@eddyskywalker/dsh-chatgpt-subscription 0.10.15 → 0.10.21

package/lib/index.js

@@ -13,12 +13,27 @@ import http, { createServer } from "node:http";
 import { WebError } from "@deepseek-ai/dsh-web";
 import { lookup } from "node:dns/promises";
 import { isIP } from "node:net";
-import { createUserMessage } from "@deepseek-ai/dsh-llm/message";
+import { boundContextSummary, createUserMessage } from "@deepseek-ai/dsh-llm/message";
 import { defineTool } from "@deepseek-ai/dsh-tools";
 import { ProxyAgent, fetch as fetch$1 } from "undici";
 import * as SettingsModule from "@deepseek-ai/dsh-settings";
 import { idleWatchdog, timeoutOf } from "@deepseek-ai/dsh-timeout";
 import { URL as URL$1, URLSearchParams as URLSearchParams$1, fileURLToPath } from "node:url";
+//#region src/host/common/output-reservation.ts
+/**
+* Keep fixed defaults only when they fit DSH compaction-basic's default policy.
+* This is host metadata, NEVER a wire cap. Callers must retain their original
+* wire fallback when the metadata is omitted. Explicit request caps still win.
+* Custom compaction policies and windows too small even at zero reservation
+* require deployment configuration; do not invent a larger context window.
+*/
+function outputReservation(contextWindow, maxTokens) {
+	const messageBudget = contextWindow - maxTokens;
+	const threshold = Math.floor(Math.min(contextWindow * .8, messageBudget - 65536));
+	const retain = Math.floor(messageBudget * .16);
+	return Number.isSafeInteger(maxTokens) && maxTokens > 0 && threshold > 0 && retain < threshold ? { defaultMaxTokens: maxTokens } : {};
+}
+//#endregion
 //#region src/shared/minimax-code-contracts.ts
 const MINIMAX_CODE_PROVIDER_ID = "minimax-code";
 const MINIMAX_CODE_PROVIDER_NAME = "MiniMax Code（编程订阅）";
@@ -78,7 +93,7 @@ const CODEX_ORIGINATOR = "opencode";
 /** OAuth presents the same identity the request path does (see above). */
 const OAUTH_ORIGINATOR = CODEX_ORIGINATOR;
 const TOKEN_REFRESH_MARGIN_MS = 6e4;
-const ROUTE_PREFIX$4 = "/api/dsh-chatgpt-subscription";
+const ROUTE_PREFIX$5 = "/api/dsh-chatgpt-subscription";
 const PLUGIN_VERSION = "0.1.0-alpha.0";
 const CODEX_CHATGPT_PROVIDER_ID = "codex-chatgpt";
 const CODEX_API_BASE = "https://chatgpt.com/backend-api/codex";
@@ -122,7 +137,9 @@ const CODEX_MODEL_CATALOG = [
 		defaultReasoningEffort: "medium",
 		reasoningProfile: "gpt-5.6",
 		supportsReasoningSummary: true,
-		fallbackModelId: "gpt-5.6-terra"
+		fallbackModelId: "gpt-5.6-terra",
+		supportsOutputVerbosity: true,
+		defaultOutputVerbosity: "low"
 	},
 	{
 		id: "gpt-6-astra",
@@ -132,7 +149,9 @@ const CODEX_MODEL_CATALOG = [
 		defaultReasoningEffort: "medium",
 		reasoningProfile: "gpt-6",
 		supportsReasoningSummary: true,
-		maxTokens: GPT_6_MAX_TOKENS
+		maxTokens: GPT_6_MAX_TOKENS,
+		supportsOutputVerbosity: true,
+		defaultOutputVerbosity: "low"
 	},
 	{
 		id: "gpt-6.1-sol",
@@ -142,7 +161,9 @@ const CODEX_MODEL_CATALOG = [
 		defaultReasoningEffort: "medium",
 		reasoningProfile: "gpt-6",
 		supportsReasoningSummary: true,
-		maxTokens: GPT_6_MAX_TOKENS
+		maxTokens: GPT_6_MAX_TOKENS,
+		supportsOutputVerbosity: true,
+		defaultOutputVerbosity: "low"
 	},
 	{
 		id: "gpt-6-sol",
@@ -152,7 +173,9 @@ const CODEX_MODEL_CATALOG = [
 		defaultReasoningEffort: "medium",
 		reasoningProfile: "gpt-6",
 		supportsReasoningSummary: true,
-		maxTokens: GPT_6_MAX_TOKENS
+		maxTokens: GPT_6_MAX_TOKENS,
+		supportsOutputVerbosity: true,
+		defaultOutputVerbosity: "low"
 	},
 	{
 		id: "gpt-6-luna",
@@ -162,7 +185,9 @@ const CODEX_MODEL_CATALOG = [
 		defaultReasoningEffort: "medium",
 		reasoningProfile: "gpt-6",
 		supportsReasoningSummary: true,
-		maxTokens: GPT_6_MAX_TOKENS
+		maxTokens: GPT_6_MAX_TOKENS,
+		supportsOutputVerbosity: true,
+		defaultOutputVerbosity: "low"
 	},
 	{
 		id: "gpt-5.6-terra",
@@ -172,7 +197,9 @@ const CODEX_MODEL_CATALOG = [
 		defaultReasoningEffort: "medium",
 		reasoningProfile: "gpt-5.6",
 		supportsReasoningSummary: true,
-		fallbackModelId: "gpt-5.5"
+		fallbackModelId: "gpt-5.5",
+		supportsOutputVerbosity: true,
+		defaultOutputVerbosity: "low"
 	},
 	{
 		id: "gpt-5.6-luna",
@@ -182,7 +209,9 @@ const CODEX_MODEL_CATALOG = [
 		defaultReasoningEffort: "medium",
 		reasoningProfile: "gpt-5.6",
 		supportsReasoningSummary: true,
-		fallbackModelId: "gpt-5.5"
+		fallbackModelId: "gpt-5.5",
+		supportsOutputVerbosity: true,
+		defaultOutputVerbosity: "low"
 	},
 	{
 		id: "gpt-5.5",
@@ -191,7 +220,9 @@ const CODEX_MODEL_CATALOG = [
 		inputModalities: ["text", "image"],
 		defaultReasoningEffort: "medium",
 		reasoningProfile: "standard",
-		supportsReasoningSummary: true
+		supportsReasoningSummary: true,
+		supportsOutputVerbosity: true,
+		defaultOutputVerbosity: "low"
 	},
 	{
 		id: "gpt-5.4",
@@ -275,7 +306,44 @@ function codexModelSupportsImageInput(model) {
 function codexModelSupportsReasoningSummary(model) {
 	return resolveCodexCatalogEntry(model).supportsReasoningSummary;
 }
-/** Per-request output cap for one model, or the pre-GPT-6 default when it declares none. */
+/**
+* One catalog entry by exact id, or undefined when this table has no record.
+*
+* Deliberately NOT {@link resolveCodexCatalogEntry}, which answers an unknown id
+* with the default entry: that substitution is the right guide for a model whose
+* capabilities only need a sensible floor, but it cannot tell "this model
+* declares verbosity support" apart from "nobody said". A wire field must not be
+* invented from silence.
+*/
+function codexCatalogEntryById(model) {
+	return CODEX_MODEL_CATALOG.find((entry) => entry.id === model);
+}
+/**
+* Verbosity to send when the user has not chosen one.
+*
+* "Follow the provider" has to mean what the provider's own client sends. The
+* official catalog declares `default_verbosity: "low"` for every model it
+* lists, so the answer is low — not the server's implicit medium, which is what
+* omitting the field gets.
+*
+* @returns the verbosity to send, or `undefined` to omit the field entirely.
+*/
+function codexDefaultOutputVerbosity(model) {
+	const entry = codexCatalogEntryById(model);
+	if (entry?.supportsOutputVerbosity !== true) return void 0;
+	return entry.defaultOutputVerbosity ?? "low";
+}
+/**
+* Declared output ceiling for one model, or the pre-GPT-6 default when it
+* declares none.
+*
+* This is NOT a request field. The subscription Responses endpoint rejects
+* `max_output_tokens` outright (400 Unsupported parameter, issue #29 — and the
+* official CLI's own request struct has no such field), so nothing is sent to
+* the wire from here. The value reaches DSH only as the adapter's
+* `defaultMaxTokens`, which compaction uses as a local completion reservation
+* when it plans a fold; that number never leaves this process.
+*/
 function codexModelMaxTokens(model) {
 	return resolveCodexCatalogEntry(model).maxTokens ?? 32768;
 }
@@ -296,8 +364,8 @@ function resolveCodexFallbackModel(model) {
 }
 //#endregion
 //#region src/host/model-catalog.ts
-const PROVIDER_ID$6 = CODEX_CHATGPT_PROVIDER_ID;
-const PROVIDER_NAME$6 = "Codex（ChatGPT 订阅）";
+const PROVIDER_ID$7 = CODEX_CHATGPT_PROVIDER_ID;
+const PROVIDER_NAME$7 = "Codex（ChatGPT 订阅）";
 function listCodexModels(preferences, live) {
 	const status = preferences?.status();
 	if (status?.enabled === false) return [];
@@ -309,7 +377,7 @@ function listCodexModels(preferences, live) {
 	const visible = new Set(status?.visibleModelIds ?? (listed ?? shippedModels()).map((entry) => entry.id));
 	const selected = (listed ?? shippedModels()).filter((entry) => visible.has(entry.id));
 	return (listed === void 0 || selected.length > 0 ? selected : shippedModels().filter((entry) => visible.has(entry.id))).map((entry) => ({
-		provider: PROVIDER_ID$6,
+		provider: PROVIDER_ID$7,
 		id: entry.id,
 		name: entry.name,
 		inputModalities: [...entry.inputModalities]
@@ -332,12 +400,12 @@ function resolveCodexModel(model, preferences, live) {
 	const statedDefault = liveEntry?.defaultReasoningEffort ?? entry.defaultReasoningEffort;
 	const defaultEffort = isCodexReasoningEffort(statedDefault) && efforts.includes(statedDefault) ? statedDefault : efforts[0];
 	return {
-		provider: PROVIDER_ID$6,
+		provider: PROVIDER_ID$7,
 		id: model,
 		name: liveEntry?.name ?? (entry.id === model ? entry.name : model),
 		inputModalities: [...liveEntry?.inputModalities ?? entry.inputModalities],
 		context: { contextWindow: configuredContextWindow ?? liveEntry?.contextWindow ?? entry.contextWindow },
-		defaultMaxTokens: codexModelMaxTokens(model),
+		...outputReservation(configuredContextWindow ?? liveEntry?.contextWindow ?? entry.contextWindow, codexModelMaxTokens(model)),
 		reasoning: {
 			efforts: efforts.map((effort) => ({
 				id: ReasoningEffortId(effort),
@@ -349,7 +417,7 @@ function resolveCodexModel(model, preferences, live) {
 }
 //#endregion
 //#region src/host/adapter.ts
-const RETRY_POLICY$5 = resolveRetryPolicy({
+const RETRY_POLICY$6 = resolveRetryPolicy({
 	mode: "normal",
 	maxRetries: 3,
 	retryableCodes: [
@@ -379,11 +447,11 @@ var CodexChatGptAdapter = class extends LlmAdapter {
 	providerInfo(provider) {
 		return {
 			id: provider,
-			name: PROVIDER_NAME$6
+			name: PROVIDER_NAME$7
 		};
 	}
 	providerRetryPolicy() {
-		return RETRY_POLICY$5;
+		return RETRY_POLICY$6;
 	}
 	imageRequestPricing(_provider, _model) {}
 	async listModels() {
@@ -2630,7 +2698,7 @@ async function oauthErrorIdentifier(response) {
 function codexPoolPath() {
 	return path.join(dshHomeDir(), "storages", "codex-pool.json");
 }
-function optionalString$7(record, key) {
+function optionalString$9(record, key) {
 	const value = record[key];
 	if (value === void 0 || value === null) return void 0;
 	if (typeof value !== "string") throw new Error("ChatGPT pool account field is invalid");
@@ -2653,11 +2721,11 @@ function parseCodexPoolData(value) {
 			addedAt: typeof raw.addedAt === "number" ? raw.addedAt : Date.now(),
 			isPrimary: raw.isPrimary === true
 		};
-		const email = optionalString$7(raw, "email") ?? credentials.email;
+		const email = optionalString$9(raw, "email") ?? credentials.email;
 		if (email !== void 0) account.email = email;
-		const planType = optionalString$7(raw, "planType") ?? credentials.planType;
+		const planType = optionalString$9(raw, "planType") ?? credentials.planType;
 		if (planType !== void 0) account.planType = planType;
-		const accountId = optionalString$7(raw, "accountId") ?? credentials.accountId;
+		const accountId = optionalString$9(raw, "accountId") ?? credentials.accountId;
 		if (accountId !== void 0) account.accountId = accountId;
 		if (typeof raw.lastUsedAt === "number") account.lastUsedAt = raw.lastUsedAt;
 		if (typeof raw.cooldownUntil === "number") account.cooldownUntil = raw.cooldownUntil;
@@ -3414,8 +3482,7 @@ function sendSearch(fetchFn, credentials, query, model, id, signal) {
 			settings: {
 				allowed_callers: ["direct"],
 				external_web_access: true
-			},
-			max_output_tokens: 4096
+			}
 		}),
 		signal
 	});
@@ -3678,8 +3745,8 @@ var ProxyManager = class {
 };
 //#endregion
 //#region src/host/antigravity/types.ts
-const PROVIDER_NAME$5 = "Antigravity";
-const PROVIDER_ID$5 = "antigravity";
+const PROVIDER_NAME$6 = "Antigravity";
+const PROVIDER_ID$6 = "antigravity";
 const STREAM_IDLE_TIMEOUT_MS$5 = 3e5;
 const STREAM_IDLE_TIMEOUT_CODE$6 = "LLM_STREAM_IDLE_TIMEOUT";
 const DISCOVERY_TIMEOUT_MS$4 = 8e3;
@@ -4143,13 +4210,13 @@ function parseAntigravityCredentials(value) {
 	if (!(credentials.access || credentials.access_token || credentials.refresh || credentials.refresh_token)) throw new Error("Antigravity credential tokens are missing");
 	return credentials;
 }
-function credentialAccount$3(filePath) {
+function credentialAccount$4(filePath) {
 	return createHash("sha256").update(path.resolve(filePath)).digest("hex");
 }
-function createCredentialBackend$3(filePath) {
+function createCredentialBackend$4(filePath) {
 	if (process.platform === "win32") return new WindowsDpapiCredentialStore(`${filePath}.dpapi`, parseAntigravityCredentials);
-	if (process.platform === "darwin") return new MacKeychainCredentialStore("dsh-antigravity", credentialAccount$3(filePath), parseAntigravityCredentials);
-	if (process.platform === "linux") return new SecretServiceCredentialStore("dsh-antigravity", credentialAccount$3(filePath), parseAntigravityCredentials);
+	if (process.platform === "darwin") return new MacKeychainCredentialStore("dsh-antigravity", credentialAccount$4(filePath), parseAntigravityCredentials);
+	if (process.platform === "linux") return new SecretServiceCredentialStore("dsh-antigravity", credentialAccount$4(filePath), parseAntigravityCredentials);
 	throw new Error("Antigravity encrypted credential storage requires Windows, macOS, or Linux.");
 }
 const credentialOperations$3 = /* @__PURE__ */ new Map();
@@ -4157,13 +4224,13 @@ const credentialOperations$3 = /* @__PURE__ */ new Map();
 var FileCredentialStore$2 = class {
 	filePath;
 	backend;
-	constructor(filePath = credentialPath$2(), backend = createCredentialBackend$3(filePath)) {
+	constructor(filePath = credentialPath$2(), backend = createCredentialBackend$4(filePath)) {
 		this.filePath = filePath;
 		this.backend = backend;
 	}
 	path() {
 		if (process.platform === "win32") return `${this.filePath}.dpapi`;
-		return `${process.platform === "darwin" ? "Keychain" : "Secret Service"}: dsh-antigravity/${credentialAccount$3(this.filePath)}`;
+		return `${process.platform === "darwin" ? "Keychain" : "Secret Service"}: dsh-antigravity/${credentialAccount$4(this.filePath)}`;
 	}
 	serialize(operation) {
 		const key = path.resolve(this.filePath);
@@ -4877,15 +4944,15 @@ async function buildResponsesPayload(options, attachments, localRawImages = {},
 		payload.tool_choice = "auto";
 		payload.parallel_tool_calls = true;
 	}
-	if (outputVerbosity !== null) payload.text = { verbosity: outputVerbosity };
+	const providerVerbosity = codexDefaultOutputVerbosity(options.model);
+	if (providerVerbosity !== void 0) payload.text = { verbosity: outputVerbosity ?? providerVerbosity };
 	if (fastMode) payload.service_tier = "priority";
-	const modelMaxTokens = codexModelMaxTokens(options.model);
-	payload.max_output_tokens = options.maxTokens === void 0 ? modelMaxTokens : Math.min(options.maxTokens, modelMaxTokens);
 	if (options.reasoningEffort !== void 0) {
 		const effort = codexWireReasoningEffort(options.model, options.reasoningEffort);
-		payload.reasoning = codexModelSupportsReasoningSummary(options.model) ? {
+		const summary = reasoningSummary === null || reasoningSummary === "none" ? void 0 : reasoningSummary;
+		payload.reasoning = summary !== void 0 && codexModelSupportsReasoningSummary(options.model) ? {
 			effort,
-			summary: reasoningSummary ?? "auto"
+			summary
 		} : { effort };
 	}
 	return payload;
@@ -6408,13 +6475,13 @@ function registerRoutes(ctx, oauth, usage, preferences, proxyManager, searchSwit
 		}
 		try {
 			switch (url.pathname) {
-				case `${ROUTE_PREFIX$4}/login/start`:
+				case `${ROUTE_PREFIX$5}/login/start`:
 					json(response, {
 						ok: true,
 						value: await oauth.startLogin()
 					});
 					return;
-				case `${ROUTE_PREFIX$4}/login/cancel`: {
+				case `${ROUTE_PREFIX$5}/login/cancel`: {
 					const loginId = field(body, "loginId");
 					if (loginId === null) throw new Error("missing loginId");
 					oauth.cancelLogin(loginId);
@@ -6424,7 +6491,7 @@ function registerRoutes(ctx, oauth, usage, preferences, proxyManager, searchSwit
 					});
 					return;
 				}
-				case `${ROUTE_PREFIX$4}/accounts`: {
+				case `${ROUTE_PREFIX$5}/accounts`: {
 					if (accountPool === void 0) throw new Error("The ChatGPT account pool is not installed.");
 					const action = field(body, "action");
 					const accountId = field(body, "accountId");
@@ -6446,7 +6513,7 @@ function registerRoutes(ctx, oauth, usage, preferences, proxyManager, searchSwit
 					});
 					return;
 				}
-				case `${ROUTE_PREFIX$4}/logout`:
+				case `${ROUTE_PREFIX$5}/logout`:
 					await oauth.logout(field(body, "accountId") ?? void 0);
 					usage.clear();
 					json(response, {
@@ -6454,7 +6521,7 @@ function registerRoutes(ctx, oauth, usage, preferences, proxyManager, searchSwit
 						value: { authenticated: false }
 					});
 					return;
-				case `${ROUTE_PREFIX$4}/token/refresh`: {
+				case `${ROUTE_PREFIX$5}/token/refresh`: {
 					const oauthStatus = await oauth.refresh();
 					json(response, {
 						ok: true,
@@ -6466,27 +6533,27 @@ function registerRoutes(ctx, oauth, usage, preferences, proxyManager, searchSwit
 					});
 					return;
 				}
-				case `${ROUTE_PREFIX$4}/quota/refresh`:
+				case `${ROUTE_PREFIX$5}/quota/refresh`:
 					if (!(await oauth.status()).authenticated) throw new Error("not authenticated");
 					json(response, {
 						ok: true,
 						value: await usage.status(true, true)
 					});
 					return;
-				case `${ROUTE_PREFIX$4}/quota/reset-credit/use`:
+				case `${ROUTE_PREFIX$5}/quota/reset-credit/use`:
 					if (!(await oauth.status()).authenticated) throw new Error("not authenticated");
 					json(response, {
 						ok: true,
 						value: await usage.consumeResetCredit()
 					});
 					return;
-				case `${ROUTE_PREFIX$4}/connection/test`:
+				case `${ROUTE_PREFIX$5}/connection/test`:
 					json(response, {
 						ok: true,
 						value: await usage.testConnection()
 					});
 					return;
-				case `${ROUTE_PREFIX$4}/preferences/update`: {
+				case `${ROUTE_PREFIX$5}/preferences/update`: {
 					const patch = readPreferencesUpdate(body, preferences.status());
 					const value = await preferences.update(patch);
 					if (patch.visibleModelIds !== void 0 || patch.enabled !== void 0 || patch.contextWindowOverrides !== void 0) ctx.emit?.("llm/adapters-updated");
@@ -6563,11 +6630,11 @@ function registerRoutes(ctx, oauth, usage, preferences, proxyManager, searchSwit
 	};
 	const disposers = [ctx.webServer.register({
 		kind: "prefix",
-		path: ROUTE_PREFIX$4,
+		path: ROUTE_PREFIX$5,
 		handler
 	}), ctx.webServer.register({
 		kind: "exact",
-		path: `${ROUTE_PREFIX$4}/login/events`,
+		path: `${ROUTE_PREFIX$5}/login/events`,
 		handler: events
 	})];
 	return () => {
@@ -6675,6 +6742,27 @@ function statusFor(error) {
 }
 //#endregion
 //#region src/host/search-provider-switcher.ts
+/**
+* Cordis fiber states, mirrored from its `FiberState` const enum.
