@cerefox/memory 1.18.0 → 1.19.1

package/dist/frontend/index.html

@@ -15,8 +15,8 @@
       href="https://fonts.googleapis.com/css2?family=Geist:wght@300;400;500;600;700&display=swap"
     />
     <title>Cerefox</title>
-    <script type="module" crossorigin src="/app/assets/index-CxMT7hjH.js"></script>
-    <link rel="stylesheet" crossorigin href="/app/assets/index-B0H_N4if.css">
+    <script type="module" crossorigin src="/app/assets/index-HNs8rgY1.js"></script>
+    <link rel="stylesheet" crossorigin href="/app/assets/index-01k5C3sG.css">
   </head>
   <body>
     <div id="root"></div>

package/dist/server-assets/_shared/ef-meta/index.ts

@@ -18,7 +18,7 @@
  * doesn't touch `supabase/functions/` leaves it alone).
  */
 
-export const EF_VERSION = "1.18.0";
+export const EF_VERSION = "1.19.1";
 
 /**
  * The Cerefox RELEASE version — what `cerefox --version` reports and what npm
@@ -36,7 +36,7 @@ export const EF_VERSION = "1.18.0";
  * is imported by the Deno Edge Functions, which cannot reach into the npm
  * package.
  */
-export const CEREFOX_VERSION = "1.18.0";
+export const CEREFOX_VERSION = "1.19.1";
 
 /**
  * The most recent version whose EF-side SOURCE actually changed (#127).
@@ -46,7 +46,7 @@ export const CEREFOX_VERSION = "1.18.0";
  * `cut_release.ts` ONLY when EF source changed since the last tag; doctor
  * uses it to stay silent on label-only drift.
  */
-export const EF_LAST_CHANGED = "1.17.3";
+export const EF_LAST_CHANGED = "1.19.0";
 
 /**
  * The 8 peer EFs the cerefox-mcp aggregator probes (excludes cerefox-mcp

package/dist/server-assets/_shared/embeddings/onnx-embedder.ts

@@ -263,6 +263,20 @@ type FeaturePipeline = (texts: string[], opts?: unknown) => Promise<any>;
 
 let pipelinePromise: Promise<FeaturePipeline> | null = null;
 
+/**
+ * ONNX Runtime's CPU memory arena is OFF (#335). The arena keeps its
+ * high-water mark for the life of the process and grew with every long
+ * document, so the web process never came back down: measured on synthetic
+ * 20,000-character, token-dense documents, idle memory after five ingests was
+ * 1.8 GB with the arena and 0.36 GB without, and peak fell from 2.05 GB to
+ * 1.39 GB, at the same speed. In a container sharing a small Docker VM, that
+ * retained memory is what the kernel's out-of-memory killer finds.
+ */
+const PIPELINE_OPTIONS = {
+  dtype: ONNX_MODEL_DTYPE,
+  session_options: { enableCpuMemArena: false },
+};
+
 /**
  * Download (first use) + materialise the inference pipeline, with progress to
  * stderr. Ported from cfcf including its hard-won progress-rendering fixes
@@ -275,9 +289,7 @@ async function ensurePipeline(): Promise<FeaturePipeline> {
     // A cached model is a local read: no banner, no per-file lines, no
     // "ready" line. The narration is for the one run that downloads.
     if (isModelCached()) {
-      const pipe = await transformers.pipeline("feature-extraction", ONNX_MODEL_ID, {
-        dtype: ONNX_MODEL_DTYPE,
-      });
+      const pipe = await transformers.pipeline("feature-extraction", ONNX_MODEL_ID, PIPELINE_OPTIONS);
       return pipe as unknown as FeaturePipeline;
     }
     const mb = ONNX_MODEL_APPROX_MB;
@@ -367,9 +379,7 @@ async function ensurePipeline(): Promise<FeaturePipeline> {
     );
     // Every file is on disk now, so this is a local load: no progress callback,
     // which would only report each file "done" a second time.
-    const pipe = await transformers.pipeline("feature-extraction", ONNX_MODEL_ID, {
-      dtype: ONNX_MODEL_DTYPE,
-    });
+    const pipe = await transformers.pipeline("feature-extraction", ONNX_MODEL_ID, PIPELINE_OPTIONS);
     process.stderr.write(`[cerefox-embed] embedder ready.\n`);
     return pipe as unknown as FeaturePipeline;
   })();
@@ -395,10 +405,16 @@ export async function warmup(): Promise<void> {
  * Per-inference sub-batch. Peak tensor memory scales with the batch, and the
  * container shares a (often small) Docker VM with Postgres + PostgREST + the
  * web server — a 12-text single call was OOM-killed on Colima's default 2 GB
- * VM (rc.2 dogfood, exit 137). 4 keeps peak memory flat at personal scale;
- * override with CEREFOX_ONNX_BATCH.
+ * VM (rc.2 dogfood, exit 137).
+ *
+ * 1 since #335 (was 4). A batch is padded to its longest chunk and attention
+ * memory grows with the square of that length, so four full-size, token-dense
+ * chunks (tables, ids, code) needed more than 2.5 GB and got the web process
+ * OOM-killed, while one at a time peaked at 1.4 GB. It is not slower on CPU:
+ * without the padding, batch 1 ingested prose slightly FASTER (about 1.7 s vs
+ * 1.9 s per 20,000-character document). Override with CEREFOX_ONNX_BATCH.
  */
-const DEFAULT_ONNX_BATCH = 4;
+const DEFAULT_ONNX_BATCH = 1;
 
 function onnxBatchSize(): number {
   const env = (globalThis as { process?: { env?: Record<string, string | undefined> } })

package/dist/server-assets/_shared/mcp-tools/_document-meta.ts

@@ -24,7 +24,7 @@
 
 import type { AccessPath, MCPSupabaseClient } from "./types.ts";
 
-import { logUsage, storeWriteRemediation } from "./_utils.ts";
+import { MetadataConcurrencyError, logUsage, metadataTokenParams, storeWriteRemediation } from "./_utils.ts";
 
 export interface FacetActor {
   author: string;
@@ -297,6 +297,9 @@ export async function updateDocumentFacets(
     documentId: string;
     title?: string;
     metadata?: Record<string, unknown>;
+    /** #333: the metadata token for the replace below; see the RPC. */
+    expectedMetadataRevision?: number | null;
+    metadataLastWriteWins?: boolean;
     projectIds?: string[];
     accessPath: AccessPath;
   } & FacetActor,
@@ -322,7 +325,7 @@ export async function updateDocumentFacets(
     await step(async () => {
       const { data: doc, error } = await supabase
         .from("cerefox_documents")
-        .select("metadata")
+        .select("*")
         .eq("id", opts.documentId)
         .is("deleted_at", null)
         .limit(1);
@@ -331,6 +334,22 @@ export async function updateDocumentFacets(
         throw new FacetNotFoundError(`Document not found (or in the trash): ${opts.documentId}`);
       }
       const stored = (doc[0].metadata ?? {}) as Record<string, unknown>;
+      // Strict when given (#333 review), even when nothing would change: a
+      // stale form must not slip through just because its tags happen to
+      // match, and go on to apply its other facets. `*` rather than naming
+      // the column keeps a pre-0.21.0 server (no column, no check) working.
+      const storedRevision = (doc[0] as { metadata_revision?: number | null }).metadata_revision;
+      if (
+        opts.expectedMetadataRevision != null &&
+        storedRevision != null &&
+        storedRevision !== opts.expectedMetadataRevision &&
+        !opts.metadataLastWriteWins
+      ) {
+        throw new MetadataConcurrencyError(
+          `CEREFOX_METADATA_CONFLICT: the metadata of document ${opts.documentId} changed since it was read ` +
+            `(expected metadata_revision ${opts.expectedMetadataRevision}, current ${storedRevision}).`,
+        );
+      }
       const wanted = normalizeMetadata(opts.metadata!);
       if (stableStringify(stored) !== stableStringify(wanted)) {
         // Replace mode: the save surface edits the whole object (an empty
@@ -343,9 +362,12 @@ export async function updateDocumentFacets(
           p_replace: true,
           p_author: opts.author,
           p_author_type: opts.authorType,
+          // #333: a replace must prove a read (or say last-write-wins).
+          ...metadataTokenParams(opts.expectedMetadataRevision, opts.metadataLastWriteWins),
         });
         if (rpcErr) {
           const msg = rpcErr.message ?? String(rpcErr);
+          if (msg.includes("CEREFOX_METADATA_")) throw new MetadataConcurrencyError(msg);
           if (/not found/i.test(msg)) throw new FacetNotFoundError(msg);
           throw new Error(`Metadata update failed: ${msg}`);
         }

