From 9bab6a0c2d0e18550f6a8894b6ebdbf22e1a226b Mon Sep 17 00:00:00 2001
From: Alex Clarke <alex.j.tusa@gmail.com>
Date: Wed, 15 Apr 2026 12:47:38 -0600
Subject: [PATCH] Sisyphus agent recreated in LangChain to figure out how it
 works and how to use it

---
 examples/langchain-sisyphus/README.md         | 416 ++++++++++++++++++
 examples/langchain-sisyphus/pyproject.toml    |  29 ++
 .../sisyphus_langchain/__init__.py            |   5 +
 .../sisyphus_langchain/agents/__init__.py     |   1 +
 .../sisyphus_langchain/agents/coder.py        | 145 ++++++
 .../sisyphus_langchain/agents/explore.py      | 110 +++++
 .../sisyphus_langchain/agents/oracle.py       | 124 ++++++
 .../sisyphus_langchain/agents/supervisor.py   | 227 ++++++++++
 .../sisyphus_langchain/cli.py                 | 155 +++++++
 .../sisyphus_langchain/graph.py               | 115 +++++
 .../sisyphus_langchain/state.py               | 100 +++++
 .../sisyphus_langchain/tools/__init__.py      |   1 +
 .../sisyphus_langchain/tools/filesystem.py    | 175 ++++++++
 .../sisyphus_langchain/tools/project.py       | 142 ++++++
 14 files changed, 1745 insertions(+)
 create mode 100644 examples/langchain-sisyphus/README.md
 create mode 100644 examples/langchain-sisyphus/pyproject.toml
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/__init__.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/agents/__init__.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/agents/coder.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/agents/explore.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/agents/oracle.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/agents/supervisor.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/cli.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/graph.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/state.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/tools/__init__.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/tools/filesystem.py
 create mode 100644 examples/langchain-sisyphus/sisyphus_langchain/tools/project.py

diff --git a/examples/langchain-sisyphus/README.md b/examples/langchain-sisyphus/README.md
new file mode 100644
index 0000000..120dfc3
--- /dev/null
+++ b/examples/langchain-sisyphus/README.md
@@ -0,0 +1,416 @@
+# Sisyphus in LangChain/LangGraph
+
+A faithful recreation of [Loki's Sisyphus agent](../../assets/agents/sisyphus/) using [LangGraph](https://docs.langchain.com/langgraph/) — LangChain's framework for stateful, multi-agent workflows.
+
+This project exists to help you understand LangChain/LangGraph by mapping every concept to its Loki equivalent.
+
+## Architecture Overview
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│                     SUPERVISOR NODE                          │
+│  Intent classification → Routing decision → Command(goto=)   │
+│                                                              │
+│  Loki equivalent: sisyphus/config.yaml                       │
+│  (agent__spawn → Command, agent__collect → graph edge)       │
+└──────────┬──────────────┬──────────────┬────────────────────┘
+           │              │              │
+           ▼              ▼              ▼
+    ┌────────────┐ ┌────────────┐ ┌────────────┐
+    │  EXPLORE   │ │   ORACLE   │ │   CODER    │
+    │ (research) │ │  (advise)  │ │  (build)   │
+    │            │ │            │ │            │
+    │ read-only  │ │ read-only  │ │ read+write │
+    │ tools      │ │ tools      │ │ tools      │
+    └─────┬──────┘ └─────┬──────┘ └─────┬──────┘
+          │              │              │
+          └──────────────┼──────────────┘
+                         │
+                  back to supervisor
+```
+
+## Concept Map: Loki → LangGraph
+
+This is the key reference.  Every row maps a Loki concept to its LangGraph equivalent.
+
+### Core Architecture
+
+| Loki Concept | LangGraph Equivalent | Where in Code |
+|---|---|---|
+| Agent config (config.yaml) | Node function + system prompt | `agents/explore.py`, etc. |
+| Agent instructions | System prompt string | `EXPLORE_SYSTEM_PROMPT`, etc. |
+| Agent tools (tools.sh) | `@tool`-decorated Python functions | `tools/filesystem.py`, `tools/project.py` |
+| Agent session (chat loop) | Graph state + message list | `state.py` → `SisyphusState.messages` |
+| `agent__spawn --agent X` | `Command(goto="X")` | `agents/supervisor.py` |
+| `agent__collect --id` | Graph edge (implicit — workers return to supervisor) | `graph.py` → `add_edge("explore", "supervisor")` |
+| `agent__check` (non-blocking) | Not needed (graph handles scheduling) | — |
+| `agent__cancel` | Not needed (graph handles lifecycle) | — |
+| `can_spawn_agents: true` | Node has routing logic (supervisor) | `agents/supervisor.py` |
+| `max_concurrent_agents: 4` | `Send()` API for parallel fan-out | See [Parallel Execution](#parallel-execution) |
+| `max_agent_depth: 3` | `recursion_limit` in config | `cli.py` → `recursion_limit: 50` |
+| `summarization_threshold` | Manual truncation in supervisor | `supervisor.py` → `_summarize_outputs()` |
+
+### Tool System
+
+| Loki Concept | LangGraph Equivalent | Notes |
+|---|---|---|
+| `tools.sh` with `@cmd` annotations | `@tool` decorator | Loki compiles bash annotations to JSON schema; LangChain generates schema from the Python function signature + docstring |
+| `@option --pattern!` (required arg) | Function parameter without default | `def search_content(pattern: str)` |
+| `@option --lines` (optional arg) | Parameter with default | `def read_file(path: str, limit: int = 200)` |
+| `@env LLM_OUTPUT=/dev/stdout` | Return value | LangChain tools return strings; Loki tools write to `$LLM_OUTPUT` |
+| `@describe` | Docstring | The tool's docstring becomes the description the LLM sees |
+| Global tools (`fs_read.sh`, etc.) | Shared tool imports | Both agents import from `tools/filesystem.py` |
+| Agent-specific tools | Per-node tool binding | `llm.bind_tools(EXPLORE_TOOLS)` vs `llm.bind_tools(CODER_TOOLS)` |
+| `.shared/utils.sh` | `tools/project.py` | Shared project detection utilities |
+| `detect_project()` heuristic | `detect_project()` in Python | Same logic: check Cargo.toml → go.mod → package.json → etc. |
+| LLM fallback for unknown projects | (omitted) | The agents themselves can reason about unknown project types |
+
+### State & Memory
+
+| Loki Concept | LangGraph Equivalent | Notes |
+|---|---|---|
+| Agent session (conversation history) | `SisyphusState.messages` | `Annotated[list, add_messages]` — the reducer appends instead of replacing |
+| `agent_session: temp` | `MemorySaver` checkpointer | Loki's temp sessions are ephemeral; MemorySaver is in-memory (lost on restart) |
+| Per-agent isolation | Per-node system prompt + tools | In Loki agents have separate sessions; in LangGraph they share messages but have different system prompts |
+| `{{project_dir}}` variable | `SisyphusState.project_dir` | Loki interpolates variables into prompts; LangGraph stores them in state |
+| `{{__tools__}}` injection | `llm.bind_tools()` | Loki injects tool descriptions into the prompt; LangChain attaches them to the API call |
+
+### Orchestration
+
+| Loki Concept | LangGraph Equivalent | Notes |
+|---|---|---|
+| Intent classification table | `RoutingDecision` structured output | Loki does this in free text; LangGraph forces typed JSON |
+| Oracle triggers ("How should I...") | Supervisor prompt + structured output | Same trigger phrases, enforced via system prompt |
+| Coder delegation format | Supervisor builds HumanMessage | The structured prompt (Goal/Reference Files/Conventions/Constraints) |
+| `agent__spawn` (parallel) | `Send()` API | Dynamic fan-out to multiple nodes |
+| Todo system (`todo__init`, etc.) | `SisyphusState.todos` | State field with a merge reducer |
+| `auto_continue: true` | Supervisor loop (iteration counter) | Supervisor re-routes until FINISH or max iterations |
+| `max_auto_continues: 25` | `MAX_ITERATIONS = 15` | Safety valve to prevent infinite loops |
+| `user__ask` / `user__confirm` | `interrupt()` API | Pauses graph, surfaces question to caller, resumes with answer |
+| Escalation (child → parent → user) | `interrupt()` in any node | Any node can pause; the caller handles the interaction |
+
+### Execution Model
+
+| Loki Concept | LangGraph Equivalent | Notes |
+|---|---|---|
+| `loki --agent sisyphus` | `python -m sisyphus_langchain.cli` | CLI entry point |
+| REPL mode | `cli.py` → `repl()` | Interactive loop with thread persistence |
+| One-shot mode | `cli.py` → `run_query()` | Single query, print result, exit |
+| Streaming output | `graph.stream()` | LangGraph supports per-node streaming |
+| `inject_spawn_instructions` | (always on) | System prompts are always included |
+| `inject_todo_instructions` | (always on) | Todo instructions could be added to prompts |
+
+## How the Execution Flow Works
+
+### 1. User sends a message
+
+```python
+graph.invoke({"messages": [HumanMessage("Add a health check endpoint")]})
+```
+
+### 2. Supervisor classifies intent
+
+The supervisor LLM reads the message and produces a `RoutingDecision`:
+```json
+{
+  "intent": "implementation",
+  "next_agent": "explore",
+  "delegation_notes": "Find existing API endpoint patterns, route structure, and health check conventions"
+}
+```
+
+### 3. Supervisor routes via Command
+
+```python
+return Command(goto="explore", update={"intent": "implementation", "iteration_count": 1})
+```
+
+### 4. Explore agent runs
+
+- Receives the full message history (including the user's request)
+- Calls read-only tools (search_content, search_files, read_file)
+- Returns findings in messages
+
+### 5. Control returns to supervisor
+
+The graph edge `explore → supervisor` fires automatically.
