@mars-sea/dsh-commandcode-provider 0.12.1 → 0.12.2

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@1.72.1) and
+* bundle's model table (`dist/cli.mjs`, re-verified at command-code@1.73.0) 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
@@ -506,7 +506,7 @@ const KNOWN_IMAGE_MODELS = /* @__PURE__ */ new Set([
 ]);
 /**
 * Models WITHOUT a zero-data-retention upstream, per the official CLI's own
-* registry (`command-code@1.72.1` `dist/cli.mjs`): `modelSupportsZdr(id)` is
+* registry (`command-code@1.73.0` `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 would drop `meituan/
@@ -788,7 +788,7 @@ const DEFAULT_MESSAGES_MAX_TOKENS = 64e3;
 * `/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@1.72.1 (2026-09-30): 86 catalog
+* Provider-tier. Re-verified at command-code@1.73.0 (2026-10-01): 86 catalog
 * ids at 53/62/76/84 cumulative, a strict superset chain — every release since
 * 1.49.0 has been additive with no tier move, and per-entry tags below name the
 * release that added each row.
@@ -916,7 +916,7 @@ function compareByPlan(a, b) {
 /**
 * Subscription plan table, synced from the official CLI bundle's plan maps
 * (located by the `"individual-go"` key in `dist/cli.mjs`, re-verified unchanged
-* through command-code@1.72.1): subscription `planId` prefix → display name and
+* through command-code@1.73.0): subscription `planId` prefix → display name and
 * the plan's monthly credit total. This is the account's own subscription
 * (from `/alpha/billing/subscriptions`) — distinct from {@link KNOWN_PLANS},
 * which maps catalog models to their minimum tier.
@@ -1186,16 +1186,19 @@ function capabilityDescription(modelId, contextWindow, now = Date.now()) {
 * rather than against hard-coded numbers, so this table and the subscription
 * table cannot drift apart silently.
 */
-const ALLOWANCE_TIERS = ["goat", "pro"];
+const ALLOWANCE_TIERS = [
+	"go",
+	"goat",
+	"pro"
+];
 /**
 * Which allowance bracket a plan-tier weight belongs to, or undefined when that
 * plan has no per-model allowance.
 *
 * Weights come from `KNOWN_SUBSCRIPTION_PLANS` (go 0 · goat 1 · pro 2 ·
 * provider 3 · max/ultra 4) and {@link PLAN_ORDER} is the same scale, which is
-* what makes the lookup exact: Go, Provider, Max and Ultra have no published
-* allowance, and answering with a neighbouring tier's figure would be a
-* fabricated number rather than a missing one.
+* what makes the lookup exact: Provider、Max、Ultra
+* 没有本地额度；缺项时不能用相邻套餐的数字代替。
 */
 function allowanceTierForWeight(tierWeight) {
 	if (tierWeight === void 0) return void 0;
@@ -1223,6 +1226,7 @@ const MODEL_PRICE_ROWS = [
 			12.5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1236,6 +1240,7 @@ const MODEL_PRICE_ROWS = [
 			12.5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1249,6 +1254,7 @@ const MODEL_PRICE_ROWS = [
 			1.25
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1262,6 +1268,7 @@ const MODEL_PRICE_ROWS = [
 			6.25
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1275,6 +1282,7 @@ const MODEL_PRICE_ROWS = [
 			6.25
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1288,6 +1296,7 @@ const MODEL_PRICE_ROWS = [
 			6.25
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1301,6 +1310,7 @@ const MODEL_PRICE_ROWS = [
 			6.25
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1314,6 +1324,7 @@ const MODEL_PRICE_ROWS = [
 			5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1327,6 +1338,7 @@ const MODEL_PRICE_ROWS = [
 			3.75
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1340,6 +1352,7 @@ const MODEL_PRICE_ROWS = [
 			2.5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1353,6 +1366,7 @@ const MODEL_PRICE_ROWS = [
 			2.5
 		],
 		allowance: {
+			go: 6,
 			goat: 10,
 			pro: 20
 		}
@@ -1370,6 +1384,7 @@ const MODEL_PRICE_ROWS = [
 			.006
 		],
 		allowance: {
+			go: 6,
 			goat: 60,
 			pro: 70
 		}
@@ -1382,6 +1397,7 @@ const MODEL_PRICE_ROWS = [
 			.07
 		],
 		allowance: {
+			go: 8,
 			goat: 20,
 			pro: 30
 		}
@@ -1399,6 +1415,7 @@ const MODEL_PRICE_ROWS = [
 			.006
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -1416,6 +1433,7 @@ const MODEL_PRICE_ROWS = [
 			.044
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -1433,6 +1451,7 @@ const MODEL_PRICE_ROWS = [
 			.006
 		],
 		allowance: {
+			go: 10,
 			goat: 60,
 			pro: 70
 		}
@@ -1450,6 +1469,7 @@ const MODEL_PRICE_ROWS = [
 			.032
 		],
 		allowance: {
+			go: 6,
 			goat: 60,
 			pro: 70
 		}
@@ -1462,6 +1482,7 @@ const MODEL_PRICE_ROWS = [
 			.5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1474,6 +1495,7 @@ const MODEL_PRICE_ROWS = [
 			.03
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1486,6 +1508,7 @@ const MODEL_PRICE_ROWS = [
 			.15
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1498,6 +1521,7 @@ const MODEL_PRICE_ROWS = [
 			.03
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1510,6 +1534,7 @@ const MODEL_PRICE_ROWS = [
 			.15
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1523,6 +1548,7 @@ const MODEL_PRICE_ROWS = [
 			.08334
 		],
 		allowance: {
+			go: 6,
 			goat: 40,
 			pro: 50
 		}
@@ -1535,6 +1561,7 @@ const MODEL_PRICE_ROWS = [
 			.15
 		],
 		allowance: {
+			go: 6,
 			goat: 40,
 			pro: 50
 		}
@@ -1547,6 +1574,7 @@ const MODEL_PRICE_ROWS = [
 			.2
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -1559,6 +1587,7 @@ const MODEL_PRICE_ROWS = [
 			.26
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -1571,6 +1600,7 @@ const MODEL_PRICE_ROWS = [
 			.26
 		],
 		allowance: {
+			go: 10,
 			goat: 70,
 			pro: 80
 		}
@@ -1583,6 +1613,7 @@ const MODEL_PRICE_ROWS = [
 			.5
 		],
 		allowance: {
+			go: 9,
 			goat: 20,
 			pro: 30
 		}
@@ -1595,6 +1626,7 @@ const MODEL_PRICE_ROWS = [
 			.26
 		],
 		allowance: {
+			go: 10,
 			goat: 20,
 			pro: 30
 		}
@@ -1607,6 +1639,7 @@ const MODEL_PRICE_ROWS = [
 			.03
 		],
 		allowance: {
+			go: 10,
 			goat: 40,
 			pro: 50
 		}
@@ -1619,6 +1652,7 @@ const MODEL_PRICE_ROWS = [
 			.075
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -1632,6 +1666,7 @@ const MODEL_PRICE_ROWS = [
 			0
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1645,6 +1680,7 @@ const MODEL_PRICE_ROWS = [
 			0
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1658,6 +1694,7 @@ const MODEL_PRICE_ROWS = [
 			0
