dsh-agent-sdk-bridge 1.1.7 → 1.1.8

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/DEVELOPMENT.md CHANGED
@@ -45,6 +45,19 @@
45
45
  仍会丢弃任何非宿主前缀的 tool_use,防止模型臆造工具名。CodeBuddy 会把 MCP server 名里的
46
46
  非字母数字字符规范成 `_`(`dsh-host` → `mcp__dsh_host__<tool>`),Qoder CLI 保留连字符,
47
47
  因此 `isHostTool()` / `hostToolName()` 同时接受两种拼写。
48
+ - **切换模型读缓存**:`resolveModel` 与 `capture()` 走 `cachedModels()`
49
+ (`src/model-cache.ts`):有快照就先返回快照,刷新在后台跑;只有从未读到过目录时才等待。
50
+ 三家的目录获取都很贵(Qoder 冷启动 CLI,最长等于 fetch 超时;CodeBuddy 要启进程;MiniMax
51
+ 走 ACP 探测),而宿主在每次切模型、切思考档位时都会重新解析一次,所以以前 TTL 一过期就会
52
+ 卡住菜单。三家 catalog 都新增了 `snapshot()`(只读内存,不发网络)供这里使用。
53
+ 同步探测(`regionBlocked`、额度)不受影响,它们仍走各自的正常读取。
54
+ - **默认思考档位取最强的一档**:`reasoningInfo()` 的第二个参数是**用户配置**的档位,不再是上游
55
+ 声明的 `defaultEffort`;留空时由 `strongestEffort()` 按 `none < minimal < low < medium <
56
+ high < xhigh < max` 取该模型支持的最强档(三家词表不同:Qoder `low|medium|high|max`、
57
+ MiniMax `low|high`、CodeBuddy 由目录声明)。宿主在请求未带档位时会应用我们声明的
58
+ `defaultEffort`(`requested ?? reasoning.defaultEffort`),所以这里选什么,实际就跑什么;
59
+ 上游自己的默认值会偏低(CodeBuddy 的 `high` 在可选 `max` 时就不再直接采用)。配置的档位若
60
+ 模型不支持,会退回最强档,而不是把这一轮判成 `UNSUPPORTED_REASONING_EFFORT`。
48
61
  - **CodeBuddy 模型选择按区域**:`codebuddyGlobalDisabledModels` /
49
62
  `codebuddyCnDisabledModels` 与 `codebuddyGlobalModelContextWindows` /
50
63
  `codebuddyCnModelContextWindows` 是四个独立字段,`CodeBuddyAdapter` 按 region 读取,
package/FAQ.md CHANGED
@@ -101,6 +101,17 @@ SQLite 依赖就在这些脚本里,普通 `npm i -g` 装出来的 `mcode` 一
101
101
  时使用该区域自己的登录态;要换站请先在原站点点「退出登录」。两个区域的模型开关与上下文窗口
102
102
  仍然分别保存,切回已登录的区域会直接恢复选择。
103
103
 
104
+ **切模型时默认选哪一档思考强度?** 选该模型支持的最强档(有 `max` 就选 `max`,否则取它声明的
105
+ 最高档;只有 `low` / `high` 的模型取 `high`)。以前沿用上游自己的默认值,而它常常偏低:
106
+ CodeBuddy 的 `GLM-5.3` 在可选 `max` 时仍声明默认 `high`。在高级设置里填了
107
+ `defaultReasoningEffort` 时以它为准;该档位模型不支持时退回最强档。
108
+
109
+ **切换模型或思考档位时会卡一下?** 宿主每次切换都会重新解析一次模型元数据(档位、上下文窗口、
110
+ 读图能力),而这三个 provider 的模型目录都要拉起各自的 CLI 或 ACP 进程才能读到。以前这次解析会
111
+ 等待目录刷新,缓存一过期就要等一次完整读取(Qoder 冷启动最久)。现在解析先用上一次成功读到的
112
+ 目录立刻返回,刷新在后台进行;只有从未读到过目录的第一次解析才会等待。目录自身的 TTL 与失败
113
+ 回退规则不变。
114
+
104
115
  **切到 MiniMax 国际站后提示「本机 IP 不在服务区域」?** 这不是登录过期,而是网络位置被拒。
105
116
  MiniMax 的边缘节点在收到跨区请求时会返回 `307`,把 `.io` 重定向到 `.minimaxi.com`(反向亦然)。
106
117
  两个域名是不同 origin,浏览器/Node 跟随重定向时 `Authorization` 头会被丢弃,于是账号接口变成空的
package/lib/index.js CHANGED
@@ -264,8 +264,8 @@ function readLive(value, fallback) {
264
264
  * Resolve one {@link Live} option that may legitimately be undefined.
265
265
  *
266
266
  * Separate from {@link readLive} because an option whose "unset" state is
267
- * itself meaningful (`defaultReasoningEffort`: empty means "follow upstream")
268
- * has no fallback to apply.
267
+ * itself meaningful (`defaultReasoningEffort`: empty means "use the model's
268
+ * strongest level") has no fallback to apply.
269
269
  *
270
270
  * @param value - the plain value or thunk, possibly undefined.
271
271
  * @returns the option's current value, or undefined when it is unset.
@@ -668,6 +668,19 @@ var QoderModelCatalog = class {
668
668
  return officialOnly((this.cached ?? sharedSnapshots.get(this.brand))?.models ?? []);
669
669
  }
670
670
  }
671
+ /**
672
+ * The last successful live list, without fetching.
673
+ *
674
+ * The model picker resolves a route on every switch, and a switch must not
675
+ * wait on a CLI cold start (up to the fetch timeout). Callers that need an
676
+ * immediate answer read this and let {@link liveModels} refresh behind them;
677
+ * `undefined` means nothing has been read yet, which only the first switch
678
+ * has to wait for.
