@jhckevin/dsh-composer-prediction 0.8.0-beta.1 → 0.8.0-beta.3

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -1,3 +1,11 @@
1
+ # 0.8.0-beta.3 — 2026-11-17
2
+
3
+ Add a saved prediction reasoning policy: off/minimum by default, explicit opt-in to parent/profile controls. Migration preserves the minimum default. Local result identities isolate policies. Switching does not automatically generate another paid prediction for the same turn. Bilingual native settings selector has an associated label, help text and explicit save.
4
+
5
+ # 0.8.0-beta.2 — 2026-11-17
6
+
7
+ Prediction and validation repair use disabled thinking when supported, otherwise the exact model minimum. Native adapter metadata is authoritative; portable Responses, Chat and Messages have protocol-specific controls and declarative aliases. Main conversation settings remain unchanged. Unknown controls stop before dispatch, with no paid probes or hidden fallback requests. Local result cache schema changes to avoid older inherited-High results. Record reasoning policy per attempt; recognize Anthropic thinking-token details without double billing.
8
+
1
9
  # 0.8.0-beta.1 — 2026-11-17
2
10
 
3
11
  First public npm preview under @jhckevin/dsh-composer-prediction. Declare a native DSH registry bundle, retain composer-prediction runtime/data identity, refresh bilingual details and installation/update guidance. Native Versions checks, compatibility, progress and rollback remain host-owned. Existing 0.7.5 prediction/cache/accounting behavior is retained.
package/MANIFEST.sha256 CHANGED
@@ -1,9 +1,9 @@
1
- aea53f66e85a7b6a9a964bfce14175571546ba2832be94d9d20348ec663a0b79 CHANGELOG.md
1
+ 4f6bb6fe382953b544d3198c32282b1647f9262232757ea57ced83b01441c627 CHANGELOG.md
2
2
  ebb4f09972aee8608be255debaf78451a68e95c290f55c240dec2ecfa16ea6be LICENSE
3
- 9da4d2582e53d09952bb153a8041c020b7460cf8af33c2eced05f57aeaeab90e PROTOCOL-COMPAT.zh-CN.md
4
- ca61bce893c75b154ffd8fc667945039f035ed04234624d23f3febbaad2ab1f5 README.md
5
- 24cf6f08fae5357b74dd4053752244cc64ca79c01ec8f501016ba07ea569b6a3 README.zh-CN.md
6
- 5318b0854b7e32afa1c63ed4e7db5594ec70924ed93f8e17ac7f7db022bb2011 SOURCE.json
3
+ 55f3d480cecada92f61849527adaad61bf58c2410098028a3b5e0b6f679655dd PROTOCOL-COMPAT.zh-CN.md
4
+ 6fc1f64e52d42248ff0e05e12aad511dfd266dcbb5373322c342f364e157cd95 README.md
5
+ 577742d5da68c5e6b4b2b23b199dfcde66ca9dd64303ad2b93a5c26c9880b8af README.zh-CN.md
6
+ 4c328b2934b4e597ff32923e4e691ad25e67f5fffe60cc0446801a474a4b07e1 SOURCE.json
7
7
  539c9ce04d26f03170ddad0d797b1149149eb7935934b6f01f8dcfbfb7e87ea4 THIRD_PARTY_NOTICES.md
8
8
  8a118cf6453b9678b4dc20df66931d923008747fdac46ef12bdb7a29b2d1163b assets/icon-source.png
9
9
  d4bf21a13ffe3aa7d5740a36212505ad05238c1bc6fdfab88b24888457c392e9 assets/icon.png
@@ -19,8 +19,8 @@ b32fc5a1f8dc40a4e30419f40b8e57361db066e908144dedd6faa99cc978b6da assets/screens
19
19
  a3e7d76d90d879dc5eafcd6103d9942ea15f8cf8014b514e23e7a5aa5252575c assets/screenshots/responsive-costs-1000-zh.png
20
20
  b5fbc70b2fcc93ef8716d17e1d518e644de11439f7f524c32f34072f7ad85c46 assets/screenshots/responsive-costs-420-en.png
21
21
  b25b33d947731153406f4134c440a78a6053191e9e1542a7f131cda347ad333a assets/screenshots/responsive-costs-420-zh.png
22
- 945812361ac47173b643e6cef1b24a1fb3246369c3181e226aaf522d01b0c5c2 assets/screenshots/responsive-settings-en.png
23
- dca2a7f970390108a855c5b05aeb92a8e50987a7fcd5dbb66bd0fdfa852a8c45 assets/screenshots/responsive-settings-zh.png
22
+ 3384be48a76f9397fc2232ce04d1b28c70af89dc07ef949fc9912e896b7bfc9a assets/screenshots/responsive-settings-en.png
23
+ da15729b9ea5a47de47e7098b3b47da86112bc0a29fee390529c5288cabbd452 assets/screenshots/responsive-settings-zh.png
24
24
  1e926df7d883be60e3430fe01aecebd7a3cf01e476da921e5d605ede20d375a2 assets/screenshots/settings-en.png
25
25
  8dac7330c582495aacc440a4a7ccbe1369568a9e324a32c4080bbf635af7fc74 assets/screenshots/settings-narrow-en.png
26
26
