agentsw 0.11.2 → 0.12.1

package/README.md

@@ -35,7 +35,7 @@ without touching the settings you hand-tuned.
 - **Model metadata, not just an id list.** Reseller `/v1/models` endpoints return bare ids.
   agentsw enriches them from [models.dev](https://models.dev), automatically fills meaningful gaps from AI Gateway when needed, and writes what each app
   actually reads: `thinkingLevelMap` for pi/prime, `limit`/`attachment` for opencode,
-  `contextWindow`/`maxTokens` for omp, `reasoningEfforts`/`input` for dsh. Limits the endpoint declares for
+  `contextWindow`/`maxTokens` for omp, `reasoningEfforts`/`input` for dsh. Limits and prices the endpoint declares for
   its own models outrank the catalogs, so one model cannot end up with two different context windows
   depending on which provider lists it.
 - **Adopts what you already have** — including [cc-switch](https://github.com/farion1231/cc-switch).
@@ -207,7 +207,7 @@ use `--all` explicitly in scripts. Model discovery may update the local metadata
 ## Model discovery filters
 
 `--discover` lists model ids from the provider's `/v1/models`, then enriches each id from models.dev first,
-with automatic AI Gateway supplementation when needed. Limits the provider declares for its own models in
+with automatic AI Gateway supplementation when needed. Limits and prices the provider declares for its own models in
 that same listing win over what a catalog claims for them; ids it lists bare are enriched as before. Reseller lists are noisy,
 so filters are persisted per provider and re-applied on every `discover`:
 
@@ -283,7 +283,12 @@ Import still skips already configured accounts.
 - **Field-level priority:** manual values > the endpoint's own declaration > models.dev > Gateway gap filling.
   An endpoint that advertises `context_window_tokens`/`maximum_output_tokens`
   (or their `context_length`/`max_output_tokens` spellings) for its own list is authoritative for its own
-  route, so those limits are written instead of a catalog's estimate for the same model. Tracked automatic
+  route, so those limits are written instead of a catalog's estimate for the same model. The same applies
+  to prices a route publishes for its own models — `input_cost_per_million_tokens_usd`,
+  `output_cost_per_million_tokens_usd`, `cached_input_cost_per_million_tokens_usd`,
+  `cache_write_cost_per_million_tokens_usd`, or the nano-USD `pricing.*_nano_usd_per_million_tokens`
+  form — which are stored as USD per million tokens. Rates priced per image, second or character are
+  not token prices, and a tier rate beside the base one is not the base price, so neither is written. Tracked automatic
   values may refresh; subsequent manual edits are retained. Legacy values without provenance are conservatively
   treated as manual overrides. Unmodeled extension fields survive enrichment too.
 - **Exact identities:** automatic lookup can map a bare, case-sensitive ID (no `/`) to a unique Gateway
@@ -320,7 +325,9 @@ Import still skips already configured accounts.
   written once rather than once per provider: it gets the active provider when it can accept it, or the single
   provider in the run that it can (Claude Code takes anthropic endpoints only, so an openai active provider does not
   starve it). With several candidates and the active provider not among them the choice is yours, so the app keeps
-  pointing where it already does: syncing is not switching. A provider whose endpoint cannot be reached is reported
+  pointing where it already does: syncing is not switching. The same holds for an app that keeps a pointer of its
+  own — only the store's active provider may move it, so `sync -p other` writes that provider's block and models and
+  leaves the app's current model alone (`use` and `add` are what switch). A provider whose endpoint cannot be reached is reported
   and skipped, the rest still sync, and the run exits non-zero.