679
+ */
680
+ snapshot() {
681
+ const cached = this.cached ?? sharedSnapshots.get(this.brand);
682
+ return cached === void 0 ? void 0 : officialOnly(cached.models);
683
+ }
671
684
  /** Why the last refresh failed, or undefined when it succeeded. */
672
685
  get error() {
673
686
  return this.failure;
@@ -765,6 +778,41 @@ function withTimeout$2(promise, ms) {
765
778
  });
766
779
  }
767
780
  //#endregion
781
+ //#region src/model-cache.ts
782
+ /**
783
+ * Cache-first catalogue reads for the model picker.
784
+ *
785
+ * The harness resolves a route every time the user switches model or reasoning
786
+ * level, and each of the three providers answers that resolution from a
787
+ * catalogue whose fetch is expensive: Qoder cold-starts its CLI (up to the
788
+ * fetch timeout), CodeBuddy boots a process, MiniMax probes over ACP. Awaiting
789
+ * that fetch inside a resolution made the picker freeze for seconds whenever
790
+ * the cache had just expired.
791
+ *
792
+ * So a resolution reads the last successful snapshot when there is one and lets
793
+ * the catalogue refresh behind it; only the very first resolution, where there
794
+ * is nothing to show, waits. A snapshot that is merely stale still describes
795
+ * the same account's models, so answering from it is no less correct than the
796
+ * previous behaviour of answering from a failed fetch's fallback.
797
+ *
798
+ * @module dsh-agent-sdk-bridge/model-cache
799
+ */
800
+ /**
801
+ * The cached list when one exists, otherwise the awaited fetch.
802
+ *
803
+ * @param snapshot - last successful catalogue, or undefined before the first.
804
+ * @param fetch - the catalogue's normal (TTL-aware, in-flight-sharing) read.
805
+ * @returns the entries to resolve against, never empty-by-surprise when a
806
+ * snapshot exists.
807
+ */
808
+ async function cachedModels(snapshot, fetch) {
809
+ if (snapshot !== void 0 && snapshot.length > 0) {
810
+ fetch().catch(() => void 0);
811
+ return snapshot;
812
+ }
813
+ return fetch();
814
+ }
815
+ //#endregion
768
816
  //#region src/tokens.ts
769
817
  /**
770
818
  * Script-aware text pricing, shared by the input and output estimators.
@@ -2869,22 +2917,72 @@ const EFFORT_NAMES = {
2869
2917
  max: "Max"
2870
2918
  };
2871
2919
  /**
2920
+ * Strength order used to pick a model's starting level.
2921
+ *
2922
+ * The three providers spell their levels differently (Qoder publishes
2923
+ * `low|medium|high|max`, MiniMax `low|high`, CodeBuddy whatever its catalog
2924
+ * declares), so the ranking is by name rather than by position in any one
2925
+ * upstream list. An id outside this table ranks below every known one, which is
2926
+ * what several sources agreeing on an unfamiliar spelling should do.
2927
+ */
2928
+ const EFFORT_ORDER = [
2929
+ "none",
2930
+ "minimal",
2931
+ "low",
2932
+ "medium",
2933
+ "high",
2934
+ "xhigh",
2935
+ "max"
2936
+ ];
2937
+ /**
2938
+ * The strongest level a model offers.
2939
+ *
2940
+ * Switching model or level must start at the most capable setting the route
2941
+ * accepts, the same rule the sibling adapter plugin uses; the previous
2942
+ * behaviour of following upstream's own default meant some models silently
2943
+ * started one or two steps down (`CodeBuddy` advertises `high` as its default
2944
+ * even when `max` is selectable).
2945
+ *
2946
+ * @param efforts - the levels the route accepts, in upstream order.
2947
+ * @returns the strongest id, or undefined for a non-reasoning model.
2948
+ */
2949
+ function strongestEffort(efforts) {
2950
+ if (efforts.length === 0) return void 0;
2951
+ let best;
2952
+ let bestRank = -1;
2953
+ for (const effort of efforts) {
2954
+ const rank = EFFORT_ORDER.indexOf(effort);
2955
+ if (best === void 0 || rank > bestRank) {
2956
+ best = effort;
2957
+ bestRank = rank;
2958
+ }
2959
+ }
2960
+ return best;
2961
+ }
2962
+ /**
2872
2963
  * Build the `reasoning` metadata block for a resolved model, or undefined
2873
2964
  * when the model does not support reasoning.
2965
+ *
2966
+ * The starting level is the strongest one the model accepts, not the level
2967
+ * upstream happens to name as its default: the user's explicit choice in the
2968
+ * sibling plugin, applied here too. An explicit configuration value still wins
2969
+ * when the model accepts it.
2970
+ *
2874
2971
  * @param efforts - CLI-reported effort ids (absent for reasoning-less models).
2875
- * @param defaultEffort - CLI-reported default effort id.
2972
+ * @param configuredEffort - user-configured starting level, when any.
2876
2973
  * @param isReasoning - whether the model supports reasoning per the CLI.
2877
2974
  * @returns the harness reasoning metadata, or undefined to omit it.
