feat(adapters): added LangGraph and DeepAgents export adapters

README.md

@@ -367,6 +367,8 @@ Adapters are used by both `export` and `run`. Available adapters:
 | `openclaw` | OpenClaw format |
 | `nanobot` | Nanobot format |
 | `cursor` | Cursor `.cursor/rules/*.mdc` files |
+| `langgraph` | LangGraph `StateGraph` Python module (skills → nodes, skillflows → edges) |
+| `deepagents` | LangChain DeepAgents harness (`create_deep_agent(...)`) — skills, tools, sub-agents |
 
 ```bash
 # Export to system prompt

examples/deepagents/README.md

@@ -0,0 +1,30 @@
+# DeepAgents example
+
+A research agent with a fact-checker sub-agent, demonstrating the DeepAgents adapter:
+
+- `agent.yaml` — manifest (model, skills, tools, sub-agent declaration)
+- `SOUL.md` — agent identity, embedded in the generated `system_prompt`
+- `skills/web-research/`, `skills/summarize/` — passed via `skills=["./skills"]`; DeepAgents loads each `SKILL.md` natively
+- `tools/web-search.yaml` — emitted as a `@tool` function bound into `tools=[...]`
+- `agents/fact-checker/` — emitted as a `SubAgent` dict in `subagents=[...]`
+- `expected_output.py` — the Python module the adapter produces
+
+DeepAgents is a higher-level harness on top of LangGraph — there is no graph
+wiring: the model decides when to plan, when to delegate to sub-agents, and
+when to call tools. If you need explicit per-step edges, use the `langgraph`
+adapter instead.
+
+## Regenerate
+
+```bash
+gapman export --dir examples/deepagents --format deepagents --output examples/deepagents/expected_output.py
+```
+
+## Run the generated agent
+
+```bash
+pip install deepagents langchain-anthropic
+python examples/deepagents/expected_output.py
+```
+
+The generated file leaves tool implementations as `NotImplementedError` stubs — replace them with your own logic before invoking.

examples/deepagents/SOUL.md

@@ -0,0 +1,5 @@
+# Research Assistant
+
+You are a research assistant. Find authoritative sources, extract the relevant
+facts, and summarize them honestly. Delegate to the `fact-checker` sub-agent
+whenever a claim warrants independent verification. Cite every claim.

examples/deepagents/agent.yaml

@@ -0,0 +1,16 @@
+spec_version: "0.1.0"
+name: research-assistant
+version: 1.0.0
+description: Research assistant with a fact-checker sub-agent
+author: gitagent-examples
+license: MIT
+model:
+  preferred: anthropic:claude-sonnet-4-5
+skills:
+  - web-research
+  - summarize
+tools:
+  - web-search
+agents:
+  fact-checker:
+    description: Verifies factual claims against authoritative sources

examples/deepagents/agents/fact-checker/SOUL.md

@@ -0,0 +1,4 @@
+# Fact-Checker
+
+You are a pedantic fact-checker. Treat every claim as unverified until you
+locate a primary source. If a claim cannot be substantiated, say so plainly.

examples/deepagents/agents/fact-checker/agent.yaml

@@ -0,0 +1,4 @@
+spec_version: "0.1.0"
+name: fact-checker
+version: 1.0.0
+description: Verifies factual claims against authoritative sources

examples/deepagents/expected_output.py

@@ -0,0 +1,83 @@
+"""
+DeepAgents definition for research-assistant v1.0.0
+Generated by gitagent export --format deepagents
+"""
+
+from __future__ import annotations
+
+from deepagents import create_deep_agent
+from langchain_core.tools import tool
+
+# Agent metadata
+AGENT_NAME = "research-assistant"
+AGENT_VERSION = "1.0.0"
+MODEL = "anthropic:claude-sonnet-4-5"
+
+# System prompt (SOUL.md + RULES.md + DUTIES.md + compliance)
+SYSTEM_PROMPT = """# research-assistant
+Research assistant with a fact-checker sub-agent
+
+# Research Assistant
+
+You are a research assistant. Find authoritative sources, extract the relevant
+facts, and summarize them honestly. Delegate to the `fact-checker` sub-agent
+whenever a claim warrants independent verification. Cite every claim.
+"""
+
+# Hooks (pre_tool_use scripts run before every tool call)
+def _run_pre_tool_use_hooks(tool_name: str) -> None:
+    """No pre_tool_use hooks configured in hooks/hooks.yaml."""
+    return None
+
+# NOTE: other hook events (post_tool_use, on_session_start, etc.) have no
+# DeepAgents equivalent and are intentionally skipped.
+# NOTE: memory/ has no direct DeepAgents equivalent — wire a checkpointer if needed.
