@mars-sea/dsh-commandcode-provider 0.13.2 → 0.13.4

package/lib/index.js

@@ -103,7 +103,7 @@ const PLAN_ORDER = Object.fromEntries(PLAN_TIER_ORDER.map((tier, index) => [tier
 * subscription-plan labels, deals, and hourly (peak/off-peak) pricing.
 *
 * Everything here is synced from official sources — the command-code CLI
-* bundle's model table (`dist/cli.mjs`, re-verified at command-code@1.82.1) and
+* bundle's model table (`dist/cli.mjs`, re-verified at command-code@1.82.5) and
 * the official plan/pricing/model docs; see the dsh-commandcode-upstream skill
 * for the extraction procedures. Keeping the snapshot in its own module
 * confines those frequent sync diffs here: src/adapter.ts holds only the stable
@@ -138,6 +138,11 @@ const KNOWN_EFFORTS = {
 		"medium",
 		"xhigh"
 	],
+	"Qwen/Qwen3.8-Max-Prime": [
+		"low",
+		"medium",
+		"xhigh"
+	],
 	"claude-fable-5-1": [
 		"low",
 		"medium",
@@ -391,6 +396,16 @@ const KNOWN_EFFORTS = {
 		"high",
 		"xhigh"
 	],
+	"meta/muse-spark-open": [
+		"low",
+		"medium",
+		"high"
+	],
+	"nvidia/nemotron-3.5-super-vl": [
+		"low",
+		"medium",
+		"high"
+	],
 	"gpt-6-astra": [
 		"low",
 		"medium",
@@ -431,7 +446,7 @@ const KNOWN_EFFORTS = {
 		"xhigh",
 		"max"
 	],
-	"inclusionai/ling-3.1-flash:free": [
+	"inclusionai/ling-3.1-flash": [
 		"low",
 		"medium",
 		"high"
@@ -453,6 +468,13 @@ const KNOWN_EFFORTS = {
 		"medium",
 		"high",
 		"xhigh"
+	],
+	"stealth/cat-jam:free": [
+		"low",
+		"medium",
+		"high",
+		"xhigh",
+		"max"
 	]
 };
 /**
@@ -480,6 +502,7 @@ const KNOWN_IMAGE_MODELS = /* @__PURE__ */ new Set([
 	"Qwen/Qwen3.8-Max",
 	"Qwen/Qwen3.8-Max-0902",
 	"Qwen/Qwen3.8-Omni-Flash",
+	"Qwen/Qwen3.8-Max-Prime",
 	"claude-fable-5-1",
 	"claude-fable-5",
 	"claude-haiku-4-5-20251001",
@@ -518,6 +541,8 @@ const KNOWN_IMAGE_MODELS = /* @__PURE__ */ new Set([
 	"meta/muse-spark-1.3",
 	"meta/muse-spark-1.3-contributor",
 	"meta/muse-wonder-1.0",
+	"meta/muse-spark-open",
+	"nvidia/nemotron-3.5-super-vl",
 	"mistral/mistral-large-4",
 	"moonshotai/Kimi-K2.5",
 	"moonshotai/Kimi-K2.6",
@@ -540,14 +565,16 @@ const KNOWN_IMAGE_MODELS = /* @__PURE__ */ new Set([
 	"z-ai/glm-5.3-flashx"
 ]);
 /**
-* Models WITHOUT a zero-data-retention upstream, per the official CLI's own
-* registry (`command-code@1.82.1` `dist/cli.mjs`): `modelSupportsZdr(id)` is
-* exactly `!nonZdrSet.has(canonicalize(id))`, and `knownModelSupportsZdr`
-* carries the same membership in the sibling route table — the UNION of both
-* is this set. Reading only the sibling route table dropped
-* `meituan/LongCat-2.0`, which sat in `modelSupportsZdr` alone through
-* 1.73.0; the 1.74.3 table lists it in NEITHER set, so as of this snapshot it
-* is no longer excluded. The official docs (commandcode.ai/docs/resources/
+* Models WITHOUT a zero-data-retention upstream. Through command-code@1.82.2
+* the CLI's own registry answered this: `modelSupportsZdr(id)` was exactly
+* `!nonZdrSet.has(canonicalize(id))`, and `knownModelSupportsZdr` carried the
+* membership in the sibling route table — the UNION of both was this set, and
+* the two anchors disagreed by one id (`meituan/LongCat-2.0:free`, which 1.74.1
+* dropped from the first anchor while the canonical helper kept it).
