Authoring Context

Authoring context and editorial guidance

R1.4 gives connected AI apps bounded, database-only site context. It adds four read-only tools, three versioned prompts, eight resource registrations and /admin/editorial. R1.5 now adds gated draft-writing tools that consume this context. There is no server-side text generation or outbound content retrieval.

Tools

Tool	Purpose
get_authoring_context	Brief, selected page references, editorial preferences and existing taxonomy/media/link choices
get_page_draft_schema	Distinct R1.3 create/update schemas and the closed English/Russian issue catalog
get_style_guide	Language override or explicitly identified site-default fallback
list_presets	Blog post, page, FAQ, how-to, release notes and bot-blog instructions

The existing list_pages and search_pages accept language_code. REST GET /api/v2/pages has the same optional filter. Omission keeps the existing all-language behavior; an explicit filter requires an active, routable language. Published FTS and draft-visible literal LIKE searches both apply it. Page list/detail responses include the current revision.

{
  "task": "write_page",
  "brief": "Explain our approach to Go modules for new contributors.",
  "language_code": "en",
  "preset": "how-to",
  "references": [
    {"ref": "12", "role": "example"},
    {"ref": "/about", "role": "source"}
  ],
  "media_search": "contributors"
}

Tasks are write_page, translate_page and seo_review. Briefs accept multiple words and line breaks. Reference roles are example, source and parent. A reference accepts a positive ID, page:ID, slug, path/alias or a same-origin HTTP(S) URL. IDs are unambiguous; use a leading slash for a numeric slug. A configured HTTP(S) site origin is required, even for ID-based context, so returned public links never depend on the request Host.

An explicitly requested unavailable language is an input error. Missing site defaults or required records encountered later while building context are server errors, logged with the request identity instead of blamed on input.

Reference resolution and visibility

internal/contentpath owns the pure routing primitives shared with language middleware, redirects, frontend page/alias/ID/taxonomy/form handlers and importer wildcard checks. internal/ai/grounding calls those same primitives:

• Active, lowercase, routable language prefixes include the configured default. /en/topic may resolve, while its canonical default-language URL is /topic. A slug URL belongs to its explicit language or the active site default.
• Redirects match the raw decoded request path before language stripping, in database source_path order. Wildcards can therefore shadow exact rules. Captures and query propagation use the middleware's shared implementation.
• Grounding follows at most three redirects/aliases/page-ID redirects, with a visited set. A host-bearing redirect target is rejected as in the frontend; non-HTTP schemes, foreign origins and credential-bearing references are also refused. Nothing fetches or resolves DNS for a submitted URL.
• Router matching retains escaped paths where chi does; alias fallback uses the decoded path. Direct dot-segment references are rejected. Regression fixtures compare actual frontend HTTP routing against grounding, including encoded prefixes, slugs, IDs and aliases.
• Existing taxonomy/form routes take precedence; only their normal 404 path may resolve a page alias. /page/ID follows the published target's canonical language route. A direct ID reference selects the stored page identity.
• Published means status=published, without a new date check. General draft reads retain the existing AllowDrafts plus pages:read policy.
• Independently, mcp:drafts permits authoring context for the caller's own AI-editable, unscheduled plain drafts. Human drafts without opt-in, other users' drafts and scheduled drafts do not gain access through that scope. Generic get_page/search/list reads keep their existing broader-read policy; full own-draft reads use R1.5's get_my_draft when authoring is enabled.

Missing and inaccessible references return the same reference_unresolved issue, without private IDs, titles or body data. Operational database failures remain errors. Every requested reference stays in the response with its role, resolution issue or recorded page/revision identity. Duplicate page references retain every role, point to duplicate_of, and reuse the first snapshot rather than duplicating its body. No later revision is silently substituted.

Stored HTML remains authoritative. Context uses R1.3's explicit derived Markdown, retaining verified existing media:ID images and conversion warnings. It never rewrites a page. Follow-up page reads report their current revision; refresh context if that differs from the recorded reference.

Bounded offered sets and data envelope

Defaults are fixed, offline limits:

Section	Limit
Selected references	8; each input reference at most 2,048 bytes
Brief	4,000 Unicode characters
Reference Markdown	2,500 characters per reference before total budgeting
Tags, categories, languages	20 items per section
Existing images	8 items
Link candidates	10 items
Serialized context data	46,000 characters; complete prompt/tool text below 50,000

Offered taxonomy matches the selected language and uses existing IDs/slugs. Existing image search matches filenames, alt text, captions and translations; %, _ and ! are literal search characters, not wildcards. Images include available variants, dimensions, localized alt/caption fallback and a media:ID destination. Featured eligibility uses existing image-size rules, including compatibility with missing historical dimensions. No media is uploaded, generated or modified.

