mindgraph-mcp 0.14.5 → 0.14.7

package/dist/cli.js

@@ -2891,6 +2891,7 @@ async function handleRetrieve(client2, args) {
     limit,
     offset,
     include_chunks,
+    include_documents,
     include_graph,
     graph_expansion_limit,
     graph_max_depth,
@@ -2908,6 +2909,10 @@ async function handleRetrieve(client2, args) {
           node_types,
           layer,
           chunk_limit: include_chunks ? limit ?? 5 : 0,
+          // Scoped by default, never walled: the coding profile drops the
+          // document/article leg (the "Bob Work" class of noise) unless the
+          // model opts in for a genuinely document-shaped question.
+          ...process.env.MINDGRAPH_PROFILE === "coding" && !include_documents ? { article_limit: 0 } : {},
           include_graph,
           graph_expansion_limit: graph_expansion_limit ?? 3,
           graph_max_depth: graph_max_depth ?? 2
@@ -3855,7 +3860,7 @@ var init_tools = __esm({
       },
       {
         name: "mindgraph_retrieve",
-        description: "Search and explore the knowledge graph. Use 'context' (default) for BM25 keyword retrieval plus bounded cheapest-first graph expansion \u2014 returns graph nodes only by default (labels, summaries, types, confidence); set include_chunks=true to also get full source text. Pass 1-3 discriminating keywords, not sentences. GOOD: 'Kissinger NATO'. BAD: 'What is Kissinger\\'s view on NATO?'. Use 'document_index' to list all ingested documents (titles, dates, UIDs) for orientation. Use 'semantic' when keywords return nothing \u2014 conceptual/fuzzy queries. Use 'hybrid' for keyword + semantic. Use 'text' for fast keyword-only. Use 'neighborhood'/'chain'/'path'/'subgraph' for graph traversal \u2014 traversal steps carry the cheapest witness path_cost (lower = stronger connection; sum of -ln(edge weight)), path_confidence, and witness depth. Use 'preferences' for advice or recommendation requests ('suggest a hotel', 'what should I read?') \u2014 it returns the user's stated/learned preferences relevant to the query (pass the topic as 'query'), so your answer reflects what they actually like instead of being generic. Other structured actions ('active_goals', 'open_questions', etc.) only when the user explicitly asks.",
+        description: "Search and explore the knowledge graph. NOT for work state: for 'what was I working on' or resuming work, use mindgraph_plan action resume_work (structural, deterministic) \u2014 this tool is semantic/keyword recall over ALL knowledge. Use 'context' (default) for BM25 keyword retrieval plus bounded cheapest-first graph expansion \u2014 returns graph nodes only by default (labels, summaries, types, confidence); set include_chunks=true to also get full source text. Pass 1-3 discriminating keywords, not sentences. GOOD: 'Kissinger NATO'. BAD: 'What is Kissinger\\'s view on NATO?'. Use 'document_index' to list all ingested documents (titles, dates, UIDs) for orientation. Use 'semantic' when keywords return nothing \u2014 conceptual/fuzzy queries. Use 'hybrid' for keyword + semantic. Use 'text' for fast keyword-only. Use 'neighborhood'/'chain'/'path'/'subgraph' for graph traversal \u2014 traversal steps carry the cheapest witness path_cost (lower = stronger connection; sum of -ln(edge weight)), path_confidence, and witness depth. Use 'preferences' for advice or recommendation requests ('suggest a hotel', 'what should I read?') \u2014 it returns the user's stated/learned preferences relevant to the query (pass the topic as 'query'), so your answer reflects what they actually like instead of being generic. Other structured actions ('active_goals', 'open_questions', etc.) only when the user explicitly asks.",
         inputSchema: {
           type: "object",
           properties: {
@@ -3940,6 +3945,10 @@ var init_tools = __esm({
               type: "number",
               description: "Pagination offset (for stale_derivations)"
             },
+            include_documents: {
+              type: "boolean",
+              description: "Include ingested documents and wiki articles in context results. The coding profile defaults these OFF so engineering questions aren't answered from unrelated ingested content; pass true when the question is about a document, spec, or article."
+            },
             include_chunks: {
               type: "boolean",
               description: "Include source document chunks in context results (default: false \u2014 set true to fetch full chunk text for citations or deep reading)"
@@ -4891,6 +4900,11 @@ Capture knowledge when the user shares something worth remembering:
 
