[Feature]: 增加手动设置上下文压缩阈值

[JohnXu22786]

What workflow is missing or painful today?

我觉得可以除了内置还可以整一个手动设置的，因为自动压缩阈值较低可以使得更加节约上下文成本。比如说有的长度是 20 万，实际上那一个在 4 万左右就够了的话，多余的会比较浪费💰

Proposed change

具体来说通过自定义一组预设定义模型的api等，包括上下文这种，这样自定义预设可以发挥更大的作用，比如一个模型可能有多种上下文设置的可能性（对于不同的项目），然后在 OH 界面里面，通过选择不同的模型，直接就是可以选到同一个模型的不同的预设。

Alternatives considered

No response

Additional context

No response

[fengjie926]

有 /compact 指令.

我觉得还是自动管理比较好, 除非你的场景问题比较发散关联性低, 还特别注重 TOKEN 成本.

根据我的使用, 我发现目前这方面离 CC 还比较远, 目前代码里面是 到达/model set模型 配置容量的 3/4 就移出掉工具调用 保留最近 6 条. 不支持CC 那种压缩后还能按需回顾. 希望官方能考虑这个智能的特性

[tjb-tech]

Implemented on main in commit 39ffc19.

OpenHarness now supports manual compact threshold tuning via config/env/profile state:

• context_window_tokens
• auto_compact_threshold_tokens

This is wired through settings -> runtime -> query -> compact, so it affects the real auto-compact decision path rather than being a no-op config field.

What changed:

• added configurable compact threshold/context window fields in settings and provider profiles
• passed them into the runtime/query/compact pipeline
• added regression coverage in tests/test_config/test_settings.py and tests/test_services/test_compact.py

Validation:

• focused unit/regression runs passed locally
• real agent-loop validation was also run on /tmp/openharness-eval
• with an aggressively low override, the real conversation emitted compact lifecycle events (hooks_start, compact_start, compact_retry, compact_failed, context_collapse_start, context_collapse_end), confirming the override is active in the live path

Note: setting the threshold very low will make compaction much more aggressive and can slow real tasks down, but the override itself is now working.