+*
+* It is a const enum, so importing the values would inline them here anyway and
+* the numbers would be duplicated in the emitted JavaScript either way. Naming
+* them locally keeps this module free of a type-only import that a consumer's
+* build would still have to resolve, and the values are asserted against the
+* installed cordis in the test below.
+*/
+const FIBER_LOADING = 1;
+const FIBER_ACTIVE = 2;
+const FIBER_UNLOADING = 5;
+/** How long to wait for a restarted entry to become active again. */
+const RESTART_TIMEOUT_MS = 1e4;
+/** Gap between restart-state samples. */
+const RESTART_POLL_MS = 10;
+function delay(ms) {
+	return new Promise((resolve) => {
+		setTimeout(resolve, ms);
+	});
+}
 var SearchProviderSwitcher = class {
 	loader;
 	originalSearchProvider;
@@ -6733,15 +6821,40 @@ var SearchProviderSwitcher = class {
 		else nextConfig.fetchProvider = nextFetch;
 		this.state = "applying";
 		try {
-			if (configured && entry.fiber) await entry.fiber.update(nextConfig, true);
+			if (configured && entry.fiber) entry.fiber.update(nextConfig, true);
 			else await entry.update({ config: nextConfig });
-			await entry.fiber?.await();
+			await this.awaitSettled(entry);
 			this.state = "applied";
 		} catch (error) {
 			this.state = "failed";
 			throw error;
 		}
 	}
+	/**
+	* Wait until the entry's fiber has finished restarting, not just until it
+	* happens not to be mid-restart when asked.
+	*
+	* `Fiber.await()` alone is a single sample: it returns at once when `inertia`
+	* is unset, which is exactly the state the fiber is in on the first tick after
+	* `update()` queues its restart. Sampling that reports success while the entry
+	* is still UNLOADING. This waits for a restart to actually be observed, then
+	* lets `Fiber.await()` drain whatever queue is left, so a slow or multi-pass
+	* restart cannot be reported as applied either.
+	*/
+	async awaitSettled(entry) {
+		const fiber = entry.fiber;
+		if (fiber === void 0) return;
+		const deadline = Date.now() + RESTART_TIMEOUT_MS;
+		let sawRestart = false;
+		while (Date.now() < deadline) {
+			const state = fiber.state;
+			if (state === FIBER_LOADING || state === FIBER_UNLOADING) sawRestart = true;
+			if (sawRestart && state === FIBER_ACTIVE) return;
+			await fiber.await();
+			if (fiber.state === FIBER_ACTIVE) return;
+			await delay(RESTART_POLL_MS);
+		}
+	}
 	findWebEntry() {
 		for (const entry of this.loader.entries()) if (entry.options.id === "web" || entry.options.name === "@deepseek-ai/dsh-web") return entry;
 		return null;
@@ -7971,7 +8084,7 @@ function buildRequest$2(options, model, projectId, runtimeModel, effort, images
 		requestId: `req_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`
 	};
 }
-function createStreamState$5() {
+function createStreamState$6() {
 	return {
 		blocks: [],
 		replayBlocks: [],
@@ -8032,7 +8145,7 @@ function processStreamLine$2(line, state) {
 	const json = line.slice(5).trim();
 	if (json === "[DONE]") {
 		state.done = true;
-		return closeStream$4(state);
+		return closeStream$5(state);
 	}
 	if (!json) return [];
 	const chunk = safeJsonParse$5(json);
@@ -8149,7 +8262,7 @@ function processStreamLine$2(line, state) {
 	}
 	return out;
 }
-function closeStream$4(state) {
+function closeStream$5(state) {
 	if (state.finished) return [];
 	if (!state.finishReason && !state.done) throw new LlmError$1("Antigravity stream ended before its terminal response", "PROVIDER_ERROR");
 	state.finished = true;
@@ -8163,7 +8276,7 @@ function closeStream$4(state) {
 		type: "finish",
 		reason,
 		replayState: {
-			response: { provider: PROVIDER_ID$5 },
+			response: { provider: PROVIDER_ID$6 },
 			blocks: state.replayBlocks
 		}
 	});
@@ -8371,7 +8484,7 @@ var AntigravityAdapter = class extends LlmAdapter {
 	providerInfo(provider) {
 		return {
 			id: provider,
-			name: PROVIDER_NAME$5
+			name: PROVIDER_NAME$6
 		};
 	}
 	providerRetryPolicy() {}
@@ -8389,7 +8502,7 @@ var AntigravityAdapter = class extends LlmAdapter {
 			name: model.name,
 			inputModalities: model.inputModalities,
 			context: { contextWindow: overrides[model.id] || model.contextWindow },
-			defaultMaxTokens: model.maxTokens,
+			...outputReservation(overrides[model.id] || model.contextWindow, model.maxTokens),
 			...model.reasoningEfforts ? { reasoningEfforts: model.reasoningEfforts } : {}
 		}));
 	}
@@ -8416,7 +8529,7 @@ var AntigravityAdapter = class extends LlmAdapter {
 			name: model.name,
 			inputModalities: model.inputModalities,
 			context: { contextWindow: overrides[model.id] || model.contextWindow },
-			defaultMaxTokens: model.maxTokens,
+			...outputReservation(overrides[model.id] || model.contextWindow, model.maxTokens),
 			...model.reasoningEfforts ? { reasoning: {
 				efforts: efforts.map((effort) => ({
 					id: ReasoningEffortId(effort),
@@ -8446,7 +8559,10 @@ var AntigravityAdapter = class extends LlmAdapter {
 			...options,
 			reasoningEffort: effectiveEffort
 		} : options;
-		yield* wrapStreamWithWatchdog((watchdogSignal) => this.requestStream(effectiveOptions, model, watchdogSignal), options.signal, STREAM_IDLE_TIMEOUT_MS$5, STREAM_IDLE_TIMEOUT_CODE$6, "Antigravity");
+		yield* wrapStreamWithWatchdog((watchdogSignal) => this.requestStream({
+			...effectiveOptions,
+			maxTokens: options.maxTokens ?? model.maxTokens
+		}, model, watchdogSignal), options.signal, STREAM_IDLE_TIMEOUT_MS$5, STREAM_IDLE_TIMEOUT_CODE$6, "Antigravity");
 	}
 	async *requestStream(options, model, signal) {
 		const fetchFn = this.options.fetchFn ?? fetch;
@@ -8519,7 +8635,7 @@ var AntigravityAdapter = class extends LlmAdapter {
 		if (!response.body) throw new LlmError("Antigravity returned empty response body", "PROVIDER_ERROR");
 		const reader = response.body.getReader();
 		const decoder = new TextDecoder();
-		const state = createStreamState$5();
+		const state = createStreamState$6();
 		let buffer = "";
 		try {
 			while (true) {
@@ -8541,7 +8657,7 @@ var AntigravityAdapter = class extends LlmAdapter {
 				const chunks = processStreamLine$2(buffer.trim(), state);
 				for (const chunk of chunks) yield chunk;
 			}
-			for (const chunk of closeStream$4(state)) yield chunk;
+			for (const chunk of closeStream$5(state)) yield chunk;
 		} finally {
 			reader.cancel().catch(() => void 0);
 		}
@@ -8606,17 +8722,17 @@ var QuotaRefresh = class {
 };
 //#endregion
 //#region src/host/antigravity/routes.ts
-function sendJson$7(response, status, body) {
+function sendJson$8(response, status, body) {
 	response.writeHead(status, { "Content-Type": "application/json" });
 	response.end(JSON.stringify(body));
 }
-function sendMethodNotAllowed$5(response) {
-	sendJson$7(response, 405, {
+function sendMethodNotAllowed$6(response) {
+	sendJson$8(response, 405, {
 		ok: false,
 		error: "Method Not Allowed"
 	});
 }
-async function readRequestJson$6(request) {
+async function readRequestJson$7(request) {
 	return new Promise((resolve, reject) => {
 		let raw = "";
 		request.on("data", (chunk) => {
@@ -8676,7 +8792,7 @@ function registerAntigravityRoutes(ctx, store, modelSettings, preferences, fetch
 			const path = new URL(request.url || "/", "http://dsh.local").pathname.replace(/^\/antigravity\/api\/?/, "");
 			try {
 				if (path === "status" || path === "") {
-					if (request.method !== "GET") return sendMethodNotAllowed$5(response);
+					if (request.method !== "GET") return sendMethodNotAllowed$6(response);
 					const credentials = await store.read();
 					const authenticated = !!(credentials?.access || credentials?.access_token);
 					const cached = getCachedQuota();
@@ -8685,7 +8801,7 @@ function registerAntigravityRoutes(ctx, store, modelSettings, preferences, fetch
 						if (cached === void 0) await quotaRefresh.run(refresh);
 						else quotaRefresh.start(refresh);
 					}
-					return sendJson$7(response, 200, {
+					return sendJson$8(response, 200, {
 						ok: true,
 						value: {
 							...await getAntigravityWebStatus(store, modelSettings, preferences, accountPool),
@@ -8694,8 +8810,8 @@ function registerAntigravityRoutes(ctx, store, modelSettings, preferences, fetch
 					});
 				}
 				if (path === "login" || path === "accounts/login") {
-					if (request.method !== "POST") return sendMethodNotAllowed$5(response);
-					return sendJson$7(response, 200, {
+					if (request.method !== "POST") return sendMethodNotAllowed$6(response);
+					return sendJson$8(response, 200, {
 						ok: true,
 						value: await beginWebLogin$2(store, fetchFn, void 0, async (creds) => {
 							await accountPool.addAccount(creds);
@@ -8703,19 +8819,19 @@ function registerAntigravityRoutes(ctx, store, modelSettings, preferences, fetch
 					});
 				}
 				if (path === "login/status") {
-					if (request.method !== "GET") return sendMethodNotAllowed$5(response);
-					return sendJson$7(response, 200, {
+					if (request.method !== "GET") return sendMethodNotAllowed$6(response);
+					return sendJson$8(response, 200, {
 						ok: true,
 						value: getWebLoginStatus$3()
 					});
 				}
 				if (path === "accounts") {
-					if (request.method === "GET") return sendJson$7(response, 200, {
+					if (request.method === "GET") return sendJson$8(response, 200, {
 						ok: true,
 						value: await accountPool.listAccounts()
 					});
 					if (request.method === "POST") {
-						const body = await readRequestJson$6(request);
+						const body = await readRequestJson$7(request);
 						const action = String(body.action || "");
 						if (action === "set-primary" && typeof body.accountId === "string") await accountPool.setPrimary(body.accountId);
 						else if (action === "set-alias" && typeof body.accountId === "string" && typeof body.alias === "string") await accountPool.setAlias(body.accountId, body.alias);
@@ -8723,17 +8839,17 @@ function registerAntigravityRoutes(ctx, store, modelSettings, preferences, fetch
 						else if (action === "strategy" && (body.strategy === "sequential" || body.strategy === "round-robin" || body.strategy === "sticky")) await accountPool.setStrategy(body.strategy);
 						else if (action === "clear-cooldown" && typeof body.accountId === "string") await accountPool.clearCooldown(body.accountId);
 						clearCachedQuota();
-						return sendJson$7(response, 200, {
+						return sendJson$8(response, 200, {
 							ok: true,
 							value: await getAntigravityWebStatus(store, modelSettings, preferences, accountPool)
 						});
 					}
-					return sendMethodNotAllowed$5(response);
+					return sendMethodNotAllowed$6(response);
 				}
 				if (path === "quota") {
-					if (request.method !== "GET" && request.method !== "POST") return sendMethodNotAllowed$5(response);
+					if (request.method !== "GET" && request.method !== "POST") return sendMethodNotAllowed$6(response);
 					const quota = await fetchAccountQuota(store, modelSettings, fetchFn, true);
-					return sendJson$7(response, 200, {
+					return sendJson$8(response, 200, {
 						ok: true,
 						value: {
 							...await getAntigravityWebStatus(store, modelSettings, preferences, accountPool),
@@ -8742,38 +8858,38 @@ function registerAntigravityRoutes(ctx, store, modelSettings, preferences, fetch
 					});
 				}
 				if (path === "settings") {
-					if (request.method !== "POST") return sendMethodNotAllowed$5(response);
-					const body = await readRequestJson$6(request);
+					if (request.method !== "POST") return sendMethodNotAllowed$6(response);
+					const body = await readRequestJson$7(request);
 					if (preferences) await preferences.update(body);
 					else await modelSettings.updateSettings(body);
 					if (body.enabledModelIds !== void 0 || body.enabled !== void 0) ctx.emit?.("llm/adapters-updated");
-					return sendJson$7(response, 200, {
+					return sendJson$8(response, 200, {
 						ok: true,
 						value: await getAntigravityWebStatus(store, modelSettings, preferences, accountPool)
 					});
 				}
 				if (path === "models") {
-					if (request.method === "GET") return sendJson$7(response, 200, {
+					if (request.method === "GET") return sendJson$8(response, 200, {
 						ok: true,
 						value: (await getAntigravityWebStatus(store, modelSettings, preferences, accountPool)).models
 					});
 					if (request.method === "POST") {
-						const body = await readRequestJson$6(request);
+						const body = await readRequestJson$7(request);
 						if (Array.isArray(body.enabledModelIds) || body.enabled !== void 0 || body.contextWindowOverrides || body.defaultReasoningEffort !== void 0) {
 							if (preferences) await preferences.update(body);
 							else await modelSettings.updateSettings(body);
 							if (body.enabledModelIds !== void 0 || body.enabled !== void 0) ctx.emit?.("llm/adapters-updated");
 						}
-						return sendJson$7(response, 200, {
+						return sendJson$8(response, 200, {
 							ok: true,
 							value: await getAntigravityWebStatus(store, modelSettings, preferences, accountPool)
 						});
 					}
-					return sendMethodNotAllowed$5(response);
+					return sendMethodNotAllowed$6(response);
 				}
 				if (path === "logout") {
-					if (request.method !== "POST") return sendMethodNotAllowed$5(response);
-					const body = await readRequestJson$6(request).catch(() => ({}));
+					if (request.method !== "POST") return sendMethodNotAllowed$6(response);
+					const body = await readRequestJson$7(request).catch(() => ({}));
 					const targetId = typeof body.accountId === "string" ? body.accountId : void 0;
 					if (targetId) await accountPool.deleteAccount(targetId);
 					else {
@@ -8782,17 +8898,17 @@ function registerAntigravityRoutes(ctx, store, modelSettings, preferences, fetch
 						else await store.delete();
 					}
 					clearCachedQuota();
-					return sendJson$7(response, 200, {
+					return sendJson$8(response, 200, {
 						ok: true,
 						value: await getAntigravityWebStatus(store, modelSettings, preferences, accountPool)
 					});
 				}
-				return sendJson$7(response, 404, {
+				return sendJson$8(response, 404, {
 					ok: false,
 					error: "not-found"
 				});
 			} catch (err) {
-				return sendJson$7(response, 500, {
+				return sendJson$8(response, 500, {
 					ok: false,
 					error: err instanceof Error ? err.message : String(err)
 				});
@@ -8804,6 +8920,20 @@ function registerAntigravityRoutes(ctx, store, modelSettings, preferences, fetch
 //#region src/host/command-code/model-catalog.ts
 /** Every model the Command Code registry describes, in registry order. */
 const COMMAND_CODE_MODELS = [
+	{
+		id: "claude-sonnet-5-5",
+		name: "Claude Sonnet 5.5",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"high",
+			"xhigh",
+			"max"
+		],
+		contextWindow: 1e6,
+		maxTokens: null
+	},
 	{
 		id: "claude-sonnet-5",
 		name: "Claude Sonnet 5",
@@ -8860,6 +8990,20 @@ const COMMAND_CODE_MODELS = [
 		contextWindow: 1e6,
 		maxTokens: null
 	},
+	{
+		id: "claude-opus-5-5",
+		name: "Claude Opus 5.5",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"high",
+			"xhigh",
+			"max"
+		],
+		contextWindow: 1e6,
+		maxTokens: null
+	},
 	{
 		id: "claude-opus-5",
 		name: "Claude Opus 5",
@@ -8924,6 +9068,48 @@ const COMMAND_CODE_MODELS = [
 		contextWindow: 105e4,
 		maxTokens: null
 	},
+	{
+		id: "gpt-6.1-sol",
+		name: "GPT-6.1 Sol",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"high",
+			"xhigh",
+			"max"
+		],
+		contextWindow: 105e4,
+		maxTokens: null
+	},
+	{
+		id: "gpt-6-sol",
+		name: "GPT-6 Sol",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"high",
+			"xhigh",
+			"max"
+		],
+		contextWindow: 105e4,
+		maxTokens: null
+	},
+	{
+		id: "gpt-6-luna",
+		name: "GPT-6 Luna",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"high",
+			"xhigh",
+			"max"
+		],
+		contextWindow: 105e4,
+		maxTokens: null
+	},
 	{
 		id: "gpt-5.6-sol",
 		name: "GPT-5.6 Sol",
@@ -9021,7 +9207,11 @@ const COMMAND_CODE_MODELS = [
 		id: "deepseek/deepseek-v4-pro",
 		name: "DeepSeek V4 Pro (latest)",
 		inputModalities: ["text"],
-		reasoningEfforts: ["high", "max"],
+		reasoningEfforts: [
+			"off",
+			"high",
+			"max"
+		],
 		contextWindow: 1e6,
 		maxTokens: null
 	},
@@ -9029,7 +9219,11 @@ const COMMAND_CODE_MODELS = [
 		id: "deepseek/deepseek-v4-flash",
 		name: "DeepSeek V4 Flash (latest)",
 		inputModalities: ["text"],
-		reasoningEfforts: ["high", "max"],
+		reasoningEfforts: [
+			"off",
+			"high",
+			"max"
+		],
 		contextWindow: 1e6,
 		maxTokens: null
 	},
@@ -9037,7 +9231,11 @@ const COMMAND_CODE_MODELS = [
 		id: "deepseek/deepseek-v4-flash-vision-exp",
 		name: "DeepSeek V4 Flash Vision (exp)",
 		inputModalities: ["text", "image"],
-		reasoningEfforts: ["high", "max"],
+		reasoningEfforts: [
+			"off",
+			"high",
+			"max"
+		],
 		contextWindow: 1e6,
 		maxTokens: null
 	},
@@ -9058,6 +9256,20 @@ const COMMAND_CODE_MODELS = [
 		name: "DeepSeek V4.1 Flash",
 		inputModalities: ["text", "image"],
 		reasoningEfforts: [
+			"off",
+			"low",
+			"high",
+			"max"
+		],
+		contextWindow: 1e6,
+		maxTokens: null
+	},
+	{
+		id: "deepseek/deepseek-v4.1-flash-fast",
+		name: "DeepSeek V4.1 Flash Fast",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"off",
 			"low",
 			"high",
 			"max"
@@ -9121,6 +9333,18 @@ const COMMAND_CODE_MODELS = [
 		contextWindow: 1048576,
 		maxTokens: 131072
 	},
+	{
+		id: "z-ai/glm-5.3-flashx",
+		name: "GLM-5.3 FlashX",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"low",
+			"high",
+			"max"
+		],
+		contextWindow: 1e6,
+		maxTokens: 131072
+	},
 	{
 		id: "zai-org/GLM-5.3",
 		name: "GLM-5.3",
@@ -9154,7 +9378,7 @@ const COMMAND_CODE_MODELS = [
 		name: "GLM-5.1",
 		inputModalities: ["text"],
 		reasoningEfforts: [],
-		contextWindow: null,
+		contextWindow: 2e5,
 		maxTokens: null
 	},
 	{
@@ -9182,35 +9406,39 @@ const COMMAND_CODE_MODELS = [
 		name: "MiniMax M2.7",
