Skip to content

Commit 3e17c54

Browse files
ranxiangleiranxianglei
committed
merge: master (PR #351 all-field compress.providers cascade) into release v1.14.26
2 parents 4c4262d + 46014bc commit 3e17c54

14 files changed

Lines changed: 2007 additions & 15 deletions

File tree

‎CONFIGURATION.md‎

Lines changed: 72 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -189,7 +189,39 @@ Core compression behavior.
189189
- **Type:** `number`
190190
- **Default:** `5`
191191
- **Status:** ACTIVE
192-
- **Description:** Floor for growth-triggered nudges, as a percentage of the model context window: a growth nudge requires context usage at or above this percentage (in addition to the growth threshold). Over-max (`maxContextLimit`) and the 98% emergency-override nudges bypass the floor. If the model context window is unknown, the floor is unresolvable and growth nudges fall back to growth-only behavior. Turn/iteration reminder nudges are governed by `minContextLimit`, not this field. The default is deliberately low: with the default `nudgeGrowthTokens` (50K), a 5% floor stays inert for typical working cycles and only binds on very large (≥2M-class) windows — a higher default (e.g. 15%) would bind on ≥400K windows and shift every compress cycle's working range upward on large-window models. Set `0` to disable the floor entirely, or raise it (e.g. 15–30%) to keep growth nudges waiting until a larger share of the window is in use.
192+
- **Description:** Floor for growth-triggered nudges, as a percentage of the model context window: a growth nudge requires context usage at or above this percentage (in addition to the growth threshold). Over-max (`maxContextLimit`) and the 98% emergency-override nudges bypass the floor. If the model context window is unknown, the floor is unresolvable and growth nudges fall back to growth-only behavior. Turn/iteration reminder nudges are governed by `minContextLimit`, not this field. The default is deliberately low: with the default `nudgeGrowthTokens` (50K), a 5% floor stays inert for typical working cycles and only binds on very large (≥2M-class) windows — a higher default (e.g. 15%) would bind on ≥400K windows and shift every compress cycle's working range upward on large-window models. Set `0` to disable the floor entirely, or raise it (e.g. 15–30%) to keep growth nudges waiting until a larger share of the window is in use. Can be narrowed per provider / per model via [`compress.providers`](#compressproviders) (issue #344).
193+
194+
#### `compress.providers`
195+
- **Type:** `Record<string, ProviderOverrides>` where `ProviderOverrides = Partial<CompressOverridableConfig> & { models?: Record<string, Partial<CompressOverridableConfig>> }` (all fields optional at both levels)
196+
- **Default:** `undefined`
197+
- **Status:** ACTIVE
198+
- **Description:** Nested per-provider / per-model overrides for **every tunable compress field**, resolved field-by-field with the cascade **model > provider > global** (mirrors the sibling project billion-context-pi, issue #344). Deeper levels only override when the field is explicitly set — unset fields never clear shallower values. `0` / `false` are explicit values, not "unset". Unknown provider/model ids fall back to the global value. Percentages and `"X%"` limits resolve against the active model's context window. Across the three config file layers (global → config dir → project) the maps deep-merge per provider/model key — a project layer can narrow one provider without wiping others configured in lower layers.
199+
- **Overridable fields:** `maxContextLimit`, `emergencyThresholdPercent`, `minNudgeContextPercent`, `nudgeFrequency`, `iterationNudgeThreshold`, `toolOutputNudgeThreshold`, `nudgeGrowthTokens`, `minNudgeGrowthRatio`, `minNudgeGrowthFloor`, `nudgeForce`, `protectedTools`, `showCompression`, `summaryBuffer`, `protectTags`, `protectUserMessages`, `maxSummaryLengthHard`, `minCompressRange`, `maxVisibleSegments`, `keepEmbedMaxChars`, `lastSegmentSoftBlock`, `preserveRecentMessages`, `preserveRecentTokens`, `preserveLastUserMessage`.
200+
- **Not overridable:** `permission` (session-level, fixed before model info is known), the deprecated `minContextLimit` / `modelMinLimits` family, the flat `modelMaxLimits` map (legacy), and `providers` itself. For `maxContextLimit` the precedence when set nested is **nested override > `modelMaxLimits` flat map > global**. `protectedTools` set here affects the compress tool and nudge-side logic; the system-prompt protected-tools listing (shown at prompt build time, before model info is available) always reflects the global value.
201+
202+
```jsonc
203+
{
204+
"compress": {
205+
"maxContextLimit": "55%",
206+
"minNudgeContextPercent": 5,
207+
"nudgeGrowthTokens": 50000,
208+
"providers": {
209+
"anthropic": {
210+
"minNudgeContextPercent": 8,
211+
"nudgeForce": "strong",
212+
"models": {
213+
"claude-sonnet-4-6": {
214+
"minNudgeContextPercent": 30,
215+
"maxContextLimit": "70%",
216+
"nudgeGrowthTokens": 20000
217+
}
218+
}
219+
}
220+
}
221+
}
222+
}
223+
```
224+
In this example, for `anthropic/claude-sonnet-4-6`: the floor is 30%, the over-max band starts at 70% of the window instead of the global 55% (a larger working range before over-max nudges kick in), the growth threshold is 20K, and nudges use the `strong` tone (inherited from the provider level). Every other Anthropic model gets the 8% floor and `strong` tone but keeps the global band and 50K growth threshold; everything else uses the pure global values. Provider keys are provider ids and model keys are model ids (as reported by the active session, e.g. `anthropic`, `claude-sonnet-4-6`).
193225