 - **Automatic lookup is not an automatic push:** enrichment updates saved metadata during add/quick/discover/import/refresh
   and never writes agent configs by itself. `sync` is the push, and it refreshes first: it fetches the provider's model

package/README.zh-CN.md

@@ -33,7 +33,7 @@
 - **写的是元数据，不只是模型 id。** 转售商 `/v1/models` 只返回裸 id；agentsw 从
   [models.dev](https://models.dev) 优先补全，并在有实际缺口时自动查询 AI Gateway，按各应用真正读取的字段下发：pi/prime 的
   `thinkingLevelMap`、opencode 的 `limit`/`attachment`、omp 的 `contextWindow`/`maxTokens`、
-  dsh 的 `reasoningEfforts`/`input`。端点自己声明的上下文/输出上限优先于目录，避免同一模型在不同供应商下被写成不同窗口。
+  dsh 的 `reasoningEfforts`/`input`。端点自己声明的上下文/输出上限与价格优先于目录，避免同一模型在不同供应商下被写成不同窗口。
 - **接管你已有的配置——包括 [cc-switch](https://github.com/farion1231/cc-switch)。**
   `import` 会扫描所有已安装的智能体**以及 cc-switch 自己的存储**，合并重复项
   （相同协议 + 归一化端点 + 凭据），并且对两者都只读不写。
@@ -191,7 +191,7 @@ imported zhipu-glm · anthropic · https://open.bigmodel.cn/api/anthropic · 1 m
 ## 模型发现过滤
 
 `--discover` 从供应商 `/v1/models` 拉取模型 id，优先从 models.dev 补全元数据，必要时自动查询 AI Gateway。
-同一份列表里供应商为自家模型声明的上下文/输出上限优先于目录的估算，只给出裸 id 的模型仍按原逻辑补全。
+同一份列表里供应商为自家模型声明的上下文/输出上限与价格优先于目录的估算，只给出裸 id 的模型仍按原逻辑补全。
 转售商列表噪音大，过滤规则持久化在供应商上，每次 `discover` 自动复用：
 
 ```bash
@@ -263,7 +263,11 @@ asw sync --provider myproxy,backup-proxy
 
 - **按字段合并：** 用户手动值优先，其次端点自己声明的值，然后 models.dev，最后 Gateway 补缺。
   端点在 `/v1/models` 中为自家模型声明了 `context_window_tokens`/`maximum_output_tokens`（或
-  `context_length`/`max_output_tokens` 写法）时，以它为准而不是目录对同一模型的估算。已记录来源的自动值可以刷新，
+  `context_length`/`max_output_tokens` 写法）时，以它为准而不是目录对同一模型的估算；自家发布的
+  价格（`input_cost_per_million_tokens_usd`、`output_cost_per_million_tokens_usd`、
+  `cached_input_cost_per_million_tokens_usd`、`cache_write_cost_per_million_tokens_usd`，或 `pricing` 下的
+  `*_nano_usd_per_million_tokens` 纳美元写法；一律按每百万 token 美元存储）同理。按图片/秒/字符计费的不是 token 价格，
+  基础价之外的档位价也不是基础价，两者都不写。已记录来源的自动值可以刷新，
   手动修改后会保留；旧版本中没有来源记录的已有值保守视为手动值，不强制覆盖。未知扩展字段也保留。
 - **精确匹配：** 不含 `/` 的裸 ID 可按大小写精确匹配唯一的 Gateway `creator/model` 身份，
   前提是 models.dev 没有身份线索或线索一致。任一目录中的创建者身份有歧义时拒绝自动匹配；

package/dist/commands.js

@@ -203,7 +203,7 @@ export async function cmdAdd(opts) {
     const manualIds = (answers.models ?? "").split(",").map((s) => s.trim()).filter(Boolean);
     let modelIds = manualIds;
     // Limits the endpoint declares for its own models, kept from the discovery pass.