 		inputModalities: ["text"],
 		reasoningEfforts: [],
-		contextWindow: null,
+		contextWindow: 2e5,
+		maxTokens: null
+	},
+	{
+		id: "MiniMaxAI/MiniMax-M2.5",
+		name: "MiniMax M2.5",
+		inputModalities: ["text"],
+		reasoningEfforts: [],
+		contextWindow: 2e5,
 		maxTokens: null
 	},
 	{
-		id: "minimax/minimax-m3-free",
-		name: "MiniMax M3 (Free)",
+		id: "xiaomi/mimo-v2.6-pro",
+		name: "MiMo V2.6 Pro",
 		inputModalities: ["text", "image"],
-		reasoningEfforts: [
-			"low",
-			"medium",
-			"high"
-		],
-		contextWindow: 1e6,
+		reasoningEfforts: [],
+		contextWindow: 1048576,
 		maxTokens: null
 	},
 	{
-		id: "minimax/minimax-m2.7-free",
-		name: "MiniMax M2.7 (Free)",
-		inputModalities: ["text"],
+		id: "xiaomi/mimo-v2.6-pro-ultraspeed",
+		name: "MiMo V2.6 Pro UltraSpeed",
+		inputModalities: ["text", "image"],
 		reasoningEfforts: [],
-		contextWindow: 197e3,
+		contextWindow: 1048576,
 		maxTokens: null
 	},
 	{
-		id: "MiniMaxAI/MiniMax-M2.5",
-		name: "MiniMax M2.5",
-		inputModalities: ["text"],
+		id: "xiaomi/mimo-v2.6-flash",
+		name: "MiMo V2.6 Flash",
+		inputModalities: ["text", "image"],
 		reasoningEfforts: [],
-		contextWindow: 2e5,
+		contextWindow: 1048576,
 		maxTokens: null
 	},
 	{
@@ -9229,6 +9457,18 @@ const COMMAND_CODE_MODELS = [
 		contextWindow: 1e6,
 		maxTokens: null
 	},
+	{
+		id: "Qwen/Qwen3.8-Omni-Flash",
+		name: "Qwen 3.8 Omni Flash",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"xhigh"
+		],
+		contextWindow: 1e6,
+		maxTokens: 131072
+	},
 	{
 		id: "Qwen/Qwen3.8-Max-0902",
 		name: "Qwen 3.8 Max 0902",
@@ -9306,7 +9546,7 @@ const COMMAND_CODE_MODELS = [
 		name: "Qwen 3.6 Max Preview",
 		inputModalities: ["text"],
 		reasoningEfforts: [],
-		contextWindow: null,
+		contextWindow: 2e5,
 		maxTokens: null
 	},
 	{
@@ -9314,17 +9554,29 @@ const COMMAND_CODE_MODELS = [
 		name: "Qwen 3.6 Plus",
 		inputModalities: ["text", "image"],
 		reasoningEfforts: [],
-		contextWindow: null,
+		contextWindow: 2e5,
 		maxTokens: null
 	},
 	{
-		id: "meituan/LongCat-2.0:free",
+		id: "meituan/LongCat-2.0",
 		name: "LongCat 2.0",
 		inputModalities: ["text"],
 		reasoningEfforts: [],
 		contextWindow: 1048576,
 		maxTokens: null
 	},
+	{
+		id: "stepfun/Step-5-Preview",
+		name: "Step 5 Preview",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"high"
+		],
+		contextWindow: 1e6,
+		maxTokens: null
+	},
 	{
 		id: "stepfun/Step-3.7-Flash",
 		name: "Step 3.7 Flash",
@@ -9338,14 +9590,6 @@ const COMMAND_CODE_MODELS = [
 		name: "Step 3.5 Flash",
 		inputModalities: ["text"],
 		reasoningEfforts: [],
-		contextWindow: 1e6,
-		maxTokens: null
-	},
-	{
-		id: "tencent/Hy3",
-		name: "Tencent Hy3 (Free)",
-		inputModalities: ["text"],
-		reasoningEfforts: [],
 		contextWindow: 262144,
 		maxTokens: null
 	},
@@ -9473,6 +9717,19 @@ const COMMAND_CODE_MODELS = [
 		contextWindow: 1e6,
 		maxTokens: null
 	},
+	{
+		id: "stealth/space-bunny-alpha",
+		name: "Space Bunny Alpha",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"high",
+			"max"
+		],
+		contextWindow: 1e6,
+		maxTokens: 524288
+	},
 	{
 		id: "poolside/laguna-s-2.1-free",
 		name: "Laguna S 2.1",
@@ -9482,18 +9739,22 @@ const COMMAND_CODE_MODELS = [
 		maxTokens: 32768
 	},
 	{
-		id: "inclusionai/ling-3.0-flash-free",
-		name: "Ling 3.0 Flash",
+		id: "inclusionai/ling-3.0-flash-sante:free",
+		name: "Ling 3.0 Flash Sante",
 		inputModalities: ["text"],
 		reasoningEfforts: [],
-		contextWindow: 256e3,
+		contextWindow: 262144,
 		maxTokens: 32768
 	},
 	{
-		id: "inclusionai/ling-3.0-flash-sante:free",
-		name: "Ling 3.0 Flash Sante",
+		id: "inclusionai/ling-3.1-flash:free",
+		name: "Ling 3.1 Flash",
 		inputModalities: ["text"],
-		reasoningEfforts: [],
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"high"
+		],
 		contextWindow: 262144,
 		maxTokens: 32768
 	},
@@ -9587,6 +9848,19 @@ const COMMAND_CODE_MODELS = [
 		],
 		contextWindow: 5e5,
 		maxTokens: null
+	},
+	{
+		id: "xai/grok-4.7",
+		name: "Grok 4.7",
+		inputModalities: ["text", "image"],
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"high",
+			"xhigh"
+		],
+		contextWindow: 5e5,
+		maxTokens: null
 	}
 ];
 /** Registry entry for one model id, or undefined when it describes no such model. */
@@ -9595,8 +9869,8 @@ function commandCodeModelDef(modelId) {
 }
 //#endregion
 //#region src/host/command-code/types.ts
-const PROVIDER_ID$4 = "command-code";
-const PROVIDER_NAME$4 = "Command Code";
+const PROVIDER_ID$5 = "command-code";
+const PROVIDER_NAME$5 = "Command Code";
 /** Browser sign-in page the CLI opens; it redirects back to our loopback server. */
 const STUDIO_PATH = "/studio/auth/cli";
 /** Query parameter that carries the loopback callback URL to the studio page. */
@@ -9674,11 +9948,11 @@ function providerUrl(env = resolveApiEnv()) {
 * @param modelId - exact model id from the live catalog or a caller.
 * @returns the wire dialect to build the request for.
 */
-function wireForModel$1(modelId) {
+function wireForModel$2(modelId) {
 	return /^claude[-/]/i.test(modelId.trim()) ? "anthropic" : "openai";
 }
 /**
-* Reasoning levels one model advertises.
+* Reasoning levels one model advertises, in DSH's vocabulary.
 *
 * The registry is the authority: a model it describes but gives no
 * `reasoningEfforts` is a non-reasoning model, which is why that answers with an
@@ -9686,9 +9960,42 @@ function wireForModel$1(modelId) {
 * describe declares nothing either — guessing from the family name is exactly
 * how `deepseek-v4-flash` got treated as a reasoner with image input when it is
 * text-only with a different effort set.
+*
+* One level is translated rather than copied. Command Code names the
+* "do not think" level `off`; DSH names the same level `none` (as the Kimi and
+* Anthropic lines here already do). Both spellings are accepted on the way in —
+* a stored selection or a session started elsewhere may carry either — but the
+* adapter only ever advertises `none`, so the picker shows the level DSH
+* understands.
+*
+* `off` is also what the CLI's `thinkingHook` keys on to send no reasoning
+* field at all, and the OpenAI-family wire has no such level, so
+* {@link wireReasoningEffort} drops it before the request is built rather than
+* transmitting a value the endpoint would reject.
 */
 function reasoningEffortsFor$1(modelId) {
-	return [...commandCodeModelDef(modelId)?.reasoningEfforts ?? []];
+	return (commandCodeModelDef(modelId)?.reasoningEfforts ?? []).map((effort) => effort === "off" ? DISABLED_THINKING_EFFORT : effort);
+}
+/**
+* The level that means "think as little as possible" on this route.
+*
+* DSH's spelling, matching the Kimi and Claude lines. Kept as a constant
+* because three separate places compare against it — the catalog translation,
+* the wire filter below, and the Anthropic thinking budget.
+*/
+const DISABLED_THINKING_EFFORT = "none";
+/**
+* Whether one effort survives onto the provider wire.
+*
+* `none` is a real DSH level but not a Command Code one: it is expressed by
+* OMITTING the field, which is exactly what the official CLI does when its
+* effort is `off`. Every other level this route advertises passes through.
+*/
+function wireReasoningEffort(effort) {
+	if (effort === void 0 || effort === null) return void 0;
+	const value = String(effort).trim();
+	if (value === "" || value === "none" || value === "off") return void 0;
+	return value;
 }
 /**
 * Accepted request modalities for one model.
@@ -9776,7 +10083,7 @@ function maxOutputTokensFor$4(modelId) {
 * from the registry rather than hand-listed, so a model cannot drift between the
 * offline fallback and its real capability entry.
 */
-const FALLBACK_MODELS$3 = COMMAND_CODE_MODELS.filter((model) => model.contextWindow !== null).map((model) => ({
+const FALLBACK_MODELS$4 = COMMAND_CODE_MODELS.filter((model) => model.contextWindow !== null).map((model) => ({
 	id: model.id,
 	name: model.name,
 	contextWindow: model.contextWindow
@@ -9785,6 +10092,7 @@ const FALLBACK_MODELS$3 = COMMAND_CODE_MODELS.filter((model) => model.contextWin
 //#region src/shared/command-code-contracts.ts
 /** Every level, in escalating order; the settings card renders exactly these. */
 const COMMAND_CODE_REASONING_EFFORTS = [
+	"none",
 	"minimal",
 	"low",
 	"medium",
@@ -9799,7 +10107,7 @@ const COMMAND_CODE_PREFERENCES_NAMESPACE = "dsh-command-code";
 function isReasoningEffort$3(value) {
 	return typeof value === "string" && COMMAND_CODE_REASONING_EFFORTS.includes(value);
 }
-const DEFAULT_ENABLED_MODEL_IDS$4 = FALLBACK_MODELS$3.map((model) => model.id);
+const DEFAULT_ENABLED_MODEL_IDS$4 = FALLBACK_MODELS$4.map((model) => model.id);
 /**
 * Bind the model selection to the DSH settings document when the harness still
 * offers one, and to the JSON file beside it otherwise: a harness without the
@@ -9865,7 +10173,7 @@ function credentialPath$1() {
 function modelSettingsPath$1() {
 	return path.join(dshHomeDir$2(), "storages", "command-code-models.json");
 }
-function optionalString$6(record, key) {
+function optionalString$8(record, key) {
 	const value = record[key];
 	if (value === void 0 || value === null) return void 0;
 	if (typeof value !== "string") throw new Error("Command Code credential payload is invalid");
@@ -9886,7 +10194,7 @@ function parseCommandCodeCredentials(value) {
 		"planLabel",
 		"planId"
 	]) {
-		const parsed = optionalString$6(record, key);
+		const parsed = optionalString$8(record, key);
 		if (parsed !== void 0) credentials[key] = parsed;
 	}
 	const authenticatedAt = record.authenticatedAt;
@@ -9901,13 +10209,13 @@ function parseCommandCodeCredentials(value) {
 	}
 	return credentials;
 }
-function credentialAccount$2(filePath) {
+function credentialAccount$3(filePath) {
 	return createHash("sha256").update(path.resolve(filePath)).digest("hex");
 }
-function createCredentialBackend$2(filePath) {
+function createCredentialBackend$3(filePath) {
 	if (process.platform === "win32") return new WindowsDpapiCredentialStore(`${filePath}.dpapi`, parseCommandCodeCredentials);
-	if (process.platform === "darwin") return new MacKeychainCredentialStore(PROVIDER_ID$4, credentialAccount$2(filePath), parseCommandCodeCredentials);
-	if (process.platform === "linux") return new SecretServiceCredentialStore(PROVIDER_ID$4, credentialAccount$2(filePath), parseCommandCodeCredentials);
+	if (process.platform === "darwin") return new MacKeychainCredentialStore(PROVIDER_ID$5, credentialAccount$3(filePath), parseCommandCodeCredentials);
+	if (process.platform === "linux") return new SecretServiceCredentialStore(PROVIDER_ID$5, credentialAccount$3(filePath), parseCommandCodeCredentials);
 	throw new Error("Command Code credential storage requires Windows, macOS, or Linux.");
 }
 const credentialOperations$2 = /* @__PURE__ */ new Map();
@@ -9915,13 +10223,13 @@ const credentialOperations$2 = /* @__PURE__ */ new Map();
 var FileCredentialStore$1 = class {
 	filePath;
 	backend;
-	constructor(filePath = credentialPath$1(), backend = createCredentialBackend$2(filePath)) {
+	constructor(filePath = credentialPath$1(), backend = createCredentialBackend$3(filePath)) {
 		this.filePath = filePath;
 		this.backend = backend;
 	}
 	path() {
 		if (process.platform === "win32") return `${this.filePath}.dpapi`;
-		return `${process.platform === "darwin" ? "Keychain" : "Secret Service"}: ${PROVIDER_ID$4}/${credentialAccount$2(this.filePath)}`;
+		return `${process.platform === "darwin" ? "Keychain" : "Secret Service"}: ${PROVIDER_ID$5}/${credentialAccount$3(this.filePath)}`;
 	}
 	serialize(operation) {
 		const key = path.resolve(this.filePath);
@@ -10136,7 +10444,7 @@ function commandCodeHeaders(apiKey, extra = {}) {
 		authorization: `Bearer ${apiKey}`,
 		"content-type": "application/json",
 		accept: "application/json",
-		"user-agent": `${PROVIDER_ID$4}/${CLI_VERSION}`,
+		"user-agent": `${PROVIDER_ID$5}/${CLI_VERSION}`,
 		[HEADER_CLI_VERSION]: CLI_VERSION,
 		[HEADER_CLI_ENVIRONMENT]: process.platform === "win32" ? "windows" : process.platform === "darwin" ? "macos" : "linux",
 		[HEADER_PROJECT_SLUG]: "dsh-chatgpt-subscription",
@@ -10172,7 +10480,7 @@ function firstString$1(record, keys) {
 		if (value !== void 0) return value;
 	}
 }
-function firstNumber$1(record, keys) {
+function firstNumber$2(record, keys) {
 	for (const key of keys) {
 		const value = asNumber$3(record[key]);
 		if (value !== void 0) return value;
@@ -10344,12 +10652,12 @@ function parseWindowLimits(payload, consumed = []) {
 		if (key === "limited" || key === "exceeded") continue;
 		const record = asRecord$9(limits[key]);
 		if (record === void 0) continue;
-		const cap = firstNumber$1(record, [
+		const cap = firstNumber$2(record, [
 			"cap",
 			"limit",
 			"total"
 		]);
-		const used = firstNumber$1(record, ["used", "consumed"]);
+		const used = firstNumber$2(record, ["used", "consumed"]);
 		if (cap === void 0 && used === void 0) continue;
 		consumed.push(record);
 		const usedFraction = cap !== void 0 && cap > 0 && used !== void 0 ? clamp01$2(used / cap) : null;
@@ -10379,9 +10687,9 @@ function parseCreditBalances(payload, consumed = []) {
 	const root = asRecord$9(payload) ?? {};
 	const credits = asRecord$9(deepValue(asRecord$9(root.data) ?? root, ["credits"]));
 	if (credits === void 0) return [];
-	const monthly = firstNumber$1(credits, ["monthlyCredits", "monthly_credits"]);
-	const purchased = firstNumber$1(credits, ["purchasedCredits", "purchased_credits"]);
-	const free = firstNumber$1(credits, ["freeCredits", "free_credits"]);
+	const monthly = firstNumber$2(credits, ["monthlyCredits", "monthly_credits"]);
+	const purchased = firstNumber$2(credits, ["purchasedCredits", "purchased_credits"]);
+	const free = firstNumber$2(credits, ["freeCredits", "free_credits"]);
 	if (monthly === void 0 && purchased === void 0 && free === void 0) return [];
 	consumed.push(credits);
 	const meters = [];
@@ -10460,27 +10768,27 @@ function sweepMeters(payload, skipIds, consumed) {
 	let anonymized = 0;
 	for (const record of records) {
 		if (consumed.has(record)) continue;
-		const limit = firstNumber$1(record, [
+		const limit = firstNumber$2(record, [
 			"limit",
 			"quota",
 			"allowance",
 			"cap",
 			"total"
 		]);
-		const used = firstNumber$1(record, [
+		const used = firstNumber$2(record, [
 			"used",
 			"consumed",
 			"spent"
 		]);
-		const remainingRaw = firstNumber$1(record, ["remaining", "left"]);
-		const balance = firstNumber$1(record, ["balance", "credits"]);
-		const usedPercentRaw = firstNumber$1(record, [
+		const remainingRaw = firstNumber$2(record, ["remaining", "left"]);
+		const balance = firstNumber$2(record, ["balance", "credits"]);
+		const usedPercentRaw = firstNumber$2(record, [
 			"usedPercent",
 			"used_percent",
 			"usagePercent",
 			"percentUsed"
 		]);
-		const remainingPercentRaw = firstNumber$1(record, [
+		const remainingPercentRaw = firstNumber$2(record, [
 			"remainingPercent",
 			"remaining_percent",
 			"percentRemaining"
@@ -10604,7 +10912,7 @@ function parseUsageWindows$1(payload) {
 	const windows = [];
 	const seen = /* @__PURE__ */ new Set();
 	for (const record of records) {
-		const percent = firstNumber$1(record, [
+		const percent = firstNumber$2(record, [
 			"usedPercent",
 			"used_percent",
 			"usagePercent",
@@ -10631,7 +10939,7 @@ function parseUsageWindows$1(payload) {
 				"window"
 			]) ?? id,
 			usedPercent: Math.round(clamp01$2(normalizePercent(percent) ?? 0) * 100),
-			windowDurationMins: firstNumber$1(record, [
+			windowDurationMins: firstNumber$2(record, [
 				"windowDurationMins",
 				"window_minutes",
 				"durationMins",
@@ -10707,7 +11015,7 @@ function parseProviderModels(payload) {
 		models.push({
 			id,
 			name: firstString$1(record, ["displayName", "name"]) ?? id,
-			contextWindow: firstNumber$1(record, [
+			contextWindow: firstNumber$2(record, [
 				"context_length",
 				"contextLength",
 				"context_window",
@@ -10725,7 +11033,7 @@ async function fetchProviderModels(options = {}) {
 	const response = await (options.fetchFn ?? fetch)(`${providerUrl(options.apiEnv ?? resolveApiEnv())}/models`, {
 		headers: {
 			accept: "application/json",
-			"user-agent": PROVIDER_NAME$4
+			"user-agent": PROVIDER_NAME$5
 		},
 		signal: timeoutSignal$3(options.signal, DISCOVERY_TIMEOUT_MS$3)
 	});
@@ -10948,7 +11256,7 @@ function buildModelOptions$2(catalog, enabledModelIds, overrides) {
 			contextWindow: overrides[model.id] && overrides[model.id] > 0 ? overrides[model.id] : contextWindow,
 			defaultMaxTokens: maxOutputTokensFor$4(model.id),
 			...efforts.length > 0 ? { reasoningEfforts: efforts } : {},
-			wire: wireForModel$1(model.id)
+			wire: wireForModel$2(model.id)
 		};
 	});
 }
@@ -11137,14 +11445,14 @@ function imageBlockToInline$3(block, images) {
 		data: resolved.data
 	} : void 0;
 }
-function textOf$4(content) {
+function textOf$5(content) {
 	if (typeof content === "string") return sanitizeText$4(content);
 	if (!Array.isArray(content)) return "";
 	const parts = [];
 	for (const block of content) {
 		if (!isRecord$18(block)) continue;
 		if (block.type === "text" && typeof block.text === "string") parts.push(sanitizeText$4(block.text));
-		else if (block.type === "tool-result") parts.push(textOf$4(block.content));
+		else if (block.type === "tool-result") parts.push(textOf$5(block.content));
 	}
 	return parts.join("");
 }
@@ -11252,7 +11560,7 @@ function leadingSystemText$4(options) {
 	if (typeof options.system === "string" && options.system.trim() !== "") parts.push(options.system);
 	for (const message of options.messages) {
 		if (message.role !== "system") continue;
-		const text = textOf$4(message.content);
+		const text = textOf$5(message.content);
 		if (text !== "") parts.push(text);
 	}
 	return parts.length === 0 ? void 0 : parts.join("\n\n");
