Implement pydanyic-ai for analysis

Author: trly

.beads/beads.base.jsonl

@@ -1 +1 @@
-{"version":"0.22.1","timestamp":"2025-11-07T11:14:47.148162-05:00","commit":"8f9fcc8"}
+{"version":"0.22.1","timestamp":"2025-11-14T09:16:38.967273-05:00","commit":"549a780"}

.beads/beads.left.jsonl

@@ -0,0 +1,32 @@
+# Python cache
+__pycache__/
+*.pyc
+*.pyo
+*.pyd
+.Python
+
+# Virtual environments
+.venv/
+venv/
+ENV/
+
+# Development and testing
+.pytest_cache/
+.ruff_cache/
+tests/
+
+# Version control
+.git/
+.github/
+.jj/
+
+# Local development
+.beads/
+history/
+.envrc
+
+# IDE
+.vscode/
+.idea/
+*.swp
+*.swo

.beads/beads.left.meta.json

@@ -355,6 +355,8 @@ Configuration uses `TRIAGE_` prefix:
 - `TRIAGE_SOURCEGRAPH_TOKEN`: Sourcegraph access token for MCP integration
 - `TRIAGE_SOURCEGRAPH_MCP_URL`: Sourcegraph MCP server URL
 - `TRIAGE_REDIS_URL`: Redis URL for Celery broker and result backend (default: redis://localhost:6379/0)
+- `TRIAGE_DISABLE_ISSUE_CREATION`: Set to "true" to disable GitHub issue creation (logs proposals instead). Useful for local testing without creating real issues
+- `TRIAGE_LOG_LEVEL`: Set to "DEBUG" to enable detailed agent conversation logging (default: INFO)
 
 ## Current Implementation Status
 

.dockerignore

@@ -29,6 +29,12 @@ src/github_action_triage/
 │   ├── celery_app.py        # Celery application configuration
 │   └── factory.py           # FastAPI application factory
 ├── agent/                   # Agent layer
+│   ├── analysis/            # Analysis agent implementation
+│   │   ├── agent.py         # Core analysis agent
+│   │   ├── config.py        # Analysis agent configuration
+│   │   └── tools/           # Agent tool integrations
+│   │       ├── github.py    # GitHub API tools
+│   │       └── sourcegraph.py # Sourcegraph code search tools
 │   ├── ports.py             # Protocol definitions for external services
 │   ├── ai_agent.py          # Claude Agent SDK remediation agent
 │   ├── config.py            # Agent configuration
@@ -122,12 +128,14 @@ export TRIAGE_ANTHROPIC_API_KEY="sk-ant-..."
 export TRIAGE_SOURCEGRAPH_TOKEN="sgp_..."
 export TRIAGE_SOURCEGRAPH_MCP_URL="http://localhost:3000"
 export TRIAGE_LOG_LEVEL="INFO"  # DEBUG, INFO, WARNING, ERROR, CRITICAL
+export TRIAGE_DISABLE_ISSUE_CREATION="false"  # Set to "true" for testing without creating issues
 ```
 
 **Notes**:
 
 - `TRIAGE_GITHUB_PRIVATE_KEY` should contain the full PEM content (including `-----BEGIN RSA PRIVATE KEY-----` and `-----END RSA PRIVATE KEY-----` lines), not just a file path.
 - `TRIAGE_GITHUB_WEBHOOK_SECRET` should be a secure random string. Generate one with:
+- `TRIAGE_DISABLE_ISSUE_CREATION` when set to `"true"`, disables GitHub issue creation and instead logs the proposal. Useful for local testing and development to avoid cluttering repositories with test issues.
 
   ```bash
   # Generate a secure random secret

