KodaCode 0.8.0 adds a safe chat mode, conversation branching and image uploads

KodaCode has released version 0.8.0 and substantially reworked how the interface behaves. The main focus is a more predictable UX: chat now follows agent logic but works only with read-only tools, and multimodal models can receive image uploads directly in the conversation.

Safe chat mode

The biggest change in the release is a new principle behind chat mode. It now mirrors the logic of agent mode but has no access to tools that can change files or project state. That makes exploring a codebase noticeably safer: you can look at the structure, read files, and study dependencies and code behavior without the risk of triggering an edit in the wrong place. For teams that need to keep analysis separate from intervention, this mode removes needless tension and makes the product's behavior clearer from the first launch.

At the same time, Edit mode has been removed. KodaCode's developers give a simple explanation: agent mode now fully covers its use cases. In practice, that means fewer modes, less confusion in the interface and less duplicated functionality. When a user has one main way to make changes and one safe way to explore a project, the barrier to entry drops. This is not the flashiest kind of update, but decisions like this usually have the biggest effect on the day-to-day experience.

Branching and history

The second major part of the release is conversation branching. If the agent goes off track, you no longer need to rewrite the request from scratch or scroll through a long session trying to find a good point to return to. You can now go back to any tool call in the history, continue from there and create a new branch without touching the original scenario. This is especially useful when you need to test several approaches in a row: one for analysis, another for a fix, a third for an alternative hypothesis.

For developers and teams, this brings several practical benefits:

• it is easier to roll back to a good step without losing context;
• it is more convenient to compare the agent's alternative actions;
• it is easier to analyze errors and repeat successful scenarios;
• it is safer to experiment with long chains of tools.

Sessions can also now be exported to JSON. This looks like a minor utility feature, but it has many practical uses: storing history, analyzing failed runs, passing cases to colleagues, and preparing reproducible examples for support or an internal knowledge base. In other words, KodaCode makes working with the agent not only interactive but also more portable. When a result can be saved in a structured form, it is easier to work with outside a single chat window.

Images in chat

A separate update concerns multimodal models. For gemini-3-flash-preview, gpt-5.3-codex, kimi-k2.5 and gpt-5.4, chat now lets you attach images. This moves some tasks away from "describe the picture in words" toward a more natural flow: the user simply shows a screenshot, diagram or mockup and immediately gets the model's answer in the right context. For development tools this is especially useful, because many problems are explained faster visually than through long text.

The new feature opens up several obvious use cases:

• analyzing screenshots of errors and UI bugs;
• reviewing mockups and interface states;
• reading diagrams, schematics and fragments of documentation;
• discussing the model's output based on an image within a single session.

It also matters that support is added not abstractly "for everyone" but for a specific set of models. That points to a more careful product approach: multimodality is enabled where it is actually supported, rather than hiding limitations behind a uniform interface. For users this is more useful than vague promises, because it is immediately clear which models can handle visual workflows without surprises.

What this means

KodaCode 0.8.0 is not about flashy demos but about control over the workflow. A safe chat mode, session branching and multimodal attachments reduce accidental errors and make agent work more predictable. For an AI development tool, this matters more than an extra "wow": users need not only a smart agent but also a clear environment where they can explore calmly, experiment and step back.