pi-multikey 1.14.0 → 1.15.0

package/README.md

@@ -40,6 +40,8 @@ Built-in presets decouple "model settings" from "keys". The data comes from b.ai
 | hy3 | 256K / 128K | text | off · low · high |
 | mimo-v2.5 | 1M / 128K | text+image | off · high (official: low/medium/high behave identically) |
 | qwen3.8-flash | 1M / 131K | text+image | off · low · medium · xhigh |
+| deepseek-v4.1-flash | 1M / 384K | text+image | off · low · high · max (official DeepSeek V4 tiers; b.ai probed live 2026-09-25) |
+| glm-5.3-flash | 1M / 131K | text+image | low · high · max (no off — GLM always thinks; b.ai probed live 2026-09-25) |
 
 > Why `null` must be explicit: pi's `getSupportedThinkingLevels` treats `mapped === null` as unsupported and hides that level, but **omitting** it is treated as supported and the level name is sent to the API verbatim; `xhigh` / `max` additionally require an explicit non-null value to be usable.
 

package/README.zh.md

@@ -41,6 +41,8 @@ DeepSeek / Tencent / 小米官方文档，并对每个 thinking 档位做过实
 | hy3 | 256K / 128K | text | off · low · high |
 | mimo-v2.5 | 1M / 128K | text+image | off · high（官方：low/medium/high 行为相同） |
 | qwen3.8-flash | 1M / 131K | text+image | off · low · medium · xhigh |
+| deepseek-v4.1-flash | 1M / 384K | text+image | off · low · high · max（官方 DeepSeek V4 档位；b.ai 已于 2026-09-25 实测） |
+| glm-5.3-flash | 1M / 131K | text+image | low · high · max（无 off——GLM 始终思考；b.ai 已于 2026-09-25 实测） |
 
 > 为什么必须显式写 `null`：pi 的 `getSupportedThinkingLevels` 把 `mapped === null`
 > 视为不支持并隐藏该档，但**省略**会被当作支持并把档名原样发给 API；

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "pi-multikey",
-	"version": "1.14.0",
+	"version": "1.15.0",
 	"description": "One pi provider backed by many API keys: automatic 429 rotation, per-request key leases for concurrent subagents, and a /multikey management TUI",
 	"keywords": [
 		"pi-package",

package/presets.ts

@@ -102,31 +102,41 @@ export const PRESETS: Preset[] = [
 				// consensus (models.dev opencode/greenpt rows + the same model
 				// behind cline-free/deepseek-v4.1-flash, verified live 2026-09-17):
 				// ctx 1M, out 384K. Text+image input (models.dev opencode row).
-				// Thinking tiers UNVERIFIED on b.ai: conservative off/high only —
-				// b.ai rejects minimal/xhigh/max with HTTP 400 (seen on mimo-v2.5).
-				// Widen the map after probing with a paid key.
+				// Thinking tiers probed live on b.ai 2026-09-25 (paid key, exact pi
+				// request shape: thinking.enabled + reasoning_effort): every level is
+				// accepted — the mimo-v2.5-style HTTP 400 for minimal/xhigh/max does
+				// NOT apply to this model. DeepSeek official exposes only
+				// low/high/max, and low vs high are behaviorally distinct in reasoning
+				// length, so minimal/medium/xhigh stay hidden (b.ai accepts them but
+				// treats them as duplicates).
 				id: "deepseek-v4.1-flash",
 				name: "DeepSeek V4.1 Flash",
 				reasoning: true,
 				input: ["text", "image"],
 				contextWindow: 1_000_000,
 				maxTokens: 384_000,
-				thinkingLevelMap: levels({ off: "none", high: "high" }),
+				thinkingLevelMap: levels({ off: "none", low: "low", high: "high", max: "max" }),
 			},
 			{
 				// b.ai /v1/models lists `glm-5.3-flash` (bare ids only — verified
 				// live 2026-09-21). Sizes from catalog consensus (models.dev
 				// zhipuai/zai rows): ctx 1M, out 128K. Upstream inputs
 				// text/image/video/pdf (pi tracks text + image, like mimo-v2.5).
-				// Thinking tiers UNVERIFIED on b.ai: conservative off/high only
-				// (see deepseek-v4.1-flash above). Widen after probing.
+				// Thinking tiers probed live on b.ai 2026-09-25 (paid key, exact pi
+				// request shape): low/high/max accepted and behaviorally distinct
+				// (completion 153/257/434 tok on a fixed prompt). medium/minimal are
+				// REJECTED (HTTP 400), and thinking:{type:"disabled"} (what pi sends
+				// for off) also 400s — so off stays hidden: GLM always thinks, with
+				// the plain no-param request defaulting to deep reasoning (verified
+				// 200). xhigh is accepted with ≈max-like depth on a single sample but
+				// upstream GLM 5.3 Flash declares no xhigh tier — hidden for now.
 				id: "glm-5.3-flash",
 				name: "GLM 5.3 Flash",
 				reasoning: true,
 				input: ["text", "image"],
 				contextWindow: 1_000_000,
 				maxTokens: 131_072,
-				thinkingLevelMap: levels({ off: "none", high: "high" }),
+				thinkingLevelMap: levels({ off: null, low: "low", high: "high", max: "max" }),
 			},
 		],
 	},
@@ -276,6 +286,7 @@ export const PRESETS: Preset[] = [
 				// (deepseek/deepseek-v4.1-flash) is usage-billed — see the comment above.
 				// Catalog (openrouter): ctx 1048576, out 384000 (both verified live), text+image.
 				// Effort tiers verified live via reasoning:{effort} (incl. "none" = reasoning off).
+				// DeepSeek official tiers are only low/high/max, so minimal/medium/xhigh stay hidden.
 				id: "cline-free/deepseek-v4.1-flash",
 				name: "DeepSeek V4.1 Flash (Free)",
 				reasoning: true,
@@ -283,7 +294,7 @@ export const PRESETS: Preset[] = [
 				contextWindow: 1_048_576,
 				maxTokens: 384_000,
 				compat: { thinkingFormat: "openrouter" },
-				thinkingLevelMap: levels({ off: "none", low: "low", medium: "medium", high: "high", xhigh: "xhigh", max: "max" }),
+				thinkingLevelMap: levels({ off: "none", low: "low", high: "high", max: "max" }),
 			},
 			{
 				// Free at its raw id (no cline-free/ prefix in the feed); 262K window measured