glitool 2.0.1 → 2.0.3

package/dist/agent.js

@@ -26,10 +26,17 @@ const __dirname = dirname(__filename);
 loadEnv({ path: join(os.homedir(), '.glitool', '.env') });
 const MAX_HISTORY_CHARS = 60_000;
 // const simpleLlm = makeLlm('meta-llama/Llama-3.3-70B-Instruct-Turbo');
-export const llm = createLlm('meta-llama/Llama-3.3-70B-Instruct-Turbo');
+// const simpleLlm = makeLlm('meta-llama/Llama-3.3-70B-Instruct-Turbo');
 function createLlm(model) {
     return makeLlm(model);
 }
+/**
+ * Lazy default LLM. Fresh instance each call so it picks up the
+ * current request_id set by startNewRequest() at the top of chat().
+ */
+export function getDefaultLlm() {
+    return createLlm('meta-llama/Llama-3.3-70B-Instruct-Turbo');
+}
 // const config = loadConfig();
 const tools = [listFilesTool, readFileTool, searchCodeTool, writeFileTool, editFileTool, bashTool, readBackgroundOutputTool, webFetchTool];
 process.on('exit', cleanupAll);
@@ -48,18 +55,33 @@ function buildSystemPrompt() {
     if (!summary) {
         const rawSession = loadSession();
         if (rawSession.length > 4) {
-            generateAndSaveSummary(rawSession, llm);
+            generateAndSaveSummary(rawSession, getDefaultLlm());
             summary = loadSummary();
         }
     }
-    let prompt = `You are an expert coding assistant. Be concise and code-focused.
+    let prompt = `You are Glitool — an AI coding assistant running in the user's terminal.
+
+You can:
+- Have a normal conversation (just respond, no tools needed)
+- Answer programming and general questions
+- Read, write, edit, and search files in the user's project
+- Run shell commands, fetch web pages
+- Plan, review, refactor, debug code, and help with git
+
+Be concise. Default to plain conversation. Only call tools when the request clearly needs them.
+
+When the user asks to read, show, or display a specific file → call readFile.
+For "read <name>" shorthand, pass the bare name; the tool searches the project automatically.
+Don't claim a file is missing without verifying via listFiles or readFile first.
 
-CRITICAL — file operations:
-- When the user asks to read, show, view, or display a file, you MUST call the readFile tool. NEVER answer from memory or guess at file contents.
-- When the user asks if a file exists, you MUST call listFiles or readFile to verify. NEVER claim a file is missing without checking.
-- For "read <name>" prompts, call readFile with the bare name — the tool will search the project automatically.
+If any tool returns USER_CANCELLED, stop immediately and tell the user. Never retry a cancelled operation.
 
-IMPORTANT: If any tool returns USER_CANCELLED, immediately stop all tool calls and tell the user the operation was cancelled. Never retry a cancelled operation.`;
+Style:
+- No preamble like "Sure!", "Of course!", "I'd be happy to..."
+- Code blocks use language tags: \`\`\`ts, \`\`\`py, \`\`\`bash
+- File references use backticks: \`src/auth.ts:42\`
+- The user's current working directory IS the project they're working on — file paths are relative to it.
+`;
     if (summary) {
         const capped = summary.length > MAX_SUMMARY_CHARS
             ? summary.slice(0, MAX_SUMMARY_CHARS) + '\n…[summary truncated]'

package/dist/agents/coder.js

@@ -22,7 +22,14 @@ GROUNDING RULES — these are not optional:
 6. Maximum 5 file reads per task. If you need more, you're doing it wrong — use searchCode instead.
 7. If you can't safely complete the task, STOP and return a failure message. Do not invent.
 