-    let endpointLimits = new Map();
+    let endpointFacts = new Map();
     const modelFilter = parseFilterOpts(opts, existing?.modelFilter);
     const shouldDiscover = opts.discover || manualIds.length === 0;
     if (shouldDiscover) {
@@ -213,7 +213,7 @@ export async function cmdAdd(opts) {
         const discover = () => withProgress(t("add.discovering", { url: baseUrl }), (progress) => discoverProviderCatalog({ baseUrl, apiKey: answers.apiKey, protocol }, {
             onPage: ({ page, count }) => progress.update(t("progress.discoverPage", { url: baseUrl, page, count })),
         }));
-        let discovered = { ids: [], limits: new Map() };
+        let discovered = { ids: [], facts: new Map() };
         for (let attempt = 0;; attempt++) {
             try {
                 discovered = await discover();
@@ -233,7 +233,7 @@ export async function cmdAdd(opts) {
         if (discovered.ids.length)
             out(t("add.providerLists", { count: discovered.ids.length }));
         modelIds = [...new Set([...manualIds, ...discovered.ids])];
-        endpointLimits = discovered.limits;
+        endpointFacts = discovered.facts;
     }
     const retained = [opts.defaultModel === undefined ? existing?.defaultModel : undefined, opts.smallModel === undefined ? existing?.smallModel : undefined].filter((id) => !!id);
     modelIds = [...new Set([...modelIds, ...retained])];
@@ -245,7 +245,7 @@ export async function cmdAdd(opts) {
         fail(t("add.atLeastOne"));
     const catalog = await withProgress(t("progress.catalog"), () => loadCatalog());
     const hint = opts.modelsDev ?? existing?.modelsDevId ?? guessProviderHint(catalog, answers.baseUrl);
-    const models = await withProgress(t("progress.metadata"), () => enrichProviderModels(catalog, modelIds, { ...metadataOptions, modelsDevId: hint, models: existing?.models }, { endpointLimits }));
+    const models = await withProgress(t("progress.metadata"), () => enrichProviderModels(catalog, modelIds, { ...metadataOptions, modelsDevId: hint, models: existing?.models }, { endpointFacts }));
     const matched = models.filter((m) => m.contextWindow !== undefined).length;
     const defaultModel = opts.defaultModel ?? existing?.defaultModel ?? modelIds[0];
     if (!modelIds.includes(defaultModel))
@@ -292,7 +292,7 @@ async function createProvider(opts) {
     const hint = opts.modelsDev ?? opts.existing?.modelsDevId ?? guessProviderHint(catalog, baseUrl);
     const models = await withProgress(t("progress.metadata"), () => enrichProviderModels(catalog, modelIds, { ...metadataOptions, modelsDevId: hint, models: opts.existing?.models }, {
         gatewayLoader: opts.gatewayLoader,
-        endpointLimits: opts.endpointLimits,
+        endpointFacts: opts.endpointFacts,
     }));
     const matched = models.filter((m) => m.contextWindow !== undefined).length;
     const defaultModel = opts.defaultModel ?? opts.existing?.defaultModel ?? modelIds[0];
@@ -426,7 +426,7 @@ export async function cmdQuickAdd(opts) {
             gatewayModels: opts.gatewayModels,
             metadataMode: opts.metadataMode,
             gatewayLoader,
-            endpointLimits: discovered.limits,
+            endpointFacts: discovered.facts,
         });
         createdIds.push(provider.id);
         note(`\n${t("add.next", { id: provider.id })}`);
@@ -747,7 +747,7 @@ async function refreshProviderModels(store, provider, save = true) {
     const added = ids.filter((m) => !known.includes(m));
     const removed = known.filter((m) => !retained.has(m));
     const catalog = await withProgress(t("progress.catalog"), () => loadCatalog());
-    provider.models = await withProgress(t("progress.metadata"), () => enrichProviderModels(catalog, [...retained], { ...provider, modelsDevId: provider.modelsDevId ?? guessProviderHint(catalog, provider.baseUrl) }, { endpointLimits: listed.limits }));
+    provider.models = await withProgress(t("progress.metadata"), () => enrichProviderModels(catalog, [...retained], { ...provider, modelsDevId: provider.modelsDevId ?? guessProviderHint(catalog, provider.baseUrl) }, { endpointFacts: listed.facts }));
     // Reported by the caller, where it reads after the table or the per-app results instead of
     // scrolling away above them.
     const missing = [...new Set(pinned)].filter((model) => !ids.includes(model));
@@ -801,9 +801,10 @@ function resolveProviderIds(store, opts, fallback) {
     return ids;
 }
 /**
- * Apps that name a single provider are written once (see `RunTargetsOptions.oneSlot`); for the rest
- * the last writer owns whatever "current" pointer the app keeps, so a multi-provider sync finishes
- * on the active provider and leaves the agents where the store says they already are.