+* command-code@1.82.4 removes that disagreement by making the verdict a catalog
+* lookup — `resolveKnownCatalogModel({model,catalog})?.zdr ?? false` — so this
+* set is now the live catalog's `zdr:false` membership and nothing more. The
+* official docs (commandcode.ai/docs/resources/
 * zdr) put coverage in prose — "99% of our models have ZDR-capable upstreams … only
 * a small handful of models are affected" — so the CLI's exclusion list is
 * the only per-model evidence there is; a ZDR request naming one of these
@@ -560,10 +587,19 @@ const KNOWN_IMAGE_MODELS = /* @__PURE__ */ new Set([
 * switch is on — the provider remains the routing authority and refuses an
 * unsupported model rather than silently dropping the privacy guarantee.
 *
-* `minimax/minimax-m3-free` is the one entry the public catalog
-* (`/provider/v1/models`) does not serve (it is CLI/pricing-visible and hidden
-* from the picker); it stays listed because the CLI carries it and a Go-plan
-* request can still name it.
+* `minimax/minimax-m3-free` is the one entry the public catalog's default view
+* (`/provider/v1/models`) does not serve through the picker (it is CLI/pricing-visible
+* and hidden from the picker); it stays listed because the CLI carries it and a Go-plan
+* request can still name it. Both this row and `stealth/cat-jam:free` ARE served
+* by the `?view=client` catalog, which is where the 1.82.4 verdict comes from.
+*
+* command-code@1.82.4 moved the authority for this verdict: `knownModelSupportsZdr`
+* no longer negates a bundled Set, it returns `resolveKnownCatalogModel({model,
+* catalog})?.zdr ?? false`, so an id the catalog does not carry (or carries with
+* `zdr:false`) is NOT covered. That inverts the default this helper was built on
+* ("not listed" answered TRUE); the snapshot keeps its negative-set shape because
+* 99% of the catalog is still covered, but membership is now read from the
+* `?view=client` catalog and confirmed against the canonical helper set.
 *
 * Keep in sync via the dsh-commandcode-upstream skill: the CLI's registry data
 * (its `zdr:{only:[…]}` provider routes and the per-provider `zdr`/`noTraining`
@@ -584,6 +620,11 @@ const KNOWN_IMAGE_MODELS = /* @__PURE__ */ new Set([
 * `listSelectableModelIds()`, so they are pruned along with the rest of the
 * tables and `supportsZeroDataRetention` now answers TRUE for them, matching
 * what the current CLI says.
+*
+* 1.82.4 then reversed 1.74.1's removal of `meituan/LongCat-2.0:free` and added
+* the `stealth/cat-jam:free` preview, taking this set to 24 members — the size
+* of the live catalog's `zdr:false` list. The canonical helper set still carries
+* 23 because `cat-jam:free` postdates the artifact's embedded snapshot.
 */
 const KNOWN_NON_ZDR_MODELS = /* @__PURE__ */ new Set([
 	"MiniMaxAI/MiniMax-M3",
@@ -595,10 +636,12 @@ const KNOWN_NON_ZDR_MODELS = /* @__PURE__ */ new Set([
 	"meta/muse-spark-1.3",
 	"meta/muse-spark-1.3-contributor",
 	"meta/muse-wonder-1.0",
+	"meituan/LongCat-2.0:free",
 	"minimax/minimax-m3-free",
 	"poolside/laguna-s-2.1-free",
 	"sakana/fugu-ultra",
 	"stealth/glyph-cluster:free",
+	"stealth/cat-jam:free",
 	"stepfun/Step-3.7-Flash",
 	"stepfun/Step-5-Preview",
 	"xai/grok-4.5",
@@ -646,8 +689,7 @@ const KNOWN_THINKING_MODELS = /* @__PURE__ */ new Set([
 	"thinkingmachines/inkling",
 	"thinkingmachines/inkling-small",
 	"poolside/laguna-s-2.1-free",
-	"meituan/LongCat-2.0",
-	"inclusionai/ling-3.0-flash-sante:free"
+	"meituan/LongCat-2.0"
 ]);
 /**
 * Catalog models the Provider API serves ONLY through `/provider/v1/messages`
@@ -711,22 +753,21 @@ function requiresMessagesEndpoint(modelId) {
 * **Generated — do not hand-edit.** `scripts/sync-output-limits.mjs` rewrites
 * this literal from models.dev; run it after a catalog sync, and use its
 * `--check` in CI. Command Code resells these models through their VENDOR's own
-* API, so the vendor's published ceiling is the value recorded: 77 of the 89
-* catalog ids as of 2026-10-14. Each row carries a comment naming the entry it
+* API, so the vendor's published ceiling is the value recorded: 76 of the 89
+* catalog ids as of 2026-10-15. Each row carries a comment naming the entry it
 * came from.