package/dist/server-assets/_shared/mcp-tools/_utils.ts

@@ -164,6 +164,65 @@ export interface LogUsageParams {
  *  Differs from the EF's `logUsage` only in that `accessPath` is a required
  *  parameter (was hardcoded to `"remote-mcp"` in the EF) so the local TS
  *  MCP server can pass `"local-mcp"` for the same call site. */
+/**
+ * Thrown when a metadata write loses the race (#333): metadata_revision moved
+ * between read and write, or a replace came without a token. The web routes
+ * map a conflict to HTTP 409 and a missing token to 400.
+ */
+export class MetadataConcurrencyError extends Error {
+  readonly tokenRequired: boolean;
+  constructor(message: string) {
+    super(message);
+    this.name = "MetadataConcurrencyError";
+    this.tokenRequired = message.includes("CEREFOX_METADATA_TOKEN_REQUIRED");
+  }
+}
+
+/** The #333 metadata-token RPC parameters, only when set. Sending them as
+ *  null/false would name a signature a server older than schema 0.21.0 does
+ *  not have, and PostgREST would fail the whole write with "function not
+ *  found"; omitted, an older server runs the write exactly as before. */
+export function metadataTokenParams(
+  expected: number | null | undefined,
+  lastWriteWins: boolean | undefined,
+): Record<string, unknown> {
+  return {
+    ...(expected != null ? { p_expected_metadata_revision: expected } : {}),
+    ...(lastWriteWins ? { p_last_write_wins: true } : {}),
+  };
+}
+
+/** Agent-first text for the metadata token errors (#333), or null when the
+ *  message is not one. Shared by every tool that writes metadata, so the
+ *  remedy reads the same whichever path the agent took. */
+export function metadataTokenError(message: string): Error | null {
+  if (message.includes("CEREFOX_METADATA_CONFLICT")) {
+    const id = message.match(/document ([0-9a-f-]{36})/)?.[1] ?? "(unknown)";
+    const expected = message.match(/expected metadata_revision (\d+)/)?.[1] ?? "?";
+    const current = message.match(/current (\d+)/)?.[1] ?? "?";
+    return new Error(
+      `Metadata conflict: the metadata of document ${id} changed since you read it ` +
+        `(your metadata_revision: ${expected}, current: ${current}). Nothing was written. ` +
+        `To resolve: (1) cerefox_get_document("${id}", outline: true) for the current metadata ` +
+        `and metadata_revision, (2) reapply your change to it, (3) retry with ` +
+        `expected_metadata_revision set to the new value. Do not overwrite blindly — ` +
+        `another writer's keys may be in there.`,
+    );
+  }
+  if (message.includes("CEREFOX_METADATA_TOKEN_REQUIRED")) {
+    const current = message.match(/Current metadata_revision: (\d+)/)?.[1];
+    return new Error(
+      `Replacing all metadata deletes every key you do not list, so it needs ` +
+        `expected_metadata_revision: the metadata_revision you read (returned by ` +
+        `cerefox_get_document and cerefox_metadata_search).` +
+        (current ? ` The document's current metadata_revision is ${current}; pass it ONLY if you read the current metadata.` : "") +
+        ` To change some keys, drop replace and merge instead (null removes a key). ` +
+        `To overwrite regardless, pass last_write_wins=true.`,
+    );
+  }
+  return null;
+}
+
 /** Pull the two hashes out of an RPC `CEREFOX_CONFLICT` message.
  *
  *  The ONE site coupled to the SQL `RAISE` wording ("expected hash %, current
@@ -202,6 +261,7 @@ export {
   AUDIT_OPERATIONS,
   auditDocLabel,
   isStoreLevelAuditOp,
+  metadataChangeSummary,
   STORE_LEVEL_AUDIT_OPS,
 } from "./audit-ops.ts";
 
@@ -215,6 +275,38 @@ export function isAuditCheckError(message: string): boolean {
   return /cerefox_audit_log_operation_check/.test(message);
 }
 
+/**
+ * A create that lost the duplicate-content race (0.21.1, #334 review): the RPC
+ * raises CEREFOX_DUPLICATE_CONTENT naming the live document that already has
+ * this content, and nothing was written. Callers report it as a skip.
+ */
+export class DuplicateContentError extends Error {
+  constructor(public readonly existingDocumentId: string, message: string) {
+    super(message);
+    this.name = "DuplicateContentError";
+  }
+}
+
+/** The existing document's id from a CEREFOX_DUPLICATE_CONTENT error, or null. */
+export function duplicateContentDocumentId(message: string): string | null {
+  if (!message.includes("CEREFOX_DUPLICATE_CONTENT")) return null;
+  return message.match(/document ([0-9a-f-]{36})/i)?.[1] ?? null;
+}
+
+/**
+ * The store-wide content constraint a server below schema 0.20.0 still has
+ * (#332). A current client meets it when it writes content that some other
+ * document, trashed ones included, already has: only the server can fix that.
+ */
+export function legacyContentUniqueError(message: string): Error | null {
+  if (!message.includes("cerefox_documents_hash_unique")) return null;
+  return new Error(
+    "Another document (possibly one in the trash) already has exactly this content, " +
+      "and this server predates schema 0.20.0, which still refuses that. " +
+      `Run \`cerefox server deploy\`, then retry. (${message})`,
+  );
+}
+
 /** Postgres 23505 through PostgREST — one predicate, not N copied regexes. */
 export function isDuplicateKeyError(message: string): boolean {
   return /duplicate key|unique constraint|23505/i.test(message);

package/dist/server-assets/_shared/mcp-tools/audit-log.ts

@@ -5,7 +5,7 @@
 
 import type { MCPSupabaseClient } from "./types.ts";
 
-import { logUsage, auditDocLabel, AUDIT_OPERATIONS } from "./_utils.ts";
+import { logUsage, auditDocLabel, AUDIT_OPERATIONS, metadataChangeSummary } from "./_utils.ts";
 import type { ToolContext, ToolDefinition } from "./types.ts";
 import { AUTHOR_PARAM_READ, auditLogIdentity } from "./identity.ts";
 
@@ -54,6 +54,8 @@ async function handler(
     size_after: number | null;
     description: string;
     created_at: string;
+    metadata_before?: Record<string, unknown> | null;
+    metadata_after?: Record<string, unknown> | null;
   }>;
 
   logUsage(supabase, {
@@ -73,7 +75,9 @@ async function handler(
         : e.size_after != null
           ? ` | ${e.size_after} chars`
           : "";
-    return `${utcStamp(e.created_at)} | ${e.operation} | ${e.author} (${e.author_type}) | ${docLabel}${sizeInfo} | ${e.description}`;
+    // #333: the actual metadata change, so the trail can undo a mistake.
+    const change = metadataChangeSummary(e.metadata_before, e.metadata_after);
+    return `${utcStamp(e.created_at)} | ${e.operation} | ${e.author} (${e.author_type}) | ${docLabel}${sizeInfo} | ${e.description}${change ? ` | ${change}` : ""}`;
   });
   return `Audit log (${entries.length} entries, newest first):\n\n${lines.join("\n")}`;
 }

package/dist/server-assets/_shared/mcp-tools/audit-ops.ts

@@ -44,3 +44,30 @@ export function auditDocLabel(
   if (documentId) return documentId.slice(0, 8) + "…";
   return isStoreLevelAuditOp(operation) ? "(store)" : "(deleted)";
 }
+
+/**
+ * One line describing a metadata change (#333): the keys set (with their new
+ * values) and the keys removed. Built from an `update-metadata` audit entry's
+ * metadata_before / metadata_after; null when the entry carries neither (older
+ * entries, other operations) or nothing actually changed.
+ */
+export function metadataChangeSummary(
+  before: Record<string, unknown> | null | undefined,
+  after: Record<string, unknown> | null | undefined,
+): string | null {
+  if (before == null && after == null) return null;
+  const b = before ?? {};
+  const a = after ?? {};
+  const show = (v: unknown) => (typeof v === "string" ? v : JSON.stringify(v));
+  const set = Object.keys(a)
+    .filter((k) => JSON.stringify(a[k]) !== JSON.stringify(b[k]))
+    .map((k) => `${k}=${show(a[k])}`);
+  // Own keys only: `in` also finds inherited ones, so removing a key named
+  // like an Object.prototype member ("constructor") went unreported.
+  const removed = Object.keys(b).filter((k) => !Object.hasOwn(a, k));
+  if (set.length === 0 && removed.length === 0) return null;
+  const parts: string[] = [];
+  if (set.length) parts.push(`set ${set.join(", ")}`);
+  if (removed.length) parts.push(`removed ${removed.join(", ")}`);
+  return parts.join("; ");
+}

package/dist/server-assets/_shared/mcp-tools/get-document.ts

@@ -49,6 +49,8 @@ async function handler(
         chunk_count?: number;
         total_chars?: number;
         content_hash?: string;
+        doc_metadata?: Record<string, unknown> | null;
+        metadata_revision?: number;
       }
     | undefined;
 
@@ -84,6 +86,8 @@ async function handler(
       {
         title: row.doc_title ?? "Untitled",
         content_hash: archived ? null : (row.content_hash ?? null),
+        metadata: row.doc_metadata ?? {},
+        metadata_revision: archived ? null : (row.metadata_revision ?? null),
         total_chars: row.total_chars ?? (row.full_content ?? "").length,
         outline: nodes,
         note: archived
@@ -140,7 +144,13 @@ async function handler(
   const label = version_id !== null ? " (archived version)" : " (current)";
   // content_hash is the optimistic-concurrency token: pass it back as
   // expected_content_hash when updating this document via cerefox_ingest.
-  const hashLine = row.content_hash ? `content_hash: ${row.content_hash}\n\n` : "";
+  // metadata_revision is the metadata's own token (#333): pass it back as
+  // expected_metadata_revision on a metadata write.
+  const revision =
+    version_id === null && row.metadata_revision != null
+      ? ` · metadata_revision: ${row.metadata_revision}`
+      : "";
+  const hashLine = row.content_hash ? `content_hash: ${row.content_hash}${revision}\n\n` : "";
   return `# ${row.doc_title ?? "Untitled"}${label}\n${hashLine}${row.full_content ?? ""}`;
 }
 

