.superbee
Bundle OKF 0.2 · 0 conceitos · Holaxis-ai/superbee-docs
Open source Repository Open in the app JSON README (API)
About
<!-- agentstate-lite:generated-index:v1 -->
# .superbee
An Open Knowledge Format bundle.
Details
- Kind
- OKF bundles
- Topic
- No topic detected
- Publisher
- holaxis-ai
- Origin
- okf_github
- Category
- dados
- Version
- 0.2
- Last push
- 2026-09-07T13:50:04Z
- Repository state
- ativo
- Language
- JavaScript
- License
- Apache-2.0
- Added
- 2026-09-08 02:23:09
- Updated
- 2026-09-08 02:23:09
- Origin id
Holaxis-ai/superbee-docs:.superbee/index.md
README
# Superbee documentation This repository contains Superbee's public, source-grounded documentation bundle and the minimal tooling that publishes it as a human site. ```text verified product evidence -> public .superbee bundle -> one superbee/publication snapshot -> codebase-documentation recipe + bundle-native publication model -> one Recipe Studio-compiled @superbee/docs-projection projection -> Portal site at docs.getsuperbee.com + conventional MkDocs reference output ``` The repository is public by design. Private planning, credentials, security work, and unpublished communications belong elsewhere. ## Documentation selection The installed `codebase-documentation` recipe defines Documentation System, Documentation Publication, Documentation Section, and Documentation Trigger Kinds. Ordinary documents under `documentation-systems/`, `documentation-publications/`, and `documentation-sections/` own the renderer-neutral product identity and selection. Sections name ordered primary pages; the publication names additional public pages that both outputs include outside navigation. Adapters must not follow links to expand that explicit selection. `portal.config.json` now carries only the publication identity and target/build overlays such as brand assets, agent guidance, diagrams, indexing, Views, routes, and deployment settings. Recipe Studio compiles the installed recipe and the linked bundle records into one projection. It rejects missing pages, duplicate selection, navigation/support overlap, invalid operational exposure, or recipe drift before either output is built. Both outputs publish a bounded `llms.txt` over that exact selection. Documentation pages advertise their byte-exact Markdown source and the root discovery index; operational maintenance records stay inspectable in the complete public bundle without entering these curated agent surfaces. The optional `documentation.guidance` binding carries the document id, heading, and label of one section of one already-selected page that `llms.txt` quotes as its "when to use" guidance. Only the pointer lives in configuration. The quoted bytes and every link inside them come from that published document, so the entry point can never become a second content authority, and the build fails closed if the heading moves, is duplicated, gains a nested heading or code fence, or links to something outside the selection. It also refuses any link or markup shape the renderer cannot resolve into an absolute published URL -- a titled or angle-bracketed inline target, a reference-style link or its definition, an autolink, a raw HTML element or comment, or an indented code block -- because a page-relative link copied into `llms.txt` resolves against `/llms.txt` instead of the page it came from. ## Agent-facing responses An unknown documentation URL returns a real 404 carrying `404.html`, whose Markdown twin `404.md` holds the same recovery facts. Both link to the documentation index, `llms.txt`, and the sitemap with absolute URLs, because a recovery response is served from whatever path the reader requested. Cloudflare selects those bytes through the `not_found_handling: "404-page"` asset setting. The Markdown access contract is explicit URLs, not `Accept:` content negotiation. The reviewed evidence in `research/agent-docs-discovery-surfaces` ruled negotiation a layer violation for this stack because it would make the edge a second renderer over digest-bound bytes, so the origin serves the `.md` sibling every page advertises and deliberately emits no `Vary: Accept`. Advertising a negotiation the origin does not perform would fragment every shared cache in front of it while changing nothing an agent receives. Portal's production verifier asserts that absence. Documentation pages carry canonical, Open Graph, Twitter card, and product JSON-LD built only from the projection's own product facts and this site's URL. No postal address, contact, social image, or legal identity is emitted here; those facts belong to the public marketing site, which owns them. ## Development Node.js 22.12 or newer is required. ```bash npm ci npm run source:sync npm run portal:build npm run mkdocs:sync npm run check ``` `npm ci` installs the exact Portal and Superbee packages recorded by the lockfile. The source sync keeps an exact Superbee checkout only for source-grounded architecture checks. The Portal build captures one source snapshot and compiles the same explicit 52-document selection (40 navigated and 12 supporting), brand asset, relationships, and eight admitted diagrams into one projection consumed by both Portal and MkDocs. The `@superbee/docs-mkdocs` package owns the pinned uv command sequence: `mkdocs:sync` installs the exact locked Python environment once, while `mkdocs:build` and repository checks are frozen and offline. Before changing the source pin or package versions, confirm that no `*.apply-intent.json` transaction journal is pending; finish or roll back that operation with the package version that created it first. Journals contain local paths and exact recovery preimages and are intentionally ignored by Git. The adoption is one atomic repository change. If the private compiler or bundle-native model must be rolled back, revert the adoption commit and run `npm ci`; that restores the prior docs-site/v2 configuration and local adapters together. Do not delete only the installed conventions or model documents while docs-site/v3 remains active, because compilation intentionally fails closed on a missing or drifted recipe. ## Release documentation Navigation and maintained pages use the stable `releases/current` and `sources/current-release` bundle identities. The reader-facing `releases/release-notes` page lists the current release and immutable prior releases, while migration guidance stays beside it in navigation. The daily release-freshness workflow compares the documented release with the public npm `latest` package, GitHub release, and exact Git tag. Run the same deterministic probe locally at any time: ```bash npm run docs:release:status ``` When it reports `update_required`, the output carries the exact package, integrity, source commit, publication date, release URL, and generated GitHub notes. It also names the reader-facing fields that still require agent judgment and the documentation-impact events to query. A verified release then updates the stable identities, creates immutable versioned records, reconciles the release archive and selection, and updates the