2878
2975
  */
2879
- function reasoningInfo(efforts, defaultEffort, isReasoning) {
2976
+ function reasoningInfo(efforts, configuredEffort, isReasoning) {
2880
2977
  if (efforts === void 0 && isReasoning === false) return void 0;
2881
2978
  const ids = efforts !== void 0 && efforts.length > 0 ? efforts : DEFAULT_REASONING_EFFORTS;
2979
+ const defaultEffort = configuredEffort !== void 0 && ids.includes(configuredEffort) ? configuredEffort : strongestEffort(ids);
2882
2980
  return {
2883
2981
  efforts: ids.map((id) => ({
2884
2982
  id: ReasoningEffortId(id),
2885
2983
  name: EFFORT_NAMES[id] ?? id
2886
2984
  })),
2887
- ...defaultEffort !== void 0 && ids.includes(defaultEffort) ? { defaultEffort: ReasoningEffortId(defaultEffort) } : {}
2985
+ ...defaultEffort === void 0 ? {} : { defaultEffort: ReasoningEffortId(defaultEffort) }
2888
2986
  };
2889
2987
  }
2890
2988
  /**
@@ -3078,11 +3176,11 @@ var QoderAdapter = class extends LlmAdapter {
3078
3176
  }
3079
3177
  async resolveModel(provider, model, _signal) {
3080
3178
  const target = resolveQoderModelId(model, this.modelAliases);
3081
- const live = (await this.catalog.liveModels()).find((entry) => entry.value === target);
3179
+ const live = (await cachedModels(this.catalog.snapshot(), () => this.catalog.liveModels())).find((entry) => entry.value === target);
3082
3180
  if (live !== void 0) {
3083
3181
  const vision = live.isVl === true;
3084
3182
  this.knownVision.set(target, vision);
3085
- const reasoning = reasoningInfo(live.efforts, live.defaultEffort ?? this.defaultReasoningEffort, live.isReasoning);
3183
+ const reasoning = reasoningInfo(live.efforts, this.defaultReasoningEffort, live.isReasoning);
3086
3184
  return {
3087
3185
  provider,
3088
3186
  id: model,
@@ -3132,7 +3230,7 @@ var QoderAdapter = class extends LlmAdapter {
3132
3230
  async capture(provider, model, signal) {
3133
3231
  const info = await this.resolveModel(provider, model, signal);
3134
3232
  const target = resolveQoderModelId(model, this.modelAliases);
3135
- const live = (await this.catalog.liveModels()).find((entry) => entry.value === target);
3233
+ const live = (await cachedModels(this.catalog.snapshot(), () => this.catalog.liveModels())).find((entry) => entry.value === target);
3136
3234
  const declared = info.context?.contextWindow;
3137
3235
  return {
3138
3236
  info,
@@ -3660,7 +3758,7 @@ var CodeBuddyCatalog = class {
3660
3758
  * context-budget choice) use this instead of awaiting a refresh.
3661
3759
  */
3662
3760
  snapshot(requested) {
3663
- return this.snapshots.get(requested ?? readLive(this.region, "global"))?.models ?? [];
3761
+ return this.snapshots.get(requested ?? readLive(this.region, "global"))?.models;
3664
3762
  }
3665
3763
  /** Models for one region; a failed refresh keeps the last good list. */
3666
3764
  async models(requested) {
@@ -4998,7 +5096,7 @@ var CodeBuddyAdapter = class extends LlmAdapter {
4998
5096
  async modelVision(model) {
4999
5097
  const target = resolveCodeBuddyModelId(model, this.aliases);
5000
5098
  const key = `${this.region}:${target}`;
5001
- const live = (await this.catalog.models().catch(() => [])).find((entry) => entry.id === target);
5099
+ const live = (await cachedModels(this.catalog.snapshot(), () => this.catalog.models())).find((entry) => entry.id === target);
5002
5100
  if (live !== void 0) {
5003
5101
  this.knownVision.set(key, live.vision);
5004
5102
  return live.vision;
@@ -5085,7 +5183,7 @@ var CodeBuddyAdapter = class extends LlmAdapter {
5085
5183
  }
5086
5184
  /** The catalog snapshot the adapter last read (empty before the first read). */
5087
5185
  models(region) {
5088
- return this.catalog.snapshot(region);
5186
+ return this.catalog.snapshot(region) ?? [];
5089
5187
  }
5090
5188
  /** Whether this model lets the user pick a context budget at all. */
5091
5189
  supportsContextWindowChoice(id, region = this.region) {
@@ -5102,10 +5200,10 @@ var CodeBuddyAdapter = class extends LlmAdapter {
5102
5200
  async resolveModel(provider, model) {
5103
5201
  const target = resolveCodeBuddyModelId(model, this.aliases);
5104
5202
  const key = `${this.region}:${target}`;
5105
- const live = (await this.catalog.models()).find((entry) => entry.id === target);
5203
+ const live = (await cachedModels(this.catalog.snapshot(), () => this.catalog.models())).find((entry) => entry.id === target);
5106
5204
  if (live !== void 0) {
5107
5205
  this.knownVision.set(key, live.vision);
5108
- const reasoning = reasoningInfo(live.reasoningEfforts, live.defaultEffort, (live.reasoningEfforts?.length ?? 0) > 0);
5206
+ const reasoning = reasoningInfo(live.reasoningEfforts, readLiveOptional(this.options.defaultReasoningEffort), (live.reasoningEfforts?.length ?? 0) > 0);
5109
5207
  return {
5110
5208
  provider,
5111
5209
  id: model,
@@ -6646,6 +6744,15 @@ var MinimaxCatalog = class {
6646
6744
  return cached?.models ?? [];
6647
6745
  }
6648
6746
  }
6747
+ /**
6748
+ * The last successful list for a region, without probing.
6749
+ *
6750
+ * The picker resolves a route on every switch; awaiting an ACP probe there
6751
+ * cost a process round-trip whenever the TTL had just expired.
