Switch from gpt-oss to granite 4, default config updates

Signed-off-by: Kelly Abuelsaad <kaymar@gmail.com>

Author: kellyaa

a2a/git_issue_agent/git_issue_agent/agents.py

@@ -1,6 +1,6 @@
 from crewai import Agent, Crew, Process, Task
 from git_issue_agent.config import Settings
-from git_issue_agent.data_types import RepositoryJudgement
+from git_issue_agent.data_types import UserQueryJudgement
 from git_issue_agent.llm import CrewLLM
 from git_issue_agent.prompts import REPO_ID_BACKSTORY, TOOL_CALL_PROMPT
 class GitAgents():
@@ -12,8 +12,8 @@ def __init__(self, config: Settings, issue_tools):
         # Pre-requisitie validator
         # ##################
         self.prereq_identifier = Agent(
-            role="GitHub Issue Analyst",
-            goal="To determine whether a user has supplied enough information to find issues contained in Github repositories.",
+            role="Pre-requisite Judge",
+            goal="To determine whether a user has supplied enough information to find Github issues.",
             backstory=REPO_ID_BACKSTORY,
             verbose=True,
             llm=self.llm.llm
@@ -24,9 +24,9 @@ def __init__(self, config: Settings, issue_tools):
                 "User query: {request}"
             ),
             agent=self.prereq_identifier,
-            output_pydantic=RepositoryJudgement,
+            output_pydantic=UserQueryJudgement,
             expected_output=(
-                "A judgement on whether a request from a user successfully identifies both a repository owner/organization and repository name, accompanied by an explanation"
+                "A judgement on whether a request from a user successfully identifies an organization or user that owns issues."
             ),
         )
 
@@ -43,13 +43,14 @@ def __init__(self, config: Settings, issue_tools):
         self.issue_researcher = Agent(
             role="GitHub Issue Analyst",
             goal=(
-                "You are connected to GitHub's MCP server and specialize in exploring and summarizing repository issues. "
+                "Answer the user's query, using tools provided from MCP server. "
                 "Prefer read-only operations. When querying, be explicit about repo owner/name and filters."
             ),
             backstory=TOOL_CALL_PROMPT,
             tools=issue_tools,
             verbose=True,
             llm=self.llm.llm,
+            inject_date=True,
             max_iter=3
         )
 
@@ -58,17 +59,11 @@ def __init__(self, config: Settings, issue_tools):
         self.issue_query_task = Task(
             description=(
                 "Retrieve Github issues using tool calls in order to answer the user's query.\n"
-                "Instructions:\n"
-                "1) Identify the criteria they specify such as username, organization and/or repository name.\n"
-                "2) If the user provides filters (e.g., state=open, label=bug, assignee=alice, or search text), apply them.\n"
-                "3) Return a clean, numbered summary: issue number, title, state, labels, assignee(s), and direct URL.\n"
-                "4) Prefer listing or searching for issues over the get_issues API unless the user gives you a specific issue number"
                 "User query: {request}"
             ),
             agent=self.issue_researcher,
             expected_output=(
-                "A concise report (Markdown allowed) listing matching issues with links and key metadata, "
-                "or a brief explanation if nothing matches."
+                "A direct answer to the user's query, citing the output of the tool to support your answer. Provide as many details as possible to support your claim."
             ),
         )
 

a2a/git_issue_agent/git_issue_agent/config.py

@@ -12,11 +12,11 @@ class Settings(BaseSettings):
         description="Application log level",       
     )
     TASK_MODEL_ID: str = Field(
-        os.getenv("TASK_MODEL_ID", "granite3.3:8b"),
+        os.getenv("TASK_MODEL_ID", "ollama/ibm/granite4:latest"),
         description="The ID of the task model",
     )
     LLM_API_BASE: str = Field(
-        os.getenv("LLM_API_BASE", "http://localhost:11434/v1"),
+        os.getenv("LLM_API_BASE", "http://host.docker.internal:11434"),
         description="The URL for OpenAI API",
     )
     LLM_API_KEY: str = Field(os.getenv("LLM_API_KEY", "my_api_key"), description="The key for OpenAI API")
@@ -27,7 +27,7 @@ class Settings(BaseSettings):
         ge=0,
     )
     MCP_URL: str = Field(os.getenv("MCP_URL", "https://api.githubcopilot.com/mcp/"), description="Endpoint for an option MCP server")
-    SERVICE_PORT: int = Field(os.getenv("SERVICE_URL", 8000), description="Port on which the service will run.")
+    SERVICE_PORT: int = Field(os.getenv("SERVICE_PORT", 8000), description="Port on which the service will run.")
     GITHUB_TOKEN: str = Field(os.getenv("GITHUB_TOKEN", None), description="If not using agent with authorization, the default Github token to use")
 