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1671,6 +1708,7 @@ const MODEL_PRICE_ROWS = [
 			0
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -1684,6 +1722,7 @@ const MODEL_PRICE_ROWS = [
 			.25
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		},
@@ -1711,6 +1750,7 @@ const MODEL_PRICE_ROWS = [
 			6.25
 		],
 		allowance: {
+			go: 6,
 			goat: 70,
 			pro: 80
 		},
@@ -1738,6 +1778,7 @@ const MODEL_PRICE_ROWS = [
 			2.5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		},
@@ -1765,6 +1806,7 @@ const MODEL_PRICE_ROWS = [
 			12.5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		},
@@ -1783,33 +1825,6 @@ const MODEL_PRICE_ROWS = [
 			25
 		] }]
 	},
-	{
-		id: "gpt-6.1-sol",
-		rates: [
-			2,
-			10,
-			.1,
-			2.5
-		],
-		allowance: {
-			goat: 20,
-			pro: 20
-		},
-		contextTiers: [{
-			"maxContext": 272e3,
-			"rates": [
-				2,
-				10,
-				.1,
-				2.5
-			]
-		}, { "rates": [
-			4,
-			15,
-			.2,
-			5
-		] }]
-	},
 	{
 		id: "gpt-6-luna",
 		rates: [
@@ -1819,6 +1834,7 @@ const MODEL_PRICE_ROWS = [
 			.125
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		},
@@ -1846,6 +1862,7 @@ const MODEL_PRICE_ROWS = [
 			2.5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		},
@@ -1864,6 +1881,34 @@ const MODEL_PRICE_ROWS = [
 			5
 		] }]
 	},
+	{
+		id: "gpt-6.1-sol",
+		rates: [
+			2,
+			10,
+			.1,
+			2.5
+		],
+		allowance: {
+			go: 6,
+			goat: 20,
+			pro: 20
+		},
+		contextTiers: [{
+			"maxContext": 272e3,
+			"rates": [
+				2,
+				10,
+				.1,
+				2.5
+			]
+		}, { "rates": [
+			4,
+			15,
+			.2,
+			5
+		] }]
+	},
 	{
 		id: "grok-4.5",
 		rates: [
@@ -1872,6 +1917,7 @@ const MODEL_PRICE_ROWS = [
 			.5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -1884,6 +1930,7 @@ const MODEL_PRICE_ROWS = [
 			.5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		},
@@ -1908,6 +1955,7 @@ const MODEL_PRICE_ROWS = [
 			.5
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		},
@@ -1932,6 +1980,7 @@ const MODEL_PRICE_ROWS = [
 			.17
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -1944,6 +1993,7 @@ const MODEL_PRICE_ROWS = [
 			.1
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -1956,6 +2006,7 @@ const MODEL_PRICE_ROWS = [
 			.1
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -1968,6 +2019,7 @@ const MODEL_PRICE_ROWS = [
 			.16
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -1980,6 +2032,7 @@ const MODEL_PRICE_ROWS = [
 			.19
 		],
 		allowance: {
+			go: 9,
 			goat: 60,
 			pro: 70
 		}
@@ -1992,6 +2045,7 @@ const MODEL_PRICE_ROWS = [
 			.38
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2004,6 +2058,7 @@ const MODEL_PRICE_ROWS = [
 			.3
 		],
 		allowance: {
+			go: 8,
 			goat: 20,
 			pro: 30
 		}
@@ -2016,6 +2071,7 @@ const MODEL_PRICE_ROWS = [
 			.006
 		],
 		allowance: {
+			go: 10,
 			goat: 50,
 			pro: 60
 		}
@@ -2033,6 +2089,7 @@ const MODEL_PRICE_ROWS = [
 			.16
 		],
 		allowance: {
+			go: 6,
 			goat: 30,
 			pro: 40
 		}
@@ -2050,6 +2107,7 @@ const MODEL_PRICE_ROWS = [
 			.4
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2062,6 +2120,7 @@ const MODEL_PRICE_ROWS = [
 			.0028
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2074,6 +2133,7 @@ const MODEL_PRICE_ROWS = [
 			.0036
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2086,6 +2146,7 @@ const MODEL_PRICE_ROWS = [
 			.036
 		],
 		allowance: {
+			go: 6,
 			goat: 10,
 			pro: 20
 		}
@@ -2098,6 +2159,7 @@ const MODEL_PRICE_ROWS = [
 			.03
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2110,6 +2172,7 @@ const MODEL_PRICE_ROWS = [
 			.06
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2127,6 +2190,7 @@ const MODEL_PRICE_ROWS = [
 			.12
 		],
 		allowance: {
+			go: 8,
 			goat: 47,
 			pro: 57
 		}
@@ -2139,6 +2203,7 @@ const MODEL_PRICE_ROWS = [
 			.15
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 20
 		}
@@ -2151,6 +2216,7 @@ const MODEL_PRICE_ROWS = [
 			.15
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2163,6 +2229,7 @@ const MODEL_PRICE_ROWS = [
 			.002
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2175,6 +2242,7 @@ const MODEL_PRICE_ROWS = [
 			.15
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2187,6 +2255,7 @@ const MODEL_PRICE_ROWS = [
 			.002
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2199,6 +2268,7 @@ const MODEL_PRICE_ROWS = [
 			.12
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2212,6 +2282,7 @@ const MODEL_PRICE_ROWS = [
 			1.63
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2224,6 +2295,7 @@ const MODEL_PRICE_ROWS = [
 			.1
 		],
 		allowance: {
+			go: 10,
 			goat: 33,
 			pro: 43
 		},
@@ -2249,6 +2321,7 @@ const MODEL_PRICE_ROWS = [
 			.038
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		},
@@ -2288,6 +2361,7 @@ const MODEL_PRICE_ROWS = [
 			3.13
 		],
 		allowance: {
+			go: 10,
 			goat: 33,
 			pro: 43
 		}
@@ -2301,6 +2375,7 @@ const MODEL_PRICE_ROWS = [
 			.5
 		],
 		allowance: {
+			go: 10,
 			goat: 33,
 			pro: 43
 		},
@@ -2327,6 +2402,7 @@ const MODEL_PRICE_ROWS = [
 			.04
 		],
 		allowance: {
+			go: 10,
 			goat: 70,
 			pro: 80
 		}
@@ -2339,6 +2415,7 @@ const MODEL_PRICE_ROWS = [
 			.016
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2352,6 +2429,7 @@ const MODEL_PRICE_ROWS = [
 			2.5
 		],
 		allowance: {
+			go: 7,
 			goat: 20,
 			pro: 30
 		}
@@ -2364,6 +2442,7 @@ const MODEL_PRICE_ROWS = [
 			.25
 		],
 		allowance: {
+			go: 7,
 			goat: 20,
 			pro: 30
 		}
@@ -2376,6 +2455,7 @@ const MODEL_PRICE_ROWS = [
 			.016
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2388,6 +2468,7 @@ const MODEL_PRICE_ROWS = [
 			.02
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2400,6 +2481,7 @@ const MODEL_PRICE_ROWS = [
 			.04
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2412,6 +2494,7 @@ const MODEL_PRICE_ROWS = [
 			.05
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2424,6 +2507,7 @@ const MODEL_PRICE_ROWS = [
 			.035
 		],
 		allowance: {
+			go: 10,
 			goat: 70,
 			pro: 80
 		}
@@ -2436,6 +2520,7 @@ const MODEL_PRICE_ROWS = [
 			.042
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2448,6 +2533,7 @@ const MODEL_PRICE_ROWS = [
 			0
 		],
 		allowance: {
+			go: 6,
 			goat: 20,
 			pro: 30
 		}
@@ -2760,6 +2846,16 @@ function installCostProjection(ctx) {
 * JSON, SSE, and pre-stream error bodies undecoded.