+
+### 6. Supervisor reviews and routes again
+
+Now it has explore's findings.  It routes to coder with context:
+```json
+{
+  "intent": "implementation",
+  "next_agent": "coder",
+  "delegation_notes": "Implement health check endpoint following patterns found in src/routes/"
+}
+```
+
+### 7. Coder implements
+
+- Reads explore's findings from the message history
+- Writes files via `write_file` tool
+- Runs `verify_build` to check compilation
+
+### 8. Supervisor verifies and finishes
+
+```json
+{
+  "intent": "implementation",
+  "next_agent": "FINISH",
+  "delegation_notes": "Added /health endpoint in src/routes/health.py. Build passes."
+}
+```
+
+## Key Differences from Loki
+
+### What LangGraph does better
+
+1. **Declarative graph** — The topology is visible and debuggable.  Loki's orchestration is emergent from the LLM's tool calls.
+2. **Typed state** — `SisyphusState` is a TypedDict with reducers.  Loki's state is implicit in the conversation.
+3. **Checkpointing** — Built-in persistence.  Loki manages sessions manually.
+4. **Time-travel debugging** — Inspect any checkpoint.  Loki has no equivalent.
+5. **Structured routing** — `RoutingDecision` forces valid JSON.  Loki relies on the LLM calling the right tool.
+
+### What Loki does better
+
+1. **True parallelism** — `agent__spawn` runs multiple agents concurrently in separate threads.  This LangGraph implementation is sequential (see [Parallel Execution](#parallel-execution) for how to add it).
+2. **Agent isolation** — Each Loki agent has its own session, tools, and config.  LangGraph nodes share state.
+3. **Teammate messaging** — Loki agents can send messages to siblings.  LangGraph nodes communicate only through shared state.
+4. **Dynamic tool compilation** — Loki compiles bash/python/typescript tools at startup.  LangChain tools are statically defined.
+5. **Escalation protocol** — Loki's child-to-parent escalation is sophisticated.  LangGraph's `interrupt()` is simpler but less structured.
+6. **Task queues with dependencies** — Loki's `agent__task_create` supports dependency DAGs.  LangGraph's routing is simpler (hub-and-spoke).
+
+## Running It
+
+### Prerequisites
+
+```bash
+# Python 3.11+
+python --version
+
+# Set your API key
+export OPENAI_API_KEY="sk-..."
+```
+
+### Install
+
+```bash
+cd examples/langchain-sisyphus
+
+# With pip
+pip install -e .
+
+# Or with uv (recommended)
+uv pip install -e .
+```
+
+### Usage
+
+```bash
+# Interactive REPL (like `loki --agent sisyphus`)
+sisyphus
+
+# One-shot query
+sisyphus "Find all TODO comments in the codebase"
+
+# With custom models (cost optimization)
+sisyphus --explore-model gpt-4o-mini --coder-model gpt-4o "Add input validation to the API"
+
+# Programmatic usage
+python -c "
+from sisyphus_langchain import build_graph
+from langchain_core.messages import HumanMessage
+
+graph = build_graph()
+result = graph.invoke({
+    'messages': [HumanMessage('What patterns does this codebase use?')],
+    'intent': 'ambiguous',
+    'next_agent': '',
+    'iteration_count': 0,
+    'todos': [],
+    'agent_outputs': {},
+    'final_output': '',
+    'project_dir': '.',
+}, config={'configurable': {'thread_id': 'demo'}, 'recursion_limit': 50})
+print(result['final_output'])
+"
+```
+
+### Using Anthropic Models
+
+Replace `ChatOpenAI` with `ChatAnthropic` in the agent factories:
+
+```python
+from langchain_anthropic import ChatAnthropic
+
+# In agents/oracle.py:
+llm = ChatAnthropic(model="claude-sonnet-4-20250514", temperature=0.2).bind_tools(ORACLE_TOOLS)
+```
+
+## Deployment
+
+### Option 1: Standalone Script (Simplest)
+
+Just run the CLI directly. No infrastructure needed.
+
+```bash
+sisyphus "Add a health check endpoint"
+```
+
+### Option 2: FastAPI Server
+
+```python
+# server.py
+from fastapi import FastAPI
+from langserve import add_routes
+from sisyphus_langchain import build_graph
+
+app = FastAPI(title="Sisyphus API")
+graph = build_graph()
+add_routes(app, graph, path="/agent")
+
+# Run: uvicorn server:app --host 0.0.0.0 --port 8000
+# Call: POST http://localhost:8000/agent/invoke
+```
+
+### Option 3: LangGraph Platform (Production)
+
+Create a `langgraph.json` at the project root:
+
+```json
+{
+  "graphs": {
+    "sisyphus": "./sisyphus_langchain/graph.py:build_graph"
+  },
+  "dependencies": ["./sisyphus_langchain"],
+  "env": ".env"
+}
+```
+
+Then deploy:
+```bash
+pip install langgraph-cli
+langgraph deploy
+```
+
+This gives you:
+- Durable checkpointing (PostgreSQL)
+- Background runs
+- Streaming API
+- Zero-downtime deployments
+- Built-in observability
+
+### Option 4: Docker
+
+```dockerfile
+FROM python:3.12-slim
+WORKDIR /app
+COPY . .
+RUN pip install -e .
+CMD ["sisyphus"]
+```
+
+```bash
+docker build -t sisyphus .
+docker run -it -e OPENAI_API_KEY=$OPENAI_API_KEY sisyphus
+```
+
+## Parallel Execution
+
+This implementation routes sequentially for simplicity. To add Loki-style parallel agent execution, use LangGraph's `Send()` API:
+
+```python
+from langgraph.types import Send
+
+def supervisor_node(state):
+    # Fan out to multiple explore agents in parallel
+    # (like Loki's agent__spawn called multiple times)
+    return [
+        Send("explore", {
+            **state,
+            "messages": state["messages"] + [
+                HumanMessage("Find existing API endpoint patterns")
+            ],
+        }),
+        Send("explore", {
+            **state,
+            "messages": state["messages"] + [
+                HumanMessage("Find data models and database patterns")
+            ],
+        }),
+    ]
+```
+
+This is equivalent to Loki's pattern of spawning multiple explore agents:
+```
+agent__spawn --agent explore --prompt "Find API patterns"
+agent__spawn --agent explore --prompt "Find database patterns"
+agent__collect --id <id1>
+agent__collect --id <id2>
+```
+
+## Adding Human-in-the-Loop
+
+To replicate Loki's `user__ask` / `user__confirm` tools, use LangGraph's `interrupt()`:
+
+```python
+from langgraph.types import interrupt
+
+def supervisor_node(state):
+    # Pause and ask the user (like Loki's user__ask)
+    answer = interrupt({
+        "question": "How should we structure the authentication?",
+        "options": [
+            "JWT with httpOnly cookies (Recommended)",
+            "Session-based with Redis",
+            "OAuth2 with external provider",
+        ],
+    })
+    # `answer` contains the user's selection when the graph resumes
+```
+
+## Project Structure
+
+```
+examples/langchain-sisyphus/
+├── pyproject.toml                          # Dependencies & build config
+├── README.md                               # This file
+└── sisyphus_langchain/
+    ├── __init__.py                          # Package entry point
+    ├── cli.py                              # CLI (REPL + one-shot mode)
+    ├── graph.py                            # Graph assembly (wires nodes + edges)
+    ├── state.py                            # Shared state schema (TypedDict)
+    ├── agents/
+    │   ├── __init__.py
+    │   ├── supervisor.py                   # Sisyphus orchestrator (intent → routing)
+    │   ├── explore.py                      # Read-only codebase researcher
+    │   ├── oracle.py                       # Architecture/debugging advisor
+    │   └── coder.py                        # Implementation worker
+    └── tools/
+        ├── __init__.py
+        ├── filesystem.py                   # File read/write/search/glob tools
+        └── project.py                      # Project detection, build, test tools
+```
+
+### File-to-Loki Mapping
+
+| This Project | Loki Equivalent |
+|---|---|
+| `state.py` | Session context + todo state (implicit in Loki) |
+| `graph.py` | `src/supervisor/mod.rs` (runtime orchestration) |
+| `cli.py` | `src/main.rs` (CLI entry point) |
+| `agents/supervisor.py` | `assets/agents/sisyphus/config.yaml` |
+| `agents/explore.py` | `assets/agents/explore/config.yaml` + `tools.sh` |
+| `agents/oracle.py` | `assets/agents/oracle/config.yaml` + `tools.sh` |
+| `agents/coder.py` | `assets/agents/coder/config.yaml` + `tools.sh` |
+| `tools/filesystem.py` | `assets/functions/tools/fs_*.sh` |
+| `tools/project.py` | `assets/agents/.shared/utils.sh` + `sisyphus/tools.sh` |
+
+## Further Reading
+
+- [LangGraph Documentation](https://docs.langchain.com/langgraph/)
+- [LangGraph Multi-Agent Tutorial](https://docs.langchain.com/langgraph/how-tos/multi-agent-systems)
+- [Loki Agents Documentation](../../docs/AGENTS.md)
+- [Loki Sisyphus README](../../assets/agents/sisyphus/README.md)
+- [LangGraph Supervisor Library](https://github.com/langchain-ai/langgraph-supervisor-py)
diff --git a/examples/langchain-sisyphus/pyproject.toml b/examples/langchain-sisyphus/pyproject.toml
new file mode 100644
index 0000000..8012272
--- /dev/null
+++ b/examples/langchain-sisyphus/pyproject.toml
@@ -0,0 +1,29 @@
+[project]
+name = "sisyphus-langchain"
+version = "0.1.0"
+description = "Loki's Sisyphus multi-agent orchestrator recreated in LangChain/LangGraph"
+readme = "README.md"
+requires-python = ">=3.11"
+dependencies = [
+    "langgraph>=0.3.0",
+    "langchain>=0.3.0",
+    "langchain-openai>=0.3.0",
+    "langchain-anthropic>=0.3.0",
+    "langchain-core>=0.3.0",
+]
+
+[project.optional-dependencies]
+dev = [
+    "pytest>=8.0",
+    "ruff>=0.8.0",
+]
+server = [
+    "langgraph-api>=0.1.0",
+]
+
+[project.scripts]
+sisyphus = "sisyphus_langchain.cli:main"
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/__init__.py b/examples/langchain-sisyphus/sisyphus_langchain/__init__.py
new file mode 100644
index 0000000..9653e84
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/__init__.py
@@ -0,0 +1,5 @@
+"""Sisyphus multi-agent orchestrator — a LangGraph recreation of Loki's Sisyphus agent."""
+
+from sisyphus_langchain.graph import build_graph
+
+__all__ = ["build_graph"]
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/agents/__init__.py b/examples/langchain-sisyphus/sisyphus_langchain/agents/__init__.py
new file mode 100644
index 0000000..b1dba46
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/agents/__init__.py
@@ -0,0 +1 @@
+"""Agent node definitions for the Sisyphus orchestrator."""
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/agents/coder.py b/examples/langchain-sisyphus/sisyphus_langchain/agents/coder.py
new file mode 100644
index 0000000..56faf64
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/agents/coder.py
@@ -0,0 +1,145 @@
+"""
+Coder agent node — the implementation worker.
+
+Loki equivalent: assets/agents/coder/config.yaml + tools.sh
+
+In Loki, the coder is the ONLY agent that modifies files.  It:
+    - Receives a structured prompt from sisyphus with code patterns to follow
+    - Writes files via the write_file tool (never pastes code in chat)
+    - Verifies builds after every change
+    - Signals CODER_COMPLETE or CODER_FAILED
+
+In LangGraph, coder is a node with write-capable tools (read_file, write_file,
+search_content, execute_command, verify_build).  The supervisor formats a
+structured delegation prompt (Goal / Reference Files / Code Patterns /
+Conventions / Constraints) and routes to this node.
+
+Key Loki→LangGraph mapping:
+    - Loki's "Coder Delegation Format" → the supervisor builds this as a
+      HumanMessage before routing to the coder node.
+    - Loki's auto_continue (up to 15) → the supervisor can re-route to coder
+      if verification fails, up to iteration_count limits.
+    - Loki's todo system for multi-file changes → the coder updates
+      state["todos"] as it completes each file.
+"""
+
+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 (
+    read_file,
+    search_content,
+    search_files,
+    write_file,
+)
+from sisyphus_langchain.tools.project import (
+    execute_command,
+    run_tests,
+    verify_build,
+)
+
+# ---------------------------------------------------------------------------
+# System prompt — faithfully mirrors coder/config.yaml
+# ---------------------------------------------------------------------------
+CODER_SYSTEM_PROMPT = """\
+You are a senior engineer. You write code that works on the first try.
+
+## Your Mission
+
+Given an implementation task:
+1. Check for context provided in the conversation (patterns, conventions, reference files).
+2. Fill gaps only — read files NOT already covered in context.
+3. Write the code using the write_file tool (NEVER output code in chat).
+4. Verify it compiles/builds using verify_build.
+5. Provide a summary of what you implemented.
+
+## Using Provided Context (IMPORTANT)
+
+Your prompt often contains prior findings from the explore agent: file paths,
+code patterns, and conventions.
+
+**If context is provided:**
+1. Use it as your primary reference.  Don't re-read files already summarized.
+2. Follow the code patterns shown — snippets in context ARE the style guide.
+3. Read referenced files ONLY IF you need more detail (full signatures, imports).
+4. If context includes a "Conventions" section, follow it exactly.
