What happened
Anomalyco released OpenCode v2.2.2 on October 29, 2026. The headline change is that experimental_image_generate is enabled by default when OpenCode detects a connected provider with image-model support. Previously, users would have needed to enable that experimental capability separately; in this release, the availability condition is the provider connection and its image capability.
The release does not say that OpenCode chooses a universal image model or that every provider will work. Instead, it states that the preferred image model can be set through experimental.image_generate.model. That makes the feature a configurable tool exposed to an agent, rather than a separate image product or a claim that the runtime itself supplies image inference. The release labels the capability experimental.
Version 2.2.2 also changes several workflow details around sessions and worktrees. On the new-session screen, clicking the current Git branch opens the worktree picker for that project directly. The release unifies timeline quotes, terminal quotes, and browser annotations around a compact editor with an optional comment, while keeping source highlights attached to local drafts. It records the start time of each session’s latest execution.
Other changes are primarily fixes and interface improvements. The release says media are preserved when sessions are forked or imported, and image-generation notices are associated with the correct background job. In the terminal interface, usage-limit errors for Go and ChatGPT are typed and displayed correctly, with reset-time and account-page information improved. The desktop interface receives changes to tabs, message actions, task display, dialogs, and attachment handling. SDK scheduling gains if and until inputs, along with schedule check results and output fields.
Taken together, these details describe a coding-agent runtime making a new modality available inside existing task sessions while refining the surrounding controls: project worktree selection, annotations, attachments, execution timing, background jobs, and provider-limit feedback. The release notes do not describe a new price, a new hosted tier, or a change to the underlying providers’ own image-model terms.
What the archive adds
We have not covered this before; it opens a new thread.
Trend signal
This release is consistent with the harness architecture outlives model choice pattern. OpenCode’s release makes the runtime responsible for discovering a provider capability, exposing it as an agent tool, routing a preferred-model setting, and reporting the resulting job in its session UI. The release alone does not establish adoption or durability, but it illustrates why the harness can matter independently of any one model: the same workflow can gain a modality as provider capabilities vary.
Why it matters
For a team building with an agentic coding environment, a default-on tool changes the practical review boundary. A task that previously involved text, files, terminal output, and browser annotations can now also invoke image generation when the provider supports it. That may be useful for creating visual assets or intermediate artifacts inside a session, but it also means teams should know which connected providers expose image models and which model OpenCode is configured to prefer.
The operational value is not simply that an agent can make images. It is that the capability sits inside a runtime already managing sessions, worktrees, attachments, background jobs, and provider feedback. Teams can therefore treat image generation as one more permissioned, observable action in an engineering workflow rather than as an ad hoc handoff to a separate interface.
Limits and uncertainty
The release notes establish that the experimental tool is enabled by default for image-capable connected providers, but they do not enumerate supported providers, supported image models, quality, latency, output formats, retention, or costs. They do not say how OpenCode determines that a provider is image-capable, whether administrators can disable the tool globally, or what authorization and audit controls apply to generated assets.
The notes also do not provide usage data, benchmarks, or evidence that the feature improves coding outcomes. “Enabled by default” should not be read as a guarantee that every OpenCode installation will have access: a compatible connected provider remains a stated condition. Teams should verify their own provider configuration and billing before assuming the tool is available or inexpensive.
Skill to build now
Audit the action surface of one OpenCode environment this week. Check which provider is connected, whether it exposes image models, and what value is set for experimental.image_generate.model. Then run a small non-production session that requests an image artifact and inspect the resulting job notice, attachment behavior, and provider usage feedback. Document who may change that preferred-model setting and whether image creation belongs in the team’s normal review flow.
Reader takeaway
OpenCode v2.2.2 turns experimental image generation into a default agent tool when the connected provider supports it. The important decision for builders is not whether images are novel, but whether their runtime configuration, provider choice, cost expectations, and review process are ready for an agent to invoke that capability inside ordinary sessions.
Open question
Which providers and image models OpenCode will support in practice, and what controls operators will have over cost and authorization, are not established by this release.