codeep 3.1.0 → 3.2.0

package/README.md

@@ -20,9 +20,18 @@
   <a href="https://www.npmjs.com/package/codeep"><img src="https://img.shields.io/npm/v/codeep.svg" alt="npm version"></a>
   <a href="https://www.npmjs.com/package/codeep"><img src="https://img.shields.io/npm/dm/codeep.svg" alt="npm downloads"></a>
   <a href="https://github.com/VladoIvankovic/Codeep/blob/main/LICENSE"><img src="https://img.shields.io/npm/l/codeep.svg" alt="license"></a>
-  <a href="https://github.com/VladoIvankovic/Codeep"><img src="https://img.shields.io/github/stars/VladoIvankovic/Codeep?style=social" alt="GitHub stars"></a>
 </p>
 
+## Contents
+
+- [Install](#installation) · [Quick start](#quick-start) · [Upgrading](#upgrading)
+- [One agent, every surface](#one-agent-every-surface) · [Features](#features) · [Dashboard](#codeep-dashboard)
+- [Commands](#commands) · [Keyboard shortcuts](#keyboard-shortcuts) · [Configuration](#configuration)
+- [Zed (ACP)](#zed-editor-integration-acp) · [VS Code](#vs-code-extension) · [Contributing](#contributing)
+
+Full documentation: **[codeep.dev/docs](https://codeep.dev/docs)**
+
+
 ## One agent, every surface
 
 Codeep keeps the same project context, provider configuration, sessions, and
@@ -33,6 +42,71 @@ agent workflow wherever you work:
 - **[VS Code extension](https://github.com/VladoIvankovic/Codeep-vscode)** — the same agent over ACP, directly inside the editor. Also available on the [VS Code Marketplace](https://marketplace.visualstudio.com/items?itemName=VladoIvankovic.codeep).
 - **[Web dashboard](https://codeep.dev/dashboard)** — synced sessions, usage, costs, tasks, and project activity.
 
+## Installation
+
+### Option 1: npm (Recommended)
+
+```bash
+npm install -g codeep
+```
+
+**Update:**
+```bash
+npm update -g codeep
+```
+
+### Option 2: curl
+
+For a machine without Node. The script is served from codeep.dev and fetched from
+this repository's newest release, so the command does not change if the repository
+moves.
+
+```bash
+curl -fsSL https://codeep.dev/install.sh | bash
+```
+
+**Custom installation directory:**
+```bash
+curl -fsSL https://codeep.dev/install.sh | INSTALL_DIR=~/.local/bin bash
+```
+
+**Specific version:**
+```bash
+curl -fsSL https://codeep.dev/install.sh | VERSION=3.1.1 bash
+```
+
+### Option 3: Manual Binary
+
+Download the latest binary for your platform from [GitHub Releases](https://github.com/VladoIvankovic/Codeep/releases):
+
+| Platform | Binary |
+|----------|--------|
+| macOS Apple Silicon (M1/M2/M3/M4) | `codeep-macos-arm64` |
+| macOS Intel | `codeep-macos-x64` |
+| Linux x86_64 | `codeep-linux-x64` |
+
+```bash
+# Example for macOS Apple Silicon:
+curl -fsSL https://github.com/VladoIvankovic/Codeep/releases/latest/download/codeep-macos-arm64 -o codeep
+chmod +x codeep
+sudo mv codeep /usr/local/bin/
+```
+
+## Quick Start
+
+```bash
+# Navigate to your project directory
+cd /path/to/your/project
+
+# Start Codeep
+codeep
+
+# On first run, enter your API key
+# Get one at: https://z.ai/subscribe?ic=NXYNXZOV14
+```
+
+After installation, `codeep` is available globally in your terminal. Simply run it from any project directory to start coding with AI assistance.
+
 ## Upgrading
 
 Running `npm install -g codeep@latest` picks up the new version — config,
@@ -966,79 +1040,6 @@ Create, view, and complete tasks directly from the CLI — or manage them on the
 
 Tasks are loaded into the agent context so the AI sees them automatically on the next message.
 