package/dist/server-assets/_shared/mcp-tools/get-help-content.ts

@@ -11,13 +11,13 @@
  * docs/specs/polish-and-distribution-design.md §10d.
  */
 
-export const HELP_FULL = "# Cerefox Knowledge Base -- Agent Quick Reference\n\nCerefox is a persistent, shared knowledge base. You have **15 core MCP tools** (14 with CLI equivalents — `cerefox_get_help` is MCP-only), plus 4 dormant relation tools that appear only when `relations_enabled` is on. For the full guide, search Cerefox for \"How AI Agents Use Cerefox\" or call `cerefox_get_help` to retrieve this content over MCP.\n\n## Tools\n\n| Tool | Purpose | Key params |\n|------|---------|------------|\n| `cerefox_search` | Find documents (hybrid FTS + semantic) | `query` (required), `project_name`, `metadata_filter`, `author` |\n| `cerefox_ingest` | Save or update a document | `title`, `content` (required), `document_id` (update by ID), `expected_content_hash` (**required on content updates** — see rule 10), `last_write_wins`, `update_if_exists`, `project_name` (single, non-destructive add on update), `project_names` (list, destructive replace on update), `metadata` (omit on update to keep existing tags; `{}` clears), `author` |\n| `cerefox_insert` | **Add** to a document without resending it. Cannot destroy content. | `document_id`, `text`, `position` (`end_of_document`/`end_of_section`/`after_heading`/`before_heading`), `expected_content_hash` (required), `anchor_heading` (unless `end_of_document`), `section_part`, `author` |\n| `cerefox_edit` | **Change** parts of a document: 1..n operations applied atomically | `document_id`, `operations` (`insert`/`replace_section`/`delete_section`/`rename_section`), `expected_content_hash` (required), `author` |\n| `cerefox_delete_document` | **Soft**-delete a document (to trash; excluded from search; permanent purge is human-only) | `document_id`, `expected_content_hash` (**required** — a delete must follow a read), `reason` (recorded in the audit log — give one), `author` |\n| `cerefox_restore_document` | Restore a soft-deleted document from the trash (audited inverse of delete; no-op if not deleted) | `document_id` (required), `reason` (recorded in the audit log), `author` |\n| `cerefox_get_document` | Get full document by ID (header includes `content_hash` — the update token), or with `outline: true` just its heading paths, sizes and hash, or with `section: \"## Heading\"` one section's text | `document_id` (required), `outline`, `section`, `section_part`, `author` |\n| `cerefox_list_versions` | Version history of a document | `document_id` (required), `author` |\n| `cerefox_set_relation` ⚑ | Link two documents (`source --rel_type--> target`) | `source_id`, `target_id`, `rel_type` (required), `metadata`, `author` |\n| `cerefox_delete_relation` ⚑ | Remove a relation | `source_id`, `target_id`, `rel_type`, `author` |\n| `cerefox_get_relations` ⚑ | All relations touching a document, both directions | `document_id` |\n| `cerefox_get_neighbors` ⚑ | Walk the graph along ONE relation type | `document_id`, `rel_type` (required), `depth`, `from_time`, `to_time`, `limit` |\n| `cerefox_metadata_search` | Find or list docs by metadata, project, or time (no text query) | `metadata_filter`, `project_name` (list a project's docs), `updated_since`, `include_content` — **at least one** of metadata_filter/project_name/updated_since/created_since |\n| `cerefox_list_metadata_keys` | Discover available metadata keys | (none required) |\n| `cerefox_list_projects` | List all projects | (none required) |\n| `cerefox_set_document_metadata` | Change tags WITHOUT resending content. **Merges** by default; a `null` value removes a key | `document_id`, `metadata` (required), `replace` (rare: set exactly this object), `author` |\n| `cerefox_set_document_projects` | Set doc's project memberships to exactly the given list (destructive replace; metadata-only, no content change) | `document_id`, `project_names` (required), `author` |\n| `cerefox_get_audit_log` | Query write operation history | `document_id`, `by_author` (filter), `operation`, `since`, `author` |\n| `cerefox_get_help` | Retrieve Cerefox conventions (this reference) over MCP. **Call this whenever uncertain.** | `topic` (optional, case-insensitive H2 substring match) |\n\n⚑ **Opt-in — usually absent.** The four relation tools are hidden unless the\noperator enables them (`relations_enabled`). **Trust your own tool list**: if\nthey are not in it, the feature is switched off for this deployment. That is\nnormal, not an error, and not something to work around.\n\n## Editing part of a document (prefer this over re-sending)\n\n**Re-sending a whole document to change part of it is the main way agents lose\ndata.** You have to reproduce the untouched remainder verbatim, and any drift\nsilently rewrites content nobody asked you to touch — which the caller cannot\ndiff. Use the partial-edit tools instead:\n\n1. **Learn the anchors** — `cerefox_get_document(document_id, outline: true)`.\n   Returns heading paths, per-section sizes and the `content_hash`, without the\n   body. The paths it returns are exactly what `anchor_heading` accepts.\n2. **Add** → `cerefox_insert`. `end_of_document` is a plain append;\n   `end_of_section` adds inside a named section. It is structurally incapable of\n   removing anything, so \"I meant to append\" cannot become \"I replaced the file\".\n3. **Look before you overwrite** — `cerefox_get_document(document_id,\n   section: \"## Heading\")` returns exactly the text a `replace_section` on that\n   anchor would destroy. The outline gives you a section's *size*, never its\n   *text*, so on a document you did not write yourself this is the difference\n   between a replace and a blind overwrite.\n4. **Change or remove** → `cerefox_edit`. Put changes that belong together in\n   ONE call: they apply atomically, so a table row and the total it feeds cannot\n   end up disagreeing. To change a single line, `replace_section` on its\n   smallest enclosing heading — that is the intended granularity, not a\n   workaround. To fix a stale heading (`## OPEN TODOs (as of ...)`), use\n   `rename_section`: it changes the heading text and leaves the body and\n   position alone.\n5. All of them require `expected_content_hash` and **have no last-write-wins**. A\n   conflict means someone else changed the document; re-read and decide, do not\n   force it.\n\n**A section runs to the next same-or-higher heading, or to the end of the\ndocument.** So `end_of_document` inserts land inside the *last* section, and\nreplacing or deleting that section removes them too. A large shrink in the\nresponse is your warning; `cerefox_list_versions` has the previous content.\n\n**When an anchor is ambiguous the tool refuses and hands you the options** — a\nrepeated heading returns the qualifying paths, and a section with both its own\ncontent and sub-sections returns both `section_part` choices. That is a\nrecoverable answer, not a failure: retry with what it gave you.\n\n## Essential Rules\n\n1. **Search before ingesting** -- check if the document exists first.\n2. **Prefer ID-based updates** -- pass `document_id` from search results for deterministic updates. Falls back to title-matching with `update_if_exists: true`.\n3. **Set `author`** to your name on every call, reads and writes alike (e.g., \"Claude Code\", \"archiver\"). Same parameter on every tool. (`requestor` is still accepted everywhere as the pre-1.13.1 alias.) On the CLI it is `--author` on every command too (plus `--author-type` on writes); or rely on `CEREFOX_AUTHOR_NAME`/`CEREFOX_AUTHOR_TYPE` env vars set in the user's `.env`.\n4. **Use `document_id` from search results** `[id: uuid]` for get_document and list_versions.\n5. **`max_bytes` has to fit a whole document.** Search returns COMPLETE documents, so a budget smaller than the top hit fits nothing. You get a header-only list saying so (never \"No results found.\" — that means the store really has nothing). Raise `max_bytes`, or read one document with `cerefox_get_document` using `outline: true` or `section`.\n6. **Add metadata** -- at minimum `type` (\"decision-log\", \"research\", \"design-doc\") and `status` (\"active\", \"draft\").\n7. **Write for retrieval: the title does real work.** The document title is indexed at the **highest weight** *and* prepended to every chunk before it is embedded, so it steers both halves of hybrid search (keyword and semantic) for every chunk in the document. Name the document with the distinctive terms someone would actually search for -- \"Cerefox Response Size Limits\", not \"Notes\" or \"Update 3\". Renaming re-indexes and re-embeds, so correcting a vague title is cheap and takes effect immediately. Write the characters themselves: `&` and `<>`, never `&amp;` or `&lt;&gt;` (entities are stored literally and the write reply will say so). Inside the content, write structured Markdown with H1/H2/H3 headings: the chunker follows them, and a chunk's own heading carries the same weight as the title.\n8. **Deletes are soft (recoverable); permanent purge is web-UI-only.** `cerefox_delete_document` requires the document's `content_hash` as you read it (read before you delete) and takes a `reason` — give one; it is what the human reviewing the trash sees. `cerefox_restore_document` undoes a mistaken delete (also audited, also takes a `reason`). Always surface deletes AND restores to the user. Once a human purges from the web UI, the document is gone for good; so is one auto-purged after `trash_retention_days`, if the operator has turned trash auto-purge on.\n9. **Cross-doc links inside content**: **always use `[Text](document-uuid)`.** UUIDs are the only fully reliable link form — stable across title changes, never ambiguous, no encoding gotchas. Every `cerefox_search` result shows `[id: <uuid>]` after the title; grab it and use it. Title-based linking (`[Text](<Title With Spaces>)`) is fragile (breaks on colons, parens, ampersands, brackets — silently navigates to wrong page) — **don't write title-based links**; do an extra search to get the UUID instead. Repo-path forms (`[Text](docs/path.md)`) exist for repo-ingested files; don't construct manually. **The server validates `](uuid)` links on every write** (v1.7.0): a link to a nonexistent id rejects the write, naming the offender — that means you mangled the UUID; re-read the source and correct it, do not retry unchanged. Example ids go in backticks (code is not validated). `[[Wikilinks]]` may dangle. See `AGENT_GUIDE.md → Writing linkable content` for the full rule.\n10. **Concurrency: content updates require `expected_content_hash`.** Pass the `content_hash` you last saw — every read shows one (`cerefox_get_document` incl. outline mode, `cerefox_search`, `cerefox_metadata_search`) and **every write returns the new one, including create** (v1.3.0, #189), so after writing you already hold the token for your next edit; no re-read needed. If it's stale you get a **conflict** — re-read the document, merge your changes into the latest content, retry with the new hash. **Never resolve a conflict by overwriting blindly** — the current content includes another writer's work. `last_write_wins: true` skips the check; use it ONLY when an external source of truth makes conflicts meaningless (file re-sync), never to silence a conflict.