Code Mode

Instead of one tool call per model turn, the agent writes one small script that calls many tools and returns a value. Loops, filtering, joining results and any multi-step job whose intermediate data need not reach the model run in one round trip.

The agent sees it as the code_mode_run tool; the gateway exposes the same thing as POST /api/code-mode/run (kurahi api runCodeMode).

What the script can do

const jobs = await tools.listJobs({ status: "running" });
return jobs.jobs.map((j) => j.id);

• The script is the body of an async function: use await and return.
• Every tool call is await tools.<name>({...}), resolving to { text, details? }.
• JavaScript or TypeScript (language: "js" | "ts"; default: detect).
• The tool list the script sees is the calling agent’s own filtered list (policy, capability and hooks already applied) minus code_mode_run itself: core tools, plugin tools and mcp:<server>:<tool>. allowTools (globs such as mcp:docs:*) narrows it further for one run.
• A .d.ts generated from those tools’ schemas is in the tool description, so the model knows the signatures.

Sandbox and limits

The script runs in a worker thread with no filesystem, network or process access of its own; the only way out is a tool call, and every tool call is audited under the run id.

Limit	Default
wall clock (timeoutMs)	60 s (1 s to 600 s per run)
tool calls per script	200
heap (old / young generation)	128 MB / 32 MB
stack	4 MB
result size	1 MB
tool result size	4 MB
log output	64 KB

Each limit can be lowered per agent; raising one is capped by a built-in ceiling.

Outcome

ok: true with value, or ok: false with error.code: invalid_script, script_error, timeout, memory, tool_call_limit, result_too_large, worker_error.

Access

The HTTP route accepts the gateway token or a paired device credential with at least the attach scope. A backend with tools: none answers 403: no tools, no script.