Playground tour

An interactive canvas for seeing the cascade engine work. Sources are nodes; each concept resolves live through resolver.mjs, and where layers disagree the conflict is surfaced on the section — never dropped.

Start it

npm run playground          # serves http://127.0.0.1:8790
# or point it at a different manifest / port:
node apps/playground/server.mjs --manifest layers.json --port 8790

Then open http://127.0.0.1:8790.

The playground server is dependency-free, same as the rest of the engine — no npm install, no CDN. It’s a thin shell over the real engine: GET /api/status is the cheap poll (per-source index progress plus a generation counter, no resolve and no tokenizing), GET /api/graph returns the source topology and concept index, GET /api/resolve?concept=<id> returns exactly what the CLI and the MCP server would. The browser is just another reader of that output — nothing is reimplemented.

Sources and context budget

The Sources tab is the configuration and accounting surface. Adding or removing a source edits apps/playground/manifest.json in place.

Add sources of three kinds:

• Local path — a directory of OKF markdown (relative to the manifest, or absolute).
• GitHub repo — cloned with git (uses your credentials, so private repos work). Give owner/name, an https URL, or an SSH URL, plus an optional branch or sub-directory. The repo may or may not contain OKF — non-OKF markdown just yields fewer concepts. Sync does a git pull. Clones live in apps/playground/.cache/ (gitignored). Only the owner/name, https, and SSH forms are accepted; other git transports are rejected.
• MCP server — a command + args that speak MCP; their graph is translated to OKF at read time. See Foreign MCP sources for the mechanics.

Caution

Adding an MCP source spawns command with args. Only add servers you trust — see the trust boundary.

Token accounting — every source and concept is tokenized with a vendored o200k_base BPE tokenizer (GPT-4o’s; a close proxy for Claude, which isn’t public). The budget bar shows each source’s share of the total context, and raw → effective shows how much the cascade dedups (for example, 772 raw tokens down to 603 after merge). No npm, no network — the tokenizer is vendored under apps/playground/vendor/tiktoken/.

You can also remove a source, change its precedence, or watch a broken source (down MCP server, missing clone) show an error status without taking down the rest of the graph.

What you’re looking at

• Canvas — each layer/source is a draggable card, arranged by precedence (highest level on top), fanning into a single Resolved concept node. Drag to pan, scroll to zoom, drag a card to move it.

• Left rail — every concept across all layers. A warning badge means the layers disagree on at least one section. The colored dots show which layers contribute.

• Inspector (right) — select a concept to see the effective merge:

  • the precedence chain (which layer outranks which),
  • frontmatter provenance (which layer set each field),
  • merged sections, each tagged with the layer that won it, and
  • conflict panels — the dissenting layers with their dates and content, surfaced inline. This is the “surface, don’t hide” policy made visible — see Conflicts and provenance.

  Click a source node instead to inspect that source’s metadata and the concepts it contributes to.

• Sync re-reads the sources from disk, so you can edit the demo OKF markdown in apps/playground/demo-layers/ and watch the cascade change on refresh.

Reviewing discrepancies

Open Review → Discrepancies for the decision workflow. It starts with the resolver’s section conflicts and also reports authored frontmatter disagreements, broken links when source coverage is settled, and contributors that changed after a prior decision.

Each item explains why it exists, the relevant source evidence, source health, owner, freshness, affected links, and every original value. For disagreements, choose one explicit disposition:

• Use this answer everywhere propagates an existing contribution to every writable contributor.
• Write a reconciled answer previews Markdown you compose before writing it.
• Keep the scoped difference writes no source content and requires a reason.
• Open source files hands unsupported, remote, and read-only values to the file/source workflow.

Broken links use a shorter flow because there is no competing answer to choose. The recommended one-click action keeps the link, records Target not created yet, changes no files, and moves the item to Acknowledged. You can instead open the source concept to edit the link immediately, or expand More options to choose another acknowledgement reason or priority.

Live decisions re-check the revision, stage recoverable originals, write under the profile lock, and retain an append-only audit trail. The public Web Demo instead shows Simulation—no files will change; its history resets on reload, and automatic rules never run there.

After three consistent manual decisions across three distinct discrepancies, ContextCake may suggest a structural rule. Approval creates a recommendation. Automatic use is a separate per-rule choice, and team promotion is separately previewed and confirmed. ContextCake does not infer permission from confidence and does not perform semantic contradiction detection.

Files mode

Toggle Files in the top bar for a live explorer/editor over the same layer bundles:

• Tree — every file under each manifest source, grouped by layer.
• Editor — CodeMirror with syntax highlighting for the file’s type. Text files (markdown, code, SVG) are editable.
• Preview — markdown rendered as prose (sanitized), SVG and raster images rendered as images, PDFs rendered by pdf.js. Use the Split / Source / Preview toggle for text files.
• Save = live re-resolve. Editing a layer’s OKF markdown and saving writes it to disk and re-runs the cascade — the conflict count on the file’s concept chip (and the canvas) updates immediately. For example, remove personal’s Choice section, save, and watch decisions/primary-db drop from 2 conflicts to 1.

Writes are sandboxed to the manifest’s layer roots (path-traversal and symlink guarded, text files only). State-changing requests require a loopback Host and same-origin (CSRF/DNS-rebinding guard), and rendered source content is sanitized (markdown via DOMPurify; SVG shown as an inert image) since a source can be any repo. The libraries — CodeMirror, marked, DOMPurify, pdf.js, and the o200k tokenizer — are vendored under apps/playground/vendor/: no CDN, no npm, works offline. The engine itself stays dependency-free.

The demo data

apps/playground/demo-layers/ holds three tiny OKF bundles wired up by apps/playground/manifest.json:

Layer	Level	Role
company	0	Org baseline
team	2	Overrides + adds
personal	3	Highest precedence

decisions/primary-db is the interesting one: personal wins Choice, team wins Rationale (both surface company as a conflict), while Ownership and Related are inherited untouched from company. Edit any of these files and hit Sync to see the merge react.

Point it at your own layers

node apps/playground/server.mjs --manifest layers.json --port 8790

As with any manifest, only point --manifest at configs you trust — an mcp layer spawns a command. See the trust boundary before connecting the playground to sources you didn’t write.

Next

• Foreign MCP sources — add a non-OKF graph as a source
• Conflicts and provenance — how the merge and surfacing rules work
• The manifest reference — the full layers.json shape