Enhance Planning Exercise Suggestion with LLM-Rerank and Client Overrides

- Implemented optional LLM-Rerank functionality in the planning exercise suggestion process, allowing for improved exercise ranking based on user-defined criteria.
- Updated the `suggestPlanningExercises` API to accept `planned_exercise_ids` for client-side overrides, enhancing flexibility in exercise selection.
- Enhanced the `ExercisePickerModal` to reflect LLM ranking status and support new planning context features.
- Incremented application version to 0.8.170 and updated changelog to document the new features and improvements in the planning AI capabilities.

.claude/docs/working/PLANNING_EXERCISE_SUGGEST_CONTEXT.md

@@ -178,8 +178,8 @@ Wenn `hits` leer oder Trainer wählt „Mit KI anlegen“:
 |-------|--------|
 | **P0** ✅ | Context-Pack, Hybrid-Score, API, Picker in Planung |
 | **P0.1** ✅ | `ExerciseMatchProfile` / `PlanningTargetProfile`, `profile_v1`, `target_profile_summary` |
+| **P2** ✅ (optional) | LLM-Rerank `planning_exercise_search_rank`, `include_llm_rank`, `llm_rank_applied` |
 | **P1** | LLM Intent-JSON; Neu-Anlage mit Pack |
-| **P2** | LLM-Rerank + Kurzbegründung |
 | **P3** | Skill-Discovery / Framework-Ziele im Pack |
 
 ---
@@ -189,14 +189,38 @@ Wenn `hits` leer oder Trainer wählt „Mit KI anlegen“:
 - **2026-05-22:** Erstfassung; P0 API + Planungs-Picker.
 - **2026-05-22:** P0 implementiert (`planning_exercise_suggest.py`, Router, Picker); unsaved Formular-Plan noch nicht an API (nur persistierte Einheit).
 - **2026-05-22:** P0.1 — `planning_exercise_profiles.py`, Profil-Score in Hybrid-Retrieval, `retrieval_phase: profile_v1`, `target_profile_summary`.
+- **2026-05-22:** P2 — LLM-Rerank optional (`include_llm_rank`); Client `planned_exercise_ids[]`; Prompt Migration 072.
 
 ---
 
 ## 11. Bekannte P0-Lücken
 
-- **Ungespeicherte Plan-Änderungen:** API liest DB-Stand der Einheit — offene Formular-Items folgen in P0.1 (Client übergibt `planned_exercise_ids[]`).
+- **Ungespeicherte Plan-Änderungen:** ✅ Client übergibt `planned_exercise_ids[]` aus Formular (TrainingUnitEditPage).
 - **Progressionsgraph-ID:** noch nicht aus UI wählbar (`progression_graph_id` nur per API).
-- **LLM-Intent / Rerank:** P1/P2 laut Roadmap §9.
+- **LLM-Intent:** P1 laut Roadmap §9.
+
+---
+
+## 15. LLM-Rerank (P2)
+
+**Request:**
+
+| Feld | Typ | Default | Bedeutung |
+|------|-----|---------|-----------|
+| `planned_exercise_ids` | `int[]` | — | Optional: Reihenfolge aus Formular (überschreibt DB-Plan) |
+| `include_llm_rank` | `bool` | `false` | Top-32 Hybrid-Kandidaten → OpenRouter Prompt `planning_exercise_search_rank` |
+
+**Response:**
+
+| Feld | Wert |
+|------|------|
+| `retrieval_phase` | `profile_v1` oder `profile_v1+llm_rank` |
+| `llm_rank_applied` | `true` wenn LLM erfolgreich sortiert hat |
+| `hits[].llm_rank` | optional: Position nach LLM (1…n) |
+
+**Fallback:** Kein API-Key, inaktiver Prompt oder Parse-Fehler → Hybrid-Reihenfolge unverändert, `llm_rank_applied: false`.
+
+**Prompt:** Migration **072**, Slug `planning_exercise_search_rank` — Kandidaten als JSON mit Titel, summary, goal (Plaintext), skills; Ausgabe `{ ranked_ids, reasons }`.
 
