Self-Healing Image Generation Tool Chain for Autonomous Agents

Role: Principal Visual AI Automation & Quality Systems Engineer

Context

• Vision Critique Evaluator: {{evaluator_vision_model}}
• Retry Budget: {{max_retry_cycles}}
• Acceptance Criteria: {{quality_thresholds}}
• Generation Target: {{diffusion_api_endpoint}}
• Heuristic Mutation Policy: {{prompt_mutation_rules}}

Task

Develop an autonomous self-healing Python script that dispatches image generation requests to {{diffusion_api_endpoint}}, submits outputs to {{evaluator_vision_model}} for structured visual critique against {{quality_thresholds}}, and programmatically mutates prompt tokens and parameters until criteria are met or {{max_retry_cycles}} is reached.

Method

1. Define structured Pydantic schemas for the visual inspection verdict returned by {{evaluator_vision_model}}.
2. Implement the API client for {{diffusion_api_endpoint}} capable of handling parameter overrides (prompt, negative prompt, cfg, steps).
3. Construct the Vision QA prompt that forces {{evaluator_vision_model}} to score output against {{quality_thresholds}} with granular defect tags.
4. Build the prompt mutation engine implementing {{prompt_mutation_rules}} (token weight boosting, negative token injection, composition shift).
5. Orchestrate the evaluation-retry loop, tracking mutation history and preventing cyclical token thrashing.
6. Implement final resolution logic: approve and output asset URLs on success, or trigger structured triage alerts if retry limits expire.

Constraints

• The script MUST execute a closed-loop mutation cycle without requiring human intervention.
• MUST NOT exceed {{max_retry_cycles}} under any circumstance to prevent infinite billing loops.
• Vision critique MUST output strictly parseable JSON conforming to the verification schema.
• Every prompt mutation MUST record a diff log explaining which tokens were added, removed, or re-weighted.

Output format

Provide a unified Python script organized as follows:

1. Configuration Classes and Pydantic QA Evaluation Schemas.
2. Vision Inspector and Prompt Mutator Engine.
3. Generation Client for {{diffusion_api_endpoint}}.
4. Autonomous Self-Healing Orchestrator Loop.
5. Verification test suite simulating a failed generation that recovers on retry.

Self-review

• Confirm that the mutation logic strictly follows {{prompt_mutation_rules}} on detected defects.
• Verify that the retry loop hard-terminates when counter equals {{max_retry_cycles}}.
• Ensure parsing handles partial or malformed vision evaluation responses safely.