How Antigravity Handles Context (Without /compact)

Оригинальный автор：RUNLOG

Первоисточник：https://www.youtube.com/embed/zRxDRU_0Bco

Claude Code developers rely on /compact when context gets bloated. But how does Google Antigravity (running Gemini 3.8 Flash) handle massive 600k+ token sessions without a compaction command? In this video, we put Antigravity through the exact same 12-constraint benchmark we used on Claude Code, tracking retention from Turn 1 (27k tokens) to Turn 24 (617k tokens). We also dissect the underlying mechanics: continuous transcript payload pruning vs batch summary compression, multimodal image preservation, and the prompt cache TTL traps that silently burn tokens. TIMESTAMPS: 00:00 - The Stop-the-World Cliff 01:05 - Continuous Pruning Architecture 02:42 - Modalities, Injected Rules, and Trigger Mechanics 04:54 - The 12-Constraint Gauntlet Benchmark 06:38 - Empirical Retention Results (Turn 1 to 24) 08:07 - Cache TTL: 1h Main vs 5m Subagent Trap 09:38 - Developer Verdict and Practical Rules RESOURCES & RECEIPTS: - Benchmark plan, harness, grader, and verified Turn 24 receipts: https://github.com/runlog-tech/runlog-dev/tree/main/11-antigravity-context-mechanics - Includes an unresolved data anomaly disclosed openly, not hidden: a further run past Turn 24 produced a token count we could not explain and chose not to publish as a verified result. Full writeup in the repo. RUNLOG: Engineering receipts over prompt hype. Subscribe for falsifiable agentic engineering data.