dsh-planner-executor

When the main model must both reason and step through tool calls, a flood of intermediate steps quickly inflates the session context, burns tokens, and forces a single model to handle both thinking and doing. dsh-planner-executor cleanly separates those two roles: the main model focuses on analysis, planning, and summarizing final results, while execution is delegated via the delegate_execution tool to an isolated sub-agent that works in its own session. No intermediate step ever enters the primary conversation; only the final report returns to the main model, maximising token savings while letting each role use the best-suited model.

The sub-agent LLM provider, execution model, per-call token budget, maximum delegation depth, persona, and tool allow/deny lists are all configurable in the dedicated Planning-and-Execution section of the Web settings panel; any field left blank falls back to the main model settings, and changes take effect on the very next delegation call. On the safety side, the sub-agent permissions are pinned to its sandbox scope, its approval policy is locked to never, and delegate_execution itself is force-denied in the sub-agent tool catalogue, so recursive delegation is structurally impossible even if someone attempts to construct the call.

Ideal for multi-model workflows where a high-reasoning model handles planning while a faster or more tool-capable model handles execution, as well as for teams that need fine-grained control over sub-agent permissions and costs. Zero build dependencies, pure JavaScript ESM, ready to use right after installation.