Docs: correct the claims, fix the rendering, add a Protocol section - #1
Merged
Merged
Conversation
The site documented a remote viewer across seven pages - a uvicorn app named vlarl_viewer, a plugrl-monitor component, and --use-remote-viewer/--viewer-host /--viewer-port flags. None of it exists: there is no viewer repository, and the env client CLI has no such flags. Readers following these instructions could only fail. Also renames the remaining "worker" wording to "env client" to match the package that ships today. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
mkdocs.yml declared no markdown_extensions at all, so every fenced block on a site made almost entirely of shell and Python rendered as bare <pre><code> - no highlighting, and nothing for content.code.copy to attach a button to. theme.features listed "navigation.instance", which mkdocs-material does not have, so it was silently ignored. Correcting it to navigation.instant turns out to break the build: mkdocs-static-i18n cannot keep the language switcher contextual with instant loading on, and --strict fails. The typo was accidentally load-bearing. Left out, with a comment saying why. Also adds site_description, repo_url and edit_uri, so pages get an "Edit this page" link. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
deploy-pages.yml only runs on a push to main, so a pull request got no validation at all and a broken build was first visible after merging. --strict makes a warning - a dead link, a page missing from the nav - fail the check rather than ship. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
The boundary between the training server and the env client is the reason this project exists, and the site never described it. A reader could learn how to add an environment or a policy, but not what the two processes say to each other, or that an env client need not be Python. The page orients rather than duplicates: SPEC.md stays in plugrl-protocol, next to the code it describes, so the two cannot drift. What is here is the exchange diagram, the three rules a first implementation usually gets wrong - strict alternation, feedback env sets that need not match the infer's, and chunk-summed reward - the conformance server, and the relationship to openpi. Both languages. mkdocs build --strict passes; nav translations go from 10 elements to 11. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
tactino
changed the title
The quickstart was the dummy policy, whose learn is a sleep. A reader following the front page of this site could confirm two processes talk to each other and nothing more - which is a connectivity check, not a quickstart, and the page did not say so. It is now FPO on HalfCheetah-v5, CPU only, which trains: episode return climbs out of the -300s in a few minutes. The connectivity check is kept below it, labelled as what it is. It also states the setting that would otherwise cost someone an afternoon. FPO learns when its rollout buffer fills or when the run ends, so at the default buffer_size=983040 a run shorter than a million steps learns exactly once, at the very end, and produces a single point rather than a curve. And it stops recommending the Ray launcher flatly. That line said only "Use plugrl-run-server-ray for Ray-based distributed launch"; the launcher needs the dppo extra, builds its worker list from the local GPU count so a multi-node cluster still sees one node, and its server speaks an older dialect of the protocol. Those are now stated, with a pointer to the spec. Both languages. mkdocs build --strict passes. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
It had no installation section at all. It opened with `plugrl-run-server`, which requires two packages that are not on PyPI and that the page never mentioned cloning. A reader arriving from the front page had nowhere to go. It now begins with the clone-and-uv-sync for both repositories, then the FPO quickstart that actually learns, then the dummy connectivity check labelled as what it is. It also names the two flags that are not optional and previously were not mentioned anywhere: --policy.device cpu, because the default is cuda and the server dies on startup without a GPU, and --algo.buffer-size, because at the default a short run learns once at the very end. The troubleshooting section now covers what a new user actually hits, including both of those. The Ray launcher is qualified here and in the user guide rather than listed as an equal alternative to plugrl-run-server. Both languages. mkdocs build --strict passes. Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
tactino
marked this pull request as ready for review
September 11, 2026 15:16
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Four changes to the documentation site: it described software that does not
exist, it had never had syntax highlighting, its build was unchecked, and it
never described the thing this project is built around.
It referenced software that does not exist
Removed. A reader following those instructions would have got nowhere, and
an artifact reviewer reads the docs before the code.
Fenced code blocks were rendering as bare
<pre><code>mkdocs.ymlhad nomarkdown_extensionssection at all — so nohighlighting, and
content.code.copyhad nothing to attach a button to, on asite that is almost entirely code samples. Added
admonition,attr_list,tocwith permalinks,pymdownx.highlight,inlinehilite,superfencesand
details.One thing deliberately left alone. The theme feature list contains
navigation.instance, which looks like a typo fornavigation.instant.Correcting it breaks
--strict: mkdocs-static-i18n cannot keep the languageswitcher contextual with instant navigation on. The typo is load-bearing, so
it stays, with a comment saying why.
The build was not checked
A
--strictbuild now runs on pull requests. Without it, a broken nav entryor a dead internal link reaches the published site.
New: a Protocol section
The boundary between the training server and the env client is the reason
this project exists, and the site never described it. A reader could learn
how to add an environment or a policy, but not what the two processes say to
each other — or that an env client need not be Python.
The page orients rather than duplicates.
SPEC.mdstays inplugrl-protocol, next to the code it describes, so the two cannot drift.What is here is the exchange diagram, the three rules a first implementation
usually gets wrong — strict alternation, feedback env sets that need not
match the infer's, chunk-summed reward — the conformance server, and the
relationship to openpi.
Both languages, as every other page.
mkdocs build --strictpasses; navtranslations go from 10 elements to 11.
🤖 Generated with Claude Code