preloader
blog post

Calliope CLI v2.3.0: Long-Running Agent Sessions, Persistent Logs, and Better Local Model Support

author image

Calliope CLI v2.3.0 is live. This release is about making autonomous terminal work reliable enough for real coding sessions: loops can now run until the job is actually done, autonomous runs leave behind a persistent session log, resumed sessions stay attached to the right project, and swarm or council workflows are stricter about working in the correct directory. On top of that, v2.3.0 adds better OpenAI-compatible local server support and a set of runtime fixes across memory, MCP, and API server behavior.

Install from npm | Full Release Notes

npm install -g @calliopelabs/cli@latest

What is Calliope CLI?

Calliope CLI is an open source, multi-model AI coding agent for the terminal. It works with hosted models, local models, project memory, tool execution, and autonomous loops, without forcing you into a browser tab or a proprietary IDE.

If you want AI that can read your repo, work through a task, run commands, remember project context, and stay in the flow of the terminal, this is what Calliope CLI is built for.


What is new in v2.3.0

Long-Running Loops That Do Not Stop Early

The biggest change in v2.3.0 is loop control. The default loop behavior now matches how people actually use coding agents during real work: keep iterating until the task is complete, and stop only when the goal is met or you explicitly interrupt it with /breakloop.

That makes a difference for longer sessions like refactors, bug hunts, dependency cleanup, or “keep going until the tests pass” work. Instead of treating autonomy as a short burst with an arbitrary ceiling, Calliope CLI now treats it like a proper long-running session.

Persistent Session Logs for Autonomous Work

Every serious autonomous workflow needs a record of what happened. v2.3.0 adds a session-persistent iteration ledger that tracks agent, loop, swarm, council, and workflow runs across the life of a session.

That log survives resume, save, and fork operations. If a run is interrupted, the session can recover it cleanly rather than pretending nothing happened. There is also a new /log command surface for checking summaries, tails, failures, or resetting the log when you want a clean slate.

For teams who want to keep long histories, the default is unlimited. If you want to cap retention, sessionLogLimit is now a configurable setting.

Workflow Execution That Stays in the Right Repo

This release tightens up session and workflow isolation in a few important places. Resumed sessions now properly reactivate the saved session context instead of drifting back to whatever directory the shell happens to be in. Swarm and council runs keep their launch context, and background task cancellation now behaves the way users expect.

In practice, that means fewer “why is this agent suddenly operating in the wrong project?” moments. If you are running autonomous work across long sessions or switching between saved threads, this is one of the most important upgrades in the release.

Better OpenAI-Compatible Local Server Support

v2.3.0 also improves support for OpenAI-compatible servers running locally or inside your own infrastructure. Calliope CLI now includes compatibility shims for LM Studio, AnythingLLM, vLLM, Jan, and LocalAI.

The point is not just checkbox compatibility. Different servers disagree on details like tool calling, max token defaults, or message formatting. Calliope now detects common server patterns and adjusts requests automatically where needed, which makes local and self-hosted model setups much less brittle.

Reliability Fixes Across Memory, MCP, and API Mode

This is also a cleanup release, and a useful one. A number of issues that mostly showed up in long sessions or shipped builds have been fixed:

  • Trust checks now work correctly in the built ESM package rather than silently falling back
  • --serve now behaves like a real server mode and no longer exits into the headless path
  • The WebSocket API path now works correctly in the published build
  • Project memory now survives prompt resets like /clear, /persona, /companion, /pack, and /intensity
  • MCP stdio environments now support ${VAR} and ${VAR:-default} expansion
  • Auto-compression now has better thresholds and failure handling for long contexts
  • macOS sleep prevention is now managed as a proper guard rather than a fire-and-forget helper

None of those fixes is flashy on its own. Together, they make the CLI much more dependable when you leave it running on real engineering work.


Why This Release Matters

There are a lot of AI coding tools that look impressive for five minutes and fall apart when you ask them to stay focused for an hour. v2.3.0 is a step in the other direction.

This release is about endurance: long loops, recoverable sessions, durable logs, better workflow isolation, and fewer edge-case failures when you move from demos to actual codebases. If you want a terminal agent that can keep working without losing the plot, this is the strongest version of Calliope CLI yet.


Install or Update

Install Calliope CLI from npm

npm install -g @calliopelabs/cli@latest
calliope

Open source. Multi-model. BYOK. Built for real terminal work.

Related Articles