Gemini and Vertex prompt cache drops out due to 4m warming interval and missing cachedContent #43628

Activity

1. github-actions

   assigned

   kitlangton

   on Aug 20, 2026
2. github-actions

   added

   needs:complianceThis means the issue will auto-close after 2 hours.This means the issue will auto-close after 2 hours.

   on Aug 20, 2026
3. github-actions commented on Aug 20, 2026

   

   github-actionsBot

   on Aug 20, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue doesn't fully meet our contributing guidelines.

   What needs to be fixed:

   • This appears to be a feature request, but it does not use the required Feature Request issue template. Please resubmit using the template, which requires: a title starting with [FEATURE]:, a verification checkbox, and a proper description of the problem/need.
   • The current title does not start with [FEATURE]: as required for feature requests.

   Please edit this issue to address the above within 2 hours, or it will be automatically closed.

   ---

   This issue might be a duplicate of existing issues. Please check:

   • Improve prompt caching for Gemini models on OpenRouter #36069: Improve prompt caching for Gemini models on OpenRouter — similarly proposes improvements to Gemini prompt cache behavior (cache breakpoints, session routing).

   If you believe this was flagged incorrectly, please let a maintainer know.
4. major

   changed the title [-]ai: improve Gemini and Vertex prompt cache efficiency with explicit cachedContent and keep-alive warmup[/-] [+]Gemini and Vertex prompt cache drops out due to 4m warming interval and missing cachedContent[/+] on Aug 20, 2026
5. github-actions

   removed

   needs:complianceThis means the issue will auto-close after 2 hours.This means the issue will auto-close after 2 hours.

   on Aug 20, 2026
6. github-actions commented on Aug 20, 2026

   

   github-actionsBot

   on Aug 20, 2026 – with GitHub Actions

   Contributor

   More actions

   Thank you for updating the issue! It now meets our contributing guidelines. We appreciate the clear bug report with steps to reproduce, version, and system info.
7. kxng0109 commented on Oct 9, 2026

   

   kxng0109

   on Oct 9, 2026

   More actions

   Adding session level evidence for the Vertex Gemini dropout. Tested route google-vertex/gemini-3.8-flash on opencode 2.0.22, auth through Application Default Credentials.

   Method: read all 29 assistant turns for one session straight from the local session database, recording provider, model, variant, input tokens, cache read tokens, cache write tokens, output tokens, thinking signatures, and timestamps.

   Results:
   Cache write equals 0 on all 29 turns, so this route relies purely on provider side implicit caching.
   7 of 29 turns billed the full context fresh with 0 cached, inputs from 24k to 61k.
   The other 22 turns hit normally.
   Misses show no link to time gaps. Gaps of 3 seconds miss while a gap of 297 seconds hits.
   Misses show no link to thinking signatures, reply sizes, or tool use. One miss had no tool calls at all, only thinking plus a short reply.
   No wall or plugin activity fired during the session, and no instruction files changed that day.
   Provider, model, and variant are identical on every turn.

   Wider comparison from the same database: miss rates of 1.9 percent and 2.8 percent on opencode provider models across thousands of turns, 0.3 percent on another model across 1500 turns, against 29 percent and 21 percent on Vertex Gemini 3.8 and 3.7 across about 60 turns. No explicit cache writes on any route, so the gap sits provider side.

   First turn miss is the normal cold start and is excluded from the above. Full per turn table available on request.
8. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   Contributor

   More actions

   Thanks for the detailed report, @major!

   In v2, session warming is opt-in. With "warming": true it uses a 4-minute idle interval by default, as described in the warming docs. That default works for providers with a ~5-minute cache TTL, but Google's implicit cache can expire sooner. The interval is configurable, so for Gemini or Vertex you can warm more often today:

   {
     "warming": {
       "interval": "2 minutes"
     }
   }

   For explicit caching, the Gemini protocol already passes through a cachedContent you provide. OpenCode doesn't create or manage Google cachedContents for you, though.

   So I'm treating this as a feature request with two parts:

   • warming defaults that depend on the provider (e.g. a shorter interval for google / google-vertex)
   • automatic explicit context caching (cachedContent) for Gemini

   The maintainers will consider it.