← Docs hub

Reader API contract (v1.2+ preview)

Status: contract-only. No server yet — today the static site is the API. This doc locks the shape so when we add a hosted / SPA reader we don't have to rewrite the content model. Freezing this now protects the build pipeline (site/ outputs) and the sibling .txt / .json files from drift (#116).

Why a contract first

llmwiki is, and will stay, static-site-first. But a few near-term bets depend on the data being reachable without HTML parsing:

• A browser extension that answers "what do I know about X" from the wiki's .json sibling of the current tab.
• A Raycast/Alfred plugin that hits manifest.json + search-index.json to open a page.
• A future lightweight SPA reader that can live on the same origin as the generated site.
• Downstream LLM agents consuming llms-full.txt + per-page .json to answer questions without pulling HTML.

Every one of those wants the same shape of data. This doc says what that shape is, so refactors of llmwiki/build.py can't silently break clients.

Shipped today (v1.0+) — read-only, file-based

The static build writes these to site/ on every llmwiki build:

These are already the API. Everything below in this doc describes the future hosted/SPA surface that will be fed by the same data shapes — no new content pipeline, just new transports.

Future endpoint contract

Every endpoint below maps 1:1 to a file that llmwiki build already produces. The server is a thin JSON wrapper; the content model is what's already on disk.

Base URL: <root>/api/v1 (TBD — static deploy keeps /api/v1/*.json as files).

GET /api/v1/bootstrap

One-shot payload the reader fetches on first load so it doesn't have to chain three requests before showing anything.

{
  "version": "1.1.0rc2",
  "generated_at": "2026-04-19T08:34:42Z",
  "stats": {
    "sessions": 647,
    "projects": 30,
    "entities": 2,
    "concepts": 0,
    "total_bytes": 62691698
  },
  "nav": [
    { "id": "home",          "label": "Home",      "href": "/" },
    { "id": "projects",      "label": "Projects",  "href": "/projects/" },
    { "id": "sessions",      "label": "Sessions",  "href": "/sessions/" },
    { "id": "models",        "label": "Models",    "href": "/models/" },
    { "id": "vs",            "label": "Compare",   "href": "/vs/" },
    { "id": "graph",         "label": "Graph",     "href": "/graph.html" },
    { "id": "changelog",     "label": "Changelog", "href": "/changelog.html" }
  ],
  "theme": {
    "accent":  "#7C3AED",
    "default": "dark"
  },
  "search": {
    "mode":   "flat",
    "chunks": "/search-chunks/",
    "index":  "/search-index.json"
  },
  "cache_tiers": ["L1", "L2", "L3", "L4"]
}

Client contract. Safe to cache for 5 minutes. Never returns partial data — if the site rebuilds mid-request, the server serves the previous full payload until the new one is ready.

GET /api/v1/article?path=<url>

The article shell already rendered as structured data — lets a SPA skip HTML parsing entirely.

{
  "url":   "sessions/llm-wiki/2026-04-17T10-12-llm-wiki-refactor.html",
  "slug":  "2026-04-17T10-12-llm-wiki-refactor",
  "title": "LLM Wiki refactor",
  "type":  "source",
  "project": "llm-wiki",
  "model": "claude-sonnet-4-6",
  "date": "2026-04-17",
  "last_updated": "2026-04-17",
  "confidence": 0.75,
  "lifecycle": "reviewed",
  "cache_tier": "L3",
  "entity_type": null,
  "tags": ["claude-code", "refactor"],
  "breadcrumbs": [
    { "label": "Home",     "href": "/" },
    { "label": "Projects", "href": "/projects/" },
    { "label": "llm-wiki", "href": "/projects/llm-wiki.html" },
    { "label": "LLM Wiki refactor" }
  ],
  "body_html": "<article>…</article>",
  "body_text": "Raw markdown body without frontmatter, suitable for LLM context.",
  "wikilinks_out": ["Obsidian", "Karpathy"],
  "wikilinks_in":  ["llm-wiki", "AndrejKarpathy"],
  "related": [
    { "slug": "2026-04-16T18-30-llm-wiki-seed", "title": "LLM Wiki seed", "score": 0.82 }
  ],
  "reading_time_minutes": 4,
  "summary": "First-paragraph summary for L2 pre-loading."
}

Required fields: url, slug, title, type, body_html, body_text, wikilinks_out. Everything else is optional and may be null/missing.