\n11. **Search: prefer a few distinctive terms; heed `below confidence`.** When nothing clears the relevance threshold, `cerefox_search` returns the closest candidates prefixed with a `below confidence` warning instead of an empty set — that flag means **weak signal, not absent knowledge**: check the candidates' scores and titles before concluding the KB lacks the content. A truly empty response means nothing even weakly related exists.\n12. **Relations express how documents relate; lifecycle tells you if knowledge is still good.** Use `cerefox_set_relation` when one document supersedes, contradicts, references, or continues another. `supersedes` marks the target **superseded**; `contradicts` marks **both** stale; `related_to`/`duplicates`/`contradicts` are symmetric (both directions written). Any other type string is accepted without special behavior. When a search result or `cerefox_get_relations` shows a neighbour marked `[superseded]` or `[stale]`, say so rather than presenting it as current.\n13. **Project memberships — non-destructive by default**: on `cerefox_ingest` updates, **`project_name` (singular) is a non-destructive add** (ensures membership, preserves others). Use **`project_names` (list)** when you want to set the doc's full project set in one call (destructive replace). For metadata-only project changes without writing content, use **`cerefox_set_document_projects(document_id, project_names)`** — that tool is the destructive-replace contract made explicit. Never call `cerefox_set_document_projects` with a single name when you mean \"add\" — that would REMOVE the doc from all other projects. When in doubt, use `cerefox_ingest` with singular `project_name`.\n\n## Update Workflow (ID-based -- preferred)\n\n```\nsearch(\"topic\") -> find doc [id: abc123] -> get_document(abc123) -> note its content_hash -> modify ->\ningest(title=\"Same Title\", content=\"...\", document_id=\"abc123\",\n       expected_content_hash=\"<the hash you read>\", author=\"my-agent\")\n```\n\nOn a **conflict** error: get_document again (fresh content + fresh hash) -> merge your changes -> retry with the new hash.\n\n## Update Workflow (title-based -- fallback)\n\n```\nsearch(\"topic\") -> find doc (note its hash) -> modify ->\ningest(title=\"Same Title\", content=\"...\", update_if_exists=true,\n       expected_content_hash=\"<the hash you read>\", author=\"my-agent\")\n```\n\n## Catch-Up Workflow\n\n```\nmetadata_search(metadata_filter={\"type\": \"decision-log\"}, updated_since=\"2026-03-28T00:00:00Z\")\n```\n\n## CLI fallback (when MCP is unavailable)\n\nIf `cerefox_search` is not in your tool list, your user has likely installed the Cerefox CLI. The canonical invocation is plain **`cerefox <subcommand>`** (the TypeScript CLI, installed via `npm install -g @cerefox/memory`). It uses a resource-verb shape (`cerefox document get`, `cerefox project list`, …).\n\nSame operations, same conventions. Full reference: [`docs/guides/cli.md`](docs/guides/cli.md). CLI flag names match MCP parameter names exactly (e.g. `metadata_filter` ↔ `--metadata-filter`); common flags also have single-letter short forms (`-f`, `-p`, `-c`, `-m`, `-u`, `-a`). Use the canonical long name (what `--help` shows) or its short form — there are no long-form aliases like `--filter` or `--count`.\n\n| MCP tool | CLI |\n|---|---|\n| `cerefox_search` | `cerefox search \"<q>\" --author \"<your-name>\"` |\n| `cerefox_ingest` (paste) | `printf '...' \\| cerefox document ingest --paste --title \"<t>\" --author \"<your-name>\" --author-type agent` |\n| `cerefox_ingest` (update by ID) | `printf '...' \\| cerefox document ingest --paste --title \"<t>\" --document-id \"<uuid>\" --expected-content-hash \"<hash>\" --author \"<your-name>\" --author-type agent` |\n| `cerefox_get_document` | `cerefox document get <id> --version-id <vid> --author \"<your-name>\"` |\n| `cerefox_list_versions` | `cerefox document version list <id> --author \"<your-name>\"` |\n| `cerefox_list_projects` | `cerefox project list --author \"<your-name>\"` |\n| `cerefox_list_metadata_keys` | `cerefox metadata keys` |\n| `cerefox_insert` | `cerefox document insert <id> -t \"<text>\" -p <position> -a \"<anchor-heading>\" -e \"<hash>\" --author \"<your-name>\" --author-type agent` |\n| `cerefox_edit` | `cerefox document edit-parts <id> --operations '<json>' -e \"<hash>\" --author \"<your-name>\" --author-type agent` |\n| `cerefox_delete_document` | `cerefox document delete <id> --reason \"<why>\" --author \"<your-name>\" --author-type agent --yes` (confirms interactively instead of requiring the hash) |\n| `cerefox_restore_document` | `cerefox document restore <id> --reason \"<why>\" --author \"<your-name>\" --author-type agent` |\n| `cerefox_metadata_search` | `cerefox metadata search --metadata-filter '<json>' --author \"<your-name>\"` (list a project: `cerefox document list --project <name>`) |\n| `cerefox_set_document_metadata` | `cerefox document set-metadata <id> --set key=value` (also `--remove key`, `--json '{...}'`, `--replace`) |\n| `cerefox_set_document_projects` | `cerefox document set-projects <id> <name...> --author \"<your-name>\" --author-type agent` (or `--clear` to remove all) |\n| `cerefox_get_audit_log` | `cerefox audit list --author \"<your-name>\"` (add `--json` for scripted access) |\n| `cerefox_get_help` | `cerefox guides show agent-quick-reference` (or `cerefox guides list` for the full bundled-docs index) |\n\n**Set identity on every call**, exactly as you would on MCP:\n- Writes (`document ingest`, `document ingest-dir`): `--author \"<your-name>\" --author-type agent`\n- Reads: `--author \"<your-name>\"` (the same flag; `--requestor` still works as a hidden alias)\n\nOr have your user set `CEREFOX_AUTHOR_NAME` / `CEREFOX_AUTHOR_TYPE` / `CEREFOX_REQUESTOR_NAME` in their `.env` to apply defaults once.\n\n## Timestamps are UTC\n\nEvery timestamp Cerefox returns — `created_at` on audit entries, version\nhistory, document metadata — is **UTC**, and now carries its `Z` marker so it\ncannot be mistaken for local time.\n\n**When you write a date into a document's CONTENT, use your own clock, not a\nCerefox timestamp.** These are different things: a timestamp records when the\nserver stored something; a date in a log entry or a heading is authored content\nand belongs to your timezone. An agent working a Pacific afternoon read\n`2026-08-11` from version history, wrote \"8/11\" into its entries, and put a\nday's work in the future — the timestamp was correct, and copying it into\ncontent was not.\n\nCerefox deliberately does not convert to local time on the API or MCP paths.\n\"Local\" has no server-side meaning: the remote MCP server runs in a cloud\nfunction whose local time *is* UTC, while a local MCP server runs in yours, so\nthe same document would report two different times depending on transport. The\nweb UI converts because a browser knows the viewer's timezone; nothing\nserver-side does.\n\n## Mistakes that have actually happened\n\nEach of these comes from a real agent session, and each is easy to make.\n\n- **`cerefox_ingest` always replaces the ENTIRE document.** Never a section.\n  Before sending, check that the tool name matches the intent: if the intent is\n  \"change one section\", the call is `cerefox_edit` with `replace_section`. A\n  section-sized edit sent as a full ingest truncated a 13,000-character index to\n  a single word. It was recovered from version history within the minute, but\n  only because it was noticed immediately.\n\n- **Do not include the anchor's own heading in your text.** `replace_section`\n  keeps the heading and `insert` places your text inside the section, so\n  including it produces two. This is now refused rather than silently applied,\n  but the shape is worth knowing: it happened twice in one session, the second\n  time while trying to repair the first. A *deeper* sub-heading inside your text\n  is fine.\n\n- **Content between sections belongs to the section ABOVE it.** A section runs\n  to the next heading of the same or higher level, so a `---` rule, a note, or\n  any trailing text sitting just above the next heading is part of the section\n  before it — even when it visually reads as belonging below. Replacing that\n  section takes it too. An agent hit exactly this: a `---` that separated two\n  major sections disappeared when the section above it was replaced. The write\n  was correct by the addressing rules; the surprise is that \"the end of this\n  section\" is further down the page than it looks. Note the loss warning will\n  not catch it if your replacement text is longer than what it replaced, since\n  there is then no net loss to report.\n\n- **To change only tags, use `cerefox_set_document_metadata`, never `cerefox_ingest`.**\n  Ingest replaces the whole document, so re-sending it to set one tag carries the\n  full transcription risk for no reason. The metadata tool merges: the keys you\n  pass are set, everything else is left alone, so you do not need to read the\n  document first and cannot drop a tag another agent set. Pass `null` as a value\n  to remove a key.\n\n- **Never partial-edit to fix a partial edit.** If a write leaves unexpected\n  structure, stop. Use `cerefox_list_versions`, retrieve the last good version,\n  and re-ingest cleanly. Repairing edits with more edits compounds the damage.\n\n- **A rejected batch is safe.** Operations in one `cerefox_edit` are\n  all-or-nothing: if any is invalid, nothing is written. A refusal costs you a\n  retry, not data — so prefer one call for changes that belong together, and do\n  not split a batch to \"make it more likely to succeed\".\n\n- **Read before replacing.** `cerefox_get_document(section: \"## Heading\")`\n  returns exactly what a `replace_section` on that anchor would overwrite. Use it\n  for any section you did not write in this session. The outline gives a\n  section's *size*, never its *text*.\n\n- **Verify after writing** — read the result back before reporting success, and\n  report what the read actually shows.\n\n- **Partial edits cannot change a document's stored TITLE.** `rename_section`\n  changes a heading inside the content; the title is a separate field and still\n  needs `cerefox_ingest`.\n\n- **If a capability seems missing from one server, suspect your client first.**\n  Local and remote run the same code. **Every `cerefox_get_help()` response\n  begins with the server's version and the operations it registers** — you do\n  not need a special topic, and the *absence* of that block is itself an answer:\n  a server that does not print it predates v1.5.0. If that\n  disagrees with your tool list, the client is holding a list it fetched before\n  an upgrade — clients cache it at connect time. Ask the user to restart the\n  client. Do not record a capability difference between servers as a fact; every\n  such report so far has been a stale client.\n- Long inline bodies can arrive with literal `\\n`/`\\\"` (the author over-escaped; Cerefox stores bytes faithfully). For long or quote-dense content, ingest from a file or build incrementally with `cerefox_insert`; read back multi-line writes.\n";
