tlongwell-block
tlongwell-block
The subagent guidance in the system prompt is redundant with comprehensive instructions embedded directly in the tool descriptions themselves. The `dynamic_task__create_task` tool already explains the isolation context, extension handling, and...
Implements ephemeral context injection at the agent boundary to automatically include timestamp and TODO content in LLM context without modifying conversation history. ## Problem Previously, accessing TODO content required explicit...
Replace UUID task IDs with session IDs to reduce LLM transcription errors. ## The Problem LLMs can mistype UUIDs when referencing tasks: - `a1b2c3d4-e5f6-7890-abcd-ef1234567890` → easy to corrupt - 36...
Some LLMs quote numeric arguments when calling MCP tools (e.g., sending `"42"` instead of `42`), which can cause tool execution failures. This change adds schema-aware coercion that automatically converts string...
Replaces subprocess-based subrecipe execution with in-process transformation, unifying the execution path between subrecipes and subagents while bringing full Recipe feature support to subagents. ### What Changed - Subrecipes now execute...
## Problem The `TaskParameter` struct in `dynamic_task_tools.rs` is essentially a duplicate of the `Recipe` struct. We manually convert from one to the other, which is unnecessary and error-prone. ## Current...
fix: Added "Merged consecutive assistant messages" to the acceptable issues for moim injection check
This happens every turn and is expected