pi-lilac-provider 1.9.0 → 1.9.2

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/deprecated-models.json

@@ -0,0 +1 @@
+{}

package/index.ts

@@ -88,6 +88,7 @@ import { getAgentDir, type ExtensionAPI, type ModelRegistry } from "@earendil-wo
 import modelsData from "./models.json" with { type: "json" };
 import customModelsData from "./custom-models.json" with { type: "json" };
 import patchData from "./patch.json" with { type: "json" };
+import deprecatedData from "./deprecated-models.json" with { type: "json" };
 import fs from "fs";
 import path from "path";
 
@@ -128,7 +129,7 @@ interface JsonModel {
   id: string;
   name: string;
   reasoning: boolean;
-  input: string[];
+  input: ("text" | "image")[];
   cost: {
     input: number;
     output: number;
@@ -163,7 +164,7 @@ interface JsonModel {
 interface PatchEntry {
   name?: string;
   reasoning?: boolean;
-  input?: string[];
+  input?: ("text" | "image")[];
   cost?: {
     input?: number;
     output?: number;
@@ -379,6 +380,35 @@ function applyPatch(model: JsonModel, patch: PatchEntry): JsonModel {
 }
 
 /** Full pipeline: base models → patch → custom → user modelOverrides → result */
+// Grace period for delisted models. When the provider API stops listing a
+// model, update-models.js moves its last-known definition into
+// deprecated-models.json (stamped with deprecatedAt) instead of dropping it.
+// For 14 days the model keeps working here so in-flight sessions and saved
+// model settings do not break; afterwards it is evicted permanently.
+const DEPRECATED_MODEL_TTL_MS = 14 * 24 * 60 * 60 * 1000;
+
+// Grace-period deprecated models with deprecation metadata stripped.
+function activeDeprecatedModels(): JsonModel[] {
+  const now = Date.now();
+  const result: JsonModel[] = [];
+  for (const entry of Object.values(deprecatedData as Record<string, JsonModel & { deprecatedAt?: string }>)) {
+    if (!entry?.id) continue;
+    const removedAt = Date.parse(entry.deprecatedAt ?? "");
+    if (Number.isNaN(removedAt) || now - removedAt > DEPRECATED_MODEL_TTL_MS) continue;
+    const model = { ...entry } as JsonModel & { deprecatedAt?: string };
+    delete model.deprecatedAt;
+    result.push(model);
+  }
+  return result;
+}
+
+// Append grace-period deprecated models the list does not already have (live data wins).
+function withDeprecated(models: JsonModel[]): JsonModel[] {
+  const seen = new Set(models.map((m) => m.id));
+  const extras = activeDeprecatedModels().filter((m) => !seen.has(m.id));
+  return extras.length > 0 ? [...models, ...extras] : models;
+}
+
 function buildModels(
   base: JsonModel[],
   custom: JsonModel[],
@@ -424,7 +454,7 @@ function buildModels(
     }
   }
 
-  return Array.from(modelMap.values());
+  return withDeprecated(Array.from(modelMap.values()));
 }
 
 // ─── Stale-While-Revalidate Model Sync ────────────────────────────────────────
@@ -450,7 +480,7 @@ 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
 

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

@@ -1,12 +1,13 @@
 {
   "name": "pi-lilac-provider",
-  "version": "1.9.0",
+  "version": "1.9.2",
   "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",
   "files": [
     "index.ts",
     "models.json",
+    "deprecated-models.json",
     "patch.json",
     "custom-models.json",
     "scripts"
@@ -29,7 +30,7 @@
     "@earendil-works/pi-coding-agent": ">=0.74.0"
   },
   "devDependencies": {
-    "@earendil-works/pi-coding-agent": "0.80.3",
+    "@earendil-works/pi-coding-agent": "0.83.0",
     "typescript": "6.0.3",
     "@types/node": "25.9.1",
     "knip": "6.14.1"

package/scripts/update-models.js

@@ -295,6 +295,67 @@ function updateReadme(models) {
 
 // ─── Main ────────────────────────────────────────────────────────────────────
 
+// Grace period for delisted models: update-models.js moves models the API no
+// longer lists into deprecated-models.json (stamped with deprecatedAt) instead
+// of dropping them; the runtime appends them back so sessions and saved model
+// settings keep working, and after 14 days they are evicted permanently.
+const DEPRECATED_MODEL_TTL_MS = 14 * 24 * 60 * 60 * 1000;
+
+/**
+ * Reconcile deprecated-models.json against the freshly fetched model list.
+ * - in old models.json but not the API: moved into the deprecated file
+ *   (deprecatedAt = now; preserved on repeat runs so the grace clock is not reset)
+ * - back in the API: resurrected (dropped from the deprecated file)
+ * - deprecatedAt older than 14 days: evicted permanently
+ * Must run BEFORE the new models.json is written; it reads the old file itself.
+ */
+function updateDeprecatedModels(modelsJsonPath, newModels) {
+  const deprecatedPath = path.join(path.dirname(modelsJsonPath), 'deprecated-models.json');
+
+  let oldModels = [];
+  try {
+    const parsed = JSON.parse(fs.readFileSync(modelsJsonPath, 'utf8'));
+    if (Array.isArray(parsed)) oldModels = parsed;
+  } catch { /* first run: no previous models.json */ }
+
+  let deprecated = {};
+  try {
+    const parsed = JSON.parse(fs.readFileSync(deprecatedPath, 'utf8'));
+    if (parsed && typeof parsed === 'object' && !Array.isArray(parsed)) deprecated = parsed;
+  } catch { /* no graveyard yet */ }
+
+  const currentIds = new Set(newModels.map(m => m.id));
+  const now = new Date().toISOString();
+  const added = [];
+  const resurrected = [];
+  const evicted = [];
+
+  for (const old of oldModels) {
+    if (old && old.id && !currentIds.has(old.id) && !deprecated[old.id]) {
+      deprecated[old.id] = { ...old, deprecatedAt: now };
+      added.push(old.id);
+    }
+  }
+
+  for (const [id, entry] of Object.entries(deprecated)) {
+    if (currentIds.has(id)) {
+      delete deprecated[id];
+      resurrected.push(id);
+      continue;
+    }
+    const removedAt = Date.parse(entry && entry.deprecatedAt ? entry.deprecatedAt : '');
+    if (Number.isNaN(removedAt) || Date.now() - removedAt > DEPRECATED_MODEL_TTL_MS) {
+      delete deprecated[id];
+      evicted.push(id);
+    }
+  }
+
+  if (added.length > 0 || resurrected.length > 0 || evicted.length > 0) {
+    fs.writeFileSync(deprecatedPath, JSON.stringify(deprecated, null, 2) + '\n');
+    console.log('Updated deprecated-models.json ' + JSON.stringify({ added, resurrected, evicted }));
+  }
+}
+
 async function main() {
   try {
     const apiModels = await fetchModels();
@@ -318,6 +379,8 @@ async function main() {
     models.sort((a, b) => a.name.localeCompare(b.name));
 
     // Save models.json (pure API output, no patch/custom baked in)
+    // Move delisted models to deprecated-models.json BEFORE models.json is overwritten
+    updateDeprecatedModels(MODELS_JSON_PATH, models);
     saveJson(MODELS_JSON_PATH, models);
 
     // Build full model list for README: base → patch → custom