fix(optimize): add TARGET fallback + enrich score feedback for GEPA

benchmark.sh: fall back to  when ATLAS_OPTIMIZE_TARGET unset
score.py: track find_optimal_crossing and SA direct as separate
  scorecard examples with diagnostic context for reflector
gate.sh: add positional-arg fallback for ATLAS_OPTIMIZE_TARGET

Build: N/A (config & Python shims only)

.atlas/benchmark.sh

@@ -14,7 +14,8 @@ OUT="$(mktemp)"
 
 # --- the repo's run command (auto-detected) --------------------------------
 # Edit this line if the campaign should run something different.
-( python3 "$ATLAS_OPTIMIZE_TARGET" ) >"$OUT" 2>&1
+TARGET="${ATLAS_OPTIMIZE_TARGET:-$1}"
+( python3 "$TARGET" ) >"$OUT" 2>&1
 RC=$?
 # ---------------------------------------------------------------------------
 

.atlas/gate.sh

@@ -8,7 +8,7 @@
 # rewrites tracked files (checkpoints, logs, and .atlas/ are already ignored).
 set -uo pipefail
 WT="${1:-$PWD}"
-TARGET="${2:-${ATLAS_OPTIMIZE_TARGET:-}}"
+TARGET="${ATLAS_OPTIMIZE_TARGET:-$2}"
 REL="${TARGET#"$WT"/}"
 IGNORE_RE='^(outputs/|out/|checkpoints/|runs/|wandb/|\.atlas/|.*\.log$|.*\.ckpt$|.*\.pt$|.*\.bin$|.*\.safetensors$)'
 CHANGED="$(git -C "$WT" status --porcelain --untracked-files=no 2>/dev/null \

.atlas/score.py

@@ -1,65 +1,73 @@
 #!/usr/bin/env python3
-"""Generated by `atlas autoresearch`. Parse one metric out of a run's output
+import argparse, json, os, re, sys
-into the Atlas benchmark contract.
-
-EDIT the METRIC line if the wrong number is being read, or rewrite parse() for
-a richer evaluator (per-example feedback is what makes GEPA converge — see
-.atlas/benchmark-contract or the atlas-optimize skill)."""
-import argparse
-import json
-import os
-import re
-import sys
-
-# autoresearch guessed this metric name from your --goal.
-METRIC = "find_optimal_crossing"          # <-- EDIT ME if this is the wrong metric
-HIGHER_IS_BETTER = False
-
-# Matches "<metric> = 0.74", "<metric>: 0.74", "<metric> 0.74".
-PATTERN = re.compile(
-    re.escape(METRIC) + r"\s*[:=]?\s*([+-]?[0-9]*\.?[0-9]+(?:[eE][+-]?[0-9]+)?)"
-)
 
+METRICS = {
+    "find_optimal_crossing": False,
+    "SA direct": False,
+}
 
 def parse(text):
-    matches = PATTERN.findall(text)
+    results = {}
-    if not matches:
+    for name, higher in METRICS.items():
-        return None
+        pat = re.compile(re.escape(name) + r"\s*[:=]?\s*([+-]?[0-9]*\.?[0-9]+)")
-    return float(matches[-1])  # last occurrence == final epoch / final eval
+        for m in pat.finditer(text):
-
+            raw = m.group(0).strip()
+            val = float(m.group(1))
+            line_before = text[max(0, m.start()-80):m.start()].split("\n")[-1].strip()
+            results[name] = (val, higher, raw, line_before)
+    return results
 
 def main():
     ap = argparse.ArgumentParser()
-    ap.add_argument("--stdout", required=True, help="file holding the run's captured output")
+    ap.add_argument("--stdout", required=True)
     args = ap.parse_args()
     text = open(args.stdout, encoding="utf-8", errors="replace").read()
-    score = parse(text)
+    parsed = parse(text)
-    if score is None:
+
-        sys.stderr.write(
+    if "find_optimal_crossing" not in parsed:
-            f"[score] could not find metric '{METRIC}' in the run output.\n"
+        sys.stderr.write(f"[score] could not find metric in output.\n")
-            "        Fix: edit .atlas/score.py (METRIC=...), or pass an explicit\n"
-            "        --benchmark to 'atlas autoresearch'.\n"
-        )
         sys.exit(3)
-    direction = "higher is better" if HIGHER_IS_BETTER else "lower is better"
+
+    primary_name = "find_optimal_crossing"
+    primary_val, _, primary_raw, ctx = parsed[primary_name]
+    direction = "lower is better"
+
+    examples = [{
+        "id": primary_name,
+        "score": primary_val,
+        "pass": True,
+        "feedback": f"{primary_raw} | context: {ctx}" if ctx else primary_raw,
+    }]
+
+    if "SA direct" in parsed:
+        sa_val, _, sa_raw, sa_ctx = parsed["SA direct"]
+        examples.append({
+            "id": "SA direct",
+            "score": sa_val,
+            "pass": sa_val < 100,
+            "feedback": f"{sa_raw} | context: {sa_ctx}" if sa_ctx else sa_raw,
+        })
+
+    diag_lines = []
+    for line in text.strip().split("\n"):
+        stripped = line.strip()
+        if "find_optimal_crossing" in stripped or "SA direct" in stripped:
+            diag_lines.append(stripped)
+    diagnostic = "; ".join(diag_lines) if diag_lines else primary_raw
+
     result = {
-        "score": score,
+        "score": primary_val,
-        "examples": [{
+        "examples": examples,
-            "id": METRIC,
+        "feedback": f"find_optimal_crossing={primary_val} ({direction}) | {diagnostic}",
-            "score": score,
-            "pass": True,
-            "feedback": f"{METRIC} = {score} (parsed from run output; {direction})",
-        }],
-        "feedback": f"Final {METRIC} = {score}.",
     }
+
     out = os.environ.get("ATLAS_OPTIMIZE_RESULT")
     payload = json.dumps(result)
     if out:
-        with open(out, "w", encoding="utf-8") as fh:
+        with open(out, "w") as fh:
             fh.write(payload)
     else:
         sys.stdout.write(payload)
 
-
 if __name__ == "__main__":
     main()