✨ Curated, hardened and animated by Shorif uddin — Banglar coder er bondhu ✨

🚀 An IDE to develop real, Gradle-based Android applications on Android devices — with a powerful, multi-provider AI coding agent that understands your project, fixes your build errors, streams answers in real time, and shows you a colored +/- diff before touching a single file.

This is the piashmsu fork of AndroidCodeStudio, heavily extended by Shorif uddin with: 7+ AI providers, build-error auto-fix, streaming responses, diff preview, editor AI actions, voice + image input, sticky-provider auto-fallback, and dozens of bug fixes & optimisations.

Active branch: devin/1777186665-ai-agent-openrouter-autofix — open PR: #1

🤖 AI Agent · ⚡ Build-error auto-fix · ⚙️ Settings → AI · 🛠 Build from source · 👤 Author

---

🤖 What's new in this fork

🌐 AI providers — what you can plug in

Provider	Status	Notes
OpenRouter	New	21+ curated free + paid models, custom vendor/model slug supported, fallback chain
OpenAI-compatible (custom URL)	New	Manually configure Base URL + API key + Model id. Works with Together, Groq, DeepInfra, Fireworks, Ollama (http://10.0.2.2:11434/v1), self-hosted vLLM, anything that speaks /v1/chat/completions
OpenAI	Existing	Bug-fixed (was sharing DeepSeek strings)
Anthropic Claude	Existing	Bug-fixed
Google Gemini	Existing	—
DeepSeek	Existing	—
xAI Grok	Existing	Bug-fixed
Local LLM	Existing	Improved integration

⚡ AI agent capability matrix

Capability	Status
Token-by-token streaming response (SSE)	New
Build-error auto-fix dialog	New
Live progress dialog with streamed AI reply, file list, activity log, error excerpt	New
Auto-rebuild after AI fix (max 3 cycles)	New
Diff preview before applying file changes (line-level, optional)	New
Editor selection actions: Explain / Refactor / Add docs / Generate test	New
Project memory (.aistudio/memory.md)	New
Quick-start templates: Calculator, Todo, Chat, Login	New
Markdown rendering for AI replies (code blocks, bold, lists, headings)	New
Token counter (per request + session total)	New
Conversation export (markdown + JSON)	New
OpenRouter fallback chain (free → free → free)	New
Crash → AI auto-prefill on next launch	New
@filename mentions in chat auto-attach file contents	New
Voice input (Bengali + English, device speech recognizer)	New
Provider-preset picker (Together / Groq / DeepInfra / Fireworks / Anyscale / Mistral / Perplexity / OpenAI / OpenRouter / Ollama / vLLM)	New
Auto-detect local Ollama	New
Settings JSON export / import	New
Offline indicator (skip 30 s watchdog when no network)	New
Sensitive-data detection in outgoing prompts	New
ai_backups/ retention (30 days / 200 files max)	New
Planning mode (think-then-plan-then-code) toggle	New
Markdown pipe-table rendering in AI replies	New
Image input (vision models, gallery picker, base64 multimodal)	New
Inline ghost text (Copilot-style)	Planned
Multi-file smart context picker	Planned
Tool calling / function calling	Planned

---

⚡ AI Fix — live build-error workflow

When a Gradle build fails:

1. The IDE captures full Gradle output, the failing task name, and the error tail.
2. A confirmation dialog asks if you want the AI to fix it.
3. Tapping Fix opens a live progress dialog (85% screen height, scrollable):
   • Header card — spinner, status text, attempt chip (e.g. Attempt 2/3)
   • Build error excerpt — extracts Kotlin e: ... and error: ... lines first, filters Gradle deprecation/help noise, deduped, max 12 lines
   • Files being modified — RecyclerView with success/fail icons, count badge
   • AI reply card — markdown-rendered streaming text (80 ms throttling, smooth)
   • Activity log — color-coded (info / file / success / warn / error) with timestamps
4. After AI replies, files are written to disk via AIFileWriter (with timestamped backup in app's filesDir/ai_backups/).
5. If Auto-rebuild after AI fix is enabled, the build is re-run automatically (up to 3 cycles).
6. Cancel button aborts a running request. 30-second watchdog logs an actionable warning if no progress arrives (likely invalid API key, dead model, or unreachable provider).
7. Provider/model badge is shown on the dialog so you always know which upstream is being hit.

---

⚙️ Settings → AI

• Provider dropdown — pick from 7+ providers
• Model / Agent box — fully editable; type any model id (e.g. deepseek/deepseek-chat-v3.1:free), tap Done, it persists
• API key dialogs — per provider
• OpenAI-compatible group — Base URL + API key + Model id, with Save endpoint button and URL normalization (/v1, /v1/chat/completions, or just domain — all work)
  • Use a preset — one-tap URL fill for Together / Groq / DeepInfra / Fireworks / Anyscale / Mistral / Perplexity / OpenAI / OpenRouter / Ollama (emulator + LAN) / vLLM
  • Auto-detect local Ollama — probes 10.0.2.2:11434, localhost:11434, LAN; on hit fills URL and lists installed models
• Custom OpenRouter model — saved name appears in the main Model/Agent box and survives app restart
• Streaming responses toggle (default ON)
• Diff preview before applying toggle (default OFF)
• Auto-rebuild after AI fix toggle (default OFF)
• Auto-switch on quota/rate-limit toggle (within-provider only by default)
• Allow cross-provider fallback toggle — default OFF. ON re-enables automatic switching from (e.g.) OpenRouter to OpenAI-compat on quota/rate-limit errors. OFF keeps you on the provider you picked and lets you fix it (add credits, change model, etc.).
• Code completion toggle
• Planning mode toggle — asks the model to emit a short plan before any code
• Export / Import settings — JSON file in Downloads/AndroidCodeStudio/, preserves all provider keys, base URLs, model ids, and toggles (schema-versioned)

💬 Chat tab extras

• Voice input — mic button next to Clear. Uses the device's speech recognizer (Bengali + English on most phones). Recognised text is appended to whatever you've already typed.
• 🖼 Image input (vision models) — image button next to the mic. Pick a photo / screenshot from the gallery; it is downscaled to ≤ 1600 px on the longest edge, JPEG-compressed (q=85, recompressed harder if still > 4 MB), and base64-attached to the next message as a multimodal image_url content part. Works with every modern vision model on OpenRouter (gpt-4o, gemini-*, claude-3.5, llama-3.2-*-vision, qwen-2-vl, …) and any OpenAI-compatible vision endpoint. A live chip below the prompt shows the attachment and lets you remove it with a tap.
• @filename.kt mentions — type @MainActivity.kt in your prompt; the IDE scans the open project, prepends the matched file(s) to the prompt as explicit context (max 8 files × 32 KB), and still passes your original intent through so the model sees both.
• Offline guard — if the device has no validated internet, the request returns immediately with a clear message instead of waiting 30 seconds.
• Sensitive-data warning — if your prompt contains what looks like a real API key / token / private-key block, a warning is shown before sending.

---

🚀 Build-speed tuning

gradle.properties is auto-tuned for new projects via the included project template + GradlePropertiesPresets:

• Gradle daemon ON
• Parallel builds ON
• Build cache ON
• Configuration cache ON
• JVM heap 4 GB + UseParallelGC
• Kotlin incremental compile ON
• AAPT2 daemon + parallel resource processing

Expected impact: clean build 4–5 min → 1–2 min, incremental rebuild 30–60 s → 5–15 s.

---

🧠 RAM / performance fixes

• Conversation history capped at 20 messages for every provider
• File-context loading capped at 40 files × 64 KB to avoid OOM on large projects
• Replaced 100 ms SharedPreferences polling loop with OnSharedPreferenceChangeListener (battery + CPU savings)
• Live progress dialog uses 80 ms coalesced UI updates for streaming (no UI jank)

---

🐞 Bug fixes (compared to upstream)

• OpenAI / Anthropic / Grok strings were copy-pasted from DeepSeek (every label said "Deepseek"). Fixed.
• 51 duplicate string IDs in Arabic strings.xml. Fixed.
• Duplicate title_build_variants resource. Fixed.
• Format string mismatches in Arabic + Chinese translations. Fixed (positional format).
• AAPT2 errors during release build. Fixed.
• jdk-compiler module had references to missing OpenJDK source files (ServerMain, SjavacServer, etc). Excluded from build.
• ClangLanguageServer stub added (file was gitignored upstream).
• OpenRouter custom model field couldn't be typed into (inputType="none"). Now editable.
• Custom OpenRouter model name didn't persist visually in the main Model/Agent box. Fixed — typed name is added to the dropdown list, persists across app restarts, and is restored from preferences.
• Chat tab Send button was off-screen on narrow phones (Export + Clear were eating row width). Fixed — Export and Clear are now icon-only, Send is anchored full-width.
• Various AI agent message-routing bugs around streaming and modification callbacks.
• Provider switch was leaving stale model ids attached (e.g. picking OpenAI-compat while an OpenRouter vendor/model slug was still active made the upstream 400). AIPreferencesFragment.handleProviderChange now persists the provider before calling setAgent, and restores the saved OpenAI-compat model id on switch.
• AIAgentManager hardcoded setProvider("gemini") on init, so auto-fix ignored the user's saved preference even though chat honoured it. It now reads Agents.getProvider(). BuildErrorAutoFixer also force-sets the saved provider before hitting the upstream, regardless of what the fallback chain selected.
• Error bubbles were inlining 500-character stack traces. They are now short ("provider / type / message"); the full trace stays in logcat.
• Model dropdown could show the same id twice when the user typed a model that was already part of the curated list — deduped.
• Preferences changes (API key, base URL, model) used to require an app restart to take effect. AIAgentManager now registers a SharedPreferences.OnSharedPreferenceChangeListener that re-initialises the active agent as soon as anything ai_agent_* / ai_provider* changes.
• Live-progress Cancel button used to leave the dialog open — user had to tap OK separately. It now auto-dismisses 1.2 s after cancellation.
• Auto-switch on errors used to silently jump to a different provider (e.g. OpenRouter → OpenAI-compat) when the upstream returned 400 about token-overflow. Now: a new ContextTooLongException distinguishes "prompt > model context window" from real account-level quota errors. On context-overflow, the agent cycles through OpenRouter's large-context models (Gemini Flash 2.0 ~1M, Gemini 1.5 ~1M, Llama 3.3 70B 128k, Llama 3.1 70B 128k, Qwen-2.5-Coder 128k, DeepSeek v3.1 64k+) on the same provider — no provider hop. Cross-provider fallback for other error types is now gated behind an explicit user toggle (default OFF).

---

📦 Original feature list

• Gradle support
• JDK 11 & JDK 17 available
• Terminal with essential packages
• Custom environment variables (for builds & terminal)
• SDK & NDK Manager (manage Android SDKs and NDKs via terminal)
• API information for classes and their members (since, removed, deprecated)
• Log reader (real-time logs for your app)
• Language servers
  • Java
  • XML
  • Kotlin
• UI Designer
  • Layout inflater
  • Resolve resource references
  • Auto-complete resource values when editing attributes
  • Drag & Drop interface
  • Visual attribute editor
  • Android Widgets
• String Translator
• Asset Studio (Drawable & Icon Maker)
• Plugin Creator (Create sub-modules or plugins inside your project)
• Git integration
• AI Agent – fully project-aware AI assistant with multi-provider support, streaming, diff preview, build-error auto-fix, editor selection actions, and project memory.

---

📲 Installation

Please install Android Code Studio from trusted sources only:

• piashmsu fork releases
• Upstream releases

• Download the APK from a trusted source.
• Follow the setup guide to install build tools.

🛠 Building from source

git clone https://github.com/piashmsu/android-code-studio.git
cd android-code-studio
git checkout devin/1777186665-ai-agent-openrouter-autofix

# Provide SDK location:
echo "sdk.dir=/path/to/your/android-sdk" > local.properties

# Debug APK (split per ABI):
./gradlew :core:app:assembleDebug

# Output:
ls core/app/build/outputs/apk/debug/
# android-code-studio-arm64-v8a-debug-*.apk
# android-code-studio-armeabi-v7a-debug-*.apk

⚠️ Limitations

• Project must use Android Gradle Plugin v7.2.0 or newer.
• SDK Manager is bundled and accessible from the terminal, but some tools (like NDK) are not built for Android.

The app is in active development. Please report issues you encounter.

🤝 Contributing

See the contributing guide.

For translations, visit the Crowdin project page.

🙏 Thanks to

• Rosemoe for the awesome CodeEditor
• Termux for the Terminal Emulator
• Bogdan Melnychuk for AndroidTreeView
• George Fraser for the Java Language Server
• FWCD for the Kotlin Language Server
• Itsaky — original developer of AndroidIDE
• Mohammed-baqer-null — maintainer of the AndroidCodeStudio upstream this fork is based on.

Thanks to all the developers who have contributed to this project.

📡 Contact

• Telegram: t.me/rv2ide
• Upstream docs: m.androidide.com

👤 Author & Credits

✨ Shorif uddin ✨

Architect of this fork — every AI feature, every bug fix, every animation on this README.

What Shorif uddin built into this fork 🤖 7+ AI providers wired in ⚡ Live build-error auto-fix 🌊 Token-by-token streaming responses 🟢 Colored `+/-` diff preview per file 🎙️ Voice input (Bengali + English) 🖼️ Image input for vision models 🧠 Sticky-provider context-overflow fallback 🛡️ Sensitive-data + offline guards 💾 Settings JSON export / import 🐞 Dozens of upstream bug fixes

Tech that powers it "Banglar coder er bondhu — phone theke real Android app banao, AI er sathe."

_{If this project helped you, drop a ⭐ on the repo — it tells Shorif uddin to keep shipping.}

---

📜 License

AndroidCodeStudio is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

AndroidCodeStudio is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with AndroidCodeStudio.  If not, see <https://www.gnu.org/licenses/>.

Any violations of the license can be reported either by opening an issue or writing a mail to us directly.