Add static hosted HTML page with PR builder forms

[bollwyvl]

Comment:

elevator pitch

Provide a low-barrier way to make precise, pre-validated admin requests.

motivation

After looking at the GH PR templates feature (suggested in #535), I was unsatisfied with the specificity of the language (as usual, not quite JSON schema),

design ideas

So I wrote a thing that:

• takes in JSON schema
• builds a form with rjsf
• makes a yaml/json/toml of the form
• makes a link for new PR based on the input, which can include exactly on file

Here's a demo for an outrageously long schema:

https://deathbeds.github.io/jupyak/shaver.html

challenges

The downside: to get the nice UI (dropdown/autocomplete), all the feedstock names would need to be embedded in the schema, e.g.

"feedstocks": {
  "type": "array",
  "items": {
     "type": "string",
     "enum": ["aalto-boss", "a-few-others", "zziplib"]
  }
}

But this might be something that could be generated in one place...

{
  "$id": "https://conda-forge.org/schema/feedstocks.schema.json",
  "type": "string",
  "enum": ["aalto-boss", "a-few-others", "zziplib"]
}

And then referenced here:

{
  "feedstocks": {
    "type": "array",
    "items": {
      "$ref": "https://conda-forge.org/schema/feedstocks.schema.json"
    }
  }
}

implementation ideas

After the... experience... with pydantic over on conda-smithy, it seems like schema-first design (but perhaps authored in YAML) to get to a well-typed TypedDict might be easier and give strictly better validation.

alternatives

• use the semi-decent Issue form to generate PRs
  • the specification language is also not-quite-schema, and not portable to anything else
  • an action would have to re-parse the generated markdown
  • the generated PR is not editable by the original owner with more complexity
• use a bot pidgin grammar
  • these are harder to discover, and don't have autocomplete (other than the bot name)

[jaimergp]

If I understood correctly:

• Define JSON schema(s) for the input files we support here
• Build a static JS/HTML page that ingests the aforementioned schema(s)
• Publish it and let folks use the web UI to create PRs (with no creds needed) via Github UI (because it's just adding a single file per PR, correct?).

If this works, it could also be a nice workflow for conda-forge-repodata-patches-feedstock (assuming we can start with existing files).

Also, this doesn't need any changes in conda-forge.github.io either, right?

PS: We do have a single JSON with all packages and feedstocks at https://github.com/conda-forge/feedstock-outputs/tree/single-file. Would that be a good starting point for the autocompletion schema?

[jaimergp]

Alternatively, provided we do generate the schemas here (so they are close to the code that ingests them), we can also add a couple pages to the website so everything "JavaScript-y" is defined in the same repo. And if we go down that route for the repodata patches repo, I'd assume there are several bits of code that would be shared.

[bollwyvl]

Well, the nice thing about JSON schema is it's not just JavaScript-y (or GitHub-y, pydantic-y, or whatever) and pretty much gives the same results in all the implementations.

But yes, the workflow would be entirely offline, static, running in the browser, but supported by all the engineering rigour we can throw at the process.

The "end product" is using pre-existing HTML URL features, so the user would be prompted in a non-scary way to do a standard GitHub login when they click on the link to propose the file, which then immediately suggests the PR workflow.

repodata patches repo

This would be possible as well, but that generated schema is gnaaarly, and would probably take some work to clean it up for human consumption. Both pydantic and msgspec support JSON Schema, but only the parts they want to.

Of note: that repo is slooooow to work with, and might benefit from a performance-focused implementation. Indeed, going from pydantic to msgspec would probably cut the runtime of that repo in half, if not more.

https://github.com/conda-forge/feedstock-outputs/blob/single-file/

At rest, no, that is just... some unlabeled JSON, which would require special JavaScript-y parsing to deal with effectively. If it was instead stored as a reusable, self-describing schema fragment (as described), and hosted at a generally-resolvable URL it could help in a number of places.

Of course, this falls down when one wants to use the values from the enum as the keys of a dict... schema only offers pattern, which would be a mighty regex indeed:

"$defs": {
  "a-valid-conda-forge-feedstock-name": {
     "type": "string",
     "pattern": "^(aalto-boss|a-few-more|zziplib)$"
  }
}

As for the outputs, I would probably make that a whole separate enum schema.