feat: Sentinel AI Learning Engine & Stitch-Loop Orchestrator by jmbish04 · Pull Request #453 · jmbish04/core-github-api

jmbish04 · 2026-03-31T20:20:50Z

Summary

Phase 1: StitchService + StitchLoopWorkflow — durable 3-step pipeline (enhance prompt → generate UX via Stitch MCP → Jules implementation)
Phase 2: 11 learning micro-domain Drizzle tables (sessions, threads, messages, enrichment, tags, insights, PR reflections, etc.)
Phase 3: Sentinel API routes (/api/sentinel/tasks, insights, orchestrate-ui, health/learning) with auth middleware
Phase 4: LearningAgent DurableObject — cron-driven ingestion, MCP enrichment, AI analysis, Contemplation Gate, vectorization
Phase 5: PR interceptors (SentinelInterceptor on open/sync, SentinelPostMerge on merge) using PAT auth for human-persona comments
Phase 6: Frontend control plane — SentinelDashboard (Recharts), SentinelKanban (5-column), repo-scoped SentinelHud

36 files changed — 28 new files, 8 modified existing files.

Test plan

wrangler dev starts without errors
pnpm run db:generate:all produces migration files for 11 new learning tables
POST /api/sentinel/health/learning returns health status
GET /api/sentinel/insights returns empty array (no data yet)
POST /api/sentinel/orchestrate-ui triggers StitchLoopWorkflow
Create test PR → verify SentinelInterceptor posts analysis comment
Navigate to /sentinel → dashboard renders with Recharts
Navigate to /sentinel/kanban → 5-column Kanban renders
Navigate to /repos/:owner/:repo/sentinel → HUD shows repo-scoped insights

🤖 Generated with Claude Code

Implements the full 6-phase Sentinel system: - Phase 1: StitchService + StitchLoopWorkflow (durable 3-step pipeline) - Phase 2: 11 learning micro-domain tables (sessions, threads, insights, reflections, etc.) - Phase 3: Sentinel API routes (/api/sentinel/tasks, insights, orchestrate, health) - Phase 4: LearningAgent DurableObject with contemplation gate + vectorization - Phase 5: PR interceptors (SentinelInterceptor + SentinelPostMerge) using PAT auth - Phase 6: Frontend control plane (Dashboard, Kanban, repo-scoped HUD) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

gemini-code-assist

Code Review

This pull request introduces the Sentinel Learning Engine, a comprehensive system for architectural pattern detection and automated UI orchestration. It includes the LearningAgent Durable Object for ingesting and analyzing conversations, new GitHub automations (SentinelInterceptor, SentinelPostMerge) for PR analysis, and the StitchLoopWorkflow for design-to-code pipelines. While the implementation is feature-rich, several critical issues were identified regarding database performance and logic. Specifically, multiple N+1 query patterns in the LearningAgent will cause performance degradation, and an empty where clause in the pattern analysis query effectively breaks the ingestion filter. Additionally, the PR diff truncation in the interceptor is too aggressive for meaningful AI analysis, and the singleton pattern used for the Stitch service may lead to environment leakage across requests in the Cloudflare Workers environment.

gemini-code-assist · 2026-03-31T20:28:10Z

src/backend/src/ai/agents/LearningAgent.ts

+    const analyzed = await db
+      .select()
+      .from(learningMessages)
+      .where(
+        and(
+          // Has AI analysis
+          // We check for non-null ai_analysis
+        )
+      )
+      .limit(20);


The where clause in this query is effectively empty because the and() function has no arguments. This will cause the query to return the first 20 messages regardless of whether they have been analyzed or not. If the first 20 messages in the table are new and unenriched, the filter at line 304 will result in an empty array, and the agent will fail to process any insights. You should add a condition like isNotNull(learningMessages.aiAnalysis).

Fixed — replaced empty and() with isNotNull(learningMessages.aiAnalysis) in commit b5495d7.

gemini-code-assist · 2026-03-31T20:28:10Z

src/backend/src/ai/agents/LearningAgent.ts

+    const completedSessions = await db
+      .select()
+      .from(julesSessions)
+      .where(eq(julesSessions.status, "completed"))
+      .orderBy(desc(julesSessions.createdAt))
+      .limit(20);
+
+    for (const session of completedSessions) {
+      // Check if thread already exists for this session
+      const existing = await db
+        .select()
+        .from(learningThreads)
+        .where(eq(learningThreads.sourceIdentifier, session.id))
+        .limit(1);
+
+      if (existing.length > 0) continue;


This loop performs a database query (learningThreads) for every completedSession found. This is an N+1 query pattern that will become a performance bottleneck as the number of sessions grows. Consider using a join or a NOT EXISTS subquery to fetch only the sessions that haven't been ingested yet in a single query.

Fixed — pre-fetches all ingested sourceIdentifiers into a Set before the loop, eliminating N+1 queries. Commit b5495d7.

gemini-code-assist · 2026-03-31T20:28:10Z

src/backend/src/ai/agents/LearningAgent.ts

+        const thread = await db
+          .select()
+          .from(learningThreads)
+          .where(eq(learningThreads.id, msg.threadId))
+          .limit(1);


This is an N+1 query pattern. You are fetching the thread details for every analyzed message inside a loop. Consider fetching all required threads in a single query before entering the loop.

