LLM/manbot
1
0
Fork 0
mirror of https://github.com/larchanka/manbot.git synced 2026-05-13 21:42:08 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
main
manbot/_board/AUDIT.md
larchanka 6e4d17b41a Rebrand agent to 🧬 ManBot 
- Update Telegram welcome message and user-facing strings
- Rename README, package description, and platform docs to ManBot
- Update _board tasks and instructions with ManBot branding
- Keep AI-Agent as project/internal name (package, folder structure)

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-02-17 16:22:29 +01:00

2.1 KiB
Raw Permalink Blame History

Audit: Model Orchestrator & Selection Logic

Audit Time: 2026-02-17 11:25:00

Overview

Investigation into how ManBot selects models for planning and execution, and whether the ModelRouter works correctly with the current configuration.

Findings

1. Model Selection Mechanism

  • The ModelRouter (model-router.ts) translates abstract complexity levels (small, medium, large) into concrete model names.
  • Configuration is loaded from config.json and merged with defaults and environment variables.
  • Models are selected correctly based on the following mapping:
    • Planner Agent: Uses the complexity level provided by the Orchestrator (currently hardcoded to medium in runTaskPipeline).
    • Generator Service: Uses modelClass from the node's input, falling back to medium.

2. Configuration Status

The current active mapping in config.json is:

  • Small: qwen3:0.6b
  • Medium: qwen3:1.7b
  • Large: qwen3:8b

3. Observed Issues / Observations

  • Planner Defaults: The Orchestrator hardcodes complexity: "medium" for the initial planning phase.
  • Node Granularity: The Planner can specify different models for different nodes in the DAG (e.g., large for generation, small for summarization).
  • Fallback Logic: If modelClass is missing from a node's input, the GeneratorService defaults to medium, ignoring the global plan complexity.

Actions to Perform

  1. [TASK] [P10-04] Improved Model Selection Fallback
    • Update GeneratorService or Executor to use the plan's global complexity as a fallback for modelClass instead of a hardcoded medium.
  2. [TASK] [P10-05] Configurable Planner Complexity
    • Make the initial planning complexity configurable in config.json instead of being hardcoded in orchestrator.ts.
  3. [AUDIT] Verification
    • Conduct a test run with a "large" complexity goal to verify that the qwen3:8b model is correctly invoked for critical nodes.