6752
+ */
6753
+ snapshot(requested) {
6754
+ return this.snapshots.get(requested ?? readLive(this.region, "cn"))?.models;
6755
+ }
6649
6756
  /** Why the last refresh failed, or undefined when it succeeded. */
6650
6757
  error(requested) {
6651
6758
  return this.failures.get(requested ?? readLive(this.region, "cn"));
@@ -7828,10 +7935,10 @@ var MinimaxAdapter = class extends LlmAdapter {
7828
7935
  const region = this.region;
7829
7936
  const target = resolveMinimaxModelId(model, this.aliases);
7830
7937
  const key = `${region}:${target}`;
7831
- const live = (await this.catalog.models(region)).find((entry) => entry.id === target);
7938
+ const live = (await cachedModels(this.catalog.snapshot(region), () => this.catalog.models(region))).find((entry) => entry.id === target);
7832
7939
  if (live !== void 0) {
7833
7940
  this.knownVision.set(key, live.vision);
7834
- const reasoning = reasoningInfo(live.reasoningEfforts, live.defaultEffort, live.reasoning);
7941
+ const reasoning = reasoningInfo(live.reasoningEfforts, readLiveOptional(this.options.defaultReasoningEffort), live.reasoning);
7835
7942
  return {
7836
7943
  provider,
7837
7944
  id: model,
@@ -8575,7 +8682,7 @@ const Config$1 = z$1.object({
8575
8682
  maxSessions: z$1.number().step(1).min(1).max(64).default(8).description("同时保持的内层会话数上限,超出后按最久未使用关闭。每个会话是一个子进程。").volatile(),
8576
8683
  modelCacheTtlSeconds: z$1.number().step(1).min(10).max(86400).default(300).description("模型列表缓存时长(秒)。调小可更快看到账号侧变化。").volatile(),
8577
8684
  modelAliases: z$1.dict(z$1.string()).default({}).description("额外的模型别名映射:DSH 侧名称 → Qoder 模型 ID。内置 deepseek-v4-* 别名仍然生效。").volatile(),
8578
- defaultReasoningEffort: z$1.string().default("").description("请求未指定思考档位时使用的默认档位;留空则用上游默认。").volatile(),
8685
+ defaultReasoningEffort: z$1.string().default("").description("默认选用的思考档位;留空则用该模型支持的最强档位。").volatile(),
8579
8686
  defaultMaxTokens: z$1.number().step(1).min(1).default(32e3).description("模型未声明输出上限时使用的输出上限。").volatile(),
8580
8687
  toolResultTimeoutSeconds: z$1.number().step(1).min(5).max(3600).default(120).description("仅用于未配对 handler 的兜底时限(秒):正常工具调用会等宿主下一次请求携带结果,不按此超时。").volatile(),
8581
8688
  catalogFetchTimeoutSeconds: z$1.number().step(1).min(1).max(300).default(20).description("拉取模型列表的时限(秒),超时后回退到内置列表。").volatile(),
@@ -131,15 +131,34 @@ export declare function cliContextWindow(live: WindowFacts | undefined, declared
131
131
  * looked like 10%.
132
132
  */
133
133
  export declare function resolveDeclaredWindow(live: WindowFacts | undefined, override: number | undefined, remembered: number | undefined, fallback: number): number;
134
+ /**
135
+ * The strongest level a model offers.
136
+ *
137
+ * Switching model or level must start at the most capable setting the route
138
+ * accepts, the same rule the sibling adapter plugin uses; the previous
139
+ * behaviour of following upstream's own default meant some models silently
140
+ * started one or two steps down (`CodeBuddy` advertises `high` as its default
141
+ * even when `max` is selectable).
142
+ *
143
+ * @param efforts - the levels the route accepts, in upstream order.
144
+ * @returns the strongest id, or undefined for a non-reasoning model.
145
+ */
146
+ export declare function strongestEffort(efforts: readonly string[]): string | undefined;
134
147
  /**
135
148
  * Build the `reasoning` metadata block for a resolved model, or undefined
136
149
  * when the model does not support reasoning.
150
+ *
151
+ * The starting level is the strongest one the model accepts, not the level
152
+ * upstream happens to name as its default: the user's explicit choice in the
153
+ * sibling plugin, applied here too. An explicit configuration value still wins
154
+ * when the model accepts it.
155
+ *
137
156
  * @param efforts - CLI-reported effort ids (absent for reasoning-less models).
138
- * @param defaultEffort - CLI-reported default effort id.
157
+ * @param configuredEffort - user-configured starting level, when any.
139
158
  * @param isReasoning - whether the model supports reasoning per the CLI.
140
159
  * @returns the harness reasoning metadata, or undefined to omit it.
141
160
  */
142
- export declare function reasoningInfo(efforts: readonly string[] | undefined, defaultEffort: string | undefined, isReasoning: boolean | undefined): LlmModelReasoningInfo | undefined;
161
+ export declare function reasoningInfo(efforts: readonly string[] | undefined, configuredEffort: string | undefined, isReasoning: boolean | undefined): LlmModelReasoningInfo | undefined;
143
162
  /**
144
163
  * The Qoder-backed adapter. Session continuity, tool parking, and feed
145
164
  * planning live here; chunk synthesis lives in the session's consumer.
@@ -126,7 +126,7 @@ export declare class CodeBuddyCatalog {
126
126
  * Callers that must answer synchronously (deciding whether a model exposes a
127
127
  * context-budget choice) use this instead of awaiting a refresh.
128
128
  */
129
- snapshot(requested?: CodeBuddyRegion): readonly CodeBuddyModel[];
129
+ snapshot(requested?: CodeBuddyRegion): readonly CodeBuddyModel[] | undefined;
130
130
  /** Models for one region; a failed refresh keeps the last good list. */
131
131
  models(requested?: CodeBuddyRegion): Promise<readonly CodeBuddyModel[]>;
132
132
  /** Why the last refresh failed, or undefined when it succeeded. */
@@ -41,8 +41,8 @@ export declare function readLive<T>(value: Live<T> | undefined, fallback: T): T;
41
41
  * Resolve one {@link Live} option that may legitimately be undefined.