@@ -11372,7 +11680,7 @@ function buildOpenAIRequest$1(options, images = NO_RESOLVED_IMAGES$4) {
 			content
 		});
 	}
-	const effort = options.reasoningEffort === void 0 ? void 0 : String(options.reasoningEffort);
+	const effort = wireReasoningEffort(options.reasoningEffort === void 0 ? void 0 : String(options.reasoningEffort));
 	const maxTokens = options.maxTokens ?? maxOutputTokensFor$4(options.model);
 	return {
 		model: options.model,
@@ -11506,7 +11814,7 @@ function buildAnthropicRequest$1(options, images = NO_RESOLVED_IMAGES$4) {
 	}
 	const system = leadingSystemText$4(options);
 	const maxTokens = options.maxTokens ?? maxOutputTokensFor$4(options.model);
-	const budget = thinkingBudgetFor$1(options.reasoningEffort === void 0 ? void 0 : String(options.reasoningEffort), maxTokens);
+	const budget = thinkingBudgetFor$1(wireReasoningEffort(options.reasoningEffort === void 0 ? void 0 : String(options.reasoningEffort)), maxTokens);
 	return {
 		model: options.model,
 		max_tokens: maxTokens,
@@ -11529,7 +11837,7 @@ function buildAnthropicRequest$1(options, images = NO_RESOLVED_IMAGES$4) {
 function buildRequest$1(options, wire, images = NO_RESOLVED_IMAGES$4) {
 	return wire === "anthropic" ? buildAnthropicRequest$1(options, images) : buildOpenAIRequest$1(options, images);
 }
-function createStreamState$4(wire) {
+function createStreamState$5(wire) {
 	return {
 		wire,
 		blocks: [],
@@ -11565,7 +11873,7 @@ function closeCurrent$3(state) {
 		block
 	}];
 }
-function closeToolCalls$3(state) {
+function closeToolCalls$4(state) {
 	const out = [];
 	for (const [wireIndex, call] of [...state.toolCalls.entries()].sort((a, b) => a[0] - b[0])) {
 		const block = {
@@ -11610,7 +11918,7 @@ function processOpenAIStreamLine$1(line, state) {
 	const payload = trimmed.slice(5).trim();
 	if (payload === "[DONE]") {
 		state.done = true;
-		return closeStream$3(state);
+		return closeStream$4(state);
 	}
 	if (payload === "") return [];
 	const chunk = safeJsonParse$4(payload);
@@ -11633,7 +11941,7 @@ function processOpenAIStreamLine$1(line, state) {
 	if (delta) {
 		const reasoning = asString$11(delta.reasoning_content) ?? asString$11(delta.reasoning);
 		if (reasoning !== void 0 && reasoning !== "") {
-			out.push(...closeToolCalls$3(state));
+			out.push(...closeToolCalls$4(state));
 			if (state.current === null || state.current.type !== "reasoning") out.push(...openTextBlock$3(state, "reasoning"));
 			state.current.text += sanitizeText$4(reasoning);
 			state.hasContent = true;
@@ -11645,7 +11953,7 @@ function processOpenAIStreamLine$1(line, state) {
 		}
 		const content = asString$11(delta.content);
 		if (content !== void 0 && content !== "") {
-			out.push(...closeToolCalls$3(state));
+			out.push(...closeToolCalls$4(state));
 			if (state.current === null || state.current.type !== "text") out.push(...openTextBlock$3(state, "text"));
 			state.current.text += sanitizeText$4(content);
 			state.hasContent = true;
@@ -11665,7 +11973,7 @@ function processOpenAIStreamLine$1(line, state) {
 	if (finish !== void 0 && finish !== "") {
 		state.finishReason = finish;
 		out.push(...closeCurrent$3(state));
-		out.push(...closeToolCalls$3(state));
+		out.push(...closeToolCalls$4(state));
 	}
 	return out;
 }
@@ -11883,7 +12191,7 @@ function processAnthropicStreamLine$1(line, state) {
 	}
 	if (type === "message_stop") {
 		state.done = true;
-		return closeStream$3(state);
+		return closeStream$4(state);
 	}
 	if (type === "error") throw new LlmError$1(`Command Code stream error: ${asString$11((isRecord$18(event.error) ? event.error : {}).message) ?? "unknown error"}`, "PROVIDER_ERROR");
 	return out;
@@ -11904,10 +12212,10 @@ function finishReasonFor$4(state) {
 	return { kind: "stop" };
 }
 /** Flush every open block, then emit usage and the terminal finish. */
-function closeStream$3(state) {
+function closeStream$4(state) {
 	if (state.finished) return [];
 	state.finished = true;
-	const out = [...closeCurrent$3(state), ...closeToolCalls$3(state)];
+	const out = [...closeCurrent$3(state), ...closeToolCalls$4(state)];
 	if (state.sawUsage) out.push({
 		type: "usage",
 		usage: tokenUsage$4(state)
@@ -11937,7 +12245,7 @@ function assertStreamComplete$4(state) {
 * the set: `INVALID_CREDENTIAL` (a rejected key fails identically on every
 * attempt) and `ABORTED` (the caller already cancelled).
 */
-const RETRY_POLICY$4 = resolveRetryPolicy({
+const RETRY_POLICY$5 = resolveRetryPolicy({
 	mode: "normal",
 	maxRetries: 3,
 	retryableCodes: [
@@ -11957,7 +12265,7 @@ function resolveDefaultReasoningEffort$3(efforts, configuredEffort) {
 	if (configuredEffort && efforts.includes(configuredEffort)) return ReasoningEffortId(configuredEffort);
 }
 /** Cooldown one rate-limited key takes when the provider states no delay. */
-const POOL_COOLDOWN_MS$3 = 15 * 6e4;
+const POOL_COOLDOWN_MS$4 = 15 * 6e4;
 var CommandCodeAdapter = class extends LlmAdapter {
 	store;
 	modelSettings;
@@ -11983,11 +12291,11 @@ var CommandCodeAdapter = class extends LlmAdapter {
 	providerInfo(provider) {
 		return {
 			id: provider,
-			name: PROVIDER_NAME$4
+			name: PROVIDER_NAME$5
 		};
 	}
 	providerRetryPolicy() {
-		return RETRY_POLICY$4;
+		return RETRY_POLICY$5;
 	}
 	imageRequestPricing() {}
 	settings() {
@@ -12003,7 +12311,7 @@ var CommandCodeAdapter = class extends LlmAdapter {
 			apiEnv: resolveApiEnv()
 		})))().catch(() => []);
 		if (live.length > 0) return live;
-		return FALLBACK_MODELS$3.map((model) => ({
+		return FALLBACK_MODELS$4.map((model) => ({
 			id: model.id,
 			name: model.name,
 			contextWindow: model.contextWindow
@@ -12039,7 +12347,7 @@ var CommandCodeAdapter = class extends LlmAdapter {
 			name: entry?.name ?? modelId,
 			inputModalities: inputModalitiesFor$1(modelId),
 			context: { contextWindow: this.contextWindowFor(modelId, entry, settings.contextWindowOverrides) },
-			defaultMaxTokens: maxOutputTokensFor$4(modelId),
+			...outputReservation(this.contextWindowFor(modelId, entry, settings.contextWindowOverrides), maxOutputTokensFor$4(modelId)),
 			...efforts.length === 0 ? {} : { reasoning: {
 				efforts: efforts.map((effort) => ({
 					id: ReasoningEffortId(effort),
@@ -12062,11 +12370,17 @@ var CommandCodeAdapter = class extends LlmAdapter {
 			...options,
 			reasoningEffort: ReasoningEffortId(String(effort))
 		};
-		yield* wrapStreamWithWatchdog((watchdogSignal) => this.requestStream(effectiveOptions, watchdogSignal), options.signal, STREAM_IDLE_TIMEOUT_MS$4, STREAM_IDLE_TIMEOUT_CODE$4, PROVIDER_NAME$4);
+		const advertised = reasoningEffortsFor$1(options.model);
+		const effectiveEffort = effectiveOptions.reasoningEffort === void 0 ? void 0 : String(effectiveOptions.reasoningEffort);
+		const resolvedOptions = effectiveEffort === void 0 || advertised.length === 0 || advertised.includes(effectiveEffort) ? effectiveOptions : {
+			...effectiveOptions,
+			reasoningEffort: void 0
+		};
+		yield* wrapStreamWithWatchdog((watchdogSignal) => this.requestStream(resolvedOptions, watchdogSignal), options.signal, STREAM_IDLE_TIMEOUT_MS$4, STREAM_IDLE_TIMEOUT_CODE$4, PROVIDER_NAME$5);
 	}
 	async *requestStream(options, signal) {
 		const fetchFn = this.options.fetchFn ?? fetch;
-		const wire = wireForModel$1(options.model);
+		const wire = wireForModel$2(options.model);
 		const requestOptions = offloadOldestRequestImages$3(normalizeGenerateOptions(options));
 		const images = await resolveRequestImages$3(requestOptions, this.options.attachments, signal);
 		const body = JSON.stringify(buildRequest$1(requestOptions, wire, images));
@@ -12080,7 +12394,7 @@ var CommandCodeAdapter = class extends LlmAdapter {
 			let apiEnv;
 			if (pool === null) {
 				const credentials = await this.store.read();
-				if (credentials === null) throw new LlmError(`Not signed in to ${PROVIDER_NAME$4}. Sign in from Settings > Command Code, or paste an API key there.`, "MISSING_CREDENTIAL");
+				if (credentials === null) throw new LlmError(`Not signed in to ${PROVIDER_NAME$5}. Sign in from Settings > Command Code, or paste an API key there.`, "MISSING_CREDENTIAL");
 				apiKey = credentials.apiKey;
 				apiEnv = credentials.apiEnv ?? resolveApiEnv();
 			} else {
@@ -12116,13 +12430,13 @@ var CommandCodeAdapter = class extends LlmAdapter {
 			detail = (await response.text().catch(() => "")).slice(0, 600);
 			if (pool === null || accountId === void 0) break;
 			if (response.status === 401 || response.status === 403) {
-				await pool.markAuthFailed(accountId, `${PROVIDER_NAME$4} rejected the stored API key (${response.status}).`, "invalid").catch(() => void 0);
+				await pool.markAuthFailed(accountId, `${PROVIDER_NAME$5} rejected the stored API key (${response.status}).`, "invalid").catch(() => void 0);
 				if (await pool.hasAnotherAvailableAccount(tried)) continue;
 				break;
 			}
 			if (response.status === 429) {
 				const after = retryAfterMs$1(response.headers);
-				await pool.markCooldown(accountId, after ?? POOL_COOLDOWN_MS$3, `${PROVIDER_NAME$4} 429`).catch(() => void 0);
+				await pool.markCooldown(accountId, after ?? POOL_COOLDOWN_MS$4, `${PROVIDER_NAME$5} 429`).catch(() => void 0);
 				if (await pool.hasAnotherAvailableAccount(tried)) continue;
 				break;
 			}
@@ -12130,21 +12444,21 @@ var CommandCodeAdapter = class extends LlmAdapter {
 		}
 		if (response === void 0 || !response.ok) {
 			const status = response?.status ?? 500;
-			if (status === 401 || status === 403) throw new LlmError(`${PROVIDER_NAME$4} rejected the stored API key (${status}). Sign in again from Settings > Command Code.${detail ? ` ${detail}` : ""}`, "INVALID_CREDENTIAL", { status });
+			if (status === 401 || status === 403) throw new LlmError(`${PROVIDER_NAME$5} rejected the stored API key (${status}). Sign in again from Settings > Command Code.${detail ? ` ${detail}` : ""}`, "INVALID_CREDENTIAL", { status });
 			if (status === 429) {
 				const after = response === void 0 ? void 0 : retryAfterMs$1(response.headers);
-				throw new LlmError(`${PROVIDER_NAME$4} rate limit or plan quota reached (429). Check the quota card in Settings > Command Code.${detail ? ` ${detail}` : ""}`, "RATE_LIMIT", {
+				throw new LlmError(`${PROVIDER_NAME$5} rate limit or plan quota reached (429). Check the quota card in Settings > Command Code.${detail ? ` ${detail}` : ""}`, "RATE_LIMIT", {
 					status: 429,
 					...after === void 0 ? {} : { providerRetryAfterMs: after }
 				});
 			}
-			if (status >= 500) throw new LlmError(`${PROVIDER_NAME$4} upstream server error (${status}): ${detail || "No response"}`, "SERVER", { status });
-			throw new LlmError(`${PROVIDER_NAME$4} API error (${status}): ${detail || "No response"}`, "PROVIDER_ERROR", { status });
+			if (status >= 500) throw new LlmError(`${PROVIDER_NAME$5} upstream server error (${status}): ${detail || "No response"}`, "SERVER", { status });
+			throw new LlmError(`${PROVIDER_NAME$5} API error (${status}): ${detail || "No response"}`, "PROVIDER_ERROR", { status });
 		}
 		if (response.body === null) throw new LlmError("Command Code returned an empty response body", "PROVIDER_ERROR");
 		const reader = response.body.getReader();
 		const decoder = new TextDecoder();
-		const state = createStreamState$4(wire);
+		const state = createStreamState$5(wire);
 		let buffer = "";
 		try {
 			while (true) {
@@ -12162,7 +12476,7 @@ var CommandCodeAdapter = class extends LlmAdapter {
 			if (buffer.trim() !== "") for (const line of buffer.split("\n")) for (const chunk of processLine$1(line, state, wire)) yield chunk;
 			if (state.finished) return;
 			assertStreamComplete$4(state);
-			for (const chunk of closeStream$3(state)) yield chunk;
+			for (const chunk of closeStream$4(state)) yield chunk;
 		} finally {
 			reader.cancel().catch(() => void 0);
 		}
@@ -12188,7 +12502,7 @@ function commandCodeAccountKey(credentials) {
 	if (credentials.userId !== void 0 && credentials.userId !== "") return `${credentials.userId}:${credentials.keyName ?? ""}`;
 	return createHash("sha256").update(credentials.apiKey).digest("hex");
 }
-function optionalString$5(record, key) {
+function optionalString$7(record, key) {
 	const value = record[key];
 	if (value === void 0 || value === null) return void 0;
 	if (typeof value !== "string") throw new Error("Command Code pool account field is invalid");
@@ -12220,7 +12534,7 @@ function parseCommandCodePoolData(value) {
 			"planId",
 			"userId"
 		]) {
-			const parsed = optionalString$5(raw, key);
+			const parsed = optionalString$7(raw, key);
 			if (parsed !== void 0) account[key] = parsed;
 		}
 		if (account.email === void 0 && credentials.email !== void 0) account.email = credentials.email;
@@ -12252,7 +12566,7 @@ var CommandCodeAccountPool = class extends AccountPoolCore {
 	constructor(options = {}) {
 		const store = options.store ?? new FileCredentialStore$1();
 		const hooks = {
-			providerId: PROVIDER_ID$4,
+			providerId: PROVIDER_ID$5,
 			displayName: "Command Code",
 			poolFile: commandCodePoolPath(),
 			keychainService: "dsh-command-code-pool",
@@ -12418,7 +12732,7 @@ function applyCors(request, response) {
 	response.setHeader("Access-Control-Allow-Methods", "GET, POST, OPTIONS");
 	response.setHeader("Access-Control-Allow-Headers", "Content-Type");
 }
-function sendJson$6(response, status, body) {
+function sendJson$7(response, status, body) {
 	response.writeHead(status, { "content-type": "application/json" });
 	response.end(JSON.stringify(body));
 }
@@ -12611,7 +12925,7 @@ function createAuthServer(port, expectedState, options = {}) {
 			try {
 				fields = fieldsFromPayload(raw, contentType);
 			} catch {
-				sendJson$6(response, 400, {
+				sendJson$7(response, 400, {
 					success: false,
 					error: "Invalid JSON"
 				});
@@ -12619,7 +12933,7 @@ function createAuthServer(port, expectedState, options = {}) {
 			}
 			if (fields.error) {
 				if (fields.state !== expectedState) {
-					sendJson$6(response, 403, {
+					sendJson$7(response, 403, {
 						success: false,
 						error: "Invalid state token"
 					});
@@ -12631,14 +12945,14 @@ function createAuthServer(port, expectedState, options = {}) {
 			}
 			const apiKey = fields.apiKey;
 			if (!apiKey) {
-				sendJson$6(response, 400, {
+				sendJson$7(response, 400, {
 					success: false,
 					error: "Missing required fields"
 				});
 				return;
 			}
 			if (fields.state !== expectedState) {
-				sendJson$6(response, 403, {
+				sendJson$7(response, 403, {
 					success: false,
 					error: "Invalid state token"
 				});
@@ -12788,18 +13102,18 @@ function isCommandCodeReasoningEffort(value) {
 	return typeof value === "string" && COMMAND_CODE_REASONING_EFFORTS.includes(value);
 }
 const MAX_BODY_BYTES$5 = 64 * 1024;
-const ROUTE_PREFIX$3 = "/command-code/api";
-function sendJson$5(response, status, body) {
+const ROUTE_PREFIX$4 = "/command-code/api";
+function sendJson$6(response, status, body) {
 	response.writeHead(status, { "Content-Type": "application/json" });
 	response.end(JSON.stringify(body));
 }
-function sendMethodNotAllowed$4(response) {
-	sendJson$5(response, 405, {
+function sendMethodNotAllowed$5(response) {
+	sendJson$6(response, 405, {
 		ok: false,
 		error: "Method Not Allowed"
 	});
 }
-async function readRequestJson$5(request) {
+async function readRequestJson$6(request) {
 	return new Promise((resolve, reject) => {
 		const chunks = [];
 		let total = 0;
@@ -12824,7 +13138,7 @@ async function readRequestJson$5(request) {
 	});
 }
 function fallbackCatalog$1() {
-	return FALLBACK_MODELS$3.map((model) => ({
+	return FALLBACK_MODELS$4.map((model) => ({
 		id: model.id,
 		name: model.name,
 		contextWindow: model.contextWindow
@@ -12841,7 +13155,7 @@ function fallbackCatalog$1() {
 function resolveEnabledModelIds$4(stored, catalog, enabled = true) {
 	if (!enabled) return [];
 	const catalogIds = catalog.map((model) => model.id);
-	const shippedDefaults = new Set(FALLBACK_MODELS$3.map((model) => model.id));
+	const shippedDefaults = new Set(FALLBACK_MODELS$4.map((model) => model.id));
 	if (stored.length > 0 && stored.length === shippedDefaults.size && stored.every((id) => shippedDefaults.has(id))) return catalogIds;
 	const known = new Set(catalogIds);
 	return stored.filter((id) => known.has(id));
@@ -12926,13 +13240,13 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 	const quotaRefresh = new QuotaRefresh();
 	return ctx.webServer.register({
 		kind: "prefix",
-		path: ROUTE_PREFIX$3,
+		path: ROUTE_PREFIX$4,
 		handler: async (request, response) => {
 			const path = new URL(request.url || "/", "http://dsh.local").pathname.replace(/^\/command-code\/api\/?/, "");
 			const method = request.method ?? "GET";
 			try {
 				if (path === "" || path === "status") {
-					if (method !== "GET") return sendMethodNotAllowed$4(response);
+					if (method !== "GET") return sendMethodNotAllowed$5(response);
 					const credentials = await store.read();
 					const targetId = await activeAccountId();
 					const cached = getCachedQuotaFor$1(targetId);
@@ -12941,7 +13255,7 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 						if (cached === void 0) await quotaRefresh.run(refresh);
 						else quotaRefresh.start(refresh);
 					}
-					return sendJson$5(response, 200, {
+					return sendJson$6(response, 200, {
 						ok: true,
 						value: {
 							...await readStatus(),
@@ -12950,12 +13264,12 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 					});
 				}
 				if (path === "login") {
-					if (method !== "POST") return sendMethodNotAllowed$4(response);
-					if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+					if (method !== "POST") return sendMethodNotAllowed$5(response);
+					if (!isSameOriginMutation(request)) return sendJson$6(response, 403, {
 						ok: false,
 						error: "Cross-origin request rejected."
