AutoScout-Engine

The "engine" to AutoScout-Lab's "fuel." AutoScout-Lab generates a new agentic-AI prototype repo every day and gives each one small daily increments. This repo runs a separate, deeper pass: it researches what's currently happening around a repo's specific problem, then advances that repo toward it — a real upgrade, not just a small tweak.

Why a separate repo, and why Groq

• Separate from AutoScout-Lab so the two systems don't compete for the same budget or step on each other's commits.
• Powered by Groq instead of Gemini — its free tier renews every day (unlike a one-time credit pool), so this can run indefinitely without ever needing a manual top-up.
• Model: llama-3.3-70b-versatile. This engine only makes ONE call a day, so a small model's higher daily-request quota was never actually useful here — even the 70B model's lower 1,000-requests/day cap is 1,000x more than needed at this cadence. What matters at one call/day is per-call token budget (12,000 tokens/minute, 2x the 8B model's 6,000) and code judgment quality, and the 70B model wins both. (A first live run on the 8B model produced a real regression — a working Gemini API call replaced with a hardcoded stub — plus a fabricated log entry, which is why quality won out over quota headroom here.)
• Runs daily — since Groq's quota resets every day, there's no scarcity reason to throttle the cadence the way NVIDIA's one-time pool would have required.

How a cycle works

1. Sync the local repo registry against GitHub's actual repo list (adds any AutoScout-generated repo not yet tracked, drops any deleted).
2. Pick the ONE repo most overdue for review — never-reviewed repos first (oldest created), then oldest-last-reviewed.
3. Fetch that repo's current files (capped to fit the model's token budget).
4. Research signals specific to that repo's problem — targeted Hacker News and GitHub searches (free, no Groq tokens spent) — so the model isn't reasoning from training-data knowledge alone.
5. Ask Groq's llama-3.3-70b-versatile to combine that research with its own knowledge and propose ONE substantial advancement — with explicit guardrails against faking/stubbing out real functionality and against inventing log history that didn't happen.
6. Commit the change straight to that repo's main, and log it in that repo's own ADVANCEMENT_LOG.md.

Setup

This repo needs two secrets (add via Settings → Secrets and variables → Actions → New repository secret — never paste them anywhere else):

• SCOUT_PAT — a GitHub PAT with repo scope (same one AutoScout-Lab uses)
• GROQ_API_KEY — a free key from console.groq.com (sign up, go to API Keys, create one — no credit card needed)

State

state/registry.jsonl tracks every repo this engine knows about and when it last reviewed each one. It's independent from AutoScout-Lab's own repos/registry.jsonl — same discovery method (GitHub repos with the description "Auto-generated AI prototype by AutoScout"), kept separate on purpose so neither system depends on the other's internals.