42
42
  *
43
43
  * Separate from {@link readLive} because an option whose "unset" state is
44
- * itself meaningful (`defaultReasoningEffort`: empty means "follow upstream")
45
- * has no fallback to apply.
44
+ * itself meaningful (`defaultReasoningEffort`: empty means "use the model's
45
+ * strongest level") has no fallback to apply.
46
46
  *
47
47
  * @param value - the plain value or thunk, possibly undefined.
48
48
  * @returns the option's current value, or undefined when it is unset.
@@ -345,6 +345,13 @@ export declare class MinimaxCatalog {
345
345
  constructor(region: Live<MinimaxRegion>, ttlMs?: Live<number>, timeoutMs?: Live<number>);
346
346
  /** Models for one region; a failed refresh keeps the last good list. */
347
347
  models(requested?: MinimaxRegion): Promise<readonly MinimaxModel[]>;
348
+ /**
349
+ * The last successful list for a region, without probing.
350
+ *
351
+ * The picker resolves a route on every switch; awaiting an ACP probe there
352
+ * cost a process round-trip whenever the TTL had just expired.
353
+ */
354
+ snapshot(requested?: MinimaxRegion): readonly MinimaxModel[] | undefined;
348
355
  /** Why the last refresh failed, or undefined when it succeeded. */
349
356
  error(requested?: MinimaxRegion): string | undefined;
350
357
  invalidate(region?: MinimaxRegion): void;
@@ -0,0 +1,27 @@
1
+ /**
2
+ * Cache-first catalogue reads for the model picker.
3
+ *
4
+ * The harness resolves a route every time the user switches model or reasoning
5
+ * level, and each of the three providers answers that resolution from a
6
+ * catalogue whose fetch is expensive: Qoder cold-starts its CLI (up to the
7
+ * fetch timeout), CodeBuddy boots a process, MiniMax probes over ACP. Awaiting
8
+ * that fetch inside a resolution made the picker freeze for seconds whenever
9
+ * the cache had just expired.
10
+ *
11
+ * So a resolution reads the last successful snapshot when there is one and lets
12
+ * the catalogue refresh behind it; only the very first resolution, where there
13
+ * is nothing to show, waits. A snapshot that is merely stale still describes
14
+ * the same account's models, so answering from it is no less correct than the
15
+ * previous behaviour of answering from a failed fetch's fallback.
16
+ *
17
+ * @module dsh-agent-sdk-bridge/model-cache
18
+ */
19
+ /**
20
+ * The cached list when one exists, otherwise the awaited fetch.
21
+ *
22
+ * @param snapshot - last successful catalogue, or undefined before the first.
23
+ * @param fetch - the catalogue's normal (TTL-aware, in-flight-sharing) read.
24
+ * @returns the entries to resolve against, never empty-by-surprise when a
25
+ * snapshot exists.
26
+ */
27
+ export declare function cachedModels<T>(snapshot: readonly T[] | undefined, fetch: () => Promise<readonly T[]>): Promise<readonly T[]>;
@@ -43,6 +43,16 @@ export declare class QoderModelCatalog {
43
43
  * settings page cannot disagree about what the account has.
44
44
  */
45
45
  liveModels(): Promise<readonly ModelInfo[]>;
46
+ /**
47
+ * The last successful live list, without fetching.
48
+ *
49
+ * The model picker resolves a route on every switch, and a switch must not
50
+ * wait on a CLI cold start (up to the fetch timeout). Callers that need an
51
+ * immediate answer read this and let {@link liveModels} refresh behind them;
52
+ * `undefined` means nothing has been read yet, which only the first switch
53
+ * has to wait for.
54
+ */
55
+ snapshot(): readonly ModelInfo[] | undefined;
46
56
  /** Why the last refresh failed, or undefined when it succeeded. */
47
57
  get error(): string | undefined;
48
58
  /**
package/package.json CHANGED
@@ -2,7 +2,7 @@
2
2
  "name": "dsh-agent-sdk-bridge",
3
3
  "displayName": "Agent SDK Bridge",
4
4
  "description": "DeepSeek Harness LLM adapter bridge for Qoder CN, Qoder International, Tencent CodeBuddy and MiniMax Code subscriptions. It reuses each provider's own SDK or official ACP server and CLI login, shows available quota, and exposes per-provider settings pages with independent international and mainland login regions.",
5
- "version": "1.1.7",
5
+ "version": "1.1.8",
6
6
  "type": "module",
7
7
  "license": "MIT",
8
8
  "author": "k1-h2",
package/src/adapter.ts CHANGED
@@ -23,6 +23,7 @@ import type { ResolvedConfig } from './config.ts'
23
23
  import type { Live } from './live.ts'
24
24
  import { readLive, readLiveOptional } from './live.ts'
25
25
  import { DEFAULT_MODEL_CACHE_TTL_MS, DEFAULT_FETCH_TIMEOUT_MS, QoderModelCatalog } from './models.ts'
26
+ import { cachedModels } from './model-cache.ts'
26
27
  import { feedImageRefs, feedToText, imageRefText, imageRefs, joinFeeds, renderInitialFeed, renderRefreshed, renderUserTurn } from './render.ts'
27
28
  import type { Feed, ImagePathResolver } from './render.ts'
28
29
  import { QoderSession, QoderSessionManager, DEFAULT_MAX_SESSIONS } from './session.ts'
@@ -216,31 +217,77 @@ const EFFORT_NAMES: Readonly<Record<string, string>> = {
216
217
  max: 'Max',
217
218
  }
218
219
 
220
+ /**
221
+ * Strength order used to pick a model's starting level.
222
+ *
223
+ * The three providers spell their levels differently (Qoder publishes
224
+ * `low|medium|high|max`, MiniMax `low|high`, CodeBuddy whatever its catalog
225
+ * declares), so the ranking is by name rather than by position in any one
226
+ * upstream list. An id outside this table ranks below every known one, which is
227
+ * what several sources agreeing on an unfamiliar spelling should do.
