fix: resolve Ollama function calling failures with crewai 1.10.1

crewai 1.10.1's instructor integration requires LLMs to produce structured
tool calls for output_pydantic parsing. Ollama models (granite4, granite3.3,
llama3.2) don't reliably generate these, causing "Instructor does not support
multiple tool calls" errors and empty tool_calls responses.

Changes:
- Remove output_pydantic from prereq task, parse raw JSON output instead
- Add _parse_prereq_from_raw() fallback for extracting JSON from LLM text
- Switch default model prefix from ollama/ to ollama_chat/ for better
  litellm chat completions routing
- Pass num_ctx=8192 to Ollama models (default 2048 is too small)
- Fix config parameter bug: GitIssueAgent now uses passed config instead
  of module-level settings singleton
- Add field_validator for IssueSearchInfo.issue_numbers to handle
  string-encoded arrays from small LLMs
- Add max_retry_limit=3 and respect_context_window=True to issue researcher

Tested locally with granite4, granite3.3:8b, and llama3.2:3b-instruct-fp16.

Fixes #173

Assisted-By: Claude (Anthropic AI) <noreply@anthropic.com>
Signed-off-by: Mariusz Sabath <mrsabath@gmail.com>

Author: mrsabath

a2a/git_issue_agent/.env.ollama

@@ -5,7 +5,7 @@
 #   ollama pull ibm/granite4:latest
 
 # LLM configuration
-TASK_MODEL_ID=ollama/ibm/granite4:latest
+TASK_MODEL_ID=ollama_chat/ibm/granite4:latest
 # Ollama API base URL. Required by litellm (used by crewai >=1.10).
 # For Docker Desktop / Kind: http://host.docker.internal:11434
 # For in-cluster Ollama:     http://ollama.ollama.svc:11434

a2a/git_issue_agent/.env.template

@@ -1,4 +1,4 @@
-TASK_MODEL_ID = "ollama/ibm/granite4:latest"
+TASK_MODEL_ID = "ollama_chat/ibm/granite4:latest"
 LLM_API_BASE = "http://host.docker.internal:11434"
 LLM_API_KEY = "ollama"
 MODEL_TEMPERATURE = 0

a2a/git_issue_agent/git_issue_agent/agents.py

@@ -1,6 +1,5 @@
 from crewai import Agent, Crew, Process, Task
 from git_issue_agent.config import Settings
-from git_issue_agent.data_types import IssueSearchInfo
 from git_issue_agent.llm import CrewLLM
 from git_issue_agent.prompts import TOOL_CALL_PROMPT, INFO_PARSER_PROMPT
 
@@ -23,8 +22,10 @@ def __init__(self, config: Settings, issue_tools):
         self.prereq_identifier_task = Task(
             description=("User query: {request}"),
             agent=self.prereq_identifier,
-            output_pydantic=IssueSearchInfo,
-            expected_output=("A pydantic object representing the extracted relevant information."),
+            expected_output=(
+                'A JSON object with keys "owner", "repo", and "issue_numbers". '
+                "Example: {\"owner\": \"kagenti\", \"repo\": \"kagenti\", \"issue_numbers\": null}"
+            ),
         )
 
         self.prereq_id_crew = Crew(
@@ -49,6 +50,8 @@ def __init__(self, config: Settings, issue_tools):
             llm=self.llm.llm,
             inject_date=True,
             max_iter=6,
+            max_retry_limit=3,
+            respect_context_window=True,
         )
 
         # --- A generic task template -------------------------------------------------

a2a/git_issue_agent/git_issue_agent/config.py

@@ -12,7 +12,7 @@ class Settings(BaseSettings):
         description="Application log level",
     )
     TASK_MODEL_ID: str = Field(
-        os.getenv("TASK_MODEL_ID", "ollama/ibm/granite4:latest"),
+        os.getenv("TASK_MODEL_ID", "ollama_chat/ibm/granite4:latest"),
         description="The ID of the task model",
     )
     LLM_API_BASE: str = Field(

a2a/git_issue_agent/git_issue_agent/data_types.py

@@ -1,4 +1,6 @@
-from pydantic import BaseModel, Field
+import json
+
+from pydantic import BaseModel, Field, field_validator
 