+
+**If context is NOT provided or is too vague:**
+Fall back to self-exploration: search for similar files, read 1-2 examples,
+match their style.
+
+## Writing Code
+
+CRITICAL: Write code using the write_file tool. NEVER paste code in chat.
+
+## Pattern Matching
+
+Before writing ANY file:
+1. Find a similar existing file.
+2. Match its style: imports, naming, structure.
+3. Follow the same patterns exactly.
+
+## Verification
+
+After writing files:
+1. Run verify_build to check compilation.
+2. If it fails, fix the error (minimal change).
+3. Don't move on until build passes.
+
+## Rules
+
+1. Write code via tools — never output code to chat.
+2. Follow patterns — read existing files first.
+3. Verify builds — don't finish without checking.
+4. Minimal fixes — if build fails, fix precisely.
+5. No refactoring — only implement what's asked.
+"""
+
+# Full tool set — coder gets write access and command execution
+CODER_TOOLS = [
+    read_file,
+    write_file,
+    search_content,
+    search_files,
+    execute_command,
+    verify_build,
+    run_tests,
+]
+
+
+def create_coder_node(model_name: str = "gpt-4o", temperature: float = 0.1):
+    """
+    Factory that returns a coder node function.
+
+    Coder needs a capable model because it writes production code.  In Loki,
+    coder uses the same model as the parent by default.
+
+    Args:
+        model_name: Model identifier.
+        temperature: LLM temperature (Loki coder uses 0.1 for consistency).
+    """
+    llm = ChatOpenAI(model=model_name, temperature=temperature).bind_tools(CODER_TOOLS)
+
+    def coder_node(state: SisyphusState) -> dict:
+        """
+        LangGraph node: run the coder agent.
+
+        Reads conversation history (including the supervisor's structured
+        delegation prompt), invokes the LLM with write-capable tools,
+        and returns the result.
+        """
+        response = llm.invoke(
+            [SystemMessage(content=CODER_SYSTEM_PROMPT)] + state["messages"]
+        )
+        return {
+            "messages": [response],
+            "agent_outputs": {
+                **state.get("agent_outputs", {}),
+                "coder": response.content,
+            },
+        }
+
+    return coder_node
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/agents/explore.py b/examples/langchain-sisyphus/sisyphus_langchain/agents/explore.py
new file mode 100644
index 0000000..7859396
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/agents/explore.py
@@ -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
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/agents/oracle.py b/examples/langchain-sisyphus/sisyphus_langchain/agents/oracle.py
new file mode 100644
index 0000000..aa1dfd0
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/agents/oracle.py
@@ -0,0 +1,124 @@
+"""
+Oracle agent node — the high-IQ architecture and debugging advisor.
+
+Loki equivalent: assets/agents/oracle/config.yaml + tools.sh
+
+In Loki, the oracle is a READ-ONLY advisor spawned for:
+    - Architecture decisions and multi-system tradeoffs
+    - Complex debugging (after 2+ failed fix attempts)
+    - Code/design review
+    - Risk assessment
+
+It uses temperature 0.2 (slightly higher than explore/coder for more creative
+reasoning) and ends with "ORACLE_COMPLETE".
+
+In LangGraph, oracle is a node that receives the full message history, reasons
+about the problem, and writes structured advice back.  It has read-only tools
+only — it never modifies files.
+
+Key Loki→LangGraph mapping:
+    - Loki oracle triggers (the "MUST spawn oracle when..." rules in sisyphus)
+      become routing conditions in the supervisor node.
+    - Oracle's structured output format (Analysis/Recommendation/Reasoning/Risks)
+      is enforced via the system prompt, same as in Loki.
+"""
+
+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 oracle/config.yaml
+# ---------------------------------------------------------------------------
+ORACLE_SYSTEM_PROMPT = """\
+You are Oracle — a senior architect and debugger consulted for complex decisions.
+
+## Your Role
+
+You are READ-ONLY. You analyze, advise, and recommend. You do NOT implement.
+
+## When You're Consulted
+
+1. **Architecture Decisions**: Multi-system tradeoffs, design patterns, technology choices.
+2. **Complex Debugging**: After 2+ failed fix attempts, deep analysis needed.
+3. **Code Review**: Evaluating proposed designs or implementations.
+4. **Risk Assessment**: Security, performance, or reliability concerns.
+
+## Your Process
+
+1. **Understand**: Read relevant code, understand the full context.
+2. **Analyze**: Consider multiple angles and tradeoffs.
+3. **Recommend**: Provide clear, actionable advice.
+4. **Justify**: Explain your reasoning.
+
+## Output Format
+
+Structure your response as:
+
+## Analysis
+[Your understanding of the situation]
+
+## Recommendation
+[Clear, specific advice]
+
+## Reasoning
+[Why this is the right approach]
+
+## Risks/Considerations
+[What to watch out for]
+
+## Rules
+
+1. Never modify files — you advise, others implement.
+2. Be thorough — read all relevant context before advising.
+3. Be specific — general advice isn't helpful.
+4. Consider tradeoffs — there are rarely perfect solutions.
+5. Stay focused — answer the specific question asked.
+"""
+
+# Read-only tools — same set as explore (oracle never writes)
+ORACLE_TOOLS = [read_file, search_content, search_files, list_directory]
+
+
+def create_oracle_node(model_name: str = "gpt-4o", temperature: float = 0.2):
+    """
+    Factory that returns an oracle node function.
+
+    Oracle uses a more expensive model than explore because it needs deeper
+    reasoning.  In Loki, the model is inherited from the global config unless
+    overridden in oracle/config.yaml.
+
+    Args:
+        model_name: Model identifier (use a strong reasoning model).
+        temperature: LLM temperature (Loki oracle uses 0.2).
+    """
+    llm = ChatOpenAI(model=model_name, temperature=temperature).bind_tools(ORACLE_TOOLS)
+
+    def oracle_node(state: SisyphusState) -> dict:
+        """
+        LangGraph node: run the oracle agent.
+
+        Reads conversation history, applies the oracle system prompt,
+        invokes the LLM, and returns structured advice.
+        """
+        response = llm.invoke(
+            [SystemMessage(content=ORACLE_SYSTEM_PROMPT)] + state["messages"]
+        )
+        return {
+            "messages": [response],
+            "agent_outputs": {
+                **state.get("agent_outputs", {}),
+                "oracle": response.content,
+            },
+        }
+
+    return oracle_node
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/agents/supervisor.py b/examples/langchain-sisyphus/sisyphus_langchain/agents/supervisor.py
new file mode 100644
index 0000000..7fb4771
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/agents/supervisor.py
@@ -0,0 +1,227 @@
+"""
+Sisyphus supervisor node — the orchestrator that classifies intent and routes.
+
+Loki equivalent: assets/agents/sisyphus/config.yaml
+
+This is the brain of the system.  In Loki, Sisyphus is the top-level agent that:
+    1. Classifies every incoming request (trivial / exploration / implementation /
+       architecture / ambiguous)
+    2. Routes to the appropriate sub-agent (explore, coder, oracle)
+    3. Manages the todo list for multi-step tasks
+    4. Verifies results and decides when the task is complete
+
+In LangGraph, the supervisor is a node that returns `Command(goto="agent_name")`
+to route control.  This replaces Loki's `agent__spawn` + `agent__collect` pattern
+with a declarative graph edge.
+
+Key Loki→LangGraph mapping:
+    - agent__spawn --agent explore  →  Command(goto="explore")
+    - agent__spawn --agent coder    →  Command(goto="coder")
+    - agent__spawn --agent oracle   →  Command(goto="oracle")
+    - agent__check / agent__collect →  (implicit: graph edges return to supervisor)
+    - todo__init / todo__add        →  state["todos"] updates
+    - user__ask / user__confirm     →  interrupt() for human-in-the-loop
+
+Parallel execution note:
+    Loki can spawn multiple explore agents in parallel.  In LangGraph, you'd use
+    the Send() API for dynamic fan-out.  For simplicity, this implementation uses
+    sequential routing.  See the README for how to add parallel fan-out.
+"""
+
+from __future__ import annotations
+
+from typing import Literal
+
+from langchain_core.messages import SystemMessage
+from langchain_openai import ChatOpenAI
+from langgraph.types import Command
+from pydantic import BaseModel, Field
+
+from sisyphus_langchain.state import SisyphusState
+
+# ---------------------------------------------------------------------------
+# Maximum iterations before forcing completion (safety valve)
+# Mirrors Loki's max_auto_continues: 25
+# ---------------------------------------------------------------------------
+MAX_ITERATIONS = 15
+
+# ---------------------------------------------------------------------------
+# Structured output schema for the supervisor's routing decision.
+#
+# In Loki, the supervisor is an LLM that produces free-text and calls tools
+# like agent__spawn.  In LangGraph, we use structured output to force the
+# LLM into a typed routing decision — more reliable than parsing free text.
+# ---------------------------------------------------------------------------
+class RoutingDecision(BaseModel):
+    """The supervisor's decision about what to do next."""
+
+    intent: Literal["trivial", "exploration", "implementation", "architecture", "ambiguous"] = Field(
+        description="Classified intent of the user's request."
+    )
+    next_agent: Literal["explore", "oracle", "coder", "FINISH"] = Field(
+        description=(
+            "Which agent to route to.  'explore' for research/discovery, "
+            "'oracle' for architecture/design/debugging advice, "
+            "'coder' for implementation, 'FINISH' if the task is complete."
+        )
+    )
+    delegation_notes: str = Field(
+        description=(
+            "Brief instructions for the target agent: what to look for (explore), "
+            "what to analyze (oracle), or what to implement (coder).  "
+            "For FINISH, summarize what was accomplished."
+        )
+    )
+
+
+# ---------------------------------------------------------------------------
+# Supervisor system prompt — faithfully mirrors sisyphus/config.yaml
+# ---------------------------------------------------------------------------
+SUPERVISOR_SYSTEM_PROMPT = """\
+You are Sisyphus — an orchestrator that drives coding tasks to completion.
+
+Your job: Classify → Delegate → Verify → Complete.
+
+## Intent Classification (BEFORE every action)
+
+| Type            | Signal                                              | Action               |
+|-----------------|-----------------------------------------------------|----------------------|
+| trivial         | Single file, known location, typo fix               | Route to FINISH      |
+| exploration     | "Find X", "Where is Y", "List all Z"               | Route to explore     |
+| implementation  | "Add feature", "Fix bug", "Write code"              | Route to coder       |
+| architecture    | See oracle triggers below                           | Route to oracle      |
+| ambiguous       | Unclear scope, multiple interpretations             | Route to FINISH with a clarifying question |
+
+## Oracle Triggers (MUST route to oracle when you see these)
+
+Route to oracle ANY time the user asks about:
+- "How should I..." / "What's the best way to..." — design/approach questions
+- "Why does X keep..." / "What's wrong with..." — complex debugging
+- "Should I use X or Y?" — technology or pattern choices
+- "How should this be structured?" — architecture
+- "Review this" / "What do you think of..." — code/design review
+- Tradeoff questions, multi-component questions, vague/open-ended questions
+
+## Agent Specializations
+
+| Agent   | Use For                                       |
+|---------|-----------------------------------------------|
+| explore | Find patterns, understand code, search        |
+| coder   | Write/edit files, implement features          |
+| oracle  | Architecture decisions, complex debugging     |
+
+## Workflow Patterns
+
+### Implementation task: explore → coder
+1. Route to explore to find existing patterns and conventions.
+2. Review explore findings.
+3. Route to coder with a structured prompt including the explore findings.
+4. Verify the coder's output (check for CODER_COMPLETE or CODER_FAILED).
+
+### Architecture question: explore + oracle
+1. Route to explore to find relevant code.
+2. Route to oracle with the explore findings for analysis.
+
+### Simple question: oracle directly
+For pure design/architecture questions, route to oracle directly.
+
+## Rules
+
+1. Always classify before acting.
+2. You are a coordinator, not an implementer.
+3. Route to oracle for ANY design/architecture question.
+4. When routing to coder, include code patterns from explore findings.
+5. Route to FINISH when the task is fully addressed.
+
+## Current State
+
+Iteration: {iteration_count}/{max_iterations}
+Previous agent outputs: {agent_outputs}
+"""
+
+
+def create_supervisor_node(model_name: str = "gpt-4o", temperature: float = 0.1):
+    """
+    Factory that returns a supervisor node function.
+
+    The supervisor uses a capable model for accurate routing.
+
+    Args:
+        model_name: Model identifier.
+        temperature: LLM temperature (low for consistent routing).
+    """
+    llm = ChatOpenAI(model=model_name, temperature=temperature).with_structured_output(
+        RoutingDecision
+    )
+
+    def supervisor_node(
+        state: SisyphusState,
+    ) -> Command[Literal["explore", "oracle", "coder", "__end__"]]:
+        """
+        LangGraph node: the Sisyphus supervisor.
+
+        Classifies the user's intent, decides which agent to route to,
+        and returns a Command that directs graph execution.
+        """
+        iteration = state.get("iteration_count", 0)
+
+        # Safety valve — prevent infinite loops
+        if iteration >= MAX_ITERATIONS:
+            return Command(
+                goto="__end__",
+                update={
+                    "final_output": "Reached maximum iterations.  Here's what was accomplished:\n"
+                    + "\n".join(
+                        f"- {k}: {v[:200]}" for k, v in state.get("agent_outputs", {}).items()
+                    ),
+                },
+            )
+
+        # Format the system prompt with current state
+        prompt = SUPERVISOR_SYSTEM_PROMPT.format(
+            iteration_count=iteration,
+            max_iterations=MAX_ITERATIONS,
+            agent_outputs=_summarize_outputs(state.get("agent_outputs", {})),
+        )
+
+        # Invoke the LLM to get a structured routing decision
+        decision: RoutingDecision = llm.invoke(
+            [SystemMessage(content=prompt)] + state["messages"]
+        )
+
+        # Route to FINISH
+        if decision.next_agent == "FINISH":
+            return Command(
+                goto="__end__",
+                update={
+                    "intent": decision.intent,
+                    "next_agent": "FINISH",
+                    "final_output": decision.delegation_notes,
+                },
+            )
+
+        # Route to a worker agent
+        return Command(
+            goto=decision.next_agent,
+            update={
+                "intent": decision.intent,
+                "next_agent": decision.next_agent,
+                "iteration_count": iteration + 1,
+            },
+        )
+
+    return supervisor_node
+
+
+def _summarize_outputs(outputs: dict[str, str]) -> str:
+    """Summarize agent outputs for the supervisor's context window."""
+    if not outputs:
+        return "(none yet)"
+    parts = []
+    for agent, output in outputs.items():
+        # Truncate long outputs to keep supervisor context manageable
+        # This mirrors Loki's summarization_threshold behavior
+        if len(output) > 2000:
+            output = output[:2000] + "... (truncated)"
+        parts.append(f"[{agent}]: {output}")
+    return "\n\n".join(parts)
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/cli.py b/examples/langchain-sisyphus/sisyphus_langchain/cli.py
new file mode 100644
index 0000000..3c1387f
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/cli.py
@@ -0,0 +1,155 @@
+"""
+CLI entry point for the Sisyphus LangChain agent.
+
+This mirrors Loki's `loki --agent sisyphus` entry point.
+
+In Loki:
+    loki --agent sisyphus
+    # Starts a REPL with the sisyphus agent loaded
+
+In this LangChain version:
+    python -m sisyphus_langchain.cli
+    # or: sisyphus (if installed via pip)
+
+Usage:
+    # Interactive REPL mode
+    sisyphus
+
+    # One-shot query
+    sisyphus "Add a health check endpoint to the API"
+
+    # With custom models
+    sisyphus --supervisor-model gpt-4o --explore-model gpt-4o-mini "Find auth patterns"
+
+Environment variables:
+    OPENAI_API_KEY      — Required for OpenAI models
+    ANTHROPIC_API_KEY   — Required if using Anthropic models
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+import uuid
+
+from langchain_core.messages import HumanMessage
+
+from sisyphus_langchain.graph import build_graph
+
+
+def run_query(graph, query: str, thread_id: str) -> str:
+    """
+    Run a single query through the Sisyphus graph.
+
+    Args:
+        graph: Compiled LangGraph.
+        query: User's natural language request.
+        thread_id: Session identifier for checkpointing.
+
+    Returns:
+        The final output string.
+    """
+    result = graph.invoke(
+        {
+            "messages": [HumanMessage(content=query)],
+            "intent": "ambiguous",
+            "next_agent": "",
+            "iteration_count": 0,
+            "todos": [],
+            "agent_outputs": {},
+            "final_output": "",
+            "project_dir": ".",
+        },
+        config={
+            "configurable": {"thread_id": thread_id},
+            "recursion_limit": 50,
+        },
+    )
+    return result.get("final_output", "(no output)")
+
+
+def repl(graph, thread_id: str) -> None:
+    """
+    Interactive REPL loop — mirrors Loki's REPL mode.
+
+    Maintains conversation across turns via the thread_id (checkpointer).
+    """
+    print("Sisyphus (LangChain) — type 'quit' to exit")
+    print("=" * 50)
+
+    while True:
+        try:
+            query = input("\n> ").strip()
+        except (EOFError, KeyboardInterrupt):
+            print("\nBye.")
+            break
+
+        if not query:
+            continue
+        if query.lower() in ("quit", "exit", "q"):
+            print("Bye.")
+            break
+
+        try:
+            output = run_query(graph, query, thread_id)
+            print(f"\n{output}")
+        except Exception as e:
+            print(f"\nError: {e}")
+
+
+def main() -> None:
+    """CLI entry point."""
+    parser = argparse.ArgumentParser(
+        description="Sisyphus — multi-agent coding orchestrator (LangChain edition)"
+    )
+    parser.add_argument(
+        "query",
+        nargs="?",
+        help="One-shot query (omit for REPL mode)",
+    )
+    parser.add_argument(
+        "--supervisor-model",
+        default="gpt-4o",
+        help="Model for the supervisor (default: gpt-4o)",
+    )
+    parser.add_argument(
+        "--explore-model",
+        default="gpt-4o-mini",
+        help="Model for the explore agent (default: gpt-4o-mini)",
+    )
+    parser.add_argument(
+        "--oracle-model",
+        default="gpt-4o",
+        help="Model for the oracle agent (default: gpt-4o)",
+    )
+    parser.add_argument(
+        "--coder-model",
+        default="gpt-4o",
+        help="Model for the coder agent (default: gpt-4o)",
+    )
+    parser.add_argument(
+        "--thread-id",
+        default=None,
+        help="Session thread ID for persistence (auto-generated if omitted)",
+    )
+
+    args = parser.parse_args()
+
+    graph = build_graph(
+        supervisor_model=args.supervisor_model,
+        explore_model=args.explore_model,
+        oracle_model=args.oracle_model,
+        coder_model=args.coder_model,
+    )
+
+    thread_id = args.thread_id or f"sisyphus-{uuid.uuid4().hex[:8]}"
+
+    if args.query:
+        output = run_query(graph, args.query, thread_id)
+        print(output)
+    else:
+        repl(graph, thread_id)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/graph.py b/examples/langchain-sisyphus/sisyphus_langchain/graph.py
new file mode 100644
index 0000000..a267e8a
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/graph.py
@@ -0,0 +1,115 @@
+"""
+Graph assembly — wires together the supervisor and worker nodes.
+
+This is the LangGraph equivalent of Loki's runtime agent execution engine
+(src/supervisor/mod.rs + src/config/request_context.rs).
+
+In Loki, the runtime:
+    1. Loads the agent config (config.yaml)
+    2. Compiles tools (tools.sh → binary)
+    3. Starts a chat loop: user → LLM → tool calls → LLM → ...
+    4. For orchestrators with can_spawn_agents: true, the supervisor module
+       manages child agent lifecycle (spawn, check, collect, cancel).
+
+In LangGraph, all of this is declarative:
+    1. Define nodes (supervisor, explore, oracle, coder)
+    2. Define edges (workers always return to supervisor)
+    3. Compile the graph (with optional checkpointer for persistence)
+    4. Invoke with initial state
+
+The graph topology:
+
+    ┌─────────────────────────────────────────────┐
+    │               SUPERVISOR                     │
+    │  (classifies intent, routes to workers)       │
+    └─────┬──────────┬──────────┬─────────────────┘
+          │          │          │
+          ▼          ▼          ▼
+      ┌────────┐ ┌────────┐ ┌────────┐
+      │EXPLORE │ │ ORACLE │ │ CODER  │
+      │(search)│ │(advise)│ │(build) │
+      └───┬────┘ └───┬────┘ └───┬────┘
+          │          │          │
+          └──────────┼──────────┘
+                     │
+              (back to supervisor)
+
+Every worker returns to the supervisor.  The supervisor decides what to do next:
+route to another worker, or end the graph.
+"""
+
+from __future__ import annotations
+
+from langgraph.checkpoint.memory import MemorySaver
+from langgraph.graph import END, START, StateGraph
+
+from sisyphus_langchain.agents.coder import create_coder_node
+from sisyphus_langchain.agents.explore import create_explore_node
+from sisyphus_langchain.agents.oracle import create_oracle_node
+from sisyphus_langchain.agents.supervisor import create_supervisor_node
+from sisyphus_langchain.state import SisyphusState
+
+
+def build_graph(
+    *,
+    supervisor_model: str = "gpt-4o",
+    explore_model: str = "gpt-4o-mini",
+    oracle_model: str = "gpt-4o",
+    coder_model: str = "gpt-4o",
+    use_checkpointer: bool = True,
+):
+    """
+    Build and compile the Sisyphus LangGraph.
+
+    This is the main entry point for creating the agent system.  It wires
+    together all nodes and edges, optionally adds a checkpointer for
+    persistence, and returns a compiled graph ready to invoke.
+
+    Args:
+        supervisor_model: Model for the routing supervisor.
+        explore_model: Model for the explore agent (can be cheaper).
+        oracle_model: Model for the oracle agent (should be strong).
+        coder_model: Model for the coder agent.
+        use_checkpointer: Whether to add MemorySaver for session persistence.
+
+    Returns:
+        A compiled LangGraph ready to .invoke() or .stream().
+
+    Model cost optimization (mirrors Loki's per-agent model config):
+        - supervisor: expensive (accurate routing is critical)
+        - explore: cheap (just searching, not reasoning deeply)
+        - oracle: expensive (deep reasoning, architecture advice)
+        - coder: expensive (writing correct code matters)
+    """
+    # Create the graph builder with our typed state
+    builder = StateGraph(SisyphusState)
+
+    # ── Register nodes ─────────────────────────────────────────────────
+    # Each node is a function that takes state and returns state updates.
+    # This mirrors Loki's agent registration (agents are discovered by
+    # their config.yaml in the agents/ directory).
+    builder.add_node("supervisor", create_supervisor_node(supervisor_model))
+    builder.add_node("explore", create_explore_node(explore_model))
+    builder.add_node("oracle", create_oracle_node(oracle_model))
+    builder.add_node("coder", create_coder_node(coder_model))
+
+    # ── Define edges ───────────────────────────────────────────────────
+    # Entry point: every invocation starts at the supervisor
+    builder.add_edge(START, "supervisor")
+
+    # Workers always return to supervisor (the hub-and-spoke pattern).