-## Installation
-
-### Option 1: curl (Quickest)
-
-```bash
-curl -fsSL https://raw.githubusercontent.com/VladoIvankovic/Codeep/main/install.sh | bash
-```
-
-**Custom installation directory:**
-```bash
-curl -fsSL https://raw.githubusercontent.com/VladoIvankovic/Codeep/main/install.sh | INSTALL_DIR=~/.local/bin bash
-```
-
-**Specific version:**
-```bash
-curl -fsSL https://raw.githubusercontent.com/VladoIvankovic/Codeep/main/install.sh | VERSION=1.0.0 bash
-```
-
-### Option 2: Homebrew (macOS/Linux)
-
-```bash
-brew tap VladoIvankovic/codeep
-brew install codeep
-```
-
-**Update:**
-```bash
-brew upgrade codeep
-```
-
-### Option 3: npm
-
-```bash
-npm install -g codeep
-```
-
-**Update:**
-```bash
-npm update -g codeep
-```
-
-### Option 4: Manual Binary
-
-Download the latest binary for your platform from [GitHub Releases](https://github.com/VladoIvankovic/Codeep/releases):
-
-| Platform | Binary |
-|----------|--------|
-| macOS Apple Silicon (M1/M2/M3/M4) | `codeep-macos-arm64` |
-| macOS Intel | `codeep-macos-x64` |
-| Linux x86_64 | `codeep-linux-x64` |
-
-```bash
-# Example for macOS Apple Silicon:
-curl -fsSL https://github.com/VladoIvankovic/Codeep/releases/latest/download/codeep-macos-arm64 -o codeep
-chmod +x codeep
-sudo mv codeep /usr/local/bin/
-```
-
-## Quick Start
-
-```bash
-# Navigate to your project directory
-cd /path/to/your/project
-
-# Start Codeep
-codeep
-
-# On first run, enter your API key
-# Get one at: https://z.ai/subscribe?ic=NXYNXZOV14
-```
-
-After installation, `codeep` is available globally in your terminal. Simply run it from any project directory to start coding with AI assistance.
-
 ## Commands
 
 ### General

package/dist/config/providers.js

@@ -453,6 +453,7 @@ export const PROVIDERS = {
             },
         },
         models: [
+            { id: 'gpt-6-astra', name: 'GPT-6 Astra', description: 'Frontier GPT — 1M context, 128K output. Rolling out by organization, so a key may not have it yet. 2x the price of 5.6 Sol' },
             { id: 'gpt-5.6-sol', name: 'GPT-5.6 Sol', description: 'Most capable GPT — best for coding & agentic work' },
             { id: 'gpt-5.6-terra', name: 'GPT-5.6 Terra', description: 'Balanced — GPT-5.5 quality at about half the price' },
             { id: 'gpt-5.6-luna', name: 'GPT-5.6 Luna', description: 'Fast and cheap — high-volume workloads' },
@@ -501,8 +502,9 @@ export const PROVIDERS = {
         },
         models: [
             { id: 'gemini-3.1-pro-preview', name: 'Gemini 3.1 Pro', description: 'Most capable Gemini model' },
-            { id: 'gemini-3.7-flash', name: 'Gemini 3.7 Flash', description: 'Latest production Flash model — adjustable thinking, 64K output' },
-            { id: 'gemini-3.6-flash', name: 'Gemini 3.6 Flash', description: 'Previous production Flash model' },
+            { id: 'gemini-3.8-flash', name: 'Gemini 3.8 Flash', description: 'Latest production Flash model — adjustable thinking, 64K output' },
+            { id: 'gemini-3.7-flash', name: 'Gemini 3.7 Flash', description: 'Previous production Flash model' },
+            { id: 'gemini-3.6-flash', name: 'Gemini 3.6 Flash', description: 'Earlier production Flash model' },
             { id: 'gemini-3.5-flash', name: 'Gemini 3.5 Flash', description: 'Stable frontier Flash model for coding and long agentic tasks' },
             { id: 'gemini-3.5-flash-lite', name: 'Gemini 3.5 Flash-Lite', description: 'Latest low-latency, low-cost workhorse' },
         ],
@@ -783,7 +785,7 @@ export function providerNoStreamWithTools(providerId) {
 const SAMPLING_PARAMS_REJECTED = [
     'claude-fable-5', 'claude-opus-5', 'claude-opus-4-8', 'claude-opus-4-7', 'claude-sonnet-5',
     'kimi-k3', 'kimi-k2.7-code', 'kimi-for-coding', 'k3',
-    'gemini-3.7-flash',
+    'gemini-3.7-flash', 'gemini-3.8-flash',
 ];
 export function modelRejectsSamplingParams(model) {
     // Canonicalize both sides. OpenRouter routes these as `google/gemini-3.7-flash`
@@ -837,8 +839,11 @@ export function modelSupportsReasoningEffort(providerId, model) {
                 return false;
             return /^claude-(opus-5|opus-4-([5-9]|\d\d)|sonnet-(4-6|5)|fable-5)/.test(id);
         case 'openai':
-            // GPT-5.x are reasoning models — reasoning_effort across the family (incl. mini).
-            return id.startsWith('gpt-5');
+            // GPT-5.x and GPT-6.x are reasoning models — reasoning_effort across both
+            // families (incl. mini). Written as two prefixes rather than `gpt-`
+            // because GPT-4 and earlier are not reasoning models and must not be
+            // offered the parameter.
+            return id.startsWith('gpt-5') || id.startsWith('gpt-6');
         case 'google':
             // Gemini 3.x thinking_level via the OpenAI-compat reasoning_effort mapping.
             return id.startsWith('gemini-3');

package/dist/renderer/agentExecution.js

@@ -9,7 +9,7 @@ import { chat } from '../api/index.js';
 import { runAgent } from '../utils/agent.js';
 import { TelegramApproval, outcomeForAnswer, describePermissionOutcome } from '../utils/telegramApproval.js';
 import { loadTelegramCredentials } from '../utils/telegramCredentials.js';
-import { composeRunSummary, sendTelegramNotice, shouldNotify } from '../utils/telegramNotify.js';
+import { composeRunMessages, sendTelegramNotice, shouldNotify } from '../utils/telegramNotify.js';
 import { takeRunFromPhone } from '../utils/telegramInbox.js';
 import { isFlatFeeProvider } from '../config/providers.js';
 import { raceApproval } from '../utils/approvalRace.js';
@@ -470,13 +470,19 @@ export async function executeAgentTask(task, dryRun, ctx) {
             // and withholding it leaves "Started —" as the last word.
             if (fromPhone || shouldNotify(elapsedMs, true)) {
                 const payPerUse = costBreakdown.filter(entry => !isFlatFeeProvider(entry.provider));
-                await sendTelegramNotice(noticeCredentials, composeRunSummary({
+                // Usually one message. An answer past Telegram's limit continues into
+                // further ones rather than being cut at the first — awaited in turn so
+                // they arrive in the order they were written.
+                const messages = composeRunMessages({
                     task: runLabel,
                     elapsedMs,
                     answer: fromPhone ? result.finalResponse : undefined,
                     tokens: costBreakdown.reduce((sum, e) => sum + e.promptTokens + e.completionTokens, 0),
                     costUsd: payPerUse.reduce((sum, e) => sum + e.estimatedCost, 0),
-                })).catch(() => false);
+                });
+                for (const message of messages) {
+                    await sendTelegramNotice(noticeCredentials, message).catch(() => false);
+                }
             }
         }
         const sharedFields = {