194226
#### `compress.nudgeGrowthTokens`
195227
- **Type:** `number`
@@ -447,6 +479,45 @@ Post-compression quality evaluation. Runs after each compression to verify summa
447479
}
448480
```
449481

482+
### Per-model growth-nudge floor (nested providers.models)
483+
```jsonc
484+
{
485+
"compress": {
486+
"minNudgeContextPercent": 5,
487+
"providers": {
488+
"anthropic": {
489+
"minNudgeContextPercent": 8,
490+
"models": {
491+
"claude-sonnet-4-6": { "minNudgeContextPercent": 30 },
492+
"claude-haiku-4-5": { "minNudgeContextPercent": 0 }
493+
}
494+
}
495+
}
496+
}
497+
}
498+
```
499+
Floors resolve as model > provider > global (field-by-field). `0` disables the floor for that model — useful for small-window models where the growth threshold alone is the right signal.
500+
501+
### Per-model tuning of any compress field (nested providers.models)
502+
The same cascade works for every tunable field, not just the floor — e.g. give one heavy model a tighter growth threshold and a lower over-max band while its siblings keep the global profile:
503+
```jsonc
504+
{
505+
"compress": {
506+
"providers": {
507+
"anthropic": {
508+
"models": {
509+
"claude-sonnet-4-6": {
510+
"nudgeGrowthTokens": 20000,
511+
"maxContextLimit": "40%"
512+
}
513+
}
514+
}
515+
}
516+
}
517+
}
518+
```
519+
See the [`compress.providers`](#compressproviders) reference for the full overridable field list.
520+
450521
### Protect sensitive files
451522
```jsonc
452523
{