 */
 const IDENTITY_ENCODING_HEADER = { "accept-encoding": "identity" };
+//#endregion
+//#region src/timeout-limits.ts
+const MAX_TIMEOUT_SECONDS = 3600;
+//#endregion
+//#region src/model-visibility.ts
+/** 网页、终端与模型目录共用的选择规则；显式开关优先于白名单。 */
+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);
+}
 const THROTTLED_CODE = "THROTTLED";
 const boundedCodes = /* @__PURE__ */ new Set([
 	"EMPTY_RESPONSE",
@@ -2945,28 +3041,10 @@ var CommandCodeAccountPool = class {
 		if (routed !== void 0) return this.pick(routed);
 		const chosen = selectActiveAccount(available, preferred);
 		if (chosen !== void 0) {
-			const explicit = this.explicitAccount(available, options?.model ?? "", preferred);
-			if (explicit !== void 0 && explicit.key !== chosen.key && this.canProbeExplicit(explicit)) {
-				const revived = await this.probeExplicit(explicit);
-				if (revived !== void 0) return this.pick(revived);
-			}
-			return this.pick(chosen);
+			const revived = await this.reviveExplicit(available, options?.model ?? "", preferred, chosen);
+			return this.pick(revived ?? chosen);
 		}
-		await Promise.all(available.map(async (account) => {
-			if (account.state?.kind === "disabled") return;
-			let probe;
-			try {
-				probe = await this.deps.probeWindow(account.key);
-			} catch {
-				return;
-			}
-			if (probe === void 0) return;
-			if (!probe.exceeded) {
-				this.states.delete(account.key);
-				return;
-			}
-			this.stampWindowMark(account, probe);
-		}));
+		await this.probeAllMarked(available);
 		const latestAccounts = await this.resolvedAccounts();
 		const revived = selectActiveAccount(tried === void 0 || tried.length === 0 ? latestAccounts : latestAccounts.filter((account) => !tried.includes(account.key)), preferred);
 		if (revived !== void 0) return this.pick(revived);
@@ -2974,6 +3052,34 @@ var CommandCodeAccountPool = class {
 		throw allAccountsUnusable(latestAccounts.length > 0 ? latestAccounts : accounts);
 	}
 	/**
+	* Which account is serving right now, for display (the settings page's
+	* account card, the sidebar usage stats) — without consuming a request's
+	* `tried` budget and without throwing when nothing can serve.
+	*
+	* Reads and self-heals exactly like {@link resolveKey} would: an `unknown`
+	* mark on the pinned account is re-probed (throttled the same way, sharing
+	* the same per-key timestamp), and a pool where nothing currently resolves
+	* gets one batch probe pass too. Without this, the badge read a possibly
+	* stale mark forever — nothing but an actual generate request ever cleared
+	* an `unknown` mark, so a page left open with no chat activity kept
+	* showing "switched to the fallback account" long after a real request
+	* would have quietly recovered (issue #51's follow-up report: the account
+	* card and the sidebar stats both stuck on the wrong account).
+	*
+	* No `model` parameter: the badge describes the pool as a whole, not one
+	* request, so model-routing rules are deliberately not consulted here —
+	* matching what the direct `selectActiveAccount` call this replaces did.
+	*/
+	async activeAccount() {
+		const accounts = await this.resolvedAccounts();
+		if (accounts.length === 0) return void 0;
+		const preferred = this.deps.preferredId?.();
+		const chosen = selectActiveAccount(accounts, preferred);
+		if (chosen !== void 0) return await this.reviveExplicit(accounts, "", preferred, chosen) ?? chosen;
+		await this.probeAllMarked(accounts);
+		return selectActiveAccount(await this.resolvedAccounts(), preferred);
+	}
+	/**
 	* Record a rejection against one key. A `throttled` rejection (the plain 429
 	* that named no window) marks the key with the `throttle` cause, leaving
 	* rotation exactly like a window mark but keeping the pool's own diagnosis at
@@ -3123,6 +3229,46 @@ var CommandCodeAccountPool = class {
 			until: 0
 		});
 	}
+	/**
+	* The explicit-account fallback re-check: when `chosen` (what rotation
+	* would currently serve) is not the account the user actually asked for,
+	* re-probe that explicit account's mark before accepting the fallback —
+	* so a rate-limit mark never demotes the user's own selection for good
+	* (issue #51). Returns the revived account, or `undefined` when there is
+	* nothing to revive (explicit account already serving, not due for a
+	* probe, or the probe found it still exceeded). Shared by {@link resolveKey}
+	* (the request path) and {@link activeAccount} (the display path), so both
+	* agree on what "currently usable" means instead of drifting apart.
+	*/
+	async reviveExplicit(accounts, model, preferred, chosen) {
+		const explicit = this.explicitAccount(accounts, model, preferred);
+		if (explicit === void 0 || explicit.key === chosen.key || !this.canProbeExplicit(explicit)) return;
+		return this.probeExplicit(explicit);
+	}
+	/**
+	* Probe every marked account's real usage window when nothing currently
+	* resolves. Disabled (401) keys are not probed — an invalid key stays
+	* invalid. A throwing probe counts as "unknown" (like a failed one): it
+	* must not turn the all-exhausted path into a raw rejection instead of
+	* RATE_LIMIT. Shared by {@link resolveKey} and {@link activeAccount}.
+	*/
+	async probeAllMarked(accounts) {
+		await Promise.all(accounts.map(async (account) => {
+			if (account.state?.kind === "disabled") return;
+			let probe;
+			try {
+				probe = await this.deps.probeWindow(account.key);
+			} catch {
+				return;
+			}
+			if (probe === void 0) return;
+			if (!probe.exceeded) {
+				this.states.delete(account.key);
+				return;
+			}
+			this.stampWindowMark(account, probe);
+		}));
+	}
 	/** Hand out the chosen account's key. */
 	pick(account) {
 		return {
@@ -3680,7 +3826,7 @@ function openRequestTiming(model, sink) {
 * `/provider/v1/messages`),
 * message conversion, SSE/JSONL parsing, the catalog + its on-disk cache, and
 * the pre-stream account rotation loop. The wire protocol is
-* reverse-engineered (command-code@1.28.4, re-verified through 1.72.1);
+* reverse-engineered (command-code@1.28.4, re-verified through 1.73.0);
 * docs/agent-reference/adapter-protocol.md holds the full protocol record.
