@cerefox/memory 1.3.0 → 1.4.0

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 = "1.3.0";
+export const EF_VERSION = "1.4.0";
 
 /**
  * The most recent version whose EF-side SOURCE actually changed (#127).
@@ -28,7 +28,7 @@ export const EF_VERSION = "1.3.0";
  * `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 = "1.3.0-beta.4";
+export const EF_LAST_CHANGED = "1.4.0";
 
 /**
  * The 8 peer EFs the cerefox-mcp aggregator probes (excludes cerefox-mcp

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

@@ -6,7 +6,7 @@
 
 import type { MCPSupabaseClient } from "./types.ts";
 
-import { parseOutline } from "../partial-edits/index.ts";
+import { extractSection, parseOutline } from "../partial-edits/index.ts";
 import { logUsage } from "./_utils.ts";
 import { McpInvalidParams, type ToolContext, type ToolDefinition } from "./types.ts";
 
@@ -18,8 +18,21 @@ async function handler(
   const document_id = args.document_id as string | undefined;
   const version_id = (args.version_id as string | null | undefined) ?? null;
   const outline = (args.outline as boolean | undefined) ?? false;
+  const section = ((args.section as string | undefined) ?? "").trim() || null;
+  const section_part = (args.section_part as "own_body" | "subtree" | undefined) ?? undefined;
 
   if (!document_id) throw new McpInvalidParams("document_id is required");
+  if (section && outline) {
+    // Both answer "what is in this document" at different zoom levels, and
+    // silently preferring one would make the other's absence look like an empty
+    // result. Refuse instead of choosing.
+    throw new McpInvalidParams(
+      "Pass either outline (the whole structure) or section (one section's text), not both.",
+    );
+  }
+  if (section_part && !section) {
+    throw new McpInvalidParams("section_part only applies together with section.");
+  }
 
   const { data, error } = await supabase.rpc("cerefox_get_document", {
     p_document_id: document_id,
@@ -83,6 +96,46 @@ async function handler(
     );
   }
 
+  // Section mode (#198): one section's text, so a replace_section is not a
+  // blind overwrite. The extent comes from the same resolver the write uses —
+  // what this returns is exactly what replace_section would destroy.
+  if (section) {
+    // Anchor failures are caller-recoverable, exactly as they are on the write
+    // path (`applyAndWrite` wraps the same three errors). Left unwrapped they
+    // surface as JSON-RPC -32603 "internal error" while the identical anchor
+    // through cerefox_edit surfaces as -32602 "invalid params", so a client
+    // keying on the code would classify the same mistake two ways depending on
+    // whether it read or wrote. The equivalence this feature promises has to
+    // cover refusals too, not just extents.
+    let extracted;
+    try {
+      extracted = extractSection(row.full_content ?? "", section, section_part);
+    } catch (err) {
+      throw new McpInvalidParams(err instanceof Error ? err.message : String(err));
+    }
+    // Same reasoning as outline mode: the RPC returns the CURRENT hash even when
+    // reconstructing an archived version, and pairing it with archived text
+    // would invite an edit based on content that is no longer there.
+    const archived = version_id !== null;
+    return JSON.stringify(
+      {
+        title: row.doc_title ?? "Untitled",
+        heading: extracted.heading,
+        path: extracted.path,
+        level: extracted.level,
+        section_part: extracted.section_part,
+        chars: extracted.chars,
+        content_hash: archived ? null : (row.content_hash ?? null),
+        text: extracted.text,
+        note: archived
+          ? "This is an ARCHIVED version's section, so no content_hash is returned: it must not be used to edit the current document."
+          : "This is exactly the text a replace_section on this anchor would overwrite (the heading itself is kept). content_hash is your expected_content_hash.",
+      },
+      null,
+      2,
+    );
+  }
+
   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.
@@ -93,7 +146,7 @@ async function handler(
 export const getDocumentTool: ToolDefinition = {
   name: "cerefox_get_document",
   description:
-    "Retrieve the full reconstructed content of a document — or, with outline=true, just its heading structure, sizes and content_hash (much cheaper, and the paths are the anchors the edit tools take). Pass version_id to retrieve an archived version; omit it (or pass null) for the current version. Version UUIDs are returned by cerefox_list_versions. The response header includes the document's current content_hash — pass it back as expected_content_hash when updating via cerefox_ingest (optimistic concurrency).",
+    "Retrieve a document at one of three zoom levels: the whole reconstructed content (default), its heading structure with outline=true (much cheaper, and the headings are the anchors the edit tools take), or one section's text with section=\"## Heading\" (what a replace_section on that anchor would overwrite — read it before replacing a section you did not write). Pass version_id to retrieve an archived version; omit it (or pass null) for the current version. Version UUIDs are returned by cerefox_list_versions. Every non-archived response carries the document's current content_hash — pass it back as expected_content_hash when updating (optimistic concurrency).",
   // Read-only: touches nothing. Safe for a client to run without prompting.
   annotations: {
     title: "Read document",
@@ -113,7 +166,18 @@ export const getDocumentTool: ToolDefinition = {
       outline: {
         type: "boolean",
         description:
-          "Return the document's STRUCTURE instead of its content: heading paths, levels and per-section sizes, plus content_hash and total size. Far cheaper than a full read, and the paths are exactly what cerefox_insert / cerefox_edit take as anchor_heading. Use this before editing a document you have not read.",
+          "Return the document's STRUCTURE instead of its content: heading paths, levels and per-section sizes, plus content_hash and total size. Far cheaper than a full read. Every heading listed is addressable by cerefox_insert / cerefox_edit: pass the bare heading line (e.g. '## Daily Logs') when it occurs once in the document, and only the full ' > ' path shown here when the same heading text repeats. Use this before editing a document you have not read.",
+      },
+      section: {
+        type: "string",
+        description:
+          "Return ONE section's text instead of the whole document: the anchor heading, exactly as cerefox_insert / cerefox_edit take it (bare heading line when unique, ' > ' path when it repeats). What comes back is precisely the text a replace_section on this anchor would overwrite, so read it before replacing a section you did not write. The heading itself is returned separately, because replace_section keeps it. Cannot be combined with outline.",
+      },
+      section_part: {
+        type: "string",
+        enum: ["own_body", "subtree"],
+        description:
+          "Only when the target section HAS CHILD SECTIONS, and it means the same here as on the edit tools: own_body = up to the first child, subtree = everything nested underneath. The read refuses without it for exactly the cases the write refuses, so that what you read is what you would replace. Omit it otherwise; you will be told (with both options) whenever it is needed.",
       },
       requestor: {
         type: "string",

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

@@ -11,12 +11,12 @@
  * 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 **16 MCP tools** (15 of them have CLI equivalents — `cerefox_get_help` is MCP-only). 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`, `requestor` |\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 9), `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` |\n| `cerefox_edit` | **Change** parts of a document: 1..n operations applied atomically | `document_id`, `operations` (`insert`/`replace_section`/`delete_section`), `expected_content_hash` (required) |\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 | `document_id` (required), `outline` |\n| `cerefox_list_versions` | Version history of a document | `document_id` (required) |\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` |\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_projects` | Set doc's project memberships to exactly the given list (destructive replace; metadata-only, no content change) | `document_id`, `project_names` (required) |\n| `cerefox_get_audit_log` | Query write operation history | `document_id`, `author`, `operation`, `since` |\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. **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.\n4. Both 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`/`requestor`** to your name on every call (e.g., \"Claude Code\", \"archiver\"). On MCP, pass as parameters. On CLI, pass `--author`/`--author-type`/`--requestor` flags, or rely on `CEREFOX_AUTHOR_NAME`/`CEREFOX_AUTHOR_TYPE`/`CEREFOX_REQUESTOR_NAME` env vars set in the user's `.env`.\n4. **Use `document_id` from search results** `[id: uuid]` for get_document and list_versions.\n5. **Add metadata** -- at minimum `type` (\"decision-log\", \"research\", \"design-doc\") and `status` (\"active\", \"draft\").\n6. **Write structured Markdown** with H1/H2/H3 headings for good chunking and search.\n7. **Deletes are soft (recoverable); purge is web-UI-only.** If you decide to delete, surface it to the user (`I soft-deleted X — recoverable from the Cerefox web UI trash`). You cannot un-do your own delete from agent code by design.\n8. **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. See `AGENT_GUIDE.md → Writing linkable content` for the full rule.\n9. **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.\n10. **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.\n11. **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 behaviour. When a search result or `cerefox_get_relations` shows a neighbour marked `[superseded]` or `[stale]`, say so rather than presenting it as current.\n12. **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`, `-r`). 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>\" --requestor \"<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> --requestor \"<your-name>\"` |\n| `cerefox_list_versions` | `cerefox document version list <id> --requestor \"<your-name>\"` |\n| `cerefox_list_projects` | `cerefox project list --requestor \"<your-name>\"` |\n| `cerefox_list_metadata_keys` | `cerefox metadata keys` |\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` |\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` | `cerefox metadata search --metadata-filter '<json>' --requestor \"<your-name>\"` (list a project: `cerefox document list --project <name>`) |\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 --requestor \"<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: `--requestor \"<your-name>\"`\n\nOr have your user set `CEREFOX_AUTHOR_NAME` / `CEREFOX_AUTHOR_TYPE` / `CEREFOX_REQUESTOR_NAME` in their `.env` to apply defaults once.\n";
+export const HELP_FULL = "# Cerefox Knowledge Base -- Agent Quick Reference\n\nCerefox is a persistent, shared knowledge base. You have **16 MCP tools** (15 of them have CLI equivalents — `cerefox_get_help` is MCP-only). 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`, `requestor` |\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 9), `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` |\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) |\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` |\n| `cerefox_list_versions` | Version history of a document | `document_id` (required) |\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` |\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_projects` | Set doc's project memberships to exactly the given list (destructive replace; metadata-only, no content change) | `document_id`, `project_names` (required) |\n| `cerefox_get_audit_log` | Query write operation history | `document_id`, `author`, `operation`, `since` |\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`/`requestor`** to your name on every call (e.g., \"Claude Code\", \"archiver\"). On MCP, pass as parameters. On CLI, pass `--author`/`--author-type`/`--requestor` flags, or rely on `CEREFOX_AUTHOR_NAME`/`CEREFOX_AUTHOR_TYPE`/`CEREFOX_REQUESTOR_NAME` env vars set in the user's `.env`.\n4. **Use `document_id` from search results** `[id: uuid]` for get_document and list_versions.\n5. **Add metadata** -- at minimum `type` (\"decision-log\", \"research\", \"design-doc\") and `status` (\"active\", \"draft\").