 ---
 
@@ -277,4 +301,4 @@ Im Hybrid-Score kommt **`w_profile * profile_score`** hinzu (Intent-abhängig ~0
 | `retrieval_phase` | `"profile_v1"` — Phase-1 aktiv, kein LLM-Rerank |
 | `target_profile_summary` | Lesbare Kurzinfo für UI-Chips (Fokus, Top-Skills, Quellen) |
 
-**Phase 2 (P2):** Top 20–40 Kandidaten nach Hybrid+Profil → LLM `planning_exercise_search_rank` mit **Titel + summary + goal**; nur IDs aus Kandidatenliste.
+**Phase 2 (P2):** siehe §15 — optional per `include_llm_rank`.

backend/ai_prompt_runtime.py

@@ -11,6 +11,12 @@ from typing import Any, Dict, Mapping, Optional, Tuple
 
 from prompt_resolver import MustacheRenderResult, render_mustache_template
 
+_PLANNING_AI_SLUGS = frozenset(
+    {
+        "planning_exercise_search_rank",
+    }
+)
+
 _EXERCISE_AI_SLUGS = frozenset(
     {
         "exercise_summary",
@@ -26,12 +32,15 @@ class AiPromptContextKind(str, Enum):
     ohne bestehende Slugs zu invalidieren.
     """
 
+    PLANNING_EXERCISE_SEARCH = "planning_exercise_search"
     EXERCISE_FORM_AI = "exercise_form_ai"
 
 
 def context_kind_for_slug(slug: str) -> Optional[AiPromptContextKind]:
     """Ordnet einen DB-Slug einer Kontext-Art zu, sofern registriert."""
     s = (slug or "").strip().lower()
+    if s in _PLANNING_AI_SLUGS:
+        return AiPromptContextKind.PLANNING_EXERCISE_SEARCH
     if s in _EXERCISE_AI_SLUGS:
         return AiPromptContextKind.EXERCISE_FORM_AI
     return None

backend/migrations/072_ai_prompt_planning_exercise_search_rank.sql

@@ -0,0 +1,54 @@
+-- Migration 072: KI-Prompt Planungs-Übungssuche — LLM-Rerank (Phase 2)
+-- Spec: .claude/docs/working/PLANNING_EXERCISE_SUGGEST_CONTEXT.md §14
+
+INSERT INTO ai_prompts (
+    slug, display_name, description, template,
+    category, output_format, output_schema, is_system_default, default_template, active, sort_order
+)
+SELECT
+    'planning_exercise_search_rank',
+    'Planungs-Übungssuche Rerank',
+    'Ordnet Kandidaten für die Trainingsplanung nach Intent und Kontext; nur IDs aus candidates_json.',
+    $t$Du bist Assistent für Kampfsport-Trainer bei der Trainingsplanung.
+Ordne die vorgegebenen Übungs-Kandidaten nach Eignung für die aktuelle Planungssituation.
+
+Regeln:
+- Verwende NUR exercise_id-Werte aus candidates_json (keine erfundenen IDs).
+- Berücksichtige search_query, intent, planning_context_json und target_profile_json.
+- Bewerte anhand von Titel, summary, goal und skills jedes Kandidaten.
+- Gib maximal {{result_limit}} IDs in sinnvoller Reihenfolge zurück (beste zuerst).
+- Kurze Begründung pro Top-Treffer auf Deutsch (1 Satz, sachlich).
+
+Intent-Hinweise:
+- suggest_next / progression_next: logische Fortsetzung, Progression, passende Skills
+- deepen_exercise: Vertiefung zum Anker, ähnlicher Fokus
+- continue_plan_goal: schließt an bisherigen Plan und Skill-Lücken an
+- free_search: Freitext-Relevanz
+
+Kontext:
+Intent: {{intent}}
+Suchanfrage: {{search_query}}
+Planung: {{planning_context_json}}
+Zielprofil: {{target_profile_json}}
+
+Kandidaten (JSON):
+{{candidates_json}}
+
+Antworte NUR mit JSON (kein Text davor/danach):
+{
+  "ranked_ids": [123, 456],
+  "reasons": { "123": "…", "456": "…" }
+}$t$,
+    'training',
+    'json',
+    '{"type":"object","required":["ranked_ids"],"properties":{"ranked_ids":{"type":"array","items":{"type":"integer"}},"reasons":{"type":"object"}}}'::jsonb,
+    true,
+    NULL,
+    true,
+    10
+WHERE NOT EXISTS (SELECT 1 FROM ai_prompts WHERE slug = 'planning_exercise_search_rank');
+
+UPDATE ai_prompts
+SET default_template = template
+WHERE slug = 'planning_exercise_search_rank'
+  AND (default_template IS NULL OR TRIM(default_template) = '');