 *
 * The adapter is deliberately free of cordis/schemastery: it receives a
@@ -3782,7 +3928,7 @@ function outputCeilingRefusal(apiBase, modelId, error) {
 function bilingual(code, en, zh, options) {
 	return new LlmError(`${en}；${zh}`, code, options);
 }
-const COMMAND_CODE_CLI_VERSION = "1.72.1";
+const COMMAND_CODE_CLI_VERSION = "1.73.0";
 const DEFAULT_API_BASE = "https://api.commandcode.ai";
 /**
 * The output budget this bundle asks for, and the ceiling it will never exceed.
@@ -3800,9 +3946,16 @@ const DEFAULT_API_BASE = "https://api.commandcode.ai";
 * (a 64 000 request answered headers in 1.28 s against 1.57 s for 131 072) but
 * it did cap what a model could ever say in one turn.
 *
-* This is only an UPPER bound. `requestContextBudget` lowers it per request
-* from the prompt actually being sent, so raising the ceiling cannot push a
-* large session over its window.
+* No client-side clamp is applied on top of this: an earlier version pre-shrank
+* `max_tokens` from an estimate of the prompt already sent (`requestContextBudget()`,
+* issue #67), but that estimate over-corrected on ordinary long conversations —
+* exactly the "output gets cut short as the conversation grows" symptom issue #74
+* reported — while a live A/B probe (2026-09-30, `stealth/space-bunny-alpha`,
+* ~19.6% over its 1,000,000-token window with no client-side shrink) showed the
+* endpoint answers a genuine overflow with an unambiguous `400 … exceeds the
+* model's context window`, which `isContextOverflowDetail()` below already
+* classifies as `CONTEXT_WINDOW_EXCEEDED` independently of this constant. See
+* [决策记录](../docs/决策记录.md).
 */
 const DEFAULT_GENERATE_MAX_TOKENS = 131072;
 /**
@@ -3815,8 +3968,6 @@ const DEFAULT_GENERATE_MAX_TOKENS = 131072;
 * `Math.min(contextLength, this)` before the ceiling above is ever consulted.
 */
 const DEFAULT_MAX_OUTPUT_TOKENS = 131072;
-/** Preserve a small usable answer budget when the heuristic overshoots. */
-const MIN_CONTEXT_OUTPUT_TOKENS = 1024;
 const MODELS_TIMEOUT_MS = 1e4;
 /** 目录短期复用；刷新失败只短暂退避，避免反复阻塞模型选择器。 */
 const MODEL_CATALOG_TTL_MS = 3e5;
@@ -4073,8 +4224,6 @@ const CLI_CONTEXT_OVERFLOW_PATTERN = /prompt is too long|context.*(length|window
 function isContextOverflowDetail(detail) {
 	return isContextWindowExceededError(detail) || CLI_CONTEXT_OVERFLOW_PATTERN.test(detail);
 }
-/** A gateway refusal observed for an over-reserved request, but not unique to that cause. */
-const AMBIGUOUS_SIZE_REJECTION = /\bmodel could not complete the request\b/i;
 /**
 * A request-side `max_tokens` refusal, told apart from a context overflow that
 * mentions the same field. Kept narrow on purpose: the overflow pattern also
@@ -4107,10 +4256,7 @@ function parseOutputCeiling(detail) {
 /**
 * The advice floor for a Provider-API timeout: doubling a 60 s budget is still
 * short for a slow upstream first token, so the suggestion never goes below 5
-* minutes. Capped by `RETRY_MAX_DELAY_MS` rather than the schema's timer
-* ceiling on purpose — a single header wait that outlasts the whole retry
-* backoff is not a budget, and clamping to the engine constant would also pull
-* a runtime dependency into the bundle.
+* minutes. 提示上限与页面可编辑范围相同，重试退避上限不约束单次请求。
 */
 const MIN_SUGGESTED_TIMEOUT_MS = 3e5;
 /**
@@ -4134,10 +4280,13 @@ const MIN_SUGGESTED_TIMEOUT_MS = 3e5;
 * @param timeoutMs - the header budget that just expired.
 */
 function headersTimeoutAdvice(protocol, timeoutMs) {
-	if (protocol === "openai") return {
-		cause: "upstream-first-token",
-		suggestionMs: Math.min(RETRY_MAX_DELAY_MS, Math.max(timeoutMs * 2, MIN_SUGGESTED_TIMEOUT_MS))
-	};
+	if (protocol === "openai") {
+		const suggestionMs = Math.min(MAX_TIMEOUT_SECONDS * 1e3, Math.max(timeoutMs * 2, MIN_SUGGESTED_TIMEOUT_MS));
+		return {
+			cause: "upstream-first-token",
+			...suggestionMs > timeoutMs ? { suggestionMs } : {}
+		};
+	}
 	return { cause: "gateway" };
 }
 /**
@@ -4423,21 +4572,21 @@ function parseCatalogResponse(value, apiBase) {
 	if (models.length === 0) throw new LlmError("Command Code returned an empty model catalog", "PROVIDER_PROTOCOL_ERROR");
 	return models;
 }
-async function readModelsCache(cachePath, apiBase) {
+async function readModelsCache(cachePath, apiBase, restorePublished = true) {
 	const parsed = JSON.parse(await readFile(cachePath, "utf-8"));
 	if (!isRecord(parsed) || parsed.version !== MODEL_CACHE_VERSION || parsed.apiBase !== apiBase || !Array.isArray(parsed.models)) throw new Error(`Invalid model cache at ${cachePath}`);
-	publishedOutputLimits.set(apiBase, /* @__PURE__ */ new Map());
+	if (restorePublished) publishedOutputLimits.set(apiBase, /* @__PURE__ */ new Map());
 	return parsed.models.map((entry) => {
 		const model = isRecord(entry) ? entry : {};
 		const endpoints = Array.isArray(model.supportedEndpoints) ? model.supportedEndpoints.filter((item) => typeof item === "string") : [];
 		const id = stringValue(model.id);
 		const published = numberValue(model.publishedMaxTokens);
-		if (id !== void 0 && published !== void 0 && published > 0) outputLimitsFor(publishedOutputLimits, apiBase).set(id, published);
+		if (restorePublished && id !== void 0 && published !== void 0 && published > 0) outputLimitsFor(publishedOutputLimits, apiBase).set(id, published);
 		const contextWindow = numberValue(model.contextWindow);
 		return {
 			...model,
 			supportedEndpoints: endpoints,
-			maxTokens: id === void 0 || contextWindow === void 0 ? numberValue(model.maxTokens) ?? 131072 : Math.min(contextWindow, modelOutputTokenLimit(apiBase, id))
+			maxTokens: id === void 0 || contextWindow === void 0 ? numberValue(model.maxTokens) ?? 131072 : Math.min(contextWindow, modelOutputTokenLimit(apiBase, id), published !== void 0 && published > 0 ? published : Infinity)
 		};
 	});
 }
@@ -5264,6 +5413,7 @@ async function buildCliBody(options, connection, facts) {
 		memory: null,
 		taste: null,
 		skills: null,
+		permissionMode: "standard",
 		params: {
 			model: options.model,
 			messages: await messagesToCC(options.messages, facts.readImage),
@@ -5363,41 +5513,6 @@ async function buildMessagesBody(options, facts) {
 	return body;
 }
 /**
-* Estimate the text the selected transport actually sends. JSON framing and
-* punctuation each get a token; ordinary ASCII words get one, with an extra
-* allowance for long identifiers. UTF-8 bytes / 4 tracks the measured CJK
-* case better than the engine's UTF-16 characters / 4. This is a heuristic,
-* never proof that a prompt cannot fit: a valid long English prompt can still
-* be larger under this estimate than under the provider's tokenizer.
-*/
-function estimateWireInputTokens(protocol, body, messages, model) {
-	const params = protocol === "cli" ? recordOrEmpty(body.params) : body;
-	const wire = JSON.stringify(params, (_key, value) => typeof value === "string" && /^data:[^,]*;base64,/.test(value) ? "[image]" : value);
-	let tokens = 0;
-	for (const match of wire.matchAll(/[A-Za-z0-9_]+|[^\x00-\x7f]+|[^\s]/gu)) {
-		const part = match[0];
-		if (/^[A-Za-z0-9_]+$/.test(part)) tokens += 1 + Math.ceil(Math.max(0, part.length - 12) / 4);
-		else if (part.charCodeAt(0) > 127) tokens += Math.ceil(Buffer.byteLength(part) / 4);
-		else tokens += 1;
-	}
-	for (const message of messages) for (const block of message.content) if (block.type === "image") tokens += priceRequestImage(block, model, true).visualTokens;
-	return tokens;
-}
-function requestContextBudget(protocol, body, messages, model, contextWindow, requestedMaxTokens) {
-	if (contextWindow === void 0) return {
-		maxTokens: requestedMaxTokens,
-		sizePressure: false
-	};
-	const estimatedInput = estimateWireInputTokens(protocol, body, messages, model);
-	const headroom = Math.min(16384, Math.max(2048, Math.ceil(contextWindow * .02)));
-	const available = Math.floor(contextWindow - estimatedInput - headroom);
-	const floor = Math.min(requestedMaxTokens, MIN_CONTEXT_OUTPUT_TOKENS, Math.max(1, Math.floor(contextWindow / 8)));
-	return {
-		maxTokens: Math.min(requestedMaxTokens, Math.max(floor, available)),
-		sizePressure: estimatedInput >= contextWindow / 4 && estimatedInput + requestedMaxTokens + headroom >= contextWindow
-	};
-}
-/**
 * One connect attempt's inputs: everything `connectGenerate` needs beyond
 * the per-attempt key and protocol. Passed explicitly (not closed over) so
 * the rotation loop's mutation of `protocol`/`body` stays visible at the
@@ -5408,6 +5523,42 @@ function protocolEndpoint(protocol, apiBase) {
 	if (protocol === "cli") return `${apiBase}/alpha/generate`;
 	return protocol === "messages" ? `${apiBase}/provider/v1/messages` : `${apiBase}/provider/v1/chat/completions`;
 }
+/** 错误正文只供诊断与分类使用，不能让坏网关占满内存或无限拖住重试。 */
+async function readGenerateErrorBody(response, abort, timeoutMs) {
+	if (response.body === null) return "";
+	const reader = response.body.getReader();
+	const decoder = new TextDecoder();
+	let text = "";
+	let bytes = 0;
+	const limit = 65536;
+	let removeAbort = () => void 0;
+	const stopped = new Promise((resolve) => {
+		if (abort.signal.aborted) resolve(void 0);
+		else abort.signal.addEventListener("abort", onAbort, { once: true });
+		function onAbort() {
+			resolve(void 0);
+		}
+		removeAbort = () => abort.signal.removeEventListener("abort", onAbort);
+	});
+	const timer = setTimeout(() => abort.abort(), Math.min(timeoutMs, 3e4));
+	try {
+		while (bytes < limit) {
+			const read = await Promise.race([reader.read(), stopped]);
+			if (read === void 0 || read.done) break;
+			const remaining = limit - bytes;
+			text += decoder.decode(read.value.subarray(0, remaining), { stream: true });
+			bytes += read.value.byteLength;
+		}
+		return text + decoder.decode();
+	} catch {
+		return text + decoder.decode();
+	} finally {
+		clearTimeout(timer);
+		removeAbort();
+		reader.cancel().catch(() => void 0);
+		reader.releaseLock();
+	}
+}
 /**
 * One pre-stream connect attempt: POST the body and wait for response
 * headers only (`requestTimeoutMs` must never bound the body stream). Returns
@@ -5450,7 +5601,7 @@ async function connectGenerate(deps, key, protocol, serializedBody) {
 		"x-command-code-version": COMMAND_CODE_CLI_VERSION,
 		"x-cli-environment": "production",
 		"x-project-slug": projectSlugFromPath(connection.workingDir),
-		"x-taste-learning": "true",
+		"x-taste-learning": "false",
 		"x-co-flag": "false"
 	} : {
 		...sharedHeaders,
@@ -5481,8 +5632,13 @@ async function connectGenerate(deps, key, protocol, serializedBody) {
 	}
 	deps.onResponse(response);
 	if (!response.ok) {
-		const errText = await response.text().catch(() => "");
-		cleanup();
+		let errText;
+		try {
+			errText = await readGenerateErrorBody(response, connectAbort, connection.requestTimeoutMs);
+			if (options.signal?.aborted) throw options.signal.reason ?? new DOMException("已取消请求", "AbortError");
+		} finally {
+			cleanup();
+		}
 		const retryAfterMs = parseRetryAfterMs(response.headers.get("retry-after"));
 		return retryAfterMs === void 0 ? {
 			status: response.status,
@@ -5511,6 +5667,7 @@ function createBlockAssembler() {
 		cliCacheWriteTokens: void 0,
 		openAiToolCalls: [],
 		messagesBlocks: /* @__PURE__ */ new Map(),
+		messagesUsage: void 0,
 		replayBlocks: []
 	};
 }
@@ -5720,6 +5877,17 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 		}
 		return this.catalogState;
 	}
+	/** 直接生成也使用本地可信目录，不额外联网；并发冷请求共用一次文件读取。 */
+	async catalogForGenerate(connection) {
+		const state = this.currentCatalogState(connection);
+		if (this.catalog.length > 0 || state.expiresAt > 0) return this.catalog;
+		state.disk ??= readModelsCache(connection.modelsCachePath, connection.apiBase, false).catch(() => []);
+		const cached = await state.disk;
+		if (this.catalogState !== state) return cached;
+		if (this.catalog.length > 0 || state.expiresAt > 0) return this.catalog;
+		if (state.inflight === void 0) this.catalog = cached;
+		return cached;
+	}
 	/** 共享一次刷新；调用者取消只结束自己的等待，不取消其他调用者的请求。 */
 	async loadCatalog(signal) {
 		signal?.throwIfAborted();
@@ -5778,21 +5946,16 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 		if (opts?.unfiltered === true) return catalog.map(toInfo).sort(compareByPlan);
 		const accesses = this.deps.options().filterModelsByPlan === false ? void 0 : await this.loadPoolBillingAccess();
 		const visible = this.deps.options().visibleModels;
-		const allow = Array.isArray(visible) && visible.length > 0 ? new Set(visible.filter((id) => typeof id === "string" && id !== "")) : void 0;
+		const allow = Array.isArray(visible) ? visible.filter((id) => typeof id === "string" && id !== "") : [];
 		const overrides = this.deps.options().modelVisibility;
-		return catalog.filter((model) => modelVisibleForAnyAccount(model.id, accesses)).filter((model) => {
-			const override = overrides?.[model.id];
-			if (typeof override === "boolean") return override;
-			return allow === void 0 || allow.has(model.id);
-		}).map(toInfo).sort(compareByPlan);
+		return catalog.filter((model) => modelVisibleForAnyAccount(model.id, accesses)).filter((model) => modelIsVisible(model.id, allow, overrides)).map(toInfo).sort(compareByPlan);
 	}
 	/**
 	* The per-model allowance bracket this account POOL falls into, or undefined
 	* when there is nothing honest to show.
 	*
-	* The pricing page publishes a per-model monthly allowance for GOAT and Pro
-	* only — its own words are "the boost is a per-model allowance, so it lives on
-	* the plans that have them" — so a pool whose highest plan is Go, Provider,
+	* The pricing page publishes per-model monthly allowances for Go, GOAT and
+	* Pro. A pool whose highest plan is Provider,
 	* Max or Ultra gets undefined rather than a neighbouring tier's figure. Taking
 	* the HIGHEST tier matches the question the picker already answers for a pool
 	* (see `modelVisibleForAnyAccount`): what the user can reach, not what the one
@@ -6234,14 +6397,12 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 		} finally {
 			finishCredentials();
 		}
-		const modelEntry = this.catalog.find((m) => m.id === options.model);
+		const catalog = await this.catalogForGenerate(connection);
+		options.signal?.throwIfAborted();
+		const modelEntry = catalog.find((m) => m.id === options.model);
 		const modelMax = Math.min(modelEntry?.maxTokens ?? Infinity, modelOutputTokenLimit(connection.apiBase, options.model));
 		const requestedMaxTokens = Math.min(options.maxTokens ?? modelMax, modelMax, DEFAULT_GENERATE_MAX_TOKENS);
-		let maxTokens = requestedMaxTokens;
-		let contextBudget = {
-			maxTokens,
-			sizePressure: false
-		};
+		const maxTokens = requestedMaxTokens;
 		const effort = options.reasoningEffort;
 		const supported = KNOWN_EFFORTS[options.model];
 		const reasoningEffort = effort && effort !== "off" && supported?.includes(effort) ? effort : void 0;
@@ -6259,8 +6420,6 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 					readImage
 				};
 				const built = target === "cli" ? await buildCliBody(requestOptions, connection, facts) : target === "messages" ? await buildMessagesBody(requestOptions, facts) : await buildOpenAIBody(requestOptions, facts);
-				contextBudget = requestContextBudget(target, built, requestOptions.messages, options.model, modelEntry?.contextWindow, requestedMaxTokens);
-				maxTokens = contextBudget.maxTokens;
 				if (target === "cli") recordOrEmpty(built.params).max_tokens = maxTokens;
 				else built.max_tokens = maxTokens;
 				return built;
@@ -6370,7 +6529,7 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 						...wait > 0 && wait <= 9e5 ? { providerRetryAfterMs: wait } : {}
 					});
 				}
-				throw generateHttpError(attempt.status, attempt.errText, attempt.retryAfterMs, providerError, contextBudget.sizePressure);
+				throw generateHttpError(attempt.status, attempt.errText, attempt.retryAfterMs, providerError);
 			}
 		} catch (error) {
 			trace.record("connect-error", {
@@ -6419,12 +6578,21 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 		let usage;
 		let usageEmitted = false;
 		let endRecorded = false;
+		let protocolStopped = false;
 		function* handle(event) {
 			captureResponseIdentifiers(responseMetadata, event);
+			if (protocol === "messages" && isRecord(event) && event.type === "message_stop") protocolStopped = true;
 			for (const chunk of handleEvent(asm, protocol, event)) if (chunk.type === "finish") finish = chunk;
 			else if (chunk.type === "usage") usage = chunk.usage;
 			else yield chunk;
 		}
+		function* handleLine(line) {
+			if (protocol === "openai" && line.trim().replace(/^data:\s*/, "") === "[DONE]") {
+				protocolStopped = true;
+				return;
+			}
+			yield* handle(parseStreamEventLine(line));
+		}
 		try {
 			for (;;) {
 				let read;
@@ -6455,7 +6623,7 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 						buffered: buffer.length
 					});
 					if (idleFired) throw bilingual("TIMEOUT", `Command Code API stream from ${connection.apiBase} was idle for ${connection.streamIdleTimeoutMs}ms (no events) and was treated as a dead connection`, `Command Code API 流式响应已 ${connection.streamIdleTimeoutMs} 毫秒无任何事件，被判定为死连接——长思考模型可在设置中调大流空闲超时`);
-					if (buffer.trim()) yield* handle(parseStreamEventLine(buffer));
+					if (buffer.trim()) yield* handleLine(buffer);
 					break;
 				}
 				const text = decoder.decode(value, { stream: true });
@@ -6471,8 +6639,11 @@ var CommandCodeAdapter = class CommandCodeAdapter extends LlmAdapter {
 				});
 				const lines = buffer.split("\n");
 				buffer = lines.pop() ?? "";
-				for (const line of lines) yield* handle(parseStreamEventLine(line));
-				if (finish !== void 0) break;
+				for (const line of lines) {
+					yield* handleLine(line);
+					if (protocolStopped) break;
+				}
+				if (protocolStopped || protocol === "cli" && finish !== void 0) break;
 			}
 			if (usage !== void 0) {
 				usageEmitted = true;
@@ -6590,7 +6761,7 @@ function parseProviderError(errText) {
 		return { message: errText };
 	}
 }
-function generateHttpError(status, errText, retryAfterMs, parsed = parseProviderError(errText), sizePressure = false) {
+function generateHttpError(status, errText, retryAfterMs, parsed = parseProviderError(errText)) {
 	const providerCode = parsed.nested === void 0 ? void 0 : parsed.code;
 	const providerDetail = parsed.nested === void 0 ? "" : [
 		parsed.code,
@@ -6604,11 +6775,12 @@ function generateHttpError(status, errText, retryAfterMs, parsed = parseProvider
 		return bilingual("PROVIDER_HTTP_ERROR", `Command Code API error ${status}: this request's max_tokens is not valid for this model` + (stated ? ` (${stated})` : "") + " — the session is not oversized, so compacting it will not help; lower \"Max output tokens\" or pick a model with a larger output ceiling", `Command Code API 返回 ${status}：本次请求的 max_tokens 对该模型不合法` + (stated ? `（${stated}）` : "") + "——会话内容并未超长，压缩上下文无效；请调低「最大输出 token」或改用输出上限更大的模型", { status });
 	}
 	if (status < 500 && isContextOverflowDetail(overflowDetail)) return bilingual(CONTEXT_WINDOW_EXCEEDED_CODE, `Command Code API error ${status}: the request exceeds the model's context window — this session is being compacted and the request retried`, `Command Code API 返回 ${status}：请求内容超出模型上下文窗口——正在压缩上下文后重试；如仍失败，请新建会话或减少上下文`, { status });
-	if (status === 400 && sizePressure && AMBIGUOUS_SIZE_REJECTION.test(providerDetail) && /\bprovider_error\b/i.test(providerDetail)) return bilingual(CONTEXT_WINDOW_EXCEEDED_CODE, "Command Code API error 400: the provider could not complete a request near the model context limit — compacting the session and retrying", "Command Code API 返回 400：请求接近模型上下文上限，服务商无法完成——正在压缩上下文后重试", { status });
 	if (status === 422 && isZdrNoProviders(providerCode, providerDetail, errText)) return bilingual("PROVIDER_HTTP_ERROR", `Command Code API error 422 (${detail}): zero data retention is enforced but no ZDR-capable upstream is available for this model — turn the ZDR setting off (or use another model)`, "Command Code API 返回 422：已开启零数据保留（ZDR），但该模型当前没有可用的 ZDR 上游——请关闭 ZDR 开关，或改用其他模型", { status: 422 });
 	if (status === 401) return bilingual("INVALID_CREDENTIAL", `Command Code API error 401 (${detail}): the API key is missing or invalid — check the key stored for COMMANDCODE_API_KEY (Models page) or the auth file`, "Command Code API 返回 401：API 密钥缺失或无效——请在设置页检查 COMMANDCODE_API_KEY 存储的密钥，或检查 auth 文件", { status: 401 });
 	if (status === 413) return bilingual("PROVIDER_HTTP_ERROR", "Command Code API error 413: the request body exceeds the provider's size limit (the cap is undocumented, measured at about 50 MB) — the session history is too large to send, usually because of accumulated image attachments. Start a new session, or drop the image-heavy part of this one, and retry", "Command Code API 返回 413：请求体超过服务端的体积上限（官方未公开，实测约 50 MB）——会话历史过大，通常是历史图片累积所致。请新建会话，或移除本会话中图片较多的部分后重试", { status: 413 });
-	return new LlmError(`Command Code API error ${status}${detail === `HTTP ${status}` ? "" : ` (${detail})`}: ${errText.slice(0, 500)}`, httpErrorCode(status), {
+	const diagnosticCode = parsed.code ?? /^\s*(MODEL_NOT_IN_PLAN|INSUFFICIENT_CREDITS|USAGE_EXCEEDED|PREMIUM_CREDITS_EXHAUSTED)\b/i.exec(parsed.message ?? "")?.[1];
+	const diagnostic = diagnosticCode !== void 0 && /^[A-Za-z0-9_.:-]{1,80}$/.test(diagnosticCode) ? `（${diagnosticCode}）` : "";
+	return new LlmError(`Command Code 请求失败（HTTP ${status}）${diagnostic}：` + (status === 429 ? "请求受到用量或速率限制，请稍后重试并检查账户额度。" : status === 408 ? "服务端处理超时，请稍后重试。" : status >= 500 ? "上游服务暂时无法完成请求，请稍后重试。" : "服务端拒绝了本次请求，请检查模型、套餐和请求设置。"), httpErrorCode(status), {
 		status,
 		...retryAfterMs !== void 0 && retryAfterMs > 0 && retryAfterMs <= 9e5 ? { providerRetryAfterMs: retryAfterMs } : {}
 	});
@@ -6856,13 +7028,13 @@ function handleOpenAIEvent(asm, event) {
 * cache_read_input_tokens, output_tokens, output_tokens_details:
 * { thinking_tokens } }`, plus gateway additions the meter ignores.
 */
-function mapMessagesUsage(raw) {
+function mapMessagesUsage(raw, previous) {
 	if (!isRecord(raw)) return void 0;
-	const inputTokens = numberValue(raw.input_tokens) ?? 0;
-	const outputTokens = numberValue(raw.output_tokens) ?? 0;
-	const cacheRead = numberValue(raw.cache_read_input_tokens) ?? 0;
-	const cacheWrite = numberValue(raw.cache_creation_input_tokens) ?? 0;
-	const reasoningTokens = numberValue((isRecord(raw.output_tokens_details) ? raw.output_tokens_details : void 0)?.thinking_tokens);
+	const inputTokens = numberValue(raw.input_tokens) ?? previous?.inputTokens ?? 0;
+	const outputTokens = numberValue(raw.output_tokens) ?? previous?.outputTokens ?? 0;
+	const cacheRead = numberValue(raw.cache_read_input_tokens) ?? previous?.cacheReadTokens ?? 0;
+	const cacheWrite = numberValue(raw.cache_creation_input_tokens) ?? previous?.cacheWriteTokens ?? 0;
+	const reasoningTokens = numberValue((isRecord(raw.output_tokens_details) ? raw.output_tokens_details : void 0)?.thinking_tokens) ?? previous?.reasoningTokens;
 	const usage = {
 		inputTokens: Math.max(0, inputTokens),
 		outputTokens,
@@ -6895,10 +7067,13 @@ function handleMessagesEvent(asm, event) {
 	switch (event.type) {
 		case "message_start": {
 			const usage = mapMessagesUsage((isRecord(event.message) ? event.message : void 0)?.usage);
-			if (usage) chunks.push({
-				type: "usage",
-				usage
-			});
+			if (usage) {
+				asm.messagesUsage = usage;
+				chunks.push({
+					type: "usage",
+					usage
+				});
+			}
 			break;
 		}
 		case "content_block_start": {
@@ -7015,18 +7190,22 @@ function handleMessagesEvent(asm, event) {
 		case "message_delta": {
 			const reason = stringValue((isRecord(event.delta) ? event.delta : void 0)?.stop_reason);
 			if (reason !== void 0) asm.finishReason = reason;
-			const usage = mapMessagesUsage(event.usage);
-			if (usage) chunks.push({
-				type: "usage",
-				usage
-			});
+			const usage = mapMessagesUsage(event.usage, asm.messagesUsage);
+			if (usage) {
+				asm.messagesUsage = usage;
+				chunks.push({
+					type: "usage",
+					usage
+				});
+			}
+			break;
+		}
+		case "message_stop":
 			chunks.push({
 				type: "finish",
-				reason: mapFinishReason(reason)
+				reason: mapFinishReason(asm.finishReason)
 			});
 			break;
-		}
-		case "message_stop":
 		case "ping": break;
 		case "error": throw streamErrorToLlmError(event.error ?? event, stringValue(event.message));
 	}
@@ -7909,7 +8088,7 @@ function applyUsageRemote(ctx, deps) {
 *
 * @module dsh-commandcode-provider/login
 */
-/** Give up on the browser after this long without a callback (mirrors the CLI). */
+/** 登录总期限包含浏览器回调、密钥验证与存储，不能在收到回调后撤掉保护。 */
 const LOGIN_TIMEOUT_MS = 12e4;
 /** First local port the flow tries (mirrors the CLI). */
 const LOGIN_START_PORT = 5959;
@@ -7947,7 +8126,7 @@ function studioBaseForApiBase(apiBase) {
 * tests). Mirrors the CLI's verdicts: 401 → invalid_key, other non-OK →
 * server_error, transport failure → network_error.
 */
-async function validateCommandApiKey(fetchImpl, apiBase, apiKey) {
+async function validateCommandApiKey(fetchImpl, apiBase, apiKey, signal) {
 	try {
 		const response = await fetchImpl(`${apiBase}/alpha/whoami`, {
 			method: "GET",
@@ -7955,8 +8134,10 @@ async function validateCommandApiKey(fetchImpl, apiBase, apiKey) {
 				"Content-Type": "application/json",
 				...IDENTITY_ENCODING_HEADER,
 				Authorization: `Bearer ${apiKey}`
-			}
+			},
+			...signal === void 0 ? {} : { signal }
 		});
+		response.body?.cancel().catch(() => void 0);
 		if (response.status === 401) return {
 			valid: false,
 			error: "invalid_key"
@@ -8002,6 +8183,7 @@ var CommandCodeLoginFlow = class {
 	statusValue = { state: "idle" };
 	server;
 	timer;
+	attemptAbort;
 	/** Settle hooks of the live attempt's callback promise. */
 	settle;
 	/**
@@ -8071,6 +8253,7 @@ var CommandCodeLoginFlow = class {
 	async startAttempt(targetRef) {
 		this.teardown();
 		const attempt = ++this.attemptSeq;
+		this.attemptAbort = new AbortController();
 		const port = await this.findPort();
 		const expectedState = this.deps.randomToken?.(32) ?? randomBytes(32).toString("base64url");
 		const settled = new Promise((resolve, reject) => {
@@ -8100,11 +8283,12 @@ var CommandCodeLoginFlow = class {
 			})
 		});
 		this.timer = setTimeout(() => {
+			if (!this.ownsAttempt(attempt)) return;
 			this.teardown();
 			this.setStatus({
 				state: "failed",
 				reason: "timeout",
-				message: "No browser callback arrived within the login window."
+				message: "登录未在规定时间内完成，请重试。"
 			});
 		}, this.deps.timeoutMs ?? 12e4);
 		this.timer.unref?.();
@@ -8272,7 +8456,8 @@ var CommandCodeLoginFlow = class {
 	settleAttempt(json, code, body, failure, credentials) {
 		json(code, body);
 		const settle = this.settle;
-		this.teardown();
+		if (failure !== void 0) this.teardown();
+		else this.closeCallback();
 		if (settle === void 0) return;
 		if (failure !== void 0) settle.reject(failure);
 		else if (credentials !== void 0) settle.resolve(credentials);
@@ -8289,10 +8474,11 @@ var CommandCodeLoginFlow = class {
 	*/
 	async complete(attempt, credentials, targetRef) {
 		if (!this.ownsAttempt(attempt)) return;
-		const validation = await validateCommandApiKey(this.deps.fetchImpl ?? fetch, this.readApiBase(), credentials.apiKey);
+		const validation = await validateCommandApiKey(this.deps.fetchImpl ?? fetch, this.readApiBase(), credentials.apiKey, this.attemptAbort?.signal);
 		if (!this.ownsAttempt(attempt)) return;
 		if (!validation.valid) {
 			const reason = validation.error === "invalid_key" ? "invalid-key" : validation.error === "network_error" ? "network" : "error";
+			this.teardown();
 			this.setStatus({
 				state: "failed",
 				reason,
@@ -8306,6 +8492,7 @@ var CommandCodeLoginFlow = class {
 			await this.deps.storeKey(credentials, targetRef);
 		} catch (error) {
 			if (!this.ownsAttempt(attempt)) return;
+			this.teardown();
 			this.setStatus({
 				state: "failed",
 				reason: "unavailable",
@@ -8314,7 +8501,7 @@ var CommandCodeLoginFlow = class {
 			return;
 		}
 		if (!this.ownsAttempt(attempt)) return;
-		this.clearTimer();
+		this.teardown();
 		this.setStatus({
 			state: "success",
 			userName: credentials.userName,
@@ -8329,6 +8516,7 @@ var CommandCodeLoginFlow = class {
 	failFrom(attempt, failure) {
 		if (!(failure instanceof LoginSettleError)) return;
 		if (!this.ownsAttempt(attempt)) return;
+		this.teardown();
 		this.setStatus({
 			state: "failed",
 			reason: failure.reason,
@@ -8341,13 +8529,19 @@ var CommandCodeLoginFlow = class {
 			this.timer = void 0;
 		}
 	}
-	/** Close the server and watchdog without touching the published status. */
-	teardown() {
-		this.clearTimer();
+	/** 消费一次回调后关闭监听；保留总期限以约束后续验证。 */
+	closeCallback() {
 		closeServer(this.server);
 		this.server = void 0;
 		this.settle = void 0;
 	}
+	/** 结束整个尝试，同时中止验证请求，迟到结果由所有权检查拦截。 */
+	teardown() {
+		this.clearTimer();
+		this.attemptAbort?.abort();
+		this.attemptAbort = void 0;
+		this.closeCallback();
+	}
 };
 /** Close one loopback server without letting "not running"/"already closed" escape. */
 function closeServer(server) {
@@ -8665,9 +8859,8 @@ function buildCommandCodeTuiSection(deps) {
 	* ONE draft: every one of them then parses that same draft on save, all N
 	* write ops address the same path, and only the LAST field's op survives —
 	* which silently rewrote the allowlist from the last catalog model instead
-	* of the one that was toggled. An override equal to what the array already
-	* says is written as a CLEAR, so toggling a model back to its inherited
-	* state leaves no residue.
+	* of the one that was toggled. 保存显式布尔值，不能因为与白名单相同就
+	* 清除用户层：清除后可能重新露出基础配置中的相反开关。
 	*/
 	const modelField = (id, hint, group) => ({
 		path: ["modelVisibility", id],
@@ -8681,11 +8874,9 @@ function buildCommandCodeTuiSection(deps) {
 			return String(listed.length === 0 || listed.includes(id));
 		},
 		parse: (text) => {
-			const on = text.trim() === "true";
-			const listed = storedIds(deps.visibleModels());
-			return on === (listed.length === 0 || listed.includes(id)) ? { kind: "clear" } : {
+			return {
 				kind: "set",
-				value: on
+				value: text.trim() === "true"
 			};
 		}
 	});
@@ -9132,6 +9323,7 @@ const Config = z.object(markVolatileFields({
 		apiKey: z.string().role("secret")
 	})),
 	activeAccount: z.string(),
+	credentialCleanupRefs: z.array(z.string().role("credential-ref")),
 	modelAccountRules: z.array(z.object({
 		models: z.array(z.string()),
 		account: z.string()
@@ -9291,7 +9483,7 @@ function apply(ctx, config) {
 	const usageReports = async () => {
 		const described = await pool.describeAccounts();
 		const byId = new Map(described.map((account) => [account.slot.id, account]));
-		const active = selectActiveAccount(await pool.resolvedAccounts(), preferredId());
+		const active = await pool.activeAccount();
 		return { accounts: await Promise.all(slots().map(async (slot) => {
 			const account = byId.get(slot.id);
 			let report;