Fixed — batch-fetches all referenced threads into a Map<id, thread> before the loop. Commit b5495d7.

gemini-code-assist · 2026-03-31T20:28:10Z

src/backend/src/automations/pr/SentinelInterceptor.ts

+
+**Diff (truncated to 3000 chars):**
+\`\`\`
+${diff.substring(0, 3000)}


Truncating the PR diff to only 3000 characters is likely too aggressive. Most meaningful PRs will have diffs significantly larger than this, and the AI will lose the context necessary to perform a valid architectural analysis. Consider increasing this limit to at least 50,000 characters or using a more sophisticated truncation strategy that preserves file headers and changed lines.

Fixed — increased diff truncation from 3,000 to 50,000 characters. Commit b5495d7.

gemini-code-assist · 2026-03-31T20:28:10Z

src/backend/src/ai/agents/LearningAgent.ts

+
+    for (const msg of analyzedWithAnalysis) {
+      try {
+        const parsed = JSON.parse(msg.aiAnalysis!);


Directly calling JSON.parse() on LLM output is risky. LLMs often wrap JSON responses in markdown code blocks (e.g., json ... ), which will cause JSON.parse to throw an error. It is safer to use a utility function that extracts the JSON content from the string before parsing.

Fixed — added safeParseJson() helper that strips markdown code fences before parsing, with graceful null return on failure. Commit b5495d7.

gemini-code-assist · 2026-03-31T20:28:11Z

src/backend/src/services/stitch/service.ts

+  private static instance: StitchService;

-  private client: StitchToolClient;
-
-  private constructor(apiKey: string) {
-    this.client = new StitchToolClient({ apiKey });
-  }
+  private constructor(private readonly env: Env) {}

  /**
-   * Returns the singleton, lazily connecting with the resolved API key.
-   * Must be called with `await` since key resolution may be async.
+   * Returns the singleton `StitchService` for the current request context.
   */
-  static async getInstance(env: Env): Promise<StitchService> {
+  public static getInstance(env: Env): StitchService {
    if (!StitchService.instance) {
-      const apiKey = await resolveApiKey(env);
-      StitchService.instance = new StitchService(apiKey);
+      StitchService.instance = new StitchService(env);
    }
    return StitchService.instance;
  }


The singleton implementation for StitchService captures the env object in a static property. In Cloudflare Workers, global state can persist across multiple requests within the same isolate. If getInstance is called with a different env (e.g., during testing or in a multi-tenant setup), it will return the instance initialized with the previous request's environment. It is better to avoid storing env in a static singleton or to ensure the instance is truly request-scoped.

Fixed — singleton now uses WeakRef<Env> comparison to detect stale env across requests in the same isolate. Commit b5495d7.

The workflow referenced `analyze_drizzle_schema.py` but the actual script is named `audit_drizzle_schema.py`. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Fix empty and() where clause in analyzePatterns — now uses isNotNull(aiAnalysis) - Eliminate N+1 query in ingestSessions — pre-fetch ingested IDs into a Set - Eliminate N+1 query in analyzePatterns — batch-fetch threads into a Map - Add safeParseJson helper to handle LLM markdown-fenced JSON output - Increase PR diff truncation from 3000 to 50000 chars for meaningful analysis - Make StitchService singleton request-scoped via WeakRef env comparison Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Merge origin/main (PR #452's Sentinel infra) into claude/sentinel-engine - Remove duplicate kebab-case schema files (ai-insights.ts, ai-pr-reflections.ts, etc.) in favor of main's camelCase versions (aiInsights.ts, aiPrReflections.ts) - Update all imports to use main's schema exports (learningAiInsights, etc.) - Merge LearningAgent: keep main's pattern detection + contemplation gate, add Sentinel pipeline routes (/ingest, /enrich, /schedule/run, /ingest-pr) - Remove duplicate v7 migration (LearningAgent already in v1_sentinel) - Restore Sentinel frontend routes in App.tsx Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Resolve conflicts from PR #450 architecture restructuring: - Integrate with mountRoutes() pattern in routes/index.ts - Add LearningAgent to ai/agents/exports.ts barrel - Add StitchLoopWorkflow to workflows/exports.ts barrel - Add Sentinel routes to GlobalRoutes.tsx and RepoRoutes.tsx - Add stitch-loop-workflow to wrangler.jsonc workflows array - Add db:auto script to package.json Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

gemini-code-assist bot reviewed Mar 31, 2026

View reviewed changes

jmbish04 and others added 5 commits March 31, 2026 13:37

fix(ci): correct schema analysis script filename in workflow

e5da12d

The workflow referenced `analyze_drizzle_schema.py` but the actual script is named `audit_drizzle_schema.py`. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Merge remote-tracking branch 'origin/main' into claude/sentinel-engine

d05eb9c

jmbish04 merged commit b93293b into main Mar 31, 2026
1 check failed

jmbish04 deleted the claude/sentinel-engine branch March 31, 2026 23:10

This was referenced Mar 31, 2026

Refactor: Implement Dual-Scope Routing and DO Abstraction #448

Closed

Implement Autonomous Vibe Coding Orchestration Pipeline #42

Closed

Conversation

jmbish04 commented Mar 31, 2026

Summary

Test plan

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

jmbish04 Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

jmbish04 Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

jmbish04 Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

jmbish04 Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

jmbish04 Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

jmbish04 Mar 31, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant