More models to work with. Tools that resolve across your installed taps. Background results that reach the agent while it's still working. Long sessions that keep useful context and saved output within reach.
Octomind 0.50.2 is out today, September 5. Since our 0.48.0 release post on August 31, we've shipped 0.49, 0.50, and two patches. Releases have been arriving faster than the release posts. This one catches you up on what changed and where you'll notice it.
More model choice in your coding sessions
Octomind's model integrations have expanded since 0.48. This release includes support for GPT-6 Astra, Claude Fable 5.1 and Mythos 5.1, and Gemini 3.8 Flash, plus a new Meta Model API provider.
The new integrations work through the same model selection and configuration you already know. As long as you have the provider credentials and model access, switch the current session with /model, or set your preferred model in the [model] profile introduced in 0.48.
For example, with your OpenAI API key configured:
/model openai:gpt-6-astraThat changes the model for the current session and saves the selection with it. Your global default stays as configured. It's a quick way to try a newly supported model on an existing task, with the conversation and project already in place.
These integrations come through Octolib, the provider library included with Octomind. See the provider setup guide for credentials and configuration.
Use tools from across your taps
Taps package agents, skills, and tool capabilities. If you maintain a company tap alongside the standard Octomind tap, you'll want to combine their capabilities in one workflow.
0.49 improves capability resolution across connected taps. Agents can resolve tool bundles from the taps you've installed, including references that select a particular organization. Dependencies install using the scripts from the tap that owns them.
For example, a team could keep a project-read capability in its own tap and use standard memory tools alongside it. An organization-qualified reference such as myorg/project-read selects that organization's capability, provided it has a connected tap; octomind/memory-read selects the baseline one.
This makes custom tooling easier to reuse: your team's integration can live in its own repository, and the agent still brings it into the same session as the rest of its tools.
The tap guide explains how to connect a repository. The capability guide covers naming and activation.
Background results arrive while the agent is working
Suppose your agent starts a background check, then continues investigating the code. The check finishes while the agent is still making tool calls. Previously, that result could sit in the inbox until the current turn ended.
In 0.50, completed background jobs, monitor updates, and tap replies can enter the running turn before the next model request. The agent can use a result as it continues the task, instead of discovering the information only after finishing its reply.
Completion checks also handle pending asynchronous work more carefully. When background work remains, the completion gate waits, and its pending state carries forward when the result resumes the session.
Together, that means delegated work becomes part of the answer at the right time. A code review can incorporate the check it started; an investigation can continue from a specialist's findings.
Longer sessions, with less unnecessary compression
A large tool response can fill a conversation quickly. Summarizing is one way to recover space, but that takes another model call.
0.50.2 trims oversized tool results before deciding whether a paid compression call is needed. It then recalculates the context size. If trimming puts the conversation below the compression trigger, the summarization call is skipped. The full output still goes to disk, so the agent can read it when needed.
The release also corrects how unsuccessful compression attempts affect later decisions. A declined or rejected summary no longer counts as a successful compression that freed space.
Two other changes since 0.48 help preserve continuity:
- Recently recalled context survives compression for a short, bounded period. Information the agent just retrieved can remain available for the next steps, within the context budget.
- Large tool outputs saved separately now live with session data. Resuming a session can retain access to those files, instead of relying on a temporary directory that may have been cleared.
For a long debugging session, that means fewer reasons to repeat a search or regenerate a large result simply because the conversation was compressed or resumed. The compression guide explains the controls available if you want to tune that behavior.
Completion checks follow the files you actually changed
Octomind's supervisor checks the working tree as part of deciding whether the requested work is complete. That check needs to follow the session's project, including when the process itself started elsewhere.
The updated checks use the session's working directory and inspect changed file contents where possible. They can distinguish an actual edit from rewriting identical content or merely staging a file. Projects outside Git also get a bounded filesystem check.
This improves the evidence behind completion decisions, especially when an agent works through several edits in one session. It complements the background-work fixes: a task only counts as finished when both the project state and any outstanding work check out.
A clearer view of usage and request costs
For signed-in users, /usage now understands the account's allowance window and uses the period label supplied by the service. Its allowance summary includes both recorded spending and reserved spending when the service provides it, so commitments from running machines show up in the total.
The per-request /report view also gets corrected cost and timing attribution. Running totals now save at request boundaries, and multiline prompts flatten into readable report rows.
Use /usage for the account, /report for individual requests, and /info for the current session's statistics. The session command reference covers each view.
Update to Octomind 0.50.2
Today's patch also makes installation more dependable. Latest-version lookup now tries GitHub's release redirect before its API fallbacks, avoiding the unauthenticated API limit on the primary lookup path. We've also reworked the documentation with clearer setup and agent workflow guidance.
To install this release on macOS or Linux:
curl -fsSL https://octomind.run/install.sh | bash -s -- --version 0.50.2
octomind --versionSee the installation guide for other platforms and methods. The full changelog records the changes across 0.49, 0.50, and both patches.
Already available in Octomind Cloud
Octomind 0.50.2 is already available in Cloud. Open Octomind Cloud to work from the browser.
The cloud experience has grown alongside these releases: images and voice give you more ways to bring context into a session, durable artifacts keep deliverables accessible while a machine sleeps, and interactive agent cards let you answer questions through buttons and forms. Each has its own walkthrough if you missed it this week.
Pick up an existing task, choose a model, and put the new release to work.
