agent-tools

TauriTavern Agent built-in tools for workspace and skill management are used by following structured tool contracts that specify canonical naming, parameter schemas, and visibility rules to ensure compliant interaction. The reference covers tool calls like skill.list, search, and read.

The correct parameter schema for workspace tool calling in TauriTavern requires matching the documented input and output contracts for list, search, read, write, apply_patch, and commit operations. Required and optional parameters must be passed exactly as specified to avoid runtime mistakes.

Recoverable versus fatal tool errors during agent execution are handled by applying specific error semantics that define whether a run continues or stops. The operational constraints distinguish recoverable errors, which allow retry, from fatal errors that terminate the constrained agent run.

Read budgets and max limits for worldinfo and skill search define the operational constraints for tool calls, restricting how much data can be retrieved per run. These limits apply to worldinfo.read_activated and skill.search to enforce deterministic input and output boundaries.

The TauriTavern agent uses both canonical tool names and model-facing aliases, which are documented in the tool catalog. Understanding these visibility rules and naming conventions is necessary to ensure correct tool behavior and profile permissions during an agent run.

Tool calls fail when writing to workspace files in a constrained agent run due to exceeding read budgets, violating read and write boundaries, or passing incorrect parameter schemas. Following the documented tool contracts and operational constraints prevents these runtime mistakes.