jcode run has no turn/step limit (--max-turns equivalent)

Summary

jcode run has no way to bound how long the agent loop runs. There is no --max-turns / --max-steps flag, and no equivalent key in config.toml. Grepping the tree at 02439b4, the only max_steps are in jcode-telemetry-core (onboarding trace) and the TUI debug bench — nothing caps the agent loop itself.

Every comparable CLI has one:

Why it matters

For unattended/batch use — CI, evaluation harnesses, anything spawning many jcode run invocations — the cap is what bounds worst-case cost and wall-clock when a task turns out to be unsolvable and the model keeps trying. Right now a single stuck run can loop until it is killed externally.

I run jcode over a benchmark of Word-editing tasks (one jcode run per task, dozens in parallel), and this is the only preset in my harness where I cannot express "give up after N turns" to the CLI itself.

Workaround, for reference

There is a workable stopgap using the existing hooks, which is what I ended up shipping:

• JCODE_HOOK_TURN_START → a script that appends one byte to a counter file per turn
• JCODE_HOOK_PRE_TOOL → a gate that reads the counter and exit 2s once it passes the cap

It works because pre_tool is the one hook that can veto, and counting turn_start keeps the unit honest (a single turn may issue several parallel tool calls, so counting calls caps early). But it only denies tools — the loop keeps running and the model has to give up on its own, so it is a budget on tool use rather than a real stop condition.

Suggestion

A --max-turns N on run (and a matching config.toml default) that ends the session with a clear stop reason. For non-interactive use it would also help if the reason were visible in --json output, so a caller can distinguish "finished" from "hit the cap" without parsing prose.

Environment: jcode 0.68.0, Linux x86_64.

Start at the implementation of jcode run and trace the agent loop through the existing JCODE_HOOK_TURN_START and JCODE_HOOK_PRE_TOOL lifecycle. Check how config.toml defaults and --json output are handled. Done means a configurable turn cap stops the session cleanly and exposes a distinct cap stop reason in non-interactive JSON output.