Clarify prompt and tool source boundaries

Co-authored-by: Tura AI <info@turaai.net>
This commit is contained in:
Yohjisakamoto 2026-07-16 18:36:48 +02:00
parent 9b51eb3942
commit 1807883153

View File

@ -6,9 +6,10 @@ Release: https://github.com/Tura-AI/tura/releases/tag/v0.1.33
Prompt architecture: https://github.com/Tura-AI/tura/blob/v0.1.33/docs/core/prompt-style.md
Tura does not use one static mega-prompt. The runtime composes several system
messages from the selected agent, frontend, persona, task type, tool catalog,
session state, and context pressure. Its first runtime-owned identity message is
generated from this template:
messages from the selected agent, frontend, persona, task type, session state,
and context pressure. It supplies the available tool catalog separately as
runtime capabilities. Its first runtime-owned identity message is generated
from this template:
You are {agent}, an agent. Current user: {user}. Runtime model: {model}. LLM
provider: {provider}. Active context limit before compaction: {context_limit}
@ -30,8 +31,8 @@ Agent behavior
- Direct text-only:
https://github.com/Tura-AI/tura/blob/v0.1.33/agents/src/direct-text-only/prompt.md
Tool instructions
-----------------
Tool and capability sources
---------------------------
- apply_patch, bash, planning, shell_command, task_status, and zsh:
https://github.com/Tura-AI/tura/tree/v0.1.33/crates/tools/src/commands