 *
-* Two of those 77 rest on more than a vendor lookup, both hand-recorded in the
-* script's `MANUAL_CEILINGS`. `gpt-6.1-sol` (128000) comes from OpenAI's own row
-* and is corroborated by every other provider carrying it.
-* `inclusionai/ling-3.1-flash:free` (32768) has no models.dev row at all, so it
-* takes the Ling family's uniform ceiling; `inclusionai/ling-3.0-flash-sante:free`
-* (32768) lost both its vendor row and its unanimous third-party vote (vercel
-* files 32000), and is pinned so a regeneration cannot drop a live model's
-* ceiling back to the default. All are still correctable — a stricter gateway
-* costs one refused request, and `streamRequest()` remembers what the endpoint
-* said.
-*
-* The 12 unlisted models are ones no vendor publishes and the aggregators
+* No row needs a hand-recorded pin as of 2026-10-16: the one that did
+* (`inclusionai/ling-3.0-flash-sante:free`, pinned at 32768 because it had lost
+* both its vendor row and its unanimous third-party vote) left the live catalog
+* with the 1.82.4 catalog migration, so the script's `MANUAL_CEILINGS` map is
+* now empty and every remaining row resolves from a vendor row or a unanimous
+* third-party vote. (`gpt-6.1-sol` (128000) never needed a pin — OpenAI's own row
+* carries it and every other provider carrying the id agrees.) All are still
+* correctable — a stricter gateway costs one refused request, and
+* `streamRequest()` remembers what the endpoint said.
+*
+* The 13 unlisted models are ones no vendor publishes and the aggregators
 * disagree about. They are NOT a gap: a refusal states the ceiling in full, and
 * `streamRequest()` reads it off, retries once and remembers it for the model
 * (issue #71). A vendor figure the gateway turns out to be stricter than costs
@@ -771,8 +812,6 @@ const MODEL_OUTPUT_TOKEN_LIMITS = /* @__PURE__ */ new Map([
 	["gpt-6-luna", 128e3],
 	["gpt-6-sol", 128e3],
 	["gpt-6.1-sol", 128e3],
-	["inclusionai/ling-3.0-flash-sante:free", 32768],
-	["inclusionai/ling-3.1-flash:free", 32768],
 	["meituan/LongCat-2.0", 131072],
 	["meta/muse-spark-1.1", 131072],
 	["meta/muse-spark-1.2", 131072],
@@ -823,15 +862,20 @@ const MODEL_OUTPUT_TOKEN_LIMITS = /* @__PURE__ */ new Map([
 * `/docs/plans/max` and `/docs/resources/pricing-limits`). Each plan's model
 * list is a superset of the one below it: Go ⊂ GOAT ⊂ Pro ⊂ Provider/Max.
 * Models absent from every plan list (Claude Opus/Fable, Fugu Ultra) are
-* Provider-tier. Re-verified at command-code@1.82.1 (2026-10-14): 89 catalog
-* ids at 54/64/79/87/89 cumulative, a strict superset chain. Mistral Large 4
+* Provider-tier. Re-verified at command-code@1.82.4 (2026-10-16) against the
+* live `?view=client` catalog: 93 ids, still a strict superset chain. Mistral
+* Large 4
 * (1.75.0) was the only addition at the 1.77.0 check; the 1.78.0/1.79.0 pair
 * added Claude Haiku 5.5 at Pro and the free Glyph Cluster preview at Go.
 * 1.79.2 lowers Haiku 5.5's minimum plan to GOAT; 1.80.0 adds Muse Wonder at
-* Max; 1.82.0 adds Gemini 4 Argon at Pro and nothing else moves. Every other
+* Max; 1.82.0 adds Gemini 4 Argon at Pro and nothing else moves. The 1.82.4
+* train adds Qwen 3.8 Max Prime at GOAT plus Muse Spark Open, Nemotron 3.5
+* Super VL and the free Cat Jam preview at Go, and drops Ling 3.0 Flash Sante,
+* which left the live catalog. Every other
 * tier is unchanged.
 * The only removals since 1.49.0 are the two retired stealth previews (Pixel
-* Canary at 1.73.1, Space Bunny Alpha at 1.74.3). Per-entry tags below name the
+* Canary at 1.73.1, Space Bunny Alpha at 1.74.3) and Ling 3.0 Flash Sante
+* (2026-10-16). Per-entry tags below name the
 * releases that added or moved each row.
 *
 * The Provider API exposes no plan metadata, so this snapshot is the source of
@@ -855,6 +899,7 @@ const KNOWN_PLANS = {
 	"Qwen/Qwen3.8-Flash": "go",
 	"Qwen/Qwen3.8-Max": "go",
 	"Qwen/Qwen3.8-Max-0902": "go",
+	"Qwen/Qwen3.8-Max-Prime": "goat",
 	"Qwen/Qwen3.8-Omni-Flash": "go",
 	"deepseek/deepseek-v4-flash-fast": "go",
 	"deepseek/deepseek-v4.1-flash": "go",
@@ -865,10 +910,10 @@ const KNOWN_PLANS = {
 	"gpt-5.6-luna": "go",
 	"gpt-6-luna": "go",
 	"meituan/LongCat-2.0": "go",
-	"inclusionai/ling-3.0-flash-sante:free": "go",
-	"inclusionai/ling-3.1-flash:free": "go",
+	"inclusionai/ling-3.1-flash": "go",
 	"meta/muse-spark-1.2-contributor": "go",
 	"meta/muse-spark-1.3-contributor": "go",
+	"meta/muse-spark-open": "go",
 	"mistral/mistral-large-4": "go",
 	"moonshotai/Kimi-K2.5": "go",
 	"moonshotai/Kimi-K2.6": "go",
@@ -876,10 +921,12 @@ const KNOWN_PLANS = {
 	"moonshotai/Kimi-K2.7-Code-Highspeed": "go",
 	"moonshotai/Kimi-K3": "go",
 	"nvidia/nemotron-3-ultra-550b-a55b": "go",
+	"nvidia/nemotron-3.5-super-vl": "go",
 	"poolside/laguna-s-2.1-free": "go",
 	"stepfun/Step-3.5-Flash": "go",
 	"stepfun/Step-3.7-Flash": "go",
 	"stepfun/Step-5-Preview": "go",
+	"stealth/cat-jam:free": "go",
 	"stealth/glyph-cluster:free": "go",
 	"tencent/hy3-paid": "go",
 	"tencent/hy4-preview": "go",
@@ -1072,11 +1119,7 @@ const KNOWN_DEALS = {
 		label: "FREE",
 		free: true
 	},
-	"inclusionai/ling-3.0-flash-sante:free": {
-		label: "FREE",
-		free: true
-	},
-	"inclusionai/ling-3.1-flash:free": {
+	"stealth/cat-jam:free": {
 		label: "FREE",
 		free: true
 	},
@@ -1721,8 +1764,8 @@ const MODEL_PRICE_ROWS = [
 		],
 		allowance: {
 			go: 10,
-			goat: 60,
-			pro: 70
+			goat: 100,
+			pro: 100
 		}
 	},
 	{
@@ -1832,8 +1875,8 @@ const MODEL_PRICE_ROWS = [
 		],
 		allowance: {
 			go: 6,
-			goat: 70,
-			pro: 80
+			goat: 20,
+			pro: 30
 		},
 		contextTiers: [{
 			"maxContext": 272e3,
@@ -2144,6 +2187,19 @@ const MODEL_PRICE_ROWS = [
 			pro: 30
 		}
 	},
+	{
+		id: "ling-3.1-flash",
+		rates: [
+			.3,
+			.9,
+			.06
+		],
+		allowance: {
+			go: 6,
+			goat: 100,
+			pro: 110
+		}
+	},
 	{
 		id: "longcat-2.0",
 		rates: [
@@ -2354,6 +2410,19 @@ const MODEL_PRICE_ROWS = [
 			pro: 30
 		}
 	},
+	{
+		id: "muse-spark-open",
+		rates: [
+			1.25,
+			4.25,
+			.15
+		],
+		allowance: {
+			go: 6,