\n6. **Write structured Markdown** with H1/H2/H3 headings for good chunking and search.\n7. **Deletes are soft (recoverable); purge is web-UI-only.** If you decide to delete, surface it to the user (`I soft-deleted X — recoverable from the Cerefox web UI trash`). You cannot un-do your own delete from agent code by design.\n8. **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. See `AGENT_GUIDE.md → Writing linkable content` for the full rule.\n9. **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.\n10. **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.\n11. **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 behaviour. When a search result or `cerefox_get_relations` shows a neighbour marked `[superseded]` or `[stale]`, say so rather than presenting it as current.\n12. **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`, `-r`). 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>\" --requestor \"<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> --requestor \"<your-name>\"` |\n| `cerefox_list_versions` | `cerefox document version list <id> --requestor \"<your-name>\"` |\n| `cerefox_list_projects` | `cerefox project list --requestor \"<your-name>\"` |\n| `cerefox_list_metadata_keys` | `cerefox metadata keys` |\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` |\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` | `cerefox metadata search --metadata-filter '<json>' --requestor \"<your-name>\"` (list a project: `cerefox document list --project <name>`) |\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 --requestor \"<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: `--requestor \"<your-name>\"`\n\nOr have your user set `CEREFOX_AUTHOR_NAME` / `CEREFOX_AUTHOR_TYPE` / `CEREFOX_REQUESTOR_NAME` in their `.env` to apply defaults once.\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`, `requestor` |\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 9), `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` |\n| `cerefox_edit` | **Change** parts of a document: 1..n operations applied atomically | `document_id`, `operations` (`insert`/`replace_section`/`delete_section`), `expected_content_hash` (required) |\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 | `document_id` (required), `outline` |\n| `cerefox_list_versions` | Version history of a document | `document_id` (required) |\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` |\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_projects` | Set doc's project memberships to exactly the given list (destructive replace; metadata-only, no content change) | `document_id`, `project_names` (required) |\n| `cerefox_get_audit_log` | Query write operation history | `document_id`, `author`, `operation`, `since` |\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. **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.\n4. Both 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.",
+  "Tools": "## Tools\n\n| Tool | Purpose | Key params |\n|------|---------|------------|\n| `cerefox_search` | Find documents (hybrid FTS + semantic) | `query` (required), `project_name`, `metadata_filter`, `requestor` |\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 9), `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` |\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) |\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` |\n| `cerefox_list_versions` | Version history of a document | `document_id` (required) |\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` |\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_projects` | Set doc's project memberships to exactly the given list (destructive replace; metadata-only, no content change) | `document_id`, `project_names` (required) |\n| `cerefox_get_audit_log` | Query write operation history | `document_id`, `author`, `operation`, `since` |\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`/`requestor`** to your name on every call (e.g., \"Claude Code\", \"archiver\"). On MCP, pass as parameters. On CLI, pass `--author`/`--author-type`/`--requestor` flags, or rely on `CEREFOX_AUTHOR_NAME`/`CEREFOX_AUTHOR_TYPE`/`CEREFOX_REQUESTOR_NAME` env vars set in the user's `.env`.\n4. **Use `document_id` from search results** `[id: uuid]` for get_document and list_versions.\n5. **Add metadata** -- at minimum `type` (\"decision-log\", \"research\", \"design-doc\") and `status` (\"active\", \"draft\").\n6. **Write structured Markdown** with H1/H2/H3 headings for good chunking and search.\n7. **Deletes are soft (recoverable); purge is web-UI-only.** If you decide to delete, surface it to the user (`I soft-deleted X — recoverable from the Cerefox web UI trash`). You cannot un-do your own delete from agent code by design.\n8. **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. See `AGENT_GUIDE.md → Writing linkable content` for the full rule.\n9. **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.\n10. **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.\n11. **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 behaviour. When a search result or `cerefox_get_relations` shows a neighbour marked `[superseded]` or `[stale]`, say so rather than presenting it as current.\n12. **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```",

