codeep 2.15.0 → 2.17.0

package/dist/utils/resourceImpact.js

@@ -0,0 +1,54 @@
+/**
+ * Broad operational-impact estimate for hosted LLM inference.
+ *
+ * This is deliberately a range, not a meter reading:
+ * - 0.3–1.5 J/token covers published H100 inference benchmarks.
+ * - 0.27–1.08 L/kWh spans efficient direct cooling and a full-stack
+ *   production estimate that also captures associated infrastructure.
+ *
+ * Model size, batching, context length, hardware, data-centre location and
+ * provider efficiency can move the real result outside this band. Local
+ * models are included because their token usage still consumes electricity,
+ * but Codeep cannot measure the device directly.
+ */
+const ENERGY_JOULES_PER_TOKEN = { low: 0.3, high: 1.5 };
+const WATER_LITRES_PER_KWH = { low: 0.27, high: 1.08 };
+export function estimateResourceImpact(totalTokens) {
+    const tokens = Number.isFinite(totalTokens) ? Math.max(0, totalTokens) : 0;
+    const energyWhLow = (tokens * ENERGY_JOULES_PER_TOKEN.low) / 3600;
+    const energyWhHigh = (tokens * ENERGY_JOULES_PER_TOKEN.high) / 3600;
+    return {
+        energyWhLow,
+        energyWhHigh,
+        waterMlLow: (energyWhLow / 1000) * WATER_LITRES_PER_KWH.low * 1000,
+        waterMlHigh: (energyWhHigh / 1000) * WATER_LITRES_PER_KWH.high * 1000,
+    };
+}
+function compact(value) {
+    if (value === 0)
+        return '0';
+    if (value < 0.01)
+        return value.toFixed(3);
+    if (value < 1)
+        return value.toFixed(2);
+    if (value < 10)
+        return value.toFixed(1);
+    return Math.round(value).toLocaleString('en-US');
+}
+export function formatResourceImpact(estimate) {
+    const energy = estimate.energyWhHigh >= 1000
+        ? `${compact(estimate.energyWhLow / 1000)}–${compact(estimate.energyWhHigh / 1000)} kWh`
+        : `${compact(estimate.energyWhLow)}–${compact(estimate.energyWhHigh)} Wh`;
+    const water = estimate.waterMlHigh >= 1000
+        ? `${compact(estimate.waterMlLow / 1000)}–${compact(estimate.waterMlHigh / 1000)} L`
+        : `${compact(estimate.waterMlLow)}–${compact(estimate.waterMlHigh)} mL`;
+    return { energy, water };
+}
+export function formatResourceImpactReport(totalTokens) {
+    const formatted = formatResourceImpact(estimateResourceImpact(totalTokens));
+    return [
+        '### Estimated compute impact',
+        `**Electricity:** ${formatted.energy}  ·  **Cooling water:** ${formatted.water}`,
+        '_Research-based range, not a provider measurement. Actual usage varies widely by model, hardware, batching, context length, data-centre location, and cooling system._',
+    ];
+}

package/dist/utils/tokenTracker.js

@@ -2,13 +2,14 @@
  * Token and cost tracking for API usage
  */
 import { AsyncLocalStorage } from 'node:async_hooks';
-// Context window sizes per model (in tokens).
-// Keep this table in lockstep with `providers.ts` — entries for models that
-// aren't in the provider catalogue only show up if a user types an id by hand
-// and produce phantom estimates against the wrong context size.
+import { formatResourceImpactReport } from './resourceImpact.js';
+// Context window sizes per model (in tokens). Primarily mirrors providers.ts;
+// retired aliases remain only where restored historical sessions still need a
+// meaningful context/cost display.
 const MODEL_CONTEXT_WINDOWS = {
     // Z.AI / ZhipuAI
-    'glm-5.2': 200_000,
+    'glm-5.2': 1_000_000,
+    'glm-5.1': 200_000,
     'glm-5-turbo': 202_752,
     // OpenAI
     'gpt-5.6-sol': 1_050_000,
@@ -19,7 +20,7 @@ const MODEL_CONTEXT_WINDOWS = {
     'gpt-5.4-mini': 400_000,
     // Anthropic
     'claude-fable-5': 1_000_000,
-    'claude-opus-4-8': 1_000_000,
+    'claude-opus-5': 1_000_000,
     'claude-sonnet-4-6': 1_000_000,
     'claude-sonnet-5': 1_000_000,
     'claude-haiku-4-5-20251001': 200_000,
@@ -28,28 +29,34 @@ const MODEL_CONTEXT_WINDOWS = {
     'deepseek-v4-flash': 1_000_000,
     // Google
     'gemini-3.1-pro-preview': 1_048_576,
-    'gemini-3.5-flash': 1_000_000,
-    'gemini-3.1-flash-lite': 1_048_576,
+    'gemini-3.6-flash': 1_048_576,
+    'gemini-3.5-flash': 1_048_576,
+    'gemini-3.5-flash-lite': 1_048_576,
     'gemini-3-flash-preview': 1_000_000,
     // MiniMax
-    'MiniMax-M3': 524_288,
-    // Kimi (Moonshot) — 256K across the K2.x line
+    'MiniMax-M3': 1_000_000,
+    // Kimi (Moonshot) — 1M on K3, 256K across K2.x
+    'kimi-k3': 1_000_000,
     'kimi-k2.7-code': 262_144,
     'kimi-k2.7-code-highspeed': 262_144,
     'kimi-k2.6': 262_144,
-    'kimi-k2.5': 262_144,
     'kimi-for-coding': 262_144,
+    'kimi-for-coding-highspeed': 262_144,
+    'k3': 1_000_000,
+    'k3-256k': 262_144,
     // Grok (xAI)
     'grok-4.5': 500_000,
     'grok-build-0.1': 256_000,
     'grok-4.3': 1_000_000,
     'grok-code-fast-1': 256_000,
     'grok-4-fast-reasoning': 2_000_000,
-    // Qwen (Alibaba) — 256K native (1M with extrapolation)
-    'qwen3-coder-plus': 262_144,
-    'qwen3-coder-next': 262_144,
-    'qwen3-coder-flash': 262_144,
+    // Qwen (Alibaba) — current hosted generation
     'qwen3.7-max': 1_000_000,
+    'qwen3.8-max-preview': 1_000_000,
+    'qwen3.7-plus': 1_000_000,
+    'qwen3.6-plus': 1_000_000,
+    'qwen3.5-plus': 1_000_000,
+    'qwen3.6-flash': 1_000_000,
     'Qwen/Qwen3-Coder-480B-A35B-Instruct': 262_144,
 };
 const DEFAULT_CONTEXT_WINDOW = 128_000;
@@ -60,58 +67,65 @@ export function getModelContextWindow(model) {
     return MODEL_CONTEXT_WINDOWS[model] ?? DEFAULT_CONTEXT_WINDOW;
 }
 // Pricing table — USD per 1M tokens. Same rule as MODEL_CONTEXT_WINDOWS:
-// only list model ids that exist in `providers.ts`, otherwise typing an id
-// by hand can produce phantom cost estimates against stale rates.
+// Primarily mirrors `providers.ts`. A few retired aliases remain so restored
+// historical sessions still show the rate that applied when they were created.
 const MODEL_PRICING = {
     // Z.AI / ZhipuAI
-    // GLM-5.2 per-token pricing isn't published yet — mirror GLM-5.1 (same tier,
-    // its successor) provisionally so /cost stays sane; update when z.ai posts it.
-    // Note: on the GLM Coding Plan (the default `z.ai` provider) billing is a flat
-    // subscription, so this only affects the pay-per-use estimate.
-    'glm-5.2': { inputPer1M: 1.00, outputPer1M: 3.20 },
+    // Coding Plan is flat-fee; these official rates apply to pay-per-use.
+    'glm-5.2': { inputPer1M: 1.40, outputPer1M: 4.40 },
+    'glm-5.1': { inputPer1M: 1.40, outputPer1M: 4.40 },
     'glm-5-turbo': { inputPer1M: 1.20, outputPer1M: 4.00 },
     // OpenAI
     'gpt-5.6-sol': { inputPer1M: 5.00, outputPer1M: 30.00 },
-    'gpt-5.6-terra': { inputPer1M: 2.50, outputPer1M: 15.00 },
-    'gpt-5.6-luna': { inputPer1M: 1.00, outputPer1M: 6.00 },
+    'gpt-5.6-terra': { inputPer1M: 2.00, outputPer1M: 12.00 },
+    'gpt-5.6-luna': { inputPer1M: 0.20, outputPer1M: 1.20 },
     'gpt-5.5': { inputPer1M: 5.00, outputPer1M: 30.00 },
     'gpt-5.4': { inputPer1M: 2.50, outputPer1M: 15.00 },
