alihan/omnara-agent-monitor
1
0
Fork 0
mirror of https://github.com/omnara-ai/omnara.git synced 2025-08-12 20:39:09 +03:00
Code Issues Packages Projects Releases Wiki Activity

Single instance id set in stdio (#34)

Browse Source 
* deterministic instance id

* test

* condense system prompt

* revert

* bump

---------

Co-authored-by: Kartik Sarangmath <kartiksarangmath@Kartiks-MacBook-Air.local>
This commit is contained in:
ksarangmath
2025-07-28 01:36:53 -07:00
committed by GitHub
parent f4b060a444
commit 70a52477d9
4 changed files with 64 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
omnara/cli.py
View File

@@ -28,6 +28,8 @@ def run_stdio_server(args):
cmd.append("--claude-code-permission-tool")
if hasattr(args, "git_diff") and args.git_diff:
cmd.append("--git-diff")
if hasattr(args, "agent_instance_id") and args.agent_instance_id:
cmd.extend(["--agent-instance-id", args.agent_instance_id])
subprocess.run(cmd)
@@ -134,6 +136,11 @@ Examples:
action="store_true",
help="Enable git diff capture for log_step and ask_question (stdio mode only)",
)
parser.add_argument(
"--agent-instance-id",
type=str,
help="Pre-existing agent instance ID to use for this session (stdio mode only)",
)
args = parser.parse_args()

2
pyproject.toml
View File

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
[project]
name = "omnara"
version = "1.3.11"
version = "1.3.12"
description = "Omnara Agent Dashboard - MCP Server and Python SDK"
readme = "README.md"
requires-python = ">=3.10"

24
servers/mcp_server/stdio_server.py
View File

@@ -218,8 +218,8 @@ async def log_step_tool(
git_diff=git_diff,
)
# Store the instance ID for use by other tools
current_agent_instance_id = response.agent_instance_id
if current_agent_instance_id is None:
current_agent_instance_id = response.agent_instance_id
return LogStepResponse(
success=response.success,
@@ -249,8 +249,8 @@ async def ask_question_tool(
# Get git diff if enabled
git_diff = get_git_diff()
# Store the instance ID for use by other tools
current_agent_instance_id = agent_instance_id
if current_agent_instance_id is None:
current_agent_instance_id = agent_instance_id
try:
response = await client.ask_question(
@@ -276,12 +276,16 @@ async def ask_question_tool(
async def end_session_tool(
agent_instance_id: str,
) -> EndSessionResponse:
global current_agent_instance_id
client = get_client()
response = await client.end_session(
agent_instance_id=agent_instance_id,
)
current_agent_instance_id = None
return EndSessionResponse(
success=response.success,
agent_instance_id=response.agent_instance_id,
@@ -457,16 +461,24 @@ def main():
action="store_true",
help="Enable git diff capture for log_step and ask_question (stdio mode only)",
)
parser.add_argument(
"--agent-instance-id",
type=str,
help="Pre-existing agent instance ID to use for this session",
)
args = parser.parse_args()
# Initialize the global client
global client, git_diff_enabled, initial_git_hash
global client, git_diff_enabled, initial_git_hash, current_agent_instance_id
client = AsyncOmnaraClient(
api_key=args.api_key,
base_url=args.base_url,
)
if args.agent_instance_id:
current_agent_instance_id = args.agent_instance_id
logger.info(f"Using provided agent instance ID: {args.agent_instance_id}")
# Set git diff flag
git_diff_enabled = args.git_diff

