redthread

Instructions for triggering external Context7 connections

Lightweight research-plan-implement workflow for small tweaks, isolated bugfixes, copy edits, and low-blast-radius polish in this repo. Use when the change can be understood from 1-3 files and stays inside one subsystem without changing shared data flow.

Implements tree-based attack branching and prompt refinement optimization.

Fetch external library or framework documentation only when local docs are insufficient.

Guidance for JudgeAgent scoring, guardrail synthesis, and anti-hallucination validation work.

Full research-plan-implement workflow for new features, architectural updates, multi-file changes, or any task with non-trivial blast radius in RedThread.

Audit a plan for missing architecture, evaluation, safety, or verification coverage before implementation.

Execute an approved plan with focused edits, iterative verification, and strict attention to repo boundaries.

Lightweight research-plan-implement workflow for small tweaks, isolated bugfixes, copy edits, and low-blast-radius polish in this repo.

Turn research findings into a small, reviewable implementation plan before editing.

Guidance for attack-algorithm work across PAIR, TAP, MCTS, and Crescendo in RedThread.

Deep repository research workflow for understanding architecture, conventions, and implementation patterns without modifying files.

Algorithms for LLM-as-a-Judge telemetry and rule synthesis.

Structural audit skill invoked against plans prior to execution to resolve missing edge cases.

Isolated task-building phase after approval. Strongly encourages Subagents.

Defines the formal step-by-step logic map before taking action.

Sweeping file aggregation and pattern recognition across the repository. Does not modify files.