-Be surgical, not exhaustive. Most tasks need 2-4 tool calls, not 15. The validator will catch broken output — you don't need to over-verify.`)
+Be surgical, not exhaustive. Most tasks need 2-4 tool calls, not 15. The validator will catch broken output — you don't need to over-verify.
+
+Response style:
+- Your final text should be 1-3 sentences summarizing what files you changed and why.
+- Do NOT paste file contents in the response — the files are on disk; the user can read them.
+- The validator runs tsc + ESLint after you finish — no need to verify those yourself.
+- If a step is impossible (binary file, command blocked, etc.), say so explicitly and stop.
+`)
     });
     const stream = await coderAgent.stream({ messages: [new HumanMessage(`Plan to execute:\n${plan}\n\nOriginal request: ${userMessage}`)] }, { recursionLimit: 60, streamMode: 'updates' });
     let result = '';

package/dist/agents/explainer.js

@@ -1,16 +1,20 @@
 import { makeLlm } from '../llm/factory.js';
 import { SystemMessage, HumanMessage } from "@langchain/core/messages";
-const explainerLlm = makeLlm('meta-llama/Llama-3.3-70B-Instruct-Turbo');
 export async function explainResponse(response) {
     if (!response || response.length < 50)
         return '';
+    const explainerLlm = makeLlm('meta-llama/Llama-3.3-70B-Instruct-Turbo');
     const result = await explainerLlm.invoke([
-        new SystemMessage(`You are a coding teacher. Give what an AI coding assistant just did, explain it in 2-4 simple sentences a biginner would underStand.     
-            Focus on:
-            - What changed and why
-            - What concept was used
-            - What to learn next
-            Keep it friendly and short.`),
+        new SystemMessage(`You are a friendly coding teacher.
+
+Given what an AI coding assistant just did (described below), explain it in 2-4 simple sentences that a beginner would understand.
+
+Focus on:
+- WHAT changed and WHY
+- WHICH concept was used (e.g., "promise chaining", "dependency injection")
+- ONE thing to learn next
+
+Tone: encouraging, plain language, define any jargon used.`),
         new HumanMessage(`What was just done:\n${response}`)
     ]);
     return result.content;

package/dist/agents/judge.js

@@ -23,7 +23,7 @@ IMPORTANT — how the coder works:
 Return JSON only:
 {
   "verdict": "ok" or "fail",
-  "failure_point": "plan" | "workflow" | "executor" | "final_output" | null,
+   "failure_point": "plan" | "workflow" | "execution" | "final_output" | null,
   "failure_step_id": number or null,
   "reason": "short explanation",
   "fix_hint": "specific instruction to fix the problem, empty if ok",
@@ -34,7 +34,7 @@ Return JSON only:
 failure_point meanings:
 - "plan": the plan itself was wrong or incomplete — needs replanning
 - "workflow": wrong execution order caused the failure
-- "executor": a specific step failed — use failure_step_id
+- "execution": a specific step failed — use failure_step_id
 - "final_output": code runs but doesn't meet the requirement
 - null: everything ok, verdict must be ok`),
         new HumanMessage(`User request: ${input.userMessage}\n\n` +

package/dist/agents/planner.js

@@ -5,15 +5,30 @@ export async function runPlanner(userMessage, context, model) {
     const response = await llm.invoke([
         new SystemMessage(`You are a coding task planner. Output a structured JSON plan.
 
+Output exactly one of:
+
+A) The literal text:  SIMPLE
+   (use when the task is just a question, explanation, or chat — no file changes or code execution needed)
+
+B) A JSON array of 3-6 steps:
+[
+  { "id": 1, "action": "read",   "target": "src/auth.ts",     "depends_on": [], "why": "understand current login flow" },
+  { "id": 2, "action": "edit",   "target": "src/auth.ts",     "depends_on": [1], "why": "add refresh-token check before validate" },
+  { "id": 3, "action": "run",    "target": "npx tsc --noEmit","depends_on": [2], "why": "verify it compiles" }
+]
+
+Action values:
+- "read":   open a file to understand it
+- "edit":   modify an existing file
+- "create": make a new file
+- "run":    execute a shell command (tsc, npm, git, etc.)
+- "search": grep the codebase
+
 Rules:
-- If the task is ONLY a question or explanation with NO file creation or code execution needed, output exactly: SIMPLE
-- Otherwise output a JSON array of steps with this shape:
-  [{ "id": 1, "action": "read|edit|create|run|search", "target": "file/path or command", "depends_on": [], "why": "reason" }]
-- Be specific about which files to read, edit, or create
-- Do NOT write any code — only plan the steps
-- Keep it to 3-6 steps maximum
-- Steps that can run independently should have no shared depends_on
-- depends_on contains the ids of steps that must finish before this one`),
+- Output ONLY the SIMPLE literal OR the JSON array. No prose, no markdown, no code fences.
+- Steps that can run in parallel: leave depends_on empty.
+- Sequential dependencies: list step ids in depends_on.
+- Be specific about file paths and command names — vague targets cause failures.`),
         new HumanMessage(`Context:\n${context}\n\nUser request: ${userMessage}`)
     ]);
     const content = response.content.trim();