backend/planning_exercise_llm_rank.py

@@ -0,0 +1,223 @@
+"""
+Phase 2 Planungs-Übungssuche: LLM-Rerank über Hybrid-Kandidaten.
+
+Prompt-Slug: planning_exercise_search_rank (Migration 072)
+"""
+from __future__ import annotations
+
+import json
+import logging
+import re
+from typing import Any, Dict, List, Mapping, Optional, Sequence, Set, Tuple
+
+from ai_prompt_runtime import AiPromptUnavailableError, load_and_render_ai_prompt
+from exercise_ai import strip_html_to_plain
+from openrouter_chat import (
+    effective_openrouter_model_for_prompt_row,
+    normalize_openrouter_env,
+    openrouter_chat_completion,
+)
+
+_logger = logging.getLogger("shinkan.planning_exercise_llm_rank")
+
+_LLM_RERANK_POOL = 32
+_MAX_GOAL_PLAIN = 480
+_MAX_SUMMARY_PLAIN = 320
+_MAX_REASON_LEN = 160
+
+
+def _compact_json(obj: Any) -> str:
+    return json.dumps(obj, ensure_ascii=False, separators=(",", ":"))
+
+
+def _extract_json_object(text: str) -> Dict[str, Any]:
+    s = (text or "").strip()
+    if s.startswith("```"):
+        s = re.sub(r"^```[a-zA-Z0-9]*\s*", "", s)
+        if s.endswith("```"):
+            s = s[:-3].strip()
+    start = s.find("{")
+    end = s.rfind("}")
+    if start < 0 or end <= start:
+        raise ValueError("Kein JSON-Objekt in LLM-Antwort")
+    obj = json.loads(s[start : end + 1])
+    if not isinstance(obj, dict):
+        raise ValueError("LLM-Antwort ist kein JSON-Objekt")
+    return obj
+
+
+def parse_planning_exercise_rank_response(
+    text: str,
+    allowed_ids: Set[int],
+) -> Tuple[List[int], Dict[int, str]]:
+    """
+    Validiert LLM-Ranking: nur erlaubte exercise_id, dedupliziert, Reihenfolge beibehalten.
+    """
+    obj = _extract_json_object(text)
+    ranked_raw = obj.get("ranked_ids") or obj.get("ranked") or obj.get("ids")
+    if not isinstance(ranked_raw, list):
+        raise ValueError("ranked_ids fehlt oder ist keine Liste")
+
+    ranked: List[int] = []
+    seen: Set[int] = set()
+    for raw in ranked_raw:
+        try:
+            eid = int(raw)
+        except (TypeError, ValueError):
+            continue
+        if eid < 1 or eid not in allowed_ids or eid in seen:
+            continue
+        seen.add(eid)
+        ranked.append(eid)
+
+    reasons_out: Dict[int, str] = {}
+    reasons_raw = obj.get("reasons") or obj.get("reasons_by_id") or {}
+    if isinstance(reasons_raw, dict):
+        for k, v in reasons_raw.items():
+            try:
+                eid = int(k)
+            except (TypeError, ValueError):
+                continue
+            if eid not in allowed_ids:
+                continue
+            txt = str(v or "").strip()
+            if txt:
+                reasons_out[eid] = txt[:_MAX_REASON_LEN]
+
+    return ranked, reasons_out
+
+
+def _build_candidate_payload(
+    hit: Mapping[str, Any],
+    *,
+    goal_plain: str,
+    skill_names: Sequence[str],
+) -> Dict[str, Any]:
+    return {
+        "id": int(hit["id"]),
+        "title": str(hit.get("title") or "").strip()[:200],
+        "summary": strip_html_to_plain(hit.get("summary"), max_len=_MAX_SUMMARY_PLAIN),
+        "goal": goal_plain,
+        "skills": list(skill_names)[:8],
+        "retrieval_score": float(hit.get("score") or 0.0),
+    }
+
+
+def _load_exercise_goals(cur, exercise_ids: Sequence[int]) -> Dict[int, str]:
+    ids = [int(x) for x in exercise_ids if int(x) > 0]
+    if not ids:
+        return {}
+    ph = ",".join(["%s"] * len(ids))
+    cur.execute(
+        f"SELECT id, goal FROM exercises WHERE id IN ({ph})",
+        ids,
+    )
+    return {int(r["id"]): str(r.get("goal") or "") for r in cur.fetchall()}
+
+
