Repository navigation
Group and restyle product API references; steady players on load - #267
Merged
Merged
Conversation
why: A workspace or MCP symbol page sat inside the article's prose styles, so its section headings computed to 32px against the core page's 18px, and under a field table it repeated every member as a full entry beneath a bordered Members bar: border on border, with the table's rows shown twice. what: - Add ApiMemberList, the linked member rows ApiReference drew inline, and draw them through it from both pages - List a product type's members as the table, then the rest as those rows; link each to its own page and keep every entry anchor as an empty target so existing links still land - Take the rows' type repeat flag from the core rule: C# properties and variants only - Add DocsLayout reference mode: the title and notice keep the prose column, the page body sits outside it; ProductApiPage uses it
why: The product page's members and heading size were only checked against the old stacked entries. what: - Add product-api-style.test.ts: reference mode keeps the body out of prose, both pages share ApiMemberList, members are rows with anchors - Update the product page tests in api-headings, enum-variants and field-table-pages to the table, rows and repeat rule of a core page - List ApiMemberList in the prose-link reader inventory
why: The MCP and workspace API sidebars listed every declaration in one A-Z list, mixing structs, functions, constants and enums, so a reader could not find a tool, a URI constant or an error type without scanning 83 rows. what: - Add product-api-sections.ts: one section vocabulary for MCP (Tools, Prompts, Resources and URIs, Views and results, Server and setup, Configuration and environment, Errors, Utilities) and workspace (Workspace files, Building and freezing, Command line, Plugins and shell, Errors, Utilities), assigned by ordered rules over kind, name, source file and the MCP tool catalog, with a small per-port override table - Feed the sections through the core TreeBucket shape to the sidebar tree, the documentation navigation and the reference index; products under eight roots stay flat; every URL is unchanged, A-Z kept inside a section - Test the assignment for rs, go, py, csharp and ts, determinism, no symbol lost or duplicated, and the tree and navigation counts - Nest the Julia extension-function namespaces inside their section
why: A landing page's Core API player was drawn at its last call and then jumped to the state before the first call when the script bound (about 300 ms after first paint), while the controls bar, hidden until then, pushed the page 49px down. Go's player also redrew its dots and count for one call fewer, since its short program plays 15 of the 16 recorded calls. The jump was visible as a flash of steps before the autoplay began. what: - Draw the player's frame and result at the state before the first call; a page with no script, and a reduced-motion reader's page until the script binds, still show the final frame and result (CSS swap) - Draw the controls bar in the markup, with the dots, fill and count for the first view's calls; the stylesheet hides it with no script - Pass the step the page is drawn at from the MCP replay, which keeps drawing its finished chat - Test that binding the script changes nothing on screen
…n test why: The core page's member rows now come from ApiMemberList, which this test's template loader stubbed as an empty component, so the Members list it reads rendered nothing. The core markup is unchanged. what: - List ApiMemberList among the real components the test renders
why: The MCP overview's worked example, the first thing on every MCP page, was drawn as the finished chat with every result open and then reset to the bare prompt when the script bound: the chat collapsed, the header appeared and each call's padding grew for the pin button, so text rewrapped. For a reader who asked for less motion the results also folded shut at bind, though the replay stays finished. what: - Draw the calls and frames out of view as data-off, not [hidden] (preflight hides [hidden] with !important), and draw the prompt alone with the empty tmux; the stylesheet shows the whole chat and the last frame where there is no script, and until bind for reduced motion - Draw the header from the first paint, hidden with no script, and give each call its pin button's room from the first paint - Keep a finished replay's results open at bind - Keep the controls bar's place but hide it until bind for a reader who asked for less motion, so its 0/N count and dots never show against a finished replay - Test that binding the script changes nothing on screen
…y the shape why: Product members are linked rows now, so the per-member disclosure of other ports is gone. check-dev found nothing to open on ClassicWorkspaceBuilder, and check-native-layout still asked product pages for a disclosure and a facts paragraph only their member entries drew. what: - Measure overflow in check-dev on the py Freezer.freeze and java WorkspaceBuilder.build pages, which carry their own In other ports table, and require that table to be present - Ask check-native-layout's py and java workspace pages for their own examples and description, and Ruby ConfigError for its bases
…tion why: The assembled-site suite read every MCP reference page's sidebar trail and required its top row to be "API reference". A product reference with eight or more roots is now grouped into sections, whose labels are the top-level rows, as the core references' groups are; the breadcrumb keeps the section after "API reference". Eight ports' MCP pages failed on that one assertion in CI (run 42440820229). what: - Accept a product section label as the tree's top row and keep it in the returned trail, so the breadcrumb comparison includes it - Keep requiring "API reference" for a flat reference
why: Shared pages such as every port's /mcp/ page linked to each port's same-product page even where the port has no page for it, giving broken links in the assembled check. The product match also missed the bare `mcp` and `workspace` routes. Guarding on productAvailable would drop an unpublished overview that does exist (Lua's MCP). what: - Add an 'excluded' product availability and productExcluded(), which productAvailable now also honours; 'unpublished' keeps its pages - Match the bare product route in pagePortLinks and drop a port's links only when it excludes the product - Test a synthetic excluded port, Lua's unpublished MCP overview and the bare route
why: A sectioned reference's tree row links to the page of its first entry, but the section group in the documentation navigation had no href, so the breadcrumb crumb for the section had none (assembled-site check "ancestor destination", run 42452517353). what: - Give each section group the href of the first entry the tree row uses, computed with the same firstEntry rule - Match a group's own rows before its link in sidebarBreadcrumbs, so a page that is both a section target and a row keeps its row crumb - Test every grouped product: section href equals the tree row target
why: The excluded-product fix widened the product match to the bare `workspace` and `mcp` routes for every use. The bare `workspace` page then required the port's workspace to be available before linking its overview, so Lua's and Elixir's unpublished workspace overviews lost their links, and a bare product page's home link moved to the product. The Perl port's review found it in the source. what: - Match the bare product root only for the exclusion check; the product used for eligibility and the home link keeps its prior, slash-only match - Test the shared workspace overview: Lua's and Elixir's unpublished roots keep their link, an excluded port's is dropped; the test fails with the widened match
This was referenced Nov 8, 2026
why: Product API pages now list a type's members as rows linking to each member's own page, instead of repeating every member's full entry under the type. The mentions inside those repeated entries were counted a second time. CI run 42472628116 counted the drop against main's push run 42421548678 in exactly the ports with an MCP or Workspace Manager reference: py -446, ruby -131, powershell -545, ts -12, rs -1028, go -94, java -64, csharp -633, cxx -115, swift -451; lua, elixir, kotlin, scala and fsharp, which have none, are unchanged, and julia rose by 159. Each member's mentions still resolve on its own page. PowerShell's floor (4752) sat above its new count (4663); every other port stays above its floor. what: - Lower the powershell floor in scripts/xref-floor.json to 4663; the other floors are unchanged
why: Product API pages now list a type's members as rows that show each member's type through ApiType, the same renderer the entries used. CI run 42481976006 counted 110 Julia type names rendered plain against a ceiling of 108; the most frequent are Base names that stay plain by design (Exception, show, validated, N, var). Every other port stayed at or under its ceiling, and Julia's resolved cross-references rose by 159 in the same change, so the two extra plain names come from the new rows rather than from a resolution regression. what: - Raise the julia ceiling in scripts/type-links-ceiling.json from 108 to 110; the other ceilings are unchanged
7 of 9 tasks
tony
temporarily deployed
to
docs-preview-cleanup
November 8, 2026 19:05 — with
GitHub Actions
Inactive
This branch was previously deployed
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
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.
Three reader-facing fixes from review of the live site after the wave at c5b56f2.
Summary
Product reference pages in the core style
https://libtmux.org/en/rs/latest/workspace/reference/src-builderror-refused/ drew "Fields" and "Member details" at 32px, where the same heading on a core page such as https://libtmux.org/en/rs/latest/reference/options-optionvaluerefusal/ is 18px, and repeated each field as a full bordered entry under a bordered bar.
ProductApiPagerendered insideDocsLayout's<article class="prose">; Tailwind Typography's.prose h2 { font-size: 2em }outranks the heading'stext-lg, andnot-proseon the wrapper does not stop it.DocsLayouttakes areferenceprop that keeps the title and product notice in the prose column and puts the body outside.prose.ApiMemberList.astroholds the compact member rows, moved out ofApiReference, so core and product pages list members the same way: the field or variant table, then the other members as linked rows. Each member keeps its old entry id, and the section and parameter ids its entry carried, as an anchor target.Sections in the product API references
https://libtmux.org/en/rs/latest/mcp/reference/mcp-views-branch/ listed 83 symbols A to Z in one bucket.
site/src/lib/product-api-sections.tsassigns each symbol of a product with eight or more public roots to a section, by ordered rules (kind, name pattern, module, the MCP tool catalog) with a few per-port overrides, and feeds the sidebar tree, the documentation sidebar's API group and the reference index. It reuses the core references'TreeBucketshape; URLs are unchanged.Players draw their first frame on the server
On a cold load of https://libtmux.org/en/ruby/latest/ the server HTML drew the Core player at its last call (16/16) and the script reset it to step 0 about 0.3 s later, while the controls bar appeared and pushed the player from 627 to 676px. The MCP replay did the same: the whole chat with every result open, then the bare prompt (531 to 571px).
CorePlayer.astroandMcpReplay.astrodraw the step the script first shows;ReplayControls.astrois in the markup from the first paint.core-player.cssandmcp-replay.cssshow the finished frame and chat to readers without the script, and to reduced-motion readers until the script binds; out-of-view calls usedata-offinstead ofhidden, which preflight hides with!important.Test plan
product-api-style,product-api-sections, and the new "binding the script changes nothing on screen" cases incore-playerandmcp-replay: each fails on main and passes hereastro check: 0 errorstestworkflow, including the assembled-site suites (product-docs.test.ts,check-type-links)