Follow Discord
Sweep 06 Nov 2026 · 01:32Z Build v2.1.319 537 read Stable v2.1.311 Latest v2.1.318 Next v2.1.319 Feeds RSS JSON llms.txt llms-full.txt Unofficial

Claude Code v2.1.267 ·

Prompt-cache diagnostics track a 'prefix-break hash' across sessions and forks

Claude Code now tracks a prefix-break hash to diagnose prompt-cache-busting changes, including across restored sessions and forked agents

Group of 3 Under the hood Internal Changes
JSON All of v2.1.267
Under the hoodTier: how much it should matter to you
2Useful: my rating, 1 to 5
2Signal: worth watching, 1 to 5
CachingArea: what it touches
Internal ChangesKind: in v2.1.267,
Internal ChangesSection of the release

Unclear What user-visible effect this has, if any, isn't stated in the finding.

What

  • The per-request cache-state tracker now computes and stores a prefixBreakHash (when hash-sample messages are available), used alongside existing message-hash tracking to help detect changes that break prompt caching. It also tracks a summaryForks map recording parent/child cache state around session summarization and forking (capped at a limit), and reports cache-break telemetry (cacheBroke, changeKinds).
  • When restoring cached session state from disk, entries now compute a hydratedHashRows value from a stored prefixBreakHash when its salt is recognized, and reset callCount/cacheDeletionsPending on restore.
  • The cache-diagnostics tracker also accepts a baselineAgentId, and when a forked agent has no prior baseline of its own, it seeds its baseline from its parent's cache state (forkAgentId/parentKey) instead of starting cold, recording where the baseline came from (baselineHashSource: memory, disk, or none).

Why

This gives better diagnostics for why prompt caching breaks, and it means forked or restored sessions no longer lose their cache-diagnostic history, reducing false alarms about the cache having broken.

How sure we are
Something disagreesSomething we can check disagrees with this entry, or the writer said they could not settle it.
The writer flagged doubtWhat user-visible effect this has, if any, isn't stated in the finding.

See this entry in the whole of v2.1.267 →

Feedback