Case study — 04
A terminal coding agent that sends routine steps to a cheap model and hard reasoning to a large one, with a plan mode that blocks every write.
A coding agent that uses one large model for every step spends budget on file reads and test runs. Its user also needs to know what it may change before it changes anything.
Routine work — reading files, running tests, small edits — goes to a fast model. Planning and hard debugging escalate to a large one, and escalation needs explicit approval.
In plan mode every write, patch, commit, and issue or PR action is blocked, and terminal commands are limited to read-only ones.
All data is kept in ~/.dhybrid/, with no server and no telemetry.
Installable with pip, released under the MIT license, and marked Beta. Token cost is measured per session through a /tokens command.