core: Registry, Skills, Router, Executor, Verifier, Runner mit Tests

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

core/src/knaster/problems.py

@@ -0,0 +1,71 @@
+"""Signale (Fehlerklassen) und ihre Erkennung."""
+
+from __future__ import annotations
+
+import json
+import re
+from enum import StrEnum
+
+
+class Signal(StrEnum):
+    NO_SKILL = "no_skill"
+    SKILL_CONFLICT = "skill_conflict"
+    SLOT_UNAVAILABLE = "slot_unavailable"
+    CONTEXT_EXCEEDED = "context_exceeded"          # präventiv vom Router erkannt
+    COMPACT_PROMPT = "compact_prompt"              # Ausweichen auf kompakte Skill-Variante
+    LARGER_SLOT = "larger_slot"                    # Ausweichen auf Slot mit mehr Kontext
+    NEEDS_SPLIT = "needs_split"
+    OUTPUT_TRUNCATED = "output_truncated"
+    DEGENERATE_OUTPUT = "degenerate_output"
+    INVALID_FORMAT = "invalid_format"
+    CONTEXT_OVERFLOW = "context_overflow"          # Engine meldet Overflow
+    ENDPOINT_ERROR = "endpoint_error"
+    VERIFIER_FAILED = "verifier_failed"
+    CLASSIFIER_LOW_CONFIDENCE = "classifier_low_confidence"
+    PROGRAM_ERROR = "program_error"
+    MISROUTE_SUSPECTED = "misroute_suspected"
+
+
+INFRA_SIGNALS = {Signal.ENDPOINT_ERROR, Signal.CONTEXT_OVERFLOW}
+QUALITY_SIGNALS = {
+    Signal.VERIFIER_FAILED,
+    Signal.INVALID_FORMAT,
+    Signal.DEGENERATE_OUTPUT,
+    Signal.OUTPUT_TRUNCATED,
+    Signal.CLASSIFIER_LOW_CONFIDENCE,
+    Signal.PROGRAM_ERROR,
+}
+
+_OVERFLOW_PATTERNS = re.compile(
+    r"context length|maximum context|too many tokens|context window|prompt is too long", re.IGNORECASE
+)
+_WORD = re.compile(r"\w+", re.UNICODE)
+
+
+def is_context_overflow_message(text: str) -> bool:
+    return bool(_OVERFLOW_PATTERNS.search(text or ""))
+
+
+def repetition_ratio(text: str, n: int = 4) -> float:
+    """Anteil der n-Gramme, die Wiederholungen sind (0 = keine)."""
+    words = _WORD.findall(text.lower())
+    if len(words) < n:
+        return 0.0
+    grams = [tuple(words[i:i + n]) for i in range(len(words) - n + 1)]
+    return 1.0 - len(set(grams)) / len(grams)
+
+
+def is_degenerate(text: str, min_words: int = 40, max_ratio: float = 0.5) -> bool:
+    return len(_WORD.findall(text)) >= min_words and repetition_ratio(text) > max_ratio
+
+
+def extract_json(text: str) -> object:
+    """Parst JSON, toleriert ```json-Codeblöcke. Wirft ValueError bei ungültigem JSON."""
+    stripped = text.strip()
+    fence = re.match(r"^```(?:json)?\s*\n(.*?)\n```\s*$", stripped, re.DOTALL)
+    if fence:
+        stripped = fence.group(1)
+    try:
+        return json.loads(stripped)
+    except json.JSONDecodeError as exc:
+        raise ValueError(str(exc)) from exc