The Agent analyses telemetry and exposes ranked issues via GET /agent/issues. In the dashboard they appear inside Issue List (SDK events) on the Security tab and feed Overview/Summary counts. Issues load when you open an integration, not only on Security.

• Errors — spikes, trends, regressions and high-impact recurring failures.
• Vulnerabilities — findings from the built-in scanner, ranked by severity.
• Performance — slow LCP/TTFB/CLS per route (e.g. Slow LCP on /).
• Instability — crash loops, error bursts and reliability signals.
• UI/UX — from TestingAI; shown in agent API but filtered from Security tab noise.

Each integration receives an engine grade from 0 to 100. The grade reflects how much actionable signal the agent is producing, not just whether problems exist:

• 85–100 Excellent — diverse, high-confidence, fixable signal.
• 70–84 Good — healthy detection with multiple actionable items.
• 50–69 Fair — detection is running but coverage can improve.
• Below 50 — the agent is mostly silent; verify your SDK is reporting data.

System Health is a 0–100 score with an A–F grade. On the Overview tab the score uses 24h telemetry compared to the previous day; issue counts in the same card use the 7-day SDK + agent window. Unlike naive scoring, penalties are capped and logarithmic so one noisy signal cannot zero the score.

How it is computed:

• Each issue contributes a penalty weighted by severity (critical >> high > medium > low > info).
• Penalties decay logarithmically — the second similar issue costs about 40% of the first, the third about 28%, and so on.
• Each category has its own cap: errors and vulnerabilities can hurt at most 30 points each, performance 18, UI 12.
• The total penalty is capped at 75 points, and the score has a hard floor of 25 — the engine never reports a panic-zero.
• Low-confidence detections automatically penalise less (×0.7 for medium, ×0.4 for low).
• When the dataset is small, the penalty is softened by a data-confidence multiplier to avoid over-reacting.

Grades:

• A (90–100) — Excellent.
• B (80–89) — Healthy.
• C (65–79) — Stable with watchpoints.
• D (50–64) — Needs attention.
• F (< 50) — At risk.

The same number is exposed to AssistAI as live context, so when you ask “how is my project doing?” the answer is grounded in the real System Health value rather than a guess.

Configure under Settings → Integrations → [integration] → Agent Autonomous:

• Suggest only — runner prepares fixes; you apply manually from Security (View Fix → Commit Correction).
• Autonomous — cron pipeline (every 5 min when enabled): locate → AI → apply on GitHub. Use Run now for an immediate cycle. Requires Pro + linked repo.

Issue tags and severity overrides in Security are stored in browser localStorage per integration (sublyzer:issueOverrides:…).

When a GitHub repository is connected, the agent indexes your code with a combination of symbol graphs (Tree-sitter), AST diffs between commits and a retrieval-augmented context layer (RAG). This lets it relate an error stack trace to the precise function and recent change that caused it, rather than guessing from a message string.

The agent also auto-detects your project type — web SPA, SSR app, REST API, mobile, or executable — and adapts its analysis strategy accordingly.

AI usage is billed under your existing plan. Each plan includes a monthly AI allowance; if you exceed it, additional usage is metered and billed transparently via Stripe with a fixed margin on top of the underlying provider cost. You can cap or pause AI per integration from the settings page.

• Auto-Patch documentation
• TestingAI runs
• Connect a GitHub repository