does this poor cache issue also affect --continue the same way as --resume?

[vyoungnyc]

is --continue also affected by cache issue like --resume if i just want to reopen claude for the last session i was working on?

use case would be making an update for my agents/commands/skills, and need to exit claude, and reopen for it to pickup the new agents, or i need to add something like an mcp and need to restart claude.

[ArkNill]

Hi @vyoungnyc, thanks for the question!

Yes, --continue is affected the same way as --resume. Both completely invalidate the prompt cache on re-entry.

This was confirmed with controlled tests in anthropics/claude-code#42338 — all three resume methods (--continue, --continue --dangerously-skip-permissions, /resume) drop cache_read to 0 and trigger a full cache_creation of the entire prompt, even when re-entering within seconds (well within the 5-minute cache TTL).

For your use case (restart to pick up new agents/MCP/skills)

• Each restart + continue costs a full cache rebuild — on a 500K token session, that's a significant chunk of your quota per restart
• Workaround: Start a fresh session and manually provide the relevant context instead of using --continue. This avoids replaying the full history as billable input.
• If you must continue: Keep sessions short before restarting so the cache rebuild is cheaper.

References

• anthropics/claude-code#42338 — Session resume invalidates entire prompt cache (all methods tested)
• anthropics/claude-code#34629 — Resume cache regression (deferred_tools_delta)
• anthropics/claude-code#42260 — Resume replays opaque thinking signatures as billable input

I also noticed your related issue at anthropics/claude-code#42277 — the rapid drain you experienced there is consistent with this cache invalidation behavior. Hope this helps!