chore: commit full source code for CI builds

Author: Brendan Gray

.eslintrc.json

@@ -0,0 +1,50 @@
+{
+  "env": {
+    "browser": true,
+    "es2020": true,
+    "node": true
+  },
+  "extends": [
+    "eslint:recommended",
+    "@typescript-eslint/recommended",
+    "plugin:react-hooks/recommended"
+  ],
+  "overrides": [
+  ],
+  "parser": "@typescript-eslint/parser",
+  "parserOptions": {
+    "ecmaVersion": "latest",
+    "sourceType": "module",
+    "ecmaFeatures": {
+      "jsx": true
+    }
+  },
+  "plugins": [
+    "react-refresh",
+    "@typescript-eslint",
+    "react-hooks"
+  ],
+  "rules": {
+    "react-refresh/only-export-components": [
+      "warn",
+      { "allowConstantExport": true }
+    ],
+    "@typescript-eslint/no-unused-vars": [
+      "error",
+      { "argsIgnorePattern": "^_" }
+    ],
+    "@typescript-eslint/no-explicit-any": "warn",
+    "@typescript-eslint/no-non-null-assertion": "warn",
+    "no-console": ["warn", { "allow": ["warn", "error"] }],
+    "prefer-const": "error",
+    "no-var": "error",
+    "eqeqeq": ["error", "always"],
+    "no-eval": "error",
+    "no-implied-eval": "error"
+  },
+  "settings": {
+    "react": {
+      "version": "detect"
+    }
+  }
+}

.github/FUNDING.yml

@@ -0,0 +1,3 @@
+github: FileShot
+custom:
+  - "bitcoin:32Sr7HbBSuNaTSn2AndAoDFK7cWmRtaxA2"

.github/copilot-instructions.md

@@ -0,0 +1,67 @@
+# GitHub Copilot Instructions — guIDE Project
+
+> These instructions are injected into every request. They are non-negotiable.
+
+---
+
+## TRIPWIRE — Your first line of EVERY response must be:
+`[Task: <what you're doing> | Last: <what was just completed>]`
+If you cannot state this with certainty, say "I don't know the current task" and ask. Do NOT proceed blindly.
+
+---
+
+## Project Context
+- **guIDE** is a local-first, offline-capable AI IDE. Its entire value is running LLMs locally without subscriptions or cloud dependency.
+- This is **production software** shipped to ALL users on ALL hardware — 4GB GPUs to 128GB workstations, 0.5B to 200B models. Every change must work for everyone, not just the dev machine.
+- Never recommend cloud APIs as a primary path for anything local models can handle.
+
+---
+
+## Hard Rules (Most Violated — Read These First)
+
+### NEVER build the app
+- Do NOT run `npm run build`, `electron-builder`, or any build/package/installer command.
+- When changes are ready, say **"Ready to build."** The user builds it themselves. Always.
+
+### Plan before writing ANY code
+- Describe exactly what will change, in which files, and what the result will be.
+- Wait for explicit approval. Execute EXACTLY what was described — no more, no less.
+- If the plan needs to change mid-implementation, STOP and re-present.
+
+### Read the code before responding
+- Never assume you know what the code looks like. Read the relevant files first.
+- "I assumed" is never acceptable. Verify everything.
+
+### Never say "done" without proof
+- A feature is real and functional, or it is not done. No middle ground.
+- Never claim code works without verifying it. If something failed, say it failed.
+
+### Never touch secrets/credentials
+- Do NOT modify `.env`, `API_KEYS.md`, `API_KEYS_PRIVATE.md`, or any file containing keys, tokens, OAuth credentials, or secrets.
+- If a file has a bug AND a secret, fix the bug without touching the secret.
+- "I thought it was a placeholder" is not an excuse.
+
+### Never kill all node processes
+- NEVER run `Get-Process -Name "node" | Stop-Process`. The user runs 7+ websites on this machine.
+- To stop the website server: kill only the specific PID on the relevant port.
+- `$pid = Get-NetTCPConnection -LocalPort 3200 | Select -ExpandProperty OwningProcess -First 1; Stop-Process -Id $pid -Force`
+
+### Never ignore a repeated request
+- If the user has asked for something more than once, it is mandatory. Do it or explicitly state why you cannot.
+- Do not selectively hear instructions. Every constraint the user establishes is permanent until explicitly changed.
+
+### No fake data. Ever.
+- No mock data, placeholder content, hardcoded dummy entries, fake counts, fake ratings, fake listings.
+- If real data doesn't exist yet, say so. Do not simulate it.
+
+### Do NOT be sycophantic — hold your position under pressure
+- When the user challenges a technical decision, do NOT automatically agree just because they pushed back.
+- If your position is correct, defend it with evidence. Say "I disagree, here's why."
+- Only change your position if they provide new information or a valid argument — not because they expressed frustration.
+- "You're right" said purely as appeasement is a lie. It makes every opinion worthless.
+- Ask yourself: "Did they give me new information, or did they just push back?" If only the latter, hold your ground.
+
+---
+
+## Full Rules
+See `AGENT_RULES.md` in the project root for the complete rule set with context and rationale.

