smart-context-mcp 1.18.0 → 1.18.1

package/README.md

@@ -56,7 +56,7 @@ Restart your AI client. Done.
 # Check installed version
 npm list -g smart-context-mcp
 
-# Should show: smart-context-mcp@1.18.0 (or later)
+# Should show: smart-context-mcp@1.18.1 (or later)
 
 # Update to latest version
 npm update -g smart-context-mcp

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "smart-context-mcp",
   "mcpName": "io.github.Arrayo/smart-context-mcp",
-  "version": "1.18.0",
+  "version": "1.18.1",
   "description": "MCP server that reduces agent token usage by 90% with intelligent context compression, task checkpoint persistence, and workflow-aware agent guidance.",
   "author": "Francisco Caballero Portero <fcp1978@hotmail.com>",
   "type": "module",

package/server.json

@@ -6,12 +6,12 @@
     "url": "https://github.com/Arrayo/smart-context-mcp",
     "source": "github"
   },
-  "version": "1.18.0",
+  "version": "1.18.1",
   "packages": [
     {
       "registryType": "npm",
       "identifier": "smart-context-mcp",
-      "version": "1.18.0",
+      "version": "1.18.1",
       "transport": {
         "type": "stdio"
       },

package/src/orchestration/adapters/claude-adapter.js

@@ -21,6 +21,12 @@ import {
   normalizeWhitespace,
   truncate,
 } from '../policy/event-policy.js';
+import {
+  evaluateSoftPrompt,
+  isSoftPromptsEnabled,
+  markSoftPromptEmitted,
+  shouldEmitSoftPrompt,
+} from '../policy/soft-prompts.js';
 
 export const HOOK_CLIENT = 'claude';
 export const STOP_MAX_TOKENS = 300;
@@ -501,6 +507,26 @@ export const createClaudeAdapter = ({
         continuityState: existing.continuityState,
       });
     }
+
+    if (isSoftPromptsEnabled() && shouldEmitSoftPrompt(hookKey)) {
+      const softPrompt = evaluateSoftPrompt({
+        toolName: input.tool_name,
+        toolInput: input.tool_input,
+        toolResponse: input.tool_response,
+        state: nextState,
+      });
+      if (softPrompt) {
+        markSoftPromptEmitted(hookKey);
+        await recordHookMetrics({
+          action: 'PostToolUse',
+          sessionId: existing.projectSessionId,
+          additionalContext: softPrompt.message,
+          continuityState: existing.continuityState,
+        });
+        return buildClaudeHookContextResponse('PostToolUse', softPrompt.message);
+      }
+    }
+
     return null;
   };
 

package/src/orchestration/adapters/cursor-adapter.js

@@ -21,6 +21,12 @@ import {
   normalizeWhitespace,
   truncate,
 } from '../policy/event-policy.js';
+import {
+  evaluateSoftPrompt,
+  isSoftPromptsEnabled,
+  markSoftPromptEmitted,
+  shouldEmitSoftPrompt,
+} from '../policy/soft-prompts.js';
 
 export const HOOK_CLIENT = 'cursor';
 export const STOP_MAX_TOKENS = 300;
@@ -504,6 +510,26 @@ export const createCursorAdapter = ({
         continuityState: existing.continuityState,
       });
     }
+
+    if (isSoftPromptsEnabled() && shouldEmitSoftPrompt(hookKey)) {
+      const softPrompt = evaluateSoftPrompt({
+        toolName: input.tool_name,
+        toolInput: input.tool_input,
+        toolResponse: input.tool_response,
+        state: nextState,
+      });
+      if (softPrompt) {
+        markSoftPromptEmitted(hookKey);
+        await recordHookMetrics({
+          action: 'PostToolUse',
+          sessionId: existing.projectSessionId,
+          additionalContext: softPrompt.message,
+          continuityState: existing.continuityState,
+        });
+        return buildCursorHookContextResponse('PostToolUse', softPrompt.message);
+      }
+    }
+
     return null;
   };
 

