Context: checkpoints and the explorer

Finish real work in a 20–30K token context, with one rule: cheap models read; the main model decides.

Checkpoints

The agent chooses what stays in its own context with the checkpoint tool: keep a result whole, keep only some of its lines, or replace it with a one-line note. Everything dropped stays in the session log and comes back with recall. The agent calls it when a step is done, and the runtime asks for one when the context reaches the trigger (axon's pruner.mode: low 48K, moderate 32K, extreme 20K tokens). Asked of the main model, a checkpoint reuses the cached prompt, so it costs about one ordinary step. If the work really needs more, the agent keeps it and the next checkpoint waits for growth beyond that.

# optional; omit to let the main model decide checkpoints
pruner:
  model: some/other-model
  provider: openrouter   # defaults to the main provider if omitted

The explorer

With explorer.model set, the agent gets an explore tool. It hands a broad question to a cheaper model that may only read and search, in its own throw-away context. The explorer reports facts with quotes; every quote is checked against the file, and only verified facts reach the agent, labelled as the explorer's reading. The agent still reads raw whatever it edits.

explorer:
  model: z-ai/glm-5.3-flash
  provider: openrouter

Auditing

Both side models are tapped into the trace lognext to the main one, so you can read what a checkpoint kept and what the explorer read. Choices made with /pruner and /explorer, or in the editor, are saved to your user config.