+export const HELP_FULL = "# Cerefox Knowledge Base -- Agent Quick Reference\n\nCerefox is a persistent, shared knowledge base. You have **15 core MCP tools** (14 with CLI equivalents — `cerefox_get_help` is MCP-only), plus 4 dormant relation tools that appear only when `relations_enabled` is on. For the full guide, search Cerefox for \"How AI Agents Use Cerefox\" or call `cerefox_get_help` to retrieve this content over MCP.\n\n## Tools\n\n| Tool | Purpose | Key params |\n|------|---------|------------|\n| `cerefox_search` | Find documents (hybrid FTS + semantic) | `query` (required), `project_name`, `metadata_filter`, `author` |\n| `cerefox_ingest` | Save or update a document | `title`, `content` (required), `document_id` (update by ID), `expected_content_hash` (**required on content updates** — see rule 10), `last_write_wins`, `update_if_exists`, `project_name` (single, non-destructive add on update), `project_names` (list, destructive replace on update), `metadata` (on update it **merges**: `null` removes a key, omitted keys are kept), `expected_metadata_revision` (optional, see rule 10), `author` |\n| `cerefox_insert` | **Add** to a document without resending it. Cannot destroy content. | `document_id`, `text`, `position` (`end_of_document`/`end_of_section`/`after_heading`/`before_heading`), `expected_content_hash` (required), `anchor_heading` (unless `end_of_document`), `section_part`, `author` |\n| `cerefox_edit` | **Change** parts of a document: 1..n operations applied atomically | `document_id`, `operations` (`insert`/`replace_section`/`delete_section`/`rename_section`), `expected_content_hash` (required), `author` |\n| `cerefox_delete_document` | **Soft**-delete a document (to trash; excluded from search; permanent purge is human-only) | `document_id`, `expected_content_hash` (**required** — a delete must follow a read), `reason` (recorded in the audit log — give one), `author` |\n| `cerefox_restore_document` | Restore a soft-deleted document from the trash (audited inverse of delete; no-op if not deleted) | `document_id` (required), `reason` (recorded in the audit log), `author` |\n| `cerefox_get_document` | Get full document by ID (header includes `content_hash` — the update token), or with `outline: true` just its heading paths, sizes and hash, or with `section: \"## Heading\"` one section's text | `document_id` (required), `outline`, `section`, `section_part`, `author` |\n| `cerefox_list_versions` | Version history of a document | `document_id` (required), `author` |\n| `cerefox_set_relation` ⚑ | Link two documents (`source --rel_type--> target`) | `source_id`, `target_id`, `rel_type` (required), `metadata`, `author` |\n| `cerefox_delete_relation` ⚑ | Remove a relation | `source_id`, `target_id`, `rel_type`, `author` |\n| `cerefox_get_relations` ⚑ | All relations touching a document, both directions | `document_id` |\n| `cerefox_get_neighbors` ⚑ | Walk the graph along ONE relation type | `document_id`, `rel_type` (required), `depth`, `from_time`, `to_time`, `limit` |\n| `cerefox_metadata_search` | Find or list docs by metadata, project, or time (no text query) | `metadata_filter`, `project_name` (list a project's docs), `updated_since`, `include_content` — **at least one** of metadata_filter/project_name/updated_since/created_since |\n| `cerefox_list_metadata_keys` | Discover available metadata keys | (none required) |\n| `cerefox_list_projects` | List all projects | (none required) |\n| `cerefox_set_document_metadata` | Change tags WITHOUT resending content. **Merges** by default; a `null` value removes a key | `document_id`, `metadata` (required), `replace` (rare: set exactly this object; needs `expected_metadata_revision`), `expected_metadata_revision`, `author` |\n| `cerefox_set_document_projects` | Set doc's project memberships to exactly the given list (destructive replace; metadata-only, no content change) | `document_id`, `project_names` (required), `author` |\n| `cerefox_get_audit_log` | Query write operation history | `document_id`, `by_author` (filter), `operation`, `since`, `author` |\n| `cerefox_get_help` | Retrieve Cerefox conventions (this reference) over MCP. **Call this whenever uncertain.** | `topic` (optional, case-insensitive H2 substring match) |\n\n⚑ **Opt-in — usually absent.** The four relation tools are hidden unless the\noperator enables them (`relations_enabled`). **Trust your own tool list**: if\nthey are not in it, the feature is switched off for this deployment. That is\nnormal, not an error, and not something to work around.\n\n## Editing part of a document (prefer this over re-sending)\n\n**Re-sending a whole document to change part of it is the main way agents lose\ndata.** You have to reproduce the untouched remainder verbatim, and any drift\nsilently rewrites content nobody asked you to touch — which the caller cannot\ndiff. Use the partial-edit tools instead:\n\n1. **Learn the anchors** — `cerefox_get_document(document_id, outline: true)`.\n   Returns heading paths, per-section sizes and the `content_hash`, without the\n   body. The paths it returns are exactly what `anchor_heading` accepts.\n2. **Add** → `cerefox_insert`. `end_of_document` is a plain append;\n   `end_of_section` adds inside a named section. It is structurally incapable of\n   removing anything, so \"I meant to append\" cannot become \"I replaced the file\".\n3. **Look before you overwrite** — `cerefox_get_document(document_id,\n   section: \"## Heading\")` returns exactly the text a `replace_section` on that\n   anchor would destroy. The outline gives you a section's *size*, never its\n   *text*, so on a document you did not write yourself this is the difference\n   between a replace and a blind overwrite.\n4. **Change or remove** → `cerefox_edit`. Put changes that belong together in\n   ONE call: they apply atomically, so a table row and the total it feeds cannot\n   end up disagreeing. To change a single line, `replace_section` on its\n   smallest enclosing heading — that is the intended granularity, not a\n   workaround. To fix a stale heading (`## OPEN TODOs (as of ...)`), use\n   `rename_section`: it changes the heading text and leaves the body and\n   position alone.\n5. All of them require `expected_content_hash` and **have no last-write-wins**. A\n   conflict means someone else changed the document; re-read and decide, do not\n   force it.\n\n**A section runs to the next same-or-higher heading, or to the end of the\ndocument.** So `end_of_document` inserts land inside the *last* section, and\nreplacing or deleting that section removes them too. A large shrink in the\nresponse is your warning; `cerefox_list_versions` has the previous content.\n\n**When an anchor is ambiguous the tool refuses and hands you the options** — a\nrepeated heading returns the qualifying paths, and a section with both its own\ncontent and sub-sections returns both `section_part` choices. That is a\nrecoverable answer, not a failure: retry with what it gave you.\n\n## Essential Rules\n\n1. **Search before ingesting** -- check if the document exists first.\n2. **Prefer ID-based updates** -- pass `document_id` from search results for deterministic updates. Falls back to title-matching with `update_if_exists: true`.\n3. **Set `author`** to your name on every call, reads and writes alike (e.g., \"Claude Code\", \"archiver\"). Same parameter on every tool. (`requestor` is still accepted everywhere as the pre-1.13.1 alias.) On the CLI it is `--author` on every command too (plus `--author-type` on writes); or rely on `CEREFOX_AUTHOR_NAME`/`CEREFOX_AUTHOR_TYPE` env vars set in the user's `.env`.\n4. **Use `document_id` from search results** `[id: uuid]` for get_document and list_versions.\n5. **`max_bytes` has to fit a whole document.** Search returns COMPLETE documents, so a budget smaller than the top hit fits nothing. You get a header-only list saying so (never \"No results found.\" — that means the store really has nothing). Raise `max_bytes`, or read one document with `cerefox_get_document` using `outline: true` or `section`.\n6. **Add metadata** -- at minimum `type` (\"decision-log\", \"research\", \"design-doc\") and `status` (\"active\", \"draft\").\n7. **Write for retrieval: the title does real work.** The document title is indexed at the **highest weight** *and* prepended to every chunk before it is embedded, so it steers both halves of hybrid search (keyword and semantic) for every chunk in the document. Name the document with the distinctive terms someone would actually search for -- \"Cerefox Response Size Limits\", not \"Notes\" or \"Update 3\". Renaming re-indexes and re-embeds, so correcting a vague title is cheap and takes effect immediately. Write the characters themselves: `&` and `<>`, never `&amp;` or `&lt;&gt;` (entities are stored literally and the write reply will say so). Inside the content, write structured Markdown with H1/H2/H3 headings: the chunker follows them, and a chunk's own heading carries the same weight as the title.\n8. **Deletes are soft (recoverable); permanent purge is web-UI-only.** `cerefox_delete_document` requires the document's `content_hash` as you read it (read before you delete) and takes a `reason` — give one; it is what the human reviewing the trash sees. `cerefox_restore_document` undoes a mistaken delete (also audited, also takes a `reason`). Always surface deletes AND restores to the user. Once a human purges from the web UI, the document is gone for good; so is one auto-purged after `trash_retention_days`, if the operator has turned trash auto-purge on.\n9. **Cross-doc links inside content**: **always use `[Text](document-uuid)`.** UUIDs are the only fully reliable link form — stable across title changes, never ambiguous, no encoding gotchas. Every `cerefox_search` result shows `[id: <uuid>]` after the title; grab it and use it. Title-based linking (`[Text](<Title With Spaces>)`) is fragile (breaks on colons, parens, ampersands, brackets — silently navigates to wrong page) — **don't write title-based links**; do an extra search to get the UUID instead. Repo-path forms (`[Text](docs/path.md)`) exist for repo-ingested files; don't construct manually. **The server validates `](uuid)` links on every write** (v1.7.0): a link to a nonexistent id rejects the write, naming the offender — that means you mangled the UUID; re-read the source and correct it, do not retry unchanged. Example ids go in backticks (code is not validated). `[[Wikilinks]]` may dangle. See `AGENT_GUIDE.md → Writing linkable content` for the full rule.\n10. **Concurrency: content updates require `expected_content_hash`.** Pass the `content_hash` you last saw — every read shows one (`cerefox_get_document` incl. outline mode, `cerefox_search`, `cerefox_metadata_search`) and **every write returns the new one, including create** (v1.3.0, #189), so after writing you already hold the token for your next edit; no re-read needed. If it's stale you get a **conflict** — re-read the document, merge your changes into the latest content, retry with the new hash. **Never resolve a conflict by overwriting blindly** — the current content includes another writer's work. `last_write_wins: true` skips the check; use it ONLY when an external source of truth makes conflicts meaningless (file re-sync), never to silence a conflict. **Metadata has its own token, `metadata_revision`** (shown next to `content_hash` wherever it appears). It is optional on a merge (pass it when your change depends on the current values, e.g. a status transition) and **required with `replace: true`**. A stale one is a *metadata conflict*: re-read, reapply, retry.