0.4.0 - Release

Add durable persistence and typed IPC contracts

This change introduces debounced persistent stores and task queues with crash recovery for the agent, Electron main process, and Roblox Studio host. It adds validation at RPC/IPC boundaries using Pydantic and Zod, refactors Electron main-process logic into service and IPC modules, and exposes typed bridge contracts through preload. The update also stabilizes public payload contracts for runtime config, context snapshots, CoT steps, and tool results, improving durability and boundary safety across the app.

0.3.1 - Release

Handle Vulkan-safe KV and auto-install PyYAML

Switch Vulkan/HIP loads to F16 KV + disable flash-attn by default to avoid known AMD garbled-output bug; allow overrides via MANGO_VULKAN_ALLOW_QKV and MANGO_VULKAN_FLASH_ATTN. Add backend-aware GGUF loader and pass backend through ModelRunner with informative logs. Add automatic PyYAML import/auto-install helper for config load/save and ensure installer explicitly installs PyYAML; update installer smoke-test to import yaml. Add tests for Vulkan/HIP behavior and update docs. Affected: runtime/README.md, runtime/python/mango_runtime/{config.py,gguf_loader.py,model_runner.py,tests/test_config.py}, scripts/build-installer.ps1.

0.3.0 - Release

SWE-bench shuffle + agent patch-mode/action-loop

Add SWE-bench shuffle support and CLI args (count, --shuffle, --shuffle-status, seed/reset), plus persisted shuffle state and helpers. Improve Docker daemon readiness checks to skip/skip-with-message when Docker is unavailable. Substantially enhance Agent behavior for existing-repo (patch) runs: introduce patch_repo_mode, tuned action-loop redirects, edit/write preferences, fail-open plan gate, lookup/read/read-loop handling, closest-edit snippet helper, and many related guards. Update prompts to reflect patch-mode rules. Add/adjust tests and baseline SWE-bench reports. Also expose pythonPackagePaths and set PYTHONPATH in Electron sidecar; bump Electron app version and update datasets

0.2.1 - Release

Add protocol-aware plan and ask modes

Introduce mode-specific agent flows for plan, ask, debug, and refactor, plus a read-only codebase research tool and slash-command UI. This changes the prompt system to prefer agent_v2 and full protocol prompt files, disables mutating tools in research-only modes, and adds model load/unload handling to the Electron sidecar. It also improves codebase lookup formatting, strengthens tool parsing for research-related actions, and adds regression tests for prompt composition, mode gating, and agent event lifecycle.

0.2.0

Fix CLI truncation and gap editing

This change adds CLI-specific write-token caps and feedback so console projects do not dump a full script into a single write_file. After a partial implementation is written, the agent switches from full rewrites to focused read/edit gap-filling, and updates the related tests and feedback prompts to enforce the safer flow.

b0.1.0

Mango 0.1.0

First public Windows installer for Mango — local coding agent for small GGUF models.

Install

1. Download Mango-Setup-0.1.0.exe
2. Install and open Mango
3. In Settings, set your local .gguf model path
4. Open a workspace and start a session

What's included

• Desktop UI (Electron)
• Bundled portable Python sidecar (no system Python required)
• Auto-update support via GitHub Releases (latest.yml)

Notes

• Windows x64 only
• Model weights are not included — use your own GGUF
• First run: pick a GGUF under Settings before chatting
• SmartScreen may warn (unsigned build) — “More info” → Run anyway

Auto-updates

Later releases: put the new Mango-Setup-*.exe, matching .blockmap, and latest.yml on the release. Installed apps check Help → Check for Updates.