lemonharness

Smart commit with Conventional Commits: groups logical changes, infers type + scope from file paths, and generates messages in `<type>(<scope>): <description>` format.

Iterative quality gate hardening loop — run the gate, parse failures, fix each failing category with a fresh-context delegate, re-run until all gates pass or diminishing returns are hit.

Relentless quality loop: alternate between a fresh-context implementer and a fresh-context reviewer until only diminishing-return issues remain. The reviewer is advisory-only (no state changes). The implementer starts fresh each cycle, unanchored from prior attempts. Terminates when max severity ≤ 3 for two consecutive cycles or the severity trend is flat.

Compact the current conversation into a handoff document for a fresh agent to pick up next session. Includes current phase, workspace state, memory state, and suggested skills.

Investigate a question against primary sources (official docs, arXiv papers, Semantic Scholar, source code, specs, first-party APIs) using the web_search tool. Writes structured findings as a Markdown file. Use when the user wants a topic researched, docs gathered, API facts investigated, or reading legwork delegated. Maps to P1 Explore phase.

Meta-cognitive improvement loop: reflect on errors, learn patterns, codify guardrails. The skill that rewrites skills. Always loaded.

Contract-first API design: schema-first (OpenAPI/gRPC/GraphQL), backward compatibility, consistent error responses, idempotency, cursor pagination, versioning. Use when designing or evolving an HTTP, gRPC, or GraphQL API.

Computational biology guardrails: stable API retrieval from biological databases, biological validity constraints (amino acid / DNA alphabets), synthesis constraints (GC content, repeats), and data provenance. Use for protein design, DNA/RNA analysis, molecular dynamics.

Database engineering: schema migrations (expand-migrate-contract), N+1 prevention, indexing discipline, connection pooling, transaction boundaries, and data integrity constraints. Use for SQL/NoSQL tasks involving persistence, ORMs, or data modelling.

Engineering guardrails for every task: TDD (red-green-refactor), simplicity (KISS/YAGNI), complexity thresholds, and code review readiness checks. Always loaded.

Failure-handling patterns for distributed systems: timeouts on every external call, retry with exponential backoff, circuit breakers, bulkheads, and graceful degradation. Use for microservices, reliability-critical components, or any system with network deps.

Distinctive visual UI design: research the subject's own visual vernacular, build a token system (color, type, layout, signature), critique against defaults, then code. Use when building or reshaping a user-facing UI where templated defaults won't do.

Game development rules: formal rule encoding, immutable state transitions, bounded AI search (minimax/MCTS), and independent win/draw/loss verification. Use for board games, card games, turn-based strategies, or any game with defined rules.

Cross-cutting execution guardrails for all tasks: workspace discipline, verification closures, quality gates, and data integrity rules. Always loaded as the base skill.

ML reproducibility and artifact discipline: train/val/test separation, random seed recording, metric validation against task goals, checkpoint management. Use for any ML, DL, or data science task.

Production observability: structured logging (JSON, trace context), RED metrics (rate/errors/duration) per endpoint, USE metrics for infrastructure, distributed tracing, health checks, SLOs and burn-rate alerting. Use for any service-oriented or production-deployed system.

Secure-by-default engineering: OWASP Top 10 prevention, input validation (allowlist, canonicalize, reject), auth (short-lived tokens, least privilege), secrets management, dependency hygiene, output encoding, security headers. Use for any task involving user input, network access, or data storage.

Disaster response and systems recovery: controlled execution paths, backup-first discipline, build/runtime probes, and checkpoint-based rollback. Use for debugging crashes, data recovery, failover scenarios, or infrastructure repair.

Testing pyramid strategy: property-based over example-based, contract testing at boundaries, mutation coverage, fuzzing for unstructured inputs. Use when correctness and regression resistance matter.

Computer vision and media processing: CLI interface stability, coordinate alignment (row-major vs column-major), shape consistency after transformations, and output validation. Use for image processing, video analysis, object detection, segmentation.

Reference for writing and editing skills well — the vocabulary and principles (predictability, information hierarchy, leading words, completion criteria, progressive disclosure) that make a skill predictable and low-load. Use when creating a new skill, reviewing an existing one, or pruning a skill that has grown stale.

Dual-representation memory and learning system for code agents. Implements event-sourced experience logging, pattern distillation, pre-action governance, and risk-sensitive retrieval.