228
+ */
229
+ const EFFORT_ORDER: readonly string[] = ['none', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max']
230
+
231
+ /**
232
+ * The strongest level a model offers.
233
+ *
234
+ * Switching model or level must start at the most capable setting the route
235
+ * accepts, the same rule the sibling adapter plugin uses; the previous
236
+ * behaviour of following upstream's own default meant some models silently
237
+ * started one or two steps down (`CodeBuddy` advertises `high` as its default
238
+ * even when `max` is selectable).
239
+ *
240
+ * @param efforts - the levels the route accepts, in upstream order.
241
+ * @returns the strongest id, or undefined for a non-reasoning model.
242
+ */
243
+ export function strongestEffort(efforts: readonly string[]): string | undefined {
244
+ if (efforts.length === 0) return undefined
245
+ let best: string | undefined
246
+ let bestRank = -1
247
+ for (const effort of efforts) {
248
+ // First occurrence wins a tie, which keeps an upstream-ordered list stable
249
+ // when it repeats a level.
250
+ const rank = EFFORT_ORDER.indexOf(effort)
251
+ if (best === undefined || rank > bestRank) {
252
+ best = effort
253
+ bestRank = rank
254
+ }
255
+ }
256
+ return best
257
+ }
258
+
219
259
  /**
220
260
  * Build the `reasoning` metadata block for a resolved model, or undefined
221
261
  * when the model does not support reasoning.
262
+ *
263
+ * The starting level is the strongest one the model accepts, not the level
264
+ * upstream happens to name as its default: the user's explicit choice in the
265
+ * sibling plugin, applied here too. An explicit configuration value still wins
266
+ * when the model accepts it.
267
+ *
222
268
  * @param efforts - CLI-reported effort ids (absent for reasoning-less models).
223
- * @param defaultEffort - CLI-reported default effort id.
269
+ * @param configuredEffort - user-configured starting level, when any.
224
270
  * @param isReasoning - whether the model supports reasoning per the CLI.
225
271
  * @returns the harness reasoning metadata, or undefined to omit it.
226
272
  */
227
273
  export function reasoningInfo(
228
274
  efforts: readonly string[] | undefined,
229
- defaultEffort: string | undefined,
275
+ configuredEffort: string | undefined,
230
276
  isReasoning: boolean | undefined,
231
277
  ): LlmModelReasoningInfo | undefined {
232
278
  if (efforts === undefined && isReasoning === false) return undefined
233
279
  const ids = efforts !== undefined && efforts.length > 0
234
280
  ? efforts
235
281
  : DEFAULT_REASONING_EFFORTS
282
+ const defaultEffort = configuredEffort !== undefined && ids.includes(configuredEffort)
283
+ ? configuredEffort
284
+ : strongestEffort(ids)
236
285
  return {
237
286
  efforts: ids.map(id => ({
238
287
  id: ReasoningEffortId(id),
239
288
  name: EFFORT_NAMES[id] ?? id,
240
289
  })),
241
- ...defaultEffort !== undefined && ids.includes(defaultEffort)
242
- ? { defaultEffort: ReasoningEffortId(defaultEffort) }
243
- : {},
290
+ ...defaultEffort === undefined ? {} : { defaultEffort: ReasoningEffortId(defaultEffort) },
244
291
  }
245
292
  }
246
293
 
@@ -492,11 +539,16 @@ export class QoderAdapter extends LlmAdapter {
492
539
  // address to find the metadata, but echo the requested id back: the seam
493
540
  // rejects an exact-model result whose id differs from the one asked for.
494
541
  const target = resolveQoderModelId(model, this.modelAliases)
495
- const live = (await this.catalog.liveModels()).find(entry => entry.value === target)
542
+ // The picker resolves on every switch and every level change. Awaiting the
543
+ // CLI catalogue here (a cold start, up to the fetch timeout) froze the menu
544
+ // whenever the cache had just expired, so a snapshot answers immediately and
545
+ // the refresh runs behind it.
546
+ const live = (await cachedModels(this.catalog.snapshot(), () => this.catalog.liveModels()))
547
+ .find(entry => entry.value === target)
496
548
  if (live !== undefined) {
497
549
  const vision = live.isVl === true
498
550
  this.knownVision.set(target, vision)
499
- const reasoning = reasoningInfo(live.efforts, live.defaultEffort ?? this.defaultReasoningEffort, live.isReasoning)
551
+ const reasoning = reasoningInfo(live.efforts, this.defaultReasoningEffort, live.isReasoning)
500
552
  return {
501
553
  provider,
502
554
  id: model,
@@ -566,7 +618,10 @@ export class QoderAdapter extends LlmAdapter {
566
618
  private async capture(provider: string, model: string, signal?: AbortSignal): Promise<CallGeneration> {
567
619
  const info = await this.resolveModel(provider, model, signal)
568
620
  const target = resolveQoderModelId(model, this.modelAliases)
569
- const live = (await this.catalog.liveModels()).find(entry => entry.value === target)
621
+ // Same cache-first rule as the resolution above: a dispatch may not stall
622
+ // on a catalogue fetch to learn window facts a snapshot already carries.
