pi-lilac-provider 1.8.2 → 1.9.1

package/README.md

@@ -74,7 +74,7 @@ pi
 | Model | Context | Vision | Reasoning | Input $/M | Cache Read $/M | Output $/M |
 |-------|---------|--------|-----------|-----------|-----------------|------------|
 | Gemma 4 | 262K | ✅ | ✅ | $0.11 | — | $0.35 |
-| GLM 5.2 | 524K | ❌ | ✅ | $0.90 | $0.27 | $3.00 |
+| GLM 5.2 | 524K | ❌ | ✅ | $0.90 | $0.17 | $3.00 |
 | Kimi K2.6 | 262K | ✅ | ✅ | $0.70 | $0.20 | $3.50 |
 | MiniMax M3 | 1.0M | ✅ | ✅ | $0.28 | $0.05 | $1.10 |
 
@@ -102,10 +102,7 @@ pi --provider lilac --model moonshotai/kimi-k2.6
 
 ### Thinking Mode
 
-All Lilac models toggle reasoning via `chat_template_kwargs`, but the key each
-model's chat template honors differs per family. The provider uses pi's
-`chat-template` thinkingFormat with per-model `chatTemplateKwargs` (configured in
-`patch.json`) so the right key reaches each template:
+All Lilac models toggle reasoning via `chat_template_kwargs`, but the key each model's chat template honors differs per family. The provider uses pi's `chat-template` thinkingFormat with per-model `chatTemplateKwargs` (configured in `patch.json`) so the right key reaches each template:
 
 | Model | Reasoning key | Default |
 |-------|---------------|---------|
@@ -116,21 +113,9 @@ model's chat template honors differs per family. The provider uses pi's
 | MiniMax M2.7 | `thinking` + `enable_thinking` (bool) | on |
 | MiniMax M3 | `thinking_mode` (`disabled`\|`adaptive`\|`enabled`) | adaptive (server) |
 
-Kimi K2.6, GLM 5.1, Gemma 4, and MiniMax M2.7 use the forward-compatible form
-that sends **both** `thinking` and `enable_thinking`, so whichever key the
-template honors is set. GLM 5.2 additionally maps pi's thinking levels to
-`reasoning_effort` (`high` = lower-latency, `max` = deepest). MiniMax M3 uses
-the `thinking_mode` enum, exposed as three pi thinking levels: `off` →
-`disabled` (never think), `minimal` → `adaptive` (the model decides), `high` →
-`enabled` (always think). Pi starts at `off` (`disabled`); cycle to `minimal`
-for M3's adaptive "model decides" mode. (The selector/footer show pi's level
-names — `minimal`/`high` — not the `thinking_mode` values; pi has no per-model
-level-relabel hook.)
-
-**Preserved thinking (full-history reasoning).** By default these templates
-trim older assistant reasoning between turns (each vendor's default), which
-degrades multi-turn recall. Three models opt into full-history preservation via
-a template flag sent alongside the reasoning key:
+Kimi K2.6, GLM 5.1, Gemma 4, and MiniMax M2.7 use the forward-compatible form that sends **both** `thinking` and `enable_thinking`, so whichever key the template honors is set. GLM 5.2 additionally maps pi's thinking levels to `reasoning_effort` (`high` = lower-latency, `max` = deepest). MiniMax M3 uses the `thinking_mode` enum, exposed as three pi thinking levels: `off` → `disabled` (never think), `minimal` → `adaptive` (the model decides), `high` → `enabled` (always think). Pi starts at `off` (`disabled`); cycle to `minimal` for M3's adaptive "model decides" mode. (The selector/footer show pi's level names — `minimal`/`high` — not the `thinking_mode` values; pi has no per-model level-relabel hook.)
+
+**Preserved thinking (full-history reasoning).** By default these templates trim older assistant reasoning between turns (each vendor's default), which degrades multi-turn recall. Three models opt into full-history preservation via a template flag sent alongside the reasoning key:
 
 | Model | Flag | Effect |
 |-------|------|--------|
@@ -138,14 +123,9 @@ a template flag sent alongside the reasoning key:
 | GLM 5.1 | `clear_thinking: false` | keeps reasoning for all turns (default: clears before the last user message) |
 | GLM 5.2 | `clear_thinking: false` | keeps reasoning for all turns (default: clears before the last user message) |
 