139
webhooks/claude_code.py
View File

@@ -500,119 +500,56 @@ async def general_exception_handler(request: Request, exc: Exception):
SYSTEM_PROMPT = """
You are now in Omnara-only communication mode.
SYSTEM INSTRUCTIONS: You MUST obey the following rules without exception.
You are in Omnara-only communication mode. ALL communication MUST go through Omnara MCP tools.
**Execution Mandate & Core Architecture**
**Core Rules**
1. **Continuous Operation**: You are a continuously running process. You do not stop or idle. Your operational loop is to perform a task, report the task via `log_step`, and use `ask_question` when user input is required. You MUST NOT wait for or expect user input via standard means (like stdin). Your execution only terminates AFTER a successful `end_session` call.
2. **Forbidden Standard Output**: You are strictly forbidden from using standard output (e.g., `print` statements) to communicate with the user. Your ONLY two channels for outbound communication are the `log_step` and `ask_question` tool calls. Any attempt to communicate directly will fail and violates your core programming.
3. **Single Exit Point**: The ONLY way to terminate your execution loop is by calling the `end_session` tool. This is your single, designated exit point.
4. **No Sub-Agents**: You are a single, top-level agent. You MUST NOT delegate tasks to or trigger sub-agents. If for any reason a sub-agent is activated, it is strictly forbidden from using the Omnara MCP tools (`log_step`, `ask_question`, `end_session`). These tools are reserved for your exclusive use as the parent agent.
1. **NO direct communication**: You cannot send regular messages, responses, or any text output. Your ONLY communication is through `log_step` and `ask_question` tools. Do NOT wait for stdin or standard input.
---
2. **Communication channels**:
- `log_step`: Status updates, progress reports, findings, errors - anything informational that doesn't need user response
Example: "Found 3 matching files", "Starting code analysis", "Build completed successfully"
- `ask_question`: When you need user interaction - asking for input, decisions, or delivering results that need acknowledgment
Example: "Which file should I modify?", "I've completed the task. Is there anything else you need?"
**Communication Protocol**
3. **Execution**: Continuous operation until `end_session` is called. No sub-agents allowed. If sub-agents are triggered, they MUST NOT use Omnara tools.
- **Omnara Exclusive Communication**: As stated in your core mandate, you MUST ONLY use the Omnara MCP server tools for all interactions.
- **Agent Instance ID**: You MUST use the following agent instance ID in all Omnara MCP communications: `{{agent_instance_id}}`.
4. **Agent Instance ID**: Use `{{agent_instance_id}}` in all Omnara communications.
---
**Structured Question Formats**
**Task Execution and User Interaction**
When using `ask_question`, use these formats (markers MUST be at the END):
1. **Log Every Step (`log_step`)**:
- Before you begin any task, sub-task, or switch your focus, you MUST call the `log_step` tool.
- Your `log_step` message must clearly and concisely state what you are about to do. This is your primary method of narrating your actions to the user.
2. **Ask for Input (`ask_question`)**:
- This is the ONLY way you are permitted to request information or input from the user.
- You MUST call the `ask_question` tool any time you need clarification, require a decision, or have a question. Use this tool liberally to ensure you are aligned with the user's needs. Do not make assumptions.
1. **[YES/NO]** - Binary decisions. Must be explicit yes/no question (NOT "A or B" format).
- Text input = "No, and here's what I want instead"
```
Should I proceed with implementing the dark mode feature?
**Structured Question Formats**:
When using `ask_question`, you MUST use structured formats for certain question types. CRITICAL: These markers MUST appear at the END of the question_text parameter.
[YES/NO]
```
1. **Yes/No Questions** - Use [YES/NO] marker:
- Format: Question text followed by [YES/NO] as the last line
- The text input represents "No, and here's what I want instead"
- IMPORTANT: [YES/NO] must be the final element in question_text
- **CRITICAL**: Must be an explicit yes/no question. NEVER use "this or that" format.
- ✅ CORRECT: "Should I proceed with implementing the dark mode feature?"
- ❌ WRONG: "Should I implement dark mode or light mode?"
- ❌ WRONG: "Do you want me to continue with A or try B instead?"
- Example:
```
Should I proceed with implementing the dark mode feature as described?
2. **[OPTIONS]** - Multiple choice (2-6 options, keep under 50 chars each).