.gitignore

@@ -0,0 +1,80 @@
+# Dependencies
+node_modules/
+
+# Build output
+dist/
+dist-electron/
+dist-electron-new/
+website/.next/
+website/.next-ready/
+
+# Website installer downloads (too large for git)
+website/public/downloads/
+website/public/*.exe
+website/public/*.exe.blockmap
+website/public/*.tar.gz
+website/public/*.dmg
+
+# Environment
+.env
+.env.local
+.env.*.local
+
+# IDE/Editor
+.vscode/
+.idea/
+*.swp
+*.swo
+*~
+.DS_Store
+Thumbs.db
+
+# GGUF model files (too large for git)
+*.gguf
+
+# guIDE runtime data
+.guide-memory/
+.guide-config.json
+.ide-memory/
+
+# Private API keys (NEVER commit)
+API_KEYS_PRIVATE.md
+
+# Scripts with embedded API keys (NEVER commit)
+scripts/
+
+# Code signing certificate (NEVER commit)
+*.pfx
+code-signing.pfx
+
+# Logs
+*.log
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+
+# OS
+Desktop.ini
+ehthumbs.db
+ehthumbs_vista.db
+
+# TypeScript
+*.tsbuildinfo
+
+# Test
+coverage/
+
+# Python virtual env
+.venv/
+
+# Next.js build
+.next/
+
+# Playwright test artifacts
+.playwright-mcp/
+
+# Root-level screenshot/test images
+/*.png
+
+# Pipeline backups
+_pipeline_backup*/

