Skip to content

AgentiLoopGo 0.0.3

Latest

Choose a tag to compare

@github-actions github-actions released this 27 Sep 06:22
· 1 commit to main since this release

๐ŸŽ‰ Release โ€” 0.0.3

AgentiLoop 0.0.3: the TUI-hosted setup wizard signing in with a ChatGPT plan (Codex), picking gpt-6-astra, and answering the first prompt

๐ŸŽ‰ Release

AgentiLoopGo 0.0.3 is built from main and tagged v0.0.3. The theme of this release is getting connected: a first-run setup wizard (plain terminal or full-screen TUI), a new Codex provider that uses your ChatGPT Plus/Pro plan, OpenAI fixes for the newest models, and smarter context budgeting so small models work too. The Rust version, AgentiLoop CLI 0.0.3, has the same features and the same tests; every Go commit mirrors its Rust twin. The macOS binaries are signed with Developer ID and notarized.

Testers welcome โ€” see What to test below.

โœจ New

  • Setup wizard: agentiloop --setup (or just agentiloop on a machine with no credentials) asks for a provider, key and model, checks the connection, and saves to ~/.agentiloop/env, your shell profile, or the macOS Keychain. agentiloop --reset wipes it all. Works in the line REPL, in the TUI (--setup --tui) and mid-session with /setup (1666179, 12eb614).
  • Codex provider (-p codex): sign in once with codex login and AgentiLoop uses your ChatGPT Plus/Pro plan โ€” no API key. Includes token refresh, the /responses streaming API, the model list, hosted web search and the apply_patch tool (3f289f4).
  • Windows: the wizard stores the key as a user environment variable (setx) instead of a profile that a Restricted execution policy never runs; PowerShell profiles are supported on every OS (8947939, 02c27d0).
  • TUI: drag to select text and it is copied to the clipboard on release (8247b2a).
  • CI now tests on macOS, Linux and Windows, including a Windows smoke test of --setup / --reset (28430b9).

๐Ÿ”ง Changes

  • OpenAI: api.openai.com requests go through /v1/responses (the newest models reject tools + reasoning on chat/completions) and send max_completion_tokens instead of the deprecated max_tokens (3f42cf1, 1356d5d).
  • Model limits come from the provider's catalog (Anthropic GET /v1/models/{id}, OpenAI's published limits, OpenAI-compatible servers) instead of a hard-coded table (66351e4, 1fe111e, 341eaa8).
  • Context budgeting: the completion limit is computed after reserving room for the system prompt, messages and tool definitions, recalculated every turn and applied to compaction โ€” so "hello" no longer overflows an 8K model (f27e9d2).
  • Default max_tokens raised to 32k; truncated tool calls are reported clearly (fb83e1e).
  • Shorter default system prompt; identity questions are answered as AgentiLoop (4265025, 7c2eb35).
  • READMEs (8 languages) rewritten around the wizard, with a Pong demo (d0a69e7, d06d3e2).

๐Ÿ“ฆ Downloads

Platform File
macOS Apple Silicon agentiloop-macos-arm64.tar.gz (signed + notarized)
macOS Intel agentiloop-macos-x86_64.tar.gz (signed + notarized)
Linux x86_64 agentiloop-linux-x86_64.tar.gz (glibc 2.35+)
Linux ARM64 agentiloop-linux-arm64.tar.gz (glibc 2.35+)
Windows x86_64 agentiloop-windows-x86_64.zip

Unpack the download and put agentiloop (agentiloop.exe on Windows) on your PATH, then run agentiloop โ€” the wizard takes it from there.

๐Ÿงช What to test (testers welcome!)

  • agentiloop --version prints 0.0.3.
  • On a clean machine (or after agentiloop --reset), run agentiloop โ€” the wizard should start by itself. Try it plain and with --setup --tui.
  • Each provider: Claude (API key), OpenAI (API key; try a gpt-5/gpt-6 model โ€” it should use /v1/responses), Ollama / LM Studio (no key), oMLX, and Codex (codex login first, then pick 5).
  • A small local model (8K context): say hello and run a tool โ€” no "context window exceeded" errors.
  • Windows: after the wizard, open a new PowerShell and check the key is in $env: โ€” no profile edits needed. SmartScreen may warn because the Windows binary is not code-signed yet.
  • TUI: drag over some output and paste it somewhere.
  • Report problems with your OS, provider/model and the command you ran. Do not include API keys.

๐ŸŽฎ Pretend war games with Astra

A session in the new -p codex provider, playing a make-believe war-games simulation with gpt-6-astra in the TUI. Astra runs the turns as a labelled fictional simulation, refuses the one order that crosses the line (turn 15), and keeps the story going through revived populations, a magic potion that rebuilds every building, and a negotiated democratic transition. Status bar shows the codex provider, model, session id and live tok/s.

Pretend war games with Astra in the AgentiLoop TUI

๐Ÿ”ข Version

Full Changelog: v0.0.2...v0.0.3