SYNAPTIC Sentinel

The vibe-coding security sentinel. Apache-2.0. Five deterministic scouts scan your repo; a multi-provider Brain Layer decides what really matters and how to fix it — all in the IDE, with cost visibility per session.

SYNAPTIC Sentinel is the security companion of the SYNAPTIC family (sibling of SYNAPTIC Expert). Where Expert helps you write AI-assisted code with traceability, Sentinel audits what gets written — your own code, your AI's code, your dependencies, your config — and explains why it matters. All capabilities are open under Apache-2.0; there is no premium tier, no proprietary gating, and the Brain Layer runs against any LLM provider you choose.

Built for three kinds of users:

• Developers shipping AI-assisted code who want to catch the classic mistakes (eval-of-user-input, SQL injection by concatenation, secrets in commits, vulnerable deps) before the PR.
• Tech leads who want SARIF in CI as a quality gate, a "living tome" of audited findings to ship as evidence, and per-session cost visibility to pick the cheapest LLM that still gives quality triage.
• AppSec engineers who want taint analysis + secrets + SCA + IaC + AI-anti-patterns in one place, with a Brain Layer that explains exploitability instead of just listing CVEs.

What it does

Five scouts, deterministic and parallel

The Scout Layer runs five auditors in parallel, each best-in-class for its category. Binaries are downloaded on-demand from their official GitHub releases with SHA-256 verification — no Docker, no Python toolchain on the client.

Multi-provider Brain Layer (BYOK any provider)

Three LLM agents wired into the scan flow:

• Triage Agent — for each finding, decides true positive / false positive / inconclusive with a confidence score and a one-sentence rationale.
• Context Agent — for confirmed true positives, explains the exploitability chain: entry point → propagation → sink → exposure.
• Remediation Agent — proposes a concrete fix, with a code snippet you can copy.

Pick a different provider per agent. Run Triage on DeepSeek v4-flash (cheap), Context on Anthropic Haiku (quality), Remediation on Ollama locally (free / private). Or use one provider for everything. Mix-and-match by cost / quality / privacy.

Cost visibility — see what each session actually costs you

At the end of every triage session you get a summary block:

Cost summary (~estimated — tokens are chars/4 proxy, ±15-20% vs provider usage):
  provider/model                           agent              calls      input     output   cost USD  latency
  anthropic/claude-haiku-4-5-20251001      triage         3 calls      450 in     180 out    $0.0014      1842ms avg
  anthropic/claude-haiku-4-5-20251001      context        2 calls      890 in     420 out    $0.0030      2104ms avg
  anthropic/claude-haiku-4-5-20251001      remediation    2 calls      870 in     510 out    $0.0034      2851ms avg
  Total: 2210 input tokens · 1110 output tokens · $0.0078 (~estimated USD)

The synaptic-sentinel cost-history [--limit N] sub-command shows a rollup across the last N sessions, grouped by provider+agent — useful to compare empirically (on your findings, not synthetic benchmarks) which provider gives you the best cost/quality trade-off.

Token counts are proxies (Math.ceil(text.length / 4)) — ±15-20% vs the provider's billed usage. Surfaced as ~estimated everywhere.

Configuring providers (three equivalent paths)

A. From the IDE — Command Palette → SYNAPTIC Sentinel: Configure Brain Layer Providers opens a Settings panel:

• Active Configuration — pick a provider + model for each agent (Triage / Context / Remediation) from a dropdown, then click Apply. The dropdown only lists providers whose API key is configured (plus Ollama if its daemon is reachable). Clicking Apply writes .sentinel/agents.yaml in your workspace — same file format the CLI reads (path B below).
• Managed Credentials — Save / Delete / Test API keys for 12 providers. Keys never cross to the webview in plain text (state is configured: boolean, real value lives in vscode.SecretStorage).
• Local Models — auto-discovers your Ollama daemon at localhost:11434 and lists models you've pulled. Models above 5 GB get a ⚠ heavy badge and the panel shows a one-time warning about RAM saturation (with a "Don't remind me again" toggle that persists cross-workspace).