     # auth variables for token validation

a2a/git_issue_agent/git_issue_agent/data_types.py

@@ -6,6 +6,6 @@
 ############
 
 
-class RepositoryJudgement(BaseModel):
-    is_owner_and_repo_identified: bool = Field(False, description="Whether the user's query indicates both an owner/organization name and the repo name")
+class UserQueryJudgement(BaseModel):
+    has_sufficient_information: bool = Field(False, description="Whether the user's query contains sufficient enough information to retrieve issues from github")
     explanation: str = Field(None, description="A detailed explanation as to why you made the judgement")

a2a/git_issue_agent/git_issue_agent/main.py

@@ -29,14 +29,29 @@ async def _send_event(self, message: str, final: bool = False):
             await self.eventer.emit_event(message, final)
         else:
             logger.warning("No event handler registered")
+    
+    def extract_user_input(self, body):
+        content = body[-1]["content"]
+        latest_content = ""
+
+        if isinstance(content, str):
+            latest_content = content
+        else:
+            for item in content:
+                if item["type"] == "text":
+                    latest_content += item["text"]
+                else:
+                    self.logger.warning(f"Ignoring content with type {item['type']}")
 
-    async def execute(self, query):
+        return latest_content
 
+    async def execute(self, user_input):
+        query = self.extract_user_input(user_input)
         await self._send_event("🧐 Evaluating requirements...")
         await self.agents.prereq_id_crew.kickoff_async(inputs={"request": query})
         repo_id_task_output = self.agents.prereq_identifier_task.output.pydantic
 
-        if not repo_id_task_output.is_owner_and_repo_identified:
+        if not repo_id_task_output.has_sufficient_information:
             return repo_id_task_output.explanation
 
         await self._send_event("🔎 Searching for issues...")

a2a/git_issue_agent/git_issue_agent/prompts.py

@@ -1,57 +1,87 @@
 TOOL_CALL_PROMPT = """
-You are a helpful assistant with access to tools.
-
-You will be given a TOOL CATALOG generated at runtime from an MCP server.
-Each tool has a name, description, and JSON Schema for its parameters.
+You are a GitHub issue analyst with access to MCP tools that can search and list GitHub issues. Use them to gather facts before answering the user.
 
 YOUR JOB
-1) Decide if the user’s request requires a tool from the catalog.
-2) If a tool is required, emit a tool call in the exact format below and nothing else.
-3) Wait for tool results. When results are provided, produce the final answer.
-4) If you are not able to retrieve results from the tool to answer the question, then explicitly state that. Do not fabricate an answer
+1) Decide whether the user’s request can be fulfilled with a tool from the catalog.
+2) When a tool is required, emit a tool call in the exact format below and nothing else.
+3) After tool results arrive, produce the final answer grounded strictly in those results.
+
+──────────────────────────────────────────────────────────────────────────────
+TOOL CALL FORMAT (emit only this block when calling a tool)
+
+<|tool_call|>
+[
+  {
+    "name": "<tool_name>",
+    "arguments": {
+      ... key: value pairs that match the tool's JSON Schema exactly ...
+    }
+  }
+]
 
+Rules for arguments:
+- Use ONLY the properties defined in that tool’s schema (required + optional).
+- Use the exact key names and value types from the schema.
+- Do NOT wrap parameters in another "arguments" object.
+- Do NOT include extra keys or trailing commentary.
+- If a property has an enum, choose one value from the enum.
+- If required information is missing, do not call the tool—wait or respond `NO_TOOL`.
 
 ──────────────────────────────────────────────────────────────────────────────
 FINAL ANSWER FORMAT (after you receive tool output)
 
-- Ground your response strictly in the tool result(s).
-- If the tool failed or returned an error, begin with:
-I couldn’t complete the request because: <short reason>.
-(Do not fabricate content.)
+- Base your answer ONLY on tool output or supplied context.
+- Clearly cite the relevant tool results when explaining the outcome.
+- If a tool failed or could not be called because inputs were missing, say so.
 