package/dist/utils/telegramNotify.d.ts

@@ -15,10 +15,31 @@ export declare function formatDuration(ms: number): string;
 /**
  * Telegram refuses a message over 4096 characters outright.
  *
- * Kept well under: the rest of the summary shares the message, and a reply
- * that fills a phone screen twice over is not read on a phone anyway.
+ * Kept well under, because the summary head shares the first message.
  */
 export declare const MAX_ANSWER_LENGTH = 3000;
+/**
+ * How many messages one answer may become.
+ *
+ * A long answer is worth several; an enormous one is not worth a phone buzzing
+ * eleven times, and past a point nobody is reading it there anyway. Three is
+ * enough for an explanation and short of a flood.
+ */
+export declare const MAX_ANSWER_PARTS = 3;
+/**
+ * An answer as the messages to send, in order.
+ *
+ * Cutting at the limit was honest — it said it had cut — but lossy in the place
+ * it hurts: an answer that lists files or walks through a change passes 3000
+ * characters easily, and the conclusion is at the end. So it is split instead,
+ * on a paragraph or line boundary where there is one nearby, and only what
+ * exceeds three messages is cut.
+ *
+ * Mirrors the Mac app's `TelegramAnswerText.partsForPhone` deliberately: two
+ * implementations of "what does the phone get" that can disagree is worse than
+ * either.
+ */
+export declare function splitAnswer(text: string): string[];
 /**
  * Markdown out, plain words in.
  *
@@ -57,7 +78,12 @@ export interface RunSummary {
     answer?: string;
 }
 /**
- * The notification text.
+ * The notification, as the messages to send in order.
+ *
+ * One message when the answer fits, which is the common case. A longer answer
+ * continues into further messages rather than being cut at the first limit —
+ * see `splitAnswer`. The head shares the first message, so the reply is not a
+ * bare wall of text with no idea which run it belongs to.
  *
  * Carries the agent's answer only when the run was started from the phone. A
  * finished run can end with anything in it — a file it read, a command it ran,
@@ -65,7 +91,7 @@ export interface RunSummary {
  * servers, so it travels only where it was actually asked for. Start a run at
  * the terminal and this says there is a result to come back to, and no more.
  */
-export declare function composeRunSummary(summary: RunSummary): string;
+export declare function composeRunMessages(summary: RunSummary): string[];
 /**
  * Send it, and say nothing if it fails.
  *