+			goat: 20,
+			pro: 30
+		}
+	},
 	{
 		id: "muse-wonder-1.0",
 		rates: [
@@ -2380,6 +2449,19 @@ const MODEL_PRICE_ROWS = [
 			pro: 30
 		}
 	},
+	{
+		id: "nemotron-3.5-super-vl",
+		rates: [
+			.2,
+			.5,
+			.04
+		],
+		allowance: {
+			go: 6,
+			goat: 20,
+			pro: 30
+		}
+	},
 	{
 		id: "qwen-3.6-max",
 		rates: [
@@ -2554,6 +2636,20 @@ const MODEL_PRICE_ROWS = [
 			pro: 30
 		}
 	},
+	{
+		id: "qwen-3.8-max-prime",
+		rates: [
+			4,
+			12,
+			.5,
+			5
+		],
+		allowance: {
+			go: 6,
+			goat: 20,
+			pro: 30
+		}
+	},
 	{
 		id: "qwen-3.8-omni-flash",
 		rates: [
@@ -2644,6 +2740,19 @@ const MODEL_PRICE_ROWS = [
 			goat: 20,
 			pro: 30
 		}
+	},
+	{
+		id: "typesafe/jev-1.14.0",
+		rates: [
+			.042,
+			0,
+			0
+		],
+		allowance: {
+			go: 6,
+			goat: 20,
+			pro: 30
+		}
 	}
 ];
 /**
@@ -3532,6 +3641,41 @@ function modelIsVisible(id, allowlist, overrides) {
 	const flag = overrides !== null && typeof overrides === "object" && !Array.isArray(overrides) ? overrides[id] : void 0;
 	return typeof flag === "boolean" ? flag : allowlist.length === 0 || allowlist.includes(id);
 }
+/** 自定义上下文长度接受的上界（token）；覆盖任何已知模型窗口仍留有余量。 */
+const CONTEXT_WINDOW_MAX = 1e7;
+/**
+* 判定一个值能否作为自定义上下文长度：边界内的安全整数，符合宿主模型元数据约束。
+*
+* 程序化构造的连接可能绕过 schemastery 归一化，所以 `resolveAdapterOptions`
+* 也要重新判定——越界或非数值直接丢弃，回落到目录公布值，而不是让请求带着
+* 一个网关会拒绝的窗口上线。
+*/
+function isValidContextWindow(value) {
+	return typeof value === "number" && Number.isSafeInteger(value) && value >= 1 && value <= 1e7;
+}
+/**
+* 只接收合法条目的按模型覆盖表；空表、非对象与非法数值一律视为未设置。
+* 配置校验会拒绝非整数或越界值；本函数额外保护绕过配置校验的程序化调用，
+* 丢弃坏条目，不把非法窗口交给宿主。返回新对象；未设置时为 `undefined`。
+*/
+function readContextWindowOverrides(raw) {
+	if (raw === null || typeof raw !== "object" || Array.isArray(raw)) return void 0;
+	const entries = Object.entries(raw).filter((entry) => entry[0] !== "" && isValidContextWindow(entry[1]));
+	return entries.length === 0 ? void 0 : Object.fromEntries(entries);
+}
+/**
+* 一次模型解析的有效上下文长度：按模型覆盖 → 全局值 → 目录公布值。
+*
+* 适配器只在有目录条目时调用（未知模型保持既有的「无上下文宣告」形状），
+* 所以 `catalog` 是已知的正数；按模型覆盖与全局值在 `resolveAdapterOptions`
+* 已清洗，这里对程序化构造的连接再判一次，坏值静默回落到目录值。
+*/
+function effectiveContextWindow(catalog, model, options) {
+	const perModel = options.contextWindowOverrides?.[model];
+	if (isValidContextWindow(perModel)) return perModel;
+	if (isValidContextWindow(options.contextWindow)) return options.contextWindow;
+	return catalog;
+}
 const THROTTLED_CODE = "THROTTLED";
 const boundedCodes = /* @__PURE__ */ new Set([
 	"EMPTY_RESPONSE",
@@ -5838,6 +5982,13 @@ function learnedOutputLimit(apiBase, model) {
 function modelOutputTokenLimit(apiBase, model) {
 	return Math.min(published.get(apiBase)?.get(model) ?? fallbackLimit(model), learnedOutputLimit(apiBase, model));
 }
+/**
+* 从本次目录记录的独立输出上限重新推导预算，不能复用已经被旧窗口压低的 maxTokens。
+* 公布值固定在该记录，不读取并发刷新后的共享目录；只继续采纳同网关更低的学习上限。
+*/
+function catalogModelMaxTokens(apiBase, model, contextWindow = model.contextWindow) {
+	return Math.min(contextWindow, model.publishedMaxTokens ?? fallbackLimit(model.id), learnedOutputLimit(apiBase, model.id));
+}
 /** 是否更新共享事实与是否值得重试是两个问题；以本次实发预算判断拒绝。 */