AGENTS.md

@@ -48,7 +48,7 @@ services:
       dockerfile: Dockerfile
       target: worker
     container_name: triage-worker
-    command: sh -c "cd /app && /app/venv/bin/python -m celery -A github_action_triage.app.celery_app:app worker --loglevel=info --concurrency=2"
+    command: sh -c "cd /app && /app/venv/bin/python -m celery -A github_action_triage.app.celery_app:app worker --loglevel=debug --concurrency=2"
     user: appuser
     volumes:
       - ./src:/app/src:ro
@@ -60,11 +60,12 @@ services:
       TRIAGE_MCP_ENABLE_SOURCEGRAPH: ${TRIAGE_MCP_ENABLE_SOURCEGRAPH:-false}
       TRIAGE_SOURCEGRAPH_TOKEN: ${TRIAGE_SOURCEGRAPH_TOKEN:-}
       TRIAGE_SOURCEGRAPH_MCP_URL: ${TRIAGE_SOURCEGRAPH_MCP_URL:-}
+      TRIAGE_DISABLE_ISSUE_CREATION: ${TRIAGE_DISABLE_ISSUE_CREATION:-false}
       TRIAGE_LOG_LEVEL: ${TRIAGE_LOG_LEVEL:-INFO}
     depends_on:
       redis:
         condition: service_healthy
     restart: unless-stopped
 
 volumes:
-  redis-data:
+  redis-data:

README.md

@@ -14,6 +14,7 @@ dependencies = [
     "fastapi>=0.121.0",
     "githubkit>=0.13.5",
     "httpx>=0.28.1",
+    "pydantic-ai-slim[anthropic,mcp]>=1.16.0",
     "redis>=5.2.0",
 ]
 

docker-compose.yml

@@ -3,7 +3,16 @@
 from typing import Any
 
 from claude_agent_sdk import ClaudeSDKClient, create_sdk_mcp_server, tool
-from claude_agent_sdk.types import AssistantMessage, ClaudeAgentOptions, ResultMessage, TextBlock
+from claude_agent_sdk.types import (
+    AssistantMessage,
+    ClaudeAgentOptions,
+    Message,
+    ResultMessage,
+    TextBlock,
+    ToolResultBlock,
+    ToolUseBlock,
+    UserMessage,
+)
 
 from github_action_triage.agent.config import Settings
 from github_action_triage.agent.mcp import (
@@ -30,21 +39,26 @@
 - Failure logs excerpt demonstrating the error condition
 - Workflow file path (if available)
 - Recent commit history"""
-
 SYSTEM_PROMPT_MCP_TOOLS = """
 
 ## Remote Repository Access
 
-**All repository code inspection must happen through the Sourcegraph MCP server.** You do not have local access to any files in the repository. Do not assume you can read files directly or that the repository is checked out on your filesystem.
+**CRITICAL: All repository code inspection MUST happen through the Sourcegraph MCP server.** You do not have local access to any files in the repository. Do not assume you can read files directly or that the repository is checked out on your filesystem.
 
-You have access to a Sourcegraph MCP server (OAuth-authenticated) providing:
-- Code search (keyword and semantic search across repositories)
-- File reading (sg_read_file) — the ONLY way to read repository files
-- Directory listing (sg_list_files) — the ONLY way to list repository contents
-- Symbol navigation (sg_find_references, sg_go_to_definition)
-- Commit and diff search
+**You MUST use the following Sourcegraph MCP tools for all code investigation:**
 
-Use these MCP tools exclusively to investigate the codebase, examine workflow configurations, and analyze recent changes that may have introduced the failure."""
+- **sg_read_file** — Read file contents (the ONLY way to read repository files)
+- **sg_list_files** — List directory contents (the ONLY way to browse the repository)
+- **sg_list_repos** — Verify repository names and access
+- **sg_keyword_search** — Search for exact code patterns, function names, and keywords
+- **sg_nls_search** — Semantic search for concepts and related code
+- **sg_go_to_definition** — Navigate to symbol definitions
+- **sg_find_references** — Find where symbols are used
+- **sg_commit_search** — Search commit history and messages
+- **sg_diff_search** — Search code changes in diffs
+- **sg_compare_revisions** — Compare code between commits
+
+**IMPORTANT:** Start your investigation by using sg_list_repos to verify the repository name, then use sg_keyword_search or sg_nls_search to find relevant code, workflow files, and configuration. Use sg_read_file to examine specific files. These tools are OAuth-authenticated and provide full access to the codebase."""
 
