pi-hypercharm-provider 1.0.3

package/.github/FUNDING.yml

@@ -0,0 +1,4 @@
+github: monotykamary
+ko_fi: monotykamary
+buy_me_a_coffee: monotykamary
+polar: monotykamary

package/.pi/messenger/session-id

@@ -0,0 +1 @@
+019eb090-74e4-7e86-998f-031e8113fb88

package/AGENTS.md

@@ -0,0 +1,56 @@
+# AGENTS.md
+
+## DO NOT EDIT — Auto-generated Files
+
+The following files are **idempotent** and regenerated by `scripts/update-models.js`. Never edit them directly — your changes will be overwritten on the next model sync.
+
+| File | Why it's auto-generated |
+|------|------------------------|
+| `models.json` | Built from the provider API. `update-models.js` fetches models, preserves curated data for known IDs, and writes this file. |
+| `README.md` (model table) | The table under `## Available Models` is replaced in-place by `update-models.js` after merging base models → patch → custom models. |
+
+## Correct Files to Edit
+
+When a model needs overrides, new properties, or corrections, edit the appropriate source file below. These are the **source of truth** that the update script reads but never writes.
+
+| File | Purpose |
+|------|---------|
+| `patch.json` | Per-model overrides keyed by model ID. Add reasoning flags, compat settings, pricing corrections, thinking level maps, etc. Applied on top of `models.json` at runtime and for README generation. |
+| `custom-models.json` | Models that don't exist in the provider API (hidden models, router endpoints, cross-provider aliases). Merged after patch. Format: array of full model objects (same schema as `models.json` entries). |
+| `index.ts` | Provider extension code. |
+| `scripts/update-models.js` | The sync script itself (edit only if changing how models are fetched/transformed). |
+
+## Data Flow
+
+```
+Provider API  ──fetch──►  models.json  ──apply──►  patch.json  ──merge──►  custom-models.json
+                                    │                            │                      │
+                                    └────────────────────────────┴──────────────────────┘
+                                                      │
+                                              README model table
+```
+
+1. `models.json` — base data from the provider API (auto-generated, DO NOT EDIT)
+2. `patch.json` — overrides applied on top (EDIT THIS for corrections/enrichments)
+3. `custom-models.json` — additional models not in the API (EDIT THIS for new models)
+4. README table — rendered from the merged result of all three (auto-generated, DO NOT EDIT)
+
+## Common Tasks
+
+### Add a compat setting or override pricing for an existing model
+→ Edit `patch.json`. Add an entry keyed by the model's `id`.
+
+### Add a model not available in the provider API
+→ Edit `custom-models.json`. Add a full model object to the array.
+
+### Update models from the provider API
+→ Run `node scripts/update-models.js` (may require an API key env var).
+
+### Regenerate the README model table
+→ Run `node scripts/update-models.js` — it updates both `models.json` and the README table.
+
+## TL;DR
+
+- **Never edit `models.json`** — edit `patch.json` instead.
+- **Never edit the README model table** — run the update script instead.
+- `patch.json` and `custom-models.json` are the source files you should modify.

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,208 @@
+<div align="center">
+
+# ✨ pi-hypercharm-provider
+
+**17+ models through [Charm Hyper](https://hyper.charm.land/)**
+
+_Hyperoptimized coding models — DeepSeek, GLM, Kimi, Qwen, MiniMax, Gemma, GPT-OSS, and Llama for [pi](https://github.com/earendil-works/pi-coding-agent)._
+
+[![pi extension](https://img.shields.io/badge/pi-extension-blueviolet)](https://github.com/earendil-works/pi-coding-agent)
+[![license](https://img.shields.io/badge/license-MIT-blue)](./LICENSE)
+
+</div>
+
+---
+
+## Features
+
+- **20+ AI Models** including DeepSeek V4 Flash/Pro, GLM 5/5.1, Kimi K2.5/K2.6, Qwen3.6/3.7, MiniMax M2.7, Gemma 4, GPT-OSS, and Llama
+- **DeepSeek Native Thinking** — Uses the `deepseek` thinking format with `reasoning_effort` and 384K max output for DeepSeek V4 models
+- **OpenAI-compatible API** via Charm Hyper's `/v1/chat/completions` endpoint
+- **Cost Tracking** with per-model pricing from the API
+- **Reasoning Models** with `reasoning_effort` parameter support on select models
+- **Vision Support** for image-capable models (Kimi K2.5/K2.6, GLM 5.1, Qwen3.6/3.7, Gemma 4)
+
+## Available Models
+
+| Model | Type | Context | Max Tokens | Input Cost | Output Cost |
+|-------|------|---------|------------|------------|-------------|
+| DeepSeek V4 Flash | Text | 1.0M | 384K | $0.14 | $0.28 |
+| DeepSeek V4 Pro | Text | 1.0M | 384K | $0.43 | $0.87 |
+| Gemma 4 26B A4B | Text + Image | 256K | 26K | $0.13 | $0.42 |
+| GLM 5.1 | Text + Image | 202K | 202K | $1.30 | $4.30 |
+| GLM-5 | Text | 203K | 20K | $0.81 | $2.62 |
+| GLM-5.2 | Text | 1.0M | 131K | $1.40 | $4.40 |
+| GPT-OSS 120B | Text | 131K | 13K | $0.17 | $0.56 |
+| Kimi K2.5 | Text + Image | 262K | 26K | $0.49 | $2.63 |
+| Kimi K2.6 | Text + Image | 262K | 262K | $0.95 | $4.00 |
+| Kimi K2.7 Code | Text | 256K | 33K | $0.95 | $4.00 |
+| Llama 3.3 70B Instruct | Text | 128K | 13K | $0.55 | $1.07 |
+| Llama 4 Maverick 17B 128E FP8 | Text | 430K | 43K | $0.27 | $0.89 |
+| MiniMax M2.7 | Text | 205K | 131K | $0.30 | $1.20 |
+| Qwen3 Coder 480B A35B INT4 | Text | 106K | 11K | $0.56 | $2.54 |
+| Qwen3 Next 80B A3B | Text | 262K | 26K | $0.13 | $1.28 |
+| Qwen3.6-Flash | Text + Image | 1.0M | 64K | $1.00 | $4.00 |
+| Qwen3.6-Max | Text + Image | 256K | 64K | $2.00 | $12.00 |
+| Qwen3.6-Plus | Text + Image | 1.0M | 64K | $2.00 | $6.00 |
+| Qwen3.7-Max | Text + Image | 1.0M | 64K | $2.50 | $7.50 |
+| Qwen3.7-Plus | Text | 1.0M | 64K | $1.20 | $4.80 |
+*Costs are per million tokens. Prices subject to change — check [hyper.charm.land](https://hyper.charm.land) for current pricing.*
+
+## Installation
+
+### Option 1: Using `pi install` (Recommended)
+
+Install directly from GitHub:
+
+```bash
+pi install https://github.com/monotykamary/pi-hypercharm-provider
+```
+
+Then set your API key and run pi:
+```bash
+# Recommended: add to auth.json
+# See Authentication section below
+
+# Or set as environment variable
+export HYPERCHARM_API_KEY=your-api-key-here
+
+pi
+```
+
+Get your API key from [hyper.charm.land](https://hyper.charm.land).