 ## Coding profile
 
+**Recall order (overrides the #1 rule above for work and code questions):**
+- "What was I working on?" / current work state / resuming anything: call \`mindgraph_plan\` action \`resume_work\` FIRST. Work state is structural; never answer it from keyword search ("recent work" famously retrieves people named Work).
+- Knowledge about specific code (a function, file, or route): \`mindgraph_code\` action \`recall\` \u2014 anchor-based, not keyword-based.
+- Broad knowledge searches: \`mindgraph_retrieve\` as usual, but when searching for work artifacts pass node_types filters (Task, Lesson, RiskAssessment, Claim) to keep personal-graph content out of engineering answers.
+
 - Use mindgraph_sync for explicit, one-way project memory-file imports. It discovers and hashes files, but you deliberately map durable assertions with record; it never writes graph state back to files.
 - Capture corrections, decisions with rationale, fixes that survived meaningful failed attempts, durable invariants/preferences, and blockers with evidence.
 - Do not capture routine mechanics, transient command output, facts cheaply derivable from code, unsupported speculation as high-confidence lessons, or completion without evidence.
@@ -4903,7 +4917,7 @@ Capture knowledge when the user shares something worth remembering:
     server = new import_server.Server(
       {
         name: "mindgraph",
-        version: "0.14.5"
+        version: "0.14.7"
       },
       {
         capabilities: {