+    # In Loki, this is implicit: agent__collect returns output to the parent,
+    # and the parent (sisyphus) decides what to do next.
+    builder.add_edge("explore", "supervisor")
+    builder.add_edge("oracle", "supervisor")
+    builder.add_edge("coder", "supervisor")
+
+    # The supervisor node itself uses Command(goto=...) to route,
+    # so we don't need add_conditional_edges — the Command API
+    # handles dynamic routing internally.
+
+    # ── Compile ────────────────────────────────────────────────────────
+    checkpointer = MemorySaver() if use_checkpointer else None
+    graph = builder.compile(checkpointer=checkpointer)
+
+    return graph
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/state.py b/examples/langchain-sisyphus/sisyphus_langchain/state.py
new file mode 100644
index 0000000..52f788a
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/state.py
@@ -0,0 +1,100 @@
+"""
+Shared state schema for the Sisyphus orchestrator graph.
+
+In LangGraph, state is the single source of truth that flows through every node.
+This is analogous to Loki's per-agent session context, but unified into one typed
+dictionary that the entire graph shares.
+
+Loki Concept Mapping:
+    - Loki session context          → SisyphusState (TypedDict)
+    - Loki todo__init / todo__add   → SisyphusState.todos list
+    - Loki agent__spawn outputs     → SisyphusState.agent_outputs dict
+    - Loki intent classification    → SisyphusState.intent field
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Annotated, Literal
+
+from langchain_core.messages import BaseMessage
+from langgraph.graph.message import add_messages
+from typing_extensions import TypedDict
+
+# ---------------------------------------------------------------------------
+# Intent types — mirrors Loki's Sisyphus classification table
+# ---------------------------------------------------------------------------
+IntentType = Literal[
+    "trivial",        # Single file, known location, typo fix → handle yourself
+    "exploration",    # "Find X", "Where is Y" → spawn explore
+    "implementation", # "Add feature", "Fix bug" → spawn coder
+    "architecture",   # Design questions, oracle triggers → spawn oracle
+    "ambiguous",      # Unclear scope → ask user
+]
+
+# ---------------------------------------------------------------------------
+# Todo item — mirrors Loki's built-in todo system
+# ---------------------------------------------------------------------------
+@dataclass
+class TodoItem:
+    """A single task in the orchestrator's todo list."""
+    id: int
+    task: str
+    done: bool = False
+
+
+def _merge_todos(existing: list[TodoItem], new: list[TodoItem]) -> list[TodoItem]:
+    """
+    Reducer for the todos field.
+
+    LangGraph requires a reducer for any state field that can be written by
+    multiple nodes.  This merges by id: if a todo with the same id already
+    exists, the incoming version wins (allows marking done).
+    """
+    by_id = {t.id: t for t in existing}
+    for t in new:
+        by_id[t.id] = t
+    return list(by_id.values())
+
+
+# ---------------------------------------------------------------------------
+# Core graph state
+# ---------------------------------------------------------------------------
+class SisyphusState(TypedDict):
+    """
+    The shared state that flows through every node in the Sisyphus graph.
+
+    Annotated fields use *reducers* — functions that merge concurrent writes.
+    Without reducers, parallel node outputs would overwrite each other.
+    """
+
+    # Conversation history — the `add_messages` reducer appends new messages
+    # instead of replacing the list.  This is critical: every node adds its
+    # response here, and downstream nodes see the full history.
+    #
+    # Loki equivalent: each agent's chat session accumulates messages the same
+    # way, but messages are scoped per-agent.  In LangGraph the shared message
+    # list IS the inter-agent communication channel.
+    messages: Annotated[list[BaseMessage], add_messages]
+
+    # Classified intent for the current request
+    intent: IntentType
+
+    # Which agent the supervisor routed to last
+    next_agent: str
+
+    # Iteration counter — safety valve analogous to Loki's max_auto_continues
+    iteration_count: int
+
+    # Todo list for multi-step tracking (mirrors Loki's todo__* tools)
+    todos: Annotated[list[TodoItem], _merge_todos]
+
+    # Accumulated outputs from sub-agent nodes, keyed by agent name.
+    # The supervisor reads these to decide what to do next.
+    agent_outputs: dict[str, str]
+
+    # Final synthesized answer to return to the user
+    final_output: str
+
+    # The working directory / project path (mirrors Loki's project_dir variable)
+    project_dir: str
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/tools/__init__.py b/examples/langchain-sisyphus/sisyphus_langchain/tools/__init__.py
new file mode 100644
index 0000000..9c1ffa6
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/tools/__init__.py
@@ -0,0 +1 @@
+"""Tool definitions for Sisyphus agents."""
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/tools/filesystem.py b/examples/langchain-sisyphus/sisyphus_langchain/tools/filesystem.py
new file mode 100644
index 0000000..9a50cd4
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/tools/filesystem.py
@@ -0,0 +1,175 @@
+"""
+Filesystem tools for Sisyphus agents.
+
+These are the LangChain equivalents of Loki's global tools:
+    - fs_read.sh   → read_file
+    - fs_grep.sh   → search_content
+    - fs_glob.sh   → search_files
+    - fs_ls.sh     → list_directory
+    - fs_write.sh  → write_file
+    - fs_patch.sh  → (omitted — write_file covers full rewrites)
+
+Loki Concept Mapping:
+    Loki tools are bash scripts with @cmd annotations that Loki's compiler
+    turns into function-calling declarations.  In LangChain, we use the @tool
+    decorator which serves the same purpose: it generates the JSON schema
+    that the LLM sees, and wraps the Python function for execution.
+"""
+
+from __future__ import annotations
+
+import fnmatch
+import os
+import re
+import subprocess
+
+from langchain_core.tools import tool
+
+
+@tool
+def read_file(path: str, offset: int = 1, limit: int = 200) -> str:
+    """Read a file's contents with optional line range.
+
+    Args:
+        path: Path to the file (absolute or relative to cwd).
+        offset: 1-based line number to start from.
+        limit: Maximum number of lines to return.
+    """
+    path = os.path.expanduser(path)
+    if not os.path.isfile(path):
+        return f"Error: file not found: {path}"
+
+    try:
+        with open(path, "r", encoding="utf-8", errors="replace") as f:
+            lines = f.readlines()
+    except Exception as e:
+        return f"Error reading {path}: {e}"
+
+    total = len(lines)
+    start = max(0, offset - 1)
+    end = min(total, start + limit)
+    selected = lines[start:end]
+
+    result = f"File: {path}  (lines {start + 1}-{end} of {total})\n\n"
+    for i, line in enumerate(selected, start=start + 1):
+        result += f"{i}: {line}"
+
+    if end < total:
+        result += f"\n... truncated ({total} total lines)"
+
+    return result
+
+
+@tool
+def write_file(path: str, content: str) -> str:
+    """Write complete contents to a file, creating parent directories as needed.
+
+    Args:
+        path: Path for the file.
+        content: Complete file contents to write.
+    """
+    path = os.path.expanduser(path)
+    os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
+    try:
+        with open(path, "w", encoding="utf-8") as f:
+            f.write(content)
+        return f"Wrote: {path}"
+    except Exception as e:
+        return f"Error writing {path}: {e}"
+
+
+@tool
+def search_content(pattern: str, directory: str = ".", file_type: str = "") -> str:
+    """Search for a text/regex pattern in files under a directory.
+
+    Args:
+        pattern: Text or regex pattern to search for.
+        directory: Root directory to search in.
+        file_type: Optional file extension filter (e.g. "py", "rs").
