AUTHERGUARD AI

# What Makes AetherGuard Unique

AetherGuard is a CLI-installed, local-first daemon that watches a repo, uses Gemini

to predict next-logical code, and audits/auto-fixes bugs and vulnerabilities in

real time. A few implementation details set it apart from typical "AI coding

assistant" side projects.

# 1. Auto-applies security fixes to live files — unattended

triggerAutoAudit() doesn't just flag critical vulnerabilities, it patches them

directly on disk when AUTO_FIX_ENABLED is on — but only after verifying the

exact old code still matches at that line:

that guard against stale or hallucinated line numbers is the detail that

separates "cool demo" from "safe to run unattended."

## 2. Debounced, watch-triggered self-healing loop

File change → chokidar fires → 2s debounce collapses rapid edits into one

audit → Gemini call → auto-fix → write back to disk, in a closed loop with no

human in it unless AUTO_FIX_ENABLED=false. Every applied fix is logged

(`autoFixLog`, capped at 200 entries) so there's an audit trail of what the AI

changed and why.

3. Model failover baked into every AI call

callGeminiWithFailover tries gemini-2.5-flash, then falls back to

gemini-2.0-flash-lite, with retry-on-transient-error logic (503, 429,

"overloaded", etc.). Most hobby AI tools call the model once and just fail;

this treats API flakiness as an expected condition.

## 4. BYOK as a first-class runtime feature, not just an env var

/api/set-key lets a user set their Gemini key at runtime through the UI — no

restart needed — and /api/key-status reports whether the key came from env

or was set live. That's built for handing the tool to other people, not just

running it yourself.

# 5. Path-traversal guard on live file read/write

readLiveFile / writeLiveFile check fullPath.startsWith(TARGET_DIR) before

touching disk. Since this tool can read and write arbitrary files in a

directory you point it at, that check is load-bearing, not decorative.

## 6. The prediction UI sells "pattern learning" at the UX level

CodeEditor.tsx badges predictions as "Pattern Matches Found" and offers

ranked alternatives (not just one completion), with an "Auto-Inject

Completion" action that splices into the exact cursor line — reinforcing the

repo-pattern-learning pitch in the UI, not just in the system prompt.

# Honest caveat: pattern learning is re-derived, not persisted

The "pattern learning" is prompt-driven — Gemini reads the full repo content

on every audit call and is instructed to detect patterns fresh each time. It's

not a trained or persisted model. In practice, "learns your patterns" means

"re-derives them from full repo context every audit," which is simpler than it

sounds and currently has no long-term memory of patterns between sessions

unless the patterns output is being stored somewhere outside server.ts.