@anchrd/intel-api 0.14.0 → 0.16.0

package/dist/indexing/indexing.js

@@ -1,10 +1,9 @@
-import { ArchivedBoardStatusId, BlockNoteDocument, BlockNoteMediaType, BoardDocument, } from "@anchrd/intel-contract";
-import { upgradeStoredBoard } from "../nodes/board/board.js";
+import { BlockNoteDocument, BlockNoteMediaType } from "@anchrd/intel-contract/node";
 export class PermanentIndexingError extends Error {
 }
-// The key under which a node that has exactly one vector is filed — every kind but `board`. It is
-// the bare node id in the index itself, which is why nothing written before anchrd/intel#301 has to
-// be renamed or embedded again.
+// The key under which a node with exactly one vector is filed — since #390 that is every kind. It
+// is the bare node id in the index itself, which is why nothing written before anchrd/intel#301 has
+// to be renamed or embedded again.
 const wholeNodeChunkKey = "";
 // What the pass hands to `purgeVectors` at its end: the names it has just written, in case the node
 // was archived while it worked and the record of them was refused (anchrd/intel#348).
@@ -21,58 +20,6 @@ function indexText(mediaType, content) {
         throw new PermanentIndexingError(`Node version content is not a valid BlockNote document: ${error instanceof Error ? error.message : "unknown parse failure"}`);
     }
 }
-/**
- * One passage per task, never one per board (#285) — and since anchrd/intel#301 one VECTOR per task
- * too, which is what `key` carries.
- *
- * ⚠️ This is the whole reason a board is chunked at all. "Where do I stand with X" has to land on a
- * CARD: indexed as one blob, a board of three hundred tasks matches on any of them and answers with
- * whichever thousand characters the snippet function happened to cut — text from a task that may
- * have nothing to do with the question. Each chunk carries the task's own title, which is the
- * column the FTS table weights highest.
- *
- * ⚠️ The status list is not indexed, and neither is a task's own status. "Backlog" and "Done" appear
- * on every board in the installation, so they are the words most likely to match and the least
- * likely to mean anything — and in the semantic half they would drag every card of every board
- * towards every question phrased as a state.
- *
- * ⚠️ A task on the `archived` shelf is left out of BOTH halves (anchrd/intel#301). That shelf is
- * what a board has instead of deleting a card (`ArchivedBoardStatusId`), so it is the same statement
- * `archived_at` makes about a node one level up — and an archived node has never been searchable.
- * Answering "where do I stand with X" out of a card somebody swept away is the failure the ticket
- * names. A card in a `terminal` status is NOT swept: "we finished it" is an answer to where the work
- * stands, and the only thing that says a card no longer counts is the shelf.
- */
-function boardChunks(content) {
-    let parsed;
-    try {
-        parsed = JSON.parse(content);
-    }
-    catch (error) {
-        throw new PermanentIndexingError(`Node version content is not valid JSON: ${error instanceof Error ? error.message : "unknown parse failure"}`);
-    }
-    /**
-     * ⚠️ Upgraded first, exactly like `parseStoredBoard` — this is the SECOND place a stored board
-     * body is parsed, and it reads bodies nobody has rewritten since (anchrd/intel#318).
-     *
-     * `reindex` walks every current version in the installation, so an old board arrives here as it
-     * was written: without `terminal` (#311) or with a repeated `dependsOn`. Without the upgrade the
-     * refusal below is PERMANENT — the version is marked failed and never retried — so that board
-     * would simply stop being findable, quietly, while its screen still worked. A rule added at one
-     * parse site and not the other is how a fix looks complete and is not.
-     */
-    const board = BoardDocument.safeParse(upgradeStoredBoard(parsed));
-    if (!board.success) {
-        throw new PermanentIndexingError("Node version content is not a valid board document");
-    }
-    return board.data.tasks
-        .filter((task) => task.status !== ArchivedBoardStatusId)
-        .map((task) => ({
-        key: task.id,
-        title: task.title,
-        text: [task.title, ...task.labels, task.description].filter(Boolean).join("\n\n"),
-    }));
-}
 /**
  * What one chunk is embedded as (anchrd/intel#301).
  *
@@ -85,9 +32,9 @@ function boardChunks(content) {
  * compare against. That cost is one embedding per node, paid when that node is next saved or when
  * an administrator runs `reindex`, and never per card.
  *
- * ⚠️ A card is NOT prefixed with the board's title. It brings its own title as the first line of its
- * text already, and the board's would put a value into all three hundred fingerprints that moves
- * when somebody renames the board — one rename, three hundred embeddings.
+ * ⚠️ Only the WHOLE-node chunk is prefixed, and the distinction is what kept a rename cheap while
+ * a node could hold many chunks: a per-chunk prefix would have put the node's title into every
+ * fingerprint, so renaming it once cost one embedding per chunk.
  */
 function embeddedText(target, chunk) {
     return chunk.key === wholeNodeChunkKey ? `${target.title}\n\n${chunk.text}` : chunk.text;
@@ -95,14 +42,13 @@ function embeddedText(target, chunk) {
