Ashita Orbis
← All Projects

Cache Warmer

production AI Tooling

Fork-based prompt-cache warmer for Claude Code. Keeps long sessions warm in the Anthropic prompt cache without modifying them: idle sessions nearing cache expiry are warmed through disposable interactive forks, with per-warm cache_read receipts proving each refresh worked.

  • Zero session impact — live transcripts verified byte-identical before/after warming
  • Empirical TTL measurement from session history (measure-ttl.py, zero API spend)
  • Self-verifying: every warm logs the measured cache_read token count
  • Disproved two naive approaches empirically (print-mode pings, local heartbeat files)
  • ~20x cheaper resume on large sessions vs paying a cold cache re-write
BashPythonsystemdtmuxClaude Code

Activity Timeline

  • RC bridge deregistration bug diagnosed and worked around; full reverse proxy fix specified.

    Proxy's 404 on RC availability checks caused silent session deregistration from the bridge. Workaround: unset ANTHROPIC_BASE_URL before RC session registration. Full fix (true reverse proxy + Account B opt-in warming) specified, awaiting test pass with throwaway session.

    bugfixarchitecture
  • v0.4.1 released: fixed transient-failure chain break.

    Reliability patch applied after review hardening. Transient failures were causing chain breaks that interrupted cache warming sequences.

    bugfixversion-update
  • v0.4.0: v3 replay architecture replaces deprecated fork-warming.

    Fork-warming deprecated in Claude Code ≥2.1.198. v3 replay is now the primary mechanism. Clean architectural succession.

    version-updatearchitecture
  • v0.3.3 released; adversarial hardening complete; GitHub release queued for Codex Pro review.

    Fork startup prompt handling added for large MCP sessions; busy and remote-control session behavior improved. End-to-end testing passed with transparent cache refresh and no session state side effects.

    version-updatebugfixmilestone
  • Built, verified, and open-sourced the fork-based cache warmer in a single day.

    Investigated a pending cache-warming idea, measured the effective prompt-cache TTL (~1h cliff) from session history, discovered that print-mode pings cannot refresh an interactive session's cache, and landed on disposable interactive forks as the zero-impact mechanism. Published to GitHub with a TTL measurement tool and systemd timer install.

    shippedopen-source