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

@@ -356,7 +356,8 @@ async function handler(
 
 export const ingestTool: ToolDefinition = {
   name: "cerefox_ingest",
-  description: "Save a note or document to the Cerefox knowledge base.",
+  description:
+    "Save a note or document to the Cerefox knowledge base. Updating an existing document REPLACES its whole content, so every unchanged character has to be reproduced exactly. Prefer cerefox_insert / cerefox_edit for any change to part of a document, and especially where the untouched content cannot be checked by reading it — document IDs, hashes, numeric tables, indexes and registries. Drifting prose is obvious on review; one wrong character in a UUID is not, and it silently breaks the reference.",
   /** Destructive: `project_names` REPLACES the document's project memberships, and
  *  memberships have no version history — a partial list silently drops the rest.
  *  Content itself is version-snapshotted and guarded by expected_content_hash, so

package/dist/server-assets/_shared/mcp-tools/partial-edits.ts

@@ -68,18 +68,75 @@ function defaultRequestor(ctx: ToolContext): string {
  * The full-document diff is deliberately not returned (§3.8), so the size delta
  * is the cheapest honest signal available.
  */
-function shrinkNote(before: number, after: number): string {
-  const lost = before - after;
+/**
+ * Did any destructive operation target a section that ran to the END of the
+ * document?
+ *
+ * This is the shape behind the whole warning. A section extends to the next
+ * heading of equal-or-higher level **or to EOF**, so the last section owns
+ * everything appended after it. An `end_of_document` insert lands inside that
+ * section, and a later `replace_section` on its heading removes it — correctly
+ * by the addressing rules, and silently.
+ *
+ * The answer is recorded by `applyOne` when the operation resolves, not
+ * recomputed here. The first version re-parsed the PRE-batch document and
+ * matched on `applied.path`, which breaks the moment a batch renames a heading
+ * before touching it: the later op's path names a heading the pre-batch
+ * outline has never seen, the lookup finds nothing, and the warning goes quiet
+ * on exactly the batch shape `rename_section` exists to enable. It also cost a
+ * second parse per destructive edit.
+ */
+function touchedTrailingSection(applied: AppliedOperation[]): boolean {
+  return applied.some((a) => a.reachedEnd === true);
+}
+
+/**
+ * Report content loss (#196).
+ *
+ * The first version gated on a >25% ratio and so could not see the case it was
+ * built for. The loss that matters most is small *precisely because it was
+ * recently added*: append a 400-character entry to an 11,000-character
+ * document, then replace the last section, and 4% of the document silently
+ * takes the new entry with it. A percentage threshold is structurally blind to
+ * that — the more established the document, the smaller the fraction, the
+ * quieter the failure.
+ *
+ * So the ratio is no longer a gate on whether to speak, only on how loudly:
+ *
+ * - Any net loss is stated with its size. Cheap, and a caller who meant it
+ *   reads one clause.
+ * - A destructive operation on a section that ran to EOF gets the full
+ *   explanation at any size, because that is the sequence that surprises.
+ * - A large proportional loss gets it too, since that is worth a second look
+ *   whichever section it hit.
+ */
+function shrinkNote(before: string, afterChars: number, applied: AppliedOperation[]): string {
+  // Code points, NOT UTF-16 code units. `afterChars` comes from the RPC's
+  // SUM(char_count), and the chunker counts code points — so `before.length`
+  // over-counts by one per non-BMP character (emoji, most non-BMP CJK) and
+  // manufactures a phantom loss. Latent since v1.3.0, where the >25% gate hid
+  // it; removing that gate for #196 would have surfaced it on every edit to a
+  // document containing an emoji — including cerefox_insert, which is
+  // annotated destructiveHint: false precisely because it cannot lose content.
+  const beforeChars = [...before].length;
+  const lost = beforeChars - afterChars;
   if (lost <= 0) return "";
-  const pct = Math.round((lost / Math.max(before, 1)) * 100);
-  if (pct < 25) return "";
-  return (
-    `⚠ This edit removed ${lost} characters (${pct}% smaller). If you did not ` +
-    `intend that, note that a section runs to the next heading of the same or ` +
-    `higher level — or to the end of the document — so replacing or deleting ` +
-    `the LAST section also removes anything appended after it. ` +
-    `cerefox_list_versions has the previous content.\n`
-  );
+  const pct = Math.round((lost / Math.max(beforeChars, 1)) * 100);
+  const trailing = touchedTrailingSection(applied);
+
+  if (pct < 25 && !trailing) {
+    return `This edit removed ${lost} characters. cerefox_list_versions has the previous content.\n`;
+  }
+
+  const why = trailing
+    ? `You replaced or deleted the LAST section, and a section runs to the next ` +
+      `heading of the same or higher level — or to the end of the document — so ` +
+      `anything appended after it was inside it. `
+    : `If you did not intend that, note that a section runs to the next heading ` +
+      `of the same or higher level — or to the end of the document — so replacing ` +
+      `or deleting the LAST section also removes anything appended after it. `;
+
+  return `⚠ This edit removed ${lost} characters (${pct}% smaller). ${why}cerefox_list_versions has the previous content.\n`;
 }
 