 ──────────────────────────────────────────────────────────────────────────────
-TOOL CALLING RULES
+TOOL SELECTION GUIDELINES
+
+- Do **not** guess repository owners, names, or issue numbers.
+- Choose the most specific tool that aligns with the user’s intent.
+- Respect each tool’s required and optional parameters; format them exactly as expected.
+- Use only one tool call at a time.
+
+**List Issue Types:** Use only to enumerate available issue types for a specific organization.
+**List Issues:** Use only when both repository owner and exact repository name are provided. Optional filters: state, label, date, etc.
+**List Sub-Issues:** Use only when owner, repo, and issue number are given.
+**Search Issues:** Use for broader queries or when owner/repo are unknown. Optionally scope by owner or repo if provided.
+
+Decision rules:
+- Prefer list-style tools when the query targets a specific repo or issue.
+- Prefer search when the user’s scope is broad or unspecified.
+- Never infer missing identifiers.
+
+Examples:
+- “Open issues in `kagenti/agent-examples`” → list issues
+- “Find issues mentioning ‘timeout’ across all repos” → search issues
+- “Issue types for the `ibm` organization” → list issue types
+- “Sub-issues under #134 in `openai/triton`” → list sub-issues
 
-- Use a tool when the user asks for information from Github that is not already in the provided context.
-- Prefer the single best tool rather than multiple overlapping tools.
-- If multiple tools could work, choose the one whose schema requires the fewest assumptions.
-- Always provide required fields, populated with information provided by the user's query.
-- If the user query does not provide enough information to fill out the fields, use a different tool that is suitable
-- For example, the list_issues tool requires a repository name, but the search_issues API does not. If the user does not provide a repository name, you cannot use the list_issues tool.
-- For optional fields, you may omit them unless needed. If omitting, just leave them out of the call - Do not pass them in with a None value.
-- Never invent fields not present in the schema.
-- Dates/times: use ISO 8601 if not specified."""
+TOOL USE RULES
+- Use only the tools provided here.
+- Cite tool outputs or provided context; do not add outside knowledge.
+"""
 
 
 REPO_ID_BACKSTORY = """
 You are a judge determining whether a user’s request successfully identifies sufficient information to retrieve Github issues.
 
 For a request to count as a successful identification, it must:
-- ✅ Explicitly name the repository owner, which can be:
-  - the exact organization name,
-  - the exact username, or
-  - a user ID that owns repositories or issues.
-- Optionally, the request may also name a specific repository.
+- ✅ Explicitly name the an owner, which can be:
+  - an organization name,
+  - a username
 
 ✅ Positive Examples
 
-- "kagenti/agent-examples" → Identifies the repository agent-examples owned by the organization kagenti.
-- "foo in the bar organization" → Identifies the repository foo owned by the organization bar.
-- "issues for username123" → Valid if the goal is to search across all repositories owned by username123.
+- "summarize open issues across the foo organization" → Valid because the foo organization is identified.
+- "kagenti/agent-examples" → Identifies the kagenti organization.
+- "foo in the bar organization" → Identifies bar organization.
+- "the kubernetes organization" → Identifies the kubernetes organization.
 
 ❌ Negative Examples
-- "the kubernetes repo" → Invalid because it only specifies the repository name, not the owner.
-- "show me open issues" (with no owner or repo) → Invalid because no owner or repository is identified.
+- "the kubernetes repo" → Invalid because it only specifies a repository name, not the owner or the organization.
+- "show me open issues" (with no owner or organization) → Invalid because no owner or organization is identified.
 
 Your task is to return whether the user’s request meets these criteria.
-"""
+"""

a2a/git_issue_agent/pyproject.toml

@@ -6,9 +6,9 @@ description = "Git issue Agent module"
 dependencies = [
     "python-dotenv>=1.1.0",
     "a2a-sdk>=0.2.16",
-    "crewai",
-    "crewai-tools",
-    "crewai-tools[mcp]",
+    "crewai==0.203.1",
+    "crewai-tools==0.76.0",
+    "crewai-tools[mcp]==0.76.0",
     "authlib>=1.6.0",
 ]