‎CONFIGURATION.zh-CN.md‎

Lines changed: 72 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -189,7 +189,39 @@ ACP 从最多三层配置文件中读取（后加载的覆盖先加载的）：
189189
- **类型：** `number`
190190
- **默认值：** `5`
191191
- **状态：** ACTIVE
192-
- **说明：** 增长触发型 nudge 的上下文使用率下限（占模型上下文窗口的百分比）：增长 nudge 要求上下文使用率达到或超过此百分比（此外还需满足增长阈值）。超过上限（`maxContextLimit`）与 98% 紧急覆盖的 nudge 不受此下限约束。若模型上下文窗口未知，则无法计算该下限，增长 nudge 回退为仅按增长触发的行为。轮次/迭代提醒 nudge 由 `minContextLimit` 控制，而非此字段。默认值刻意设低：在默认 `nudgeGrowthTokens`（50K）下，5% 下限对典型工作周期不生效，仅在非常大的（≥2M 级）窗口上才生效——更高的默认值（如 15%）会在 ≥400K 窗口上生效，并推高大型窗口模型上每个压缩周期的工作区间。设为 `0` 可完全禁用该下限，或调高（如 15–30%）使增长 nudge 等待更大的窗口占用比例。
192+
- **说明：** 增长触发型 nudge 的上下文使用率下限（占模型上下文窗口的百分比）：增长 nudge 要求上下文使用率达到或超过此百分比（此外还需满足增长阈值）。超过上限（`maxContextLimit`）与 98% 紧急覆盖的 nudge 不受此下限约束。若模型上下文窗口未知，则无法计算该下限，增长 nudge 回退为仅按增长触发的行为。轮次/迭代提醒 nudge 由 `minContextLimit` 控制，而非此字段。默认值刻意设低：在默认 `nudgeGrowthTokens`（50K）下，5% 下限对典型工作周期不生效，仅在非常大的（≥2M 级）窗口上才生效——更高的默认值（如 15%）会在 ≥400K 窗口上生效，并推高大型窗口模型上每个压缩周期的工作区间。设为 `0` 可完全禁用该下限，或调高（如 15–30%）使增长 nudge 等待更大的窗口占用比例。可通过 [`compress.providers`](#compressproviders) 按 provider / 按模型细化（issue #344）。
193+
194+
#### `compress.providers`
195+
- **类型：** `Record<string, ProviderOverrides>`，其中 `ProviderOverrides = Partial<CompressOverridableConfig> & { models?: Record<string, Partial<CompressOverridableConfig>> }`（两级均所有字段可选）
196+
- **默认值：** `undefined`
197+
- **状态：** ACTIVE
198+
- **说明：** 对**所有可调 compress 字段**的嵌套按 provider / 按模型覆盖，逐字段按 **模型 > provider > 全局** 级联解析（与姊妹项目 billion-context-pi 一致，issue #344）。深层仅在该字段被显式设置时才覆盖——未设置的字段不会清空浅层取值；`0` / `false` 是显式值，而非“未设置”。未知的 provider/model id 回退到全局值。百分比与 `"X%"` 限额按当前激活模型的上下文窗口换算。在三个配置文件层（全局 → 配置目录 → 项目）之间，该映射按 provider/model 键深度合并——项目层可以只细化某个 provider 而不清掉低层配置的其他 provider。
199+
- **可覆盖字段：** `maxContextLimit`、`emergencyThresholdPercent`、`minNudgeContextPercent`、`nudgeFrequency`、`iterationNudgeThreshold`、`toolOutputNudgeThreshold`、`nudgeGrowthTokens`、`minNudgeGrowthRatio`、`minNudgeGrowthFloor`、`nudgeForce`、`protectedTools`、`showCompression`、`summaryBuffer`、`protectTags`、`protectUserMessages`、`maxSummaryLengthHard`、`minCompressRange`、`maxVisibleSegments`、`keepEmbedMaxChars`、`lastSegmentSoftBlock`、`preserveRecentMessages`、`preserveRecentTokens`、`preserveLastUserMessage`。
200+
- **不可覆盖：** `permission`（会话级，在得知模型信息前已固定）、已废弃的 `minContextLimit` / `modelMinLimits` 系列、旧版扁平 `modelMaxLimits` 映射、以及 `providers` 本身。`maxContextLimit` 在嵌套层设置时的优先级为 **嵌套覆盖 > `modelMaxLimits` 扁平映射 > 全局**。在此设置的 `protectedTools` 影响压缩工具与 nudge 侧逻辑；系统提示词中的受保护工具列表（在提示词构建时生成，早于模型信息可用）始终反映全局值。
201+
202+
```jsonc
203+
{
204+
"compress": {
205+
"maxContextLimit": "55%",
206+
"minNudgeContextPercent": 5,
207+
"nudgeGrowthTokens": 50000,
208+
"providers": {
209+
"anthropic": {
210+
"minNudgeContextPercent": 8,
211+
"nudgeForce": "strong",
212+
"models": {
213+
"claude-sonnet-4-6": {
214+
"minNudgeContextPercent": 30,
215+
"maxContextLimit": "70%",
216+
"nudgeGrowthTokens": 20000
217+
}
218+
}
219+
}
220+
}
221+
}
222+
}
223+
```
224+
此例中，对 `anthropic/claude-sonnet-4-6`：下限为 30%，超限（over-max）区间从窗口的 70% 开始（而非全局 55%，即更大的工作区间），增长阈值为 20K，nudge 语气为 `strong`（继承自 provider 层）。其他 Anthropic 模型获得 8% 下限与 `strong` 语气，但保留全局的区间与 50K 增长阈值；其余全部使用纯全局值。provider 键为 provider id，model 键为模型 id（取自当前激活会话上报的标识，如 `anthropic`、`claude-sonnet-4-6`）。
193225

