1. Worktrees as billing amplifier: 92x cache-create on one account

    Claude Code keys its cacheable prefix on the literal current directory, not on the repository. A 30-day reconstruction across 25 cwd values in one repo logged a cache-creation-to-input-token ratio of 92x against an expected ratio under 2x. Daily samples ran 366x to 1,403x.

  2. Cold-start cache miss: 1 in 3 on 2.1.104

    Developer analysis finds Claude Code places CLAUDE.md, skills, and tool schemas in the messages[0] user block where jitter invalidates the cache. Fresh sessions pay full base rate on turn 1 roughly 1-in-3 on 2.1.104.

  3. v2.1.108: two fixes, five unaddressed

    Telemetry-gated TTL downgrade and cache_creation inflation both fixed, two new warnings, five structural cache issues unchanged.

  4. Gemma 4 on iPhone: $0/query, Pay in Heat

    $0/query, no API call, no billing dashboard. Google AI Edge Gallery runs Gemma 4 locally on iPhone 15 Pro or better — you pay in battery and a warm pocket.