Code Mode · S-Code Docs

Code Mode lets the model use one JavaScript program to read several files, search a workspace, and return only the useful parts of the results. Each tool still appears in the conversation, with its own status under a Code Mode parent. A caught tool error remains visible even when the program succeeds.

Enable it

Code Mode is optional and disabled by default. Add this to your S-Code configuration and restart the service:

[model]
code_mode = true

The model receives an additional execute tool in coding, planning and review sessions. Direct tools stay available. Code Mode works with providers that can call ordinary JSON tools; it does not require a provider-specific code execution API. Whether a particular model uses it effectively still needs evaluation.

What the model sends

The execute arguments contain a code string. Its body is JavaScript inside an async function. For example:

const files = await Promise.all([
  tools.read_file({ path: "src/main.rs", max_bytes: 12000 }),
  tools.read_file({ path: "Cargo.toml", max_bytes: 12000 }),
]);
text(files);

The available functions are tools.read_file, tools.list_files, tools.search_text, tools.git_status and tools.git_diff. They use the same argument objects and result shapes as the direct tools. text(value) selects what returns to the model; otherwise raw child results stay out of the model's conversation. Rejected tools reject their promises. Always await every call.

Edits, commands, MCP tools and operations requiring approval use direct tool calls. Code Mode cannot suspend for approval. It returns a direct-call hint and creates no approval request. The host checks policy before hooks and again after hooks modify arguments; normal sensitive-path checks still apply.

Execution limits

Each program runs in a fresh QuickJS process with a 64 MiB JavaScript heap limit, a 512 KiB JavaScript stack limit, a 30-second deadline, and no filesystem, network, environment, process, module-loader or timer bindings. Source and selected output are each limited to 32 KiB. Tool arguments are limited to 64 KiB and each result transfer to 1 MiB. At most four child tools run concurrently; there are at most 32 child calls per turn and four active workers per daemon. These heap limits do not represent a total operating-system process RSS cap.

Stopping a turn cancels the program and prevents further child calls. Blocking host Git reads may finish under their existing timeout; their late results are discarded. Installed hooks continue to be trusted host code under their existing execution rules. A daemon restart marks interrupted programs, children and associated active turns cancelled. Scripts are never automatically replayed.

Tool visibility

The host assigns each program and child a stable tool ID. It persists the child before validation, hooks or execution, then publishes lifecycle events. The parent ID is stored with the child, so a history page containing only a child still identifies it as a Code Mode call. Events contain tool names and redacted display details, rather than file contents or the complete script output.

This follows the parent/child tracing approach used in OpenAI programmatic tool calling and Anthropic programmatic tool calling. OpenCode's implementation also reports nested tool start/end activity from its host bridge. S-Code keeps these lifecycle records in its normal tool history so live UI and refreshed history share the same identities.

Code Mode can reduce model round trips and returned context for batch reads. It is not inherently faster for a single tool, and this release does not claim a measured quality or token advantage across models.