\n11. **Search: prefer a few distinctive terms; heed `below confidence`.** When nothing clears the relevance threshold, `cerefox_search` returns the closest candidates prefixed with a `below confidence` warning instead of an empty set — that flag means **weak signal, not absent knowledge**: check the candidates' scores and titles before concluding the KB lacks the content. A truly empty response means nothing even weakly related exists.\n12. **Relations express how documents relate; lifecycle tells you if knowledge is still good.** Use `cerefox_set_relation` when one document supersedes, contradicts, references, or continues another. `supersedes` marks the target **superseded**; `contradicts` marks **both** stale; `related_to`/`duplicates`/`contradicts` are symmetric (both directions written). Any other type string is accepted without special behavior. When a search result or `cerefox_get_relations` shows a neighbour marked `[superseded]` or `[stale]`, say so rather than presenting it as current.\n13. **Project memberships — non-destructive by default**: on `cerefox_ingest` updates, **`project_name` (singular) is a non-destructive add** (ensures membership, preserves others). Use **`project_names` (list)** when you want to set the doc's full project set in one call (destructive replace). For metadata-only project changes without writing content, use **`cerefox_set_document_projects(document_id, project_names)`** — that tool is the destructive-replace contract made explicit. Never call `cerefox_set_document_projects` with a single name when you mean \"add\" — that would REMOVE the doc from all other projects. When in doubt, use `cerefox_ingest` with singular `project_name`.\n\n## Update Workflow (ID-based -- preferred)\n\n```\nsearch(\"topic\") -> find doc [id: abc123] -> get_document(abc123) -> note its content_hash -> modify ->\ningest(title=\"Same Title\", content=\"...\", document_id=\"abc123\",\n       expected_content_hash=\"<the hash you read>\", author=\"my-agent\")\n```\n\nOn a **conflict** error: get_document again (fresh content + fresh hash) -> merge your changes -> retry with the new hash.\n\n## Update Workflow (title-based -- fallback)\n\n```\nsearch(\"topic\") -> find doc (note its hash) -> modify ->\ningest(title=\"Same Title\", content=\"...\", update_if_exists=true,\n       expected_content_hash=\"<the hash you read>\", author=\"my-agent\")\n```\n\n## Catch-Up Workflow\n\n```\nmetadata_search(metadata_filter={\"type\": \"decision-log\"}, updated_since=\"2026-03-28T00:00:00Z\")\n```\n\n## CLI fallback (when MCP is unavailable)\n\nIf `cerefox_search` is not in your tool list, your user has likely installed the Cerefox CLI. The canonical invocation is plain **`cerefox <subcommand>`** (the TypeScript CLI, installed via `npm install -g @cerefox/memory`). It uses a resource-verb shape (`cerefox document get`, `cerefox project list`, …).\n\nSame operations, same conventions. Full reference: [`docs/guides/cli.md`](docs/guides/cli.md). CLI flag names match MCP parameter names exactly (e.g. `metadata_filter` ↔ `--metadata-filter`); common flags also have single-letter short forms (`-f`, `-p`, `-c`, `-m`, `-u`, `-a`). Use the canonical long name (what `--help` shows) or its short form — there are no long-form aliases like `--filter` or `--count`.\n\n| MCP tool | CLI |\n|---|---|\n| `cerefox_search` | `cerefox search \"<q>\" --author \"<your-name>\"` |\n| `cerefox_ingest` (paste) | `printf '...' \\| cerefox document ingest --paste --title \"<t>\" --author \"<your-name>\" --author-type agent` |\n| `cerefox_ingest` (update by ID) | `printf '...' \\| cerefox document ingest --paste --title \"<t>\" --document-id \"<uuid>\" --expected-content-hash \"<hash>\" --author \"<your-name>\" --author-type agent` |\n| `cerefox_get_document` | `cerefox document get <id> --version-id <vid> --author \"<your-name>\"` |\n| `cerefox_list_versions` | `cerefox document version list <id> --author \"<your-name>\"` |\n| `cerefox_list_projects` | `cerefox project list --author \"<your-name>\"` |\n| `cerefox_list_metadata_keys` | `cerefox metadata keys` |\n| `cerefox_insert` | `cerefox document insert <id> -t \"<text>\" -p <position> -a \"<anchor-heading>\" -e \"<hash>\" --author \"<your-name>\" --author-type agent` |\n| `cerefox_edit` | `cerefox document edit-parts <id> --operations '<json>' -e \"<hash>\" --author \"<your-name>\" --author-type agent` |\n| `cerefox_delete_document` | `cerefox document delete <id> --reason \"<why>\" --author \"<your-name>\" --author-type agent --yes` (confirms interactively instead of requiring the hash) |\n| `cerefox_restore_document` | `cerefox document restore <id> --reason \"<why>\" --author \"<your-name>\" --author-type agent` |\n| `cerefox_metadata_search` | `cerefox metadata search --metadata-filter '<json>' --author \"<your-name>\"` (list a project: `cerefox document list --project <name>`) |\n| `cerefox_set_document_metadata` | `cerefox document set-metadata <id> --set key=value` (also `--remove key`, `--json '{...}'`, `--replace`) |\n| `cerefox_set_document_projects` | `cerefox document set-projects <id> <name...> --author \"<your-name>\" --author-type agent` (or `--clear` to remove all) |\n| `cerefox_get_audit_log` | `cerefox audit list --author \"<your-name>\"` (add `--json` for scripted access) |\n| `cerefox_get_help` | `cerefox guides show agent-quick-reference` (or `cerefox guides list` for the full bundled-docs index) |\n\n**Set identity on every call**, exactly as you would on MCP:\n- Writes (`document ingest`, `document ingest-dir`): `--author \"<your-name>\" --author-type agent`\n- Reads: `--author \"<your-name>\"` (the same flag; `--requestor` still works as a hidden alias)\n\nOr have your user set `CEREFOX_AUTHOR_NAME` / `CEREFOX_AUTHOR_TYPE` / `CEREFOX_REQUESTOR_NAME` in their `.env` to apply defaults once.\n\n## Timestamps are UTC\n\nEvery timestamp Cerefox returns — `created_at` on audit entries, version\nhistory, document metadata — is **UTC**, and now carries its `Z` marker so it\ncannot be mistaken for local time.\n\n**When you write a date into a document's CONTENT, use your own clock, not a\nCerefox timestamp.** These are different things: a timestamp records when the\nserver stored something; a date in a log entry or a heading is authored content\nand belongs to your timezone. An agent working a Pacific afternoon read\n`2026-08-11` from version history, wrote \"8/11\" into its entries, and put a\nday's work in the future — the timestamp was correct, and copying it into\ncontent was not.\n\nCerefox deliberately does not convert to local time on the API or MCP paths.\n\"Local\" has no server-side meaning: the remote MCP server runs in a cloud\nfunction whose local time *is* UTC, while a local MCP server runs in yours, so\nthe same document would report two different times depending on transport. The\nweb UI converts because a browser knows the viewer's timezone; nothing\nserver-side does.\n\n## Mistakes that have actually happened\n\nEach of these comes from a real agent session, and each is easy to make.\n\n- **`cerefox_ingest` always replaces the ENTIRE document.** Never a section.\n  Before sending, check that the tool name matches the intent: if the intent is\n  \"change one section\", the call is `cerefox_edit` with `replace_section`. A\n  section-sized edit sent as a full ingest truncated a 13,000-character index to\n  a single word. It was recovered from version history within the minute, but\n  only because it was noticed immediately.\n\n- **Do not include the anchor's own heading in your text.** `replace_section`\n  keeps the heading and `insert` places your text inside the section, so\n  including it produces two. This is now refused rather than silently applied,\n  but the shape is worth knowing: it happened twice in one session, the second\n  time while trying to repair the first. A *deeper* sub-heading inside your text\n  is fine.\n\n- **Content between sections belongs to the section ABOVE it.** A section runs\n  to the next heading of the same or higher level, so a `---` rule, a note, or\n  any trailing text sitting just above the next heading is part of the section\n  before it — even when it visually reads as belonging below. Replacing that\n  section takes it too. An agent hit exactly this: a `---` that separated two\n  major sections disappeared when the section above it was replaced. The write\n  was correct by the addressing rules; the surprise is that \"the end of this\n  section\" is further down the page than it looks. Note the loss warning will\n  not catch it if your replacement text is longer than what it replaced, since\n  there is then no net loss to report.\n\n- **To change only tags, use `cerefox_set_document_metadata`, never `cerefox_ingest`.**\n  Ingest replaces the whole document, so re-sending it to set one tag carries the\n  full transcription risk for no reason. The metadata tool merges: the keys you\n  pass are set, everything else is left alone, so you do not need to read the\n  document first and cannot drop a tag another agent set. Pass `null` as a value\n  to remove a key.\n\n- **Never partial-edit to fix a partial edit.** If a write leaves unexpected\n  structure, stop. Use `cerefox_list_versions`, retrieve the last good version,\n  and re-ingest cleanly. Repairing edits with more edits compounds the damage.\n\n- **A rejected batch is safe.** Operations in one `cerefox_edit` are\n  all-or-nothing: if any is invalid, nothing is written. A refusal costs you a\n  retry, not data — so prefer one call for changes that belong together, and do\n  not split a batch to \"make it more likely to succeed\".\n\n- **Read before replacing.** `cerefox_get_document(section: \"## Heading\")`\n  returns exactly what a `replace_section` on that anchor would overwrite. Use it\n  for any section you did not write in this session. The outline gives a\n  section's *size*, never its *text*.\n\n- **Verify after writing** — read the result back before reporting success, and\n  report what the read actually shows.\n\n- **Partial edits cannot change a document's stored TITLE.** `rename_section`\n  changes a heading inside the content; the title is a separate field and still\n  needs `cerefox_ingest`.\n\n- **If a capability seems missing from one server, suspect your client first.**\n  Local and remote run the same code. **Every `cerefox_get_help()` response\n  begins with the server's version and the operations it registers** — you do\n  not need a special topic, and the *absence* of that block is itself an answer:\n  a server that does not print it predates v1.5.0. If that\n  disagrees with your tool list, the client is holding a list it fetched before\n  an upgrade — clients cache it at connect time. Ask the user to restart the\n  client. Do not record a capability difference between servers as a fact; every\n  such report so far has been a stale client.\n- Long inline bodies can arrive with literal `\\n`/`\\\"` (the author over-escaped; Cerefox stores bytes faithfully). For long or quote-dense content, ingest from a file or build incrementally with `cerefox_insert`; read back multi-line writes.\n";
 
