2 Commits

Author	SHA1	Message	Date
Adriano	ba4eb09a71	feat(phase-2.5): Task 6 cost_kind attribution + fees_eat_alpha threshold CLI ... Task 6 del piano Phase 2.5 (deferito → ora completato): - CostRecord: nuovo campo call_kind (default "hypothesis") - CostTracker.record: accetta call_kind opzionale, summary include by_call_kind breakdown (hypothesis vs mutation) - Schema cost_records: aggiunta colonna call_kind TEXT NOT NULL DEFAULT 'hypothesis' + migration soft via ALTER TABLE in init_schema (silently catched per DB pre-Task 6) - Repository.save_cost_record: nuova arg call_kind opzionale - mutate_prompt_llm: accetta cost_tracker/repo/run_id opzionali e logga la call mutator con call_kind="mutation" quando sink presente - weighted_random_mutate, next_generation: propagano cost sink - orchestrator.run_phase1: passa cost_tracker+repo+run_id a next_generation solo se prompt_mutation_weight > 0 Esposto fees_eat_alpha_threshold come parametro AdversarialAgent (default 0.5 = comportamento Phase 1.5 invariato), propagato via RunConfig.fees_eat_alpha_threshold e flag CLI --fees-eat-alpha-threshold. Abilita ablation con soglia 0.7-0.8 senza modificare codice — adversarial finding dominante in tutti i run Phase 2/2.5 (50+ HIGH per run). Tests (+4 → 186 totale): - test_cost_tracker: default call_kind="hypothesis"; breakdown by_call_kind con hypothesis+mutation - test_mutation_prompt_llm: logging mutation cost con sink completo; backward compat senza sink (no errore) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-12 10:42:13 +02:00
Adriano	c38311e5fa	feat(phase-2.5): mutate_prompt_llm operator + weighted dispatcher + GA wiring ... Implementazione completa Phase 2.5 (LLM prompt mutator) seguendo il piano in docs/superpowers/plans/2026-05-11-mutate-prompt-llm-phase-2-5.md. Nuovo modulo src/multi_swarm/genome/mutation_prompt_llm.py: - 6 mutation instructions (tighten_threshold, swap_comparator, add_condition, remove_condition, change_timeframe, add_temporal_gate) - mutate_prompt_llm(g, llm, rng, mutator_tier=ModelTier.B): clona genome con tier B per la call mutator, costruisce system+user prompt con istruzione scelta random, estrae prompt da tag <prompt>...</prompt>, valida - is_valid_prompt(): 3 check (lunghezza >= 50, keyword tecnica, diff > 5% Levenshtein-like via difflib.SequenceMatcher) - Fallback random_mutate su qualsiasi validation fail O LLM exception Esteso src/multi_swarm/genome/mutation.py con weighted_random_mutate dispatcher: con probabilità prompt_mutation_weight invoca mutate_prompt_llm, altrimenti random_mutate. Backward compat: llm=None oppure weight=0 → solo scalare. Integrazione GA loop + RunConfig: - GAConfig.prompt_mutation_weight: float = 0.0 (default off) - next_generation(llm=...) propagato all'invocazione mutator - RunConfig.prompt_mutation_weight con stesso default - run_phase1: passa llm a next_generation solo se weight > 0 - scripts/run_phase1.py: flag CLI --prompt-mutation-weight Tests (+18, 175 totale): - tests/unit/test_mutation_prompt_llm.py (12): extract_prompt, is_valid_prompt 3 check, operator success + fallback su 3 modi (invalid/identical/exception), tier B per LLM call, istruzione scelta dal pool - tests/unit/test_mutation_dispatcher.py (4): weight 0/1/None + distribuzione 30/70 su 1000 estrazioni con tolleranza ±5% - tests/integration/test_ga_loop_with_prompt_mutator.py (2): loop con weight=1.0 produce prompt evoluti; backward compat weight=0 invariato Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>	2026-05-11 23:49:46 +02:00