RFC: optional add-ons for reproducibility, token cost, and cross-run memory (happy to PR in small pieces)

Hi, and thanks for AI-Scientist-v2. We've been running it and built a set of optional, default-off add-ons in a fork, and would like to know whether you'd be open to upstreaming some of them as small, independent PRs before we send anything.

Fork: https://github.com/EvolvingAgentsLabs/AI-Scientist-v2 (the full design doc lives in UPGRADE_PLAN.md).

Motivation

Running v2 we hit four recurring pain points (consistent with the README's own notes on v2's open-ended, lower-success-rate nature):

1. Reproducibility — a successful experiment is hard to reproduce exactly (temperature, the exact code+reasoning that produced it).
2. Token cost — the improve/debug loop re-sends the full experiment script each iteration.
3. Backtracking — when a debug branch hits a dead end there's no clean "restore the last good state."
4. No memory across runs — each run re-learns the same failures (e.g. "this optimizer/lr diverges on this data").

What we built (each is opt-in via config; with flags off, the pipeline is byte-for-byte unchanged; each has offline unit tests)

Ranked by how self-contained they are (easiest to review first):

#	Add-on	New deps?	Surface area
A	strict-patch editing — in the improve step the model emits a line-precise patch instead of a full rewrite; deterministic applier with safe fallback to full code. ~10x fewer output tokens on large files.	none	ai_scientist/dialects/, one branch in parallel_agent._improve, flag agent.code.patch_mode
B	discipline-aware reviewers — infer the paper's field and tailor the reviewer persona (numerical-stability / statistical-rigor / proof-rigor / physical-correctness lenses). Additive wrapper over perform_review.	none	ai_scientist/reviewers/
C	measurement harness — scripts/measure_run.py reports $/run, tokens/stage, buggy-vs-good, debug-depth, from the artifacts a run already writes.	none	new script + a token-summary dump in save_run
D	role-isolated improve pipeline — Researcher (frontier) → Coder (small/local) in separate contexts, so a mid-size/local model can do the mechanical coding cheaply.	none	ai_scientist/pipeline/
E	cross-run "lab memory" — an offline pass distills durable axioms (successes/failures/constraints) from past journals into an append-only wiki that future runs pull into their prompts.	none	ai_scientist/dream/, flag use_lab_memory
F	multidimensional VCS of the search tree — each BFTS node becomes a commit of (code + hypothesis + models + trace); the best node is "frozen" into a deterministic, replayable recipe shippable as supplementary material; real rollback.	optional dep on agentvcs (separate OSS project), soft-imported and fail-safe	ai_scientist/agentvcs_bridge.py, hooks in journal.py/agent_manager.py/config.py, flag agentvcs.enabled

What we're asking

• Would you accept any of A–E as small standalone PRs? They add no dependencies and default to off.
• F is the most opinionated and carries an optional dependency on a separate project of ours; we'd only propose it if there's interest, and would keep it strictly opt-in. Totally fine if it stays in our fork.

We'll open one PR per item, each branched off main, minimal diff, with tests. Happy to adjust scope/naming to your conventions. Thanks!