 function outputCeilingRefusal(apiBase, model, sent, error) {
 	if (!(error instanceof LlmError) || !OUTPUT_LIMIT_REJECTION.test(error.message)) return void 0;
@@ -5993,7 +6144,7 @@ var GatewayFacts = class {
 * per-request options thunk and an API-key resolver from the plugin entry
 * (src/index.ts), so a settings change reaches the very next request.
 */
-const COMMAND_CODE_CLI_VERSION = "1.82.1";
+const COMMAND_CODE_CLI_VERSION = "1.82.5";
 const DEFAULT_API_BASE = "https://api.commandcode.ai";
 /**
 * The output budget this bundle asks for, and the ceiling it will never exceed.
@@ -7656,7 +7807,7 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 				provider,
 				id: model.id,
 				name: `${model.name} (CC)`,
-				description: capabilityDescription(model.id, model.contextWindow),
+				description: capabilityDescription(model.id, effectiveContextWindow(model.contextWindow, model.id, connection)),
 				inputModalities: vision ? ["text", "image"] : ["text"]
 			};
 		};
@@ -7696,16 +7847,17 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 		const entry = this.gatewayFacts.knownModel(connection, model) ?? (await this.gatewayFacts.refresh(connection, signal)).find((m) => m.id === model);
 		const efforts = KNOWN_EFFORTS[model];
 		const vision = KNOWN_IMAGE_MODELS.has(model);
+		const contextWindow = entry === void 0 ? void 0 : effectiveContextWindow(entry.contextWindow, model, connection);
 		return {
 			provider,
 			id: model,
 			name: entry ? `${entry.name} (CC)` : model,
-			description: capabilityDescription(model, entry?.contextWindow),
+			description: capabilityDescription(model, contextWindow),
 			inputModalities: vision ? ["text", "image"] : ["text"],
 			toolUpdate: "addition-only",
-			...entry ? {
-				context: { contextWindow: entry.contextWindow },
-				defaultMaxTokens: Math.min(entry.maxTokens, DEFAULT_GENERATE_MAX_TOKENS)
+			...entry !== void 0 && contextWindow !== void 0 ? {
+				context: { contextWindow },
+				defaultMaxTokens: Math.min(catalogModelMaxTokens(connection.apiBase, entry, contextWindow), DEFAULT_GENERATE_MAX_TOKENS)
 			} : {},
 			...efforts ? { reasoning: { efforts: efforts.map((effort) => ({
 				id: ReasoningEffortId(effort),
@@ -8128,7 +8280,8 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 		const catalog = await facts.catalog;
 		options.signal?.throwIfAborted();
 		const modelEntry = catalog.find((row) => row.id === options.model);
-		const initialMax = Math.min(options.maxTokens ?? Infinity, modelEntry?.maxTokens ?? facts.fallbackLimit, DEFAULT_GENERATE_MAX_TOKENS);
+		const modelMax = modelEntry === void 0 ? facts.fallbackLimit : catalogModelMaxTokens(connection.apiBase, modelEntry, effectiveContextWindow(modelEntry.contextWindow, options.model, connection));
+		const initialMax = Math.min(options.maxTokens ?? Infinity, modelMax, DEFAULT_GENERATE_MAX_TOKENS);
 		const budget = {
 			max: initialMax,
 			sent: initialMax
@@ -8168,8 +8321,8 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 		}
 		options.signal?.throwIfAborted();
 		let maxTokens = Math.min(budget.max, learnedOutputLimit(connection.apiBase, options.model));
