Release LocalAgent Benchmark v0.8.0 · gloutchov/LocalAgentBenchmark

Italiano

Nuovo caso thinking_challenge: pianificazione esatta multi-vincolo con budget, rischio, capacità dei team, dipendenze, conflitti, categorie obbligatorie e tie-break deterministico.
Nuovo profilo indipendente thinking; il caso entra anche in standard e full, mentre smoke e showcase restano invariati.
Grader deterministico da 100 punti con scenari nascosti, baseline iniziale 5/100, controllo anti-hardcoding e soluzione generica calibrata a 100/100.
Dashboard ed export separano classifiche e fattori per run e modalità thinking; thinking e showcase restano fuori dal funnel smoke → standard → full.
Protocollo A/B simmetrico off/medium, senza fallback condizionali o retry selettivi.

Compileall completato; 106 test Python verdi con 12 skip ambientali motivati.
8 test JavaScript verdi e tutti i 6 casi validati.
Verifica Playwright completata su desktop e mobile, temi chiaro/scuro, tastiera e sole richieste locali.
Esperimento reale con Pi 0.85.1, Ollama 0.34.2 e qwen3.5:9b-mlx: tre ripetizioni medium e tre off, sandbox Seatbelt richiesta, integrità valida e zero retry.
medium: punteggi 60/0/0, qualità media 20,0 e completion 33%; off: 22/0/10, qualità media 10,7 e completion 0%.
CI verde su macOS, Ubuntu e Windows per PR #2 (35457236695) e main (35457439164).

Il runtime Python resta privo di dipendenze esterne e supporta macOS, Windows e Linux secondo i backend documentati.
Tutte le task del campione reale hanno terminato per limite length; tre campioni per cella mostrano un vantaggio descrittivo per medium, ma non consentono inferenza statistica.
I run in results/ sono output locali potenzialmente sensibili, ignorati da Git e non inclusi nella release.
La release distribuisce il repository sorgente. Non esistono artifact binari di progetto, quindi non si applicano checksum di artifact; GitHub genera automaticamente gli archivi sorgente.

New thinking_challenge case: exact multi-constraint planning with budgets, risk, team capacities, dependencies, conflicts, required categories, and deterministic tie-breaking.
New independent thinking profile; the case is also included in standard and full, while smoke and showcase remain unchanged.
Deterministic 100-point grader with hidden scenarios, a 5/100 initial baseline, anti-hardcoding coverage, and a generic reference solution calibrated at 100/100.
Dashboard and exports keep rankings and score factors separate by run and thinking mode; thinking and showcase stay outside the smoke → standard → full funnel.
Symmetric off/medium A/B protocol without conditional fallbacks or selective retries.

Compileall passed; 106 Python tests passed with 12 documented environment-dependent skips.
8 JavaScript tests passed and all 6 cases validated.
Playwright verification covered desktop and mobile layouts, light/dark themes, keyboard operation, and local-only requests.
Real experiment with Pi 0.85.1, Ollama 0.34.2, and qwen3.5:9b-mlx: three medium and three off repetitions, required Seatbelt sandbox, valid integrity, and zero retries.
medium: scores 60/0/0, mean quality 20.0, completion 33%; off: 22/0/10, mean quality 10.7, completion 0%.
CI passed on macOS, Ubuntu, and Windows for PR #2 (35457236695) and main (35457439164).

The Python runtime remains dependency-free and supports macOS, Windows, and Linux according to the documented backends.
Every task in the real sample stopped at the length limit; three samples per cell show a descriptive advantage for medium, but do not support statistical inference.
Runs under results/ are potentially sensitive local output, ignored by Git and excluded from the release.
This release distributes the source repository. There are no project binary artifacts, so artifact checksums do not apply; GitHub automatically provides source archives.