Task Lifecycle

This page documents the exact lifecycle implemented in orchestrator.py.

1. Provider Initialization

The selected provider (anthropic, openai, gemini) is instantiated once and reused by:

• task analyzer
• recommendation reasoning
• test generator
• completeness judge in scoring

2. Task Analysis

The analyzer converts raw text into TaskDefinition fields:

• type
• domain
• requirements (language/frameworks/capabilities)
• complexity
• success_criteria
• testable_assertions

3. Stack Recommendation

Recommendation combines:

• optional Tavily web search results
• MCP server candidates from upstream registries
• skills from upstream GitHub catalogs
• LLM synthesis into a StackRecommendation

4. Execution Attempt

For each attempt:

• runtime files (task.json, stack.json) are written
• agent image is selected or built
• container executes and writes artifacts to /workspace/output
• artifacts are collected into ExecutionResult.artifacts

5. Test Generation and Run

• tests are generated from task metadata + artifacts
• tests are written to /workspace/tests/test_output.py
• tests run in a separate container with network_mode=none

6. Scoring

Scoring computes weighted dimensions and pass/fail against threshold.

7. Retry

If score is below threshold and retries remain:

• failure dimensions are analyzed
• model upgrade and/or re-search hints are applied
• next attempt runs with adjusted stack

The final response returns the best attempt, not necessarily the last attempt.