+
+### Option 2: Manual Clone
+
+1. Clone this repository:
+   ```bash
+   git clone https://github.com/monotykamary/pi-hypercharm-provider.git
+   cd pi-hypercharm-provider
+   ```
+
+2. Set your HyperCharm API key:
+   ```bash
+   # Recommended: add to auth.json
+   # See Authentication section below
+
+   # Or set as environment variable
+   export HYPERCHARM_API_KEY=your-api-key-here
+   ```
+
+3. Run pi with the extension:
+   ```bash
+   pi -e /path/to/pi-hypercharm-provider
+   ```
+
+## Usage
+
+After loading the extension, select a model with:
+
+```
+/model hypercharm deepseek-v4-pro
+```
+
+Or start pi directly with a HyperCharm model:
+
+```bash
+pi --provider hypercharm --model deepseek-v4-pro
+```
+
+### Reasoning Effort
+
+For reasoning models that support `reasoning_effort`, control thinking depth:
+
+```bash
+pi --provider hypercharm --model deepseek-v4-pro --reasoning-effort max
+```
+
+Values: `none`, `low`, `medium`, `high`, `max`
+
+### Thinking Mode
+
+DeepSeek V4 models use the `deepseek` thinking format — the same native format as the [pi-deepseek-provider](https://github.com/monotykamary/pi-deepseek-provider). This sends `thinking: {type: "enabled/disabled"}` plus `reasoning_effort` mapped via `thinkingLevelMap` (`high` → `"high"`, `max` → `"max"`). Replayed assistant messages include empty `reasoning_content` as required by DeepSeek's API.
+
+## Authentication
+
+The HyperCharm API key can be configured in multiple ways (resolved in this order):
+
+1. **`auth.json`** (recommended) — Add to `~/.pi/agent/auth.json`:
+   ```json
+   { "hypercharm": { "type": "api_key", "key": "your-api-key" } }
+   ```
+   The `key` field supports literal values, env var names, and shell commands (prefix with `!`). See [pi's auth file docs](https://github.com/badlogic/pi-mono) for details.
+2. **Runtime override** — Use the `--api-key` CLI flag
+3. **Environment variable** — Set `HYPERCHARM_API_KEY`
+
+Get your API key from [hyper.charm.land](https://hyper.charm.land).
+
+## Environment Variables
+
+| Variable | Required | Description |
+|----------|----------|-------------|
+| `HYPERCHARM_API_KEY` | No | Your Charm Hyper API key (fallback if not in auth.json) |
+
+## Configuration
+
+Add to your pi configuration for automatic loading:
+
+```json
+{
+  "extensions": [
+    "/path/to/pi-hypercharm-provider"
+  ]
+}
+```
+
+### Compat Settings
+
+Charm Hyper uses an OpenAI-compatible API. Compatibility settings are managed via `patch.json`:
+
+**DeepSeek models** use the `deepseek` thinking format (identical to the native DeepSeek API):
+
+- **`thinkingFormat: "deepseek"`** — Sends `thinking: {type: "enabled/disabled"}` to toggle thinking mode
+- **`supportsReasoningEffort: true`** — Supports `reasoning_effort: "high" | "max"` via `thinkingLevelMap`
+- **`requiresReasoningContentOnAssistantMessages: true`** — Replayed assistant messages include empty `reasoning_content` when reasoning is enabled
+- **`maxTokensField: "max_tokens"`** — Uses `max_tokens` (not `max_completion_tokens`)
+
+**Other reasoning models** (GLM, Kimi, Qwen, MiniMax, Gemma, GPT-OSS) use the `openai` thinking format:
+
+- **`thinkingFormat: "openai"`** — Returns `reasoning_content` in OpenAI format
+- **`supportsReasoningEffort: true`** — Accepts `reasoning_effort` parameter on supported models
+- **`maxTokensField: "max_tokens"`** — Uses `max_tokens` (not `max_completion_tokens`)
+- **`supportsDeveloperRole: true`** — Developer role messages are accepted
+- **`supportsStore: false`** — The `store` parameter is not supported
+
+> **Note:** The HyperCharm `/v1/models` endpoint reports `supports_reasoning: true` for all models,
+> including those that don't natively support extended thinking (e.g., Llama 3.3 70B). `models.json`
+> curates reasoning flags based on known model capabilities; `patch.json` adds compat overrides.
+
+### Patch Overrides
+
+The `patch.json` file contains overrides applied on top of `models.json` data:
+
+- Correcting API-derived values (e.g., DeepSeek `maxTokens` from 8000 to 384000)
+- Adding compat settings that the API doesn't provide
+- Setting `thinkingFormat: "deepseek"` + `thinkingLevelMap` for DeepSeek V4 models
+- Overriding pricing when official rates change
+
+## Updating Models
+
+Run the update script to fetch the latest models from the HyperCharm API:
+
+```bash
+export HYPERCHARM_API_KEY=your-api-key
+node scripts/update-models.js
+```
+
+This will:
+1. Fetch models from `https://hyper.charm.land/v1/models`
+2. Preserve curated data (reasoning, vision, DeepSeek compat) from existing `models.json`
+3. Apply overrides from `patch.json`
+4. Remove custom models now available upstream from `custom-models.json`
+5. Update `models.json` and the README model table
+
+A GitHub Actions workflow runs this daily and creates a PR if models have changed.
+
+## License
+
+MIT

package/custom-models.json

@@ -0,0 +1 @@
+[]

package/index.ts

@@ -0,0 +1,371 @@
+/**
+ * HyperCharm Provider Extension
+ *
+ * Registers HyperCharm (hyper.charm.land) as a custom provider using the
+ * openai-completions API. Base URL: https://hyper.charm.land/v1
+ *
+ * HyperCharm provides hyperoptimized coding models via an OpenAI-compatible API.
+ * The /v1/models endpoint returns structured metadata including reasoning flags,
+ * pricing, context windows, and max output tokens.
+ *
+ * Note: The API's `supports_reasoning` flag is unreliable for some models (e.g.,
+ * it reports true for Llama 3.3 70B which doesn't support extended thinking).
+ * The models.json embeds curated reasoning flags; patch.json corrects compat.
+ *
+ * Model resolution strategy: Stale-While-Revalidate
+ *   1. Serve stale immediately: disk cache → embedded models.json (zero-latency)
+ *   2. Revalidate in background: live API /v1/models → merge with embedded → cache → hot-swap
+ *   3. patch.json + custom-models.json applied on top of whichever source won
+ *
+ * Merge order: [live|cache|embedded] → apply patch.json → merge custom-models.json
+ *
+ * Usage:
+ *   # Option 1: Store in auth.json (recommended)
+ *   # Add to ~/.pi/agent/auth.json:
+ *   #   "hypercharm": { "type": "api_key", "key": "your-api-key" }
+ *
+ *   # Option 2: Set as environment variable
+ *   export HYPERCHARM_API_KEY=your-api-key
+ *
+ *   # Run pi with the extension
+ *   pi -e /path/to/pi-hypercharm-provider
+ *
+ * Then use /model to select from available models.
+ *
+ * @see https://hyper.charm.land
+ */
+
+import { getAgentDir, type ExtensionAPI, type ModelRegistry } from "@earendil-works/pi-coding-agent";
+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 fs from "fs";
+import path from "path";
+
+// ─── Types ────────────────────────────────────────────────────────────────────
+
+interface JsonModel {
+  id: string;
+  name: string;
+  reasoning: boolean;
+  input: string[];
+  cost: {
+    input: number;
+    output: number;
+    cacheRead: number;
+    cacheWrite: number;
+  };
+  contextWindow: number;
+  maxTokens: number;
+  thinkingLevelMap?: Record<string, string | null>;
+  compat?: {
+    supportsDeveloperRole?: boolean;
+    supportsStore?: boolean;
+    maxTokensField?: "max_completion_tokens" | "max_tokens";
+    thinkingFormat?: "openai" | "zai" | "qwen" | "qwen-chat-template" | "deepseek";
+    supportsReasoningEffort?: boolean;
+    requiresReasoningContentOnAssistantMessages?: boolean;
+  };
+}
+
+interface PatchEntry {
+  name?: string;
+  reasoning?: boolean;
+  input?: string[];
+  cost?: {
+    input?: number;
+    output?: number;
+    cacheRead?: number;
+    cacheWrite?: number;
+  };
+  contextWindow?: number;
+  maxTokens?: number;
+  thinkingLevelMap?: Record<string, string | null>;
+  compat?: Record<string, unknown>;
+}
+
+type PatchData = Record<string, PatchEntry>;
+
+// ─── Patch Application ────────────────────────────────────────────────────────
+
+function applyPatch(model: JsonModel, patch: PatchEntry): JsonModel {
+  const result = { ...model };
+
+  if (patch.name !== undefined) result.name = patch.name;
+  if (patch.reasoning !== undefined) result.reasoning = patch.reasoning;
+  if (patch.input !== undefined) result.input = patch.input;
+  if (patch.contextWindow !== undefined) result.contextWindow = patch.contextWindow;
+  if (patch.maxTokens !== undefined) result.maxTokens = patch.maxTokens;
+  if (patch.thinkingLevelMap !== undefined) result.thinkingLevelMap = { ...patch.thinkingLevelMap };
+
+  if (patch.cost) {
+    result.cost = {
+      input: patch.cost.input ?? result.cost.input,
+      output: patch.cost.output ?? result.cost.output,
+      cacheRead: patch.cost.cacheRead ?? result.cost.cacheRead,
+      cacheWrite: patch.cost.cacheWrite ?? result.cost.cacheWrite,
+    };
+  }
+  if (patch.compat) {
+    result.compat = { ...(result.compat || {}), ...patch.compat };
+  }
+
+  if (!result.reasoning && result.compat?.thinkingFormat) {
+    delete result.compat.thinkingFormat;
+  }
+  if (result.compat && Object.keys(result.compat).length === 0) {
+    delete result.compat;
+  }
+
+  return result;
+}
+
+/** Full pipeline: base models → patch → custom → result */
+function buildModels(base: JsonModel[], custom: JsonModel[], patch: PatchData): JsonModel[] {
+  const modelMap = new Map<string, JsonModel>();
+
+  for (const model of base) {
+    modelMap.set(model.id, model);
+  }
+
+  for (const [id, patchEntry] of Object.entries(patch)) {
+    const existing = modelMap.get(id);
+    if (existing) {
+      modelMap.set(id, applyPatch(existing, patchEntry));
+    }
+  }
+
+  for (const model of custom) {
+    const existing = modelMap.get(model.id);
+    const patchEntry = patch[model.id];
+    if (existing && patchEntry) {
+      modelMap.set(model.id, applyPatch(model, patchEntry));
+    } else if (existing) {
+      modelMap.set(model.id, model);
+    } else if (patchEntry) {
+      modelMap.set(model.id, applyPatch(model, patchEntry));
+    } else {
+      modelMap.set(model.id, model);
+    }
+  }
+
+  const result = Array.from(modelMap.values());
+
+  // Ensure DeepSeek reasoning models have required compat settings.
+  // Live-fetched models from the SWR pipeline may not have these set.
+  for (const model of result) {
+    if (!model.reasoning) continue;
+    if (isDeepSeekModel(model.id)) {
+      if (!model.compat) {
+        model.compat = {
+          thinkingFormat: "deepseek",
+          maxTokensField: "max_tokens",
+          supportsDeveloperRole: true,
+          supportsStore: false,
+          supportsReasoningEffort: true,
+          requiresReasoningContentOnAssistantMessages: true,
+        };
+      } else {
+        if (model.compat.thinkingFormat === undefined) {
+          model.compat.thinkingFormat = "deepseek";
+        }
+        if (model.compat.supportsReasoningEffort === undefined) {
+          model.compat.supportsReasoningEffort = true;
+        }
+        if ((model.compat as any).requiresReasoningContentOnAssistantMessages === undefined) {
+          (model.compat as any).requiresReasoningContentOnAssistantMessages = true;
+        }
+      }
+      if (!model.thinkingLevelMap) {
+        model.thinkingLevelMap = {
+          minimal: null, low: null, medium: null, high: "high", xhigh: "max",
+        };
+      }
+    }
+  }
+
+  return result;
+}
+
+function isDeepSeekModel(id: string): boolean {
+  return /^deepseek-v/.test(id);
+}
+
+// ─── Stale-While-Revalidate Model Sync ────────────────────────────────────────
+
+const PROVIDER_ID = "hypercharm";
+const BASE_URL = "https://hyper.charm.land/v1";
+const MODELS_URL = `${BASE_URL}/models`;
+const CACHE_DIR = path.join(getAgentDir(), "cache");
+const CACHE_PATH = path.join(CACHE_DIR, `${PROVIDER_ID}-models.json`);
+const LIVE_FETCH_TIMEOUT_MS = 8000;
+
+/** Transform a model from the HyperCharm /v1/models API to JsonModel format. */
+function transformApiModel(apiModel: any): JsonModel | null {
+  if (!apiModel.id) return null;
+
+  const cost = apiModel.cost?.usd || {};
+  const toPerM = (v: any) => {
+    const n = typeof v === "string" ? parseFloat(v) : (v || 0);
+    // API returns $/M directly; round to 6 decimals to preserve sub-cent cache prices.
+    return Math.round(n * 1e6) / 1e6;
+  };
+
+  return {
+    id: apiModel.id,
+    name: apiModel.display_name || apiModel.id,
+    reasoning: false, // API supports_reasoning is unreliable; patch.json corrects
+    input: ["text"],
+    cost: {
+      input: toPerM(cost["1m_in"]),
+      output: toPerM(cost["1m_out"]),
+      cacheRead: toPerM(cost["1m_in_cache"]),
+      cacheWrite: toPerM(cost["1m_out_cache"]),
+    },
+    contextWindow: apiModel.context_window || 0,
+    maxTokens: apiModel.max_output_tokens || 0,
+  };
+}
+
+async function fetchLiveModels(apiKey: string, signal?: AbortSignal): Promise<JsonModel[] | null> {
+  try {
+    const response = await fetch(MODELS_URL, {
+      headers: { Authorization: `Bearer ${apiKey}` },
+      signal: signal ? AbortSignal.any([AbortSignal.timeout(LIVE_FETCH_TIMEOUT_MS), signal]) : AbortSignal.timeout(LIVE_FETCH_TIMEOUT_MS),
+    });
+    if (!response.ok) return null;
+    const data = await response.json();
+    const apiModels = Array.isArray(data) ? data : (data.data || []);
+    if (!Array.isArray(apiModels) || apiModels.length === 0) return null;
+    return apiModels.map(transformApiModel).filter((m): m is JsonModel => m !== null);
+  } catch {
+    return null;
+  }
+}
+
+function loadCachedModels(): JsonModel[] | null {
+  try {
+    const data = JSON.parse(fs.readFileSync(CACHE_PATH, "utf8"));
+    return Array.isArray(data) ? data : null;
+  } catch {
+    return null;
+  }
+}
+
+function cacheModels(models: JsonModel[]): void {
+  try {
+    fs.mkdirSync(CACHE_DIR, { recursive: true });
+    fs.writeFileSync(CACHE_PATH, JSON.stringify(models, null, 2) + "\n");
+  } catch {
+    // Cache write failure is non-fatal
+  }
+}
+
+function mergeWithEmbedded(liveModels: JsonModel[], embeddedModels: JsonModel[]): JsonModel[] {
+  const embeddedMap = new Map(embeddedModels.map(m => [m.id, m]));
+  const seen = new Set<string>();
+  const result: JsonModel[] = [];
+  for (const liveModel of liveModels) {
+    const embedded = embeddedMap.get(liveModel.id);
+    seen.add(liveModel.id);
+    if (embedded) {
+      // Self-heal: live API pricing is authoritative field-by-field. Prefer the
+      // live cost when the API reports it (non-zero); fall back to embedded when
+      // the API is silent (0) so curated cacheRead/cacheWrite isn't clobbered and
+      // providers whose /models endpoint exposes no pricing keep their curated
+      // cost. Curation (reasoning/input/compat/name) still wins via ...embedded.
