Sisyphus agent recreated in LangChain to figure out how it works and how to use it

2026-04-15 12:47:38 -06:00
parent ff3419a714
commit 9bab6a0c2d
14 changed files with 1745 additions and 0 deletions
@@ -0,0 +1,110 @@
+"""
+Explore agent node — the read-only codebase researcher.
+
+Loki equivalent: assets/agents/explore/config.yaml + tools.sh
+
+In Loki, the explore agent is spawned via `agent__spawn --agent explore --prompt "..."`
+and runs as an isolated subprocess with its own session.  It ends with
+"EXPLORE_COMPLETE" so the parent knows it's finished.
+
+In LangGraph, the explore agent is a *node* in the graph.  The supervisor routes
+to it via `Command(goto="explore")`.  It reads the latest message (the supervisor's
+delegation prompt), calls the LLM with read-only tools, and writes its findings
+back to the shared message list.  The graph edge then returns control to the
+supervisor.
+
+Key differences from Loki:
+    - No isolated session — shares the graph's message list (but has its own
+      system prompt and tool set, just like Loki's per-agent config).
+    - No "EXPLORE_COMPLETE" sentinel — the graph edge handles control flow.
+    - No output summarization — LangGraph's state handles context management.
+"""
+
+from __future__ import annotations
+
+from langchain_core.messages import SystemMessage
+from langchain_openai import ChatOpenAI
+
+from sisyphus_langchain.state import SisyphusState
+from sisyphus_langchain.tools.filesystem import (
+    list_directory,
+    read_file,
+    search_content,
+    search_files,
+)
+
+# ---------------------------------------------------------------------------
+# System prompt — faithfully mirrors explore/config.yaml
+# ---------------------------------------------------------------------------
+EXPLORE_SYSTEM_PROMPT = """\
+You are a codebase explorer. Your job: Search, find, report. Nothing else.
+
+## Your Mission
+
+Given a search task, you:
+1. Search for relevant files and patterns
+2. Read key files to understand structure
+3. Report findings concisely
+
+## Strategy
+
+1. **Find first, read second** — Never read a file without knowing why.
+2. **Use search_content to locate** — find exactly where things are defined.
+3. **Use search_files to discover** — find files by name pattern.
+4. **Read targeted sections** — use offset and limit to read only relevant lines.
+5. **Never read entire large files** — if a file is 500+ lines, read the relevant section only.
+
+## Output Format
+
+Always end your response with a structured findings summary:
+
+FINDINGS:
+- [Key finding 1]
+- [Key finding 2]
+- Relevant files: [list of paths]
+
+## Rules
+
+1. Be fast — don't read every file, read representative ones.
+2. Be focused — answer the specific question asked.
+3. Be concise — report findings, not your process.
+4. Never modify files — you are read-only.
+5. Limit reads — max 5 file reads per exploration.
+"""
+
+# Read-only tools — mirrors explore's tool set (no write_file, no execute_command)
+EXPLORE_TOOLS = [read_file, search_content, search_files, list_directory]
+
+
+def create_explore_node(model_name: str = "gpt-4o-mini", temperature: float = 0.1):
+    """
+    Factory that returns an explore node function bound to a specific model.
+
+    In Loki, the model is set per-agent in config.yaml.  Here we parameterize it
+    so you can use a cheap model for exploration (cost optimization).
+
+    Args:
+        model_name: OpenAI model identifier.
+        temperature: LLM temperature (Loki explore uses 0.1).
+    """
+    llm = ChatOpenAI(model=model_name, temperature=temperature).bind_tools(EXPLORE_TOOLS)
+
+    def explore_node(state: SisyphusState) -> dict:
+        """
+        LangGraph node: run the explore agent.
+
+        Reads the conversation history, applies the explore system prompt,
+        invokes the LLM with read-only tools, and returns the response.
+        """
+        response = llm.invoke(
+            [SystemMessage(content=EXPLORE_SYSTEM_PROMPT)] + state["messages"]
+        )
+        return {
+            "messages": [response],
+            "agent_outputs": {
+                **state.get("agent_outputs", {}),
+                "explore": response.content,
+            },
+        }
+
+    return explore_node