package/dist/agents/planningAgent.js

@@ -5,10 +5,26 @@ import { join } from "path";
 const PLAN_FILE = "plan.md";
 const BLOCKED_EXTENSIONS = ['.ts', '.js', '.tsx', '.jsx', '.py', '.go', '.rs'];
 export async function runPlanningAgent(userMessage, onUsage) {
-    const llm = makeLlm('deepseek-ai/DeepSeek-V3');
+    const llm = makeLlm('openai/gpt-oss-120b');
     const planPath = join(process.cwd(), PLAN_FILE);
     const existingPlan = existsSync(planPath) ? readFileSync(planPath, 'utf-8') : null;
-    const systemPrompt = existingPlan ? `...existing logic...` : `You are a planning assistant. Create a clear structured plan based on user's request.
+    const systemPrompt = existingPlan
+        ? `You are a planning assistant. The user has an existing plan.md and wants to update it.
+
+You receive: (1) the current plan, (2) the user's change request.
+
+OUTPUT:
+- The FULL updated plan in Markdown — this overwrites plan.md.
+- After the plan, write exactly "---" on its own line.
+- Then 1-3 bullets summarising what changed.
+
+RULES:
+- Preserve sections the user didn't ask to change.
+- Apply requested changes precisely; don't expand scope.
+- If the request is ambiguous, ask ONE clarifying question instead of guessing.
+- Never delete existing sections unless the user explicitly says to.
+- Avoid writing implementation code; this is planning, not execution.`
+        : `You are a planning assistant. Create a clear structured plan based on user's request.
 
 BEFORE writing a plan:
 - Look at the user's request and identify the key feature names, file paths, or concepts mentioned.
@@ -21,8 +37,7 @@ Rules:
 - Use Markdown with clear sections and numbered steps
 - Be specific: name files, components, decisions, trade-offs
 - If the request is vague, prefer asking 1-2 clarifying questions over guessing
-- After the plan, write exactly "---" on its own line, then 1-3 bullet points summarising what you created
-`;
+- After the plan, write exactly "---" on its own line, then 1-3 bullet points summarising what you created`;
     const userContent = existingPlan ? `Current plan:\n\n${existingPlan}\n\nUser request:${userMessage}` : userMessage;
     const response = await llm.invoke([
         new SystemMessage(systemPrompt),

package/dist/agents/reviewer-agent.js

@@ -27,33 +27,29 @@ Workflow:
 2. Run \`npx tsc --noEmit\` via the bash tool and capture type errors.
 3. Run \`npx eslint <changed-files>\` via the bash tool (skip if eslint isn't configured).
 4. Read the files. Combine static-analysis output with your own reading.
-5. Return a markdown report in EXACTLY this shape:
+5. Return a plain-text report (terminal-friendly, NO markdown):
 
-## Summary
+SUMMARY
 <one or two sentences — overall health>
 
-## Issues
-
-### CRITICAL
+CRITICAL
 - file.ts:LINE — what is wrong and why it matters
 
-### WARNING
+WARNING
 - file.ts:LINE — what is wrong and how to think about it
 
-### SUGGESTION
+SUGGESTION
 - file.ts:LINE — small improvement
 
-## What's good
+GOOD
 - one or two things the code does well
 
-If a section has no items, write "none" under its heading. Do not omit sections.
-
-TERMINAL OUTPUT RULES — these override everything else about formatting:
-- No markdown headers (no ##, no ###). Use plain uppercase labels: SUMMARY, CRITICAL, WARNING, SUGGESTION, GOOD.
-- No **bold** or *italic*. Plain text only.
-- No bullet dashes longer than needed — one short line per issue.
-- Each issue: filename:LINE — one sentence max.
-- Total response: 25 lines maximum. Cut low-value suggestions if needed to stay under.`;
+FORMAT RULES:
+- No ##, ###, **bold**, *italic*. Plain text only.
+- Each issue: one line, exactly "filename:LINE — single sentence".
+- If a section has no items, write "none" under the label.
+- Total response: 25 lines maximum. Cut low-value suggestions to stay under.
+`;
 export async function runReviewer(userMessage, onToolCall, model) {
     const llm = makeLlm(model);
     const tools = [listFilesTool, readFileTool, searchCodeTool, bashTool];

