[FEATURE]: Modify system prompt to improve cache hit rate

[BruceXSK]

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

I previously submitted an issue mentioning the impact of AGENTS.md on cache hit rates. See: #29584

Additionally, I have found that the environment information and the skills list in the system prompt also affect the cache hit rate.

Currently, the environment information in the system prompt includes: model id & provider, working directory, workspace root folder, git repo flag, platform, today's date.

Below is my analysis of each field's impact on cache hits:

• ✅ model id & provider: It only changes when switching models, and different models have different KV caches, so switching models naturally leads to cache misses. So this field is fine.
• ✅ working directory: Once a session is created, this field does not change dynamically, and it is information the LLM needs to know.
• ✅ platform: This field is stable, would not change during the lifetime of a session, and is also information the LLM needs to know.
• ❌ workspace root folder & git repo flag: In my testing, in a non-git repo directory, workspace root folder is /, and git repo flag is false. While the session remains in the foreground and I run git init in that directory, neither field changes. However, once I exit opencode and resume the session after git init or rm -r .git, workspace root folder becomes the same as the working directory, and git repo flag becomes true, causing all subsequent token caches to miss.
• ❌ today's date: After crossing midnight, this field changes, causing all subsequent token caches to miss. DeepSeek servers may keep caches on disk for several hours or even days; the cache miss caused by the date change is unnecessary.

Therefore, I suggest the following regarding the env section of the system prompt:

• Remove workspace root folder and git repo flag. The agent can use git commands to check whether the current path is a git repo. I don't think this information needs to be in the system prompt.
• Remove today's date. The agent can also query this using commands. I don't think this information needs to be in the system prompt either.

In addition to the modifications to the env section of the system prompt, I also suggest placing the skill list before env. Because the skill list is more stable than env (e.g., working directory), this will yield better cache hit rates when using opencode in different working paths.

Related code packages/opencode/src/session/system.ts:42-80

OpenCode version: 1.15.10

[chrisjimallen]

Adding quantitative data to this issue: the <available_skills> block introduced in v1.16.0 causes a 5-7x cost increase on DeepSeek due to defeated prompt caching.

Measured impact (single DeepSeek v4-pro user):

Metric	Pre-upgrade (May, v1.15.x)	Post-upgrade (June 8, v1.16.x)
Cache miss rate (pro)	1-2%	10-15%
Daily cost	~$0.36/day	~$5.03/day (one day)
Cache miss tokens/day	1-2M	8-14M

Why it matters for cost: DeepSeek charges cache misses at 120x the cache-hit rate ($0.000000435 vs $0.000000003625 per token). The skill listing block pushes the system prompt past the cache boundary.

Effective workaround (verified -- miss rate dropped from 14% to 1.59% after applying):

{
  "agent": {
    "build": {
      "tools": {
        "skill": false
      }
    },
    "plan": {
      "tools": {
        "skill": false
      }
    }
  }
}

Restart opencode after applying. Verify via DeepSeek usage CSV at platform.deepseek.com.

[0byte-coding]

Critical Issue: System Prompt Changes Break Entire Session Cache

This is a serious performance and cost problem that goes beyond just AGENTS.md updates. When ANY system prompt component changes during an active session, it causes a complete cache miss and forces the entire session context to be re-sent to the API.

The Problem

When system prompt is modified (whether due to AGENTS.md changes, env variable changes like workspace root or git status, or date changes at midnight), the architecture forces a re-transmission of all previously cached context:

[System Prompt - ORIGINAL - 200K tokens cached]
[User Message 1 - cached]
[Assistant Response 1 - cached]
[User Message 2 - cached]
[Assistant Response 2 - cached]
... (many more turns)

[System Prompt CHANGES] ❌ ENTIRE CACHE INVALIDATED
[System Prompt - MODIFIED]
[User Message 1 - RE-SENT, no cache]
[Assistant Response 1 - RE-SENT, no cache]
[User Message 2 - RE-SENT, no cache]
[Assistant Response 2 - RE-SENT, no cache]
... (entire session re-sent)

This is not feasible for long-running sessions with 200K+ tokens of context.

Solution

Don't Update System Prompt for Existing Sessions

When rules/AGENTS.md change during an active session, the AI agent already knows what changed because it made the change itself (or saw it in the context). There's no need to update the system prompt:

• If the agent modifies a rule via edit command → it knows the change
• If the user modifies AGENTS.md → communicate via <system-reminder> (like Claude Code does)
• The system prompt should only reflect the original session state; rule changes are part of the conversation history