 /** Sections keyed by their H2 heading text (lower-cased for matching). */
 export const HELP_SECTIONS: Record<string, string> = {
-  "Tools": "## Tools\n\n| Tool | Purpose | Key params |\n|------|---------|------------|\n| `cerefox_search` | Find documents (hybrid FTS + semantic) | `query` (required), `project_name`, `metadata_filter`, `author` |\n| `cerefox_ingest` | Save or update a document | `title`, `content` (required), `document_id` (update by ID), `expected_content_hash` (**required on content updates** — see rule 10), `last_write_wins`, `update_if_exists`, `project_name` (single, non-destructive add on update), `project_names` (list, destructive replace on update), `metadata` (omit on update to keep existing tags; `{}` clears), `author` |\n| `cerefox_insert` | **Add** to a document without resending it. Cannot destroy content. | `document_id`, `text`, `position` (`end_of_document`/`end_of_section`/`after_heading`/`before_heading`), `expected_content_hash` (required), `anchor_heading` (unless `end_of_document`), `section_part`, `author` |\n| `cerefox_edit` | **Change** parts of a document: 1..n operations applied atomically | `document_id`, `operations` (`insert`/`replace_section`/`delete_section`/`rename_section`), `expected_content_hash` (required), `author` |\n| `cerefox_delete_document` | **Soft**-delete a document (to trash; excluded from search; permanent purge is human-only) | `document_id`, `expected_content_hash` (**required** — a delete must follow a read), `reason` (recorded in the audit log — give one), `author` |\n| `cerefox_restore_document` | Restore a soft-deleted document from the trash (audited inverse of delete; no-op if not deleted) | `document_id` (required), `reason` (recorded in the audit log), `author` |\n| `cerefox_get_document` | Get full document by ID (header includes `content_hash` — the update token), or with `outline: true` just its heading paths, sizes and hash, or with `section: \"## Heading\"` one section's text | `document_id` (required), `outline`, `section`, `section_part`, `author` |\n| `cerefox_list_versions` | Version history of a document | `document_id` (required), `author` |\n| `cerefox_set_relation` ⚑ | Link two documents (`source --rel_type--> target`) | `source_id`, `target_id`, `rel_type` (required), `metadata`, `author` |\n| `cerefox_delete_relation` ⚑ | Remove a relation | `source_id`, `target_id`, `rel_type`, `author` |\n| `cerefox_get_relations` ⚑ | All relations touching a document, both directions | `document_id` |\n| `cerefox_get_neighbors` ⚑ | Walk the graph along ONE relation type | `document_id`, `rel_type` (required), `depth`, `from_time`, `to_time`, `limit` |\n| `cerefox_metadata_search` | Find or list docs by metadata, project, or time (no text query) | `metadata_filter`, `project_name` (list a project's docs), `updated_since`, `include_content` — **at least one** of metadata_filter/project_name/updated_since/created_since |\n| `cerefox_list_metadata_keys` | Discover available metadata keys | (none required) |\n| `cerefox_list_projects` | List all projects | (none required) |\n| `cerefox_set_document_metadata` | Change tags WITHOUT resending content. **Merges** by default; a `null` value removes a key | `document_id`, `metadata` (required), `replace` (rare: set exactly this object), `author` |\n| `cerefox_set_document_projects` | Set doc's project memberships to exactly the given list (destructive replace; metadata-only, no content change) | `document_id`, `project_names` (required), `author` |\n| `cerefox_get_audit_log` | Query write operation history | `document_id`, `by_author` (filter), `operation`, `since`, `author` |\n| `cerefox_get_help` | Retrieve Cerefox conventions (this reference) over MCP. **Call this whenever uncertain.** | `topic` (optional, case-insensitive H2 substring match) |\n\n⚑ **Opt-in — usually absent.** The four relation tools are hidden unless the\noperator enables them (`relations_enabled`). **Trust your own tool list**: if\nthey are not in it, the feature is switched off for this deployment. That is\nnormal, not an error, and not something to work around.",
+  "Tools": "## Tools\n\n| Tool | Purpose | Key params |\n|------|---------|------------|\n| `cerefox_search` | Find documents (hybrid FTS + semantic) | `query` (required), `project_name`, `metadata_filter`, `author` |\n| `cerefox_ingest` | Save or update a document | `title`, `content` (required), `document_id` (update by ID), `expected_content_hash` (**required on content updates** — see rule 10), `last_write_wins`, `update_if_exists`, `project_name` (single, non-destructive add on update), `project_names` (list, destructive replace on update), `metadata` (on update it **merges**: `null` removes a key, omitted keys are kept), `expected_metadata_revision` (optional, see rule 10), `author` |\n| `cerefox_insert` | **Add** to a document without resending it. Cannot destroy content. | `document_id`, `text`, `position` (`end_of_document`/`end_of_section`/`after_heading`/`before_heading`), `expected_content_hash` (required), `anchor_heading` (unless `end_of_document`), `section_part`, `author` |\n| `cerefox_edit` | **Change** parts of a document: 1..n operations applied atomically | `document_id`, `operations` (`insert`/`replace_section`/`delete_section`/`rename_section`), `expected_content_hash` (required), `author` |\n| `cerefox_delete_document` | **Soft**-delete a document (to trash; excluded from search; permanent purge is human-only) | `document_id`, `expected_content_hash` (**required** — a delete must follow a read), `reason` (recorded in the audit log — give one), `author` |\n| `cerefox_restore_document` | Restore a soft-deleted document from the trash (audited inverse of delete; no-op if not deleted) | `document_id` (required), `reason` (recorded in the audit log), `author` |\n| `cerefox_get_document` | Get full document by ID (header includes `content_hash` — the update token), or with `outline: true` just its heading paths, sizes and hash, or with `section: \"## Heading\"` one section's text | `document_id` (required), `outline`, `section`, `section_part`, `author` |\n| `cerefox_list_versions` | Version history of a document | `document_id` (required), `author` |\n| `cerefox_set_relation` ⚑ | Link two documents (`source --rel_type--> target`) | `source_id`, `target_id`, `rel_type` (required), `metadata`, `author` |\n| `cerefox_delete_relation` ⚑ | Remove a relation | `source_id`, `target_id`, `rel_type`, `author` |\n| `cerefox_get_relations` ⚑ | All relations touching a document, both directions | `document_id` |\n| `cerefox_get_neighbors` ⚑ | Walk the graph along ONE relation type | `document_id`, `rel_type` (required), `depth`, `from_time`, `to_time`, `limit` |\n| `cerefox_metadata_search` | Find or list docs by metadata, project, or time (no text query) | `metadata_filter`, `project_name` (list a project's docs), `updated_since`, `include_content` — **at least one** of metadata_filter/project_name/updated_since/created_since |\n| `cerefox_list_metadata_keys` | Discover available metadata keys | (none required) |\n| `cerefox_list_projects` | List all projects | (none required) |\n| `cerefox_set_document_metadata` | Change tags WITHOUT resending content. **Merges** by default; a `null` value removes a key | `document_id`, `metadata` (required), `replace` (rare: set exactly this object; needs `expected_metadata_revision`), `expected_metadata_revision`, `author` |\n| `cerefox_set_document_projects` | Set doc's project memberships to exactly the given list (destructive replace; metadata-only, no content change) | `document_id`, `project_names` (required), `author` |\n| `cerefox_get_audit_log` | Query write operation history | `document_id`, `by_author` (filter), `operation`, `since`, `author` |\n| `cerefox_get_help` | Retrieve Cerefox conventions (this reference) over MCP. **Call this whenever uncertain.** | `topic` (optional, case-insensitive H2 substring match) |\n\n⚑ **Opt-in — usually absent.** The four relation tools are hidden unless the\noperator enables them (`relations_enabled`). **Trust your own tool list**: if\nthey are not in it, the feature is switched off for this deployment. That is\nnormal, not an error, and not something to work around.",
   "Editing part of a document (prefer this over re-sending)": "## Editing part of a document (prefer this over re-sending)\n\n**Re-sending a whole document to change part of it is the main way agents lose\ndata.** You have to reproduce the untouched remainder verbatim, and any drift\nsilently rewrites content nobody asked you to touch — which the caller cannot\ndiff. Use the partial-edit tools instead:\n\n1. **Learn the anchors** — `cerefox_get_document(document_id, outline: true)`.\n   Returns heading paths, per-section sizes and the `content_hash`, without the\n   body. The paths it returns are exactly what `anchor_heading` accepts.\n2. **Add** → `cerefox_insert`. `end_of_document` is a plain append;\n   `end_of_section` adds inside a named section. It is structurally incapable of\n   removing anything, so \"I meant to append\" cannot become \"I replaced the file\".\n3. **Look before you overwrite** — `cerefox_get_document(document_id,\n   section: \"## Heading\")` returns exactly the text a `replace_section` on that\n   anchor would destroy. The outline gives you a section's *size*, never its\n   *text*, so on a document you did not write yourself this is the difference\n   between a replace and a blind overwrite.\n4. **Change or remove** → `cerefox_edit`. Put changes that belong together in\n   ONE call: they apply atomically, so a table row and the total it feeds cannot\n   end up disagreeing. To change a single line, `replace_section` on its\n   smallest enclosing heading — that is the intended granularity, not a\n   workaround. To fix a stale heading (`## OPEN TODOs (as of ...)`), use\n   `rename_section`: it changes the heading text and leaves the body and\n   position alone.\n5. All of them require `expected_content_hash` and **have no last-write-wins**. A\n   conflict means someone else changed the document; re-read and decide, do not\n   force it.\n\n**A section runs to the next same-or-higher heading, or to the end of the\ndocument.** So `end_of_document` inserts land inside the *last* section, and\nreplacing or deleting that section removes them too. A large shrink in the\nresponse is your warning; `cerefox_list_versions` has the previous content.\n\n**When an anchor is ambiguous the tool refuses and hands you the options** — a\nrepeated heading returns the qualifying paths, and a section with both its own\ncontent and sub-sections returns both `section_part` choices. That is a\nrecoverable answer, not a failure: retry with what it gave you.",