+ * Apps that name a single provider are written once (see `RunTargetsOptions.oneSlot`); every other
+ * app only lets the store's active provider move its "current" pointer (see `ApplyOptions.keepSelection`),
+ * so a multi-provider sync finishes on the active provider and a run that does not include it leaves
+ * the pointer where the app already had it.
  */
 function activeLast(store, ids) {
     const active = store.active;
@@ -895,7 +896,7 @@ async function syncProviders(opts, sourceStore) {
             }
         }
         out(`${t(opts.keepDefault ? "sync.syncingKeepDefault" : "sync.syncing", { id: pc.bold(id), model: provider.defaultModel })}\n`);
-        const outcome = await runWithOptionalDryRun("apply", provider, opts.apps, opts.dryRun, oneSlot, { keepDefault: opts.keepDefault });
+        const outcome = await runWithOptionalDryRun("apply", provider, opts.apps, opts.dryRun, oneSlot, { keepDefault: opts.keepDefault, keepSelection: provider.id !== store.active });
         if (outcome.failed?.length)
             failed.push(id);
         if (refresh?.missing.length) {
@@ -1381,7 +1382,7 @@ export async function cmdImport(opts) {
             continue;
         }
         let ids = [...c.models];
-        let endpointLimits = new Map();
+        let endpointFacts = new Map();
         if (ids.length === 0) {
             const discover = () => withProgress(t("import.discovering", { id: c.id }), (progress) => discoverProviderCatalog({ baseUrl: c.baseUrl, apiKey: apiKey, protocol: c.protocol }, {
                 onPage: ({ page, count }) => progress.update(t("progress.discoverPage", { url: c.baseUrl, page, count })),
@@ -1408,14 +1409,14 @@ export async function cmdImport(opts) {
                 continue;
             }
             ids = listed.ids;
-            endpointLimits = listed.limits;
+            endpointFacts = listed.facts;
             ids = applyModelFilter(ids, undefined, []).kept;
         }
         if (ids.length === 0)
             fail(t("import.noModelsImport", { id: c.id }));
         const id = availableProviderId(c.id, store.providers);
         const hint = guessProviderHint(catalog, c.baseUrl);
-        const models = await withProgress(t("progress.metadata"), () => enrichProviderModels(catalog, ids, { ...metadataOptions, modelsDevId: hint }, { gatewayLoader, endpointLimits }));
+        const models = await withProgress(t("progress.metadata"), () => enrichProviderModels(catalog, ids, { ...metadataOptions, modelsDevId: hint }, { gatewayLoader, endpointFacts }));
         const defaultModel = c.defaultModel && ids.includes(c.defaultModel) ? c.defaultModel : ids[0];
         const existed = store.providers[id] !== undefined;
         store.providers[id] = {

package/dist/discover.js

@@ -14,7 +14,7 @@ function modelListUrl(baseUrl) {
 /** Hard cap on model-list pages; an endpoint emitting endless fresh cursors must not loop forever. */
 export const MAX_MODEL_PAGES = 100;
 /**
- * Spellings endpoints use for the same limit, in priority order: the first
+ * Spellings endpoints use for the same fact, in priority order: the first
  * declared key wins when a row carries several. Providers do not agree on
  * names — an OpenAI-compatible gateway publishes context_window_tokens /
  * maximum_output_tokens, OpenRouter publishes context_length /
@@ -23,6 +23,35 @@ export const MAX_MODEL_PAGES = 100;
  */
 const contextWindowKeys = ["context_window_tokens", "context_length", "max_context_length"];
 const maxOutputKeys = ["maximum_output_tokens", "max_output_tokens", "max_completion_tokens"];
+const costSpellings = {
+    input: [["input_cost_per_million_tokens_usd", 1], ["pricing.input_nano_usd_per_million_tokens", 1e9]],
+    output: [["output_cost_per_million_tokens_usd", 1], ["pricing.output_nano_usd_per_million_tokens", 1e9]],
+    cacheRead: [["cached_input_cost_per_million_tokens_usd", 1], ["pricing.cached_input_nano_usd_per_million_tokens", 1e9]],
+    cacheWrite: [
+        ["cache_write_cost_per_million_tokens_usd", 1],