623
+ const live = (await cachedModels(this.catalog.snapshot(), () => this.catalog.liveModels()))
624
+ .find(entry => entry.value === target)
570
625
  const declared = info.context?.contextWindow
571
626
  return {
572
627
  info,
@@ -18,6 +18,7 @@ import type { CodeBuddyRegion } from './brand.ts'
18
18
  import { codeBuddyColdStream, CodeBuddyAccountReader, CodeBuddyAuth, CodeBuddyCatalog, CodeBuddySessionManager, CODEBUDDY_DEFAULT_CONTEXT_WINDOW, CODEBUDDY_DEFAULT_MAX_TOKENS, CODEBUDDY_CACHE_TTL_MS, CODEBUDDY_FETCH_TIMEOUT_MS } from './codebuddy.ts'
19
19
  import type { CodeBuddyModel, CodeBuddyUsageSnapshot } from './codebuddy.ts'
20
20
  import { reasoningInfo, planContinuation } from './adapter.ts'
21
+ import { cachedModels } from './model-cache.ts'
21
22
  import type { ImageRequestReader } from './adapter.ts'
22
23
  import type { Live } from './live.ts'
23
24
  import { readLive, readLiveOptional } from './live.ts'
@@ -199,7 +200,8 @@ export class CodeBuddyAdapter extends LlmAdapter {
199
200
  private async modelVision(model: string): Promise<boolean> {
200
201
  const target = resolveCodeBuddyModelId(model, this.aliases)
201
202
  const key = `${this.region}:${target}`
202
- const live = (await this.catalog.models().catch(() => [])).find(entry => entry.id === target)
203
+ const live = (await cachedModels(this.catalog.snapshot(), () => this.catalog.models()))
204
+ .find(entry => entry.id === target)
203
205
  if (live !== undefined) {
204
206
  this.knownVision.set(key, live.vision)
205
207
  return live.vision
@@ -301,7 +303,7 @@ export class CodeBuddyAdapter extends LlmAdapter {
301
303
 
302
304
  /** The catalog snapshot the adapter last read (empty before the first read). */
303
305
  private models(region: CodeBuddyRegion): readonly CodeBuddyModel[] {
304
- return this.catalog.snapshot(region)
306
+ return this.catalog.snapshot(region) ?? []
305
307
  }
306
308
 
307
309
  /** Whether this model lets the user pick a context budget at all. */
@@ -321,10 +323,13 @@ export class CodeBuddyAdapter extends LlmAdapter {
321
323
  override async resolveModel(provider: string, model: string): Promise<LlmResolvedModelInfo> {
322
324
  const target = resolveCodeBuddyModelId(model, this.aliases)
323
325
  const key = `${this.region}:${target}`
324
- const live = (await this.catalog.models()).find(entry => entry.id === target)
326
+ // Cache first: a model switch must not wait on a catalogue fetch (which
327
+ // boots a CodeBuddy process) just because the TTL has expired.
328
+ const live = (await cachedModels(this.catalog.snapshot(), () => this.catalog.models()))
329
+ .find(entry => entry.id === target)
325
330
  if (live !== undefined) {
326
331
  this.knownVision.set(key, live.vision)
327
- const reasoning = reasoningInfo(live.reasoningEfforts, live.defaultEffort, (live.reasoningEfforts?.length ?? 0) > 0)
332
+ const reasoning = reasoningInfo(live.reasoningEfforts, readLiveOptional(this.options.defaultReasoningEffort), (live.reasoningEfforts?.length ?? 0) > 0)
328
333
  return {
329
334
  provider,
330
335
  id: model,
package/src/codebuddy.ts CHANGED
@@ -352,8 +352,8 @@ export class CodeBuddyCatalog {
352
352
  * Callers that must answer synchronously (deciding whether a model exposes a
353
353
  * context-budget choice) use this instead of awaiting a refresh.
354
354
  */
355
- snapshot(requested?: CodeBuddyRegion): readonly CodeBuddyModel[] {
356
- return this.snapshots.get(requested ?? readLive(this.region, 'global'))?.models ?? []
355
+ snapshot(requested?: CodeBuddyRegion): readonly CodeBuddyModel[] | undefined {
356
+ return this.snapshots.get(requested ?? readLive(this.region, 'global'))?.models
357
357
  }
358
358
 
359
359
  /** Models for one region; a failed refresh keeps the last good list. */
package/src/config.ts CHANGED
@@ -247,7 +247,7 @@ export const Config = z.object({
247
247
  maxSessions: z.number().step(1).min(1).max(64).default(8).description('同时保持的内层会话数上限,超出后按最久未使用关闭。每个会话是一个子进程。').volatile(),
248
248
  modelCacheTtlSeconds: z.number().step(1).min(10).max(86_400).default(300).description('模型列表缓存时长(秒)。调小可更快看到账号侧变化。').volatile(),
249
249
  modelAliases: z.dict(z.string()).default({}).description('额外的模型别名映射:DSH 侧名称 → Qoder 模型 ID。内置 deepseek-v4-* 别名仍然生效。').volatile(),
250
- defaultReasoningEffort: z.string().default('').description('请求未指定思考档位时使用的默认档位;留空则用上游默认。').volatile(),
250
+ defaultReasoningEffort: z.string().default('').description('默认选用的思考档位;留空则用该模型支持的最强档位。').volatile(),
251
251
  defaultMaxTokens: z.number().step(1).min(1).default(32_000).description('模型未声明输出上限时使用的输出上限。').volatile(),
252
252
  toolResultTimeoutSeconds: z.number().step(1).min(5).max(3_600).default(120).description('仅用于未配对 handler 的兜底时限(秒):正常工具调用会等宿主下一次请求携带结果,不按此超时。').volatile(),
253
253
  catalogFetchTimeoutSeconds: z.number().step(1).min(1).max(300).default(20).description('拉取模型列表的时限(秒),超时后回退到内置列表。').volatile(),
package/src/live.ts CHANGED
@@ -47,8 +47,8 @@ export function readLive<T>(value: Live<T> | undefined, fallback: T): T {
47
47
  * Resolve one {@link Live} option that may legitimately be undefined.