exact site version with one idempotent command: ```bash cp examples/release-input.example.json /tmp/superbee-release.json # Fill the handoff with facts and verification from the completed release. npm run docs:release -- --manifest /tmp/superbee-release.json npm run docs:release:check ``` The JSON file is ephemeral release-process input, not another persisted documentation authority. The command writes release documents through Superbee, refuses to alter existing version history, and updates the bundle-native Documentation System and Publication records when needed. It converges to a no-op when retried and never invents the summary, changes, user action, compatibility, recovery, supported platforms, or verification performed. Review those fields against the release and query the returned impact events before opening the documentation pull request. The repository check rejects a package version pinned in ordinary pages; exact versions remain available in release, evidence, and migration records. ## Deployment Portal builds the immutable public site artifact into `dist`. The public `@superbee/portal-cloudflare/static-assets` adapter then verifies that complete artifact and atomically assembles `deploy`: every inventoried byte plus only the host configuration Cloudflare consumes and does not serve. The two directories stay separate because `dist` is inventory-exact. Portal must be able to reject any extra or changed byte in it. The package derives `_headers` from the artifact's declared response policy and exact media types. It also derives every canonical 307 HTML alias from the artifact's hosting requirements. This repository contributes only two site-specific rules: `/docs` and `/docs/` return 301 redirects to the canonical root. Each names one exact path, so every other missing route still reaches the published recovery body as a real 404. `scripts/deployment-assets.mjs` is the small consumer binding for that policy. It does not duplicate Cloudflare limits, MIME tables, integrity checks, or filesystem replacement logic. `scripts/cloudflare-worker.mjs` is similarly small. It exports the package's verified public Portal handler under a stable Wrangler entry path. Portal owns byte validation, route admission, recovery, opaque downloads, View isolation and bridge behavior, and the live deployment-effect identity. Wrangler keeps presentation routes asset-first so generated headers and canonical redirects apply, then sends `/data/*`, `/bundle/*`, and the View bridge to the Worker for verified dynamic policy. The MkDocs adapter independently materializes a conventional reference site under ignored `.tmp/mkdocs/site` from the same owned projection; it is validated but not deployed by this repository. Validate the exact artifact, Cloudflare assembly, and staged runtime locally: ```bash npm run portal:build npm run cloudflare:check npm run cloudflare:reconciliation:check ``` After a deployment, compare the live origin against the exact artifact this repository built: ```bash npm run verify:production -- --base https://docs.getsuperbee.com --dist dist ``` The command is a thin consumer of Portal's host-neutral verifier. Portal checks exact status, bytes, media types, response headers, audience policy, fallback behavior, content negotiation, canonical HTML routes, and their exact 307 aliases from the artifact's hosting requirements. This repository adds only its page metadata, advertised Markdown-alternate, and two guessed `/docs` entry-route assertions to the same receipt. No static-byte or canonical-redirect capability is waived. Production activation belongs to `.github/workflows/verify-production.yml`, not to Cloudflare Git builds or a bare Wrangler command. Its uncredentialed build job checks out the exact `main` commit, expands repository history, runs the deterministic build and staging checks, and uploads only the completed `dist` artifact. The reconciliation job downloads those exact bytes, assembles the package-owned Cloudflare layer without rebuilding the site, and re-resolves `origin/main` immediately before activation. A changed desired commit fails closed. Only the reconciliation step receives `CLOUDFLARE_API_TOKEN`. It calls the package's public reconciler through `scripts/reconcile-cloudflare.mjs` with an immutable source, site, and toolchain provenance tuple. The reconciler inspects provider generation, stages and digests the complete activation unit, activates with strict generation protection, then externally verifies the live effect. If a new activation fails verification and the prior generation was verified, the package rolls back and verifies recovery. The workflow concurrency group never cancels an in-flight run, so one production target has one serialized activation stream. The workflow then runs the Docs-specific verifier over every expected status, byte digest, media type, response header, redirect, fallback, and admitted View. An `if: always()` step uploads both the reconciliation receipt and the Docs verification receipt for 30 days, including bounded preflight failures and verified rollback outcomes. Its daily scheduled run uses probe mode and does not mutate production, so deployment drift is detected against the current exact artifact. Cloudflare's repository integration must remain disabled while this workflow is the production writer. Enabling both would create two independent activation authorities. The `npm run cloudflare:check` command remains a credential-free dry run, not deployment agreement, and this repository intentionally exposes no `cloudflare:deploy` or `cloudflare:preview` script. `main` is the production branch. ## Diagrams Mermaid does not execute inside Superbee Markdown. Diagram source is compiled into a source-bound, admitted static SVG whose exact committed bytes both documentation outputs consume: ```bash npm run diagram:build # apply source -> public bundle static SVG + v3 receipt npm run diagram:check # prove checked-in source/SVG/receipt agreement npm run portal:preview # inspect the exact Portal artifact locally ``` The exact `@superbee/docs-tooling` package owns both reusable compilation and the explicit, namespace-bounded `superbee-docs diagram apply` publication conductor; this repository owns its source, manifest, configuration, and receipt. Diagram source uses directive-free flowchart syntax without font overrides and printable ASCII in v1. Browser layout geometry may differ slightly across operating systems, so checks re-render for safety and accessibility while source-bound projection metadata detects drift; both output adapters separately pin the exact committed static SVG bytes. The old View HTML and registrations remain byte-for-byte predecessor evidence; they are not maintained publication output. If `diagram:build` reports an interrupted apply, run `npm run diagram:rollback` before retrying. That repository command supplies the required root and config arguments. Apply never removes stale publications merely because the local receipt no longer lists them; pruning is a separate, explicitly authorized lifecycle operation.