feat: semantic context injection via Chroma on UserPromptSubmit (#1568)

* feat: semantic context injection via Chroma on every UserPromptSubmit

On each prompt, queries ChromaDB for the top-N most relevant past
observations and injects them as additionalContext. Replaces the
recency-based "last N observations" approach with relevance-based
semantic search.

Changes:
- session-init.ts: After session init, query /api/context/semantic
  with user's prompt text. If results found, return as
  hookSpecificOutput with hookEventName 'UserPromptSubmit'.
- SearchRoutes.ts: New GET /api/context/semantic endpoint that queries
  SearchManager with format='json' and formats results as markdown.
- SettingsDefaultsManager.ts: New settings CLAUDE_MEM_SEMANTIC_INJECT
  (default: true) and CLAUDE_MEM_SEMANTIC_INJECT_LIMIT (default: 5).

Key behaviors:
- Fires on every UserPromptSubmit (not just SessionStart)
- Minimum prompt length: 20 chars (skips "ok", "yes", etc.)
- Skips media-only prompts
- Graceful degradation: if worker/Chroma unavailable, no injection
- Survives /clear: re-injects on next prompt (not session-bound)
- Uses workerHttpRequest (v10.6.3 API, not raw fetch)

Production data (23 days, 3,400+ observations):
- Before: 8 most recent observations (often irrelevant to current topic)
- After: 5 most relevant observations (semantic match)
- Token cost: ~1800 → ~800-1200 per injection

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: address CodeRabbit review on PR #1568

- session-init: don't skip semantic injection when contextInjected=true
  (only skip agent re-init, semantic lookup must run every prompt)
- session-init: normalize SEMANTIC_INJECT toggle via String().toLowerCase()
- semantic endpoint: change from GET to POST to avoid URL-length limits
  and prompt exposure in access logs. Handler accepts both body and query
  for backwards compatibility.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Alessandro Costa <alessandro@claudio.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

src/cli/handlers/session-init.ts

@@ -87,17 +87,18 @@ export const sessionInitHandler: EventHandler = {
 
     // Skip SDK agent re-initialization if context was already injected for this session (#1079)
     // The prompt was already saved to the database by /api/sessions/init above —
-    // no need to re-start the SDK agent on every turn
-    if (initResult.contextInjected) {
+    // no need to re-start the SDK agent on every turn.
+    // Note: we do NOT return here — semantic injection below must run on every prompt.
+    const skipAgentInit = Boolean(initResult.contextInjected);
+    if (skipAgentInit) {
       logger.info('HOOK', `INIT_COMPLETE | sessionDbId=${sessionDbId} | promptNumber=${promptNumber} | skipped_agent_init=true | reason=context_already_injected`, {
         sessionId: sessionDbId
       });
-      return { continue: true, suppressOutput: true };
     }
 
     // Only initialize SDK agent for Claude Code (not Cursor)
     // Cursor doesn't use the SDK agent - it only needs session/observation storage
-    if (input.platform !== 'cursor' && sessionDbId) {
+    if (!skipAgentInit && input.platform !== 'cursor' && sessionDbId) {
       // Strip leading slash from commands for memory agent
       // /review 101 -> review 101 (more semantic for observations)
       const cleanedPrompt = prompt.startsWith('/') ? prompt.substring(1) : prompt;
@@ -115,14 +116,58 @@ export const sessionInitHandler: EventHandler = {
         // Log but don't throw - SDK agent failure should not block the user's prompt
         logger.failure('HOOK', `SDK agent start failed: ${response.status}`, { sessionDbId, promptNumber });
       }
-    } else if (input.platform === 'cursor') {
+    } else if (!skipAgentInit && input.platform === 'cursor') {
       logger.debug('HOOK', 'session-init: Skipping SDK agent init for Cursor platform', { sessionDbId, promptNumber });
     }
 
+    // Semantic context injection: query Chroma for relevant past observations
+    // and inject as additionalContext so Claude receives relevant memory each prompt.
+    // Controlled by CLAUDE_MEM_SEMANTIC_INJECT setting (default: true).
+    const semanticInject =
+      String(settings.CLAUDE_MEM_SEMANTIC_INJECT).toLowerCase() === 'true';
+    let additionalContext = '';
+
+    if (semanticInject && prompt && prompt.length >= 20 && prompt !== '[media prompt]') {
+      try {
+        const limit = settings.CLAUDE_MEM_SEMANTIC_INJECT_LIMIT || '5';
+        const semanticRes = await workerHttpRequest('/api/context/semantic', {
+          method: 'POST',
+          headers: { 'Content-Type': 'application/json' },
+          body: JSON.stringify({ q: prompt, project, limit })
+        });
+        if (semanticRes.ok) {
+          const data = await semanticRes.json() as { context: string; count: number };
+          if (data.context) {
+            additionalContext = data.context;
+            logger.debug('HOOK', `Semantic injection: ${data.count} observations for prompt`, {
+              sessionId: sessionDbId, count: data.count
+            });
+          }
+        }
+      } catch (e) {
+        // Graceful degradation — semantic injection is optional
+        logger.debug('HOOK', 'Semantic injection unavailable', {
+          error: e instanceof Error ? e.message : String(e)
+        });
+      }
+    }
+
     logger.info('HOOK', `INIT_COMPLETE | sessionDbId=${sessionDbId} | promptNumber=${promptNumber} | project=${project}`, {
       sessionId: sessionDbId
     });
 
+    // Return with semantic context if available
+    if (additionalContext) {
+      return {
+        continue: true,
+        suppressOutput: true,
+        hookSpecificOutput: {
+          hookEventName: 'UserPromptSubmit',
+          additionalContext
+        }
+      };
+    }
+
     return { continue: true, suppressOutput: true };
   }
 };