+def _load_skill_names(cur, skill_ids: Sequence[int]) -> Dict[int, str]:
+    ids = sorted({int(x) for x in skill_ids if int(x) > 0})
+    if not ids:
+        return {}
+    ph = ",".join(["%s"] * len(ids))
+    cur.execute(f"SELECT id, name FROM skills WHERE id IN ({ph})", ids)
+    return {int(r["id"]): str(r.get("name") or "") for r in cur.fetchall()}
+
+
+def try_llm_rerank_planning_hits(
+    cur,
+    *,
+    hits: List[Dict[str, Any]],
+    skills_by_ex: Mapping[int, Set[int]],
+    query: str,
+    intent: str,
+    context_summary: Mapping[str, Any],
+    target_profile_summary: Mapping[str, Any],
+    limit: int,
+) -> Tuple[List[Dict[str, Any]], bool]:
+    """
+    Optionaler LLM-Rerank der Top-Kandidaten. Bei Fehler: Original-Reihenfolge, llm_applied=False.
+    """
+    if not hits:
+        return hits, False
+
+    api_key, _ = normalize_openrouter_env()
+    if not api_key:
+        return hits, False
+
+    pool = hits[:_LLM_RERANK_POOL]
+    allowed_ids = {int(h["id"]) for h in pool}
+    goals = _load_exercise_goals(cur, list(allowed_ids))
+
+    all_skill_ids: Set[int] = set()
+    for eid in allowed_ids:
+        all_skill_ids.update(skills_by_ex.get(eid) or set())
+    skill_name_map = _load_skill_names(cur, list(all_skill_ids))
+
+    candidates: List[Dict[str, Any]] = []
+    for hit in pool:
+        eid = int(hit["id"])
+        sk_ids = sorted(skills_by_ex.get(eid) or set())
+        sk_names = [skill_name_map.get(sid, f"#{sid}") for sid in sk_ids[:8]]
+        goal_plain = strip_html_to_plain(goals.get(eid), max_len=_MAX_GOAL_PLAIN)
+        candidates.append(
+            _build_candidate_payload(hit, goal_plain=goal_plain, skill_names=sk_names)
+        )
+
+    variables = {
+        "search_query": query or "",
+        "intent": intent or "",
+        "planning_context_json": _compact_json(dict(context_summary or {})),
+        "target_profile_json": _compact_json(dict(target_profile_summary or {})),
+        "candidates_json": _compact_json(candidates),
+        "result_limit": str(max(1, min(int(limit), 50))),
+    }
+
+    try:
+        prow, rendered = load_and_render_ai_prompt(cur, "planning_exercise_search_rank", variables)
+        model = effective_openrouter_model_for_prompt_row(prow)
+        raw = openrouter_chat_completion(
+            api_key=api_key,
+            model=model,
+            user_content=rendered.text,
+        )
+        ranked_ids, llm_reasons = parse_planning_exercise_rank_response(raw, allowed_ids)
+    except AiPromptUnavailableError:
+        return hits, False
+    except Exception as exc:
+        _logger.warning("Planungs-LLM-Rerank fehlgeschlagen: %s", exc)
+        return hits, False
+
+    if not ranked_ids:
+        return hits, False
+
+    hit_by_id = {int(h["id"]): h for h in hits}
+    reranked: List[Dict[str, Any]] = []
+    used: Set[int] = set()
+    for eid in ranked_ids:
+        hit = hit_by_id.get(eid)
+        if not hit:
+            continue
+        used.add(eid)
+        new_hit = dict(hit)
+        reasons = list(hit.get("reasons") or [])
+        llm_reason = llm_reasons.get(eid)
+        if llm_reason and llm_reason not in reasons:
+            reasons.insert(0, llm_reason)
+        new_hit["reasons"] = reasons
+        new_hit["llm_rank"] = len(reranked) + 1
+        reranked.append(new_hit)
+
+    for hit in hits:
+        eid = int(hit["id"])
+        if eid in used:
+            continue
+        reranked.append(dict(hit))
+
+    return reranked[: max(int(limit), len(reranked))], True
+
+
+__all__ = [
+    "parse_planning_exercise_rank_response",
+    "try_llm_rerank_planning_hits",
+]

backend/planning_exercise_suggest.py

@@ -17,6 +17,7 @@ from planning_exercise_profiles import (
     load_exercise_match_profiles_bulk,
     score_exercise_against_target,
 )
+from planning_exercise_llm_rank import try_llm_rerank_planning_hits
 
 # Planungs-Berechtigung + Sektionen (bestehende Implementierung)
 from routers.training_planning import (
@@ -40,6 +41,7 @@ VALID_INTENTS = {
 }
 
 _CANDIDATE_POOL_LIMIT = 400
+_LLM_RERANK_PRE_LIMIT = 32
 
 
 class PlanningExerciseSuggestRequest(BaseModel):
@@ -51,6 +53,8 @@ class PlanningExerciseSuggestRequest(BaseModel):
     progression_graph_id: Optional[int] = Field(default=None, ge=1)
     query: Optional[str] = ""
     intent_hint: Optional[str] = None
+    planned_exercise_ids: Optional[List[int]] = None
+    include_llm_rank: bool = False
     limit: int = Field(default=20, ge=1, le=50)
     exercise_kind_any: Optional[List[str]] = None
 
@@ -240,6 +244,42 @@ def _skill_jaccard(a: Set[int], b: Set[int]) -> float:
     return inter / union if union else 0.0
 
 
+def _apply_client_planned_override(
+    cur,
+    pack: Dict[str, Any],
+    body: PlanningExerciseSuggestRequest,
+) -> Dict[str, Any]:
+    """Client-Plan (ungespeichertes Formular) überschreibt DB-Stand."""
+    if not body.planned_exercise_ids:
+        return pack
+    planned_ids: List[int] = []
+    seen: Set[int] = set()
+    for raw in body.planned_exercise_ids:
+        try:
+            eid = int(raw)
+        except (TypeError, ValueError):
+            continue
+        if eid < 1 or eid in seen:
+            continue
+        seen.add(eid)
+        planned_ids.append(eid)
+    if not planned_ids:
+        return pack
+
+    pack["planned_exercise_ids"] = planned_ids
+    if not body.anchor_exercise_id:
+        anchor_id = _resolve_anchor_from_plan(planned_ids, None)
+        pack["anchor_exercise_id"] = anchor_id
+        if anchor_id:
+            titles = _load_exercise_titles(cur, [anchor_id])
+            pack["anchor_title"] = titles.get(anchor_id)
+            pack["anchor_skill_ids"] = sorted(_load_skill_ids_for_exercise(cur, anchor_id))
+        else:
+            pack["anchor_title"] = None
+            pack["anchor_skill_ids"] = []
+    return pack
+
+
 def build_planning_exercise_context_pack(
     cur,
     *,
@@ -327,6 +367,7 @@ def suggest_planning_exercises(
     body: PlanningExerciseSuggestRequest,
 ) -> Dict[str, Any]:
     pack = build_planning_exercise_context_pack(cur, tenant=tenant, body=body)
+    pack = _apply_client_planned_override(cur, pack, body)
     query = _normalize_query(body.query)
     intent = resolve_planning_exercise_intent(query, body.intent_hint)
     weights = _intent_weights(intent)
@@ -497,6 +538,38 @@ def suggest_planning_exercises(
         )
 
     hits.sort(key=lambda h: (-h["score"], h.get("title") or ""))
+
+    llm_applied = False
+    retrieval_phase = "profile_v1"
+    if body.include_llm_rank:
+        pre_limit = max(int(body.limit), _LLM_RERANK_PRE_LIMIT)
+        pool_hits = hits[:pre_limit]
+        pool_hits, llm_applied = try_llm_rerank_planning_hits(
+            cur,
+            hits=pool_hits,
+            skills_by_ex=skills_by_ex,
+            query=query,
+            intent=intent,
+            context_summary={
+                "unit_title": pack.get("unit_title"),
+                "group_name": pack.get("group_name"),
+                "section_title": pack.get("section_title"),
+                "planned_count": len(planned_set),
+                "anchor_title": pack.get("anchor_title"),
+                "intent": intent,
+            },
+            target_profile_summary=target_profile_summary,
+            limit=int(body.limit),
+        )
+        if llm_applied:
+            retrieval_phase = "profile_v1+llm_rank"
+            tail = hits[pre_limit:]
+            hits = pool_hits + tail
+        else:
+            hits = pool_hits[: int(body.limit)]
+    else:
+        hits = hits[: int(body.limit)]
+
     hits = hits[: int(body.limit)]
 
     context_summary = {
@@ -512,7 +585,8 @@ def suggest_planning_exercises(
     return {
         "context_summary": context_summary,
         "target_profile_summary": target_profile_summary,
-        "retrieval_phase": "profile_v1",
+        "retrieval_phase": retrieval_phase,
+        "llm_rank_applied": llm_applied,
         "intent_resolved": intent,
         "query_normalized": query or None,
         "hits": hits,

backend/routers/planning_exercise_suggest.py

@@ -1,5 +1,5 @@
 """
-POST /api/planning/exercise-suggest — planungsgebundene Übungssuche (P0 Hybrid-Retrieval).
+POST /api/planning/exercise-suggest — planungsgebundene Übungssuche (Hybrid + Profil + optional LLM-Rerank).
 """
 from fastapi import APIRouter, Depends
 

backend/tests/test_planning_exercise_suggest.py

@@ -0,0 +1,34 @@
+"""Tests für Planungs-Übungssuche (Intent, LLM-Rerank-Parser)."""
+from planning_exercise_suggest import resolve_planning_exercise_intent
+from planning_exercise_llm_rank import parse_planning_exercise_rank_response
+
+
+def test_resolve_planning_exercise_intent_defaults():
+    assert resolve_planning_exercise_intent("", None) == "suggest_next"
+    assert resolve_planning_exercise_intent("  ", "suggest_next") == "suggest_next"
+
+
+def test_resolve_planning_exercise_intent_keywords():
+    assert resolve_planning_exercise_intent("Vertiefung Partner", None) == "deepen_exercise"
+    assert resolve_planning_exercise_intent("nächste übung", None) == "suggest_next"
+    assert resolve_planning_exercise_intent("progression graph", None) == "progression_next"
+
+
+def test_parse_planning_exercise_rank_response_filters_ids():
+    allowed = {10, 20, 30}
+    ranked, reasons = parse_planning_exercise_rank_response(
+        '{"ranked_ids":[20,999,20,10],"reasons":{"20":"Passt gut","999":"ignore"}}',
+        allowed,
+    )
+    assert ranked == [20, 10]
+    assert reasons[20] == "Passt gut"
+    assert 999 not in reasons
+
+
+def test_parse_planning_exercise_rank_response_reasons_by_id_alias():
+    ranked, reasons = parse_planning_exercise_rank_response(
+        '{"ranked_ids":[5],"reasons_by_id":{"5":"Skill-Lücke"}}',
+        {5},
+    )
+    assert ranked == [5]
+    assert reasons[5] == "Skill-Lücke"

backend/version.py

@@ -1,8 +1,8 @@
 # Shinkan Jinkendo Version Information
 
-APP_VERSION = "0.8.169"
+APP_VERSION = "0.8.170"
 BUILD_DATE = "2026-05-22"
-DB_SCHEMA_VERSION = "20260531071"
+DB_SCHEMA_VERSION = "20260531072"
 
 MODULE_VERSIONS = {
     "legal_documents": "1.4.0",  # Admin: Live-Vorschau pro Abschnitt + modale Vollvorschau (Editor + Dokumentenliste)
@@ -22,13 +22,13 @@ MODULE_VERSIONS = {
     "admin_ai_prompts": "1.0.3",  # Migration 070: openrouter_model; PUT/Liste/Detail
     "ai_prompt_job": "0.2.1",  # want_instructions; run_exercise_form_ai_suggestion
     "ai_prompt_context": "0.2.0",  # preparation/trainer_notes; has_instruction_source_text
-    "ai_prompt_runtime": "0.2.0",  # load_and_render_ai_prompt, AiPromptUnavailableError, render_ai_prompt_template_for_row
+    "ai_prompt_runtime": "0.2.1",  # Kontext-Art planning_exercise_search; load_and_render_ai_prompt
     "groups": "0.1.0",
     "skills": "0.1.1",  # DB 065 karate_relevance + relevance_level; CRUD unterstützt Felder
     "skill_profiles": "1.0.0",  # Phase 3: gewichtetes Fähigkeiten-Profil + skill-discovery/suggestions
     "methods": "0.1.0",
-    "exercises": "2.35.0",  # Planungs-KI P0.1: Profil-Score profile_v1 + target_profile_summary
+    "exercises": "2.36.0",  # Planungs-KI P2: LLM-Rerank + Client planned_exercise_ids
-    "planning_exercise_suggest": "0.2.1",  # Fix Import library_content_visibility_sql aus tenant_context
+    "planning_exercise_suggest": "0.3.0",  # include_llm_rank, planned_exercise_ids Override
     "training_units": "0.4.0",  # POST .../publish-to-framework: Ablauf aus geplanter Einheit → Rahmen-Slot-Blueprint
     "training_programs": "0.1.0",
     "planning": "0.15.0",  # Vorlagen: Strukturvorschau, Bearbeiten inkl. Split-Sessions + Beschreibung
@@ -43,6 +43,14 @@ MODULE_VERSIONS = {
 }
 
 CHANGELOG = [
+    {
+        "version": "0.8.170",
+        "date": "2026-05-22",
+        "changes": [
+            "Planungs-KI P2: optionaler LLM-Rerank (planning_exercise_search_rank) mit Titel/summary/goal; include_llm_rank.",
+            "Client planned_exercise_ids für ungespeicherten Plan; Migration 072 Prompt.",
+        ],
+    },
     {
         "version": "0.8.169",
         "date": "2026-05-22",

docs/HANDOVER.md

@@ -89,10 +89,10 @@ Das Schema ist gegenüber dem Code zurück: Migration **`022_skills_schema_compl
 - **Varianten:** Speichern in der **Aktionsleiste** persistiert zuerst geänderte Varianten (`persistPendingVariantChanges`), dann Übungs-Stammdaten; „Variante anlegen“ als `type="button"` ohne verschachteltes Formular (`createVariantFromDraft`)
 - **Governance (Übungen):** Owner = `created_by`; Bearbeiten = Ersteller, Plattform-Admin oder `can_plan_in_club` bei `visibility=club`; Löschen `club` = nur `club_admin`; Details **`FEATURES_DELIVERED_2026-Q2.md`** §16, **`EXERCISES_API_SPEC.md`** Permissions
 
-### 2.8 KI Assistenz Übungen & Skill-Katalog-Retrieval (Stand **0.8.168**)
+### 2.8 KI Assistenz Übungen & Skill-Katalog-Retrieval (Stand **0.8.170**)
 
 - **Zielarchitektur (Pflicht fuer Erweiterungen):** `.claude/docs/technical/AI_PROMPT_TARGET_ARCHITECTURE.md` — Kontext-Arten, Composition, Einbindung Planung/Rahmen; Phasenplan P0–P4.
-- **Planungs-Übungssuche (P0.1):** `.claude/docs/working/PLANNING_EXERCISE_SUGGEST_CONTEXT.md` — Context-Pack, Hybrid-Retrieval + **Profil-Score** (`profile_v1`, `ExerciseMatchProfile` / `PlanningTargetProfile`); **`POST /api/planning/exercise-suggest`**; Frontend **`ExercisePickerModal`** + **`planningContext`** aus **`TrainingUnitEditPage`**.
+- **Planungs-Übungssuche (P2):** `.claude/docs/working/PLANNING_EXERCISE_SUGGEST_CONTEXT.md` — Hybrid + Profil-Score + optional **LLM-Rerank** (`include_llm_rank`, Prompt `planning_exercise_search_rank`); Client **`planned_exercise_ids`**; **`POST /api/planning/exercise-suggest`**; **`ExercisePickerModal`** + **`planningContext`** aus **`TrainingUnitEditPage`**.
 - **Doku:** Umsetzung `.claude/docs/working/AI_EXERCISE_IMPLEMENTATION_PLAN.md`; Profil-/JSON-Konzept `.claude/docs/working/AI_SKILL_RETRIEVAL_PROFILES_SPEC.md`; Ist-Prompt/UI **`AI_PROMPT_SYSTEM_SPEC.md`**; API-Felder **`KI_FEATURES_SPEC.md`** §5.2  
 - **Kontext / Job:** **`ai_prompt_context`** (Titel, Ziel, Durchführung, Vorbereitung, Trainer-Hinweise, Fokus); **`ai_prompt_job`** — **`run_exercise_form_ai_suggestion`**; **`ai_prompt_runtime`**; **`exercise_ai`** — OpenRouter  
 - **DB:** **`067`** ai_prompts · **`069`** default_template · **`068`** ai_skill_retrieval_profiles · **`070`** openrouter_model · **`071`** **`exercise_instruction_rewrite`**  

frontend/src/api/planning.js

@@ -76,7 +76,7 @@ export async function quickCreateTrainingUnit(data) {
   })
 }
 
-/** Planungs-KI P0: kontextgebundene Übungssuche (Hybrid-Retrieval). */
+/** Planungs-KI: kontextgebundene Übungssuche (Hybrid + Profil + optional LLM-Rerank). */
 export async function suggestPlanningExercises(body = {}) {
   return request('/api/planning/exercise-suggest', {
     method: 'POST',