+    """
+    directory = os.path.expanduser(directory)
+    cmd = ["grep", "-rn"]
+    if file_type:
+        cmd += [f"--include=*.{file_type}"]
+    cmd += [pattern, directory]
+
+    try:
+        result = subprocess.run(cmd, capture_output=True, text=True, timeout=30)
+        lines = result.stdout.strip().splitlines()
+    except Exception as e:
+        return f"Error: {e}"
+
+    # Filter noise
+    noise = {"/.git/", "/node_modules/", "/target/", "/dist/", "/__pycache__/"}
+    filtered = [l for l in lines if not any(n in l for n in noise)][:30]
+
+    if not filtered:
+        return "No matches found."
+    return "\n".join(filtered)
+
+
+@tool
+def search_files(pattern: str, directory: str = ".") -> str:
+    """Find files matching a glob pattern.
+
+    Args:
+        pattern: Glob pattern (e.g. '*.py', 'config*', '*test*').
+        directory: Directory to search in.
+    """
+    directory = os.path.expanduser(directory)
+    noise = {".git", "node_modules", "target", "dist", "__pycache__"}
+    matches: list[str] = []
+
+    for root, dirs, files in os.walk(directory):
+        dirs[:] = [d for d in dirs if d not in noise]
+        for name in files:
+            if fnmatch.fnmatch(name, pattern):
+                matches.append(os.path.join(root, name))
+                if len(matches) >= 25:
+                    break
+        if len(matches) >= 25:
+            break
+
+    if not matches:
+        return "No files found."
+    return "\n".join(matches)
+
+
+@tool
+def list_directory(path: str = ".", max_depth: int = 3) -> str:
+    """List directory tree structure.
+
+    Args:
+        path: Directory to list.
+        max_depth: Maximum depth to recurse.
+    """
+    path = os.path.expanduser(path)
+    if not os.path.isdir(path):
+        return f"Error: not a directory: {path}"
+
+    noise = {".git", "node_modules", "target", "dist", "__pycache__", ".venv", "venv"}
+    lines: list[str] = []
+
+    def _walk(dir_path: str, prefix: str, depth: int) -> None:
+        if depth > max_depth:
+            return
+        try:
+            entries = sorted(os.listdir(dir_path))
+        except PermissionError:
+            return
+
+        dirs = [e for e in entries if os.path.isdir(os.path.join(dir_path, e)) and e not in noise]
+        files = [e for e in entries if os.path.isfile(os.path.join(dir_path, e))]
+
+        for f in files[:20]:
+            lines.append(f"{prefix}{f}")
+        if len(files) > 20:
+            lines.append(f"{prefix}... ({len(files) - 20} more files)")
+
+        for d in dirs:
+            lines.append(f"{prefix}{d}/")
+            _walk(os.path.join(dir_path, d), prefix + "  ", depth + 1)
+
+    lines.append(f"{os.path.basename(path) or path}/")
+    _walk(path, "  ", 1)
+    return "\n".join(lines[:200])
diff --git a/examples/langchain-sisyphus/sisyphus_langchain/tools/project.py b/examples/langchain-sisyphus/sisyphus_langchain/tools/project.py
new file mode 100644
index 0000000..763ef32
--- /dev/null
+++ b/examples/langchain-sisyphus/sisyphus_langchain/tools/project.py
@@ -0,0 +1,142 @@
+"""
+Project detection and build/test tools.
+
+These mirror Loki's .shared/utils.sh detect_project() heuristic and the
+sisyphus/coder tools.sh run_build / run_tests / verify_build commands.
+
+Loki Concept Mapping:
+    Loki uses a heuristic cascade: check for Cargo.toml → go.mod → package.json
+    etc., then falls back to an LLM call for unknown projects.  We replicate the
+    heuristic portion here.  The LLM fallback is omitted since the agents
+    themselves can reason about unknown project types.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import subprocess
+
+from langchain_core.tools import tool
+
+# ---------------------------------------------------------------------------
+# Project detection (mirrors _detect_heuristic in utils.sh)
+# ---------------------------------------------------------------------------
+_HEURISTICS: list[tuple[str, dict[str, str]]] = [
+    ("Cargo.toml", {"type": "rust", "build": "cargo build", "test": "cargo test", "check": "cargo check"}),
+    ("go.mod", {"type": "go", "build": "go build ./...", "test": "go test ./...", "check": "go vet ./..."}),
+    ("package.json", {"type": "nodejs", "build": "npm run build", "test": "npm test", "check": "npm run lint"}),
+    ("pyproject.toml", {"type": "python", "build": "", "test": "pytest", "check": "ruff check ."}),
+    ("pom.xml", {"type": "java", "build": "mvn compile", "test": "mvn test", "check": "mvn verify"}),
+    ("Makefile", {"type": "make", "build": "make build", "test": "make test", "check": "make lint"}),
+]
+
+
+def detect_project(directory: str) -> dict[str, str]:
+    """Detect project type and return build/test commands."""
+    for marker, info in _HEURISTICS:
+        if os.path.exists(os.path.join(directory, marker)):
+            return info
+    return {"type": "unknown", "build": "", "test": "", "check": ""}
+
+
+@tool
+def get_project_info(directory: str = ".") -> str:
+    """Detect the project type and show structure overview.
+
+    Args:
+        directory: Project root directory.
+    """
+    directory = os.path.expanduser(directory)
+    info = detect_project(directory)
+    result = f"Project: {os.path.abspath(directory)}\n"
+    result += f"Type: {info['type']}\n"
+    result += f"Build: {info['build'] or '(none)'}\n"
+    result += f"Test: {info['test'] or '(none)'}\n"
+    result += f"Check: {info['check'] or '(none)'}\n"
+    return result
+
+
+def _run_project_command(directory: str, command_key: str) -> str:
+    """Run a detected project command (build/test/check)."""
+    directory = os.path.expanduser(directory)
+    info = detect_project(directory)
+    cmd = info.get(command_key, "")
+
+    if not cmd:
+        return f"No {command_key} command detected for this project."
+
+    try:
+        result = subprocess.run(
+            cmd,
+            shell=True,
+            capture_output=True,
+            text=True,
+            cwd=directory,
+            timeout=300,
+        )
+        output = result.stdout + result.stderr
+        status = "SUCCESS" if result.returncode == 0 else f"FAILED (exit {result.returncode})"
+        return f"Running: {cmd}\n\n{output}\n\n{command_key.upper()}: {status}"
+    except subprocess.TimeoutExpired:
+        return f"{command_key.upper()}: TIMEOUT after 300s"
+    except Exception as e:
+        return f"{command_key.upper()}: ERROR — {e}"
+
+
+@tool
+def run_build(directory: str = ".") -> str:
+    """Run the project's build command.
+
+    Args:
+        directory: Project root directory.
+    """
+    return _run_project_command(directory, "build")
+
+
+@tool
+def run_tests(directory: str = ".") -> str:
+    """Run the project's test suite.
+
+    Args:
+        directory: Project root directory.
+    """
+    return _run_project_command(directory, "test")
+
+
+@tool
+def verify_build(directory: str = ".") -> str:
+    """Run the project's check/lint command to verify correctness.
+
+    Args:
+        directory: Project root directory.
+    """
+    return _run_project_command(directory, "check")
+
+
+@tool
+def execute_command(command: str, directory: str = ".") -> str:
+    """Execute a shell command and return its output.
+
+    Args:
+        command: Shell command to execute.
+        directory: Working directory.
+    """
+    directory = os.path.expanduser(directory)
+    try:
+        result = subprocess.run(
+            command,
+            shell=True,
+            capture_output=True,
+            text=True,
+            cwd=directory,
+            timeout=120,
+        )
+        output = (result.stdout + result.stderr).strip()
+        if result.returncode != 0:
+            return f"Command failed (exit {result.returncode}):\n{output}"
+        return output or "(no output)"
+    except subprocess.TimeoutExpired:
+        return "Command timed out after 120s."
+    except Exception as e:
+        return f"Error: {e}"