Architecture

RedTeam is a skills-only package: markdown instructions + small Node scripts for context loading and command pinning. No server, no API, no model calls from code.

Context diagram

User → AI harness (Claude / Cursor / ChatGPT / Codex / Copilot)
         ↓ invokes
       /redteam <command>
         ↓ loads
       skill/SKILL.md + reference/<command>.md
         ↓ optional
       scripts/context.mjs → CONTEXT.md
         ↓ optional
       .redteam/reviews/*.md + .redteam/sessions/*.md (persisted output)

Components

Component	Responsibility	Depends on
skill/SKILL.md	Command router and principles	reference/*.md
skill/reference/	Per-command flows and TTP catalog	Handbook v10 concepts
skill/scripts/context.mjs	Load CONTEXT.md from project	Node 22.12+
skill/scripts/pin.mjs	Pin command shortcuts	Harness dirs
.redteam/	Decision context and review artifacts	User project
cli/bin/cli.js	GitHub-backed npx installer	scripts/build.mjs
chatgpt/INSTRUCTIONS.md	Custom GPT system prompt	—
.claude-plugin/	Claude Code marketplace manifest	plugin/
scripts/build.mjs	Copy skill to provider directories and dist/ (incl. ChatGPT knowledge bundle)	skill/
scripts/prepare-docs-site.mjs	Stage docs/ for Starlight	docs/
scripts/check-docs-site.mjs	Validate built docs site links	npm run docs:site
scripts/test-cli-install.mjs	Smoke-test CLI install in a temp project	cli/bin/cli.js
src/content/docs/	Starlight site content (guide, handbook, synced developer docs)	Astro / Starlight

Data model

RedTeam has no database. State lives in the user’s project as markdown and JSON files:

Entity	Location	Role
Decision context	CONTEXT.md or .redteam/CONTEXT.md	Background the model loads for anchored reviews
Project config	.redteam/config.json	Shared settings; config.local.json is gitignored
Review artifacts	.redteam/reviews/<slug>-<command>.md	Persisted command output for retrospectives
Session notes	.redteam/sessions/<date>-<slug>.md	Facilitated session records (groupthink, ideate, converge)
Command flows	skill/reference/<command>.md	Read-only technique instructions
TTP catalog	skill/reference/ttp-catalog.md	Browseable technique index

Context resolution order: root CONTEXT.md → .redteam/ → .agents/context/ → docs/ → $REDTEAM_CONTEXT_DIR.

Domain language and boundaries

Domain concept	Meaning in this project	Boundary / owner
Command	A /redteam <name> entry point with a reference flow	skill/SKILL.md + skill/reference/
Technique	A handbook structured thinking method (premortem, ACH, …)	Handbook + TTP catalog
Review artifact	Persisted markdown output of a command	.redteam/reviews/
Decision context	Background constraints for anchored reviews	CONTEXT.md / .redteam/
Provider bundle	Harness-specific copy of the skill	scripts/build.mjs outputs
Developer docs	In-repo BDD lifecycle documentation	docs/

Key flows

Chat-only review

1. User pastes plan and runs /redteam challenge
2. Skill loads principles + challenge.md reference
3. Model produces structured adversarial review in chat

Project-anchored review

1. User runs /redteam init → writes CONTEXT.md
2. Later: /redteam premortem launch → context.mjs loads CONTEXT.md
3. Output saved to .redteam/reviews/launch-premortem.md

Install

1. npx --yes github:CurateLabs/RedTeam install → cli/bin/cli.js runs build.mjs, detects harness dirs (.cursor, .claude, .agents, .github, .gemini) or honors --providers= / --scope=global, and copies the skill in
2. Creates .redteam/ scaffold (reviews/, sessions/, config.json, CONTEXT.template.md) if missing
3. User reloads harness

Cross-cutting concerns

• Error handling: Skill scripts fail fast with clear messages; missing CONTEXT.md does not block chat-only reviews
• Configuration: .redteam/config.json (shared), config.local.json (gitignored)
• Security: No network calls in skill scripts; no exploit content
• Observability: No runtime telemetry; review artifacts in .redteam/reviews/ and GitHub issues serve as the feedback trail. See OBSERVABILITY.md.

Decisions

Significant architectural choices are recorded as ADRs in adrs/. Create the next one with docslime add adr <short-slug>.

Decision	Status	Notes
Skills-only execution (no server, no model calls from code)	Accepted, ADR pending	See overview above
LLM-only quality gate (no deterministic validator)	Accepted, ADR pending	Quality depends on model following reference flows
Handbook techniques summarized, not reproduced verbatim	Accepted, ADR pending	See strategy/source-lineage.md

Risks & trade-offs

• LLM-only execution — quality depends on model following reference flows; no deterministic validator
• Handbook attribution — techniques summarized, not reproduced verbatim
• Multi-provider drift — npm run build must keep provider copies in sync with skill/