Skip to content

0.10.0

Choose a tag to compare

View all tags
@suhail-akhtar suhail-akhtar released this 05 Sep 02:59
· 24 commits to main since this release
v0.10.0
04c97e0

0.10.0

Two client-reported problems fixed at their cause, a new provider, and a store that tests can no longer reach.

Added

  • Moonshot Kimi as a provider of its own: Kimi K3 (1M context), K2.7 Code (256K, the default) and K2.6, in the CLI, the web client and the VS Code panel. Built from the platform's documentation: the chain of thought comes back as reasoning_content and is replayed as the docs require; reasoning is controlled the way each model wants (reasoning_effort on K3, a thinking switch on K2.6, nothing on K2.7 Code); cache hits are priced at the published fraction. Set MOONSHOT_API_KEY, or add it under Settings → Providers. aico -m kimi-k3 routes there on the model name alone.
  • The context window learns from use. A prompt the model accepts that is larger than an assumed window raises the figure, marked inferred on the meter and said once in the conversation.
  • A ContextWindow tool. The model can read the figure and where it came from, record one the user states, and is refused anything below what has already been seen to work — or under 8,000 tokens.
  • AICO_HOME moves the whole store. Every test and probe now runs on a scratch store seeded with a copy of your settings and removed on exit. scripts/prune-test-projects.mjs lists and removes what earlier runs left behind.
  • The skill bench in Settings → Skills, with a truer optimiser (candidates, a result cache, patience, preserve hints) — see the changelog.

Fixed

  • Cost climbing with every word on an OpenAI-compatible endpoint. Gateways that report a running usage total on every streamed chunk were being summed chunk by chunk. Usage is now reported once per request.
  • Compaction firing far too often on an assumed window. Detection now asks the instance that actually routes the request, matches ids forgivingly, and the meter is a button: click it, type 1m or 128k, done. The same control is in the VS Code panel.
  • Two context-window writes close together no longer lose one.
  • Picking a model for a provider in Settings now applies and closes the list.
  • The website header no longer breaks its links onto two lines or overflows on small screens; the homepage figures and install commands are current.

Verified

Harness 2,495 / 0 · window floor 18 / 0 · panel tunnel 81 / 0 · VS Code panel 31 / 0 in a real editor · web live 194 / 0 · web layout 52 / 0 · web unit 197 / 0.

Install

npx github:suhail-akhtar/aico#v0.10.0 serve
npm install -g github:suhail-akhtar/aico#v0.10.0

The VS Code extension aico-vscode-0.6.3.vsix is attached; install it with Extensions: Install from VSIX….

Assets 3
Loading