Skip to content

v2.0.0 — 15 cache-stability fixes

Choose a tag to compare

@cnighswonger cnighswonger released this 17 Apr 16:41
· 245 commits to main since this release

Major release — 7 new cache-stability fixes from @deafsquad, expanding the interceptor from 8 to 15 fixes. Combined stack reduces first-request cache creation by up to 99.8% on affected accounts (940K → 1.7K tokens measured).

New fixes

  • smoosh_split — Universal un-smoosh: peels trailing <system-reminder> content out of tool_result.content strings (PR #26)
  • session_start_normalize — Rewrites SessionStart:resume:startup, strips volatile session-id and timestamps (PR #27)
  • continue_trailer_strip — Removes the --continue resume-nudge text block (PR #28)
  • deferred_tools_restore — Snapshots and restores MCP deferred-tools block across reconnect race (PR #29)
  • reminder_strip — Drops bookkeeping system-reminder blocks (token usage, budget, turn counters) (PR #30)
  • cache_control_normalize — Pins cache_control marker at canonical position (PR #31)
  • tool_use_input_normalize — Strips non-schema keys from tool_use.input, canonicalizes key order (PR #32)

Opus 4.7 advisory

Opus 4.7 burns Q5h at ~2.4x the rate of 4.6 due to invisible adaptive thinking tokens. Workaround: CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1. See Discussion #25.

Stats

  • 146 tests, all passing
  • Compatible with CC v2.1.112
  • 10 contributors

Full changelog: CHANGELOG.md