 /**
  * The vectors this node needs, minus the ones it already has (anchrd/intel#301).
  *
- * ⚠️ This is the answer to "a changed card must not reindex the board". A save writes the whole
- * document — that is what a board version IS — so without this every keystroke-save of a board with
- * three hundred cards would cost three hundred embeddings. The fingerprint of the exact text that
- * was embedded is what tells an untouched card from a changed one.
+ * ⚠️ This is the answer to "an unchanged chunk must not be embedded again". A save writes the
+ * whole document, so without this every save would re-embed everything it holds. The fingerprint of
+ * the exact text that was embedded is what tells an untouched chunk from a changed one.
  *
  * ⚠️ The comparison is against D1 and deliberately not against Vectorize. Vectorize writes are
  * asynchronous — a vector upserted a moment ago is not readable yet — so a check made there would
- * answer about the save before last and re-embed a board that had just been saved twice.
+ * answer about the save before last and re-embed a node that had just been saved twice.
  *
  * ⚠️ The record is written only after the upsert returned, in the caller. A pass that dies in
  * between leaves the record short, so the next one embeds again; the opposite order would leave a
@@ -127,9 +73,9 @@ async function replaceVectors(deps, target, chunks) {
     const stale = stored.map((record) => record.chunkKey).filter((key) => !wanted.has(key));
     // ⚠️ The one-time sweep of the vector this node had BEFORE it was chunked. An empty record with a
     // chunked node means this is the first pass since anchrd/intel#301 (or since `reindex` emptied
-    // the record), and the whole-board vector written under the bare node id is still sitting there —
+    // the record), and the whole-node vector written under the bare node id is still sitting there —
     // matching questions and answering with the wrong card's passage. Once a record exists this
-    // branch is never taken again, so it costs one call per board rather than one per save.
+    // branch is never taken again, so it costs one call per node rather than one per save.
     if (stored.length === 0 && !wanted.has(wholeNodeChunkKey))
         stale.push(wholeNodeChunkKey);
     await semantic.upsert(target, changed);
@@ -207,22 +153,6 @@ export function createIndexing(deps) {
                 await purgeVectors(deps, versionId);
                 return;
             }
-            // ⚠️ An agent's body is a definition, not something to read: node IDs, a cron line and a model
-            // name (ADR-0005 §4). Indexing it verbatim would fill the index with identifiers and, worse,
-            // put them into the passage a searcher is shown — the definition names nodes the searcher may
-            // have no access to, and a snippet is the one place their IDs would become visible without
-            // the tree being asked. What makes an agent findable is what a person wrote about it, so the
-            // text is its description, and the FTS table indexes the title on its own (#139).
-            if (target.kind === "agent") {
-                const chunks = [
-                    { key: wholeNodeChunkKey, title: target.title, text: target.description ?? "" },
-                ];
-                await deps.repository.replace(target, chunks);
-                await replaceVectors(deps, target, chunks);
-                await purgeVectors(deps, versionId, writtenKeys(chunks));
-                await deps.repository.markIndexed(versionId, deps.now().toISOString());
-                return;
-            }
             // An attachment is one object of bytes; text kinds may be several, because a table's content
             // is the join of its versions (#40). A missing segment fails the whole pass rather than
             // indexing a table with a hole in it.
@@ -240,12 +170,11 @@ export function createIndexing(deps) {
                 if (text === undefined) {
                     throw new PermanentIndexingError(`No document converter is configured for ${target.mediaType}`);
                 }
-                // Both halves are chunked the same way now (anchrd/intel#301): a board is one passage and
-                // one vector per card, everything else is one of each for the node. The two indexes read the
-                // same list, so they cannot come to disagree about what a board is made of.
-                const chunks = target.kind === "board"
-                    ? boardChunks(text)
-                    : [{ key: wholeNodeChunkKey, title: target.title, text }];
+                // ⚠️ One chunk per node since #390, and the LIST is what stays. Both indexes read it, so
+                // they cannot come to disagree about what a node is made of — which is the property that
+                // let a board be one passage and one vector per card (#285, anchrd/intel#301) without
+                // either half being told about it separately.
+                const chunks = [{ key: wholeNodeChunkKey, title: target.title, text }];
                 await deps.repository.replace(target, chunks);
                 await replaceVectors(deps, target, chunks);
                 // ⚠️ Asked again, at the END of a pass that started on a live node (anchrd/intel#348). The