 					});
-					return sendJson$5(response, 200, {
+					return sendJson$6(response, 200, {
 						ok: true,
 						value: await beginWebLogin$3(store, {
 							fetchFn,
@@ -12964,25 +13278,25 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 					});
 				}
 				if (path === "login/status") {
-					if (method !== "GET") return sendMethodNotAllowed$4(response);
-					return sendJson$5(response, 200, {
+					if (method !== "GET") return sendMethodNotAllowed$5(response);
+					return sendJson$6(response, 200, {
 						ok: true,
 						value: getWebLoginStatus()
 					});
 				}
 				if (path === "login/apikey") {
-					if (method !== "POST") return sendMethodNotAllowed$4(response);
-					if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+					if (method !== "POST") return sendMethodNotAllowed$5(response);
+					if (!isSameOriginMutation(request)) return sendJson$6(response, 403, {
 						ok: false,
 						error: "Cross-origin request rejected."
 					});
-					const body = await readRequestJson$5(request);
+					const body = await readRequestJson$6(request);
 					const account = await saveApiKey(store, typeof body.apiKey === "string" ? body.apiKey : "", {
 						fetchFn,
 						...accountPool === void 0 ? {} : { onSave: (credentials) => accountPool.addAccount(credentials) }
 					});
 					clearCachedQuota$2();
-					return sendJson$5(response, 200, {
+					return sendJson$6(response, 200, {
 						ok: true,
 						value: {
 							...await readStatus(),
@@ -12991,13 +13305,13 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 					});
 				}
 				if (path === "connection/test") {
-					if (method !== "POST") return sendMethodNotAllowed$4(response);
-					if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+					if (method !== "POST") return sendMethodNotAllowed$5(response);
+					if (!isSameOriginMutation(request)) return sendJson$6(response, 403, {
 						ok: false,
 						error: "Cross-origin request rejected."
 					});
 					const credentials = await store.read();
-					if (credentials === null) return sendJson$5(response, 400, {
+					if (credentials === null) return sendJson$6(response, 400, {
 						ok: false,
 						error: "Not signed in."
 					});
@@ -13006,7 +13320,7 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 						fetchFn,
 						apiEnv: credentials.apiEnv ?? resolveApiEnv()
 					});
-					return sendJson$5(response, 200, {
+					return sendJson$6(response, 200, {
 						ok: true,
 						value: {
 							connected: true,
@@ -13024,24 +13338,24 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 					});
 				}
 				if (path === "quota") {
-					if (method !== "GET" && method !== "POST") return sendMethodNotAllowed$4(response);
+					if (method !== "GET" && method !== "POST") return sendMethodNotAllowed$5(response);
 					await fetchAccountQuota$2(quotaStore, fetchFn, true, await activeAccountId());
-					return sendJson$5(response, 200, {
+					return sendJson$6(response, 200, {
 						ok: true,
 						value: await readStatus()
 					});
 				}
 				if (path === "models" || path === "settings") {
-					if (method === "GET") return sendJson$5(response, 200, {
+					if (method === "GET") return sendJson$6(response, 200, {
 						ok: true,
 						value: await readStatus()
 					});
-					if (method !== "POST") return sendMethodNotAllowed$4(response);
-					if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+					if (method !== "POST") return sendMethodNotAllowed$5(response);
+					if (!isSameOriginMutation(request)) return sendJson$6(response, 403, {
 						ok: false,
 						error: "Cross-origin request rejected."
 					});
-					const body = await readRequestJson$5(request);
+					const body = await readRequestJson$6(request);
 					const patch = {};
 					if (typeof body.enabled === "boolean") patch.enabled = body.enabled;
 					if (Array.isArray(body.enabledModelIds)) patch.enabledModelIds = body.enabledModelIds.filter((id) => typeof id === "string");
@@ -13058,14 +13372,14 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 					if (preferences) await preferences.update(patch);
 					else await modelSettings.updateSettings(patch);
 					if (patch.enabledModelIds !== void 0 || patch.enabled !== void 0) ctx.emit?.("llm/adapters-updated");
-					return sendJson$5(response, 200, {
+					return sendJson$6(response, 200, {
 						ok: true,
 						value: await readStatus()
 					});
 				}
 				if (path === "catalog/refresh") {
-					if (method !== "POST") return sendMethodNotAllowed$4(response);
-					if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+					if (method !== "POST") return sendMethodNotAllowed$5(response);
+					if (!isSameOriginMutation(request)) return sendJson$6(response, 403, {
 						ok: false,
 						error: "Cross-origin request rejected."
 					});
@@ -13075,7 +13389,7 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 						force: true,
 						apiEnv: resolveApiEnv()
 					});
-					return sendJson$5(response, 200, {
+					return sendJson$6(response, 200, {
 						ok: true,
 						value: await readStatus()
 					});
@@ -13083,7 +13397,7 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 				if (path === "accounts") {
 					if (method === "GET") {
 						const data = accountPool === void 0 ? null : await accountPool.read().catch(() => null);
-						return sendJson$5(response, 200, {
+						return sendJson$6(response, 200, {
 							ok: true,
 							value: {
 								accounts: accountPool === void 0 ? [] : await accountPool.listAccounts().catch(() => []),
@@ -13092,16 +13406,16 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 							}
 						});
 					}
-					if (method !== "POST") return sendMethodNotAllowed$4(response);
-					if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+					if (method !== "POST") return sendMethodNotAllowed$5(response);
+					if (!isSameOriginMutation(request)) return sendJson$6(response, 403, {
 						ok: false,
 						error: "Cross-origin request rejected."
 					});
-					if (accountPool === void 0) return sendJson$5(response, 400, {
+					if (accountPool === void 0) return sendJson$6(response, 400, {
 						ok: false,
 						error: "Account pool is not installed."
 					});
-					const body = await readRequestJson$5(request);
+					const body = await readRequestJson$6(request);
 					const action = typeof body.action === "string" ? body.action : "";
 					const accountId = typeof body.accountId === "string" ? body.accountId : void 0;
 					if (action === "set-primary" && accountId !== void 0) await accountPool.setPrimary(accountId);
@@ -13112,20 +13426,20 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 					else if (action === "relogin" && accountId !== void 0) await accountPool.clearAuthFailed(accountId);
 					clearCachedQuota$2();
 					clearCachedCatalog$3();
-					return sendJson$5(response, 200, {
+					return sendJson$6(response, 200, {
 						ok: true,
 						value: await readStatus()
 					});
 				}
 				if (path === "logout") {
-					if (method !== "POST") return sendMethodNotAllowed$4(response);
-					if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+					if (method !== "POST") return sendMethodNotAllowed$5(response);
+					if (!isSameOriginMutation(request)) return sendJson$6(response, 403, {
 						ok: false,
 						error: "Cross-origin request rejected."
 					});
 					if (accountPool === void 0) await store.delete();
 					else {
-						const body = await readRequestJson$5(request).catch(() => ({}));
+						const body = await readRequestJson$6(request).catch(() => ({}));
 						const data = await accountPool.read().catch(() => null);
 						const target = typeof body.accountId === "string" ? body.accountId : data?.activeAccountId ?? data?.accounts.find((account) => account.isPrimary)?.id ?? data?.accounts[0]?.id;
 						if (target === void 0) await store.delete();
@@ -13133,17 +13447,17 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 					}
 					clearCachedQuota$2();
 					clearCachedCatalog$3();
-					return sendJson$5(response, 200, {
+					return sendJson$6(response, 200, {
 						ok: true,
 						value: await readStatus()
 					});
 				}
-				return sendJson$5(response, 404, {
+				return sendJson$6(response, 404, {
 					ok: false,
 					error: "not-found"
 				});
 			} catch (error) {
-				return sendJson$5(response, 500, {
+				return sendJson$6(response, 500, {
 					ok: false,
 					error: error instanceof Error ? error.message : String(error)
 				});
@@ -13152,6 +13466,1377 @@ function registerCommandCodeRoutes(ctx, store, modelSettings, preferences, optio
 	});
 }
 //#endregion
+//#region src/host/ollama/types.ts
+/**
+* Static facts about the Ollama provider API.
+*
+* Ollama publishes three surfaces, and this line speaks two of them:
+*
+* - the **native API** (`https://ollama.com/api`) — Ollama's own wire format,
+*   with `/api/chat` and `/api/tags`;
+* - the **OpenAI-compatible API** (`https://ollama.com/v1`) — a documented
+*   subset of the OpenAI API, which is what a coding agent wants by default
+*   because its tool-calling and streaming shapes are the familiar ones.
+*
+* A third surface, the Anthropic-compatible `/v1/messages`, exists and is not
+* used here: the Claude line already transcribes the Anthropic wire format
+* against a snapshot it can verify, and routing Ollama's subset of it through
+* the same mapper would claim capabilities Ollama does not document.
+*
+* DOCUMENTED CLOUD LIMITS (docs/api/openai-compatibility, docs/cloud):
+*
+* - No stateful Responses. Only the stateless form works.
+* - No built-in web search through `/v1/responses`.
+* - No custom/freeform tool-call replay.
+*
+* Those are server-side capability limits, not choices made here, so this line
+* sends ordinary `tools` arrays and streams ordinary `tool_calls` back, and
+* makes no promise about replaying a custom tool call across turns. Saying so
+* in the settings card is the honest rendering of a documented boundary.
+*
+* Authentication is a static API key in `Authorization: Bearer`, not an OAuth
+* flow: keys do not expire and are revoked from the account settings page.
+*/
+const PROVIDER_ID$4 = "ollama";
+const PROVIDER_NAME$4 = "Ollama";
+/** Direct cloud access, per docs/api/introduction. */
+const CLOUD_BASE_URL = "https://ollama.com";
+/** Native surface. */
+const NATIVE_CHAT_PATH = "/api/chat";
+const NATIVE_TAGS_PATH = "/api/tags";
+const OPENAI_CHAT_PATH = "/chat/completions";
+/**
+* Default context window for a model whose catalog entry states none.
+*
+* Ollama's `/api/tags` returns names and families, not context windows, so a
+* fixed conservative value is the honest number here. It is deliberately small
+* enough to be true of every cloud model rather than flattering to one.
+*/
+const DEFAULT_CONTEXT_WINDOW$4 = 128e3;
+/**
+* Output cap used when neither the caller nor the catalog states one.
+*
+* Ollama does not document a per-model maximum output length, and unlike the
+* Claude line there is no snapshot to read one from, so this is a request-level
+* ceiling rather than a claim about what the model can do.
+*/
+const DEFAULT_MAX_OUTPUT_TOKENS = 8192;
+/** How long a catalog sync may take before the cached list is used instead. */
+const CATALOG_TIMEOUT_MS = 15e3;
+/**
+* Cooldown one account takes when Ollama refuses it without stating a delay.
+*
+* A 429 here is a real quota or rate limit, not a blip, so the account is parked
+* for long enough that the next request goes to a different one.
+*/
+const POOL_COOLDOWN_MS$3 = 15 * 6e4;
+/**
+* NO fallback model names.
+*
+* An earlier revision hard-coded `gpt-oss:120b-cloud` and `gpt-oss:20b-cloud`
+* here, and they reached the model picker. That was wrong twice over: this line
+* has no model table of its own, so every name in it would be a guess about what
+* some account is entitled to, and a guess that lands in the picker is a model a
+* user can select and then fail on. The other lines can fall back to hard-coded
+* names because theirs are transcribed from a real table - Claude's is its own
+* frozen catalog. Ollama's truth is `/api/tags`, and there is nothing honest to
+* stand in for it before the first sync.
+*
+* So the catalog starts empty and the card says the list has not been synced yet,
+* with the sync button right beside it. A user who has a key is one click from a
+* list that is actually theirs.
+*/
+const FALLBACK_MODELS$3 = Object.freeze([]);
+/**
+* Resolve the wire shape for a model.
+*
+* `openai` first, always: the OpenAI-compatible surface is the one whose tool
+* and streaming semantics this plugin's mapper already implements faithfully,
+* and it is what an agent wants. `native` is the fallback for the documented
+* cases where the OpenAI surface refuses a request, so a model the service
+* accepts natively still works instead of failing outright.
+*/
+function wireForModel$1(_model) {
+	return "openai";
+}
+/** Context window for a model, falling back to the conservative constant. */
+function contextWindowFor(model) {
+	const stated = model?.contextWindow;
+	return typeof stated === "number" && Number.isFinite(stated) && stated > 0 ? Math.floor(stated) : DEFAULT_CONTEXT_WINDOW$4;
+}
+/** Bearer headers for a stored key. */
+function ollamaHeaders(apiKey) {
+	return {
+		authorization: `Bearer ${apiKey}`,
+		"content-type": "application/json"
+	};
+}
+//#endregion
+//#region src/host/ollama/token-store.ts
+function credentialPath$4() {
+	return path.join(dshHomeDir$2(), "storages", "ollama-credentials.json");
+}
+function modelSettingsPath$6() {
+	return path.join(dshHomeDir$2(), "storages", "ollama-models.json");
+}
+/** Stable keyring account name for one credential file. */
+function credentialAccount$2(filePath) {
+	return createHash("sha256").update(path.resolve(filePath)).digest("hex");
+}
+function optionalString$6(record, key) {
+	const value = record[key];
+	if (value === void 0 || value === null) return void 0;
+	if (typeof value !== "string") throw new Error("Ollama credential payload is invalid");
+	return value;
+}
+function parseOllamaCredentials(value) {
+	if (typeof value !== "object" || value === null || Array.isArray(value)) throw new Error("Ollama credential payload is invalid");
+	const record = value;
+	const apiKey = record.apiKey;
+	if (typeof apiKey !== "string" || apiKey.trim() === "") throw new Error("Ollama credential is missing its API key");
+	const credentials = { apiKey };
+	const alias = optionalString$6(record, "alias");
+	if (alias !== void 0) credentials.alias = alias;
+	const addedAt = record.addedAt;
+	if (typeof addedAt === "number" && Number.isFinite(addedAt)) credentials.addedAt = addedAt;
+	return credentials;
+}
+/**
+* The OS credential backend for this platform.
+*
+* Same matrix as every other line here: DPAPI on Windows, Keychain on macOS,
+* Secret Service on Linux. The plaintext JSON file is a migration source only,
+* never the live store.
+*/
+function createCredentialBackend$2(filePath) {
+	if (process.platform === "win32") return new WindowsDpapiCredentialStore(`${filePath}.dpapi`, parseOllamaCredentials);
+	if (process.platform === "darwin") return new MacKeychainCredentialStore(PROVIDER_ID$4, credentialAccount$2(filePath), parseOllamaCredentials);
+	if (process.platform === "linux") return new SecretServiceCredentialStore(PROVIDER_ID$4, credentialAccount$2(filePath), parseOllamaCredentials);
+	throw new Error("Ollama credential storage requires Windows, macOS, or Linux.");
+}
+/** Encrypted credential store for the single-key (no pool) configuration. */
+var FileCredentialStore$5 = class {
+	filePath;
+	backend;
+	constructor(filePath = credentialPath$4(), backend = createCredentialBackend$2(filePath)) {
+		this.filePath = filePath;
+		this.backend = backend;
+	}
+	path() {
+		return this.filePath;
+	}
+	async read() {
+		return this.backend.load();
+	}
+	async write(credentials) {
+		await fs.mkdir(path.dirname(this.filePath), { recursive: true });
+		await this.backend.save(credentials);
+	}
+	async clear() {
+		await this.backend.clear();
+	}
+};
+/**
+* Model settings, bound to the DSH settings document when the harness offers one
+* and to a JSON file beside it otherwise.
+*/
+var FileModelSettingsStore$5 = class {
+	filePath;
+	cached = null;
+	loadPromise = null;
+	constructor(filePath = modelSettingsPath$6()) {
+		this.filePath = filePath;
+	}
+	path() {
+		return this.filePath;
+	}
+	status() {
+		return this.cached ?? {
+			enabled: true,
+			enabledModelIds: [],
+			catalogModels: [],
+			defaultReasoningEffort: null
+		};
+	}
+	async read() {
+		if (this.cached !== null) return this.cached;
+		if (this.loadPromise !== null) return this.loadPromise;
+		this.loadPromise = this.load().finally(() => {
+			this.loadPromise = null;
+		});
+		return this.loadPromise;
+	}
+	async load() {
+		try {
+			const content = await fs.readFile(this.filePath, "utf8");
+			const parsed = JSON.parse(content);
+			if (typeof parsed === "object" && parsed !== null) {
+				const record = parsed;
+				const enabledModels = Array.isArray(record.enabledModelIds) ? record.enabledModelIds.filter((id) => typeof id === "string") : [];
+				const catalogModels = Array.isArray(record.catalogModels) ? record.catalogModels.filter((entry) => typeof entry === "object" && entry !== null && typeof entry.id === "string") : [];
+				this.cached = {
+					enabled: record.enabled !== false,
+					enabledModelIds: enabledModels,
+					catalogModels,
+					defaultReasoningEffort: typeof record.defaultReasoningEffort === "string" ? record.defaultReasoningEffort : null
+				};
+				return this.cached;
+			}
+		} catch {}
+		this.cached = {
+			enabled: true,
+			enabledModelIds: [],
+			catalogModels: [],
+			defaultReasoningEffort: null
+		};
+		return this.cached;
+	}
+	async write(settings) {
+		await fs.mkdir(path.dirname(this.filePath), { recursive: true });
+		const tmp = `${this.filePath}.tmp.${Date.now()}`;
+		await fs.writeFile(tmp, JSON.stringify(settings, null, 2), "utf8");
+		await fs.rename(tmp, this.filePath);
+		this.cached = settings;
+	}
+	async update(patch) {
+		const current = await this.read();
+		const next = {
+			enabled: patch.enabled ?? current.enabled,
+			enabledModelIds: patch.enabledModelIds ?? current.enabledModelIds,
+			catalogModels: current.catalogModels,
+			defaultReasoningEffort: patch.defaultReasoningEffort === void 0 ? current.defaultReasoningEffort : patch.defaultReasoningEffort
+		};
+		await this.write(next);
+		return next;
+	}
+	/** Record a catalog sync without touching the user's enabled selection. */
+	async storeCatalog(models) {
+		const next = {
+			...await this.read(),
+			catalogModels: models
+		};
+		await this.write(next);
+		return next;
+	}
+};
+//#endregion
+//#region src/host/ollama/client.ts
+/** Headers both surfaces require, per docs/api/authentication. */
+function headersFor(credentials, accept) {
+	return {
+		...ollamaHeaders(credentials.apiKey),
+		accept
+	};
+}
+/**
+* Build the request body for one wire shape.