+      result.push({
+        ...liveModel,
+        ...embedded,
+        cost: {
+          input: liveModel.cost.input || embedded.cost.input,
+          output: liveModel.cost.output || embedded.cost.output,
+          cacheRead: liveModel.cost.cacheRead || embedded.cost.cacheRead,
+          cacheWrite: liveModel.cost.cacheWrite || embedded.cost.cacheWrite,
+        },
+        contextWindow: liveModel.contextWindow || embedded.contextWindow,
+      });
+    } else {
+      result.push(liveModel);
+    }
+  }
+  // Append any embedded models that the live API didn't return
+  for (const em of embeddedModels) {
+    if (!seen.has(em.id)) {
+      result.push(em);
+    }
+  }
+  return result;
+}
+
+function loadStaleModels(embeddedModels: JsonModel[]): JsonModel[] {
+  const cached = loadCachedModels();
+  if (!cached || cached.length === 0) return embeddedModels;
+
+  // Merge embedded models that are missing from cache (newly added models)
+  const cachedMap = new Map(cached.map(m => [m.id, m]));
+  for (const em of embeddedModels) {
+    if (!cachedMap.has(em.id)) {
+      cached.push(em);
+    }
+  }
+  return cached;
+}
+
+async function revalidateModels(apiKey: string | undefined, embeddedModels: JsonModel[], signal?: AbortSignal): Promise<JsonModel[] | null> {
+  if (!apiKey) return null;
+  const liveModels = await fetchLiveModels(apiKey, signal);
+  if (!liveModels || liveModels.length === 0) return null;
+  const merged = mergeWithEmbedded(liveModels, embeddedModels);
+  cacheModels(merged);
+  return merged;
+}
+
+// ─── API Key Resolution (via ModelRegistry) ────────────────────────────────────
+
+let cachedApiKey: string | undefined;
+let revalidateAbort: AbortController | null = null;
+
+async function resolveApiKey(modelRegistry: ModelRegistry): Promise<void> {
+  cachedApiKey = await modelRegistry.getApiKeyForProvider("hypercharm") ?? undefined;
+}
+
+// ─── Extension Entry Point ────────────────────────────────────────────────────
+
+export default function (pi: ExtensionAPI) {
+  const embeddedModels = modelsData as JsonModel[];
+  const customModels = customModelsData as JsonModel[];
+  const patches = patchData as PatchData;
+
+  const staleBase = loadStaleModels(embeddedModels);
+  const staleModels = buildModels(staleBase, customModels, patches);
+
+  pi.registerProvider("hypercharm", {
+    baseUrl: BASE_URL,
+    apiKey: "$HYPERCHARM_API_KEY",
+    api: "openai-completions",
+    models: staleModels,
+  });
+
+  pi.on("session_start", async (_event, ctx) => {
+    revalidateAbort?.abort();
+    revalidateAbort = new AbortController();
+    const signal = revalidateAbort.signal;
+    resolveApiKey(ctx.modelRegistry).then(() => {
+      revalidateModels(cachedApiKey, embeddedModels, signal).then((freshBase) => {
+        if (freshBase && !signal.aborted) {
+          pi.registerProvider("hypercharm", {
+            baseUrl: BASE_URL,
+            apiKey: "$HYPERCHARM_API_KEY",
+            api: "openai-completions",
+            models: buildModels(freshBase, customModels, patches),
+          });
+        }
+      });
+    });
+  });
+
+  pi.on("session_shutdown", () => {
+    revalidateAbort?.abort();
+  });
+}