B. From .sentinel/agents.yaml in your project root (versionable):

agents:
  triage:
    provider: deepseek
    model: deepseek-v4-flash
  context:
    provider: anthropic
    model: claude-haiku-4-5-20251001
  remediation:
    provider: ollama
    model: gemma3:4b

C. From the CLI — repeatable --agent-provider <agent>=<provider>/<model> flag:

synaptic-sentinel triage \
  --agent-provider triage=deepseek/deepseek-v4-flash \
  --agent-provider context=anthropic/claude-haiku-4-5-20251001 \
  --agent-provider remediation=ollama/gemma3:4b

API keys come from SENTINEL_<PROVIDER>_API_KEY environment variables (SENTINEL_ANTHROPIC_API_KEY, SENTINEL_OPENAI_API_KEY, SENTINEL_DEEPSEEK_API_KEY, etc). The legacy ANTHROPIC_API_KEY is also accepted for backward compatibility with v0.2.0. Keys never appear in command-line arguments; they always travel via environment variables or vscode.SecretStorage.

Backward-compatible: users on v0.2.0 with only ANTHROPIC_API_KEY set continue to work identically — Anthropic Haiku 4.5 runs all three agents (the implicit fallback).

Memory of the swarm

The three agents share a learning store on disk in your repo (.sentinel/colony.db — same .sentinel/ directory that holds agents.yaml): a triage pattern seen with strong evidence is pre-classified on the next scan without spending an LLM token. Backward-compat: if your repo still has the legacy .synaptic-sentinel/colony.db from v0.3.5 or earlier, the CLI keeps reading it and emits a log suggesting you move it; no automatic migration to avoid data loss.

Turnkey from install to first scan

1. Install the extension.
2. Command Palette → "SYNAPTIC Sentinel: Install Scanners" — downloads and verifies the five scout binaries to a per-user cache (~/.synaptic-sentinel/scanners). Once.
3. Command Palette → "SYNAPTIC Sentinel: Scan Workspace" — findings appear as inline diagnostics, in the Problems panel, and in the living tome side view.
4. (Optional, BYOK) Configure Brain Layer Providers → Triage Findings (Brain Layer) to enrich findings with classification + context + remediation.

Commands

Plus Code Actions on each finding: mark as false positive (suppressed in future scans), and copy suggested remediation.

CI-native

Sentinel's CLI bundle ships with the extension and is also runnable standalone. The CLI exports the audit tome to JSON, HTML (an audit report you can hand to a reviewer), and SARIF 2.1.0 (GitHub Code Scanning, Azure DevOps). The --fail-on <severity> flag turns the scan into a CI gate (exit code 2 if there are findings at or above the threshold).

The synaptic-sentinel cost-history sub-command lets you wire cost reports into your CI summary too.

Privacy and data flow

• Your code never leaves your machine for the deterministic scans. The five scouts run locally as child processes.
• For the Brain Layer (optional), each finding's snippet goes directly to the LLM provider you chose — no proxy, no middleman, no Synaptic backend. BYOK. With Ollama / LM Studio / vLLM, the code never leaves your machine at all.
• The audit memory (colony.db) lives in your repo's .sentinel/ directory (alongside agents.yaml). You decide whether to commit it. Workspaces from v0.3.5 or earlier: if your repo has .synaptic-sentinel/colony.db, the CLI keeps reading the legacy path (dual-read, no auto-migration to avoid data loss).

Known Limitations

Honest caveats from the first real cross-provider benchmark — none affect your real source code, but you should know them:

1. OpenAI gpt-5* models need a higher max_completion_tokens. Out-of-the-box gpt-5-nano exhausts its reasoning-token budget at 1024 tokens and returns empty content. Either raise the cap manually or pick a different OpenAI model for now.
2. Local LLMs with >5 GB models can saturate RAM on consumer hardware during long batch operations (pnpm benchmark:run). Normal triage on a handful of findings is fine. For benchmark batches, prefer the ≤3 GB tier — the Configure Brain Layer Providers panel now shows each Ollama model's size and a ⚠ badge for models above 5 GB (with a "Don't remind me again" option once you've made your peace with the trade-off). Recommended lightweight models: gemma3:4b (~3 GB), qwen2.5-coder:7b (~4.4 GB), llama3.2:3b (~2 GB).
3. Free-tier quotas exhaust quickly on Groq (100K tokens/day) and Gemini (RPM-limited). Fine for everyday triage on a few findings; use the paid tier for full benchmark runs.
4. Token counts are proxies (chars/4, ±15-20% vs the provider's billed usage). Cost USD is ~estimated everywhere it appears.

See the CHANGELOG v0.3.0 Known Issues section for the full list with rationale.

License

All packages — extension, Scout Layer, Brain Layer — are licensed under Apache License 2.0. Source code, including the three LLM agents, is open and redistributable.

Part of the SYNAPTIC family — Apache-2.0 toolkit for engineers shipping AI-assisted code with traceability.