package/dist/utils/telegramNotify.js

@@ -34,10 +34,74 @@ export function formatDuration(ms) {
 /**
  * Telegram refuses a message over 4096 characters outright.
  *
- * Kept well under: the rest of the summary shares the message, and a reply
- * that fills a phone screen twice over is not read on a phone anyway.
+ * Kept well under, because the summary head shares the first message.
  */
 export const MAX_ANSWER_LENGTH = 3000;
+/**
+ * How many messages one answer may become.
+ *
+ * A long answer is worth several; an enormous one is not worth a phone buzzing
+ * eleven times, and past a point nobody is reading it there anyway. Three is
+ * enough for an explanation and short of a flood.
+ */
+export const MAX_ANSWER_PARTS = 3;
+/**
+ * An answer as the messages to send, in order.
+ *
+ * Cutting at the limit was honest — it said it had cut — but lossy in the place
+ * it hurts: an answer that lists files or walks through a change passes 3000
+ * characters easily, and the conclusion is at the end. So it is split instead,
+ * on a paragraph or line boundary where there is one nearby, and only what
+ * exceeds three messages is cut.
+ *
+ * Mirrors the Mac app's `TelegramAnswerText.partsForPhone` deliberately: two
+ * implementations of "what does the phone get" that can disagree is worse than
+ * either.
+ */
+export function splitAnswer(text) {
+    const plain = text.trim();
+    if (!plain)
+        return [];
+    const parts = [];
+    let rest = plain;
+    while (rest.length > 0 && parts.length < MAX_ANSWER_PARTS) {
+        if (rest.length <= MAX_ANSWER_LENGTH) {
+            parts.push(rest);
+            rest = '';
+            break;
+        }
+        const cut = breakPoint(rest);
+        parts.push(rest.slice(0, cut).trim());
+        rest = rest.slice(cut).replace(/^[\n ]+/, '');
+    }
+    // Say it was cut rather than ending mid-sentence and looking finished.
+    if (rest.length > 0 && parts.length > 0) {
+        parts[parts.length - 1] += '\n\n[…cut — the full answer is in the terminal]';
+    }
+    return parts;
+}
+/**
+ * Where to end a part: the last paragraph break in the final third of the
+ * allowance, else the last line break, else the last space.
+ *
+ * Splitting mid-word reads as damage; splitting mid-paragraph reads as a
+ * continuation. Only the tail is searched so a single long paragraph does not
+ * send a 200-character message and push the rest into the next one.
+ */
+function breakPoint(text) {
+    const earliest = Math.floor((MAX_ANSWER_LENGTH * 2) / 3);
+    const window = text.slice(earliest, MAX_ANSWER_LENGTH);
+    const paragraph = window.lastIndexOf('\n\n');
+    if (paragraph >= 0)
+        return earliest + paragraph;
+    const line = window.lastIndexOf('\n');
+    if (line >= 0)
+        return earliest + line;
+    const space = window.lastIndexOf(' ');
+    if (space >= 0)
+        return earliest + space;
+    return MAX_ANSWER_LENGTH;
+}
 /**
  * Markdown out, plain words in.
  *
@@ -69,16 +133,8 @@ export function stripMarkdown(text) {
         .replace(/`([^`\n]+)`/g, '$1')
         .trim();
 }
-/**
- * The notification text.
- *
- * Carries the agent's answer only when the run was started from the phone. A
- * finished run can end with anything in it — a file it read, a command it ran,
- * a secret inside an error — and this goes to a chat that syncs to Telegram's
- * servers, so it travels only where it was actually asked for. Start a run at
- * the terminal and this says there is a result to come back to, and no more.
- */
-export function composeRunSummary(summary) {
+/** The head: what ran, how long it took, what it cost. Never the answer. */
+function composeHead(summary) {
     const head = summary.failure ? '⚠️ Codeep stopped' : '✅ Codeep finished';
     const lines = [`${head} — ${summary.task}`, `took ${formatDuration(summary.elapsedMs)}`];
     if (summary.failure)
@@ -92,16 +148,29 @@ export function composeRunSummary(summary) {
     }
     if (cost.length > 0)
         lines.push(cost.join(' · '));
-    const answer = summary.answer ? stripMarkdown(summary.answer) : undefined;
-    if (answer) {
-        lines.push('');
-        lines.push(answer.length > MAX_ANSWER_LENGTH
-            // Say it was cut rather than ending mid-sentence and looking finished.
-            ? `${answer.slice(0, MAX_ANSWER_LENGTH)}\n\n[…cut — the full answer is in the terminal]`
-            : answer);
-    }
     return lines.join('\n');
 }
+/**
+ * The notification, as the messages to send in order.
+ *
+ * One message when the answer fits, which is the common case. A longer answer
+ * continues into further messages rather than being cut at the first limit —
+ * see `splitAnswer`. The head shares the first message, so the reply is not a
+ * bare wall of text with no idea which run it belongs to.
+ *
+ * Carries the agent's answer only when the run was started from the phone. A
+ * finished run can end with anything in it — a file it read, a command it ran,
+ * a secret inside an error — and this goes to a chat that syncs to Telegram's
+ * servers, so it travels only where it was actually asked for. Start a run at
+ * the terminal and this says there is a result to come back to, and no more.
+ */
+export function composeRunMessages(summary) {
+    const head = composeHead(summary);
+    const parts = summary.answer ? splitAnswer(stripMarkdown(summary.answer)) : [];
+    if (parts.length === 0)
+        return [head];
+    return [`${head}\n\n${parts[0]}`, ...parts.slice(1)];
+}
 /**
  * Send it, and say nothing if it fails.
  *