  baab48c764bd0e1c93db00c1a10a66c7665a75009bdc366f30e57f8df4427f66 assets/screenshots/settings-narrow-zh.png
@@ -30,17 +30,17 @@ baab48c764bd0e1c93db00c1a10a66c7665a75009bdc366f30e57f8df4427f66 assets/screens
30
30
  6f24c57b686d7f884f89f0be79c71dd4c7c6f8bc7e26fa5c22b5bc34edfec61a assets/screenshots/wrap-540-bottom.png
31
31
  65f0f2aec4e8c3baff80ef5325cf0aa493e37434a889c8c24b827dbf9442b86c assets/screenshots/wrap-540-top.png
32
32
  d0b27a7a9e861db505e1a7b03d2612cd58c4f3af32b8e66c6dc4bd14418e419d cordis.patch.yml
33
- ceabec8b8038498ec52583e2c736c69a1c47f3df755225ab773d90b4c7d71d52 docs/README.en.md
34
- 3030921fc7ee379d065752ad31ad95a3fd718df57ba686f8b8cf11e02b738780 docs/README.zh-CN.md
35
- 7952c7f60762a06438f4e923a09706e41bb3b3e4a84f55dbe26197a6188f8596 examples/profiles/deepseek-flash.json
36
- 1ec94fe79a348c5eabe187c4054a5c506639d7fdf26177f1f29531575c07c8bd examples/profiles/deepseek-v4-pro.json
33
+ 46ecc773fc7c3d721bee424efc65e63f2f0729258a3219acf02a1c96c606af1f docs/README.en.md
34
+ b5412c9f0a783841c292591b964bbfa28cc7e5df2635159c38d53442c7566a28 docs/README.zh-CN.md
35
+ d0149256d513fc98f469f6658eaa2c94996216b5cf962dacb9c3590eed5ed8e3 examples/profiles/deepseek-flash.json
36
+ e30e73f7b4d9217a4467324b452ec08b9e303e7aaa40a12d90292e73ace6d237 examples/profiles/deepseek-v4-pro.json
37
37
  03b3344ff72632fd5d894cb7c9d60a29c66dd83b05d44be8a5f657048e0c2d29 examples/profiles/generic-anthropic-messages.json
38
38
  797f84a30869f7646ae51d41225e3dcd517d109f9125c9a258d94b6fa421c7e0 examples/profiles/generic-chat-completions.json
39
39
  d0383b36002d035577d4ccb6d8617507268ced3e32e338995595bb57dc626d8f examples/profiles/generic-responses.json
40
40
  6c46ebd45fb215fbba851e55770628593f7d95566dd56d838b23c783b335f8dd examples/profiles/responses-custom.json
41
41
  15d1323eb515dd5f10240e62083fc7333523b74799f6452d708880f2e26f5c07 instruction_prediction.md
42
- ea9844299ed6d7f12769997e32788e552e1cf4f453c4cfcd6a1f72c2eefd5a76 lib/client.js
43
- d84b9255410a67df4a430031add39a7433baa2c3137f7848c35b01deec265d6c lib/client.js.map
42
+ 59cf34a520c7e05a3367864de20ddf36349e6ed9f9af63bd746c9c31e8a3577f lib/client.js
43
+ 90150157625cb054f1d4778bbbd6b94bba02cbd674a0fa06ceea4a925be02a18 lib/client.js.map
44
44
  c43240a988bb12e7dbf749fa9a53ec37a3060bdbdcc65ebebff248f672db6d07 lib/dsh-native.mjs
45
45
  1ee244c7672499bc3913bf6e9d1c92fe69af895eb52c8ca188b73c13da57da17 lib/dsh-native.mjs.map
46
46
  d884840a5871239ce3e105de464c35014de7dcbb8f5793aa213b8588cb74de0d lib/schema.mjs
@@ -59,11 +59,11 @@ c71d239df91726fc519c6eb72d318ec65820627232b2f796219e87dcf35d0ab4 licenses/huggi
59
59
  3f1189b28e3866e0d979968d466b78f813f76827cfdca1fbb124cc0a5c8841f8 licenses/zod-LICENSE
60
60
  3ff5c092b06f081efe1afe65e96d052ea7f1d6aa515aafb83fd832bbdac4a0c2 locale/en.json
61
61
  e86336a34ea404863d745ec512db7dc3b3540f57e377612b9cf3c239d86a8d74 locale/zh.json
62
- 2b5c903d50c0a88b7c911f8a214df0059142770e37c6d130210237733f0bdb54 package.json
62
+ 5449ba3b934e835c501f7e0cabc1d304f8e5f7da8dd876d3a3ae95c541de8ba7 package.json
63
63
  a332b0fcfdf49318d8f279d889bc38563ce1101d218de6443e3117f2dcd4b97a pricing/deepseek-flash.json
64
64
  5360068efc903eca97f8c391a0004f10f317d24de59cd48c40f2a86525aabe60 pricing/deepseek-v4-pro.json
65
65
  9552ee74970a65513d706a478a17d7949f0470364640ab94fa574e5b412eba1d src/client/blacklist.cjs
66
- ea17daf59fb1806f7e12271d3b38b1609b526665fcba2af206a9aa7b9e6af0a7 src/client/brand.cjs
66
+ f951aae3e645f39ce96d3cf2cd5b82ec7cf89226db2c8b18fdcdaa2de9d75d3f src/client/brand.cjs
67
67
  bc0e77bbe36895ce752913ed0ad39682d92d17d109ff1f502f62dd538d08a66b src/client/build.mjs
68
68
  4f95894b07526f5c6e5169ea67e20548286d3ad41d0374b4f18d9b012517f40e src/client/controller.cjs
69
69
  15c2e515b563d3243df07e7c175ab5f2ab22bbe1217506935b433e5f2629878d src/client/controls.cjs
@@ -71,30 +71,31 @@ bc0e77bbe36895ce752913ed0ad39682d92d17d109ff1f502f62dd538d08a66b src/client/bui
71
71
  9e2e570e0d1c0448d80abbb4a2c4bccedca5a7e058625f279751eac270e3f1bc src/client/guards.cjs
72
