Multiagents

Adaptive multi-agent terminal system for:

• knowledge lookup
• document summary
• tabular data analysis
• code-oriented execution

The runtime is designed as a dynamic loop, not a fixed keyword router.

Quick Start

python3 terminal_chat.py --workspace .

--workspace . means all file search/read/write/tool execution is scoped to the current directory.

Core Runtime Loop

The orchestrator follows:

1. Goal Understanding
2. Plan
3. Tool Call
4. Observation
5. Reflection
6. Recovery/Replan (if needed)
7. Continue or Finish

Main implementation: DynamicLoopOrchestrator in multi_agent_system.py.

New Reliability Logic (Reflect + Recovery + Replan)

The system now includes explicit failure handling instead of stopping after repetitive errors.

• Failure classification (missing_input, parse_error, timeout, policy_block, etc.)
• Recovery step synthesis based on failure category
• Plan injection (insert recovery steps into current plan and continue)
• Reflection logging in results (reflections) and event stream (reflect phase)
• Clarify only after recovery attempts are exhausted

Key methods:

• _classify_tool_failure(...)
• _build_recovery_steps(...)
• _inject_recovery_steps(...)
• _reflect_and_recover(...)

These methods are in multi_agent_system.py under DynamicLoopOrchestrator.

Rule-Based Fast Path (Deterministic Intents)

For high-frequency lookup tasks (for example JSON key search), the orchestrator uses a deterministic pre-plan before LLM planning.

Current rule:

• detect token/key lookup intent
• run:
  • list_workspace_files for *.json
  • search_workspace_text for the target token

This avoids fragile repeated trial calls and improves consistency.

Tooling Highlights

• read_document_file: parse .docx/.pdf/.txt/.md
• read_spreadsheet_preview / profile_tabular_columns: CSV/XLSX preview and profiling
• analyze_tabular_with_python: writes and executes a temporary Python analyzer script, then returns structured results
• read_code_file / read_code_span / replace_text_in_file / run_shell_command: code-task workflow

Terminal UX

terminal_chat.py has English-first startup/help/fallback messages and renders:

• execution summary
• observations
• reflections
• event phases

Example Queries

• do you find this key d5bbc8180dba11ecb1e81171463288e9 in the json file
• analyze this file ./sample_data.xlsx
• summarize /absolute/path/to/file.pdf
• check multi_agent_system.py and run python3 -m py_compile multi_agent_system.py

Notes

• This project executes commands locally in the workspace (not containerized by default).
• Safety checks are enforced for risky shell patterns and policy-blocked tool calls.
• High-risk actions can require explicit approval (approved=true depending on tool permission policy).