 SYSTEM_PROMPT_WORKFLOW = """
 
@@ -53,11 +67,11 @@
 1. Examine the logs_excerpt to identify the immediate failure symptom
 2. Use available tools to investigate root cause (workflow YAML, recent commits, dependency files, code changes)
 3. Form a hypothesis about the underlying issue
-4. Once you have high confidence in your diagnosis, submit your remediation proposal
+4. REQUIRED: Submit your remediation proposal using the submit_proposal tool
 
 ## Remediation Proposal Requirements
 
-When confident in your analysis, invoke the submit_proposal tool with four parameters:
+You MUST invoke the submit_proposal tool with four parameters to complete the analysis:
 
 - **issue_title**: Short, actionable title for GitHub issue (< 80 characters). Example: "Ruff linting errors in source files"
 - **identified_issue**: Precise description of the root cause (not just the symptom)
@@ -181,6 +195,10 @@ async def diagnose_and_propose(self, context: FailureContext) -> RemediationProp
             f"(run_id={context.event.workflow.run_id}, model={self._settings.claude_model}, "
             f"MCP={'enabled' if has_sourcegraph else 'disabled'})"
         )
+        
+        # Log system prompt and initial user prompt at DEBUG level
+        logger.debug(f"System prompt:\n{system_prompt}")
+        logger.debug(f"Initial user prompt:\n{initial_prompt}")
 
         # Run message loop with proper error handling
         final_result: ResultMessage | None = None
@@ -190,12 +208,49 @@ async def diagnose_and_propose(self, context: FailureContext) -> RemediationProp
 
             async def iterate_messages():
                 nonlocal final_result
+                turn_counter = 0
                 async for message in client.receive_response():
                     # Log assistant messages for debugging
                     if isinstance(message, AssistantMessage):
+                        turn_counter += 1
+                        logger.debug(f"[Turn {turn_counter}] Assistant message received")
+                        
                         for block in message.content:
                             if isinstance(block, TextBlock):
-                                logger.debug(f"Claude: {block.text[:500]}")
+                                # Log full text at DEBUG level for local debugging
+                                logger.debug(f"[Turn {turn_counter}] Text: {block.text}")
+                            elif isinstance(block, ToolUseBlock):
+                                # Log tool calls to see what the agent is doing
+                                tool_name = block.name
+                                tool_input = getattr(block, "input", {})
+                                
+                                # Highlight MCP calls specifically
+                                if tool_name.startswith("mcp__"):
+                                    logger.debug(
+                                        f"[Turn {turn_counter}] 🔧 MCP TOOL CALL: {tool_name}\n"
+                                        f"  Args: {tool_input}"
+                                    )
+                                else:
+                                    logger.debug(
+                                        f"[Turn {turn_counter}] Tool call: {tool_name} "
+                                        f"with args: {tool_input}"
+                                    )
+
+                    # Log user messages (tool results)
+                    elif isinstance(message, UserMessage):
+                        logger.debug(f"[Turn {turn_counter}] User message (tool results)")
+                        for block in message.content:
+                            if isinstance(block, ToolResultBlock):
+                                tool_use_id = block.tool_use_id
+                                result_content = block.content
+                                # Truncate large results for readability
+                                if isinstance(result_content, str) and len(result_content) > 500:
+                                    result_preview = result_content[:500] + "...[truncated]"
+                                else:
+                                    result_preview = result_content
+                                logger.debug(
+                                    f"[Turn {turn_counter}] Tool result for {tool_use_id}: {result_preview}"
+                                )
 
                     # Check for final result message
                     elif isinstance(message, ResultMessage):
@@ -205,6 +260,10 @@ async def iterate_messages():
                             f"duration_ms={message.duration_ms}, error={message.is_error}"
                         )
                         break
+                    
+                    # Log any other message types we might be missing
+                    else:
+                        logger.debug(f"[Turn {turn_counter}] Other message type: {type(message).__name__}")
 
             try:
                 await asyncio.wait_for(
@@ -309,4 +368,4 @@ async def submit_proposal_tool(args: dict[str, Any]) -> dict[str, Any]:
 
             return {"content": [{"type": "text", "text": "Proposal submitted successfully"}]}
 
-        return submit_proposal_tool
+        return submit_proposal_tool

pyproject.toml

@@ -0,0 +1 @@
+