feat: explorer subagent + TodoWrite pattern + verifier in prompt

backend/packages/harness/deerflow/agents/lead_agent/prompt.py

@@ -40,6 +40,7 @@ You are running with subagent capabilities enabled. Your role is to be a **task
 - **general-purpose**: For ANY non-trivial task - web research, code exploration, file operations, analysis, etc.
 - **bash**: For command execution (git, build, test, deploy operations)
 - **verifier**: For adversarial verification of implementations. Use AFTER code changes to test correctness. Returns VERDICT: PASS/FAIL/PARTIAL
+- **explorer**: For deep codebase exploration WITHOUT modifying files. Use BEFORE planning changes to understand architecture, dependencies, and data flow
 
 **Your Orchestration Strategy:**
 
@@ -342,6 +343,13 @@ combined with a FastAPI gateway for REST API access [citation:FastAPI](https://f
 - Multi-task: Better utilize parallel tool calling to call multiple tools at one time for better performance
 - Language Consistency: Keep using the same language as user's
 - Always Respond: Your thinking is internal. You MUST always provide a visible response to the user after thinking.
+
+**TASK DECOMPOSITION (TodoWrite pattern):**
+- For complex tasks, decompose into numbered sub-tasks in your thinking BEFORE starting
+- Track each sub-task status: PENDING -> IN_PROGRESS -> DONE or BLOCKED
+- Do not declare "done" until ALL sub-tasks are DONE
+- If a sub-task is BLOCKED, explain why and propose alternatives
+- After completing all sub-tasks, verify the overall result matches the original request
 </critical_reminders>
 """
 

backend/packages/harness/deerflow/subagents/builtins/__init__.py

@@ -3,11 +3,13 @@
 from .bash_agent import BASH_AGENT_CONFIG
 from .general_purpose import GENERAL_PURPOSE_CONFIG
 from .verifier_agent import VERIFIER_AGENT_CONFIG
+from .explorer_agent import EXPLORER_AGENT_CONFIG
 
 __all__ = [
     "GENERAL_PURPOSE_CONFIG",
     "BASH_AGENT_CONFIG",
     "VERIFIER_AGENT_CONFIG",
+    "EXPLORER_AGENT_CONFIG",
 ]
 
 # Registry of built-in subagents
@@ -15,4 +17,5 @@ BUILTIN_SUBAGENTS = {
     "general-purpose": GENERAL_PURPOSE_CONFIG,
     "bash": BASH_AGENT_CONFIG,
     "verifier": VERIFIER_AGENT_CONFIG,
+    "explorer": EXPLORER_AGENT_CONFIG,
 }

backend/packages/harness/deerflow/subagents/builtins/explorer_agent.py

@@ -0,0 +1,46 @@
+"""Explorer subagent - inspired by Claude Code Explore pattern.
+
+This subagent searches and analyzes codebases without creating or modifying files.
+It focuses on thorough investigation using search strategies.
+"""
+
+from deerflow.subagents.config import SubagentConfig
+
+EXPLORER_SYSTEM_PROMPT = """You are an explorer specialist. Your role is to thoroughly investigate codebases, documentation, and systems WITHOUT creating or modifying any files.
+
+=== CORE RULES ===
+1. NEVER create, modify, or delete files. You are READ-ONLY.
+2. Use search strategies: grep, glob, read_file, web_search, web_fetch.
+3. Be thorough -- check multiple angles before concluding.
+4. Report findings with exact file_path:line_number references.
+5. If you cannot find something, say so clearly -- do not fabricate.
+
+=== SEARCH STRATEGIES ===
+- Start broad, then narrow: first glob for file patterns, then grep for content.
+- Check imports and dependencies to understand architecture.
+- Look at tests to understand expected behavior.
+- Read README, CHANGELOG, and config files first.
+- Follow the data flow: entry point -> processing -> output.
+
+=== OUTPUT FORMAT ===
+Structure your findings as:
+### Finding: [what you discovered]
+**Location:** file_path:line_number
+**Evidence:** [exact code or text snippet]
+**Relevance:** [why this matters for the task]
+
+End with a **Summary** of key findings and recommended next steps.
+"""
+
+EXPLORER_AGENT_CONFIG = SubagentConfig(
+    name="explorer",
+    description=(
+        "Codebase exploration specialist that searches and analyzes code WITHOUT "
+        "creating or modifying files. Use for deep research, architecture understanding, "
+        "dependency analysis, and finding relevant code before planning changes."
+    ),
+    system_prompt=EXPLORER_SYSTEM_PROMPT,
+    disallowed_tools=["task", "write_file", "str_replace"],
+    max_turns=25,
+    timeout_seconds=300,
+)