     'gpt-5.4-mini': { inputPer1M: 0.75, outputPer1M: 4.50 },
     // Anthropic
     'claude-fable-5': { inputPer1M: 10.00, outputPer1M: 50.00 },
-    'claude-opus-4-8': { inputPer1M: 5.00, outputPer1M: 25.00 },
+    'claude-opus-5': { inputPer1M: 5.00, outputPer1M: 25.00 },
     'claude-sonnet-4-6': { inputPer1M: 3.00, outputPer1M: 15.00 },
     'claude-sonnet-5': { inputPer1M: 3.00, outputPer1M: 15.00 },
     'claude-haiku-4-5-20251001': { inputPer1M: 1.00, outputPer1M: 5.00 },
     // DeepSeek (cache-miss input pricing)
-    'deepseek-v4-pro': { inputPer1M: 1.74, outputPer1M: 3.48 },
+    'deepseek-v4-pro': { inputPer1M: 0.435, outputPer1M: 0.87 },
     'deepseek-v4-flash': { inputPer1M: 0.14, outputPer1M: 0.28 },
     // Google
     'gemini-3.1-pro-preview': { inputPer1M: 2.00, outputPer1M: 12.00 },
+    'gemini-3.6-flash': { inputPer1M: 1.50, outputPer1M: 7.50 },
     'gemini-3.5-flash': { inputPer1M: 1.50, outputPer1M: 9.00 },
-    'gemini-3.1-flash-lite': { inputPer1M: 0.25, outputPer1M: 1.50 },
+    'gemini-3.5-flash-lite': { inputPer1M: 0.30, outputPer1M: 2.50 },
     'gemini-3-flash-preview': { inputPer1M: 0.50, outputPer1M: 3.00 },
     // MiniMax
     'MiniMax-M3': { inputPer1M: 0.60, outputPer1M: 2.40 },
     // Kimi (Moonshot) — pay-per-use cache-miss rates; `kimi-for-coding` is the
     // subscription alias (flat-fee in reality, priced notionally like K2.7 Code).
-    'kimi-k2.7-code': { inputPer1M: 0.60, outputPer1M: 2.50 },
-    'kimi-k2.7-code-highspeed': { inputPer1M: 0.60, outputPer1M: 2.50 },
-    'kimi-k2.6': { inputPer1M: 0.55, outputPer1M: 2.20 },
-    'kimi-k2.5': { inputPer1M: 0.40, outputPer1M: 1.90 },
-    'kimi-for-coding': { inputPer1M: 0.60, outputPer1M: 2.50 },
+    'kimi-k3': { inputPer1M: 3.00, outputPer1M: 15.00 },
+    'kimi-k2.7-code': { inputPer1M: 0.95, outputPer1M: 4.00 },
+    // Kimi doesn't publish a distinct high-speed price in its main table.
+    // Leave that variant unpriced rather than presenting an invented estimate.
+    'kimi-k2.6': { inputPer1M: 0.95, outputPer1M: 4.00 },
+    'kimi-for-coding': { inputPer1M: 0.95, outputPer1M: 4.00 },
+    'kimi-for-coding-highspeed': { inputPer1M: 0.95, outputPer1M: 4.00 },
+    'k3': { inputPer1M: 3.00, outputPer1M: 15.00 },
+    'k3-256k': { inputPer1M: 3.00, outputPer1M: 15.00 },
     // Grok (xAI)
     'grok-4.5': { inputPer1M: 2.00, outputPer1M: 6.00 },
     'grok-build-0.1': { inputPer1M: 1.00, outputPer1M: 2.00 },
     'grok-4.3': { inputPer1M: 1.25, outputPer1M: 2.50 },
     'grok-code-fast-1': { inputPer1M: 0.20, outputPer1M: 1.50 },
     'grok-4-fast-reasoning': { inputPer1M: 0.20, outputPer1M: 0.50 },
-    // Qwen (Alibaba) — qwen3-coder-* 0–256K tier; the Coding Plan is flat-fee so
-    // this only affects the pay-per-use estimate.
-    'qwen3-coder-plus': { inputPer1M: 0.28, outputPer1M: 1.65 },
-    'qwen3-coder-next': { inputPer1M: 0.28, outputPer1M: 1.65 },
-    'qwen3-coder-flash': { inputPer1M: 0.10, outputPer1M: 0.50 },
+    // Qwen (Alibaba) — list prices for the first context tier. Coding Plan
+    // variants are flat-fee; these rates describe pay-per-use calls.
+    // `qwen3.8-max-preview` is Token-Plan-only (credit-metered, promotional
+    // preview rate); Alibaba publishes no pay-per-use per-token price for it.
+    // Leave it unpriced rather than borrowing the GA qwen3.8-max rate.
     'qwen3.7-max': { inputPer1M: 2.50, outputPer1M: 7.50 },
+    'qwen3.7-plus': { inputPer1M: 0.40, outputPer1M: 1.60 },
+    'qwen3.6-plus': { inputPer1M: 0.40, outputPer1M: 2.40 },
+    'qwen3.5-plus': { inputPer1M: 0.40, outputPer1M: 2.40 },
+    'qwen3.6-flash': { inputPer1M: 0.25, outputPer1M: 1.50 },
     // ModelScope free tier — no per-token charge.
     'Qwen/Qwen3-Coder-480B-A35B-Instruct': { inputPer1M: 0, outputPer1M: 0 },
 };
@@ -362,6 +376,7 @@ export function formatCostReport() {
             lines.push(`**Estimated savings vs no caching:** $${cache.estimatedSavingsUsd.toFixed(4)}`);
         }
     }
+    lines.push('', ...formatResourceImpactReport(stats.totalTokens));
     // Models with no pricing entry don't contribute to cost — flag so users
     // aren't surprised the total looks low.
     const untracked = breakdown.filter(b => b.estimatedCost === 0 && (b.promptTokens + b.completionTokens) > 0);

package/dist/utils/webFetch.d.ts

@@ -0,0 +1,101 @@
+/**
+ * `@web <url>` inline context — fetch a web page and attach its text
+ * content to the prompt, the same way `@file` attaches a file.
+ *
+ * Supported forms (anywhere in the message, like file mentions):
+ *   @web https://example.com/docs       → full URL
+ *   @web example.com/docs               → https:// is auto-prepended
+ *   @web http://localhost:3000/api      → dev servers work too
+ *
+ * The fetch is best-effort:
+ *   - HTML is stripped to readable text (tags, scripts, styles removed).
+ *   - Output is capped at `MAX_WEB_BYTES` (default 32 KB) so a single
+ *     page can't blow the context window.
+ *   - Markdown conversion is lightweight (headings, links, lists) — we
+ *     don't run a full HTML→Markdown pipeline; the goal is "agent can
+ *     read the page", not "pretty render".
+ *   - Failures (network, non-2xx, non-text content type) surface as
+ *     inline notifications, same as missing files.
+ *
+ * The fetcher is async, so `expandWebMentions` is async — unlike the
+ * sync `expandMentions` for files. Callers await it.
+ */
+/** Max bytes of text we'll inline from a fetched page (32 KB). */
+export declare const MAX_WEB_BYTES: number;
+/** Result of expanding `@web` mentions in a prompt. */
+export interface WebExpansionResult {
+    /** The prompt with fetched page text prepended. */
+    enrichedPrompt: string;
+    /** Successfully fetched pages. */
+    loaded: Array<{
+        url: string;
+        title: string;
+        content: string;
+    }>;
+    /** Mentions that couldn't be fetched, with a human-readable reason. */
+    failures: Array<{
+        mention: string;
+        reason: string;
+    }>;
+}
+/** One `@web` mention match. */
+interface WebToken {
+    /** Full match including `@web `, for display. */
+    raw: string;
+    /** The URL (normalized: `https://` prepended if missing a scheme). */
+    url: string;
+    /** Start index of `raw` in the source. */
+    start: number;
+    /** End index (exclusive). */
+    end: number;
+}
+/**
+ * Extract all `@web` mentions from `text`. Pure (no network).
+ * Returns them in document order.
+ */
+export declare function extractWebMentions(text: string): WebToken[];
+export interface WebFetchOptions {
+    /**
+     * The fetch implementation. Defaults to the global `fetch` (Node 18+).
+     * Injected so tests can mock without hitting the network.
+     */
+    fetchImpl?: typeof fetch;
+}
+/**
+ * Expand all `@web` mentions in `prompt`: fetch each URL, convert the
+ * HTML to readable text, and prepend it as a `[Web pages]` block.
+ * Failures are collected, not thrown.
+ */
+export declare function expandWebMentions(prompt: string, opts?: WebFetchOptions): Promise<WebExpansionResult>;
+/**
+ * Reset the session web cache. Public so callers (e.g. `/web-cache clear`
+ * command, or tests) can force a fresh fetch.
+ */
+export declare function clearWebCache(): void;
+/**
+ * Stats about the session web cache — for `/web-cache status`.
+ */
+export declare function webCacheStats(): {
+    entries: number;
+    maxEntries: number;
+    ttlMinutes: number;
+};
+/**
+ * Convert HTML to readable plain text + a title.
+ *
+ * Lightweight — no full parser dependency. Strips `<script>`, `<style>`,
+ * and tags, collapses whitespace, extracts `<title>` and the first
+ * `<h1>` as the page title. Good enough for the agent to read docs;
+ * not a faithful rendering.
+ */
+export declare function htmlToText(html: string): {
+    title: string;
+    content: string;
+};
+/** Format the `[Web pages]` block prepended to the enriched prompt. */
+export declare function formatWebBlock(pages: Array<{
+    url: string;
+    title: string;
+    content: string;
+}>): string;
+export {};