Custom API Optimizer

Problem: The Custom API Gap

In WorkBuddy custom API mode, the framework forwards input to a third-party model and returns that model's output. Reasoning, Tool Calling, and Image Input can be enabled, but official docs do not promise that custom models match built-in models exactly in the Agent Loop. In practice, third-party models often lack the user's domain method, output preferences, and acceptance habits, while short prompts do not reliably reproduce the workflow.

Workflow: Configuration-Driven Injection

The skill centers on profile.json: onboarding selects use case, enhancements, and style, then distills experience rules, style constraints, and template requirements into a ≤800-character expert_prompt. Before running, it checks models.json for custom or third-party endpoints and asks whether to enable optimization.

• Domain hit: inject the expert prompt when task keywords match domain_keywords.
• Style enhancement: generate instructions from structured_output, pareto, fast_direct, or detailed.
• Retrieval enhancement: call rag_search or conversation_search on domain tasks when enabled in preferred_tools.
• Plan contract: for medium or complex tasks, declare Goal, Input, Dependencies, Steps, Risks, Tool Needs, and Acceptance, then validate field presence.

Boundaries

It does not simulate reasoning chains or hard-code tool names. Orchestration, tool use, and reasoning still depend on the framework and model capability flags. Simple tasks skip injections, and unconfirmed runs execute the original task to avoid false triggers.