72
  bf62a0934dea387498cad3c808c352d7a4654f51e15ff30ad56f257cadc845a5 src/client/host-rpc.mjs
73
73
  51b7527a8de5c83a6ca7bab81c9f6428dbcb96f3b28931abd29fd30e59e97a21 src/client/index.cjs
74
- a0ac17677fc7a5934d17818fd0b62977781bbff32c32a31ed3f169f67488587c src/client/locales.cjs
74
+ 9dd6204392686b9ecef7db2e0e1055106962a25e49df967581076b39e6505de9 src/client/locales.cjs
75
75
  0f9c7c18effbc90741978cc7fe84d941d169bc24cf413ea1e484146866957a60 src/client/onboarding.cjs
76
76
  eb3567078dc210283410dbde0a2ddff05c29fb5f44d2aeedf9fc3226c6c93f70 src/client/placeholder.cjs
77
- 49014821f2d4afb8b7767e0361e46360b885276312bd1820eaa20cf421baf613 src/client/policy-fields.cjs
78
- 6c00daf080bc73f013f22bdd8838bb79aafcf37528928e7291c87b8e9b5f9809 src/client/settings.cjs
77
+ d87aeec17088fdac64cd1e95abc93e91b2672ea2c1cc789118269aac31d2fe62 src/client/policy-fields.cjs
78
+ 686fe08fa004f01cc75660affd93d859fc8dad7f8e582d9209aed3641d79e080 src/client/settings.cjs
79
79
  84c66848a1d9d908a959105be656230f87e74dd95bde964c0caeb0ea5d0a8e68 src/core/balance-watchdog.mjs
80
- d838d2846d89b6041b05205d80f7593cb562be28744f20e19b8c8cd8249f7158 src/core/billing.mjs
80
+ 9c0696533c96c33b8c192fe5398b45b1ba6587a1ff756f781de4a0ecc5cb87eb src/core/billing.mjs
81
81
  e75c4bfb8ebb5d29fc2a8cf5705f323e54e4d20be31b9e7d0cd0a9ed75ab76f5 src/core/blacklist.mjs
82
82
  8505236445caecab736cc4bce85b96f4d212cbf8e75d09073a9d379f9ecd2d7f src/core/budget.mjs
83
83
  b781129d86972e218085ee3acdeb9b905272fc13569769d29abc2aed9ec3bc2e src/core/context-events.mjs
84
84
  08fa74d953b514bb725d08e1b83567169de31de1d50ebda27da2e74018467169 src/core/context-mapping.mjs
85
85
  cf8201530df87d6d8c30c9196a02b2d8c2439d009087862dc760a4fc1227a0d5 src/core/deepseek-balance.mjs
86
86
  0ac89e51101510ae01420b49fa63956af590b1cf80f29d06fb91faf273526483 src/core/dsh-projection.mjs
87
- 7b6b25af39d7eb64167aa917e9f6b87a1d0539f5e535eeb123a2ccd0e0e469ee src/core/engine.mjs
87
+ 3135d7c137ff927491d8d761fc5459a4e3dbddea02d8f5e89985b44c13b178a5 src/core/engine.mjs
88
88
  99a0d72e3736858ada41a71cdb5e56c2fdc1eeb990b67bdbae1eac3f1c1248db src/core/event-coordinator.mjs
89
89
  b304803697cc1ff49334c2bd3776fdef1fc42607de3449e84f6637a0980cc272 src/core/lifecycle.mjs
90
90
  1b2210d68df4acefbe118ae9623416e3edc6bf4291b775942f0236fb57d99e10 src/core/loop-watchdog.mjs
91
91
  e4fd34fe898aaa85131c8f50f5ece445710133b7bcd8329c2c0751857cef58d8 src/core/ownership.mjs
92
- c38d16113b55f586165d440cfb0f83c68572dc842f88f747e6d712dad3fe4e70 src/core/policy.mjs
93
- d21f96331c2764c0117364feba7a67c327c8a26e31514137c68f8040f0c8d99f src/core/profiles.mjs
92
+ 75de9bdd3576aae03b1ef5a82e410f416d6bfd1d4a0b75d595b98db717a94cb1 src/core/policy.mjs
93
+ ed7fb40e234dd35696f8e778d18937ea55dcf2ad1f3c1a8c1e861a7b1e16da03 src/core/profiles.mjs
94
94
  64744e70e85d0a21908dfb76ea8eecbd0237820210fdc372450877ccd7a86e1b src/core/protocol-compat.mjs
95
- 4b3445a6b6c588246d85571aadfc2e3d043f63c5c964d72b6c78cfcb24ee7672 src/core/recovery-selector.mjs
96
- 0be849e70e27ae8778b8e86c43520b78824ff7aff2b2c9a4ea166cc75a7b9cf8 src/core/recovery.mjs
95
+ 0ad4d005f94f5727e864d049519b1ab9739398e510a4aab8edb73c8992a9d504 src/core/reasoning-policy.mjs
96
+ 66aedb94897ff196a88b1a45f6af1c415a11cbaecfdc7557eeeac8985dbf4190 src/core/recovery-selector.mjs
97
+ 52c2cd799e77f325bb5f47f5d19283bf796d5831ec1cca7fb3118608daa32c40 src/core/recovery.mjs
97
98
  764edb77856496eae3362af7e888bf364c007e119e3dd1c7b1bbe82b173d7b94 src/core/thread-accounting.mjs
98
99
  cf99dbee81740da7af9b3d882df560ca270a8b1510fa6eff875f8e9a5687544e src/core/token-counter.mjs
99
- 679abf05263bddb82a4bb2296627afc4769a4d105deaa0bed5231dae15780c54 src/core/transports.mjs
100
- fb349e7ab6072a54d1faa735b2170c9d2c1b1eec4568fa78a3bd5db93fa31974 src/index.mjs
100
+ 28cfa99bbc4fc740a4902e1aaefeb43c083859f066059b55781c289545217ab6 src/core/transports.mjs
101
+ 203876df917cf5d6bd054468b7457379d09a0038017a29360f83d6b04c179648 src/index.mjs
@@ -22,7 +22,7 @@ capabilities 允许:
22
22
 
