vrr/agent-zero
2
0
Fork 0
mirror of https://github.com/agent0ai/agent-zero.git synced 2026-10-25 23:59:43 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 2
agent-zero/extensions/python/agent_init/AGENTS.md
linkliti fa150f15b9 feat(agent-init): seed subordinate agents with the greeting few-shot example 
The synthetic Hello! user turn was limited to the main agent, so subordinate agents started their history without the few-shot example of the response tool-call JSON. Remove the agent-number guard so every agent receives the seed turn, which the initial_seed metadata marker already keeps out of chat naming for delegated children.

Validation: pytest tests/test_initial_message.py plugins/_chat_naming/tests/test_chat_naming.py (20 passed) on the framework runtime.
2026-10-13 18:06:47 +03:00

2.4 KiB
Raw Permalink Blame History

Agent Init Extensions DOX

Purpose

  • Own backend extensions that run when an agent context initializes.

Ownership

  • Ordered Python files own initial UI message setup and profile settings load behavior.

Local Contracts

  • Keep initialization idempotent for contexts that may be restored or reloaded.
  • Preserve ordering between initial message creation and profile settings loading.
  • _15_load_profile_settings.py merges settings from low to high priority: plugin-wide files, bundled profile, plugin profiles, user profile, project-wide file, then project profile. Each plugin layer applies bundled roots before user roots with a stable root-priority sort, preserving same-root discovery order and cached lookup lists. Global framework and usr/settings.json files remain excluded; malformed layers log an error and leave other layers usable. Profile and MCP configuration inherit when not explicitly overridden.
  • Keep a placeholder user turn (fw.initial_user_message.md) ahead of the AI greeting (fw.initial_message.md) so output_langchain never pops the greeting as a leading AIMessage; do not remove either prompt without replacing the turn-order guarantee.
  • Minify the initial AI message JSON before storage; preserve raw text as the displayed greeting when JSON parsing fails.
  • _10_initial_message.py passes LLMResult.non_llm() to hist_add_ai_response so the Responses-API state seam runs uniformly; the sentinel carries no response_id and marks mode=""/state="off" so stored metadata does not claim a Responses-API turn.
  • The synthetic Hello! user turn has no UI log. The assistant greeting is displayed as a response log with the same ID as its history message, preserving the log/history link for branching. Keep the user-first seed without duplicating the visible greeting on reload.
  • Every agent receives the seed turn, including subordinate agents, where it serves as a few-shot example of the response tool-call JSON; it is marked with initial_seed metadata so consumers such as chat naming can exclude the synthetic turn from user-derived processing.

Work Guidance

  • Coordinate changes with profile loading, settings resolution, and startup smoke checks.

Verification

  • Smoke-test new chat/context initialization after changes.
  • Run tests/test_plugin_asset_precedence.py and tests/test_subagent_profiles.py for profile-settings changes.

Child DOX Index

No child DOX files.