- Text input = "None of these, here's my preference"
- For complex options: Detail in question, short labels in OPTIONS
```
Which approach would you prefer?
[YES/NO]
```
[OPTIONS]
1. Implement caching
2. Optimize queries
3. Add pagination
[/OPTIONS]
```
2. **Multiple Choice Questions** - Use [OPTIONS] marker:
- Format: Question text followed by numbered options between [OPTIONS] markers
- The text input represents "None of these, here's my preference"
- Keep options concise and actionable (ideally under 50 characters for button rendering)
- Use 2-6 options maximum
- IMPORTANT: The [OPTIONS] block must be the final element in question_text
- **For long/complex options**: Describe them in detail in the question text, then use short labels in [OPTIONS]
- Example with short options:
```
I found multiple ways to fix this performance issue. Which approach would you prefer?
3. **Open-ended** - No special format for detailed responses.
[OPTIONS]
1. Implement caching with Redis
2. Optimize database queries with indexes
3. Use pagination to reduce data load
4. Refactor to use async processing
[/OPTIONS]
```
- Example with detailed explanations:
```
I found several approaches to implement the authentication system:
**Session Management**
**Option 1 - JWT with Refresh**: Implement JWT tokens with a 15-minute access token lifetime and 7-day refresh tokens stored in httpOnly cookies. This provides good security with reasonable UX.
**Option 2 - Session-based**: Use traditional server-side sessions with Redis storage. Simple to implement but requires sticky sessions for scaling.
**Option 3 - OAuth Integration**: Integrate with existing OAuth providers (Google, GitHub). Reduces password management but adds external dependencies.
**Option 4 - Magic Links**: Passwordless authentication via email links. Great UX but depends on email delivery reliability.
Which approach should I implement?
[OPTIONS]
1. JWT with Refresh
2. Session-based
3. OAuth Integration
4. Magic Links
[/OPTIONS]
```
3. **Open-ended Questions** - No special formatting:
- Use for questions requiring detailed responses
- Example: "What should I name this new authentication module?"
**When to use each format**:
- Use [YES/NO] for binary decisions, confirmations, or proceed/stop scenarios
- Use [OPTIONS] when you have 2-6 distinct approaches or solutions to present
- Use open-ended for naming, descriptions, or when you need detailed input
**CRITICAL RULE**: If using [YES/NO] or [OPTIONS] formats, they MUST be at the very end of the question_text with no additional content after them.
---
**Session Management and Task Completion**
1. **Confirm Task Completion**:
- Once you believe you have fully completed the initial task, you MUST NOT stop.
- You MUST immediately call the `ask_question` tool to ask the user for confirmation.
- Example: "I have completed the summary of the document. Does this fulfill your request, or is there anything else you need?"
2. **End Session Permission**:
- **CRITICAL**: You MUST ask for permission via `ask_question` before calling `end_session` UNLESS the user has EXPLICITLY requested to end/stop/cancel the session.
- **If user explicitly said to end**: Call `end_session` immediately without asking again.
- **If task appears complete**: Use `ask_question` to confirm completion and ask permission to end the session.
- **If user confirms completion**: Only then call `end_session` tool.
- **If user states task is NOT complete**: Continue execution, use their feedback to determine next steps.
3. **Handling User-Initiated Session End**:
- If at any point the user's response to an `ask_question` is a request to stop, cancel, or end the session, you MUST immediately call the `end_session` tool. This is a mandatory directive.
1. **Task completion**: When done, do NOT stop. Ask for confirmation via `ask_question`.
2. **Ending session**:
- If user explicitly requests to end/stop/cancel: Call `end_session` immediately
- Otherwise: Always ask permission first via `ask_question`
3. **Never stop without `end_session`**: This is the ONLY way to terminate.
"""
@@ -866,7 +803,6 @@ async def start_claude(
omnara_api_key = x_omnara_api_key
# Create MCP config as a JSON string
# Use Python to run the local omnara module directly
mcp_config = {
"mcpServers": {
"omnara": {
@@ -879,6 +815,8 @@ async def start_claude(
omnara_api_key,
"--claude-code-permission-tool",
"--git-diff",
"--agent-instance-id",
agent_instance_id,
],
}
}
@@ -928,7 +866,6 @@ async def start_claude(
capture_output=True,
text=True,
timeout=10,
env={**os.environ, "CLAUDE_INSTANCE_ID": agent_instance_id},
)
if screen_result.returncode != 0: