `op1`'s nor `foreach`'s sub-command, so parsing continues one level up and `summary` is recognised as sub-command of `submodule`.
I'm not sure of the technical implication of that though, and it would certainly make for a potentially confusing UX.
<br/>
*Notes/TBD*:
- Clearly, some of these rules (in particular the "chained sub-commands must be leaves" rule) are somewhat arbitrary. I'm stating them as starting point, to trigger conversation. In particular, as @epage mention, there might be technical reasons to prefer one way or the other, and I'm fine with it.
- As @brownjohnf mentioned, I completely left aside the question of find-like argument chaining. I think starting with sub-command chaining is the correct strategy, and it covers most use-cases in a way that, I believe, is more trendy by today's standards.
- I kept using the `chain` terminology, for the lack of a better one. Although I agree with @brownjohnf that the use-case may or may not care about order, Clap should definitely keep the ordering properly, for those use-cases which need it (e.g. pipelines). `multiple` may be alternative, but it sound a bit too generic and is less explicit about the ordering aspect. I don't have another suggestion at this point.
|
- Should `prog chain0 --help` list all the chained commands like
`prog --help`?
I think it should function the same as if there were no chaining.
- Should both levels of usage show all the chained commands?
I don't actually remember how this works today with subcommands, but I'd
say the behavior should remain the same.
- The parser for a chained command will need access to the parent
command (for command flags, knowing whether an arg is a command or
value) which will run into multiple borrows of a mutable value (`prog`
owns `chain0`, `chain0` will need to be mutable for the lazy building,
`prog` and `chain0` will both need to be accessible when parsing)
I'm not sure I'm following, but I'd assumed that this would be
recursive, and the current context ((sub)command) wouldn't know about
the parent. On failure or end of args it would just return whatever it
did parse about itself + remaining args it didn't recognize. Then the
parent command would resume attempting to parse it, which might recurse
down into another subcommand, or might resume parsing flags belonging to
that command.
- In general, I'm leaning towards the recursive approach. Its the
simplest, has no overhead for unchained subcommands (the more common
case), and the builder API doesn't need to be as ergonomic as that is
what the builder API is for
I think that makes sense to me.
- The main complication is the derive API knowing when to do this
(it'll have to capture the state of any chaining) and knowing when to
stop if we support unchained subcommands within a chained subcommand.
I'm not following this though. I'm pretty ignorant here, but I'd assume
the derive API (I assume by that you mean the macro?) would just enable
recursive parsing whenever it encountered a `Vec<T>`.
- Subcommands under chained subcommands would get complicated if we
move forward with deferred initialization of `Command` because we can't
tell if subcommands are present until we build it, but we shouldn't
build it until after we've copied
- This doesn't work for any call that wants to just do a
`prog.build()` as that shouldn't recurse infinitely
Also not following this, but I think it would be reasonable to say that
you can't re-nest the same subcommands under themselves. i.e., your
interface needs to be an acyclic graph. You can only have one subcommand
field in any clap Command, and you can't have a subcommand `T` as be a
subcommand of itself (including transitively).
|
Clap help behaves that way to do.
See |
@brownjohnf I agree with your (hot) take and we may certainly lift that "no nesting of chained sub-command" rule. I've included it here because of my understanding that it simplified implementation on Click's side, but I understand things might be entirely different on Clap's side. |
In thinking on chained subcommands, I realized that one challenge is clap is smarter than that allows. At least for optional positionals and variable length positionals, clap checks for if a subcommand name exists among them and will then treat it as a subcommand. At least for variable-length options, clap has a setting to check for subcommand names among them. This requires subcommands to know every subsubcommand name and alias, without getting into subcommand short flags and long flags. This presents an implementation challenge because
So we either need a solution that can have consistency out the gate or we need to be very restrictive and disallow any case that requires knowing about chained commands so we can just pop back up and continue processing. I'm assuming that would be too restrictive for this to be feasible: we'd be giving users a taste of something that won't work in a lot of their cases and it'd be better to not have the feature in that case. |
At least for variable-length options, clap has a setting to check for
subcommand names among them. This requires subcommands to know every
subsubcommand name and alias, without getting into subcommand short
flags and long flags.
So you're saying that if I have subcommand foo, clap would extract that
from the variable length args like:
```console
$ mycli -v --names bar foo
# Cli { verbose: true, names: [bar], subcommand: foo }
```
I think I understand the issue there. I'm not sure how to get around
that with multi-subcommands short of propagating the knowledge, as you
describe.
be very restrictive and disallow any case that requires
knowing about chained commands so we *can* just pop back up and
continue processing.
What would that restriction look like? You mean stop allowing putting
subcommands after variable length args? At least for myself, I'd rather
have a single layer of variable length subcommands than none at all, but
it does feel hacky to have that limitation.
Since we have no multi-subcommand support today, what about having a
config option to enable it that disables the "smart" detection of
subcommands after variable length args? That would be a way of opting
into the feature without breaking any existing CLIs in the wild. Not
ideal, but maybe a useful compromise? It would mean you had to avoid
variable length args before subcommands, put them in as delimited flags
using `Arg::value_delimiter`, or use a custom terminator with
`Arg::value_terminator`.
Given that we're wrestling with this in general, is it worth at least
pondering multi/positional flags as well? It's not clear to me if one is
a superset of the other problem, or if they're entirely orthogonal.
|
This isn't an acceptable option
|
> Since we have no multi-subcommand support today, what about having a
> config option to enable it that disables the "smart" detection of
> subcommands after variable length args?
This isn't an acceptable option
Okay, I think that's a helpful constraint. That means that we have no
choice but to figure out a way for subcommands to have the necessary
context to recognize sibling/parent options in addition to their own
options, correct?
That does make me lean towards considering positional flags as part of
this broader design effort, since that would also maybe be context that
subcommands would need?
…On 3/30/23 12:59, Ed Page wrote:
> Since we have no multi-subcommand support today, what about having a
config option to enable it that disables the "smart" detection of
subcommands after variable length args?
This isn't an acceptable option
- Being a runtime setting means we'd be bloating the code and API for everyone for this less common feature
- We try to limit what gets a feature flag because those don't have an ideal workflow for end users
- As mentioned in my previous post, consistency is important. You shouldn't be required to lose out on clap features to be able to use other features.
|
Something I realized is command chaining would get us a form of "argument chaining" for free because users could specify the start of an argument chain using |
I also have this use case. I am writing a tool for mass editing docx files. docx files are zip files with lots of XMLs and once unpacked and parsed, it is very desirable to perform multiple operations on the contents instead of spreading it out into multiple processes which would waste a lot of computation power. The CLI should look something like this: I would like to be able to use a |
I would like to also share my use case, where I need subcommand chaining with nested subcommands. I need to construct a complex structure from the command line, here is an extract: struct Transaction {
signer_account_id: String,
actions: Vec<Action>,
}
enum Action {
GrantAccess {
account_id: String,
permissions: Permissions,
},
Transfer {
receiver_account_id: String,
amount: u64,
},
}
enum Permissions {
FullAccess,
LimitedAccess { allowed_actions: Vec<AllowedActionKind> },
}
enum AllowedActionKind {
GrantAccess,
Transfer,
} Here is how I want the CLI to look like:
You may argue that nobody will write such a long command, and I totally agree, but in my case near-cli-rs has interactive mode which extends clap (there is interactive-clap helper), and guides users through the available options with prompts, and then prints the final command to be reused later. Currently, I copy-paste the code (see The pattern I use there is basically: struct ConstructTransaction {
#[clap(subcommand)]
next: AddActionOrSubmit_1,
} And then there are copies of: enum AddActionOrSubmit_1 {
AddAction(AddAction),
Submit(super::Submit),
}
struct AddAction {
#[clap(subcommand)]
action: Action,
}
enum Action {
GrantPermissions(GrantPermissionsAction),
Transfer(TransferAction),
}
...
struct TransferAction {
receiver_account_id: String,
amount: u64,
#[clap(subcommand)]
next: AddActionOrSubmit_2, // <- This is where "recursion" happens unless I manually unroll the code, and this is where it explodes when I add 5+ levels or unrolled copies
} |
Unsure which aspect doesn't scale for you but #4792 will help clap scale with a lot of subcommands with a lot of arguments. For v4 it might have to be opt-in for derive users but then on by default for v5. |
In search for this issue/feature, first thing I actually found searching was this thread, where a user has an issue identical to mine (and everyone else's in this thread...):
I reckon that's more modest in comparison to other use cases in this thread. It'd be plenty for my uses anyway. Using a workaround, they solved it as:
Could someone perhaps shed some light on this workaround: is it a viable one, while the present issue is in progress? Or is there a better approach? The |
I've actually independently resorted to the same workaround, the only difference being that I've used the word So the command would look like this:
|
Trying to re-summarize click APIOn the parent, set @click.group(chain=True)
def cli():
pass
@cli.command('sdist')
def sdist():
click.echo('sdist called')
@cli.command('bdist_wheel')
def bdist_wheel():
click.echo('bdist_wheel called') How it looks in helpParent command: TBD Chained command: TBD Behavior
|
Trying to re-summarize and polish the proposal APISetting it on the parent command is most likely the natural thing to do. Parent commands will need to know it for the help. When parsing child commands, we'll need to know its set on the parent. This can be handled in our parsing book keeping. Most likely this would end up looking like: impl Command {
pub fn chain_subcommands(self, yes: bool) -> Self
} Likely the easiest way to implement this in Most likely this would end up looking like: impl ArgMatches {
pub fn subcommands(&self) -> impl Iterator<Item=(&str,, &ArgMatches)>;
pub fn remove_subcommands(&mut self) -> impl Iterator<Item=(Str, ArgMatches)>;
pub fn subcommands_present(&self) -> bool;
} We'd likely keep the singular forms for ease of working with like we do for arguments. We could probably deprecate For derive users, I would expect this to translate to: #[derive(Parser)]
struct Cli {
#[command(subcommand)]
#[command(subcommand_required = true, args_required_else_help = true)]
subcommand: Vec<Command>,
}
enum Command {
New,
Build,
Clean,
} We'll consider the subcommand optional (zero-element We can key off of the use of How it looks in helpParent command: No different aside from the usage.
I'm thinking that if Chained command: Either
BehaviorConceptually, the easiest is: when evaluating a potential subcommand, chain up.
In most cases, nested subcommands under a chained subcommand could be confusing and we'd likely want to steer people away from doing that. However, using short/long flags for chained subcommands to create argument groups within chained subcommands would be a legitimate use case. However, this isn't the highest priority and if it isn't in the initial release but can be added later, we won't block this feature on it. Guarantees
|
Many years later, I'm coming back to this issue. After finally wanting to convert all my hacky old For the CLIs I'm working on, chaining commands is the thing I need. I fully understand, that there's many edge cases as mentioned in this thread. However, the CLIs I'm working on doesn't run into them. So now that I still can't use The ones I tested, that didn't support multiple adjacent commands are: However, use bpaf::Bpaf;
#[derive(Bpaf, Clone, Debug)]
#[bpaf(options)]
pub struct Options {
#[bpaf(external, many)]
cmd: Vec<Cmd>,
}
#[derive(Bpaf, Clone, Debug)]
pub enum Cmd {
#[bpaf(command, adjacent)]
Foo { a: bool },
#[bpaf(command, adjacent)]
Bar { b: bool },
#[bpaf(command, adjacent)]
Baz { c: bool },
}
fn main() {
let opt = options().run();
println!("{:#?}", opt);
} Which then when executing: [program] foo -a bar bar -b baz -c baz Outputs the following: Options {
cmd: [
Foo {
a: true,
},
Bar {
b: false,
},
Bar {
b: true,
},
Baz {
c: true,
},
Baz {
c: false,
},
],
} That being said, I only have mere minutes of experience using I'm not advocating the use of |
I'm not aware of |
@epage I'm confused why that's the take away. Anyways, I initially preferred |
Thanks for the clarification. That wasn't meant to be a "take away" but it was a leap that wasn't sufficient on its own with what was said so I was wanting to better understand. |
@epage if you're curious about the use case, I have various tools that operate on data or allow performing multiple actions sequentially. [program] [--dry-run] \
[cmd] [cmd-options...] \
[cmd] [cmd-options...] \
[cmd] [cmd-options...] Actual example for rendering 2 images: draw \
size 100 100 \
fill 0 0 50 50 "#FF0000" \
fill 50 0 50 50 "#00FF00" \
fill 0 50 50 50 "#0000FF" \
fill 50 50 50 50 "#FFFFFF" \
export --output "image.png" \
fill 25 25 50 50 "#FFFF00" \
export --output "image2.png" \ In reality, I could create config formats instead. However, many times I'm executing a single command, so having to create a config is tedious, e.g.: draw --input "image.png" --output "image-inverted.png" \
invert Additionally, both supporting a CLI and config doubles the code. So many times, I end up with a CLI and then create a |
I can second this exact use case. I'm considering a rust rewrite of vpype, which is very similar to @vallentin's example (albeit targeted at vector graphics for the plotter-based art niche). Example:
|
Previously all the params had to be unique between services because clap didn't yet support what we want: clap-rs/clap#2222 It still doesn't, but we can hack our way by parsing the args string manually, and passing different parts to different sub-parsers. $ cargo run -- --help Cabal Online Replacement Services Usage: cabal-mgr [OPTIONS] [-s|--service <SERVICE1> <SERVICE1_OPTIONS>] [-s|--service <SERVICE2...>] Services: crypto-mgr RockAndRoll equivalent event-mgr EventMgr equivalent proxy Options: -r, --resources-dir <RESOURCES_DIR> [default: .] -h, --help Print help (see a summary with '-h') -V, --version Print version Examples: ./cabal_mgr -s crypto-mgr --service event-mgr
Describe your use case
I would like to re-write an existing internal Python tool in rust. For this python tool we use click with the command chaining feature (https://click.palletsprojects.com/en/7.x/commands/#multi-command-chaining) this allows similar to some bigger CLI tools (e.g.
gradle
orsetup.py
) to execute multiple commands in one go. In our example we have a simple build and deploy tool and people often call it ether via./tool build start
or via./tool start logs
etc to trigger different functionality in one go. Wherebuild
would build our system;start
would startup the code locally andlogs
would tail the logs.Describe the solution you'd like
A simple switch that allows me to enable command chaining. I am ok with that prevents me from using sub commands and instead I can only use one level of commands. Examples of implementations are:
The text was updated successfully, but these errors were encountered: