[Feature]: SYSTEM prompt无条件注入"Normal: After completing a task, compress"

[Lisc1985]

Problem

注入链路：
L7236: BUNDLED_EDITABLE_PROMPTS.system = SYSTEM (含"Normal: After completing a task...")
↓
L7250: createBundledRuntimePrompts() → { system: BUNDLED_EDITABLE_PROMPTS.system }
↓
L7428: PromptStore.reload() → this.runtimePrompts = nextPrompts (无customPrompts时直接赋值)
↓
L8686: createSystemPromptHandler() → prompts.getRuntimePrompts()
↓
L8687-8692: renderSystemPrompt(runtimePrompts, ...) → [prompts.system.trim(), ...extensions].join()
↓
L8694: output.system[last] += "\n\n" + newPrompt (追加到system message末尾)
跳过条件只有4个（L8668-8684）：

1. isSubAgent && !allowSubAgents — 子agent未允许
2. INTERNAL_AGENT_SIGNATURES — 内部压缩agent
3. effectivePermission === "deny" — 压缩权限拒绝
4. 无system message时push新条目（L8696），但仍注入
   没有任何context使用率检查。

我跑长任务，一直发这个system promp，跑一半被压缩了，告诉我跑完了。
环境：opencode 1.17.13

Proposed Solution

是不是给个开关或者参数，比如:LLM每次调度/每次会话/关闭。每轮会话都注入对GLM5.2这种执行力强的模型来说，很容易跑一半完成一个子任务就压缩了。

Alternatives Considered

No response

Additional Context

No response

[ranxianglei]

Please upgrade to version 1.8.0.

The issue you raised has been thoroughly tested over several days and is now resolved. The original DCP implementation used overly aggressive prompt compression. We have now adjusted it to a more balanced approach—one that remains moderate while ensuring neutral compression behavior.

After extensive testing, here are the observed results for GLM-5.2:

• Previously, context usage was capped below 10%.
• With the new configuration, compression is now demand-driven:
  • Simple tasks still stay under 10%.
  • Complex tasks typically use around 20%.
  • Very long tasks generally fall within the 30–40% range.

Importantly, throughout the entire task process, critical information and fine-grained details in the context are preserved with minimal loss.

If you're interested in the details, please review the associated pull request here:
#47

Feel free to report any issues you encounter—we're ready to assist.

@Lisc1985

[ranxianglei]

1.8.1 has been further optimized. Please upgrade to version 1.8.1 or higher. #51

[Lisc1985]

收到，拉完代码试了一下，确实已经解决了。压缩效果不错，加油。
@ranxianglei