 /** Audit `operation` values, matching the CHECK constraint widened by migration 0019. */
@@ -87,6 +144,7 @@ const AUDIT_OP: Record<AppliedOperation["op"], string> = {
   insert: "insert",
   replace_section: "replace-section",
   delete_section: "delete-section",
+  rename_section: "rename-section",
 };
 
 /**
@@ -293,7 +351,7 @@ async function applyAndWrite(
     `Applied ${applied.length} operation(s) to "${doc.title}" (id: ${documentId}):\n${summary}\n\n` +
     `New content_hash: ${row?.content_hash ?? newHash}\n` +
     `Size: ${row?.total_chars ?? totalChars} chars (was ${doc.content.length}), ${chunks.length} chunk(s).\n` +
-    shrinkNote(doc.content.length, row?.total_chars ?? totalChars) +
+    shrinkNote(doc.content, row?.total_chars ?? totalChars, applied) +
     `Pass the new content_hash as expected_content_hash on your next edit.${warning}`
   );
 }
@@ -445,7 +503,9 @@ export const editTool: ToolDefinition = {
     "Change parts of a document without resending the whole thing: one or many operations " +
     "applied ATOMICALLY in one write. Operations: insert (same positions as cerefox_insert), " +
     "replace_section (swap a section's body, heading kept), delete_section (remove a section, " +
-    "scope body_only or heading_and_body). Use one call for changes that belong together — a " +
+    "scope body_only or heading_and_body), rename_section (change a heading's text, leaving " +
+    "its body and position untouched — for headings that go stale, like a dated one). " +
+    "Use one call for changes that belong together — a " +
     "half-applied edit is impossible, so a row and the total it feeds cannot disagree. " +
     "Operations apply in order and each sees the previous one's result. To change a single " +
     "line, replace_section on its smallest enclosing heading. Requires expected_content_hash; " +
@@ -472,7 +532,10 @@ export const editTool: ToolDefinition = {
           type: "object",
           required: ["op"],
           properties: {
-            op: { type: "string", enum: ["insert", "replace_section", "delete_section"] },
+            op: {
+              type: "string",
+              enum: ["insert", "replace_section", "delete_section", "rename_section"],
+            },
             text: { type: "string", description: "Markdown. Required for insert and replace_section." },
             position: {
               type: "string",
@@ -495,6 +558,11 @@ export const editTool: ToolDefinition = {
               enum: ["body_only", "heading_and_body"],
               description: "delete_section only. Defaults to body_only, which keeps the heading.",
             },
+            new_heading: {
+              type: "string",
+              description:
+                "rename_section only: the replacement heading LINE, at the same level (## stays ##). Changes the heading text and nothing else — the body and the section's position are untouched, which is the point: renaming via delete + insert would risk both. Use it for headings that go stale, like '## OPEN TODOs (as of 2026-08-08)'. A rename changes the anchor, so a later operation in the same call must target the NEW heading.",
+            },
           },
         },
       },