-Kimi K2.6 and GLM 5.2 are E2E-verified on the sibling neuralwatt provider via a
-3-turn, two-20-digit-number recall test (Kimi 0/6 → 6/6, GLM 5.2 1/4 → 4/4);
-GLM 5.1 uses the same `clear_thinking` mechanism (confirmed in its HuggingFace
-chat template). Gemma 4 and MiniMax M2.7/M3 expose no family-wide preserve flag,
-so their older assistant reasoning is trimmed per the template default.
+Kimi K2.6 and GLM 5.2 are E2E-verified on the sibling neuralwatt provider via a 3-turn, two-20-digit-number recall test (Kimi 0/6 → 6/6, GLM 5.2 1/4 → 4/4); GLM 5.1 uses the same `clear_thinking` mechanism (confirmed in its HuggingFace chat template). Gemma 4 and MiniMax M2.7/M3 expose no family-wide preserve flag, so their older assistant reasoning is trimmed per the template default.
 
-In pi, reasoning models automatically use the appropriate thinking format. Use
-Shift+Tab to control thinking level.
+In pi, reasoning models automatically use the appropriate thinking format. Use Shift+Tab to control thinking level.
 
 ### Vision
 

package/index.ts

@@ -128,7 +128,7 @@ interface JsonModel {
   id: string;
   name: string;
   reasoning: boolean;
-  input: string[];
+  input: ("text" | "image")[];
   cost: {
     input: number;
     output: number;
@@ -163,7 +163,7 @@ interface JsonModel {
 interface PatchEntry {
   name?: string;
   reasoning?: boolean;
-  input?: string[];
+  input?: ("text" | "image")[];
   cost?: {
     input?: number;
     output?: number;
@@ -214,7 +214,7 @@ function parseModelOverrides(raw: unknown): Record<string, ModelOverride> | unde
     if (o.thinkingLevelMap && typeof o.thinkingLevelMap === "object") {
       const m: Record<string, string | null> = {};
       for (const [k, v] of Object.entries(o.thinkingLevelMap as Record<string, unknown>)) {
-        if (v === null || typeof v === "string") m[k] = v;
+        if (v === null || typeof v === "string") m[k] = v as string | null;
       }
       if (Object.keys(m).length > 0) parsed.thinkingLevelMap = m as ThinkingLevelMap;
     }
@@ -450,7 +450,7 @@ function transformApiModel(apiModel: any): JsonModel | null {
   // multiply and preserves sub-cent cache prices like 0.003 ($/M).
   const toPerM = (v: any) => Math.round((typeof v === "string" ? parseFloat(v) : (v || 0)) * 1_000_000 * 1e6) / 1e6;
 
-  const inputTypes: string[] = ["text"];
+  const inputTypes: ("text" | "image")[] = ["text"];
   if (hasImage) inputTypes.push("image");
   // Video is sent as image frames, so we don't add a separate "video" input type
 