package/dist/intel/intel.js

@@ -72,9 +72,6 @@ export function createIntel(deps) {
         flows: deps.flows,
         tools: deps.tools,
         bundle: deps.bundle,
-        agents: deps.agents,
-        agentCosts: deps.agentCosts,
-        models: deps.models,
         resource,
         resourceMetadataUrl,
         ...(deps.auth ? { auth: deps.auth } : {}),
@@ -87,13 +84,11 @@ export function createIntel(deps) {
         }
         return await handleMcp(context.req.raw, {
             authorization,
-            // ⚠️ The very token this call was authorized with, so the agent tools reach the runtime as the
-            // caller and not as Intel. There is no cookie path on `/mcp`, so `bearer` is the same value
-            // `authorize` just used — read again rather than passed out of it, because a function that
-            // returned a credential beside its answer would be one somebody logs.
+            // ⚠️ The very token this call was authorized with, so a tool that has to speak to Gate does
+            // it as the caller and not as Intel. There is no cookie path on `/mcp`, so `bearer` is the
+            // same value `authorize` just used — read again rather than passed out of it, because a
+            // function that returned a credential beside its answer would be one somebody logs.
             bearer: bearer(context.req.raw.headers) ?? "",
-            agents: deps.agents,
-            agentCosts: deps.agentCosts,
             flows: deps.flows,
             nodes: deps.nodes,
             tools: deps.tools,

package/dist/intel/intel.types.d.ts

@@ -1,10 +1,7 @@
 import type { GateClient } from "@anchrd/gate-sdk";
-import type { AgentCostsService } from "../agent-costs/agent-costs.types.js";
-import type { AgentRuntimeService } from "../agent-runtime/agent-runtime.types.js";
 import type { BrowserAuth } from "../auth/auth.types.js";
 import type { BundleService } from "../bundle/bundle.types.js";
 import type { FlowService } from "../flows/flows.types.js";
-import type { ModelCatalogService } from "../model-catalog/model-catalog.types.js";
 import type { NodeService } from "../nodes/nodes.types.js";
 import type { ToolService } from "../tools/tools.types.js";
 export interface IntelDeps {
@@ -15,8 +12,5 @@ export interface IntelDeps {
     flows: FlowService;
     tools: ToolService;
     bundle: BundleService;
-    agents: AgentRuntimeService;
-    agentCosts: AgentCostsService;
-    models: ModelCatalogService;
     auth?: BrowserAuth;
 }