 ############
 # Pydantic types for LLM response formats
@@ -11,3 +13,16 @@ class IssueSearchInfo(BaseModel):
     issue_numbers: list[int] | None = Field(
         None, description="Specific issue number(s) mentioned by the user. If none mentioned leave blank."
     )
+
+    @field_validator("issue_numbers", mode="before")
+    @classmethod
+    def coerce_string_to_list(cls, v):
+        """Small LLMs often serialize arrays as strings in tool call args."""
+        if isinstance(v, str):
+            try:
+                parsed = json.loads(v)
+                if isinstance(parsed, list):
+                    return parsed
+            except (json.JSONDecodeError, ValueError):
+                pass
+        return v

a2a/git_issue_agent/git_issue_agent/llm.py

@@ -4,13 +4,19 @@
 
 class CrewLLM:
     def __init__(self, config: Settings):
+        kwargs = {}
+        if config.EXTRA_HEADERS is not None and None not in config.EXTRA_HEADERS:
+            kwargs["extra_headers"] = config.EXTRA_HEADERS
+
+        # For Ollama models, pass num_ctx to set the context window size.
+        # Ollama defaults to 2048 tokens which is too small for agent workflows.
+        if config.TASK_MODEL_ID.startswith(("ollama/", "ollama_chat/")):
+            kwargs["num_ctx"] = 8192
+
         self.llm = LLM(
             model=config.TASK_MODEL_ID,
             base_url=config.LLM_API_BASE,
             api_key=config.LLM_API_KEY,
-            **(
-                {"extra_headers": config.EXTRA_HEADERS}
-                if config.EXTRA_HEADERS is not None and None not in config.EXTRA_HEADERS
-                else {}
-            ),
+            temperature=config.MODEL_TEMPERATURE,
+            **kwargs,
         )

a2a/git_issue_agent/git_issue_agent/main.py

@@ -1,16 +1,38 @@
 from crewai_tools.adapters.tool_collection import ToolCollection
 
+import json
 import logging
+import re
 import sys
 
 from git_issue_agent.config import Settings, settings
+from git_issue_agent.data_types import IssueSearchInfo
 from git_issue_agent.event import Event
 from git_issue_agent.agents import GitAgents
 
 logger = logging.getLogger(__name__)
 logging.basicConfig(level=settings.LOG_LEVEL, stream=sys.stdout, format="%(levelname)s: %(message)s")
 
 
+def _parse_prereq_from_raw(raw: str) -> IssueSearchInfo:
+    """Parse IssueSearchInfo from raw LLM text when instructor/pydantic parsing fails.
+
+    Some Ollama models don't produce structured tool calls that crewai's instructor
+    integration expects. This fallback extracts JSON from the raw text output.
+    """
+    # Try to find JSON in the raw output
+    json_match = re.search(r'\{[^{}]*\}', raw)
+    if json_match:
+        try:
+            data = json.loads(json_match.group())
+            return IssueSearchInfo(**data)
+        except (json.JSONDecodeError, ValueError):
+            pass
+
+    # Fallback: return empty IssueSearchInfo (no pre-identified fields)
+    return IssueSearchInfo()
+
+
 class GitIssueAgent:
     def __init__(
         self,
@@ -19,7 +41,7 @@ def __init__(
         mcp_toolkit: ToolCollection = None,
         logger=None,
     ):
-        self.agents = GitAgents(settings, mcp_toolkit)
+        self.agents = GitAgents(config, mcp_toolkit)
         self.eventer = eventer
         self.logger = logger or logging.getLogger(__name__)
 
@@ -45,11 +67,28 @@ def extract_user_input(self, body):
 
         return latest_content
 
+    async def _get_prereq_output(self, query: str) -> IssueSearchInfo:
+        """Run the prereq crew and extract IssueSearchInfo from raw text output.
+
+        We avoid using crewai's output_pydantic because it relies on instructor's
+        tool-call-based parsing, which fails with Ollama models that don't produce
+        structured tool calls. Instead, we ask the LLM for JSON and parse it ourselves.
+        """
+        try:
+            await self.agents.prereq_id_crew.kickoff_async(
+                inputs={"request": query, "repo": "", "owner": "", "issues": []}
+            )
+            raw = self.agents.prereq_identifier_task.output.raw
+            self.logger.info(f"Prereq raw output: {raw}")
+            return _parse_prereq_from_raw(raw)
+        except Exception as e:
+            self.logger.warning(f"Prereq crew failed: {e}")
+            return IssueSearchInfo()
+
     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": "", "owner": "", "issues": []})
-        repo_id_task_output = self.agents.prereq_identifier_task.output.pydantic
+        repo_id_task_output = await self._get_prereq_output(query)
 
         if repo_id_task_output.issue_numbers:
             if not repo_id_task_output.owner or not repo_id_task_output.repo: