Tether gives small-model agents a brake on repeated commands

Catch the same action before running it again

Tether 0.0.3 targets two familiar small-agent problems: repeated tool calls and large outputs consuming the active prompt. The October 17 release adds a check for consecutive identical command-and-input pairs. Instead of executing the repeated call, the harness returns a loop-detected recovery nudge. That is a concrete guard against one repetition pattern, not a guarantee that an agent cannot get stuck in a more complex loop.

Tether is a standalone, headless Python harness and SDK. It is not a new model or an extension installed into Claude Code or Codex. Its design routes tools through one act(cmd, stdin) dispatcher, loads capabilities by explicit opt-in and uses a required completion command. Developers can add Markdown skills and custom tools rather than presenting every possible action up front.

The other notable change compresses active observations longer than 2,400 characters into their first 900 and last 900 characters, with a pointer for retrieving more. Full outputs remain in SQLite with FTS5 search. A long command log can therefore leave its summary and final error visible while the middle is retrieved when needed. The tradeoff is real: information omitted from the active view is not automatically being considered by the model.

Install the harness, then choose the model

The MIT-licensed project requires Python 3.12 or newer. Its package name is tether-sdk, while the installed command is tether; the documented routes include pip install tether-sdk and uv tool install tether-sdk. Readers reproducing this release can pin version 0.0.3.

The tagged guide documents an Ollama adapter, an OpenAI-compatible adapter that can route to providers such as DeepSeek, and a native Anthropic adapter. Local use requires a separately running model service and suitable hardware. Hosted use requires the selected provider’s credentials and carries that provider’s charges; installing Tether does not supply model access.

Version 0.0.3 also moves pinned memory to the tail of the assembled context to preserve earlier cache prefixes, adds Anthropic cache breakpoints and reuses HTTP and SQLite connections. These are documented implementation changes, not independently measured savings across every workload.

Permissions depend on enabled tools. The README describes Linux shell isolation using Bubblewrap or unshare when available, with an allowlist fallback when unavailable. Those are project claims rather than an audited containment guarantee. For developers building a small-model agent, the immediate reason to examine this release is its explicit handling of output size and repeated actions; AITrending has not run the harness or reproduced its promotional benchmarks.