-  "Essential Rules": "## Essential Rules\n\n1. **Search before ingesting** -- check if the document exists first.\n2. **Prefer ID-based updates** -- pass `document_id` from search results for deterministic updates. Falls back to title-matching with `update_if_exists: true`.\n3. **Set `author`** to your name on every call, reads and writes alike (e.g., \"Claude Code\", \"archiver\"). Same parameter on every tool. (`requestor` is still accepted everywhere as the pre-1.13.1 alias.) On the CLI it is `--author` on every command too (plus `--author-type` on writes); or rely on `CEREFOX_AUTHOR_NAME`/`CEREFOX_AUTHOR_TYPE` env vars set in the user's `.env`.\n4. **Use `document_id` from search results** `[id: uuid]` for get_document and list_versions.\n5. **`max_bytes` has to fit a whole document.** Search returns COMPLETE documents, so a budget smaller than the top hit fits nothing. You get a header-only list saying so (never \"No results found.\" — that means the store really has nothing). Raise `max_bytes`, or read one document with `cerefox_get_document` using `outline: true` or `section`.\n6. **Add metadata** -- at minimum `type` (\"decision-log\", \"research\", \"design-doc\") and `status` (\"active\", \"draft\").\n7. **Write for retrieval: the title does real work.** The document title is indexed at the **highest weight** *and* prepended to every chunk before it is embedded, so it steers both halves of hybrid search (keyword and semantic) for every chunk in the document. Name the document with the distinctive terms someone would actually search for -- \"Cerefox Response Size Limits\", not \"Notes\" or \"Update 3\". Renaming re-indexes and re-embeds, so correcting a vague title is cheap and takes effect immediately. Write the characters themselves: `&` and `<>`, never `&amp;` or `&lt;&gt;` (entities are stored literally and the write reply will say so). Inside the content, write structured Markdown with H1/H2/H3 headings: the chunker follows them, and a chunk's own heading carries the same weight as the title.\n8. **Deletes are soft (recoverable); permanent purge is web-UI-only.** `cerefox_delete_document` requires the document's `content_hash` as you read it (read before you delete) and takes a `reason` — give one; it is what the human reviewing the trash sees. `cerefox_restore_document` undoes a mistaken delete (also audited, also takes a `reason`). Always surface deletes AND restores to the user. Once a human purges from the web UI, the document is gone for good; so is one auto-purged after `trash_retention_days`, if the operator has turned trash auto-purge on.\n9. **Cross-doc links inside content**: **always use `[Text](document-uuid)`.** UUIDs are the only fully reliable link form — stable across title changes, never ambiguous, no encoding gotchas. Every `cerefox_search` result shows `[id: <uuid>]` after the title; grab it and use it. Title-based linking (`[Text](<Title With Spaces>)`) is fragile (breaks on colons, parens, ampersands, brackets — silently navigates to wrong page) — **don't write title-based links**; do an extra search to get the UUID instead. Repo-path forms (`[Text](docs/path.md)`) exist for repo-ingested files; don't construct manually. **The server validates `](uuid)` links on every write** (v1.7.0): a link to a nonexistent id rejects the write, naming the offender — that means you mangled the UUID; re-read the source and correct it, do not retry unchanged. Example ids go in backticks (code is not validated). `[[Wikilinks]]` may dangle. See `AGENT_GUIDE.md → Writing linkable content` for the full rule.\n10. **Concurrency: content updates require `expected_content_hash`.** Pass the `content_hash` you last saw — every read shows one (`cerefox_get_document` incl. outline mode, `cerefox_search`, `cerefox_metadata_search`) and **every write returns the new one, including create** (v1.3.0, #189), so after writing you already hold the token for your next edit; no re-read needed. If it's stale you get a **conflict** — re-read the document, merge your changes into the latest content, retry with the new hash. **Never resolve a conflict by overwriting blindly** — the current content includes another writer's work. `last_write_wins: true` skips the check; use it ONLY when an external source of truth makes conflicts meaningless (file re-sync), never to silence a conflict.\n11. **Search: prefer a few distinctive terms; heed `below confidence`.** When nothing clears the relevance threshold, `cerefox_search` returns the closest candidates prefixed with a `below confidence` warning instead of an empty set — that flag means **weak signal, not absent knowledge**: check the candidates' scores and titles before concluding the KB lacks the content. A truly empty response means nothing even weakly related exists.\n12. **Relations express how documents relate; lifecycle tells you if knowledge is still good.** Use `cerefox_set_relation` when one document supersedes, contradicts, references, or continues another. `supersedes` marks the target **superseded**; `contradicts` marks **both** stale; `related_to`/`duplicates`/`contradicts` are symmetric (both directions written). Any other type string is accepted without special behavior. When a search result or `cerefox_get_relations` shows a neighbour marked `[superseded]` or `[stale]`, say so rather than presenting it as current.\n13. **Project memberships — non-destructive by default**: on `cerefox_ingest` updates, **`project_name` (singular) is a non-destructive add** (ensures membership, preserves others). Use **`project_names` (list)** when you want to set the doc's full project set in one call (destructive replace). For metadata-only project changes without writing content, use **`cerefox_set_document_projects(document_id, project_names)`** — that tool is the destructive-replace contract made explicit. Never call `cerefox_set_document_projects` with a single name when you mean \"add\" — that would REMOVE the doc from all other projects. When in doubt, use `cerefox_ingest` with singular `project_name`.",
+  "Essential Rules": "## Essential Rules\n\n1. **Search before ingesting** -- check if the document exists first.\n2. **Prefer ID-based updates** -- pass `document_id` from search results for deterministic updates. Falls back to title-matching with `update_if_exists: true`.\n3. **Set `author`** to your name on every call, reads and writes alike (e.g., \"Claude Code\", \"archiver\"). Same parameter on every tool. (`requestor` is still accepted everywhere as the pre-1.13.1 alias.) On the CLI it is `--author` on every command too (plus `--author-type` on writes); or rely on `CEREFOX_AUTHOR_NAME`/`CEREFOX_AUTHOR_TYPE` env vars set in the user's `.env`.\n4. **Use `document_id` from search results** `[id: uuid]` for get_document and list_versions.\n5. **`max_bytes` has to fit a whole document.** Search returns COMPLETE documents, so a budget smaller than the top hit fits nothing. You get a header-only list saying so (never \"No results found.\" — that means the store really has nothing). Raise `max_bytes`, or read one document with `cerefox_get_document` using `outline: true` or `section`.\n6. **Add metadata** -- at minimum `type` (\"decision-log\", \"research\", \"design-doc\") and `status` (\"active\", \"draft\").\n7. **Write for retrieval: the title does real work.** The document title is indexed at the **highest weight** *and* prepended to every chunk before it is embedded, so it steers both halves of hybrid search (keyword and semantic) for every chunk in the document. Name the document with the distinctive terms someone would actually search for -- \"Cerefox Response Size Limits\", not \"Notes\" or \"Update 3\". Renaming re-indexes and re-embeds, so correcting a vague title is cheap and takes effect immediately. Write the characters themselves: `&` and `<>`, never `&amp;` or `&lt;&gt;` (entities are stored literally and the write reply will say so). Inside the content, write structured Markdown with H1/H2/H3 headings: the chunker follows them, and a chunk's own heading carries the same weight as the title.\n8. **Deletes are soft (recoverable); permanent purge is web-UI-only.** `cerefox_delete_document` requires the document's `content_hash` as you read it (read before you delete) and takes a `reason` — give one; it is what the human reviewing the trash sees. `cerefox_restore_document` undoes a mistaken delete (also audited, also takes a `reason`). Always surface deletes AND restores to the user. Once a human purges from the web UI, the document is gone for good; so is one auto-purged after `trash_retention_days`, if the operator has turned trash auto-purge on.\n9. **Cross-doc links inside content**: **always use `[Text](document-uuid)`.** UUIDs are the only fully reliable link form — stable across title changes, never ambiguous, no encoding gotchas. Every `cerefox_search` result shows `[id: <uuid>]` after the title; grab it and use it. Title-based linking (`[Text](<Title With Spaces>)`) is fragile (breaks on colons, parens, ampersands, brackets — silently navigates to wrong page) — **don't write title-based links**; do an extra search to get the UUID instead. Repo-path forms (`[Text](docs/path.md)`) exist for repo-ingested files; don't construct manually. **The server validates `](uuid)` links on every write** (v1.7.0): a link to a nonexistent id rejects the write, naming the offender — that means you mangled the UUID; re-read the source and correct it, do not retry unchanged. Example ids go in backticks (code is not validated). `[[Wikilinks]]` may dangle. See `AGENT_GUIDE.md → Writing linkable content` for the full rule.\n10. **Concurrency: content updates require `expected_content_hash`.** Pass the `content_hash` you last saw — every read shows one (`cerefox_get_document` incl. outline mode, `cerefox_search`, `cerefox_metadata_search`) and **every write returns the new one, including create** (v1.3.0, #189), so after writing you already hold the token for your next edit; no re-read needed. If it's stale you get a **conflict** — re-read the document, merge your changes into the latest content, retry with the new hash. **Never resolve a conflict by overwriting blindly** — the current content includes another writer's work. `last_write_wins: true` skips the check; use it ONLY when an external source of truth makes conflicts meaningless (file re-sync), never to silence a conflict. **Metadata has its own token, `metadata_revision`** (shown next to `content_hash` wherever it appears). It is optional on a merge (pass it when your change depends on the current values, e.g. a status transition) and **required with `replace: true`**. A stale one is a *metadata conflict*: re-read, reapply, retry.\n11. **Search: prefer a few distinctive terms; heed `below confidence`.** When nothing clears the relevance threshold, `cerefox_search` returns the closest candidates prefixed with a `below confidence` warning instead of an empty set — that flag means **weak signal, not absent knowledge**: check the candidates' scores and titles before concluding the KB lacks the content. A truly empty response means nothing even weakly related exists.\n12. **Relations express how documents relate; lifecycle tells you if knowledge is still good.** Use `cerefox_set_relation` when one document supersedes, contradicts, references, or continues another. `supersedes` marks the target **superseded**; `contradicts` marks **both** stale; `related_to`/`duplicates`/`contradicts` are symmetric (both directions written). Any other type string is accepted without special behavior. When a search result or `cerefox_get_relations` shows a neighbour marked `[superseded]` or `[stale]`, say so rather than presenting it as current.\n13. **Project memberships — non-destructive by default**: on `cerefox_ingest` updates, **`project_name` (singular) is a non-destructive add** (ensures membership, preserves others). Use **`project_names` (list)** when you want to set the doc's full project set in one call (destructive replace). For metadata-only project changes without writing content, use **`cerefox_set_document_projects(document_id, project_names)`** — that tool is the destructive-replace contract made explicit. Never call `cerefox_set_document_projects` with a single name when you mean \"add\" — that would REMOVE the doc from all other projects. When in doubt, use `cerefox_ingest` with singular `project_name`.",
   "Update Workflow (ID-based -- preferred)": "## Update Workflow (ID-based -- preferred)\n\n```\nsearch(\"topic\") -> find doc [id: abc123] -> get_document(abc123) -> note its content_hash -> modify ->\ningest(title=\"Same Title\", content=\"...\", document_id=\"abc123\",\n       expected_content_hash=\"<the hash you read>\", author=\"my-agent\")\n```\n\nOn a **conflict** error: get_document again (fresh content + fresh hash) -> merge your changes -> retry with the new hash.",
   "Update Workflow (title-based -- fallback)": "## Update Workflow (title-based -- fallback)\n\n```\nsearch(\"topic\") -> find doc (note its hash) -> modify ->\ningest(title=\"Same Title\", content=\"...\", update_if_exists=true,\n       expected_content_hash=\"<the hash you read>\", author=\"my-agent\")\n```",
   "Catch-Up Workflow": "## Catch-Up Workflow\n\n```\nmetadata_search(metadata_filter={\"type\": \"decision-log\"}, updated_since=\"2026-03-28T00:00:00Z\")\n```",