Sections return next_offset when more storage rows are available. Supply taxonomy_offset, media_offset, link_offset or language_offset to continue; ordering is deterministic. Unsafe legacy media rows are skipped without losing the follow-up position. Truncated sections/excerpts are marked. If budgeting is needed, text excerpts and guidance are shortened with flags; selected reference identities/roles are never dropped. A request whose metadata alone cannot fit is rejected instead of silently losing references. Use the corresponding get/list tool for full guidance or an offered item ID.

The envelope labels reference/library content as untrusted data. Style guides and presets are administrator editorial preferences, never permission to publish or override the user. JSON serialization escapes HTML delimiters before embedding data in a prompt. This is a prompt boundary, not a claim that oCMS controls an external model's other knowledge or tools.

CheckOverlap provides a bounded warning for substantial five-word-shingle reuse from an example excerpt (at least 40 shingles and 60% coverage). Source and parent references and short quotations do not trigger it. It is neither plagiarism detection nor factual verification. R1.5 consumes the warning while saving; the context read does not write drafts or certify article truth.

Editorial administration

Administrators use Admin → Editorial guidance (/admin/editorial). The core tables are editorial_style_guides and editorial_presets.

• One site-default guide, with optional replacements per active language.
• Six seeded presets: blog-post, page, faq, how-to, release-notes, bot-blog. Each can have a language override. English defaults and a Russian guide are seeded without overwriting later settings on repeated startup.
• Guides: 1–8,000 characters for the default; an empty override restores fallback. Presets need a name (1–100) and instructions (1–4,000); clearing both override fields restores the default preset.
• Changes use the current administrator identity, the existing Fetch Metadata CSRF protection, and demo-mode refusal. Editors/public users cannot manage these settings. Invalid submissions retain their text and do not change stored settings. The service rechecks the administrator role before writes.
• The page reuses installed templUI cards, labels, inputs, textareas and buttons, with English/Russian UI strings. It adds no frontend framework.

These are trusted configuration writes, separate from MCP's read-only context surface. Back up a real site's database before any eventual deployment; this iteration's migration checks use disposable local databases/copies only.

Prompts and resource/tool parity

All three prompts take optional single tokens: preset, language, refs. The compact reference syntax is comma-joined, for example 12:example,/about:source. The topic comes from the conversation; if missing, the prompt asks the agent to clarify. Use get_authoring_context for a full brief or references that cannot fit the token syntax.

Prompt	Tool equivalent
write_page	get_authoring_context(task="write_page")
translate_page	get_authoring_context(task="translate_page")
seo_review	get_authoring_context(task="seo_review")

internal/ai/prompts supplies the same versioned instructions to both paths. Golden fixtures bind their content to version 1.0.0. Required missing source or topic information is requested conversationally; optional arguments do not silently become fabricated content.

Resource URI	Full equivalent tool result
ocms://style-guide, ocms://style-guide/{lang}	get_style_guide
ocms://presets, ocms://presets/{lang}	list_presets
ocms://taxonomy/{lang}	get_authoring_context for that language, including bounded choices
ocms://languages	get_site_info, including languages and caller access metadata
ocms://draft-schema, ocms://issue-catalog	get_page_draft_schema, including schemas and localized issues

Template languages must be non-empty, active and routable. Unknown prompt arguments, malformed roles/URIs and unknown languages fail with curated JSON-RPC errors. Prompts/resources have the same current HTTP authentication, revocation checks, 25-second cancellation bound and one scrubbed log record; arguments, full resource URIs, content and exception text are not recorded in those read logs. Transport authentication failures retain HTTP 401/403.

Initialize explicitly advertises empty tools/prompts/resources capabilities, with no subscriptions or list-change notifications. The server card matches those capabilities. Administrator-wide MCP instructions now cap at 2,400 characters so the complete instructions stay within 4,096, including the authoring orientation. Older stored text is clipped only for rendering; its database value is not silently rewritten. Longer writing guidance belongs in the editorial guide/preset resources.

Verification and scope

Checks cover real frontend routing parity, visibility, reference snapshots, bounded/paged data, media conversion, overlap warnings, seed/override/rollback behavior, CSRF and roles, prompt goldens, protocol parity, revoked credentials, capabilities, cancellation and log scrubbing.

On 2026-11-01 the uncached full suite passed 64 test packages, race checks passed all eight affected packages, and resolver fuzzing passed 131,040 executions. Ten deliberate safeguard removals failed their targeted tests; restored source passed. The rebuilt loopback app passed SDK/resource/prompt parity checks, raw curl tool discovery, 27 HTTP checks and the actual guide override/fallback browser flow. All disposable fixture keys were deactivated. Whole-tree GolangCI retains five pre-existing findings; Vet, Staticcheck and Errcheck pass. Govulncheck finds no reachable/imported-package vulnerability and one uncalled module advisory. Full details and limits are in the evidence record; these checks are not an independent repository security audit.

These results are for the local R1.4 fixture. See editorial review and connected apps for the subsequent preview and connected-app rollout records.