194226
#### `compress.nudgeGrowthTokens`
195227
- **类型：** `number`
@@ -447,6 +479,45 @@ ACP 从最多三层配置文件中读取（后加载的覆盖先加载的）：
447479
}
448480
```
449481

482+
### 按模型设置增长 nudge 下限（嵌套 providers.models）
483+
```jsonc
484+
{
485+
"compress": {
486+
"minNudgeContextPercent": 5,
487+
"providers": {
488+
"anthropic": {
489+
"minNudgeContextPercent": 8,
490+
"models": {
491+
"claude-sonnet-4-6": { "minNudgeContextPercent": 30 },
492+
"claude-haiku-4-5": { "minNudgeContextPercent": 0 }
493+
}
494+
}
495+
}
496+
}
497+
}
498+
```
499+
下限按 **模型 > provider > 全局** 逐字段解析。`0` 表示为该模型禁用下限——适用于小窗口模型，仅靠增长阈值触发即可。
500+
501+
### 按模型调优任意 compress 字段（嵌套 providers.models）
502+
同样的级联适用于所有可调字段，而不仅是下限——例如给某个重度使用的模型设置更紧的增长阈值与更低的超限区间，而同 provider 的其他模型保持全局配置：
503+
```jsonc
504+
{
505+
"compress": {
506+
"providers": {
507+
"anthropic": {
508+
"models": {
509+
"claude-sonnet-4-6": {
510+
"nudgeGrowthTokens": 20000,
511+
"maxContextLimit": "40%"
512+
}
513+
}
514+
}
515+
}
516+
}
517+
}
518+
```
519+
完整可覆盖字段列表见 [`compress.providers`](#compressproviders) 参考节。
520+
450521
### 保护敏感文件
451522
```jsonc
452523
{

‎README.md‎

Lines changed: 42 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -344,6 +344,21 @@ Each level overrides the previous, so project settings take priority over global
344344
// "openai/gpt-5.3-codex": 50000,
345345
// "anthropic/claude-sonnet-4.6": "25%"
346346
// },
347+
// Nested per-provider/per-model overrides for ANY compress field
348+
// (23 fields: thresholds, nudge behavior, protection, ...).
349+
// Resolution is per field: model > provider > global. Unknown
350+
// provider/model IDs fall back to the global value.
351+
// "providers": {
352+
// "anthropic": {
353+
// "nudgeGrowthTokens": 50000,
354+
// "models": {
355+
// "claude-sonnet-4.6": {
356+
// "maxContextLimit": "70%",
357+
// "minNudgeContextPercent": 10
358+
// }
359+
// }
360+
// }
361+
// },
347362
// How often the context-limit nudge fires (1 = every fetch, 5 = every 5th)
348363
"nudgeFrequency": 5,
349364
// Start adding compression reminders after this many
@@ -403,6 +418,33 @@ Each level overrides the previous, so project settings take priority over global
403418
404419
</details>
405420