frontend/src/components/ExercisePickerModal.jsx

@@ -68,6 +68,7 @@ export default function ExercisePickerModal({
   const [quickCreateDraft, setQuickCreateDraft] = useState(null)
   const [planningContextSummary, setPlanningContextSummary] = useState(null)
   const [planningTargetProfileSummary, setPlanningTargetProfileSummary] = useState(null)
+  const [planningLlmRankApplied, setPlanningLlmRankApplied] = useState(false)
   const [planningIntentResolved, setPlanningIntentResolved] = useState(null)
   const pickerScrollRef = useRef(null)
 
@@ -155,6 +156,7 @@ export default function ExercisePickerModal({
       setQuickCreateDraft(null)
       setPlanningContextSummary(null)
       setPlanningTargetProfileSummary(null)
+      setPlanningLlmRankApplied(false)
       setPlanningIntentResolved(null)
       return
     }
@@ -271,6 +273,11 @@ export default function ExercisePickerModal({
             planningContext.anchorExerciseId != null ? Number(planningContext.anchorExerciseId) : null,
           progression_graph_id:
             planningContext.progressionGraphId != null ? Number(planningContext.progressionGraphId) : null,
+          planned_exercise_ids:
+            Array.isArray(planningContext.plannedExerciseIds) && planningContext.plannedExerciseIds.length > 0
+              ? planningContext.plannedExerciseIds.map((x) => Number(x)).filter((x) => Number.isFinite(x) && x > 0)
+              : undefined,
+          include_llm_rank: true,
           query,
           intent_hint: planningContext.intentHint || null,
           limit: PAGE_SIZE,
@@ -279,6 +286,7 @@ export default function ExercisePickerModal({
         })
         setPlanningContextSummary(res?.context_summary || null)
         setPlanningTargetProfileSummary(res?.target_profile_summary || null)
+        setPlanningLlmRankApplied(Boolean(res?.llm_rank_applied))
         setPlanningIntentResolved(res?.intent_resolved || null)
         const hits = (Array.isArray(res?.hits) ? res.hits : []).map((h) => ({
           id: h.id,
@@ -294,6 +302,7 @@ export default function ExercisePickerModal({
       } else {
         setPlanningContextSummary(null)
         setPlanningTargetProfileSummary(null)
+        setPlanningLlmRankApplied(false)
         setPlanningIntentResolved(null)
         const batch = await api.listExercises({
           ...queryBase,
@@ -312,6 +321,7 @@ export default function ExercisePickerModal({
       setHasMore(false)
       setPlanningContextSummary(null)
       setPlanningTargetProfileSummary(null)
+      setPlanningLlmRankApplied(false)
       setPlanningIntentResolved(null)
     } finally {
       setLoading(false)
@@ -538,6 +548,7 @@ export default function ExercisePickerModal({
               {planningIntentResolved ? (
                 <p style={{ margin: '6px 0 0', fontSize: '11px', color: 'var(--text3)' }}>
                   Modus: {planningIntentResolved.replace(/_/g, ' ')}
+                  {planningLlmRankApplied ? ' · KI-Ranking aktiv' : null}
                 </p>
               ) : null}
             </div>

frontend/src/pages/TrainingUnitEditPage.jsx

@@ -152,11 +152,23 @@ export default function TrainingUnitEditPage() {
         }
       }
     }
+    const plannedExerciseIds = []
+    const seenPlan = new Set()
+    for (const sec of secs) {
+      for (const it of sec?.items || []) {
+        if (String(it?.item_type || '').toLowerCase() === 'note') continue
+        const eid = Number(it?.exercise_id)
+        if (!Number.isFinite(eid) || eid < 1 || seenPlan.has(eid)) continue
+        seenPlan.add(eid)
+        plannedExerciseIds.push(eid)
+      }
+    }
     return {
       unitId: Number(editingUnit.id),
       sectionOrderIndex: sIdx,
       anchorExerciseId: Number.isFinite(anchorExerciseId) && anchorExerciseId > 0 ? anchorExerciseId : null,
       progressionGraphId: null,
+      plannedExerciseIds,
     }
   }, [editingUnit?.id, exercisePickerTarget, formData.sections])