AutoSaddler is a system that improves LLM-agent harnesses by diagnosing execution traces and applying structured changes to prompts, tools, middleware, and agent-loop logic. It evaluates candidate updates for their ability to generalize beyond the traces that motivated them. The catalogue add-ons represent workflows for using AutoSaddler.
This repository also configures its own agents. See what AutoSaddler tells them →
Skill Claude CodeCodex ✓ vendor
Selected during the capability phase — adds new tools, exposes new parameters, fixes tool implementations, or removes infrastructure limitations to expand what the agent can do.
Skill Claude Code ✓ vendor
Use to systematically diagnose why scenarios fail — identifies root causes by analyzing agent traces, codebase, and behavioral patterns.
Skill Claude CodeCodex ✓ vendor
Use to systematically analyze the full evolution history without output truncation — produces a structured summary of what matters for the current iteration.
Skill Claude Code ✓ vendor
Use after applying any patch to thoroughly verify the agent codebase has no bugs — checks syntax, imports, runtime behavior, docstrings, templates, hooks, and logical correctness.
Skill Claude Code ✓ vendor
Selected during the steering phase — refines agent behavior by modifying prompts, tool descriptions, or hook reminders without changing executable code.
Skill Claude CodeCodex ✓ vendor
Use before proposing a mutation to diagnose measured failures from evaluator outcomes, traces, history, and the controlling candidate surfaces.
Skill Claude CodeCodex ✓ vendor
Use after applying a candidate change to verify its diff, scope, syntax, integration behavior, and consistency before returning it.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: