Skip to content

Specification · RFC

RFC-0011: Public Product And Documentation Website

Version: 0.4.0 | Status: normative | Phase: stable Owners: @Zihua Wu


[RFC-0011:C-SUMMARY] Summary (Informative)

Section titled “[RFC-0011:C-SUMMARY] Summary (Informative) ”

InferLab publishes one public website from its filtered public-repository snapshot. The website combines a concise product introduction with operator documentation and the current architecture and specification reference.

Scope: This RFC covers the public site’s information architecture, content authority, static publication, search, and automatic delivery from the public repository to its default GitHub Pages project URL. It does not define a hosted InferLab service, browser-based experiment control, telemetry collection, historical documentation versions, a custom domain, internal execution-history publication, or automation of the operator-controlled repository publication act.

Rationale: The repository already contains the product contract and operating guidance, but a public entry point must explain the workflow without creating a second authority for facts already owned by operator documentation, support matrices, RFCs, and ADRs.

Since: v0.1.0


[RFC-0011:C-PUBLIC-SITE] Product And Documentation Experience (Normative)

Section titled “[RFC-0011:C-PUBLIC-SITE] Product And Documentation Experience (Normative) ”

The public website MUST present InferLab’s product positioning and concrete operator workflow separately from the detailed documentation reference.

The human-facing product name MUST use the exact spelling InferLab in site-owned interface text and authoritative public prose projected by the site. Machine-facing commands, package distribution names, local storage paths, and protocol or evidence literals whose exact spelling is independently specified MUST retain that spelling. Public repository names and their derived default project-site URLs MUST use the repository’s current canonical spelling.

The product experience MUST provide direct paths to installation, documentation, the source repository, and published releases.

The documentation experience MUST organize public content into getting started, concepts, guides, reference, and architecture and specification areas.

The architecture and specification area MUST expose the current rendered RFC and ADR corpus present in the public snapshot.

Site search MUST cover the public documentation corpus, including RFC and ADR pages.

Every navigation, search, and other interactive product control MUST be keyboard focusable and operable without a pointing device.

When the documentation experience offers light and dark theme states, selecting either state MUST apply one coherent color palette to navigation, sidebars, controls, and primary content. A theme change MUST NOT leave primary content using the opposite theme’s page background or text palette.

At every viewport width of at least 320 CSS pixels, navigation and primary product content MUST remain available without page-level horizontal overflow.

The website MUST remain view-only and MUST NOT start, stop, mutate, or upload InferLab workspace state.

Rationale: A product entry point and a technical reference serve different reading intents. Keeping both in one searchable, view-only site makes the product understandable without implying a hosted control plane or weakening the file-first workflow defined by RFC-0002, RFC-0003, and RFC-0005. One display spelling prevents the site shell and its projected authorities from presenting competing brand identities while independently specified machine interfaces remain stable. Following the repository’s canonical spelling keeps default GitHub Pages routing valid after a repository rename. Coherent theme states keep the same documentation content readable without allowing brand overrides to defeat the selected presentation. The explicit interaction and viewport bounds make accessibility and responsive behavior independently verifiable.

Since: v0.1.0

[RFC-0011:C-CONTENT-AUTHORITY] Published Content Authority (Normative)

Section titled “[RFC-0011:C-CONTENT-AUTHORITY] Published Content Authority (Normative) ”

The public website MUST consume content from the repository’s filtered public snapshot. It MUST NOT consume source artifacts outside the public projection or internal execution-history renders.

The public snapshot’s rendered RFC and ADR files MUST remain authoritative for normative behavior and architecture decisions. RFC pages on the site MUST use the current-state projection whose private changelog sections were removed before the snapshot entered the public repository.

The backend support matrix at plugins/inferlab/skills/inferlab/references/backend-support.md MUST remain authoritative for publicly qualified backend capabilities and the version-correspondence table. It lives in the agent skill’s reference directory so that an installed plugin carries it beside the skill.

The workspace-authoring reference set under plugins/inferlab/skills/inferlab/references/ consists of workspace-authoring.md, workspace-definition.md, execution-authoring.md, eval-authoring.md, and bench-authoring.md. This set MUST remain the single editable authority for workspace-authoring operator guidance. The workspace-authoring.md index MUST link to each of the other four members and identify their scopes as follows: workspace definitions, placement, upgrades, and validation belong to workspace-definition.md; profiling, runtime images, and invocation patches belong to execution-authoring.md; Eval tasks, datasets, and inference requests belong to eval-authoring.md; and serving-Bench load, sources, sessions, metrics, and SLOs belong to bench-authoring.md.

RFC-0001 and its topic RFCs MUST remain authoritative for command, workspace, support, record, and architecture behavior. README.md, the workspace-authoring reference set, the backend support matrix, and docs/tui.md MUST remain authoritative only for their installation entry points and operator guidance text; they MUST NOT override RFC-owned behavior semantics.

The website documentation projection MUST be limited to README.md, every member of the workspace-authoring reference set, the backend support matrix, docs/tui.md, docs/rfc/*.md, and docs/adr/*.md from the public snapshot. The website MUST project every workspace-authoring reference member as its own page and MUST NOT maintain an independently editable concatenated workspace-authoring copy. Each projected source page MUST preserve the authoritative prose, code, tables, link labels, and every heading other than the leading level-one heading. The projection MAY represent that leading level-one heading as page-title metadata and omit it from the projected body only when the rendered page exposes the same heading exactly once as its semantic level-one page title. The projection MAY add presentation metadata and route placement, including mechanical remapping of relative link destinations to the corresponding website route or public repository source.

For every documentation page, the site MUST also publish a plain-text Markdown alternate at the page route with a .md suffix, carrying the page title and the same body the page renders. The site MUST publish llms.txt at its root, an index following the llms.txt convention that links every documentation page’s Markdown alternate, and llms-full.txt, the concatenation of those alternates. These text outputs MUST be generated from the projected sources during the build and MUST NOT be independently editable; they are not a concatenated copy in the sense prohibited above.

The site build MAY add navigation, visual components, assets, and site-owned product summaries. Those additions MUST NOT redefine command behavior, support status, record semantics, or architecture facts, and MUST NOT establish a competing product contract.

The site MUST represent only the current public documentation corpus. “Current” means the latest successfully deployed public main snapshot that has not been superseded by a newer successful main deployment. The site MUST NOT expose per-release documentation snapshots or a version selector.

Internal execution-history records MUST NOT be published as product documentation.

The website source and generated output MUST NOT contain credentials, private downstream revisions, machine-local paths, model-weight locators, private host identities, or unpublished backend qualification evidence.

Rationale: The website is a presentation surface rather than a content authority. A closed source set makes drift detectable; consuming the already-filtered public snapshot avoids exposing private governance state and revision history; and the authority split keeps operator prose from competing with the normative specification. Sharing task-focused authoring sources across distribution and presentation removes a monolithic duplicate without weakening either surface. Reusing each source H1 as its page title preserves its content and semantic role without presenting it twice. Mechanical link remapping preserves navigability after those sources move into the website information architecture without creating editable content copies. RFC-0008:C-AGENT-PLUGIN owns how the agent plugin carries this reference set.

Since: v0.1.0

[RFC-0011:C-DELIVERY] Static Publication And Validation (Normative)

Section titled “[RFC-0011:C-DELIVERY] Static Publication And Validation (Normative) ”

The website MUST build as static files without a required application server or runtime data service.

Every public-repository pull request that changes website inputs or website implementation MUST validate a locked production build before merge.

Private governance validation, rendering, filtering, scanning, operator review, and the operator-initiated public snapshot push MUST remain outside the automatic website deployment flow. Automatic deployment MUST NOT perform or authorize the private-to-public snapshot act.

After an operator pushes a filtered snapshot to the public repository’s main branch, the public repository MUST publish the site automatically only after its required Rust, Python, and website validation gates succeed for that public revision.

The deployed static artifact MUST be built from the same public repository revision that satisfied those gates.

A deployment for an older public main revision MUST NOT replace a successfully deployed newer main revision.

Publication MUST use repository-scoped ephemeral credentials and MUST NOT require an operator-maintained long-lived deployment credential.

Public deployment MUST operate at the repository’s default GitHub Pages project path, currently /InferLab/, without requiring a custom domain. In the public repository’s GitHub Actions environment, production generation MUST derive the GitHub Pages project path from the repository-name component of GITHUB_REPOSITORY. Outside that environment, production-equivalent local generation MUST use an explicit fallback repository name with the same canonical spelling.

For generated page-navigation and canonical URLs beneath /InferLab/, the URL path component MUST end in / before any query or fragment, such as /InferLab/docs/; asset URLs are excluded from this page-URL rule. Local development route matching MUST accept the equivalent page route with or without its trailing slash.

Rationale: Static output matches the site’s view-only role and minimizes operating burden. Separating the manual snapshot act from its automatic public-repository consequence preserves operator authority, while revision gating and monotonic deployment prevent the site from advertising an unvalidated or superseded public state. Deriving the case-sensitive project path from the canonical repository identity prevents repository renames from leaving generated assets at a stale path, while an explicit local fallback keeps production-equivalent validation available outside GitHub Actions. A canonical URL form prevents duplicate published identities, while forgiving local matching avoids a development-only 404 that the static host does not use as the canonicalization authority.

Since: v0.1.0