48
48
  *
49
49
  * Separate from {@link readLive} because an option whose "unset" state is
50
- * itself meaningful (`defaultReasoningEffort`: empty means "follow upstream")
51
- * has no fallback to apply.
50
+ * itself meaningful (`defaultReasoningEffort`: empty means "use the model's
51
+ * strongest level") has no fallback to apply.
52
52
  *
53
53
  * @param value - the plain value or thunk, possibly undefined.
54
54
  * @returns the option's current value, or undefined when it is unset.
@@ -25,6 +25,7 @@ import {
25
25
  import type { MinimaxModel, MinimaxPromptBlock, MinimaxUsageSnapshot } from './minimax.ts'
26
26
  import { resolveMinimaxCli } from './minimax-cli.ts'
27
27
  import { reasoningInfo, planContinuation } from './adapter.ts'
28
+ import { cachedModels } from './model-cache.ts'
28
29
  import type { ImageRequestReader } from './adapter.ts'
29
30
  import type { Live } from './live.ts'
30
31
  import { readLive, readLiveOptional } from './live.ts'
@@ -255,10 +256,12 @@ export class MinimaxAdapter extends LlmAdapter {
255
256
  const region = this.region
256
257
  const target = resolveMinimaxModelId(model, this.aliases)
257
258
  const key = `${region}:${target}`
258
- const live = (await this.catalog.models(region)).find(entry => entry.id === target)
259
+ // Cache first: resolving a route may not wait on an ACP probe.
260
+ const live = (await cachedModels(this.catalog.snapshot(region), () => this.catalog.models(region)))
261
+ .find(entry => entry.id === target)
259
262
  if (live !== undefined) {
260
263
  this.knownVision.set(key, live.vision)
261
- const reasoning = reasoningInfo(live.reasoningEfforts, live.defaultEffort, live.reasoning)
264
+ const reasoning = reasoningInfo(live.reasoningEfforts, readLiveOptional(this.options.defaultReasoningEffort), live.reasoning)
262
265
  return {
263
266
  provider,
264
267
  id: model,
package/src/minimax.ts CHANGED
@@ -1527,6 +1527,16 @@ export class MinimaxCatalog {
1527
1527
  }
1528
1528
  }
1529
1529
 
1530
+ /**
1531
+ * The last successful list for a region, without probing.
1532
+ *
1533
+ * The picker resolves a route on every switch; awaiting an ACP probe there
1534
+ * cost a process round-trip whenever the TTL had just expired.
1535
+ */
1536
+ snapshot(requested?: MinimaxRegion): readonly MinimaxModel[] | undefined {
1537
+ return this.snapshots.get(requested ?? readLive(this.region, 'cn'))?.models
1538
+ }
1539
+
1530
1540
  /** Why the last refresh failed, or undefined when it succeeded. */
1531
1541
  error(requested?: MinimaxRegion): string | undefined {
1532
1542
  return this.failures.get(requested ?? readLive(this.region, 'cn'))
@@ -0,0 +1,39 @@
1
+ /**
2
+ * Cache-first catalogue reads for the model picker.
3
+ *
4
+ * The harness resolves a route every time the user switches model or reasoning
5
+ * level, and each of the three providers answers that resolution from a
6
+ * catalogue whose fetch is expensive: Qoder cold-starts its CLI (up to the
7
+ * fetch timeout), CodeBuddy boots a process, MiniMax probes over ACP. Awaiting
8
+ * that fetch inside a resolution made the picker freeze for seconds whenever
9
+ * the cache had just expired.
10
+ *
11
+ * So a resolution reads the last successful snapshot when there is one and lets
12
+ * the catalogue refresh behind it; only the very first resolution, where there
13
+ * is nothing to show, waits. A snapshot that is merely stale still describes
14
+ * the same account's models, so answering from it is no less correct than the
15
+ * previous behaviour of answering from a failed fetch's fallback.
16
+ *
17
+ * @module dsh-agent-sdk-bridge/model-cache
18
+ */
19
+
20
+ /**
21
+ * The cached list when one exists, otherwise the awaited fetch.
22
+ *
23
+ * @param snapshot - last successful catalogue, or undefined before the first.
24
+ * @param fetch - the catalogue's normal (TTL-aware, in-flight-sharing) read.
25
+ * @returns the entries to resolve against, never empty-by-surprise when a
26
+ * snapshot exists.
27
+ */
28
+ export async function cachedModels<T>(
29
+ snapshot: readonly T[] | undefined,
30
+ fetch: () => Promise<readonly T[]>,
31
+ ): Promise<readonly T[]> {
32
+ if (snapshot !== undefined && snapshot.length > 0) {
33
+ // Kick the refresh but do not wait for it: the caller needs an answer now,
34
+ // and a failure here changes nothing because the snapshot stays published.
35
+ void fetch().catch(() => undefined)
36
+ return snapshot
37
+ }
38
+ return fetch()
39
+ }
package/src/models.ts CHANGED
@@ -101,6 +101,20 @@ export class QoderModelCatalog {
101
101
  }
102
102
  }
103
103
 
104
+ /**
105
+ * The last successful live list, without fetching.
106
+ *
107
+ * The model picker resolves a route on every switch, and a switch must not
108
+ * wait on a CLI cold start (up to the fetch timeout). Callers that need an
109
+ * immediate answer read this and let {@link liveModels} refresh behind them;
110
+ * `undefined` means nothing has been read yet, which only the first switch
111
+ * has to wait for.
112
+ */
113
+ snapshot(): readonly ModelInfo[] | undefined {
114
+ const cached = this.cached ?? sharedSnapshots.get(this.brand)
115
+ return cached === undefined ? undefined : officialOnly(cached.models)
116
+ }
117
+
104
118
  /** Why the last refresh failed, or undefined when it succeeded. */
105
119
  get error(): string | undefined {
106
120
  return this.failure