Module tool 

Tools: the things an agent can actually do.

A tool is a name, a description, a JSON Schema, and an async function. The registry holds them; MCP servers and native Rust functions both land here as the same trait object, so the agent loop never learns the difference.

Modules

ask

Asking the user a question, as a tool.

builtin

Built-in tools. No server required — these are ordinary Rust functions.

recall

Recall: search this conversation’s full recorded history.

skill

The skill tool: level 2 of progressive disclosure.

todo

A task list the agent maintains for itself.

Structs

Capabilities

What a tool can do — the vocabulary MCP standardized (readOnly, destructive, openWorld) plus the two axes that decide whether an agent can be turned into an exfiltration tool.

CarriedState

A tool’s own state, on its way across a compaction.

ModeApprover

Answers from the configured PermissionMode without asking anyone.

Registry

ToolCtx

What a tool is allowed to touch.

ToolOutput

Enums

Decision

The decision an approver hands back for one pending call.

Constants

SPILL_FLOOR_BYTES

Floor under a result’s share of the turn budget. A wide batch must not starve every result down to a marker with no content: below this, the division stops and the total budget is allowed to overrun instead.

Traits

Approver

Gates tool calls that aren’t read-only. The CLI implements this with a terminal prompt; a headless caller can auto-allow or auto-deny.

Tool

Functions

cap_result

Cut an oversized tool result down to cap bytes, saving the full output where the model can get it back.