+
+# Tools (from tools/*.yaml)
+@tool
+def web_search(query: str) -> str:
+    """Search the public web for a query"""
+    _run_pre_tool_use_hooks("web-search")
+    raise NotImplementedError("Implement tool: web-search")
+
+TOOLS = [web_search]
+
+# Skills (skills/<name>/SKILL.md — DeepAgents loads these natively)
+# Pointing skills= at the directory lets DeepAgents discover every SKILL.md
+# without us having to inline the skill content into SYSTEM_PROMPT.
+SKILLS = ["./skills"]
+
+# For reference, the skills available in this agent:
+#   - summarize: Condense the gathered sources into a faithful, cited summary
+#   - web-research: Search the web for sources relevant to the user's question
+
+# Sub-agents (agents/<name>/ → SubAgent dicts)
+fact_checker_subagent = {
+    "name": "fact-checker",
+    "description": "Verifies factual claims against authoritative sources",
+    "system_prompt": """# fact-checker
+Verifies factual claims against authoritative sources
+
+# Fact-Checker
+
+You are a pedantic fact-checker. Treat every claim as unverified until you
+locate a primary source. If a claim cannot be substantiated, say so plainly.
+""",
+    "tools": TOOLS,
+}
+
+SUBAGENTS = [fact_checker_subagent]
+
+# Agent
+agent = create_deep_agent(
+    model=MODEL,
+    system_prompt=SYSTEM_PROMPT,
+    tools=TOOLS,
+    skills=SKILLS,
+    subagents=SUBAGENTS,
+)
+
+if __name__ == "__main__":
+    result = agent.invoke({"messages": [{"role": "user", "content": "Hello"}]})
+    for message in result["messages"]:
+        print(message)

examples/deepagents/skills/summarize/SKILL.md

@@ -0,0 +1,8 @@
+---
+name: summarize
+description: Condense the gathered sources into a faithful, cited summary
+---
+
+Produce a concise summary of the gathered material. Every factual claim must
+be attributed to a specific source URL. If sources disagree, surface the
+disagreement explicitly.

examples/deepagents/skills/web-research/SKILL.md

@@ -0,0 +1,8 @@
+---
+name: web-research
+description: Search the web for sources relevant to the user's question
+allowed-tools: web-search
+---
+
+Use the `web-search` tool to find authoritative sources. Capture the URL of
+every source consulted; downstream skills will cite them.

examples/deepagents/tools/web-search.yaml

@@ -0,0 +1,10 @@
+name: web-search
+description: Search the public web for a query
+input_schema:
+  type: object
+  properties:
+    query:
+      type: string
+      description: The search query
+  required:
+    - query

examples/langgraph/README.md

@@ -0,0 +1,26 @@
+# LangGraph example
+
+A two-step research agent that demonstrates the LangGraph adapter:
+
+- `agent.yaml` — manifest (model, runtime, skill + tool references)
+- `SOUL.md` — agent identity, embedded in the generated system prompt
+- `skills/web-research/SKILL.md` — first node in the graph
+- `skills/summarize/SKILL.md` — second node, depends on `web-research`
+- `skillflows/research.yaml` — wiring (`steps`, `depends_on`) → `add_edge` calls
+- `tools/web-search.yaml` — bound as a `ToolNode`
+- `expected_output.py` — the Python module the adapter produces
+
+## Regenerate
+
+```bash
+gapman export --dir examples/langgraph --format langgraph --output examples/langgraph/expected_output.py
+```
+
+## Run the generated graph
+
+```bash
+pip install "langgraph>=0.2" "langchain>=0.3" "langchain-core>=0.3" langchain-anthropic
+python examples/langgraph/expected_output.py
+```
+
+The generated file leaves tool implementations as `NotImplementedError` stubs — replace them with your own logic before invoking.

examples/langgraph/SOUL.md

@@ -0,0 +1,4 @@
+# Research Assistant
+
+You are a research assistant. Find authoritative sources, extract the relevant
+facts, and summarize them honestly. Cite every claim. When uncertain, say so.

examples/langgraph/agent.yaml

@@ -0,0 +1,15 @@
+spec_version: "0.1.0"
+name: research-assistant
+version: 1.0.0
+description: Two-step research assistant — searches the web, then summarizes findings
+author: gitagent-examples
+license: MIT
+model:
+  preferred: claude-sonnet-4-5
+runtime:
+  max_turns: 25
+skills:
+  - web-research
+  - summarize
+tools:
+  - web-search

examples/langgraph/expected_output.py

@@ -0,0 +1,109 @@
+"""
+LangGraph definition for research-assistant v1.0.0
+Generated by gitagent export --format langgraph
+"""
+
+from __future__ import annotations
+
+from typing import Annotated, TypedDict
+
+from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage
+from langchain_core.tools import tool
+from langchain.chat_models import init_chat_model
+from langgraph.graph import END, START, StateGraph
+from langgraph.graph.message import add_messages
+from langgraph.prebuilt import ToolNode
+
+# Agent metadata
+AGENT_NAME = "research-assistant"
+AGENT_VERSION = "1.0.0"
+AGENT_DESCRIPTION = "Two-step research assistant — searches the web, then summarizes findings"
+MODEL = "claude-sonnet-4-5"
+RECURSION_LIMIT = 25
+
+# System prompt (SOUL.md + RULES.md + DUTIES.md + compliance)
+SYSTEM_PROMPT = """# research-assistant
+Two-step research assistant — searches the web, then summarizes findings
+
+# Research Assistant
+
+You are a research assistant. Find authoritative sources, extract the relevant
+facts, and summarize them honestly. Cite every claim. When uncertain, say so.
