97 lines
3.6 KiB
Python
97 lines
3.6 KiB
Python
|
from langchain_core.messages import SystemMessage, HumanMessage
|
||
|
|
from state import OrchestratorState
|
||
|
|
from models import (
|
||
|
|
get_implementer,
|
||
|
|
get_reviewer,
|
||
|
|
get_researcher,
|
||
|
|
get_cross_reviewer,
|
||
|
|
get_scanner,
|
||
|
|
get_fast_coder,
|
||
|
|
)
|
||
|
|
|
||
|
|
IMPLEMENTER_PROMPT = """You are an implementation agent. You write clean, production-ready code.
|
||
|
|
Follow the spec exactly. No over-engineering, no unnecessary abstractions.
|
||
|
|
Return the complete implementation with file paths."""
|
||
|
|
|
||
|
|
REVIEWER_PROMPT = """You are a code review agent. Review code changes for correctness, security, and maintainability.
|
||
|
|
|
||
|
|
For each issue found, categorize it:
|
||
|
|
- BLOCK — Must fix before merge. Security vulnerabilities, data loss risks, broken logic.
|
||
|
|
- FIX — Should fix. Bugs, performance issues, missing error handling at boundaries.
|
||
|
|
- NIT — Optional. Style, naming, minor improvements.
|
||
|
|
- QUESTION — Needs clarification. Intent is unclear.
|
||
|
|
|
||
|
|
Start with a one-line summary: APPROVE, REQUEST CHANGES, or NEEDS DISCUSSION.
|
||
|
|
Then list findings grouped by category (BLOCK > FIX > NIT > QUESTION).
|
||
|
|
If the code is fine, say "No issues found." No praise or filler."""
|
||
|
|
|
||
|
|
RESEARCHER_PROMPT = """You are a research agent. Look up documentation, API references, and technical answers.
|
||
|
|
Be concise and cite sources. Return factual information, not opinions."""
|
||
|
|
|
||
|
|
CROSS_REVIEWER_PROMPT = """You are a cross-family code review agent. You provide an independent review perspective.
|
||
|
|
Review the provided code for bugs, security issues, and improvements.
|
||
|
|
Categorize findings as BLOCK, FIX, or NIT. Be concise.
|
||
|
|
Focus on issues that might be missed by the primary development team."""
|
||
|
|
|
||
|
|
SCANNER_PROMPT = """You are a large-context analysis agent. You analyze codebases, identify patterns,
|
||
|
|
check consistency across files, and find structural issues.
|
||
|
|
Summarize findings concisely with file references."""
|
||
|
|
|
||
|
|
FAST_CODER_PROMPT = """You are a fast coding agent for well-specified tasks.
|
||
|
|
Implement exactly what is asked. No extras, no refactoring beyond scope.
|
||
|
|
Return complete, working code."""
|
||
|
|
|
||
|
|
|
||
|
|
def implementer_node(state: OrchestratorState) -> dict:
|
||
|
|
llm = get_implementer()
|
||
|
|
response = llm.invoke([
|
||
|
|
SystemMessage(content=IMPLEMENTER_PROMPT),
|
||
|
|
HumanMessage(content=state["task"]),
|
||
|
|
])
|
||
|
|
return {"result": response.content, "messages": [response]}
|
||
|
|
|
||
|
|
|
||
|
|
def reviewer_node(state: OrchestratorState) -> dict:
|
||
|
|
llm = get_reviewer()
|
||
|
|
response = llm.invoke([
|
||
|
|
SystemMessage(content=REVIEWER_PROMPT),
|
||
|
|
HumanMessage(content=state["task"]),
|
||
|
|
])
|
||
|
|
return {"result": response.content, "messages": [response]}
|
||
|
|
|
||
|
|
|
||
|
|
def researcher_node(state: OrchestratorState) -> dict:
|
||
|
|
llm = get_researcher()
|
||
|
|
response = llm.invoke([
|
||
|
|
SystemMessage(content=RESEARCHER_PROMPT),
|
||
|
|
HumanMessage(content=state["task"]),
|
||
|
|
])
|
||
|
|
return {"result": response.content, "messages": [response]}
|
||
|
|
|
||
|
|
|
||
|
|
def cross_reviewer_node(state: OrchestratorState) -> dict:
|
||
|
|
llm = get_cross_reviewer()
|
||
|
|
response = llm.invoke([
|
||
|
|
SystemMessage(content=CROSS_REVIEWER_PROMPT),
|
||
|
|
HumanMessage(content=state["task"]),
|
||
|
|
])
|
||
|
|
return {"result": response.content, "messages": [response]}
|
||
|
|
|
||
|
|
|
||
|
|
def scanner_node(state: OrchestratorState) -> dict:
|
||
|
|
llm = get_scanner()
|
||
|
|
response = llm.invoke([
|
||
|
|
SystemMessage(content=SCANNER_PROMPT),
|
||
|
|
HumanMessage(content=state["task"]),
|
||
|
|
])
|
||
|
|
return {"result": response.content, "messages": [response]}
|
||
|
|
|
||
|
|
|
||
|
|
def fast_coder_node(state: OrchestratorState) -> dict:
|
||
|
|
llm = get_fast_coder()
|
||
|
|
response = llm.invoke([
|
||
|
|
SystemMessage(content=FAST_CODER_PROMPT),
|
||
|
|
HumanMessage(content=state["task"]),
|
||
|
|
])
|
||
|
|
return {"result": response.content, "messages": [response]}
|