-		const effort = options.reasoningEffort;
 		const supported = KNOWN_EFFORTS[options.model];
+		const effort = options.purpose === "session-title" && supported?.includes("off") ? "off" : options.reasoningEffort;
 		const declared = effort !== void 0 && supported?.includes(effort) ? effort : void 0;
 		const reasoningEffort = effort === "off" && declared === void 0 ? void 0 : declared;
 		const systemText = [options.system ?? "", ...options.messages.filter((m) => m.role === "system").map((m) => m.content.map(blockText).filter(Boolean).join("\n"))].filter(Boolean).join("\n\n");
@@ -8427,12 +8580,33 @@ function formatMoney(value) {
 function formatMoneyExact(value) {
 	return `$${value.toFixed(4)}`;
 }
-function formatTokensCompact(value) {
-	if (value >= 1e9) return `${(value / 1e9).toFixed(1)}B`;
-	if (value >= 1e6) return `${(value / 1e6).toFixed(1)}M`;
-	if (value >= 1e3) return `${(value / 1e3).toFixed(1)}K`;
+/**
+* 紧凑数字：K/M/B 三级，一切档位决策的唯一实现。
+*
+* 档位必须**回头判断**：先除以档位再四舍五入会算出 `999_999 → "1000.0K"`、
+* `999_999_999 → "1000.0M"` 这类既超长、又落在并不存在的档位上的结果，
+* 因此每个档位算完后若已经够到下一档，就改用下一档重算一次。
+*
+* @param value - 非负计数或词元数。
+* @param keepZeroDecimal - 整数值是否保留一位小数。词元显示沿用 `3.0B` 的旧样式；
+*   计数显示去掉多余的 `.0`，见 {@link formatCompactCount}。
+* @returns 紧凑文本；不足 1000 时原样返回该整数。
+*/
+function compact(value, keepZeroDecimal) {
+	const rounded = (scaled) => Math.round(scaled * 10) / 10;
+	const text = (scaled) => {
+		const fixed = rounded(scaled);
+		return keepZeroDecimal ? fixed.toFixed(1) : String(fixed);
+	};
+	if (value >= 1e9) return `${text(value / 1e9)}B`;
+	if (value >= 1e6) return rounded(value / 1e6) >= 1e3 ? `${text(value / 1e9)}B` : `${text(value / 1e6)}M`;
+	if (value >= 1e3) return rounded(value / 1e3) >= 1e3 ? `${text(value / 1e6)}M` : `${text(value / 1e3)}K`;
 	return String(value);
 }
+/** 紧凑词元数：与 {@link formatCompactCount} 共用档位，但整数值保留一位小数（`3.0B`）。 */
+function formatTokensCompact(value) {
+	return compact(value, true);
+}
 /** The input is already in percent units; callers add the percent sign. */
 function formatSuccessRate(value) {
 	return String(Number(value.toFixed(2)));
@@ -10760,6 +10934,8 @@ const Config = z.object(markVolatileFields({
 	filterModelsByPlan: z.boolean(),
 	visibleModels: z.array(z.string()),
 	modelVisibility: z.dict(z.boolean()),
+	contextWindow: z.number().min(1).max(CONTEXT_WINDOW_MAX).step(1),
+	contextWindowOverrides: z.dict(z.number().min(1).max(CONTEXT_WINDOW_MAX).step(1)),
 	webSearch: z.boolean().default(true),
 	showSidebarQuota: z.boolean().default(false),
 	accounts: z.array(z.object({
@@ -10805,6 +10981,8 @@ function resolveAdapterOptions(config) {
 		filterModelsByPlan: config.filterModelsByPlan ?? true,
 		visibleModels: Array.isArray(config.visibleModels) ? config.visibleModels.filter((id) => typeof id === "string" && id !== "") : void 0,
 		modelVisibility: readModelVisibility(config.modelVisibility),
+		contextWindow: isValidContextWindow(config.contextWindow) ? config.contextWindow : void 0,
+		contextWindowOverrides: readContextWindowOverrides(config.contextWindowOverrides),
 		zdr: config.zdr === true
 	};
 }