23
23
  通用路由默认不继承父模型的推理配置,不添加服务端会话 ID、上次 response_id、厂商缓存提示。OpenAI 官方 origin 会启用它自己的缓存 key/retention 及 store=false、background=false 字段;第三方需要声明能力。原生 Responses wire 快照已有的缓存字段原样保留,store/background 被强制关闭;持久 conversation 写入始终禁止。声明 storage=false 表示不发送该字段,不能保证第三方服务端不留存数据。
24
24
 
25
- Chat 的 request.maxTokenField 可选 max_tokens(默认)或 max_completion_tokens;统一用 request.maxTokens 设值。Messages 的 thinking 必须显式启用 reasoning 并配置 request.thinking。未知能力字段、非法鉴权、密钥内嵌均在读取配置时拒绝。鉴权只解析当前配置的 credential.env / secretRef,不复用其他端点凭证;不跟随 HTTP 重定向。设置“重新加载配置”会刷新价格绑定与配置作用域。
25
+ Chat 的 request.maxTokenField 可选 max_tokens(默认)或 max_completion_tokens;统一用 request.maxTokens 设值。Beta 2 由预测策略按模型选择最低思考控制并覆盖旧 request.thinking/reasoning 高档;未知别名用 reasoningControl 声明能力。未知能力字段、非法鉴权、密钥内嵌均在读取配置时拒绝。鉴权只解析当前配置的 credential.env / secretRef,不复用其他端点凭证;不跟随 HTTP 重定向。设置“重新加载配置”会刷新价格绑定与配置作用域。
26
26
 
27
27
  ## 上下文与缓存
28
28
  - 保留父快照的当前上下文;翻译后的历史、工具定义和顺序确定,预测指令只追加尾部。
@@ -61,3 +61,20 @@ pricing.versions 填真实生效日期、币种和每百万 token 单价 input /
61
61
  - https://developers.openai.com/api/reference/resources/responses/methods/create
62
62
  - https://platform.claude.com/docs/en/api/messages
63
63
  - https://platform.claude.com/docs/en/build-with-claude/prompt-caching
64
+
65
+ ### 预测思考默认策略
66
+
67
+ Beta 2 的预测与校验修复不再继承主会话的 High:能关闭就关闭,不能关闭则使用模型实际支持的最低档。主会话设置不变。原生 DSH 接入读取适配器按最低到最高排列的精确模型能力;若适配器没有公布思考控制,则不发送虚构参数,该路径无法保证服务器关闭思考。
68
+
69
+ - DeepSeek 原生 Messages:`reasoningEffort: "off"`,实际发送 `thinking: {"type":"disabled"}`。DeepSeek 官方 Responses 使用 `reasoning.effort: "none"`,Chat 使用 `thinking.type: "disabled"`。省略参数不等于关闭。
70
+ - OpenAI Responses/Chat:按模型选择 none/minimal/最低有效档。例如 GPT-5.1/5.5 为 none,GPT-5 为 minimal,GPT-6 Astra/6.1 Sol 为 low,GPT-5.2 Pro 为 medium,GPT-5 Pro 仅支持 high,后者无法强制关闭。
71
+ - Anthropic:支持关闭的型号使用 disabled;必须思考的型号使用 adaptive + low。Sonnet 5.5 使用 between_tools + low,不能声称完全零思考。不用旧版 budget_tokens 冒充关闭思考。
72
+
73
+ 第三方模型别名需要在配置文件声明实际能力,例如 Responses/Chat:`"reasoningControl":{"efforts":["none","low","high"]}`;Messages:`"reasoningControl":{"thinkingModes":["disabled","adaptive"],"efforts":["low","high"]}`。明确配置 `capabilities.reasoning:false` 表示该路由不接受思考控制,因此不发送这些字段。未知能力在本地拦截,不做收费探测,也不自动换协议。错误的能力声明仍可能被服务商拒绝。
74
+
75
+ 思考配置属于请求条件。第一次从 High 切换到关闭/最低档可能损失父请求的缓存命中;之后稳定复用预测请求可建立自己的缓存,以实际 usage 为准。Beta 2 的本地结果缓存与旧 High 预测隔离。
76
+
77
+ 核对日期 2026-11-17:[DeepSeek](https://api-docs.deepseek.com/guides/thinking_mode)、[OpenAI](https://developers.openai.com/api/docs/guides/reasoning)、[Anthropic](https://platform.claude.com/docs/en/build-with-claude/thinking)。
78
+
79
+
80
+ Beta 3:在设置 → 下一条消息预测 → 后端与缓存中可修改“预测思考策略”。默认“关闭/最低”;选择“按后端配置”并保存后,使用父请求或所选后端配置文件的思考参数。修复请求沿用该策略。切换配置不会自动为当前轮再次付费生成;如需重做,手动触发。
package/README.md CHANGED
@@ -4,10 +4,10 @@
4
4
 
5
5
  An unofficial DeepSeek Harness plugin that suggests the user's next message after a normal assistant completion. Press **Tab** to accept an editable draft, then decide whether to send it.
6
6
 
7
- **Version: 0.8.0-beta.1 · public npm preview**
7
+ **Version: 0.8.0-beta.3 · public npm preview**
8
8
  **Package: @jhckevin/dsh-composer-prediction**
9
9
 
10
- ![Composer preview](https://unpkg.com/@jhckevin/dsh-composer-prediction@0.8.0-beta.1/assets/screenshots/composer-en.png)
10
+ ![Composer preview](https://unpkg.com/@jhckevin/dsh-composer-prediction@0.8.0-beta.3/assets/screenshots/composer-en.png)
11
11
 
12
12
  *Feature images show actual plugin components with demonstration data, not private conversations or live bills.*
13
13
 