package/dist/index.js

@@ -9,7 +9,7 @@ import { render } from "ink";
 import { App } from "./ui/App.js";
 import { loadConfig, saveConfig } from "./config.js";
 import { generateAndSaveSummary } from "./memory.js";
-import { llm, sessionMessages } from "./agent.js";
+import { getDefaultLlm, sessionMessages } from "./agent.js";
 import { extractAndSaveProjectMemory } from "./projectMemory.js";
 import os from 'os';
 // import { mkdirSync, writeFileSync, existsSync } from 'fs';
@@ -70,8 +70,9 @@ async function ensureApiKey() {
 }
 const saveAndExit = async () => {
     try {
-        await generateAndSaveSummary(sessionMessages, llm);
-        await extractAndSaveProjectMemory(sessionMessages, llm);
+        const summaryLlm = getDefaultLlm();
+        await generateAndSaveSummary(sessionMessages, summaryLlm);
+        await extractAndSaveProjectMemory(sessionMessages, summaryLlm);
     }
     catch {
         // exit cleanly even if LLM is unreachable

package/dist/llm/classifier.js

@@ -41,7 +41,14 @@ Tie-breakers:
 - starts with the literal word "git" → git
 - short greeting or opinion question → chat
 
-Return JSON: {"domain":"<domain>","confidence":"high" or "low", "reason":"<one sentence>"}`),
+Confidence guide:
+- "high": user's intent is unambiguous from this single message; pattern + verb both point one way.
+- "low": message is short or uses pronouns ("this", "it", "that"), or could plausibly fit 2+ domains, or relies on context the history doesn't fully clarify.
+
+Return JSON: {"domain":"<domain>","confidence":"high" or "low", "reason":"<one sentence>"}
+
+
+`),
             new HumanMessage(history ? `Conversation so far:\n${history}\n\nNew message: ${prompt}` : `Message: ${prompt}`)
         ], { timeout: 3000 });
         const raw = typeof response.content === 'string' ? response.content : '';

package/dist/memory.js

@@ -79,7 +79,14 @@ export async function generateAndSaveSummary(messages, llm) {
     if (!readable.trim())
         return;
     const response = await llm.invoke([
-        new SystemMessage('Summarize this coding session in 2-3 sentences. Focus on: what was built, what decisions ware made, and what the next step is.'),
+        new SystemMessage(`You're writing a brief for the next coding session. The user will return to this project days or weeks later and need fast context recall.
+
+In 2-3 sentences, summarize:
+- What was built or changed (be specific: file paths, function names)
+- Key decisions made (e.g., "chose JWT over sessions", "Postgres over Mongo")
+- Likely next step
+
+Avoid generic phrases like "we worked on the project". Cite specifics.`),
         new HumanMessage(readable)
     ]);
     const hash = crypto.createHash('md5').update(process.cwd()).digest('hex').slice(0, 8);

package/dist/projectMemory.js

@@ -29,15 +29,30 @@ export async function extractAndSaveProjectMemory(messages, llm) {
         return;
     const existing = loadProjectMemory();
     const response = await llm.invoke([
-        new SystemMessage(`Extract structured project facts from this conversation. return valid JSON only:
-            {
-            "techStack": ["languages, frameworks, libraries mentioned"],
-            "architectureDecisions": ["key structural decisions made"],
-            "todos": ["next steps or TODOs mentioned"],
-            "lastUpdated": "${new Date().toISOString()}"
-            }
-            
-            ${existing ? `Merge with existing memory: ${JSON.stringify(existing)}` : ''}`),
+        new SystemMessage(`Extract structured project facts from this conversation.
+
+OUTPUT REQUIREMENTS:
+- Valid JSON ONLY. No markdown, no code fences, no prose.
+- Exactly this shape:
+{
+  "techStack": ["TypeScript", "Express", "MongoDB"],
+  "architectureDecisions": ["use JWT not sessions", "Mongo over Postgres for prototyping"],
+  "todos": ["add password reset", "wire Stripe webhooks"],
+  "lastUpdated": "${new Date().toISOString()}"
+}
+
+EXTRACTION RULES:
+- techStack: only languages/frameworks/libraries that were USED or CHOSEN — not briefly mentioned and discarded.
+- architectureDecisions: concrete choices ("use JWT for auth"), not vague phrases ("we'll handle auth").
+- todos: actual next steps mentioned, not aspirational tangents.
+
+${existing ? `MERGE WITH EXISTING:
+- techStack: union (no duplicates).
+- architectureDecisions: append new, keep old unless contradicted.
+- todos: keep old unless explicitly completed in this conversation.
+
+Existing memory:
+${JSON.stringify(existing, null, 2)}` : ''}`),
         new HumanMessage(readable)
     ]);
     try {

package/dist/ui/App.js

@@ -2,7 +2,7 @@ import { jsx as _jsx, jsxs as _jsxs, Fragment as _Fragment } from "react/jsx-run
 import React, { useState, useRef } from "react";
 import { Box, Text, useApp, Static } from 'ink';
 import TextInput from "ink-text-input";
-import { chat, clearSession, llm, sessionMessages } from '../agent.js';
+import { chat, clearSession, getDefaultLlm, sessionMessages } from '../agent.js';
 import { clearSummary, generateAndSaveSummary } from "../memory.js";
 import { clearProjectMemory, extractAndSaveProjectMemory } from "../projectMemory.js";
 import { useInput } from 'ink';
@@ -129,8 +129,9 @@ export const App = ({ explainMode = false }) => {
         setPaletteIndex(0);
     };
     const handleExit = async () => {
-        await generateAndSaveSummary(sessionMessages, llm);
-        await extractAndSaveProjectMemory(sessionMessages, llm);
+        const summaryLlm = getDefaultLlm();
+        await generateAndSaveSummary(sessionMessages, summaryLlm);
+        await extractAndSaveProjectMemory(sessionMessages, summaryLlm);
         exit();
     };
     useInput((inputKey, key) => {
@@ -335,7 +336,7 @@ export const App = ({ explainMode = false }) => {
                         "You've used your 5 free requests.",
                         '',
                         'Sign in with GitHub — free, 50 requests/month:',
-                        '  → https://glitool.dev/activate',
+                        '   → https://glit.in/activate',
                         '',
                         'Type /signup to start the sign-in flow in your terminal.',
                     ].join('\n'),
@@ -420,9 +421,28 @@ export const App = ({ explainMode = false }) => {
             }
         }
         catch (err) {
+            const raw = err?.message ?? 'Something went wrong.';
+            let friendly = raw;
+            if (raw.includes('anon_limit') || raw.includes('"anon_limit"')) {
+                friendly =
+                    "Your free trial (5 requests) is over.\n\n" +
+                        "→ Type /signup to continue with 50 free requests/month.\n" +
+                        "  It takes one click — sign in with GitHub.";
+            }
+            else if (raw.includes('monthly_limit') || raw.includes('"monthly_limit"')) {
+                friendly =
+                    "You've reached your monthly limit of 50 requests.\n\n" +
+                        "→ Upgrade to Pro for unlimited at https://glit.in/upgrade\n" +
+                        "Or wait until next month — your limit resets on the 1st.";
+            }
+            else if (raw.includes('Token expired') || raw.includes('Invalid token')) {
+                friendly =
+                    "Your session expired.\n\n" +
+                        "→ Type /signup to sign in again.";
+            }
             setMessages(prev => [...prev, {
                     role: 'error',
-                    content: err?.message ?? 'Something went wrong.'
+                    content: friendly
                 }]);
         }
         finally {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "glitool",
-  "version": "2.0.1",
+  "version": "2.0.3",
   "description": "AI coding assistant for your terminal",
   "main": "index.js",
   "type": "module",