Explore Code

display_name: Rigor Improve
short_description: Rigor Improve implementation leaf mode for isolated exploratory code adaptations.
default_prompt: On an isolated branch or worktree, make exploratory code adaptations conservatively, summarize the changes, and write CHANGESET.md TOP_RUNS.md and status.json into explore_outputs.


# Rigor Improve Code Policy

## Purpose

Use this skill only when exploratory code changes have been explicitly authorized.

## Requirements

- keep work on an isolated branch or worktree
- record `current_research` and experiment branch
- record source repository references when transplanting modules
- prefer the smallest viable adaptation over broad rewrites
- treat results as exploratory candidates, not trusted conclusions

## Avoid

- modifying the trusted baseline by default
- claiming reproduction success from exploratory changes
- freeform large-scale refactors
- using this skill as the end-to-end `current_research` explore orchestrator


#!/usr/bin/env python3
"""Build a conservative exploratory code-change plan."""

from __future__ import annotations

import argparse
import json
import re
from pathlib import Path
from typing import Any, Dict, List


SKIP_PARTS = {
    "__pycache__",
    ".git",
    "repro_outputs",
    "train_outputs",
    "analysis_outputs",
    "debug_outputs",
    "explore_outputs",
    "tmp",
}

CODE_SUFFIXES = {".py", ".yaml", ".yml", ".json", ".toml", ".ini"}
MODEL_PATTERN = re.compile(r"(model|network|backbone|encoder|decoder|adapter|lora|head|loss)", re.IGNORECASE)
TRAIN_PATTERN = re.compile(r"(train|trainer|optim|loss|config)", re.IGNORECASE)
TASK_KEYWORDS = {
    "classification": ("class", "imagenet", "knn", "linear", "log_regression"),
    "segmentation": ("seg", "segment", "mask", "ade20k", "m2f", "mask2former"),
    "detection": ("det", "detect", "detr", "coco", "box"),
    "depth": ("depth", "nyu", "dpt", "depther"),
    "text": ("text", "token", "clip", "dinotxt"),
    "pretrain": ("pretrain", "ssl", "teacher", "student", "gram", "distillation"),
}
COMMON_TOKENS = {"py", "yaml", "yml", "json", "toml", "ini", "run", "train", "eval", "config", "configs"}


def load_variant_spec(path: str) -> Dict[str, Any]:
    if not path:
        return {}
    return json.loads(Path(path).resolve().read_text(encoding="utf-8-sig"))


def load_structured_payload(path: str) -> Any:
    if not path:
        return {}
    return json.loads(Path(path).resolve().read_text(encoding="utf-8-sig"))


def normalize_task_family(value: Any) -> str:
    return str(value or "").strip().lower()


def focus_tokens(current_research: str, task_family: str) -> List[str]:
    tokens: List[str] = []
    for part in re.split(r"[^a-zA-Z0-9]+", current_research.lower()):
        if part and part not in COMMON_TOKENS and len(part) > 2:
            tokens.append(part)
    if task_family:
        tokens.append(task_family)
        tokens.extend(TASK_KEYWORDS.get(task_family, ()))
    ordered: List[str] = []
    for token in tokens:
        if token not in ordered:
            ordered.append(token)
    return ordered[:20]


def score_path(rel: str, task_family: str, tokens: List[str]) -> int:
    score = 0
    if MODEL_PATTERN.search(rel):
        score += 5
    if TRAIN_PATTERN.search(rel):
        score += 3
    if rel.endswith(".py"):
        score += 1
    lower = rel.lower()
    for token in TASK_KEYWORDS.get(task_family, ()):
        if token in lower:
            score += 4
    for token in tokens:
        if token in lower:
            score += 2
    if current_research_dir(rel, tokens):
        score += 3
    return score


def current_research_dir(rel: str, tokens: List[str]) -> bool:
    lower = rel.lower()
    slash_hits = [token for token in tokens if token in lower]
    return len(slash_hits) >= 2


def collect_candidate_edit_targets(repo: Path, current_research: str, task_family: str) -> List[str]:
    tokens = focus_tokens(current_research, task_family)
    scored: List[tuple[int, str]] = []