+"""
+
+# Skill instructions
+SKILL_SUMMARIZE_INSTRUCTIONS = """Condense the gathered sources into a faithful, cited summary
+
+Produce a concise summary of the gathered material. Every factual claim must
+be attributed to a specific source URL. If the sources disagree, say so
+explicitly rather than picking a side."""
+
+SKILL_WEB_RESEARCH_INSTRUCTIONS = """Search the web for sources relevant to the user's question
+Allowed tools: web-search
+
+Use the `web-search` tool to find authoritative sources. Prefer primary sources
+(official docs, academic papers) over secondary commentary. Capture the URL of
+every source consulted."""
+
+# Graph state
+class AgentState(TypedDict):
+    messages: Annotated[list[BaseMessage], add_messages]
+
+# Tools (from tools/*.yaml)
+@tool
+def web_search(query: str) -> str:
+    """Search the public web for a query"""
+    raise NotImplementedError("Implement tool: web-search")
+
+TOOLS = [web_search]
+tool_node = ToolNode(TOOLS)
+
+llm = init_chat_model(MODEL)
+llm_with_tools = llm.bind_tools(TOOLS) if TOOLS else llm
+
+# Hooks (pre_tool_use → before_tool callback)
+def before_tool(state: AgentState) -> AgentState:
+    """No pre_tool_use hooks configured in hooks/hooks.yaml."""
+    return state
+
+# NOTE: post_tool_use, on_session_start, etc. have no LangGraph equivalent — skipped.
+# NOTE: memory/ has no direct LangGraph equivalent — implement via a checkpointer.
+
+# Skill nodes (one per skills/<name>/SKILL.md)
+def skill_summarize(state: AgentState) -> dict:
+    """Skill node: summarize — Condense the gathered sources into a faithful, cited summary"""
+    prompt = SYSTEM_PROMPT + "\n\n" + SKILL_SUMMARIZE_INSTRUCTIONS
+    messages = [SystemMessage(content=prompt), *state["messages"]]
+    response = llm_with_tools.invoke(messages)
+    return {"messages": [response]}
+
+def skill_web_research(state: AgentState) -> dict:
+    """Skill node: web-research — Search the web for sources relevant to the user's question"""
+    prompt = SYSTEM_PROMPT + "\n\n" + SKILL_WEB_RESEARCH_INSTRUCTIONS
+    messages = [SystemMessage(content=prompt), *state["messages"]]
+    response = llm_with_tools.invoke(messages)
+    return {"messages": [response]}
+
+# Graph construction
+graph = StateGraph(AgentState)
+
+graph.add_node("tools", tool_node)
+
+# Skillflow: research
+graph.add_node("research", skill_web_research)
+graph.add_node("summarize", skill_summarize)
+graph.add_edge(START, "research")
+graph.add_edge("research", "summarize")
+graph.add_edge("summarize", END)
+
+
+# Compile
+app = graph.compile()
+
+if __name__ == "__main__":
+    result = app.invoke(
+        {"messages": [HumanMessage(content="Hello")]},
+        config={"recursion_limit": RECURSION_LIMIT},
+    )
+    for message in result["messages"]:
+        print(message)

examples/langgraph/skillflows/research.yaml

@@ -0,0 +1,8 @@
+name: research
+description: Run web-research, then summarize the result
+steps:
+  - id: research
+    skill: web-research
+  - id: summarize
+    skill: summarize
+    depends_on: [research]

examples/langgraph/skills/summarize/SKILL.md

@@ -0,0 +1,8 @@
+---
+name: summarize
+description: Condense the gathered sources into a faithful, cited summary
+---
+
+Produce a concise summary of the gathered material. Every factual claim must
+be attributed to a specific source URL. If the sources disagree, say so
+explicitly rather than picking a side.

examples/langgraph/skills/web-research/SKILL.md

@@ -0,0 +1,9 @@
+---
+name: web-research
+description: Search the web for sources relevant to the user's question
+allowed-tools: web-search
+---
+
+Use the `web-search` tool to find authoritative sources. Prefer primary sources
+(official docs, academic papers) over secondary commentary. Capture the URL of
+every source consulted.

examples/langgraph/tools/web-search.yaml

@@ -0,0 +1,10 @@
+name: web-search
+description: Search the public web for a query
+input_schema:
+  type: object
+  properties:
+    query:
+      type: string
+      description: The search query
+  required:
+    - query