@@ -63,13 +63,13 @@ New chats are enabled by default. Optional catch-up uses the last **user send**,
63
63
 
64
64
  ## Cache reuse and accounting
65
65
 
66
- The official DeepSeek parent path preserves effective context, tools, conditioning and affinity, appending prediction instructions at the tail. Existing parent compaction is inherited. Context windows and repair alternatives compare costs using available evidence; no fixed cache-hit or savings guarantee is made.
66
+ The official DeepSeek parent path preserves effective history, tools and affinity; prediction-only reasoning uses the minimum supported control, appending prediction instructions at the tail. Existing parent compaction is inherited. Context windows and repair alternatives compare costs using available evidence; no fixed cache-hit or savings guarantee is made.
67
67
 
68
68
  Mapping checks follow sends, normal completion, turn/token growth and structural events, not second-based polling. Estimated retained prefix below 50% after an external edit pauses that chat.
69
69
 
70
70
  **Prediction share = Prediction / (main + Prediction).**
71
71
 
72
- ![Responsive accounting preview](https://unpkg.com/@jhckevin/dsh-composer-prediction@0.8.0-beta.1/assets/screenshots/responsive-costs-420-en.png)
72
+ ![Responsive accounting preview](https://unpkg.com/@jhckevin/dsh-composer-prediction@0.8.0-beta.3/assets/screenshots/responsive-costs-420-en.png)
73
73
 
74
74
  Open the horizontal Prediction costs tab. Amounts use the same thread, currency and coverage interval, including every retry and repair attempt. Quotes are captured at request start; configured time-zone-aware peak/off-peak schedules apply. Third-party endpoints require their own prices.
75
75
 
@@ -79,7 +79,7 @@ Complete usage is priced by its components without charging reasoning twice. Mis
79
79
 
80
80
  Open **Settings → Composer Prediction** for behavior, backend profiles, context budget, prefix threshold, length preference, default 999-token visible-text limit, language checks, truncation, retries, catch-up, exclusions, user instructions and accounting.
81
81
 
82
- ![Settings preview](https://unpkg.com/@jhckevin/dsh-composer-prediction@0.8.0-beta.1/assets/screenshots/settings-en.png)
82
+ ![Settings preview](https://unpkg.com/@jhckevin/dsh-composer-prediction@0.8.0-beta.3/assets/screenshots/settings-en.png)
83
83
 
84
84
  Controls reuse native DSH primitives. UI text and details follow the active English/Chinese host locale. Prediction language is configured independently.
85
85
 
@@ -102,3 +102,20 @@ This is not an official DeepSeek/OpenAI product. Generic Responses is not Codex'
102
102
  Built and tested remotely. Synthetic HTTP tests and browser component images do not certify all vendors or replace native desktop acceptance. Release records distinguish static checks, controlled tests, public registry checks and native UI verification.
103
103
 
104
104
  See SOURCE.json, THIRD_PARTY_NOTICES.md, licenses and LICENSE. Maintainer: npm account jhckevin. Release history: CHANGELOG.md.
105
+
106
+ ### Prediction reasoning defaults
107
+
108
+ Beta 2 does not inherit the main conversation's High effort. Prediction and validation repair disable thinking when supported, otherwise use the lowest advertised effort. The main conversation is unchanged. Native DSH adapters provide ordered exact-model metadata; adapters with no reasoning knob receive no fabricated effort parameter (server-side thinking cannot be guaranteed off on such a route).
109
+
110
+ - DeepSeek native Messages: `reasoningEffort: "off"` → `thinking: {"type":"disabled"}`. Official DeepSeek Responses: `reasoning.effort: "none"`; Chat: `thinking.type: "disabled"`.
111
+ - OpenAI Responses/Chat: model-specific `none`, `minimal`, or the lowest accepted effort. For example GPT-5.1/5.5 use none; GPT-5 uses minimal; GPT-6 Astra/6.1 Sol use low; GPT-5.2 Pro uses medium; GPT-5 Pro only supports high. These latter models cannot be forced off.
112
+ - Anthropic: supported models use disabled; models requiring adaptive thinking use adaptive + low. Sonnet 5.5 uses between_tools + low, which is not a promise of zero reasoning. Legacy budget mode is never enabled to emulate off.
113
+
114
+ For unknown third-party aliases, declare actual backend capabilities in the profile, for example `"reasoningControl":{"efforts":["none","low","high"]}` for Responses/Chat or `"reasoningControl":{"thinkingModes":["disabled","adaptive"],"efforts":["low","high"]}` for Messages. Explicit `capabilities.reasoning:false` means the route has no control, so no thinking fields are sent. Missing knowledge stops locally before dispatch; no paid parameter probes or automatic protocol downgrades are performed. Incorrect declarations may still be rejected by the provider.
115
+
116
+ Reasoning settings change request conditioning. The first off/minimum request can lose the parent's High cache; stable subsequent prediction requests can build their own cache. Actual usage is authoritative; no cache-hit guarantee is made. Beta 2 partitions local results from older High predictions.
117
+
118
+ Sources checked 2026-11-17: [DeepSeek](https://api-docs.deepseek.com/guides/thinking_mode), [OpenAI](https://developers.openai.com/api/docs/guides/reasoning), [Anthropic](https://platform.claude.com/docs/en/build-with-claude/thinking).
119
+
120
+
121
+ Beta 3: Settings → Composer predictions → Backend & cache exposes Prediction reasoning. Off / minimum is the default. Explicitly select Use backend configuration and save to use parent request or profile reasoning controls, including custom values. Repairs follow that policy. Changing it does not automatically regenerate the current turn; retry manually if needed.
package/README.zh-CN.md CHANGED
@@ -4,7 +4,7 @@
4
4
 