+*
+* The two surfaces are not translations of each other — the native API nests the
+* assistant's tool calls under `message.tool_calls` and expects results under a
+* `tool` role without ids, while the OpenAI surface uses `tool_calls` and
+* `tool_call_id`. Mapping them through one shared builder would have to guess at
+* the fields the other surface does not carry, so each is written out where its
+* own rules stay visible.
+*/
+function buildBody(wire, request) {
+	if (wire === "openai") return buildOpenAIBody(request);
+	return buildNativeBody(request);
+}
+function buildOpenAIBody(request) {
+	const messages = request.messages.map((message) => {
+		if (message.role === "tool") return {
+			role: "tool",
+			tool_call_id: message.toolCallId ?? "",
+			content: message.content
+		};
+		if (message.role === "assistant" && message.toolCalls !== void 0 && message.toolCalls.length > 0) return {
+			role: "assistant",
+			content: message.content,
+			tool_calls: message.toolCalls.map((call) => ({
+				id: call.id,
+				type: "function",
+				function: {
+					name: call.name,
+					arguments: call.arguments
+				}
+			}))
+		};
+		return {
+			role: message.role,
+			content: message.content
+		};
+	});
+	const body = {
+		model: request.model,
+		messages,
+		stream: true,
+		max_tokens: request.maxOutputTokens ?? 8192
+	};
+	if (request.tools !== void 0 && request.tools.length > 0) body.tools = request.tools.map((tool) => ({
+		type: "function",
+		function: {
+			name: tool.name,
+			description: tool.description ?? "",
+			parameters: tool.parameters
+		}
+	}));
+	if (request.temperature !== void 0) body.temperature = request.temperature;
+	return body;
+}
+function buildNativeBody(request) {
+	const messages = request.messages.map((message) => {
+		if (message.role === "tool") return {
+			role: "tool",
+			content: message.content
+		};
+		if (message.role === "assistant" && message.toolCalls !== void 0 && message.toolCalls.length > 0) return {
+			role: "assistant",
+			content: message.content,
+			tool_calls: message.toolCalls.map((call) => ({ function: {
+				name: call.name,
+				arguments: call.arguments
+			} }))
+		};
+		return {
+			role: message.role,
+			content: message.content
+		};
+	});
+	const body = {
+		model: request.model,
+		messages,
+		stream: true
+	};
+	const options = { num_predict: request.maxOutputTokens ?? 8192 };
+	if (request.temperature !== void 0) options.temperature = request.temperature;
+	body.options = options;
+	if (request.tools !== void 0 && request.tools.length > 0) body.tools = request.tools.map((tool) => ({
+		type: "function",
+		function: {
+			name: tool.name,
+			description: tool.description ?? "",
+			parameters: tool.parameters
+		}
+	}));
+	return body;
+}
+/** The endpoint one request goes to. */
+function chatUrl(wire) {
+	return wire === "openai" ? `${CLOUD_BASE_URL}${OPENAI_CHAT_PATH}` : `${CLOUD_BASE_URL}${NATIVE_CHAT_PATH}`;
+}
+/**
+* Start one streaming chat call.
+*
+* The HTTP status is surfaced through a promise rather than by handing back the
+* Response, because the caller has to decide between retrying this account and
+* failing the turn, and both need the status before any body is read.
+*/
+function startChat(fetchFn, credentials, wire, request) {
+	let resolveStatus = () => void 0;
+	let rejectStatus = () => void 0;
+	const status = new Promise((resolve, reject) => {
+		resolveStatus = resolve;
+		rejectStatus = reject;
+	});
+	const opened = fetchFn(chatUrl(wire), {
+		method: "POST",
+		headers: headersFor(credentials, "text/event-stream"),
+		body: JSON.stringify(buildBody(wire, request)),
+		signal: request.signal
+	}).then((response) => {
+		resolveStatus(response.status);
+		return response;
+	}, (error) => {
+		rejectStatus(error);
+		throw error;
+	});
+	return {
+		events: (async function* run() {
+			const response = await opened;
+			if (!response.ok) {
+				const detail = await safeErrorText(response);
+				yield {
+					type: "error",
+					message: `Ollama request failed (${response.status}): ${detail}`
+				};
+				return;
+			}
+			yield* readStream(response, wire);
+		})(),
+		status
+	};
+}
+async function safeErrorText(response) {
+	try {
+		return (await response.text()).slice(0, 500);
+	} catch {
+		return "";
+	}
+}
+/** Parse the SSE stream of either surface. */
+async function* readStream(response, wire) {
+	const body = response.body;
+	if (body === null) return;
+	const decoder = new TextDecoder();
+	const reader = body.getReader();
+	let buffer = "";
+	const pending = /* @__PURE__ */ new Map();
+	try {
+		for (;;) {
+			const { done, value } = await reader.read();
+			if (done) break;
+			buffer += decoder.decode(value, { stream: true });
+			let newline = buffer.indexOf("\n");
+			while (newline !== -1) {
+				const line = buffer.slice(0, newline).trim();
+				buffer = buffer.slice(newline + 1);
+				newline = buffer.indexOf("\n");
+				if (line === "" || !line.startsWith("data:")) continue;
+				const payload = line.slice(5).trim();
+				if (payload === "" || payload === "[DONE]") continue;
+				for (const event of parseEvent(payload, wire, pending)) yield event;
+			}
+		}
+	} finally {
+		reader.releaseLock?.();
+	}
+	for (const call of pending.values()) yield {
+		type: "tool_call",
+		call: call.arguments === "" ? {
+			...call,
+			arguments: "{}"
+		} : call
+	};
+}
+function parseEvent(payload, wire, pending) {
+	let parsed;
+	try {
+		parsed = JSON.parse(payload);
+	} catch {
+		return [];
+	}
+	if (typeof parsed !== "object" || parsed === null) return [];
+	const record = parsed;
+	if (typeof record.error === "string") return [{
+		type: "error",
+		message: record.error
+	}];
+	return wire === "openai" ? parseOpenAIChunk(record, pending) : parseNativeChunk(record, pending);
+}
+function parseOpenAIChunk(record, pending) {
+	const events = [];
+	const choices = Array.isArray(record.choices) ? record.choices : [];
+	for (const choice of choices) {
+		if (typeof choice !== "object" || choice === null) continue;
+		const entry = choice;
+		const delta = entry.delta;
+		if (typeof delta === "object" && delta !== null) {
+			const fields = delta;
+			if (typeof fields.content === "string" && fields.content !== "") events.push({
+				type: "text",
+				text: fields.content
+			});
+			if (Array.isArray(fields.tool_calls)) for (const raw of fields.tool_calls) accumulateToolCall(raw, pending, true);
+		}
+		if (typeof entry.finish_reason === "string" && entry.finish_reason !== "") events.push({
+			type: "done",
+			finishReason: entry.finish_reason
+		});
+	}
+	const usage = readUsage$1(record);
+	if (usage !== null) events.push({
+		type: "usage",
+		usage
+	});
+	return events;
+}
+function parseNativeChunk(record, pending) {
+	const events = [];
+	const message = record.message;
+	if (typeof message === "object" && message !== null) {
+		const fields = message;
+		if (typeof fields.content === "string" && fields.content !== "") events.push({
+			type: "text",
+			text: fields.content
+		});
+		if (Array.isArray(fields.tool_calls)) for (const raw of fields.tool_calls) accumulateToolCall(raw, pending, false);
+	}
+	if (record.done === true) events.push({
+		type: "done",
+		finishReason: "stop"
+	});
+	const usage = readNativeUsage(record);
+	if (usage !== null) events.push({
+		type: "usage",
+		usage
+	});
+	return events;
+}
+/**
+* Read token counts off an OpenAI-compatible chunk.
+*
+* Ollama names them the OpenAI way here (`prompt_tokens` / `completion_tokens`),
+* while the native surface uses `prompt_eval_count` / `eval_count`. Both are
+* accepted on both paths because the service documents the native names in its
+* own OpenAPI and an OpenAI-compatible alias in the compatibility layer, and a
+* counter that silently read zero on the wrong surface would be worse than none.
+*
+* Returns null when the chunk carries neither pair, which is every content chunk.
+*/
+function readUsage$1(record) {
+	const container = typeof record.usage === "object" && record.usage !== null ? record.usage : record;
+	const input = firstNumber$1(container, [
+		"prompt_tokens",
+		"prompt_eval_count",
+		"input_tokens"
+	]);
+	const output = firstNumber$1(container, [
+		"completion_tokens",
+		"eval_count",
+		"output_tokens"
+	]);
+	if (input === void 0 && output === void 0) return null;
+	return {
+		inputTokens: input ?? 0,
+		outputTokens: output ?? 0
+	};
+}
+/** The native surface's names, read from the chunk itself. */
+function readNativeUsage(record) {
+	const input = firstNumber$1(record, [
+		"prompt_eval_count",
+		"prompt_tokens",
+		"input_tokens"
+	]);
+	const output = firstNumber$1(record, [
+		"eval_count",
+		"completion_tokens",
+		"output_tokens"
+	]);
+	if (input === void 0 && output === void 0) return null;
+	return {
+		inputTokens: input ?? 0,
+		outputTokens: output ?? 0
+	};
+}
+function firstNumber$1(record, keys) {
+	for (const key of keys) {
+		const value = record[key];
+		if (typeof value === "number" && Number.isFinite(value) && value >= 0) return Math.floor(value);
+	}
+}
+/**
+* Fold one streamed tool-call delta into the pending set.
+*
+* Nothing is emitted here: a call is only complete once the stream has ended, and
+* emitting a half-received argument string would have the caller act on a tool
+* invocation that does not exist yet. The native surface sends no id, so a stable
+* synthetic one is minted from the delta index — the only handle both surfaces
+* agree on for pairing a call with its result.
+*/
+function accumulateToolCall(raw, pending, openaiShape) {
+	if (typeof raw !== "object" || raw === null) return;
+	const record = raw;
+	const index = typeof record.index === "number" ? record.index : 0;
+	const holder = typeof record.function === "object" && record.function !== null ? record.function : record;
+	const name = typeof holder.name === "string" ? holder.name : void 0;
+	const args = typeof holder.arguments === "string" ? holder.arguments : void 0;
+	const id = openaiShape && typeof record.id === "string" ? record.id : void 0;
+	if (name === void 0 && args === void 0 && id === void 0) return;
+	const call = pending.get(index) ?? {
+		id: id ?? `call_${index}`,
+		name: "",
+		arguments: ""
+	};
+	if (id !== void 0 && id !== "") call.id = id;
+	if (name !== void 0 && name !== "") call.name = name;
+	if (args !== void 0) call.arguments += args;
+	pending.set(index, call);
+}
+/**
+* Read the model catalog.
+*
+* `/api/tags` is the native surface's own list and needs no translation; the
+* OpenAI `/models` endpoint would be a subset of it. Returns an empty list rather
+* than throwing, so a failed sync degrades to the cached selection instead of
+* making the line unusable.
+*/
+async function loadCatalog$1(fetchFn, credentials) {
+	const response = await fetchFn(`${CLOUD_BASE_URL}${NATIVE_TAGS_PATH}`, {
+		method: "GET",
+		headers: headersFor(credentials, "application/json"),
+		signal: AbortSignal.timeout(CATALOG_TIMEOUT_MS)
+	});
+	if (!response.ok) return [];
+	const data = await response.json();
+	if (typeof data !== "object" || data === null) return [];
+	const record = data;
+	const models = Array.isArray(record.models) ? record.models : [];
+	const result = [];
+	for (const entry of models) {
+		if (typeof entry !== "object" || entry === null) continue;
+		const fields = entry;
+		const name = typeof fields.name === "string" ? fields.name : void 0;
+		const model = typeof fields.model === "string" ? fields.model : void 0;
+		const id = name ?? model;
+		if (id === void 0 || id === "") continue;
+		result.push({
+			id,
+			name,
+			fetchedAt: Date.now()
+		});
+	}
+	return result;
+}
+//#endregion
+//#region src/host/ollama/mapper.ts
+function createStreamState$4() {
+	return {
+		nextBlockIndex: 0,
+		textBlock: null,
+		toolCalls: /* @__PURE__ */ new Map(),
+		hasContent: false
+	};
+}
+/** Fold one event into chunks, opening and closing blocks as the stream needs. */
+function applyEvent(state, event) {
+	if (event.type === "text") return applyText(state, event.text);
+	if (event.type === "tool_call") return applyToolCall(state, event.call);
+	return [];
+}
+function applyText(state, text) {
+	state.hasContent = true;
+	if (state.textBlock === null) {
+		const index = state.nextBlockIndex;
+		state.nextBlockIndex += 1;
+		state.textBlock = {
+			index,
+			text
+		};
+		return [{
+			type: "block-start",
+			index,
+			blockType: "text"
+		}, {
+			type: "text-delta",
+			index,
+			text
+		}];
+	}
+	state.textBlock.text += text;
+	return [{
+		type: "text-delta",
+		index: state.textBlock.index,
+		text
+	}];
+}
+function applyToolCall(state, call) {
+	state.hasContent = true;
+	const out = [];
+	out.push(...closeText(state));
+	const index = state.nextBlockIndex;
+	state.nextBlockIndex += 1;
+	out.push({
+		type: "block-start",
+		index,
+		blockType: "tool-call"
+	});
+	state.toolCalls.set(call.id, {
+		call,
+		blockIndex: index
+	});
+	out.push({
+		type: "tool-call-delta",
+		index,
+		id: toToolCallId(call.id),
+		name: call.name,
+		argumentsDelta: call.arguments === "" ? "{}" : call.arguments
+	});
+	return out;
+}
+function closeText(state) {
+	if (state.textBlock === null) return [];
+	const { index, text } = state.textBlock;
+	state.textBlock = null;
+	return [{
+		type: "block-end",
+		index,
+		block: {
+			type: "text",
+			text
+		}
+	}];
+}
+function closeToolCalls$3(state) {
+	const out = [];
+	for (const entry of [...state.toolCalls.values()]) {
+		const { call, blockIndex } = entry;
+		const block = {
+			type: "tool-call",
+			id: toToolCallId(call.id),
+			name: call.name,
+			arguments: call.arguments === "" ? "{}" : call.arguments
+		};
+		out.push({
+			type: "block-end",
+			index: blockIndex,
+			block
+		});
+	}
+	state.toolCalls.clear();
+	return out;
+}
+/**
+* Close the turn.
+*
+* A stream that ends while a text block is still open is a real possibility - the
+* model can be cut off mid-sentence - so the block is closed with what arrived
+* rather than dropped, and the text the user sees is the text that streamed.
+*/
+function closeStream$3(state) {
+	const sawToolCall = state.toolCalls.size > 0;
+	const out = [...closeText(state), ...closeToolCalls$3(state)];
+	out.push({
+		type: "finish",
+		reason: sawToolCall ? { kind: "tool-calls" } : { kind: "stop" }
+	});
+	return out;
+}
+//#endregion
+//#region src/host/ollama/adapter.ts
+/**
+* Retry policy for the `ollama` route.
+*
+* A 5xx from a shared inference service says nothing about this key, so it is
+* classified as `SERVER` and given the same bounded backoff the other lines use.
+* Deliberately outside the set: `INVALID_CREDENTIAL` (a rejected key fails
+* identically every time, and the pool rotates past it instead) and `ABORTED`.
+*/
+const RETRY_POLICY$4 = resolveRetryPolicy({
+	mode: "normal",
+	maxRetries: 3,
+	retryableCodes: [
+		"RATE_LIMIT",
+		"SERVER",
+		"TIMEOUT",
+		"TRANSPORT"
+	],
+	backoff: {
+		initialDelayMs: 1500,
+		maxDelayMs: 15e3,
+		jitterRatio: .2
+	}
+}, "dsh-chatgpt-subscription.ollama.retry");
+var OllamaAdapter = class extends LlmAdapter {
+	store;
+	modelSettings;
+	options;
+	/**
+	* Rotation pool, or null for the single stored key.
+	*
+	* Only the plugin entry installs it, because the entry owns the storage; an
+	* adapter built without one must keep reading the one credential file rather
+	* than minting pool state that would outlive the process.
+	*/
+	accountPool;
+	constructor(store = new FileCredentialStore$5(), modelSettings = new FileModelSettingsStore$5(), options = {}, accountPool) {
+		super();
+		this.store = store;
+		this.modelSettings = modelSettings;
+		this.options = options;
+		this.accountPool = accountPool ?? null;
+	}
+	providerInfo(provider) {
+		return {
+			id: provider,
+			name: PROVIDER_NAME$4
+		};
+	}
+	providerRetryPolicy() {
+		return RETRY_POLICY$4;
+	}
+	imageRequestPricing() {}
+	settings() {
+		return this.modelSettings.read();
+	}
+	/**
+	* Catalog for the picker: the live `/api/tags` listing when a key can reach it,
+	* the shipped fallback otherwise, narrowed by the user's enabled selection.
+	*
+	* Ollama's model set moves quickly, so a stale local table would be the wrong
+	* default; the fallback exists only so the line is usable before the first
+	* successful sync.