package/dist/index.js

@@ -1975,7 +1975,7 @@ var TOOLS = [
   },
   {
     name: "mindgraph_retrieve",
-    description: "Search and explore the knowledge graph. Use 'context' (default) for BM25 keyword retrieval plus bounded cheapest-first graph expansion \u2014 returns graph nodes only by default (labels, summaries, types, confidence); set include_chunks=true to also get full source text. Pass 1-3 discriminating keywords, not sentences. GOOD: 'Kissinger NATO'. BAD: 'What is Kissinger\\'s view on NATO?'. Use 'document_index' to list all ingested documents (titles, dates, UIDs) for orientation. Use 'semantic' when keywords return nothing \u2014 conceptual/fuzzy queries. Use 'hybrid' for keyword + semantic. Use 'text' for fast keyword-only. Use 'neighborhood'/'chain'/'path'/'subgraph' for graph traversal \u2014 traversal steps carry the cheapest witness path_cost (lower = stronger connection; sum of -ln(edge weight)), path_confidence, and witness depth. Use 'preferences' for advice or recommendation requests ('suggest a hotel', 'what should I read?') \u2014 it returns the user's stated/learned preferences relevant to the query (pass the topic as 'query'), so your answer reflects what they actually like instead of being generic. Other structured actions ('active_goals', 'open_questions', etc.) only when the user explicitly asks.",
+    description: "Search and explore the knowledge graph. NOT for work state: for 'what was I working on' or resuming work, use mindgraph_plan action resume_work (structural, deterministic) \u2014 this tool is semantic/keyword recall over ALL knowledge. Use 'context' (default) for BM25 keyword retrieval plus bounded cheapest-first graph expansion \u2014 returns graph nodes only by default (labels, summaries, types, confidence); set include_chunks=true to also get full source text. Pass 1-3 discriminating keywords, not sentences. GOOD: 'Kissinger NATO'. BAD: 'What is Kissinger\\'s view on NATO?'. Use 'document_index' to list all ingested documents (titles, dates, UIDs) for orientation. Use 'semantic' when keywords return nothing \u2014 conceptual/fuzzy queries. Use 'hybrid' for keyword + semantic. Use 'text' for fast keyword-only. Use 'neighborhood'/'chain'/'path'/'subgraph' for graph traversal \u2014 traversal steps carry the cheapest witness path_cost (lower = stronger connection; sum of -ln(edge weight)), path_confidence, and witness depth. Use 'preferences' for advice or recommendation requests ('suggest a hotel', 'what should I read?') \u2014 it returns the user's stated/learned preferences relevant to the query (pass the topic as 'query'), so your answer reflects what they actually like instead of being generic. Other structured actions ('active_goals', 'open_questions', etc.) only when the user explicitly asks.",
     inputSchema: {
       type: "object",
       properties: {
@@ -2060,6 +2060,10 @@ var TOOLS = [
           type: "number",
           description: "Pagination offset (for stale_derivations)"
         },
+        include_documents: {
+          type: "boolean",
+          description: "Include ingested documents and wiki articles in context results. The coding profile defaults these OFF so engineering questions aren't answered from unrelated ingested content; pass true when the question is about a document, spec, or article."
+        },
         include_chunks: {
           type: "boolean",
           description: "Include source document chunks in context results (default: false \u2014 set true to fetch full chunk text for citations or deep reading)"
@@ -2923,6 +2927,7 @@ async function handleRetrieve(client2, args) {
     limit,
     offset,
     include_chunks,
+    include_documents,
     include_graph,
     graph_expansion_limit,
     graph_max_depth,
@@ -2940,6 +2945,10 @@ async function handleRetrieve(client2, args) {
           node_types,
           layer,
           chunk_limit: include_chunks ? limit ?? 5 : 0,
+          // Scoped by default, never walled: the coding profile drops the
+          // document/article leg (the "Bob Work" class of noise) unless the
+          // model opts in for a genuinely document-shaped question.
+          ...process.env.MINDGRAPH_PROFILE === "coding" && !include_documents ? { article_limit: 0 } : {},
           include_graph,
           graph_expansion_limit: graph_expansion_limit ?? 3,
           graph_max_depth: graph_max_depth ?? 2
@@ -3929,6 +3938,11 @@ if (PROFILE === "coding") {
 
 ## Coding profile
 
+**Recall order (overrides the #1 rule above for work and code questions):**
+- "What was I working on?" / current work state / resuming anything: call \`mindgraph_plan\` action \`resume_work\` FIRST. Work state is structural; never answer it from keyword search ("recent work" famously retrieves people named Work).
+- Knowledge about specific code (a function, file, or route): \`mindgraph_code\` action \`recall\` \u2014 anchor-based, not keyword-based.
+- Broad knowledge searches: \`mindgraph_retrieve\` as usual, but when searching for work artifacts pass node_types filters (Task, Lesson, RiskAssessment, Claim) to keep personal-graph content out of engineering answers.
+
 - Use mindgraph_sync for explicit, one-way project memory-file imports. It discovers and hashes files, but you deliberately map durable assertions with record; it never writes graph state back to files.
 - Capture corrections, decisions with rationale, fixes that survived meaningful failed attempts, durable invariants/preferences, and blockers with evidence.
 - Do not capture routine mechanics, transient command output, facts cheaply derivable from code, unsupported speculation as high-confidence lessons, or completion without evidence.
@@ -3941,7 +3955,7 @@ if (PROFILE === "coding") {
 var server = new import_server.Server(
   {
     name: "mindgraph",
-    version: "0.14.5"
+    version: "0.14.7"
   },
   {
     capabilities: {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mindgraph-mcp",
-  "version": "0.14.5",
+  "version": "0.14.7",
   "description": "MCP server for MindGraph — connect Claude Desktop, Claude Code, and any MCP client to your knowledge graph",
   "repository": {
     "type": "git",