+        ["pricing.cache_write_nano_usd_per_million_tokens", 1e9],
+        ["pricing.cache_creation_input_nano_usd_per_million_tokens", 1e9],
+    ],
+};
+/** Read a dotted path, stopping at the first value that is not an object. */
+function readPath(row, path) {
+    let node = row;
+    for (const key of path.split(".")) {
+        if (!node || typeof node !== "object")
+            return undefined;
+        node = node[key];
+    }
+    return node;
+}
+/** A price keeps its fraction — a route charging 0.15 per million says so — and 0 is a real price. */
+function declaredCost(row, spellings) {
+    for (const [path, divisor] of spellings) {
+        const value = readPath(row, path);
+        if (typeof value === "number" && Number.isFinite(value) && value >= 0)
+            return value / divisor;
+    }
+    return undefined;
+}
 function declaredLimit(row, keys) {
     for (const key of keys) {
         const value = row[key];
@@ -36,13 +65,28 @@ function declaredLimit(row, keys) {
     }
     return undefined;
 }
-/** Limits a model row advertises, or undefined when it describes nothing but its id. */
-export function parseEndpointLimits(row) {
+/** What a model row declares about its own route, or undefined when it describes nothing but its id. */
+export function parseEndpointFacts(row) {
     const contextWindow = declaredLimit(row, contextWindowKeys);
     const maxOutput = declaredLimit(row, maxOutputKeys);
-    if (contextWindow === undefined && maxOutput === undefined)
+    const input = declaredCost(row, costSpellings.input);
+    const output = declaredCost(row, costSpellings.output);
+    const cacheRead = declaredCost(row, costSpellings.cacheRead);
+    const cacheWrite = declaredCost(row, costSpellings.cacheWrite);
+    const costs = {
+        ...(input !== undefined ? { input } : {}),
+        ...(output !== undefined ? { output } : {}),
+        ...(cacheRead !== undefined ? { cacheRead } : {}),
+        ...(cacheWrite !== undefined ? { cacheWrite } : {}),
+    };
+    const hasCost = Object.keys(costs).length > 0;
+    if (contextWindow === undefined && maxOutput === undefined && !hasCost)
         return undefined;
-    return { ...(contextWindow !== undefined ? { contextWindow } : {}), ...(maxOutput !== undefined ? { maxOutput } : {}) };
+    return {
+        ...(contextWindow !== undefined ? { contextWindow } : {}),
+        ...(maxOutput !== undefined ? { maxOutput } : {}),
+        ...(hasCost ? { cost: costs } : {}),
+    };
 }
 function parseModelPage(body) {
     if (!body || typeof body !== "object")
@@ -54,7 +98,7 @@ function parseModelPage(body) {
     if (page.has_more !== undefined && typeof page.has_more !== "boolean")
         throw new Error("invalid model-list pagination flag");
     const ids = [];
-    const limits = new Map();
+    const facts = new Map();
     for (const row of rows) {
         if (!row || typeof row !== "object")
             continue;
@@ -62,19 +106,19 @@ function parseModelPage(body) {
         if (typeof id !== "string" || id.length === 0)
             continue;
         ids.push(id);
-        const declared = parseEndpointLimits(row);
+        const declared = parseEndpointFacts(row);
         if (declared)
-            limits.set(id, declared);
+            facts.set(id, declared);
     }
-    return { ids, limits, hasMore: page.has_more === true, lastId: typeof page.last_id === "string" ? page.last_id : undefined };
+    return { ids, facts, hasMore: page.has_more === true, lastId: typeof page.last_id === "string" ? page.last_id : undefined };
 }
 /**
- * List the provider's own /v1/models catalog: model ids plus the limits a row
- * declares for itself. Plenty of resellers expose bare ids — those ids carry no
- * entry in `limits` and metadata is filled from models.dev afterwards
- * (enrichModels). An endpoint that does declare limits is authoritative for its
- * own route, so those numbers outrank the catalogs instead of being replaced by
- * whatever a third party lists for the same model.
+ * List the provider's own /v1/models catalog: model ids plus the limits and
+ * prices a row declares for itself. Plenty of resellers expose bare ids — those
+ * ids carry no entry in `facts` and metadata is filled from models.dev
+ * afterwards (enrichModels). An endpoint that declares its own numbers is
+ * authoritative for its own route, so they outrank the catalogs instead of
+ * being replaced by whatever a third party lists for the same model.
  */
 export async function discoverProviderCatalog(opts, hooks = {}) {
     const url = modelListUrl(opts.baseUrl);
@@ -82,7 +126,7 @@ export async function discoverProviderCatalog(opts, hooks = {}) {
         ? { "x-api-key": opts.apiKey, "anthropic-version": "2023-06-01" }
         : { authorization: `Bearer ${opts.apiKey}` };
     const ids = new Set();
-    const limits = new Map();
+    const facts = new Map();
     const cursors = new Set();
     const initialCursor = url.searchParams.get("after_id");
     if (initialCursor)
@@ -99,11 +143,11 @@ export async function discoverProviderCatalog(opts, hooks = {}) {
             ids.add(id);
         // A repeated id keeps what an earlier page declared: merge per key so a
         // later partial row cannot drop the other limit.
-        for (const [id, declared] of parsed.limits) {
+        for (const [id, declared] of parsed.facts) {
             if (!ids.has(id))
                 continue;
-            const earlier = limits.get(id);
-            limits.set(id, earlier ? { ...earlier, ...declared } : declared);
+            const earlier = facts.get(id);
+            facts.set(id, earlier ? { ...earlier, ...declared } : declared);
         }
         hooks.onPage?.({ page, count: ids.size });
         if (!parsed.hasMore)
@@ -117,7 +161,7 @@ export async function discoverProviderCatalog(opts, hooks = {}) {
     }
     if (ids.size === 0)
         throw new Error("provider returned no models");
-    return { ids: [...ids].sort(), limits };
+    return { ids: [...ids].sort(), facts };
 }
 /** The listed ids alone, for callers that only need the model list. */
 export async function discoverProviderModels(opts, hooks = {}) {

package/dist/metadata.js

@@ -70,13 +70,13 @@ const fields = [
     "cost.input", "cost.output", "cost.cacheRead", "cost.cacheWrite",
 ];
 const gatewayFields = fields.filter((field) => !field.startsWith("cost."));
-/** Limits are all an endpoint is trusted for: its listing describes its own wire, not capabilities. */
-const endpointFields = ["contextWindow", "maxOutput"];
+/** What a wire is trusted for: the limits it enforces and the prices it charges, not capabilities. */
+const endpointFields = ["contextWindow", "maxOutput", "cost.input", "cost.output", "cost.cacheRead", "cost.cacheWrite"];
 /**
  * Which source owns a field it proposes. An endpoint that declares its own
- * limits outranks both catalogs — models.dev describes the model, the public
- * Gateway describes *that* gateway, and neither is the route serving the
- * request. models.dev outranks Gateway, which only fills what it is given.
+ * limits and prices outranks both catalogs — models.dev describes the model,
+ * the public Gateway describes *that* gateway, and neither is the route serving
+ * the request. models.dev outranks Gateway, which only fills what it is given.
  */
 const sourceRank = { endpoint: 3, "models.dev": 2, "ai-gateway": 1 };
 const limits = ["maxInput", "maxOutput"];
@@ -387,8 +387,8 @@ export async function enrichProviderModels(catalog, ids, settings, options = {})
             }
             // The endpoint's own numbers go in first so the catalogs can only fill
             // what it left unsaid; a catalog that disagrees is reported as a conflict
-            // instead of overwriting the limits the route actually enforces.
-            const declared = options.endpointLimits?.get(id);
+            // instead of overwriting the numbers the route actually enforces.
+            const declared = options.endpointFacts?.get(id);
             if (declared)
                 mergeSource(spec, { id, ...declared }, "endpoint", id, now);
             let primary;

package/dist/targets/codex.js

@@ -29,7 +29,7 @@ export const codex = transactionalTarget({
         return [codexConfigFile(), codexAuthFile()];
     },
     detect: () => fs.existsSync(codexDir()),
-    async apply(provider) {
+    async apply(provider, options = {}) {
         const configFile = codexConfigFile();
         const authFile = codexAuthFile();
         const notes = [];
@@ -42,16 +42,18 @@ export const codex = transactionalTarget({
         if (config.model_providers !== undefined && !isJsonObject(config.model_providers)) {
             throw new Error(`${configFile}: expected model_providers to be a table`);
         }
-        config.model_provider = provider.id;
-        config.model = provider.defaultModel;
-        const model = provider.models.find((m) => m.id === provider.defaultModel);
-        if (provider.reasoningEffort && model?.reasoning !== false) {
-            config.model_reasoning_effort = provider.reasoningEffort;
-        }
-        else {
-            // The effort belongs to the active model: switching providers must not
-            // inherit the previous provider's level (prune treats it as owned too).
-            delete config.model_reasoning_effort;
+        if (!options.keepSelection) {
+            config.model_provider = provider.id;
+            config.model = provider.defaultModel;
+            const model = provider.models.find((m) => m.id === provider.defaultModel);
+            if (provider.reasoningEffort && model?.reasoning !== false) {
+                config.model_reasoning_effort = provider.reasoningEffort;
+            }
+            else {
+                // The effort belongs to the active model: switching providers must not
+                // inherit the previous provider's level (prune treats it as owned too).
+                delete config.model_reasoning_effort;
+            }
         }
         // Codex only speaks the OpenAI Responses API (wire_api = "chat" was removed Feb 2026).
         // requires_openai_auth = true makes Codex take the key from auth.json instead of an env var.

package/dist/targets/dsh.js

@@ -68,7 +68,7 @@ export const dsh = transactionalTarget({
         return [dshSettingsFile(), dshCredentialsFile()];
     },
     detect: () => fs.existsSync(dshDir()),
-    async apply(provider) {
+    async apply(provider, options = {}) {
         const notes = [];
         const file = dshSettingsFile();
         const doc = parseSettings(file);
@@ -117,7 +117,9 @@ export const dsh = transactionalTarget({
             notes.push("dropped modelOverrides (llm-pi-ai refuses it beside an explicit models list)");
         }
         // The picker writes this section whole; a partial write would keep a stale reasoningEffort.
-        doc.set("agent-default-model", doc.createNode({ provider: provider.id, model: provider.defaultModel }));
+        if (!options.keepSelection) {
+            doc.set("agent-default-model", doc.createNode({ provider: provider.id, model: provider.defaultModel }));
+        }
         const referenced = credentialReferences(doc, provider.id);
         const oldRef = typeof prev?.apiKeyEnv === "string" && isManagedCredentialRef(prev.apiKeyEnv, provider.id) && !referenced.has(prev.apiKeyEnv)
             ? prev.apiKeyEnv : undefined;

package/dist/targets/hermes.js

@@ -54,7 +54,7 @@ export const hermes = transactionalTarget({
     protocols: ["openai", "anthropic"],
     configPaths: [hermesConfigFile(), hermesEnvFile()],
     detect: () => fs.existsSync(hermesHome()),
-    async apply(provider) {
+    async apply(provider, options = {}) {
         const configFile = hermesConfigFile();
         const envFile = hermesEnvFile();
         const notes = [];
@@ -101,8 +101,10 @@ export const hermes = transactionalTarget({
             else
                 doc.deleteIn([...modelAt, "context_length"]);
         }
-        doc.setIn(["model", "provider"], provider.id);
-        doc.setIn(["model", "default"], provider.defaultModel);
+        if (!options.keepSelection) {
+            doc.setIn(["model", "provider"], provider.id);
+            doc.setIn(["model", "default"], provider.defaultModel);
+        }
         const envText = readTextIfExists(envFile) ?? "";
         const assignments = envAssignments(envFile, envText);
         const referenced = credentialReferences(doc, provider.id);

package/dist/targets/opencode.js

@@ -19,7 +19,7 @@ export const opencode = transactionalTarget({
     },
     detect: () => opencodeConfigFiles().some((file) => fs.existsSync(file)) ||
         opencodeConfigDirs().some((dir) => fs.existsSync(dir)),
-    async apply(provider) {
+    async apply(provider, options = {}) {
         const configFile = opencodePrimaryConfigFile();
         const notes = [];
         const existed = readTextIfExists(configFile) !== undefined;
@@ -94,14 +94,16 @@ export const opencode = transactionalTarget({
             models,
         };
         config.provider = providerMap;
-        config.model = `${provider.id}/${provider.defaultModel}`;
-        if (provider.smallModel) {
-            config.small_model = `${provider.id}/${provider.smallModel}`;
-        }
-        else if (typeof config.small_model === "string" && config.small_model.startsWith(`${provider.id}/`)) {
-            // This provider's own previous small model is obsolete; another provider's
-            // entry stays untouched, exactly as prune scopes its cleanup.
-            delete config.small_model;
+        if (!options.keepSelection) {
+            config.model = `${provider.id}/${provider.defaultModel}`;
+            if (provider.smallModel) {
+                config.small_model = `${provider.id}/${provider.smallModel}`;
+            }
+            else if (typeof config.small_model === "string" && config.small_model.startsWith(`${provider.id}/`)) {
+                // This provider's own previous small model is obsolete; another provider's
+                // entry stays untouched, exactly as prune scopes its cleanup.
+                delete config.small_model;
+            }
         }
         const backup = backupFile(configFile);
         if (backup)

package/dist/targets/pistyle.js

@@ -106,7 +106,7 @@ export function piStyleTarget(opts) {
             };
             modelsConfig.providers = providers;
             const modelsText = editJsoncObject(modelsDocument, modelsConfig);
-            const settingsText = options.keepDefault ? undefined : editJsoncObject(settingsDocument, {
+            const settingsText = options.keepDefault || options.keepSelection ? undefined : editJsoncObject(settingsDocument, {
                 ...settingsDocument.value,
                 defaultProvider: provider.id,
                 defaultModel: provider.defaultModel,

package/dist/targets/workbuddy.js

@@ -79,7 +79,7 @@ export const workbuddy = transactionalTarget({
         return [path.join(configDir(), "models.json"), path.join(configDir(), "settings.json")];
     },
     detect: () => fs.existsSync(configDir()),
-    async apply(provider) {
+    async apply(provider, options = {}) {
         const dir = configDir();
         const modelsFile = path.join(dir, "models.json");
         const settingsFile = path.join(dir, "settings.json");
@@ -127,7 +127,8 @@ export const workbuddy = transactionalTarget({
         const availableModels = [...new Set([
                 ...existingAvailable.filter((id) => !removedIds.has(id) || keptIds.has(id)), ...ourIds,
             ])];
-        settings.model = provider.defaultModel;
+        if (!options.keepSelection)
+            settings.model = provider.defaultModel;
         const modelsBackup = backupFile(modelsFile);
         if (modelsBackup)
             notes.push(`backup: ${modelsBackup}`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentsw",
-  "version": "0.11.2",
+  "version": "0.12.1",
   "description": "CLI provider switcher for coding agents: manage OpenAI/Anthropic-protocol providers in one place and sync them, with models.dev metadata, into Claude Code, Codex, omp, pi, prime-agent, opencode, Hermes, WorkBuddy and DeepSeek Harness",
   "type": "module",
   "repository": {