5
5
  Composer Prediction 是 DeepSeek Harness 的非官方插件。助手正常完成回复后,它预测用户可能发送的下一条消息,显示在空输入框内;按 **Tab** 接受为可编辑草稿,由你决定是否发送。
6
6
 
7
- **版本:0.8.0-beta.1 · npm 公开预览版**
7
+ **版本:0.8.0-beta.3 · npm 公开预览版**
8
8
  **包名:@jhckevin/dsh-composer-prediction**
9
9
 
10
10
  ![输入框预测示意](./assets/screenshots/composer-zh.png)
@@ -114,3 +114,20 @@ dsh plugin --profile desktop add @jhckevin/dsh-composer-prediction@latest
114
114
  本版在远程服务器构建并通过自动回归和包内容检查;测试中的合成 HTTP 响应及浏览器组件图不能冒充所有服务商实测或完整原生桌面验收。发布验证记录区分静态、受控测试、registry 实测与本机 UI。
115
115
 
116
116
  源码与版权归属见 SOURCE.json、THIRD_PARTY_NOTICES.md、licenses 和 LICENSE。维护者:npm 账号 jhckevin。版本记录见 CHANGELOG.md。
117
+
118
+ ### 预测思考默认策略
119
+
120
+ Beta 2 的预测与校验修复不再继承主会话的 High:能关闭就关闭,不能关闭则使用模型实际支持的最低档。主会话设置不变。原生 DSH 接入读取适配器按最低到最高排列的精确模型能力;若适配器没有公布思考控制,则不发送虚构参数,该路径无法保证服务器关闭思考。
121
+
122
+ - DeepSeek 原生 Messages:`reasoningEffort: "off"`,实际发送 `thinking: {"type":"disabled"}`。DeepSeek 官方 Responses 使用 `reasoning.effort: "none"`,Chat 使用 `thinking.type: "disabled"`。省略参数不等于关闭。
123
+ - OpenAI Responses/Chat:按模型选择 none/minimal/最低有效档。例如 GPT-5.1/5.5 为 none,GPT-5 为 minimal,GPT-6 Astra/6.1 Sol 为 low,GPT-5.2 Pro 为 medium,GPT-5 Pro 仅支持 high,后者无法强制关闭。
124
+ - Anthropic:支持关闭的型号使用 disabled;必须思考的型号使用 adaptive + low。Sonnet 5.5 使用 between_tools + low,不能声称完全零思考。不用旧版 budget_tokens 冒充关闭思考。
125
+
126
+ 第三方模型别名需要在配置文件声明实际能力,例如 Responses/Chat:`"reasoningControl":{"efforts":["none","low","high"]}`;Messages:`"reasoningControl":{"thinkingModes":["disabled","adaptive"],"efforts":["low","high"]}`。明确配置 `capabilities.reasoning:false` 表示该路由不接受思考控制,因此不发送这些字段。未知能力在本地拦截,不做收费探测,也不自动换协议。错误的能力声明仍可能被服务商拒绝。
127
+
128
+ 思考配置属于请求条件。第一次从 High 切换到关闭/最低档可能损失父请求的缓存命中;之后稳定复用预测请求可建立自己的缓存,以实际 usage 为准。Beta 2 的本地结果缓存与旧 High 预测隔离。
129
+
130
+ 核对日期 2026-11-17:[DeepSeek](https://api-docs.deepseek.com/guides/thinking_mode)、[OpenAI](https://developers.openai.com/api/docs/guides/reasoning)、[Anthropic](https://platform.claude.com/docs/en/build-with-claude/thinking)。
131
+
132
+
133
+ Beta 3:在设置 → 下一条消息预测 → 后端与缓存中可修改“预测思考策略”。默认“关闭/最低”;选择“按后端配置”并保存后,使用父请求或所选后端配置文件的思考参数。修复请求沿用该策略。切换配置不会自动为当前轮再次付费生成;如需重做,手动触发。
package/SOURCE.json CHANGED
@@ -12,37 +12,40 @@
12
12
  "licenseSources": {
13
13
  "js-tiktoken": "https://raw.githubusercontent.com/dqbd/tiktoken/main/LICENSE"
14
14
  },
15
- "version": "0.8.0-beta.1",
15
+ "version": "0.8.0-beta.3",
16
16
  "codexReferenceCommit": "d2f76324729d7c4a2b3225cfe42c7dbf162d98b7",
17
17
  "codexRemotePromptIncluded": false,
18
18
  "builtOn": "remote desktop-research-wsl2; no local development or compilation",
19
19
  "remoteNodeVersion": "24.22.2",
20
20
  "artifacts": {
21
21
  "lib/dsh-native.mjs.map": "1ee244c7672499bc3913bf6e9d1c92fe69af895eb52c8ca188b73c13da57da17",
22
- "lib/client.js": "ea9844299ed6d7f12769997e32788e552e1cf4f453c4cfcd6a1f72c2eefd5a76",
22
+ "lib/client.js": "59cf34a520c7e05a3367864de20ddf36349e6ed9f9af63bd746c9c31e8a3577f",
23
23
  "lib/tokenizer-runtime.mjs": "6f3181637842da225049f78d0f9534c2aae98f71a22ef4d924fc604a899c620a",
24
24
  "lib/dsh-native.mjs": "c43240a988bb12e7dbf749fa9a53ec37a3060bdbdcc65ebebff248f672db6d07",
25
25
  "lib/schema.mjs.map": "27f2ae8243eae4e473cebb83f00718b3b934851c290f8418331841078b18de19",
26
- "lib/client.js.map": "d84b9255410a67df4a430031add39a7433baa2c3137f7848c35b01deec265d6c",
26
+ "lib/client.js.map": "90150157625cb054f1d4778bbbd6b94bba02cbd674a0fa06ceea4a925be02a18",
27
27
  "lib/schema.mjs": "d884840a5871239ce3e105de464c35014de7dcbb8f5793aa213b8588cb74de0d"
28
28
  },
29
29
  "verification": {
30
- "remoteTests": 274,
31
- "passed": 274,
30
+ "remoteTests": 312,
31
+ "passed": 312,
32
32
  "failed": 0,
33
33
  "desktopTarget": "0.2.2-rc.1",
34
- "browserComponents": "actual React and pinned official DSH primitives; demo data",
34
+ "browserComponents": "beta3 actual React/native primitives with demo data",
35
35
  "screenshots": 16,
36
36
  "browserGeometryVerified": true,
37
37
  "fullNativeDesktopAcceptance": "not claimed; fixed 64-item acceptance denominator unchanged",
38
- "npmPublication": "prepared public npm preview; registry verification recorded separately",
38
+ "npmPublication": "not yet published; verify registry before installation",
39
39
  "modelNetworkInRegressionTests": "mocked",
40
40
  "nativeFlowBrowserWidths": [
41
41
  1160,
42
42
  540
43
43
  ],
44
44
  "versionAlgorithms": 5,
45
- "browserResponsiveChecks": 20
45
+ "browserResponsiveChecks": 20,
46
+ "minimumReasoningTests": 33,
47
+ "nativeHostMinimumIntegration": true,
48
+ "paidModelCalls": 0
46
49
  },
47
50
  "lifecycle": {
48
51
  "generation": "user-submit then normal assistant completion",
@@ -82,5 +85,16 @@
82
85
  "runtimeComponent": "composer-prediction",
83
86
  "previousLocalPackage": "@local/dsh-composer-prediction",
84
87
  "dataDirectoryUnchanged": true
88
+ },
89
+ "reasoningPolicy": {
90
+ "version": "minimum-v1",
91
+ "default": "disabled-if-supported-otherwise-model-minimum",
92
+ "parentConfigurationUnchanged": true,
93
+ "unknownPortableAliases": "pre-dispatch-declaration-required",
94
+ "capabilityProbes": "none",
95
+ "nativeNoControl": "omit-unadvertised-effort; server thinking unknown",
96
+ "configurable": true,
97
+ "setting": "reasoningMode: minimum(default) or configured",
98
+ "minimumCacheCompatibility": "beta2 results reused"
85
99
  }
86
100
  }
package/docs/README.en.md CHANGED
@@ -4,7 +4,7 @@
4
4
 