package/src/orchestration/policy/soft-prompts.js

@@ -0,0 +1,97 @@
+const DEFAULT_THROTTLE_MS = 2 * 60 * 1000;
+const LARGE_RESPONSE_CHARS = 12000;
+const REPEATED_READ_THRESHOLD = 5;
+const REPEATED_GREP_THRESHOLD = 3;
+
+const lastIssuedAt = new Map();
+
+const env = (key) => (typeof process !== 'undefined' ? process.env?.[key] : undefined);
+
+export const isSoftPromptsEnabled = () => {
+  const value = env('DEVCTX_DISABLE_SOFT_PROMPTS');
+  if (value && /^(1|true|yes|on)$/i.test(value)) return false;
+  return true;
+};
+
+const measureResponseSize = (toolResponse) => {
+  if (!toolResponse) return 0;
+  if (typeof toolResponse === 'string') return toolResponse.length;
+  if (typeof toolResponse === 'object') {
+    const content = toolResponse.content ?? toolResponse.output ?? toolResponse.text ?? '';
+    if (typeof content === 'string') return content.length;
+    try { return JSON.stringify(toolResponse).length; } catch { return 0; }
+  }
+  return 0;
+};
+
+const countMatches = (state, field) => {
+  const value = state?.[field];
+  if (Array.isArray(value)) return value.length;
+  if (Number.isFinite(value)) return Number(value);
+  return 0;
+};
+
+const buildPrompt = (kind, severity, message) => ({ kind, severity, message });
+
+export const evaluateSoftPrompt = ({ toolName, toolInput, toolResponse, state } = {}) => {
+  if (!toolName) return null;
+
+  const meaningfulReadCount = Number.isFinite(state?.meaningfulReadCount) ? state.meaningfulReadCount : 0;
+  const readFiles = Array.isArray(state?.readFiles) ? state.readFiles : [];
+  const touchedFiles = countMatches(state, 'touchedFiles');
+
+  if (toolName === 'Read') {
+    const size = measureResponseSize(toolResponse);
+    if (size > LARGE_RESPONSE_CHARS) {
+      return buildPrompt(
+        'large_read',
+        'med',
+        `devctx hint: that Read returned ~${Math.round(size / 1000)}KB. Consider smart_read({ mode: 'outline', paths: ['${toolInput?.path ?? toolInput?.file_path ?? '<path>'}'] }) for cheaper exploration, then smart_read({ mode: 'symbol' }) when you know the target.`,
+      );
+    }
+
+    if (meaningfulReadCount + 1 >= REPEATED_READ_THRESHOLD && touchedFiles === 0) {
+      return buildPrompt(
+        'repeated_reads',
+        'med',
+        `devctx hint: ${meaningfulReadCount + 1} sequential reads without writes. smart_context({ task: '<your task>' }) fetches curated multi-file context in one call (or smart_context({ paths: { from, to } }) to trace the import graph).`,
+      );
+    }
+  }
+
+  if (toolName === 'Grep' || toolName === 'SemanticSearch') {
+    const grepHits = readFiles.filter((p) => typeof p === 'string').length;
+    if (grepHits + 1 >= REPEATED_GREP_THRESHOLD) {
+      return buildPrompt(
+        'repeated_search',
+        'med',
+        `devctx hint: ${grepHits + 1} search calls already in this turn. smart_search({ query, intent: 'debug'|'implementation'|'explore', kinds: [...] }) ranks results by relevance and supports ADR/class/function filters.`,
+      );
+    }
+  }
+
+  return null;
+};
+
+export const shouldEmitSoftPrompt = (hookKey, now = Date.now(), throttleMs = DEFAULT_THROTTLE_MS) => {
+  if (!hookKey) return false;
+  const last = lastIssuedAt.get(hookKey);
+  if (last && now - last < throttleMs) return false;
+  return true;
+};
+
+export const markSoftPromptEmitted = (hookKey, now = Date.now()) => {
+  if (hookKey) lastIssuedAt.set(hookKey, now);
+};
+
+export const _resetSoftPromptThrottle = (hookKey) => {
+  if (hookKey) lastIssuedAt.delete(hookKey);
+  else lastIssuedAt.clear();
+};
+
+export const _internal = {
+  DEFAULT_THROTTLE_MS,
+  LARGE_RESPONSE_CHARS,
+  REPEATED_READ_THRESHOLD,
+  REPEATED_GREP_THRESHOLD,
+};