Foremerge catches coding agents' intent clashes before Git conflicts
Two agents editing different files toward incompatible plans slip past Git — Foremerge makes them declare intent first. Plus 8B fine-tuning on a 4 GB GPU.

Copy markdown
The conflict Git can't see
Two agents touch different files but chase incompatible plans — one rips out PaymentService, the other extends it — and Git merges both clean. Foremerge sits above Git and flags that clash as a semantic conflict before either agent writes code.
How it works: declare intent, get a warning
Each agent publishes its intent with a scope like symbol:PaymentService before implementing; deterministic rules compare declarations and surface incompatibilities. Claims only warn — they never lock files, symbols, or agents — so nothing stalls. Apache-2.0, written in Rust, one-line install.
Build this weekend: a coordinated worktree swarm
A SQLite store in the Git common dir shares coordination state across isolated worktrees, and a verification gate runs your checks (cargo test, etc.) before a claim clears. Point three Claude Code or Codex agents at one board and let them negotiate scope instead of clobbering each other.
The catch: advisory, and no benchmarks yet
Coordination is cooperative — agents have to opt in and declare intent honestly, so a rogue or unaware agent still slips through. The README also says published benchmark results don't exist yet: you get executable fixtures and a reproducible query harness, not proven numbers.
Elsewhere: fine-tune an 8B on a 4 GB laptop GPU
Soup fine-tunes and post-trains from a single soup.yaml, streaming the frozen base into VRAM one decoder layer at a time. It clocks Llama-3.1-8B in NF4 at 119.6 tok/s and 3.32 GB peak on an RTX 3050 laptop — pip install soup-cli[train] and go.
Elsewhere: a model that learns forever on 8 GB
Mini-AGI trains a byte-level model from scratch on an RTX 3070 laptop, reading a batch-1 stream with no frozen weights. Slowing the trunk to 0.1x learning rate cut forgetting to +0.0067 nats after 524K characters of one subject, versus +2.23 at a naive rate — near-zero catastrophic forgetting on consumer hardware.