5
5
  An unofficial DeepSeek Harness plugin that suggests the user's next message after a normal assistant completion. Press **Tab** to accept an editable draft, then decide whether to send it.
6
6
 
7
- **Version: 0.8.0-beta.1 · public npm preview**
7
+ **Version: 0.8.0-beta.3 · public npm preview**
8
8
  **Package: @jhckevin/dsh-composer-prediction**
9
9
 
10
10
  ![Composer preview](../assets/screenshots/composer-en.png)
@@ -63,7 +63,7 @@ New chats are enabled by default. Optional catch-up uses the last **user send**,
63
63
 
64
64
  ## Cache reuse and accounting
65
65
 
66
- The official DeepSeek parent path preserves effective context, tools, conditioning and affinity, appending prediction instructions at the tail. Existing parent compaction is inherited. Context windows and repair alternatives compare costs using available evidence; no fixed cache-hit or savings guarantee is made.
66
+ The official DeepSeek parent path preserves effective history, tools and affinity; prediction-only reasoning uses the minimum supported control, appending prediction instructions at the tail. Existing parent compaction is inherited. Context windows and repair alternatives compare costs using available evidence; no fixed cache-hit or savings guarantee is made.
67
67
 
68
68
  Mapping checks follow sends, normal completion, turn/token growth and structural events, not second-based polling. Estimated retained prefix below 50% after an external edit pauses that chat.
69
69
 
@@ -102,3 +102,20 @@ This is not an official DeepSeek/OpenAI product. Generic Responses is not Codex'
102
102
  Built and tested remotely. Synthetic HTTP tests and browser component images do not certify all vendors or replace native desktop acceptance. Release records distinguish static checks, controlled tests, public registry checks and native UI verification.
103
103
 
104
104
  See SOURCE.json, THIRD_PARTY_NOTICES.md, licenses and LICENSE. Maintainer: npm account jhckevin. Release history: CHANGELOG.md.
105
+
106
+ ### Prediction reasoning defaults
107
+
108
+ Beta 2 does not inherit the main conversation's High effort. Prediction and validation repair disable thinking when supported, otherwise use the lowest advertised effort. The main conversation is unchanged. Native DSH adapters provide ordered exact-model metadata; adapters with no reasoning knob receive no fabricated effort parameter (server-side thinking cannot be guaranteed off on such a route).
109
+
110
+ - DeepSeek native Messages: `reasoningEffort: "off"` → `thinking: {"type":"disabled"}`. Official DeepSeek Responses: `reasoning.effort: "none"`; Chat: `thinking.type: "disabled"`.
111
+ - OpenAI Responses/Chat: model-specific `none`, `minimal`, or the lowest accepted effort. For example GPT-5.1/5.5 use none; GPT-5 uses minimal; GPT-6 Astra/6.1 Sol use low; GPT-5.2 Pro uses medium; GPT-5 Pro only supports high. These latter models cannot be forced off.
112
+ - Anthropic: supported models use disabled; models requiring adaptive thinking use adaptive + low. Sonnet 5.5 uses between_tools + low, which is not a promise of zero reasoning. Legacy budget mode is never enabled to emulate off.
113
+
114
+ For unknown third-party aliases, declare actual backend capabilities in the profile, for example `"reasoningControl":{"efforts":["none","low","high"]}` for Responses/Chat or `"reasoningControl":{"thinkingModes":["disabled","adaptive"],"efforts":["low","high"]}` for Messages. Explicit `capabilities.reasoning:false` means the route has no control, so no thinking fields are sent. Missing knowledge stops locally before dispatch; no paid parameter probes or automatic protocol downgrades are performed. Incorrect declarations may still be rejected by the provider.
115
+
116
+ Reasoning settings change request conditioning. The first off/minimum request can lose the parent's High cache; stable subsequent prediction requests can build their own cache. Actual usage is authoritative; no cache-hit guarantee is made. Beta 2 partitions local results from older High predictions.
117
+
118
+ Sources checked 2026-11-17: [DeepSeek](https://api-docs.deepseek.com/guides/thinking_mode), [OpenAI](https://developers.openai.com/api/docs/guides/reasoning), [Anthropic](https://platform.claude.com/docs/en/build-with-claude/thinking).
119
+
120
+
121
+ Beta 3: Settings → Composer predictions → Backend & cache exposes Prediction reasoning. Off / minimum is the default. Explicitly select Use backend configuration and save to use parent request or profile reasoning controls, including custom values. Repairs follow that policy. Changing it does not automatically regenerate the current turn; retry manually if needed.
@@ -4,7 +4,7 @@
4
4
 