Client contract. The reader MUST gracefully render when optional fields are missing (a newly ingested page may not have confidence or cache_tier yet).

GET /api/v1/search?q=<query>&type=<optional>&project=<optional>

Thin wrapper over the existing client-side index + chunks. Returns the matches the palette would surface.

{
  "query": "karpathy",
  "mode":  "flat",
  "total": 12,
  "hits": [
    {
      "id":    "session:llm-wiki/2026-04-16T18-30-llm-wiki-seed",
      "url":   "sessions/llm-wiki/2026-04-16T18-30-llm-wiki-seed.html",
      "title": "LLM Wiki seed",
      "type":  "source",
      "project": "llm-wiki",
      "snippet": "Karpathy's pattern spells out what…",
      "score":   0.91,
      "headings": [
        { "depth": 2, "text": "Summary" },
        { "depth": 3, "text": "Karpathy's pattern" }
      ]
    }
  ],
  "facets": {
    "entity_type": { },
    "lifecycle":   { },
    "tags":        { },
    "confidence":  { "none": 647 }
  }
}

Mode. "flat" vs "tree" — the client-side router today picks the mode by heuristic (#53 lands the auto-router). The server MUST return the same mode it used so the client can tell the user in the palette footer.

Client contract. hits is capped at 100; the client does its own pagination. score is 0–1 but not calibrated — use for ranking, not thresholds.

POST /api/v1/sync (internal only)

Trigger a rebuild without waiting for the next watcher tick. Used by /wiki-sync after a successful ingest.

POST /api/v1/sync
Authorization: Bearer <local-token>

{
  "reason": "ingest",
  "pages_changed": ["sources/llm-wiki-refactor.md"]
}

Response:

{
  "accepted": true,
  "build_id": "2026-04-19T10:22:01Z",
  "eta_seconds": 2
}

Auth. Local bearer token only — this endpoint is never exposed to the public internet. manifest.json is the read-side proof that the build finished (its generated_at advances).

Data model invariants

Anything a client can depend on:

1. Slugs are stable. A page's slug is set at ingest and never changes on rebuild. Renames produce a new slug and a redirect stub.
2. Timestamps are UTC ISO-8601 with Z suffix. Never local time.
3. cache_tier is always one of L1, L2, L3, L4 (#52). Missing = treat as L3.
4. lifecycle is always one of draft, reviewed, verified, stale, archived (#11).
5. confidence is always in [0, 1] or missing. Never percent.
6. entity_type (when set) is one of person, org, tool, concept, api, library, project (#137).
7. Wikilinks resolve to slugs, not URLs. [[Karpathy]] → "Karpathy" — the client resolves to a URL via the index.
8. Frontmatter is authoritative for metadata. The body is authoritative for prose.

Versioning

• /api/v1/* is the long-term contract. Breaking changes bump to /v2/ and keep /v1/ live for one minor version.
• Additive-only changes (new optional fields, new top-level keys on bootstrap) don't bump the version.
• Rename of an existing required field is a breaking change.

Content negotiation

Today's static site already does this implicitly:

• curl .../page.html → HTML
• curl .../page.txt → plain text
• curl .../page.json → structured

The future server will keep those three paths exactly as-is. Accept: application/json on .html routes should redirect to the .json sibling rather than serving JSON on the HTML URL — that way caches and proxies stay simple.

Migration path — static → hosted

1. Today: llmwiki build writes the JSON/txt files. External tools read them directly. (Done — #116 is this doc.)
2. v1.2: Add a tiny llmwiki serve --api flag that wraps the same files behind /api/v1/* paths so the reader SPA can fetch them uniformly in dev. No new data, just routing.
3. v1.3+: If a hosted multi-tenant reader ships, the server reuses the same routes with per-user auth. The content pipeline doesn't change.

At no point does the contract require a rewrite of llmwiki/build.py — every endpoint maps to something build.py already emits.

Related

• llmwiki/build.py — produces every file referenced above
• llmwiki/exporters.py — llms.txt + JSON-LD + per-page siblings
• docs/reference/cache-tiers.md — cache_tier invariant (#52)
• docs/design/brand-system.md — theme tokens returned by /bootstrap
• #116 — this issue
• #112 — reader-first article shell (one client of this contract)