Repository navigation
Releases
Belmont Tools 0.4.10 (Windows)
Belmont Tools 0.4.10 — Windows
The current release. Two ways to install:
Belmont-Tools-Setup.exe(149 MB) — NSIS installer. Recommended: installs permanently, appears in the Start Menu, and comes with an uninstaller.Belmont-Tools-Portable.exe(149 MB) — portable build, a single file with no installation. Good for trying it out first, or carrying on a USB stick.
The filenames carry no version number on purpose, so that
/releases/latest/download/Belmont-Tools-Setup.exe
always points at the newest build and never has to be updated anywhere.
What changed since 0.4.9
Two bug fixes. Both were reported as one symptom — "DeepSeek always errors once the context reaches 20-30%" — and neither turned out to have anything to do with context size. A 2.3 MB request carrying 517k tokens was accepted without complaint; the size theory was measured and discarded.
Tool results that drifted away from their call. Every model API requires each tool call to be answered by a result sitting immediately behind it. History recorded through the Claude Code path does not guarantee that: subagent results often arrive after Claude has already moved on, so the stored order becomes call, call, call → result, result → another assistant message → the straggler. DeepSeek rejects that outright, and because the flaw is saved to disk the rejection is permanent — a project simply stops working with DeepSeek, Kimi, GLM, or the Claude API even though there is plenty of context left. Five of the seven projects on the development machine were already in this state.
History is now repaired on its way to the provider: results are pulled from wherever they ended up and glued behind their call, calls whose result was never stored get a placeholder, and consecutive messages with the same role are merged. Nothing on disk is rewritten, so the chat still shows what actually happened, and projects that are already broken start working again with no data migration.
Reasoning that was never sent back. Found because the first fix worked: histories that now passed the tool check were refused for a different reason — DeepSeek's thinking models require the previous turn's reasoning_content to be returned whenever a request ends with tool results. Mapped against the live API: an empty string satisfies it, and only the thinking models care — deepseek-flash and deepseek-v4-pro fail, deepseek-v4-flash-vision-exp does not. In other words, every tool call with those two models failed at the first step, regardless of history length.
Reasoning is now kept alongside the message it belongs to and passed back. The requirement is not keyed to model names — model lists change faster than this app ships, and third-party gateways resell the same models under different names — so the app learns it from the rejection itself and remembers it per endpoint and model. That costs one extra round trip, once per model per app session.
One side effect worth knowing: session files grow slightly, because reasoning is now stored with the history.
Requirements: Windows 10/11 64-bit, plus an API key from one of the providers (or a Claude Code login). Node.js is only needed if you want to use MCP servers over stdio.
Settings and history live in %APPDATA%/belmont-tools/ — installing 0.4.10 over an earlier version erases nothing.
Windows SmartScreen will warn about an unknown publisher: these builds are not code-signed.