cross-task-learner_skill

Overview

This skill enables Ralph loops to learn from similar past tasks and share reusable patterns across concurrent or sequential loops. It extracts error/fix pairs, success approaches, anti-patterns, and code templates on loop completion, and injects ranked, relevant patterns at loop start to accelerate resolution and reduce repetition. Version 2.0 adds multi-loop awareness with loop_id lineage and aggregated effectiveness metrics.

How this skill works

On loop completion the extractor scans loop state, iterations, debug memory, and reflections to identify error signatures, successful step sequences, failure modes, and code templates, then stores generalized patterns in a shared registry. On loop start the injector embeds the task description, performs semantic matching against the registry, ranks patterns by relevance and success rate, and injects the top-k items into the loop context while tracking usage and outcomes for later aggregation.

When to use it

• When a Ralph loop finishes and you want to capture reusable fixes and approaches
• When starting a new loop and you want guidance from prior successful workflows
• To avoid repeating known anti-patterns across teams or sessions
• When measuring cross-loop effectiveness or improving agent heuristics
• During large refactors or debugging waves where recurring errors appear

Best practices

• Enable automatic extraction only for completed loops with sufficient telemetry to avoid noise
• Set conservative thresholds for pattern promotion (min success rate, usage count)
• Prefer templates and patterns with higher usage_count and recent last_used timestamps
• Merge similar patterns above a similarity threshold to prevent duplication
• Track per-loop application outcomes to refine success_rate and effectiveness trends

Example use cases

• Automatically inject a null-check template when a new TypeError pattern matches the task
• Surface an established multi-step process for fixing TypeScript type errors at loop start
• Flag and avoid a commonly observed anti-pattern (premature optimization) during refactors
• Aggregate success_rate trends to prioritize high-impact templates in CI-driven loops
• Manually extract patterns from a critical failure loop for knowledge sharing across teams