onlyterp/prompt-cache-skills/continue-gemini-explicit

Claude Code and agent skill by onlyterp/prompt-cache-skills

Trust tier 2
  • community
LLM review pending
Source check pending

Checking that the SKILL.md file is still in the repository.

Continue's Gemini provider doesn't use the cachedContents API at all. Add explicit caching for sessions over the minimum token threshold.

Install

npx vskill@latest install onlyterp/prompt-cache-skills/continue-gemini-explicit

SKILL.md

Continue: implement Gemini explicit caching

Target

packages/openai-adapters/src/apis/Gemini.ts in continuedev/continue.

Symptom

The Gemini provider class does not use the cachedContents API at all. Gemini 2.5+ has implicit caching that fires automatically for byte-stable prefixes (free, best-effort), but explicit cachedContents gives guaranteed cost reduction (0.25x input price on Pro) and controllable TTL. Continue leaves this entire mechanism unused.

Result: long-running Gemini sessions with the same system prompt + tools (the exact use case caching is for) get only implicit-cache luck, not the guaranteed 75% discount.

Fix

Add a _maybeCreateCache helper and reuse the cached content across calls within a session:

diff
--- a/packages/openai-adapters/src/apis/Gemini.ts
+++ b/packages/openai-adapters/src/apis/Gemini.ts
@@
export class GeminiApi implements BaseLlmApi {
apiBase: string = "https://generativelanguage.googleapis.com/v1beta/";
private genAI: GoogleGenAI;
+ private cacheName: string | null = null;
+ private cacheSystemHash: string | null = null;
@@
+ private estimateTokens(text: string): number {
+ // Conservative ~4 chars/token; fine for threshold gating
+ return Math.ceil(text.length / 4);
+ }
+
+ private async _maybeCreateCache(
+ systemInstruction: string,
+ tools?: any[],
+ ): Promise<string | null> {
+ if (!systemInstruction) return null;
+
+ // Min tokens: 4096 for Pro variants, 1024 for Flash
+ const isPro = this.config.model.toLowerCase().includes("pro");
+ const minTokens = isPro ? 4096 : 1024;
+ if (this.estimateTokens(systemInstruction) < minTokens) return null;
+
+ // Dedupe: only create a cache if system has changed
+ const hash = crypto.createHash("sha256")
+ .update(systemInstruction + JSON.stringify(tools ?? []))
+ .digest("hex").slice(0, 16);
+ if (this.cacheName && this.cacheSystemHash === hash) {
+ return this.cacheName;
+ }
+
+ try {
+ const cache = await this.genAI.caches.create({
+ model: `models/${this.config.model}`,
+ config: {
+ systemInstruction,
+ tools,
+ ttl: "3600s", // 1 hour
+ },
+ });
+ this.cacheName = cache.name ?? null;
+ this.cacheSystemHash = hash;
+ return this.cacheName;
+ } catch (e) {
+ // Cache creation can fail (e.g. content under min). Fall back
+ // to implicit caching by returning null.
+ return null;
+ }
+ }

Then in the generateContent call path:

diff
+ const cachedContent = await this._maybeCreateCache(
+ systemInstruction,
+ tools,
+ );
+
const response = await this.genAI.models.generateContent({
model: this.config.model,
contents: messages,
+ config: cachedContent
+ ? { cachedContent }
+ : undefined,
});

Verify

  1. Start a Continue chat with gemini-2.5-pro or gemini-3-pro-preview

and a system prompt over 4096 tokens (or load a big AGENTS.md).

  1. Capture wire.
  2. First call: should see a POST /v1beta/cachedContents create call

followed by a generateContent call referencing the cache name.

  1. Second call: only generateContent, reusing the same

cachedContent reference.

  1. Inspect response usageMetadata.cachedContentTokenCount —

should be > 0 on both calls.

Background

Gemini has two caching paths: implicit (automatic, free, best-effort) and explicit (cachedContents API, guaranteed discount, configurable TTL, costs storage per hour). For agent loops with a stable large system prompt, explicit is strictly better — you trade pennies of storage for dollars of discount.

Minimums:

  • Gemini 3.5 Flash / 2.5 Flash: 1024 tokens
  • Gemini 3 Pro Preview / 2.5 Pro: 4096 tokens

Below threshold, caches.create() returns 400 — the fallback to implicit (return null) handles this gracefully.

See docs/concepts/gemini.md. Full audit: audits/continue.md.

Security

  • Security scanPassed, 100 of 100Pattern checks on the skill's files.
  • Model reviewNot runA model review is scheduled for the next rescan.

Quality evaluation

  • Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

Source

Popularity

Repository stars
114
Repository forks
9
7-day trend score
24

Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Codex
  • Gemini CLI
  • Cline
  • Amp
  • Roo Code
  • Goose
  • Continue
  • Kilo Code
  • Junie
  • Trae
  • OpenHands
  • Qwen Code
  • AI Agent
  • Tabnine Chat
  • Sourcegraph Cody
  • Amazon Q
  • Firebase Genkit
  • Sweep
  • Devin
  • Phind
  • Mentat
  • Claude Desktop
  • Aider
  • GPT Pilot
  • Plandex
  • bolt.new
  • v0
  • Replit Agent
  • Anse
  • TypingMind
  • BetterChatGPT
  • LibreChat
  • Warp AI
  • Pieces
  • Fig AI
Embed badge
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/onlyterp/prompt-cache-skills/continue-gemini-explicit/badge)](https://verified-skill.com/skills/onlyterp/prompt-cache-skills/continue-gemini-explicit)

Something wrong with this skill or its page? Tell us.