5
5
  Composer Prediction 是 DeepSeek Harness 的非官方插件。助手正常完成回复后,它预测用户可能发送的下一条消息,显示在空输入框内;按 **Tab** 接受为可编辑草稿,由你决定是否发送。
6
6
 
7
- **版本:0.8.0-beta.1 · npm 公开预览版**
7
+ **版本:0.8.0-beta.3 · npm 公开预览版**
8
8
  **包名:@jhckevin/dsh-composer-prediction**
9
9
 
10
10
  ![输入框预测示意](../assets/screenshots/composer-zh.png)
@@ -114,3 +114,20 @@ dsh plugin --profile desktop add @jhckevin/dsh-composer-prediction@latest
114
114
  本版在远程服务器构建并通过自动回归和包内容检查;测试中的合成 HTTP 响应及浏览器组件图不能冒充所有服务商实测或完整原生桌面验收。发布验证记录区分静态、受控测试、registry 实测与本机 UI。
115
115
 
116
116
  源码与版权归属见 SOURCE.json、THIRD_PARTY_NOTICES.md、licenses 和 LICENSE。维护者:npm 账号 jhckevin。版本记录见 CHANGELOG.md。
117
+
118
+ ### 预测思考默认策略
119
+
120
+ Beta 2 的预测与校验修复不再继承主会话的 High:能关闭就关闭,不能关闭则使用模型实际支持的最低档。主会话设置不变。原生 DSH 接入读取适配器按最低到最高排列的精确模型能力;若适配器没有公布思考控制,则不发送虚构参数,该路径无法保证服务器关闭思考。
121
+
122
+ - DeepSeek 原生 Messages:`reasoningEffort: "off"`,实际发送 `thinking: {"type":"disabled"}`。DeepSeek 官方 Responses 使用 `reasoning.effort: "none"`,Chat 使用 `thinking.type: "disabled"`。省略参数不等于关闭。
123
+ - OpenAI Responses/Chat:按模型选择 none/minimal/最低有效档。例如 GPT-5.1/5.5 为 none,GPT-5 为 minimal,GPT-6 Astra/6.1 Sol 为 low,GPT-5.2 Pro 为 medium,GPT-5 Pro 仅支持 high,后者无法强制关闭。
124
+ - Anthropic:支持关闭的型号使用 disabled;必须思考的型号使用 adaptive + low。Sonnet 5.5 使用 between_tools + low,不能声称完全零思考。不用旧版 budget_tokens 冒充关闭思考。
125
+
126
+ 第三方模型别名需要在配置文件声明实际能力,例如 Responses/Chat:`"reasoningControl":{"efforts":["none","low","high"]}`;Messages:`"reasoningControl":{"thinkingModes":["disabled","adaptive"],"efforts":["low","high"]}`。明确配置 `capabilities.reasoning:false` 表示该路由不接受思考控制,因此不发送这些字段。未知能力在本地拦截,不做收费探测,也不自动换协议。错误的能力声明仍可能被服务商拒绝。
127
+
128
+ 思考配置属于请求条件。第一次从 High 切换到关闭/最低档可能损失父请求的缓存命中;之后稳定复用预测请求可建立自己的缓存,以实际 usage 为准。Beta 2 的本地结果缓存与旧 High 预测隔离。
129
+
130
+ 核对日期 2026-11-17:[DeepSeek](https://api-docs.deepseek.com/guides/thinking_mode)、[OpenAI](https://developers.openai.com/api/docs/guides/reasoning)、[Anthropic](https://platform.claude.com/docs/en/build-with-claude/thinking)。
131
+
132
+
133
+ Beta 3:在设置 → 下一条消息预测 → 后端与缓存中可修改“预测思考策略”。默认“关闭/最低”;选择“按后端配置”并保存后,使用父请求或所选后端配置文件的思考参数。修复请求沿用该策略。切换配置不会自动为当前轮再次付费生成;如需重做,手动触发。
@@ -9,7 +9,9 @@
9
9
  "credential": {
10
10
  "secretRef": "DEEPSEEK_API_KEY"
11
11
  },
12
- "request": {},
12
+ "request": {
13
+ "reasoningEffort": "off"
14
+ },
13
15
  "cache": {},
14
16
  "pricing": {
15
17
  "schemaVersion": 1,
@@ -9,7 +9,9 @@
9
9
  "credential": {
10
10
  "secretRef": "DEEPSEEK_API_KEY"
11
11
  },
12
- "request": {},
12
+ "request": {
13
+ "reasoningEffort": "off"
14
+ },
13
15
  "cache": {},
14
16
  "pricing": {
15
17
  "schemaVersion": 1,