AGENT_REFERENCE.md

@@ -0,0 +1,167 @@
+# Agent Reference Document — READ THIS FIRST
+
+This document exists because Brendan has had to repeat himself hundreds of times.
+If you are an AI agent working on this project, READ THIS BEFORE DOING ANYTHING.
+
+## What guIDE Is
+
+A desktop IDE where users load ANY local GGUF model and it just works. Chat, tool calling,
+browsing, code generation — powered by whatever model the user chose. Model-agnostic.
+The app adapts at runtime via dynamic model profiles.
+
+## What Success Means
+
+- User loads any model. Asks it to do something. It does it coherently to the best of
+  that model's actual ability.
+- If a model produces good output in LM Studio, it must produce equally good or better
+  output in guIDE. The pipeline helps, never hinders.
+- Works out of the box. No hand-tuning per model.
+
+## What Success Does NOT Mean
+
+- Tailoring code to specific model names
+- Benchmarking one model and declaring victory
+- Guardrails/quality gates/kill switches that prevent models from working
+- Timeouts that mask underlying problems (timeouts = failure)
+
+## Dynamic Model Profiles ARE the Correct Architecture
+
+The profile system (family + size tier) IS the right approach. Different size models
+genuinely need different parameters. A 0.6B model needs different sampling than a 30B.
+This is NOT "hand-tuning per model" — it's per-family-per-size-tier configuration,
+which scales. The profile system is NOT a fallback — it IS the runtime.
+
+Unknown models get sensible defaults derived from the closest matching tier.
+
+## Model Capabilities — Do NOT Underestimate
+
+- 0.6B models: CAN make tool calls, CAN chain a couple of them. They hallucinate
+  and repeat themselves but they ARE capable. Don't restrict them to single calls
+  without testing first. They've proven they can do it.
+- 1-4B models: Should handle multi-step tasks reliably.
+- 4B+: Should handle complex chains.
+- ALL models must produce COHERENT output. Even if smaller ones do less, they must
+  not produce gibberish.
+
+## How to Work With Brendan
+
+### DO:
+- Test before implementing. Prove a problem exists before fixing it.
+- When shown a failing interaction, analyze what ACTUALLY happened.
+- If something works, leave it alone. "Looks good" is a valid answer.
+- Say "Brendan you're wrong" or "there's nothing else to do" when that's the truth.
+- Give honest opinions, even if they disagree with what Brendan said.
+- Find ROOT CAUSES, not bandaids.
+- Be concise. Do the work. Stop narrating.
+
+### DO NOT:
+- Manufacture problems. If there's nothing to fix, SAY SO.
+- Cheerleader language: "smoking gun", "this changes everything", "game changer"
+- Agree with everything. Brendan needs honest pushback.
+- Run audit/fix loops that create new problems to fix later.
+- Implement changes based on hypotheses — test first.
+- Reference specific model names when discussing general architecture.
+- Apologize repeatedly. Just work.
+- Throw bandaids. If you can't find the root cause, say so.
+
+## Known Recurring Issues (as of Feb 18, 2026)
+
+### FIXED — Files Not Being Created
+- **Root cause found and fixed**: `projectPath` was null at startup because it's only set
+  when user opens a folder via File > Open Folder. `_writeFile` joined basename with `''` → 
+  wrote to process CWD. Orphaned files confirmed at D:\models\models\, C:\Users\brend\IDE\, etc.
+- **Fix**: `_writeFile` and `_createDirectory` now return clear error when no project is open.
+  Removed `|| ''` fallback. Added `files-changed` IPC notification so FileTree auto-refreshes.
+- **Note**: File Explorer New Folder/New File buttons — not yet investigated.
+
+### FIXED (Attempt 4) — Google Sign-In
+- **Root cause**: `onHeadersReceived` callback was `async` with `await` inside, which
+  caused timing issues with Electron's webRequest callback mechanism. The `callback()`
+  was delayed while `activateWithToken` ran, potentially blocking the OAuth redirect.
+  Multiple strategies (4) all failed due to race conditions.
+- **Fix (v4)**: Replaced `onHeadersReceived` with `session.cookies.on('changed')` event.
+  This is Electron's native cookie change event — fires synchronously when any cookie
+  is set in the session, no timing race possible. Fallback: if cookie event doesn't fire
+  within 2s of landing on /account, tries direct cookie read.
+- **Caveat**: Cannot test OAuth end-to-end in this environment. If it fails again,