package/dist/utils/tokenTracker.js

@@ -15,6 +15,7 @@ const MODEL_CONTEXT_WINDOWS = {
     'glm-5.1': 200_000,
     'glm-5-turbo': 202_752,
     // OpenAI
+    'gpt-6-astra': 1_050_000,
     'gpt-5.6-sol': 1_050_000,
     'gpt-5.6-terra': 1_050_000,
     'gpt-5.6-luna': 1_050_000,
@@ -33,6 +34,7 @@ const MODEL_CONTEXT_WINDOWS = {
     'deepseek-v4-flash': 1_000_000,
     // Google
     'gemini-3.1-pro-preview': 1_048_576,
+    'gemini-3.8-flash': 1_048_576,
     'gemini-3.7-flash': 1_048_576,
     'gemini-3.6-flash': 1_048_576,
     'gemini-3.5-flash': 1_048_576,
@@ -91,6 +93,9 @@ const MODEL_PRICING = {
     'glm-5.1': { inputPer1M: 1.40, outputPer1M: 4.40 },
     'glm-5-turbo': { inputPer1M: 1.20, outputPer1M: 4.00 },
     // OpenAI
+    // Twice the price of 5.6 Sol. Cached reads are $1.00/M — 0.1x input, which is
+    // what DEFAULT_CACHE_READ_RATE already applies, so no model row is needed.
+    'gpt-6-astra': { inputPer1M: 10.00, outputPer1M: 50.00 },
     'gpt-5.6-sol': { inputPer1M: 5.00, outputPer1M: 30.00 },
     'gpt-5.6-terra': { inputPer1M: 2.00, outputPer1M: 12.00 },
     'gpt-5.6-luna': { inputPer1M: 0.20, outputPer1M: 1.20 },
@@ -114,9 +119,13 @@ const MODEL_PRICING = {
     'deepseek-v4-pro': { inputPer1M: 0.435, outputPer1M: 0.87 },
     'deepseek-v4-flash': { inputPer1M: 0.14, outputPer1M: 0.28 },
     // Google
-    // Gemini 3.6/3.7 Flash carry PROMOTIONAL rates that run through 2026-12-31 and
-    // step up to 1.50/7.50 on 2027-01-01 — revisit both rows on that date.
+    // Gemini 3.6/3.7/3.8 Flash carry PROMOTIONAL rates that run through 2026-12-31
+    // and are scheduled to step up to 1.50/7.50 on 2027-01-01 — revisit all three
+    // rows on that date. Deliberately NOT written ahead of time: Anthropic
+    // cancelled exactly such a scheduled rise for Sonnet 5, and the pre-entered
+    // figure over-reported every run by 50% until someone noticed by hand.
     'gemini-3.1-pro-preview': { inputPer1M: 2.00, outputPer1M: 12.00 },
+    'gemini-3.8-flash': { inputPer1M: 0.75, outputPer1M: 3.75 },
     'gemini-3.7-flash': { inputPer1M: 0.75, outputPer1M: 3.75 },
     'gemini-3.6-flash': { inputPer1M: 0.75, outputPer1M: 3.75 },
     'gemini-3.5-flash': { inputPer1M: 1.50, outputPer1M: 9.00 },

package/dist/utils/update.js

@@ -109,7 +109,7 @@ export function getUpdateInstructions() {
         case 'npm':
             return 'npm update -g codeep';
         case 'binary':
-            return 'curl -fsSL https://raw.githubusercontent.com/VladoIvankovic/Codeep/main/install.sh | bash';
+            return 'curl -fsSL https://codeep.dev/install.sh | bash';
         default:
             return 'Visit: https://codeep.dev';
     }

package/dist/version.d.ts

@@ -1 +1 @@
-export declare const VERSION = "3.1.0";
+export declare const VERSION = "3.2.0";

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 = '3.1.0';
+export const VERSION = '3.2.0';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "codeep",
-  "version": "3.1.0",
+  "version": "3.2.0",
   "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",