421+
### Per-Provider / Per-Model Overrides
422+
423+
Any `compress` field can be overridden per provider and per model via the nested `compress.providers` map:
424+
425+
```jsonc
426+
{
427+
"compress": {
428+
"maxContextLimit": "80%",
429+
"providers": {
430+
"anthropic": {
431+
"nudgeGrowthTokens": 20000,
432+
"models": {
433+
"claude-sonnet-4.6": { "maxContextLimit": "70%", "nudgeForce": "strong" }
434+
}
435+
},
436+
"openai": { "nudgeGrowthTokens": 40000 }
437+
}
438+
}
439+
}
440+
```
441+
442+
Resolution is **per field**: model > provider > global. Unknown provider/model IDs fall back to the global value. A nested `maxContextLimit` also wins over the legacy flat `modelMaxLimits` map. Overrides deep-merge across the three config layers (global → config dir → project) per provider/model key.
443+
444+
Not overridable here: `permission`, the deprecated `minContextLimit` family, and the flat `model*Limits` maps themselves.
445+
446+
See the [`compress.providers`](./CONFIGURATION.md#compressproviders) reference in CONFIGURATION.md for the full 23-field list and recipes.
447+
406448
### Prompt Overrides
407449

408450
ACP exposes six editable prompts:

‎README.zh-CN.md‎

Lines changed: 42 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -297,6 +297,21 @@ ACP 使用自己的配置文件，按以下顺序搜索：
297297
// "openai/gpt-5.3-codex": 50000,
298298
// "anthropic/claude-sonnet-4.6": "25%"
299299
// },
300+
// 嵌套的 provider/model 级覆盖：可覆盖任意 compress 字段
301+
// （23 项：阈值、nudge 行为、保护策略等）。
302+
// 逐字段解析优先级：model > provider > 全局。
303+
// 未知的 provider/model ID 回退到全局值。
304+
// "providers": {
305+
// "anthropic": {
306+
// "nudgeGrowthTokens": 50000,
307+
// "models": {
308+
// "claude-sonnet-4.6": {
309+
// "maxContextLimit": "70%",
310+
// "minNudgeContextPercent": 10
311+
// }
312+
// }
313+
// }
314+
// },
300315
// How often the context-limit nudge fires (1 = every fetch, 5 = every 5th)
301316
"nudgeFrequency": 5,
302317
// Start adding compression reminders after this many
@@ -355,6 +370,33 @@ ACP 使用自己的配置文件，按以下顺序搜索：
355370
356371
</details>
357372

373+
### 按 Provider / 按模型覆盖
374+
375+
任意 `compress` 字段都可通过嵌套的 `compress.providers` 按 provider 和按模型覆盖：
376+
377+
```jsonc
378+
{
379+
"compress": {
380+
"maxContextLimit": "80%",
381+
"providers": {
382+
"anthropic": {
383+
"nudgeGrowthTokens": 20000,
384+
"models": {
385+
"claude-sonnet-4.6": { "maxContextLimit": "70%", "nudgeForce": "strong" }
386+
}
387+
},
388+
"openai": { "nudgeGrowthTokens": 40000 }
389+
}
390+
}
391+
}
392+
```
393+
394+
解析为**逐字段**优先级：model > provider > 全局。未知的 provider/model ID 回退到全局值。嵌套的 `maxContextLimit` 同时优先于旧版扁平 `modelMaxLimits` 映射。多层配置（全局 → 配置目录 → 项目）按 provider/model 键深合并。
395+
396+
不可覆盖：`permission`、已废弃的 `minContextLimit` 系列，以及扁平 `model*Limits` 映射自身。
397+
398+
完整 23 项字段清单与配方见 CONFIGURATION.zh-CN.md 的 [`compress.providers`](./CONFIGURATION.zh-CN.md#compressproviders) 参考节。
399+
358400
### Prompt 覆盖
359401

360402
ACP 暴露六个可编辑的 prompt：

0 commit comments

Comments
 (0)