CHER-1.0 Build Roadmap

Executive Summary

4-week build plan converting the CHER architecture into a working multi-agent system. Covers technology stack, file structure, build sequence, code examples, testing checkpoints, and hardware migration.

Core Framework: LangGraph

Full control over agent routing logic
Explicit graph structure — easier debugging
Native LangChain integration
Model-agnostic — swap without framework changes
Superior state management for complex workflows

Alternative considered: CrewAI (simpler but less flexible)

Technology Stack

┌──────────────────────────────────────┐ │ CHER Technology Stack │ ├──────────────────────────────────────┤ │ Orchestration: LangGraph + LangChain │ │ Models: LM Studio / Ollama │ │ Knowledge: PANIC (SQLite) │ │ Interface: Open WebUI / Streamlit│ │ Storage: SQLite │ └──────────────────────────────────────┘

Key Dependencies

pip install --break-system-packages \
  langchain langchain-community \
  langgraph langchain-openai \
  streamlit sqlalchemy pydantic

Project Structure

~/CHER-1.0/
├── config/
│   ├── agents.yaml
│   ├── models.yaml
│   └── tools.yaml
├── cher/
│   ├── supervisor.py
│   ├── agents/
│   │   ├── code_agent.py
│   │   ├── research_agent.py
│   │   ├── content_agent.py
│   │   └── deploy_agent.py
│   ├── tools/
│   │   ├── script_executor.py
│   │   ├── file_ops.py
│   │   └── panic_query.py
│   ├── routing/
│   │   ├── planner.py
│   │   └── router.py
│   └── state/
│       └── graph_state.py

├── data/
│   ├── state.db
│   └── logs/
│       └── execution.log
├── tests/
│   ├── test_supervisor.py
│   ├── test_agents.py
│   └── test_tools.py
└── ui/
    └── streamlit_app.py

~/Agentics/panic_agent/
├── ingest.py
├── embed.py
├── panic_load.py
├── panic_agent4.py
├── panic_query.py
├── app.py
└── config.json

▶ PHASE 1 — FOUNDATION (Week 1) | Goal: Basic supervisor routing to a single agent

Day 1–2: Setup & Configuration

Create project structure
Install dependencies
Configure LM Studio connection
Test model connectivity

config/models.yaml

models:
  supervisor:
    provider: lm_studio
    base_url: http://localhost:1234/v1
    model: mistral-nemo-12b-instruct
    temperature: 0.1
    max_tokens: 2000
  worker:
    provider: lm_studio
    base_url: http://localhost:1234/v1
    model: meta-llama-3-8b-instruct
    temperature: 0.3
    max_tokens: 1500

✓ Checkpoint: Run connection test, see "connection successful"

Day 3–4: Supervisor Routing

class CHERSupervisor:
  def route(self, user_request: str) -> dict:
    chain = self.routing_prompt | self.llm
    response = chain.invoke(
      {"user_request": user_request}
    )
    return json.loads(response.content)

✓ Checkpoint: Supervisor routes 5 request types correctly

Day 5–7: Code Agent

class CodeAgent:
  def execute(self, task: str) -> dict:
    chain = self.prompt | self.llm
    response = chain.invoke({"task": task})
    return {
      "filename": self._extract_filename(
        response.content),
      "code": self._extract_code(
        response.content),
      "explanation": ...
    }

✓ Checkpoint: Code agent generates working Python functions

Week 1 Success Criteria

✅ LM Studio connected

✅ Supervisor routing verified

✅ Code agent functional

✅ All Phase 1 tests pass

Test Protocol

pytest tests/test_phase1.py
# ✓ LM Studio connection
# ✓ Supervisor routing
# ✓ Code agent execution

▶ PHASE 2 — TOOL INTEGRATION (Week 2) | Goal: Agents execute existing scripts

Day 8–10: Script Executor

class ScriptExecutor:
  allowed_scripts = {
    "bmb":   scripts_dir / "BMB.py",
    "risk":  scripts_dir / "risk.py",
  }

  def run(self, script_name, args=None):
    result = subprocess.run(
      ["python", script_path] + args,
      capture_output=True,
      timeout=300
    )
    return {"success": result.returncode == 0,
            "stdout": result.stdout}

✓ Checkpoint: Can trigger BMB.py run from CHER

Day 13–14: PANIC Integration

class PANICQuery:
  def get_context(self, query, top_k=3):
    # Filename match first
    doc_id = self._find_doc_by_name(query)
    if doc_id:
      return self._chunks_for_doc(doc_id)
    # Fall back to semantic search
    return self.search(query, top_k)

✓ Checkpoint: Agents query PANIC for client context

Week 2 Success Criteria

✅ Script executor runs BMB, risk.py

✅ Agents write files to outputs

✅ PANIC KB accessible

✅ End-to-end audit test passes

▶ PHASE 3 — MULTI-AGENT COORDINATION (Week 3) | Goal: Agents collaborate on complex tasks

Day 15–17: LangGraph State Machine

class AgentState(TypedDict):
  messages:      Sequence[BaseMessage]
  current_agent: str
  task:          str
  context:       dict
  results:       dict
  next_action:   str

Graph Construction

workflow = StateGraph(AgentState)
workflow.add_node("supervisor", ...)
workflow.add_node("code_agent", ...)
workflow.add_node("research_agent", ...)
workflow.add_node("review", ...)
workflow.set_entry_point("supervisor")
graph = workflow.compile()

✓ Checkpoint: Supervisor → Research → Content → Review completes

Day 18–21: Remaining Agents + Testing

Research Agent — queries PANIC, synthesizes findings
Content Agent — document and report generation
Deploy Agent — file ops, script execution

Week 3 Success Criteria

✅ 3+ agents coordinate

✅ State flows correctly

✅ Agents share context

✅ Audit → report workflow

pytest tests/test_phase3.py
# ✓ Multi-agent coordination
# ✓ State management
# ✓ Complex workflows

▶ PHASE 4 — INTERFACE & POLISH (Week 4) | Goal: Production-ready UI and error handling

Day 22–24: Streamlit / Open WebUI

@st.cache_resource
def load_cher():
  return create_cher_graph(
    supervisor, agents, tools)

if prompt := st.chat_input("Task..."):
  result = graph.invoke({"task": prompt})
  response = result["results"]["output"]
  st.write(response)

✓ Checkpoint: Chat interface operational

Day 25–26: Logging

logging.basicConfig(
  filename="data/logs/execution.log",
  level=logging.INFO,
  format='%(asctime)s %(levelname)s %(message)s'
)

Quick Reference — Start CHER

cd ~/CHER-1.0
source ~/.ai_env/bin/activate
streamlit run ui/streamlit_app.py

Quick Reference — PANIC Load

source ~/.ai_env/bin/activate && \
python3 ~/Agentics/panic_agent/\
panic_load.py ~/path/to/folder/

Week 4 Success Criteria

✅ UI functional

✅ Errors logged and handled

✅ Documentation complete

✅ Ready for Beast deployment

Hardware Migration — HAVOC → Beast

No Code Changes Required

Same Python environment
Same model endpoints
Same file paths

Migration Steps

rsync ~/CHER-1.0/ to Beast
Install dependencies
Configure LM Studio on Beast
Test with same commands

What Changes on Beast

Response time: ~30s → ~3s
Parallel agents: 1 → 4–6
RAM: 16GB → 128GB
GPU: Vega 8 iGPU → Dual RX 7900 XTX
Embedding: CPU sequential → GPU batched

Common Pitfalls & Solutions

LM Studio connection fails → Verify server on port 1234, check firewall

Agents return gibberish → Adjust temperature, check model quantization

State not persisting → Check SQLite write permissions, verify checkpoint config

Scripts won't execute → Verify paths in tools.yaml, check Python env

Slow response on HAVOC → Expected — reduce model size or await Beast

Embedding too slow → Switch to sentence-transformers batch mode (done)

CHER-1.0 BUILD ROADMAP