+	*/
+	async catalog() {
+		const load = this.options.loadCatalog ?? loadCatalog$1;
+		const fetchFn = this.options.fetchFn ?? fetch;
+		const credentials = await this.anyCredentials();
+		if (credentials !== null) {
+			const live = await load(fetchFn, credentials).catch(() => []);
+			if (live.length > 0) {
+				await this.modelSettings.storeCatalog(live).catch(() => void 0);
+				return live;
+			}
+		}
+		const cached = this.modelSettings.status().catalogModels;
+		if (cached.length > 0) return cached;
+		return [...FALLBACK_MODELS$3];
+	}
+	/** Any usable key, for the catalog call; the pool decides real routing later. */
+	async anyCredentials() {
+		if (this.accountPool !== null) return (await this.accountPool.getEffectiveCredential(/* @__PURE__ */ new Set(), this.options.fetchFn ?? fetch).catch(() => null))?.credentials ?? null;
+		return this.store.read().catch(() => null);
+	}
+	async listModels(provider) {
+		const prov = provider || "ollama";
+		const settings = await this.settings();
+		if (settings.enabled === false) return [];
+		const catalog = await this.catalog();
+		const enabled = new Set(settings.enabledModelIds);
+		return (settings.enabledModelIds.length === 0 ? catalog : catalog.filter((model) => enabled.has(model.id))).map((model) => ({
+			provider: prov,
+			id: model.id,
+			name: model.name ?? model.id,
+			inputModalities: ["text", "image"]
+		}));
+	}
+	async resolveModel(provider, modelId, signal) {
+		if (signal?.aborted) throw new LlmError("Ollama model resolution aborted", "ABORTED");
+		const entry = (await this.catalog()).find((model) => model.id === modelId);
+		return {
+			provider,
+			id: modelId,
+			name: entry?.name ?? modelId,
+			inputModalities: ["text", "image"],
+			context: { contextWindow: contextWindowFor(entry) },
+			...outputReservation(contextWindowFor(entry), DEFAULT_MAX_OUTPUT_TOKENS)
+		};
+	}
+	async prepareCall(provider, model, signal) {
+		return {
+			model: await this.resolveModel(provider, model, signal),
+			stream: (options) => this.stream(options)
+		};
+	}
+	async *stream(options) {
+		yield* wrapStreamWithWatchdog((watchdogSignal) => this.requestStream(options, watchdogSignal), options.signal, STREAM_IDLE_TIMEOUT_MS$4, STREAM_IDLE_TIMEOUT_CODE$4, PROVIDER_NAME$4);
+	}
+	async *requestStream(options, signal) {
+		const fetchFn = this.options.fetchFn ?? fetch;
+		const wire = wireForModel$1(options.model);
+		const request = toOllamaRequest(options);
+		const pool = this.accountPool;
+		const tried = /* @__PURE__ */ new Set();
+		let lastStatus;
+		let lastDetail = "";
+		let accountId;
+		for (;;) {
+			let credentials;
+			if (pool === null) {
+				const stored = await this.store.read();
+				if (stored === null) throw new LlmError(`No ${PROVIDER_NAME$4} API key configured. Add one from Settings > ${PROVIDER_NAME$4} (create a key at https://ollama.com/settings/keys).`, "MISSING_CREDENTIAL");
+				credentials = stored;
+			} else {
+				const effective = await pool.getEffectiveCredential(tried, fetchFn);
+				accountId = effective.account.id;
+				tried.add(accountId);
+				credentials = effective.credentials;
+			}
+			const call = startChat(fetchFn, credentials, wire, {
+				...request,
+				signal
+			});
+			const state = createStreamState$4();
+			let openedStatus;
+			call.status.then((value) => {
+				openedStatus = value;
+			});
+			let spent = null;
+			let failed = false;
+			for await (const event of call.events) {
+				if (event.type === "error") {
+					lastStatus = openedStatus;
+					lastDetail = event.message;
+					failed = true;
+					break;
+				}
+				if (event.type === "usage") {
+					spent = event.usage;
+					continue;
+				}
+				if (event.type === "text" || event.type === "tool_call") {
+					for (const chunk of applyEvent(state, event)) yield chunk;
+					continue;
+				}
+			}
+			if (!failed) {
+				for (const chunk of closeStream$3(state)) yield chunk;
+				if (pool !== null && accountId !== void 0 && spent !== null) await pool.recordUsage(accountId, spent.inputTokens, spent.outputTokens);
+				return;
+			}
+			if (pool === null || accountId === void 0) break;
+			if (lastStatus === 401 || lastStatus === 403) {
+				await pool.markAuthFailed(accountId, `${PROVIDER_NAME$4} rejected the API key (${lastStatus}).`, "invalid").catch(() => void 0);
+				if (await pool.hasAnotherAvailableAccount(tried)) continue;
+				break;
+			}
+			if (lastStatus === 429) {
+				await pool.markCooldown(accountId, POOL_COOLDOWN_MS$3, `${PROVIDER_NAME$4} 429`).catch(() => void 0);
+				if (await pool.hasAnotherAvailableAccount(tried)) continue;
+				break;
+			}
+			break;
+		}
+		const status = lastStatus;
+		if (status === 401 || status === 403) throw new LlmError(`${PROVIDER_NAME$4} rejected the API key (${status}). Replace it from Settings > ${PROVIDER_NAME$4}.${lastDetail ? ` ${lastDetail}` : ""}`, "INVALID_CREDENTIAL", { status });
+		if (status === 429) throw new LlmError(`${PROVIDER_NAME$4} rate limit or plan quota reached (429). Add another key, or wait for the cooldown shown in Settings > ${PROVIDER_NAME$4}.${lastDetail ? ` ${lastDetail}` : ""}`, "RATE_LIMIT", { status: 429 });
+		throw new LlmError(`${PROVIDER_NAME$4} request failed${status === void 0 ? "" : ` (${status})`}.${lastDetail}`, status !== void 0 && status >= 500 ? "SERVER" : "INVALID_REQUEST", status === void 0 ? {} : { status });
+	}
+};
+/** Project DSH's request shape onto Ollama's, per surface. */
+function toOllamaRequest(options) {
+	const messages = [];
+	for (const message of options.messages) {
+		if (message.role === "tool") {
+			messages.push({
+				role: "tool",
+				content: textOf$4(message.content),
+				toolCallId: message.toolCallId
+			});
+			continue;
+		}
+		const toolCalls = assistantToolCalls(message.content);
+		if (toolCalls.length > 0) {
+			messages.push({
+				role: "assistant",
+				content: textOf$4(message.content),
+				toolCalls
+			});
+			continue;
+		}
+		messages.push({
+			role: message.role === "developer" ? "system" : message.role,
+			content: textOf$4(message.content)
+		});
+	}
+	const request = {
+		model: options.model,
+		messages
+	};
+	if (options.maxTokens !== void 0) request.maxOutputTokens = options.maxTokens;
+	if (options.temperature !== void 0) request.temperature = options.temperature;
+	const tools = options.tools;
+	if (Array.isArray(tools) && tools.length > 0) request.tools = tools.map((tool) => ({
+		name: tool.name,
+		...typeof tool.description === "string" ? { description: tool.description } : {},
+		parameters: tool.parameters ?? {
+			type: "object",
+			properties: {}
+		}
+	}));
+	return request;
+}
+/** Tool calls an assistant turn carries, read from its content blocks. */
+function assistantToolCalls(content) {
+	if (!Array.isArray(content)) return [];
+	const calls = [];
+	for (const block of content) {
+		if (typeof block !== "object" || block === null) continue;
+		const fields = block;
+		if (fields.type !== "tool-call") continue;
+		const id = fields.id;
+		const name = fields.name;
+		if (typeof id !== "string" || typeof name !== "string") continue;
+		calls.push({
+			id,
+			name,
+			arguments: typeof fields.arguments === "string" ? fields.arguments : JSON.stringify(fields.arguments ?? {})
+		});
+	}
+	return calls;
+}
+function textOf$4(content) {
+	if (typeof content === "string") return content;
+	if (Array.isArray(content)) return content.map((part) => {
+		if (typeof part === "string") return part;
+		if (typeof part === "object" && part !== null) {
+			const text = part.text;
+			if (typeof text === "string") return text;
+		}
+		return "";
+	}).join("");
+	return "";
+}
+//#endregion
+//#region src/host/ollama/routes.ts
+const ROUTE_PREFIX$3 = "/ollama/api";
+function sendJson$5(response, status, body) {
+	const payload = JSON.stringify(body);
+	response.writeHead(status, { "content-type": "application/json; charset=utf-8" });
+	response.end(payload);
+}
+function sendMethodNotAllowed$4(response) {
+	sendJson$5(response, 405, {
+		ok: false,
+		error: "Method not allowed."
+	});
+}
+async function readRequestJson$5(request) {
+	const chunks = [];
+	let size = 0;
+	for await (const chunk of request) {
+		const buffer = chunk;
+		size += buffer.length;
+		if (size > 8192) throw new Error("Request body is too large.");
+		chunks.push(buffer);
+	}
+	if (chunks.length === 0) return {};
+	const parsed = JSON.parse(Buffer.concat(chunks).toString("utf8"));
+	if (typeof parsed !== "object" || parsed === null || Array.isArray(parsed)) throw new Error("Request body must be a JSON object.");
+	return parsed;
+}
+/** Whether one account can serve a request right now. */
+function isUsable(account, now) {
+	if (account.authStatus !== void 0 && account.authStatus !== "ok") return false;
+	if (typeof account.cooldownUntil === "number" && account.cooldownUntil > now) return false;
+	return true;
+}
+/** Assemble the whole status payload the Ollama tab renders. */
+async function getOllamaWebStatus(options) {
+	const { accountPool, modelSettings } = options;
+	const data = await accountPool.read().catch(() => null);
+	const accounts = await accountPool.listAccounts().catch(() => []);
+	const now = Date.now();
+	const pool = {
+		accounts,
+		rotationStrategy: data?.rotationStrategy ?? "sequential",
+		...data?.activeAccountId === void 0 ? {} : { activeAccountId: data.activeAccountId }
+	};
+	const settings = await modelSettings.read().catch(() => ({
+		enabled: true,
+		enabledModelIds: [],
+		catalogModels: [],
+		defaultReasoningEffort: null
+	}));
+	return {
+		pool,
+		models: settings.catalogModels.map((model) => ({
+			id: model.id,
+			...model.name === void 0 ? {} : { name: model.name }
+		})),
+		enabledModelIds: settings.enabledModelIds,
+		usable: accounts.some((account) => isUsable(account, now)),
+		catalogSynced: settings.catalogModels.length > 0
+	};
+}
+/**
+* Register the Ollama settings API.
+*
+* The surface is deliberately the same one every other line exposes - status,
+* accounts, models - so the shared settings card and the client API helper
+* work against it without a per-provider special case.
+*/
+function registerOllamaRoutes(ctx, options) {
+	const { accountPool, modelSettings } = options;
+	const fetchFn = options.fetchFn ?? fetch;
+	return ctx.webServer.register({
+		kind: "prefix",
+		path: ROUTE_PREFIX$3,
+		handler: async (request, response) => {
+			const path = new URL(request.url || "/", "http://dsh.local").pathname.replace(/^\/ollama\/api\/?/, "");
+			const method = request.method ?? "GET";
+			if (path === "" || path === "status") return sendJson$5(response, 200, {
+				ok: true,
+				value: await getOllamaWebStatus(options)
+			});
+			if (path === "accounts") {
+				if (method === "GET") return sendJson$5(response, 200, {
+					ok: true,
+					value: await getOllamaWebStatus(options)
+				});
+				if (method !== "POST") return sendMethodNotAllowed$4(response);
+				if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+					ok: false,
+					error: "Cross-origin request rejected."
+				});
+				let body;
+				try {
+					body = await readRequestJson$5(request);
+				} catch (error) {
+					return sendJson$5(response, 400, {
+						ok: false,
+						error: error instanceof Error ? error.message : String(error)
+					});
+				}
+				const action = typeof body.action === "string" ? body.action : "";
+				const accountId = typeof body.accountId === "string" ? body.accountId : void 0;
+				try {
+					if (action === "add") {
+						const apiKey = typeof body.apiKey === "string" ? body.apiKey.trim() : "";
+						if (apiKey === "") return sendJson$5(response, 400, {
+							ok: false,
+							error: "API key is required."
+						});
+						const alias = typeof body.alias === "string" && body.alias.trim() !== "" ? body.alias.trim() : void 0;
+						const credentials = {
+							apiKey,
+							addedAt: Date.now()
+						};
+						if (alias !== void 0) credentials.alias = alias;
+						return sendJson$5(response, 200, {
+							ok: true,
+							value: { id: (await accountPool.addAccount(credentials, alias)).id }
+						});
+					}
+					if (accountId === void 0) return sendJson$5(response, 400, {
+						ok: false,
+						error: "accountId is required."
+					});
+					if (action === "set-primary") await accountPool.setPrimary(accountId);
+					else if (action === "set-alias" && typeof body.alias === "string") await accountPool.setAlias(accountId, body.alias);
+					else if (action === "delete") await accountPool.deleteAccount(accountId);
+					else if (action === "clear-cooldown") await accountPool.clearCooldown(accountId);
+					else if (action === "strategy" && (body.strategy === "sequential" || body.strategy === "round-robin" || body.strategy === "sticky")) await accountPool.setStrategy(body.strategy);
+					else return sendJson$5(response, 400, {
+						ok: false,
+						error: "Unsupported account action."
+					});
+				} catch (error) {
+					return sendJson$5(response, 400, {
+						ok: false,
+						error: error instanceof Error ? error.message : String(error)
+					});
+				}
+				return sendJson$5(response, 200, {
+					ok: true,
+					value: await getOllamaWebStatus(options)
+				});
+			}
+			if (path === "models") {
+				if (method === "GET") {
+					const settings = await modelSettings.read();
+					return sendJson$5(response, 200, {
+						ok: true,
+						value: {
+							models: settings.catalogModels,
+							enabled: settings.enabled,
+							enabledModelIds: settings.enabledModelIds
+						}
+					});
+				}
+				if (method !== "POST") return sendMethodNotAllowed$4(response);
+				if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+					ok: false,
+					error: "Cross-origin request rejected."
+				});
+				const body = await readRequestJson$5(request).catch(() => ({}));
+				if (typeof body.enabled === "boolean") await modelSettings.update({ enabled: body.enabled });
+				if (Array.isArray(body.enabledModelIds)) await modelSettings.update({ enabledModelIds: body.enabledModelIds.filter((id) => typeof id === "string") });
+				return sendJson$5(response, 200, {
+					ok: true,
+					value: await modelSettings.read()
+				});
+			}
+			if (path === "catalog/refresh") {
+				if (method !== "POST") return sendMethodNotAllowed$4(response);
+				if (!isSameOriginMutation(request)) return sendJson$5(response, 403, {
+					ok: false,
+					error: "Cross-origin request rejected."
+				});
+				let credentials = null;
+				try {
+					credentials = (await accountPool.getEffectiveCredential(void 0, fetchFn)).credentials;
+				} catch {
+					credentials = null;
+				}
+				if (credentials === null) return sendJson$5(response, 400, {
+					ok: false,
+					error: "Add an API key before syncing models."
+				});
+				const models = await loadCatalog$1(fetchFn, credentials).catch(() => []);
+				if (models.length === 0) return sendJson$5(response, 502, {
+					ok: false,
+					error: "Could not read the model list from Ollama."
+				});
+				await modelSettings.storeCatalog(models);
+				return sendJson$5(response, 200, {
+					ok: true,
+					value: { models }
+				});
+			}
+		}
+	});
+}
+//#endregion
+//#region src/host/ollama/account-pool.ts
+/** Encrypted pool file this line owns. */
+function ollamaPoolPath() {
+	return path.join(dshHomeDir(), "storages", "ollama-pool.json");
+}
+/**
+* Stable identity of one key.
+*
+* Ollama hands the user a bare key and nothing else — no account id, no email —
+* so the key's own digest is the only identity available. It is never written
+* anywhere: it exists to tell two pasted keys apart inside the pool document,
+* and it is not a credential and cannot be used as one.
+*/
+function ollamaAccountKey(credentials) {
+	return createHash("sha256").update(credentials.apiKey).digest("hex");
+}
+function optionalString$5(record, key) {
+	const value = record[key];
+	if (value === void 0 || value === null) return void 0;
+	if (typeof value !== "string") throw new Error("Ollama pool account field is invalid");
+	return value;
+}
+/** Non-negative integer, or undefined for anything else. */
+function counter(value) {
+	if (typeof value !== "number" || !Number.isFinite(value) || value < 0) return void 0;
+	return Math.floor(value);
+}
+/** Read one account's counters, dropping the whole record if it is unusable. */
+function readUsage(value) {
+	if (typeof value !== "object" || value === null || Array.isArray(value)) return void 0;
+	const record = value;
+	const inputTokens = counter(record.inputTokens);
+	const outputTokens = counter(record.outputTokens);
+	const requestCount = counter(record.requestCount);
+	if (inputTokens === void 0 || outputTokens === void 0 || requestCount === void 0) return void 0;
+	const lastCountedAt = counter(record.lastCountedAt);
+	return {
+		inputTokens,
+		outputTokens,
+		requestCount,
+		...lastCountedAt === void 0 ? {} : { lastCountedAt }
+	};
+}
+/** Validate and normalize one whole pool document. */
+function parseOllamaPoolData(value) {
+	if (typeof value !== "object" || value === null || Array.isArray(value)) throw new Error("Ollama pool payload is invalid");
+	const record = value;
+	const accounts = [];
+	for (const item of Array.isArray(record.accounts) ? record.accounts : []) {
+		if (typeof item !== "object" || item === null) continue;
+		const raw = item;
+		if (typeof raw.id !== "string") continue;
+		const credentials = parseOllamaCredentials(raw.credentials);
+		const account = {
+			id: raw.id,
+			alias: typeof raw.alias === "string" && raw.alias !== "" ? raw.alias : "Ollama 账号",
+			credentials,
+			addedAt: typeof raw.addedAt === "number" ? raw.addedAt : Date.now(),
+			isPrimary: raw.isPrimary === true
+		};
+		const lastModelId = optionalString$5(raw, "lastModelId");
+		if (lastModelId !== void 0) account.lastModelId = lastModelId;
+		const usage = readUsage(raw.usage);
+		if (usage !== void 0) account.usage = usage;
+		if (typeof raw.lastUsedAt === "number") account.lastUsedAt = raw.lastUsedAt;
+		if (typeof raw.cooldownUntil === "number") account.cooldownUntil = raw.cooldownUntil;
+		if (typeof raw.cooldownReason === "string") account.cooldownReason = raw.cooldownReason;
+		if (raw.authStatus === "expired" || raw.authStatus === "invalid") account.authStatus = raw.authStatus;
+		if (typeof raw.authFailedReason === "string") account.authFailedReason = raw.authFailedReason;
+		accounts.push(account);
+	}
+	const result = {
+		version: 1,
+		rotationStrategy: normalizeRotationStrategy(record.rotationStrategy),
+		accounts
+	};
+	if (typeof record.activeAccountId === "string") result.activeAccountId = record.activeAccountId;
+	return result;
+}
+/**
+* Ollama's account pool.
+*
+* An Ollama credential is a static key that never expires, so this wrapper
+* supplies identity, the display alias a user chose, and the mirror of the
+* single-credential file — no refresh logic is needed anywhere.
+*/
+var OllamaAccountPool = class extends AccountPoolCore {
+	store;
+	constructor(options = {}) {
+		const store = options.store ?? new FileCredentialStore$5();
+		const hooks = {
+			providerId: PROVIDER_ID$4,
+			displayName: PROVIDER_NAME$4,
+			poolFile: ollamaPoolPath(),
+			keychainService: "dsh-ollama-pool",
+			parsePoolData: parseOllamaPoolData,
+			dedupeKey: ollamaAccountKey,
+			defaultAlias: (credentials, position) => credentials.alias ?? `Ollama 账号 ${position}`,
+			createAccount: ({ id, alias, credentials, addedAt, isPrimary }) => ({
+				id,
+				alias,
+				credentials,
+				addedAt,
+				isPrimary
+			}),
+			legacyAccount: async () => {
+				const stored = await store.read();
+				if (stored === null) return null;
+				return {
+					id: "acc_primary",
+					alias: stored.alias ?? "Ollama 账号 1",
+					credentials: stored,
+					addedAt: stored.addedAt ?? Date.now(),
+					isPrimary: true
+				};
+			},
+			mirrorPrimary: async (credentials) => {
+				if (credentials === null) {
+					await store.clear();
+					return;
+				}
+				await store.write(credentials);
+			},
+			extendSummary: (account, base) => {
+				const extra = {};
+				if (account.lastModelId !== void 0) extra.planLabel = account.lastModelId;
+				if (account.usage !== void 0) extra.usage = { ...account.usage };
+				return Object.keys(extra).length === 0 ? base : {
+					...base,
+					...extra
+				};
+			},
+			...options.backend === void 0 ? {} : { backend: options.backend },
+			...options.maxAccounts === void 0 ? {} : { maxAccounts: options.maxAccounts }
+		};
+		super(hooks);
+		this.store = store;
+	}
+	/**
+	* Pick the key for the next request, skipping accounts already tried.
+	*
+	* Ollama's credential carries no deployment dimension - there is one cloud
+	* endpoint and every key is valid for it - so the account and its key are
+	* returned together and nothing has to be attached to the credential.
+	*/
+	async getEffectiveCredential(excludeIds, fetchFn = fetch) {
+		return this.getEffectiveAccount(excludeIds, fetchFn);
+	}
+	/** The pre-pool credential file this pool mirrors its primary account into. */
+	mirrorStore() {
+		return this.store;
+	}
+	async recordUsage(accountId, inputTokens, outputTokens) {
+		if (inputTokens <= 0 && outputTokens <= 0) return;
+		await this.updatePool((data) => {
+			const account = data.accounts.find((entry) => entry.id === accountId);
+			if (account === void 0) return false;
+			const previous = account.usage;
+			account.usage = {
+				inputTokens: (previous?.inputTokens ?? 0) + inputTokens,
+				outputTokens: (previous?.outputTokens ?? 0) + outputTokens,
+				requestCount: (previous?.requestCount ?? 0) + 1,
+				lastCountedAt: Date.now()
+			};
+			return true;
+		}).catch(() => void 0);
+	}
+};
+//#endregion
 //#region src/host/kimi-code/model-catalog.ts
 /**
 * The four ids the subscription serves today.
