Multi-Phase Orchestration Pipeline
Every feature request flows through a structured pipeline with safety gates, isolated execution, and automatic verification. No manual intervention required.
How the pipeline works
1
Prompt
Build context-rich prompt with project state, file contents, and coding standards
2
Stream
Stream response from local LLM with structured output constraints
3
Parse & Review
8-point safety check on proposed steps before any execution
4
Execute
Run commands and apply file changes in isolated executor container
5
Review & Fix
Evaluate results against acceptance criteria; auto-fix if needed
Technical details
Structured Output
The operator uses Ollama's grammar-constrained output to guarantee valid JSON from every LLM response. No more parse failures from think blocks or markdown wrappers.
json
{
"steps": [
{ "type": "command", "command": "npm init -y" },
{ "type": "file", "path": "src/auth.ts", "action": "create",
"content": "// JWT authentication module..." },
{ "type": "verify", "command": "npm test" }
]
}Orchestration Limits
Built-in safeguards prevent runaway loops:
- max_auto_fix_depth = 5 - Maximum nesting for fix children
- max_orchestration_attempts = 3 - Failed run limit per request
- max_review_fix_cycles = 3 - Review → fix → re-execute loop cap
- max_review_followup_rounds = 3 - INFO_NEEDED round limit