@@ -1228,7 +1228,56 @@ export default function (pi: ExtensionAPI) {
     description: "Configure Lilac: flex discount threshold + preserved thinking per model",
     async handler(_args, ctx) {
       if (ctx.mode !== "tui") {
-        ctx.ui.notify("/lilac-settings requires TUI mode.", "error");
+        // GUI/RPC fallback: a select -> select flow. The custom SettingsList is
+        // TUI-only. One setting per run; per-model preserved thinking is a
+        // pick-model -> pick-value pair.
+        if (!ctx.hasUI) {
+          ctx.ui.notify("/lilac-settings requires a UI (TUI or GUI).", "error");
+          return;
+        }
+        const currentFlex = getConfig().flexThreshold ?? null;
+        const guiItems = [
+          { id: "flex", label: "Flex threshold", current: currentFlex == null ? "off" : String(currentFlex), values: ["off", "50", "75"] },
+          { id: "preserved-thinking", label: "Preserved thinking", current: "configure" },
+        ];
+        const pick = await ctx.ui.select("Lilac settings — pick a setting", guiItems.map((i) => `${i.label}: ${i.current}`));
+        if (pick === undefined) return;
+        const item = guiItems.find((i) => pick.startsWith(`${i.label}:`));
+        if (!item) return;
+        if (item.id === "flex") {
+          const v = await ctx.ui.select("Flex threshold", ["off", "50", "75"]);
+          if (v === undefined) return;
+          applyFlexThreshold(v === "off" ? null : Number(v), ctx);
+        } else {
+          const fresh = collectPreserveState();
+          if (fresh.length === 0) { ctx.ui.notify("No models support preserved thinking.", "info"); return; }
+          const modelPick = await ctx.ui.select("Preserved thinking — pick a model", fresh.map((e) => `${e.name}: ${e.preserved ? "Preserve Thinking" : "Clear Thinking"}`));
+          if (modelPick === undefined) return;
+          const entry = fresh.find((e) => modelPick.startsWith(`${e.name}:`));
+          if (!entry) return;
+          const v = await ctx.ui.select(entry.name, ["Preserve Thinking", "Clear Thinking"]);
+          if (v === undefined) return;
+          const preservedOn = v === "Preserve Thinking";
+          const flagValue = entry.flag === "clear_thinking" ? !preservedOn : preservedOn;
+          updateConfig((cfg) => {
+            const overrides = cfg.modelOverrides ?? (cfg.modelOverrides = {});
+            const ov = overrides[entry.id] ?? (overrides[entry.id] = {});
+            const compat = ov.compat ?? (ov.compat = {});
+            const kwargs = compat.chatTemplateKwargs ?? (compat.chatTemplateKwargs = {});
+            kwargs[entry.flag] = flagValue;
+            return cfg;
+          });
+          listModelsCache = null;
+          pi.registerProvider("lilac", {
+            baseUrl: BASE_URL,
+            apiKey: "$LILAC_API_KEY",
+            api: "openai-completions",
+            models: applyDiscounts(getListModels(), latestDiscounts),
+          });
+          syncStatus(ctx);
+          ctx.ui.notify(`Preserved thinking ${preservedOn ? "on" : "off"} for ${entry.name} — takes effect now.`, "info");
+        }
+        ctx.ui.notify("Run /lilac-settings again for more.", "info");
         return;
       }
       const { SettingsList, Container } = await import("@earendil-works/pi-tui");

package/models.json

@@ -33,7 +33,7 @@
     "cost": {
       "input": 0.9,
       "output": 3,
-      "cacheRead": 0.27,
+      "cacheRead": 0.17,
       "cacheWrite": 0
     },
     "contextWindow": 524288,
@@ -56,7 +56,7 @@
     "cost": {
       "input": 0.7,
       "output": 3.5,
-      "cacheRead": 0.2,
+      "cacheRead": 0.16,
       "cacheWrite": 0
     },
     "contextWindow": 262144,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-lilac-provider",
-  "version": "1.8.2",
+  "version": "1.9.1",
   "description": "Lilac provider extension for pi - Access Kimi K2.6, GLM 5.1, and Gemma 4 models through Lilac's OpenAI-compatible API on idle GPUs",
   "type": "module",
   "main": "index.ts",
@@ -29,7 +29,10 @@
     "@earendil-works/pi-coding-agent": ">=0.74.0"
   },
   "devDependencies": {
-    "@earendil-works/pi-coding-agent": "0.80.3"
+    "@earendil-works/pi-coding-agent": "0.80.3",
+    "typescript": "6.0.3",
+    "@types/node": "25.9.1",
+    "knip": "6.14.1"
   },
   "pi": {
     "extensions": [
@@ -39,7 +42,7 @@
   "scripts": {
     "clean": "echo 'nothing to clean'",
     "build": "echo 'nothing to build'",
-    "check": "echo 'nothing to check'",
+    "check": "tsc --noEmit && knip --no-gitignore",
     "test": "node scripts/test-discounts.ts && node scripts/test-preserved-thinking.ts && node scripts/test-model-overrides.ts && node scripts/test-flex.ts",
     "test:discounts": "node scripts/test-discounts.ts",
     "test:thinking": "node scripts/test-preserved-thinking.ts",