@@ -17188,7 +18873,6 @@ var KimiCodeAdapter = class extends LlmAdapter {
 			name: entry?.name ?? modelId,
 			inputModalities: inputModalitiesForEntry(modelId, catalog),
 			context: { contextWindow: this.contextWindowFor(modelId, entry, settings.contextWindowOverrides) },
-			defaultMaxTokens: maxOutputTokensFor$3(modelId, this.contextWindowFor(modelId, entry, settings.contextWindowOverrides)),
 			...efforts.length === 0 ? {} : { reasoning: {
 				efforts: efforts.map((effort) => ({
 					id: ReasoningEffortId(effort),
@@ -21658,9 +23342,8 @@ var MinimaxCodeAdapter = class extends LlmAdapter {
 	* One model's resolved metadata, read through the current selection.
 	*
 	* The effective context window is the override when one is saved, and the
-	* output cap is sized against that same number — a window the card shows and a
-	* cap the request path computes from a different one is precisely the drift
-	* this single resolver exists to prevent.
+	* wire output cap is sized against that same number in requestStream. It is
+	* deliberately not exposed as a fixed defaultMaxTokens reservation here.
 	*/
 	async resolveModel(provider, modelId, signal) {
 		if (signal?.aborted === true) throw new LlmError("MiniMax Code model resolution aborted", "ABORTED");
@@ -21675,7 +23358,6 @@ var MinimaxCodeAdapter = class extends LlmAdapter {
 			name: entry?.name ?? modelId,
 			inputModalities: modalitiesForModel(modelId),
 			context: { contextWindow },
-			defaultMaxTokens: maxOutputTokensFor$1(modelId, contextWindow),
 			...efforts.length === 0 ? {} : { reasoning: {
 				efforts: efforts.map((effort) => ({
 					id: ReasoningEffortId(effort),
@@ -26378,7 +28060,7 @@ var WorkBuddyAdapter = class extends LlmAdapter {
 			name: entry.id === modelId ? entry.name : modelId,
 			inputModalities: entry.supportsImage ? ["text", "image"] : ["text"],
 			context: { contextWindow: contextWindow || 128e3 },
-			defaultMaxTokens: maxOutputTokensFor$2(modelId, catalog),
+			...outputReservation(contextWindow || 128e3, maxOutputTokensFor$2(modelId, catalog)),
 			...efforts.length === 0 ? {} : { reasoning: {
 				efforts: efforts.map((effort) => ({
 					id: ReasoningEffortId(effort),
@@ -26433,7 +28115,13 @@ var WorkBuddyAdapter = class extends LlmAdapter {
 		const fetchFn = this.options.fetchFn ?? fetch;
 		const requestOptions = offloadOldestRequestImages$1(normalizeGenerateOptions(options));
 		const images = await resolveRequestImages$1(requestOptions, this.options.attachments, signal);
-		const body = JSON.stringify(buildChatRequest(requestOptions, images));
+		const settings = await this.settings();
+		const credentials = await this.credentials(settings);
+		const catalog = credentials === null ? FALLBACK_MODELS$1 : await this.catalog(credentials);
+		const body = JSON.stringify(buildChatRequest({
+			...requestOptions,
+			maxTokens: requestOptions.maxTokens ?? maxOutputTokensFor$2(options.model, catalog)
+		}, images));
 		const pool = this.accountPool;
 		const tried = /* @__PURE__ */ new Set();
 		let response;
@@ -27404,17 +29092,29 @@ function loopbackRedirectUri(port = DEFAULT_CALLBACK_PORT) {
 * LOCAL ADDITIONS — the rows the reference snapshot predates
 * ---------------------------------------------------------------------------
 *
-* `claude-opus-5-5` and `claude-sonnet-5-5` are the two such rows today. The
-* mirror list in the test is `LOCALLY_CURATED_MODEL_IDS`; the two must be
-* changed together. Sonnet 5.5's own reasons are recorded at its row; what
-* follows below is Opus 5.5's.
-*
-* It is CURATED rather than transcribed because the snapshot this table copies
-* has no entry for it, and the one thing a transcription table must never do is
-* invent a row and let it read as if it had been read from somewhere. An
-* invented row that mimics the format of a checked one is worse than a missing
-* row: it is unverifiable and it looks verified. So the row below is marked at
-* the row itself, the test asserts the id is on the curated list, and the test
+* TWO KINDS OF ROW ARE NOT FULLY TRANSCRIBED, and they are different things.
+* Do not collapse them.
+*
+* 1. `claude-sonnet-5-5` — the snapshot this table was copied from has no
+*    entry for it, so the whole row is CURATED: every value below comes from
+*    the vendor's own documentation and no snapshot field can check it. The
+*    mirror list in the test is `LOCALLY_CURATED_MODEL_IDS`.
+*
+* 2. `claude-opus-5-5` — a NEWER snapshot (pi-ai >= 0.87.1) does carry it, so
+*    the row sits in the snapshot's own position and EVERY field is checked
+*    against that snapshot. Exactly one field is not: `thinkingMode`, recorded
+*    in the test as `SNAPSHOT_AGREES_BUT_CURATED_WINS`. The snapshot would
+*    classify it 'mid-convo', whose form forces `output_config.effort = 'high'`
+*    when the caller names none, while the vendor documents this model's
+*    default effort as MEDIUM. Transcribing it would silently outrank the user
+*    and think — and bill — harder than asked, so the documented 'adaptive'
+*    stands. The test asserts the row still disagrees with the snapshot, so a
+*    future snapshot that agrees fails loudly and retires the entry.
+*
+* Both must be changed together with the mirror list in the test. An invented
+* row that mimics the format of a checked one is worse than a missing row: it
+* is unverifiable and it looks verified. So each row is marked at the row
+* itself, the test asserts the id is on the curated list, and the test
 * asserts the curated list is exactly the ids the snapshot lacks — which keeps
 * the lock narrow instead of merely weaker.
 *
@@ -27439,10 +29139,12 @@ function loopbackRedirectUri(port = DEFAULT_CALLBACK_PORT) {
 * 2. `thinkingMode: 'adaptive'`, NOT `'mid-convo'`. The two are easy to confuse
 *    here because every model in this line that refuses a temperature is also,
 *    so far, a managed-effort model. That is a coincidence of `compat`, not a
-*    rule: this row carries no `supportsMidConvoEffort` flag, and 'mid-convo'
+*    rule. A newer snapshot DOES flag this row `supportsMidConvoEffort` and
+*    would therefore classify it 'mid-convo' — which is exactly why this is the
+*    one declared exception rather than a transcription: 'mid-convo'
 *    additionally forces `output_config = { effort: 'high' }` when the caller
-*    names no effort. That forced high is correct for the rows that DO carry the
-*    flag, because the vendor makes `high` their default effort; this model's
+*    names no effort. That forced high is correct for the rows whose documented
+*    default effort is high, and wrong for this one; this model's
 *    documented default effort is `medium`. Labelling it 'mid-convo' would
 *    silently override the vendor's own default and make every request think —
 *    and cost — harder than the user asked for.
@@ -27459,8 +29161,8 @@ function loopbackRedirectUri(port = DEFAULT_CALLBACK_PORT) {
 * the user has to reach.
 */
 /**
-* Sixteen rows: the snapshot's own 14, in the snapshot's own declaration order,
-* followed by the curated rows the snapshot predates (see LOCAL ADDITIONS above).
+* Sixteen rows: the snapshot's own 15, in the snapshot's own declaration order,
+* followed by the one row the snapshot still predates (see LOCAL ADDITIONS above).
 *
 * The order matters and is not cosmetic. The test asserts the snapshot's ids
 * appear in the table as a SUBSEQUENCE, so a curated row may be appended or
@@ -27649,6 +29351,25 @@ const CLAUDE_MODELS = Object.freeze([
 		],
 		canDisableThinking: false
 	},
+	{
+		id: "claude-opus-5-5",
+		name: "Claude Opus 5.5",
+		contextWindow: 1e6,
+		maxTokens: 128e3,
+		supportsImage: true,
+		supportsTemperature: false,
+		thinkingMode: "adaptive",
+		reasoningEfforts: [
+			"low",
+			"medium",
+			"high",
+			"xhigh",
+			"max"
+		],
+		canDisableThinking: false,
+		minCliVersion: "2.1.280",
+		bindsThinkingToPrefix: true
+	},
 	{
 		id: "claude-sonnet-4-5",
 		name: "Claude Sonnet 4.5 (latest)",
@@ -27712,25 +29433,6 @@ const CLAUDE_MODELS = Object.freeze([
 		],
 		canDisableThinking: true
 	},
-	{
-		id: "claude-opus-5-5",
-		name: "Claude Opus 5.5",
-		contextWindow: 1e6,
-		maxTokens: 128e3,
-		supportsImage: true,
-		supportsTemperature: false,
-		thinkingMode: "adaptive",
-		reasoningEfforts: [
-			"low",
-			"medium",
-			"high",
-			"xhigh",
-			"max"
-		],
-		canDisableThinking: false,
-		minCliVersion: "2.1.280",
-		bindsThinkingToPrefix: true
-	},
 	{
 		id: "claude-sonnet-5-5",
 		name: "Claude Sonnet 5.5",
@@ -32254,11 +33956,9 @@ function cancelLogin() {
 *   importing it before it exists would make this adapter uncompilable, and
 *   importing it afterwards changes nothing, because the only methods used are
 *   the five in {@link ClaudeAccountPoolLike}.
-* - A2. The live `GET /v1/models` listing is NOT read on the request path. It
-*   is authoritative for the context window alone (see client.ts), and the only
-*   model fact the body depends on — the thinking form — comes from the frozen
-*   catalog. Loading it per request would add a cached round trip that cannot
-*   change a byte of the body.
+* - A2. When no fixed output default was materialized, the request path reads
+*   the cached catalog to preserve its wire cap independently of compaction
+*   reservation metadata. Thinking forms still come from the frozen catalog.
 * - A3. A 429 with no `retry-after` and no reset instant cools the account for
 *   {@link POOL_COOLDOWN_MS}. That is deliberately shorter than the 5-hour
 *   window: an account parked for hours after a burst limit recovers would cost
@@ -32501,7 +34201,7 @@ var ClaudeAdapter = class ClaudeAdapter extends LlmAdapter {
 			name: entry.name,
 			inputModalities: entry.supportsImage ? ["text", "image"] : ["text"],
 			context: { contextWindow: contextWindow || 2e5 },
-			defaultMaxTokens: maxOutputTokensFor(modelId, catalog),
+			...outputReservation(contextWindow || 2e5, maxOutputTokensFor(modelId, catalog)),
 			...efforts.length === 0 ? {} : { reasoning: {
 				efforts: efforts.map((effort) => ({
 					id: ReasoningEffortId(effort),
@@ -32584,7 +34284,11 @@ var ClaudeAdapter = class ClaudeAdapter extends LlmAdapter {
 	}
 	async attemptRequest(credentials, requestOptions, images, toolNames, thinking, settings, signal, fetchFn) {
 		const cacheTtl = ClaudeAdapter.cacheTtlFor(settings);
-		const payload = buildClaudeRequestBody(requestOptions, images, {
+		const catalog = requestOptions.maxTokens === void 0 ? await this.catalog(credentials) : void 0;
+		const payload = buildClaudeRequestBody({
+			...requestOptions,
+			maxTokens: requestOptions.maxTokens ?? maxOutputTokensFor(requestOptions.model, catalog)
+		}, images, {
 			toolNames,
 			cacheControl: true,
 			cacheTtl
@@ -37709,6 +39413,38 @@ function installRelayProbe(ctx, options) {
 }
 //#endregion
 //#region src/host/reasoning-collapse-guard/index.ts
+/**
+* Reasoning-collapse guard.
+*
+* A long reasoning stream can degenerate into repetition: it cycles through a
+* handful of short phrases ("Let me call. Go. Calling. Go.") without emitting a
+* tool call or concluding anything, and runs until the output-token ceiling
+* truncates it. One archived session burned 128,000 output tokens that way and
+* returned an empty answer.
+*
+* The guard scores n-gram uniqueness over a trailing window of reasoning text.
+* Healthy deliberation keeps nearly every n-gram unique and scores near zero; a
+* degenerate stream scores near one. On a hit it stops yielding chunks, so the
+* in-flight request ends, and lets the turn resume on a fresh step.
+*
+* Seam choice: the guard wraps `llm/stream`, not `agent/assistant-stream`.
+* `agent/assistant-stream` is emit-mode — it can observe but cannot stop a
+* stream — and it does not exist before harness 0.1.5, which would make the
+* guard a silent no-op on every generation this plugin supports below that.
+* `llm/stream` is a waterfall whose `(options, next) => AsyncIterable<StreamChunk>`
+* signature is byte-identical from 0.1.2-alpha.5 through 0.2.0-rc.1, and
+* returning early from the listener is what actually ends the stream.
+*
+* The guard subscribes to that waterfall and never reassigns `ctx.llm.stream`.
+* The harness publishes `stream` as a one-argument method that enters the
+* waterfall itself, so a property rewrite carrying a second `next` parameter
+* breaks every caller dispatching it as a method, and misses the prepared
+* call that skips the published method entirely.
+*
+* The guard rewrites nothing and appends nothing to the aborted attempt. Its
+* only model-visible input is the single resume message it queues afterwards,
+* on a fresh turn.
+*/
 const DEFAULT_GUARD_OPTIONS = {
 	windowChars: 4096,
 	minWindowChars: 1500,
@@ -37778,17 +39514,35 @@ function createBreakerState() {
 	};
 }
 /**
-* The resume message is a literal user message so the guard loads on every
-* generation without importing a generation-bound message factory.
+* The resume message the guard queues onto the kept inbox.
+*
+* It is built with the harness message factory and carries this package's own
+* source kind, because the harness rejects a message that is not fully
+* identified: a steer payload with no `id` is refused by
+* `assertMessageEventShape` (`session event … lacks an identified message`)
+* and one with no `source` by the same check (`… has invalid source`). Either
+* refusal is thrown from `session.append` at the very moment the guard's
+* microtask steers — after the abort has already been taken — so the failure
+* surfaced as a repairable-splice violation, the attempt was already cancelled,
+* and the turn ended with the user told nothing.
+*
+* `createUserMessage` is available on every supported generation, and the
+* `dsh-chatgpt-subscription` kind is this package's own entry in the
+* merge-extensible `MessageSourceMap`, so the resume carries real provenance
+* instead of impersonating a human turn.
 */
 function createResumeMessage(text) {
-	return {
-		role: "user",
+	return createUserMessage({
 		content: [{
 			type: "text",
 			text
-		}]
-	};
+		}],
+		source: {
+			kind: PLUGIN_MESSAGE_SOURCE_KIND,
+			form: "notice",
+			summary: boundContextSummary("Reasoning collapsed; the attempt was stopped and resumed.")
+		}
+	});
 }
 function appendWindow(current, addition, limit) {
 	const combined = current + addition;
@@ -38007,12 +39761,12 @@ function apply(ctx, pluginConfig = {}) {
 		const claimAntigravityRoute = () => {
 			if (antigravityRegistration !== void 0) return;
 			try {
-				antigravityRegistration = ctx.llm.registerAdapter([PROVIDER_ID$5], antigravityAdapter);
-				if (antigravityConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] Antigravity route "${PROVIDER_ID$5}" is now served by this plugin`);
+				antigravityRegistration = ctx.llm.registerAdapter([PROVIDER_ID$6], antigravityAdapter);
+				if (antigravityConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] Antigravity route "${PROVIDER_ID$6}" is now served by this plugin`);
 				antigravityConflict = null;
 			} catch (error) {
 				antigravityConflict = error instanceof Error ? error.message : String(error);
-				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$5}" is already owned by another adapter; Antigravity models keep being served by that one until its configuration is removed (${antigravityConflict})`);
+				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$6}" is already owned by another adapter; Antigravity models keep being served by that one until its configuration is removed (${antigravityConflict})`);
 			}
 		};
 		claimAntigravityRoute();
@@ -38031,6 +39785,12 @@ function apply(ctx, pluginConfig = {}) {
 		}, commandCodeAccountPool);
 		let commandCodeRegistration;
 		let commandCodeConflict = null;
+		const ollamaStore = new FileCredentialStore$5();
+		const ollamaModelSettings = new FileModelSettingsStore$5();
+		const ollamaAccountPool = new OllamaAccountPool({ store: ollamaStore });
+		const ollamaAdapter = new OllamaAdapter(ollamaStore, ollamaModelSettings, { fetchFn: proxyFetch }, ollamaAccountPool);
+		let ollamaRegistration;
+		let ollamaConflict = null;
 		const kimiCodeAdapter = new KimiCodeAdapter(kimiCodeStore, kimiCodeModelSettings, kimiCodePreferences, {
 			fetchFn: proxyFetch,
 			attachments: ctx.attachments,
@@ -38056,15 +39816,32 @@ function apply(ctx, pluginConfig = {}) {
 		const kimiCodeRouteWatch = typeof ctx.on === "function" ? ctx.on("llm/adapters-updated", () => {
 			claimKimiCodeRoute();
 		}) : void 0;
+		const claimOllamaRoute = () => {
+			if (ollamaRegistration !== void 0) return;
+			try {
+				ollamaRegistration = ctx.llm.registerAdapter([PROVIDER_ID$4], ollamaAdapter);
+				if (ollamaConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] ${PROVIDER_NAME$4} route "${PROVIDER_ID$4}" is now served by this plugin`);
+				ollamaConflict = null;
+			} catch (error) {
+				ollamaConflict = error instanceof Error ? error.message : String(error);
+				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$4}" is already owned by another adapter; ${PROVIDER_NAME$4} models keep being served by that one until its configuration is removed (${ollamaConflict})`);
+			}
+		};
+		claimOllamaRoute();
+		const disposeOllamaRoutes = registerOllamaRoutes(ctx, {
+			accountPool: ollamaAccountPool,
+			modelSettings: ollamaModelSettings,
+			fetchFn: proxyFetch
+		});
 		const claimCommandCodeRoute = () => {
 			if (commandCodeRegistration !== void 0) return;
 			try {
-				commandCodeRegistration = ctx.llm.registerAdapter([PROVIDER_ID$4], commandCodeAdapter);
-				if (commandCodeConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] ${PROVIDER_NAME$4} route "${PROVIDER_ID$4}" is now served by this plugin`);
+				commandCodeRegistration = ctx.llm.registerAdapter([PROVIDER_ID$5], commandCodeAdapter);
+				if (commandCodeConflict !== null) ctx.logger.info(`[dsh-chatgpt-subscription] ${PROVIDER_NAME$5} route "${PROVIDER_ID$5}" is now served by this plugin`);
 				commandCodeConflict = null;
 			} catch (error) {
 				commandCodeConflict = error instanceof Error ? error.message : String(error);
-				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$4}" is already owned by another adapter; ${PROVIDER_NAME$4} models keep being served by that one until its configuration is removed (${commandCodeConflict})`);
+				ctx.logger.warn(`[dsh-chatgpt-subscription] provider route "${PROVIDER_ID$5}" is already owned by another adapter; ${PROVIDER_NAME$5} models keep being served by that one until its configuration is removed (${commandCodeConflict})`);
 			}
 		};
 		claimCommandCodeRoute();
@@ -38247,7 +40024,7 @@ function apply(ctx, pluginConfig = {}) {
 			});
 		};
 		const disposeRoutes = registerRoutes(ctx, oauth, usage, preferences, proxyManager, searchSwitcher, readRouteAudit, codexAccountPool);
-		const disposeAdapter = ctx.llm.registerAdapter([PROVIDER_ID$6], adapter);
+		const disposeAdapter = ctx.llm.registerAdapter([PROVIDER_ID$7], adapter);
 		const disposeImageTool = ctx.tools.register(createCodexImageTool(oauth, ctx.attachments, { fetchFn: proxyFetch }));
 		const disposeVideoTool = ctx.tools.register(createKimiVideoTool(ctx, { fetchFn: proxyFetch }));
 		ctx.inject(["web"], (ctx) => {
@@ -38273,6 +40050,9 @@ function apply(ctx, pluginConfig = {}) {
 			releaseHandle(antigravityRouteWatch);
 			antigravityRegistration?.();
 			antigravityRegistration = void 0;
+			disposeOllamaRoutes();
+			ollamaRegistration?.();
+			ollamaRegistration = void 0;
 			disposeCommandCodeRoutes();
 			releaseHandle(commandCodeRouteWatch);
 			commandCodeRegistration?.();