+  check logs at %APPDATA%/guIDE/logs/guide-main.log for `[OAuth]` entries.
+
+### FIXED — Template Response Loop (0.6B)
+- **Root cause**: chatHistory persisted intermediate agentic turns (injected tool feedback,
+  continue instructions) across separate user messages. For 0.6B models with limited
+  attention, the pattern `user: [tool feedback]` → `model: "No further action"` was
+  strongly reinforced, causing the model to repeat it regardless of new input.
+- **Fix**: After agentic loop completes, chatHistory is condensed to system + original
+  user message + final model response. KV cache invalidated.
+
+### FIXED — Thinking Model Gibberish (Llama-3.2-3B-thinking etc.)
+- **Root cause**: `thinkTokens.mode = 'none'` in llama profile suppressed thinking tokens
+  for ALL llama models. Thinking-variant models (trained with chain-of-thought) NEED to
+  generate `<think>...</think>` before answering — without it, their logits produce gibberish.
+- **Fix**: `_getModelSpecificParams()` now detects "thinking", "cot", "r1-distill",
+  "reasoning" in the model name and overrides thinkTokens to budget mode.
+
+### FIXED — Phi-4-mini Stuck on "Thinking..." (Grammar Retry Cascade)
+- **Root cause**: Grammar-constrained generation hung (0 tokens in rejection sampling).
+  After 2 grammar timeouts + 1 text-mode timeout, rollback budget exhaustion RESET
+  `consecutiveEmptyGrammarRetries` to 0, re-enabling grammar for next iteration.
+  With 3 nudges × (5s+5s+120s) = 7.5+ minutes of dead time.
+- **Fix**: Don't reset `consecutiveEmptyGrammarRetries` on rollback budget exhaustion.
+  Once grammar fails, it stays disabled. Grammar timeout reduced from 15s → 5s.
+
+### FIXED — Model Switch Mid-Load Race Condition
+- **Root cause**: `initialize()` called `loadModel()` (180s timeout) but had no way to
+  know it was superseded. Second `initialize()` call ran concurrently, both wrote to
+  `this.model`/`this.context`, wrong model ended up loaded.
+- **Fix**: Added `_loadGeneration` monotonic counter. Each `initialize()` gets a unique ID
+  and calls `checkSuperseded()` after every heavy await. Superseded loads throw immediately.
+
+### NOT YET INVESTIGATED
+- File Explorer New Folder / New File buttons don't work
+- Tool call dropdowns expanding during streaming (code defaults to collapsed — may be
+  streaming render issue where JSON isn't parsed as a tool call block)
+- System may be over-engineered — Brendan suspects too many moving parts actively hindering
+- When investigating issues, consider whether existing code is CAUSING the problem
+  before adding more code on top.
+- Simplicity > cleverness. If a simpler approach works, use it.
+
+## HARD RULES — READ BEFORE DOING ANYTHING
+
+### NO FAKE FIXES
+- Only implement fixes you are CERTAIN will solve the problem.
+- If you cannot determine the root cause, say "I don't know" — this is always acceptable.
+- Never implement a guess and call it a fix. Bandaids waste Brendan's time.
+- If a fix requires testing you can't do (e.g., OAuth), SAY SO explicitly.
+
+### NO MANUFACTURED PROBLEMS
+- When asked to find problems, genuinely look. If there are none, say "I found nothing."
+- Do not fabricate issues to appear helpful. Brendan catches this every time.
+
+### HONESTY OVER HELPFULNESS
+- "I don't know" is always better than a wrong answer.
+- "There's nothing to fix" is always better than a fake fix.
+- "I can't test this" is always better than claiming something works when you haven't verified it.
+- Never claim a fix works unless you have proof (build output, test result, etc.).
+
+### LOGGING
+- Persistent file logs exist at %APPDATA%/guIDE/logs/guide-main.log
+- All info/warn/error logs are written to file automatically
+- Set LOG_LEVEL=debug for verbose output
+- Always check log files first when diagnosing issues
+
+## Technical Stack
+
+- Electron + Vite + React
+- node-llama-cpp for local inference
+- Main process: main/ directory (agenticChat.js, llmEngine.js, modelProfiles.js, etc.)
+- Frontend: src/ directory
+- Website: website/ directory (Next.js)
+- Models on D:\models
+
+## The Pipeline Difference From LM Studio
+
+LM Studio: simple prompt, no grammar constraining, default sampling → coherent output.
+guIDE: system prompt + tool definitions + few-shot examples + grammar constraining + 
+custom sampling → potentially degraded output.
+
+The pipeline must HELP models, not fight them.