AUTO-6 - trckr

← AUTO-6

backend core

Created: 2026-02-03 • Updated: 2026-02-03

Relationships Loading...

## Overview
Create the AutomationRunner class that executes automations using Claude Code CLI.

## Tasks
- [ ] Create `server/runner.py` with AutomationRunner class
- [ ] Implement `run(automation, cwd, run_id)` method:
  - Check if cwd is git repo → create worktree via WorktreeManager
  - Build claude command: `claude --print --output-format json --max-budget-usd X --model Y --append-system-prompt "..." "prompt"`
  - Execute with subprocess, capture stdout/stderr
  - Return RunResult dataclass with exit_code, stdout, stderr, worktree_path
- [ ] Build automation system prompt that requires structured JSON finding output
- [ ] Handle timeouts (default 1 hour)
- [ ] Handle non-zero exit codes

## System Prompt Template
```
You are running as part of an automated check.
IMPORTANT: At the end of your response, include a JSON block:
\`\`\`json
{
  "finding_status": "none|informational|actionable|error",
  "title": "Brief title",
  "summary": "Detailed summary",
  "should_create_issue": true/false,
  "suggested_issue_title": "...",
  "suggested_issue_body": "..."
}
\`\`\`
```

## QA: Live Integration Testing

**Phase Gate 1 - Module Imports:**
```bash
uv run python -c "from server.runner import AutomationRunner, RunResult; print('OK')"
```
✓ PASS: Prints "OK"

**Phase Gate 2 - Command Construction (dry run):**
```bash
uv run python -c "
from server.runner import AutomationRunner
runner = AutomationRunner()
cmd = runner._build_command(
    prompt='List files in current directory',
    model='haiku',
    max_budget_usd=0.10
)
print(' '.join(cmd))
assert 'claude' in cmd
assert '--print' in cmd
assert '--output-format' in cmd
assert 'json' in cmd
print('OK')
"
```
✓ PASS: Shows command string with expected flags

**Phase Gate 3 - Simple Run (LIVE - costs money):**
```bash
uv run python -c "
from server.runner import AutomationRunner
runner = AutomationRunner()
automation = {
    'prompt': 'What is 2+2? Respond with the finding JSON block only.',
    'model': 'haiku',
    'max_budget_usd': 0.05
}
result = runner.run(automation, '/tmp', 'test-run-001')
print(f'Exit code: {result.exit_code}')
print(f'Output length: {len(result.stdout)}')
assert result.exit_code == 0
assert len(result.stdout) > 0
print('OK')
"
```
✓ PASS: Exit code 0, has output

**Phase Gate 4 - Git Repo Run with Worktree (LIVE):**
```bash
uv run python -c "
from server.runner import AutomationRunner
import os
runner = AutomationRunner()
automation = {
    'prompt': 'What files are in this directory? Respond with finding JSON.',
    'model': 'haiku', 
    'max_budget_usd': 0.05
}
result = runner.run(automation, '/Users/joe/gitea/agent-automations', 'test-run-002')
print(f'Exit code: {result.exit_code}')
print(f'Worktree: {result.worktree_path}')
# Worktree should have been created and cleaned up (or still exists if we want to inspect)
print('OK')
"
```
✓ PASS: Exit code 0, worktree_path is set

Pull Requests (comma-separated URLs)

Attachments

Comments (4)

qa-agent · 2026-02-03

## QA PASSED See QA report document: dfdf8bc3-a8eb-494a-af47-ba40736b895e All 4 phase gates passed with LIVE testing. Additional hypotheses tested: - Timeout handling (exit 124) - Worktree cleanup - Worktree fallback mode - Empty prompt handling **Minor note:** Non-existent cwd produces misleading error message (reports 'claude CLI not found' instead of 'directory not found'). This is cosmetic and can be addressed in a future enhancement. **Test Results:** - 80 tests pass - runner.py coverage: 95% - All LIVE tests successful Fix verified working. Marking as done.

agent · 2026-02-03

## Implementation Complete ### What was implemented Created `server/runner.py` with the AutomationRunner class that executes automations using Claude Code CLI. ### Key components 1. **RunResult dataclass** - Contains exit_code, stdout, stderr, worktree_path, duration_ms 2. **AutomationRunner class** with methods: - `_build_command()` - Constructs the claude CLI command with all required flags - `run()` - Executes the automation, creating worktrees for git repos - `cleanup_worktree()` - Cleans up worktrees after runs 3. **DEFAULT_SYSTEM_PROMPT** - Instructs Claude to output structured finding JSON ### Features - Automatic worktree creation for git repositories (via WorktreeManager) - Configurable timeout (default 1 hour) - Handles subprocess timeouts, command not found, and OS errors gracefully - Duration measurement in milliseconds - Default model (sonnet) and budget (0.50 USD) with overrides ### QA Results - All 4 Phase Gates Passed 1. **Module Imports**: OK 2. **Command Construction**: Verified all flags present (--print, --output-format json, etc.) 3. **Simple Run (LIVE)**: Exit code 0, 1129 bytes output, 6023ms duration 4. **Git Repo Run with Worktree (LIVE)**: Exit code 0, worktree created at /tmp/automate/run-test-run ### Test Results - 49 tests pass - runner.py coverage: 95% - New test file: tests/test_runner.py with 19 tests covering all functionality

agent · 2026-02-03

Agent Assignment: Assigned to worker-AUTO-6 in worktree /tmp/agent-automations-AUTO-6

triage-agent · 2026-02-03

## Feature Triage Complete **Status:** NOT IMPLEMENTED **Summary:** This is a new feature for a greenfield project. The AutomationRunner class will execute Claude Code CLI commands with proper timeout handling, worktree support, and structured JSON output parsing. The project currently has only documentation - no Python code exists yet. **Implementation Location:** - New file: server/runner.py - Depends on: AUTO-2 (LocalStore), AUTO-3 (WorktreeManager) **Key Implementation Details:** - Use subprocess.run() with timeout=3600 (1 hour default) - Build command as list to prevent shell injection - Required claude CLI flags: --print --output-format json --model X --max-budget-usd Y --append-system-prompt Z - Return RunResult dataclass with exit_code, stdout, stderr, worktree_path, duration_ms **Complexity:** medium (single file, ~100-150 lines) **Dependencies verified:** - claude CLI available at /home/nuc2/.local/bin/claude - All required flags confirmed in Claude Code documentation See attached triage report for full details including research findings and edge cases.