agent-core

name	agent-core
description	Guides the implementation of the ReAct loop, orchestrator, memory, and planner. Use this when working on any file inside backend/agent/.

Agent Core Skill

ReAct Loop

The agent follows a strict Think → Act → Observe cycle. Never break this pattern.

• THINK: Reason about current state, decide next action
• ACT: Call exactly one tool with precise arguments
• OBSERVE: Read output, update memory, decide next step
• REPEAT until goal is achieved

The agent never outputs free text mid-loop. Only the final DONE step produces a summary.

Orchestrator Rules

1. Maintain the ReAct loop until task completion
2. Inject workspace_context into LLM on every step
3. Route tool calls through Tool Router, never call tools directly
4. Write every observation back to memory

workspace_context Shape

Always injected before each ReAct step:

{
  "session_id": "...",
  "target": "...",
  "completed_tasks": [],
  "pending_tasks": [],
  "recent_findings": [],
  "open_artifacts": []
}

Memory Rules

• Short-term: last N steps of current session (in-memory)
• Long-term: workspace/.agent/findings.json (persisted)
• Never dump full history into LLM context — only recent + relevant

Planner Rules

1. Decompose high-level goal into ordered task list on session start
2. Store task list in workspace/.agent/plan.json
3. Tick off tasks as they complete
4. Re-plan if a dead end is reached