codeep 3.0.0 → 3.1.0

package/dist/utils/tokenTracker.js

@@ -22,6 +22,7 @@ const MODEL_CONTEXT_WINDOWS = {
     'gpt-5.4': 1_050_000,
     'gpt-5.4-mini': 400_000,
     // Anthropic
+    'claude-fable-5-1': 1_000_000,
     'claude-fable-5': 1_000_000,
     'claude-opus-5': 1_000_000,
     'claude-sonnet-4-6': 1_000_000,
@@ -97,10 +98,11 @@ const MODEL_PRICING = {
     'gpt-5.4': { inputPer1M: 2.50, outputPer1M: 15.00 },
     'gpt-5.4-mini': { inputPer1M: 0.75, outputPer1M: 4.50 },
     // Anthropic
+    'claude-fable-5-1': { inputPer1M: 10.00, outputPer1M: 50.00 },
     'claude-fable-5': { inputPer1M: 10.00, outputPer1M: 50.00 },
     '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-sonnet-5': { inputPer1M: 2.00, outputPer1M: 10.00 },
     'claude-haiku-4-5-20251001': { inputPer1M: 1.00, outputPer1M: 5.00 },
     // DeepSeek (cache-miss input pricing)
     // DeepSeek moved to peak / off-peak billing on 2026-08-16, with off-peak at
@@ -206,11 +208,18 @@ export function recordTokenUsage(usage, model, provider, actualCostUsd) {
  */
 export function extractOpenAIUsage(data) {
     if (data?.usage) {
-        // OpenAI-protocol `prompt_tokens` is INCLUSIVE of cached prompt tokens
-        // (DeepSeek/OpenAI report cache hits in prompt_tokens_details.cached_tokens).
+        // OpenAI-protocol `prompt_tokens` is INCLUSIVE of cached prompt tokens.
         // Surface them so getCostBreakdown bills cache reads at the discounted
         // rate instead of the full cache-miss input rate.
-        const cached = data.usage.prompt_tokens_details?.cached_tokens || 0;
+        //
+        // Two shapes are in the wild. OpenAI, DeepSeek and Qwen's text models nest
+        // it under prompt_tokens_details; Kimi returns it at the top level, and
+        // Alibaba's own docs say some Qwen models still do and will be migrated
+        // later. Reading only the nested form zeroed every Kimi cache hit, so the
+        // cached portion of a run billed at the full cache-miss rate — five times
+        // what it costs — with nothing anywhere to say so.
+        const nested = data.usage.prompt_tokens_details?.cached_tokens;
+        const cached = (typeof nested === 'number' ? nested : data.usage.cached_tokens) || 0;
         return {
             promptTokens: data.usage.prompt_tokens || 0,
             completionTokens: data.usage.completion_tokens || 0,
@@ -242,6 +251,36 @@ export function extractAnthropicUsage(data) {
     }
     return null;
 }
+/**
+ * What a provider charges to read a cached token, as a fraction of its own
+ * cache-miss input rate.
+ *
+ * 0.1 is Anthropic's ratio and used to be applied to everyone. Kimi lists
+ * $0.19 against a $0.95 cache-miss rate, and Alibaba prices Qwen's implicit
+ * cache at 20% of input — both 0.2, so every cached token on those two was
+ * billed at half what it actually costs.
+ */
+/**
+ * Models whose cache-read rate is not their provider's usual one.
+ *
+ * Fable 5.1 reads a cached token at 0.025× the base input price, where every
+ * other Anthropic model charges 0.1×. Checked before the provider map, because
+ * this is a property of the model and not of the account it runs under.
+ */
+const MODEL_CACHE_READ_RATE = {
+    'claude-fable-5-1': 0.025,
+};
+const CACHE_READ_RATE = {
+    'kimi': 0.2,
+    'kimi-api': 0.2,
+    'qwen': 0.2,
+    'qwen-api': 0.2,
+    'qwen-cn': 0.2,
+    'qwen-cn-api': 0.2,
+    'qwen-token-plan': 0.2,
+};
+/** Anthropic's ratio, and the safest guess for a provider we have not priced. */
+const DEFAULT_CACHE_READ_RATE = 0.1;
 /**
  * Get cost breakdown grouped by provider/model.
  *
@@ -269,16 +308,20 @@ export function getCostBreakdown(startIndex = 0) {
         else {
             const pricing = MODEL_PRICING[record.model];
             if (pricing) {
-                // Anthropic prompt caching: cache_creation_input is billed at 1.25×
-                // the base input rate, cache_read_input at 0.1×. The remaining
-                // (uncached) prompt tokens bill at the standard 1.0× rate.
+                // Cache writes bill at 1.25× the base input rate (Anthropic's, and the
+                // only provider here that charges for them at all); cache reads bill
+                // at whatever fraction the provider charges. The remaining (uncached)
+                // prompt tokens bill at the standard 1.0× rate.
                 const cacheCreate = record.cacheCreationTokens ?? 0;
                 const cacheRead = record.cacheReadTokens ?? 0;
+                const cacheReadRate = MODEL_CACHE_READ_RATE[record.model]
+                    ?? CACHE_READ_RATE[record.provider?.trim().toLowerCase()]
+                    ?? DEFAULT_CACHE_READ_RATE;
                 const uncachedPrompt = Math.max(0, record.promptTokens - cacheCreate - cacheRead);
                 existing.estimatedCost +=
                     (uncachedPrompt / 1_000_000) * pricing.inputPer1M
                         + (cacheCreate / 1_000_000) * pricing.inputPer1M * 1.25
-                        + (cacheRead / 1_000_000) * pricing.inputPer1M * 0.1
+                        + (cacheRead / 1_000_000) * pricing.inputPer1M * cacheReadRate
                         + (record.completionTokens / 1_000_000) * pricing.outputPer1M;
             }
         }

package/dist/utils/tools.js

@@ -57,7 +57,12 @@ export const AGENT_TOOLS = {
     },
     execute_command: {
         name: 'execute_command',
-        description: 'Execute a shell command. Use for npm, git, build tools, tests, etc.',
+        // Naming a few examples read as an exhaustive list: the agent refused
+        // `sleep`, which IS on the allowlist, explaining that its tool was
+        // "limited to package managers and version control". It believed the
+        // description over its own capability. Say the shape instead, and let
+        // the runtime refuse — it already answers with a specific reason.
+        description: 'Execute a shell command. A safety allowlist applies and covers far more than build tooling — package managers, git, test runners, and ordinary utilities such as ls, cat, grep, find, echo, date, sleep and curl. Do not assume a command is forbidden: try it, and a refusal will say so.',
         parameters: {
             command: { type: 'string', description: 'The command to run (e.g., npm, git, node)', required: true },
             args: { type: 'array', description: 'Command arguments as array (e.g., ["install", "lodash"])', required: false },

package/dist/version.d.ts

@@ -1 +1 @@
-export declare const VERSION = "3.0.0";
+export declare const VERSION = "3.1.0";

package/dist/version.js

@@ -1,4 +1,4 @@
 // AUTO-GENERATED by scripts/gen-version.js — do not edit by hand.
 // Baked from package.json at build time so the bun-compiled binary reports
 // the right version (it has no package.json on disk to read at runtime).
-export const VERSION = '3.0.0';
+export const VERSION = '3.1.0';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "codeep",
-  "version": "3.0.0",
+  "version": "3.1.0",
   "description": "AI-powered coding assistant built for the terminal. Multiple LLM providers, project-aware context, and a seamless development workflow.",
   "type": "module",
   "main": "dist/index.js",