The problem

gstack has 11 specialized skills, but you can only use one at a time. When you need security AND risk AND code review on the same PR, you run them sequentially, copy-paste findings between sessions, and synthesize the output yourself. You are the message bus.

With Claude Code Agent Teams (experimental, v2.1.32+), teammates can message each other directly. This PR makes every gstack skill a first-class Agent Team teammate.

What this adds

Every skill now works in two modes — automatically

The {{PREAMBLE}} (shared by all skills) now detects if it's running as a teammate:

Standalone (user invokes directly):
  /cso → outputs findings to conversation

As teammate (lead spawns in a team):
  /cso → messages findings to /risk teammate
       → /risk incorporates into risk register
       → /risk messages to /board teammate
       → /board synthesizes into executive brief

No per-skill changes needed. The preamble handles mode detection, communication protocol, task claiming, and teammate discovery.

/team — 7 pre-built team configurations

How /team diligence actually works

Lead spawns 5 teammates, each reads their SKILL.md:

  "vc"   reads /vc/SKILL.md    → moat analysis, velocity scorecard
  "cfo"  reads /cfo/SKILL.md   → cost model, build-vs-buy
  "cso"  reads /cso/SKILL.md   → OWASP audit, STRIDE model

  vc, cfo, cso work in parallel (no dependencies)

  "risk" reads /risk/SKILL.md  → waits for cso findings
                                → incorporates security into risk register
                                → messages risk register to board

  "board" reads /board/SKILL.md → waits for vc + cfo + risk
                                 → synthesizes into 2-page executive brief

Teammates message each other directly. The lead synthesizes final output.

team/TEAMS.md — coordination reference

Every teammate reads this to understand:

• Message format: FROM, STATUS, TOP FINDINGS, FULL REPORT, ACTION NEEDED
• Dependency graph: ASCII diagram showing who messages whom
• Shared state: .gstack/ directory map for cross-teammate file access
• Anti-patterns: don't edit same files, don't broadcast everything, don't skip the lead

What's NOT in this PR

• No .claude/settings.json committed — users opt in with:

  { "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }

• No changes to existing skill logic — only the preamble gets teammate awareness
• No dependency on PRs feat: add 5 engineering-adjacent skills (conflicts, cso, risk, escalation, ai-hybrid) #106 or feat: add 6 business and communications skills (cfo, vc, board, media, comms, pr-comms) #107 — works with Garry's existing 11 skills

Integration

• Preamble change regenerates all 12 existing SKILL.md files (additive only)
• /team skill follows standard .tmpl → SKILL.md pipeline
• CLAUDE.md updated with Agent Teams section and dependency graph
• LLM-as-Judge evals for orchestration quality, TEAMS.md clarity, preamble actionability, spawn prompt sufficiency

Test plan

• bun test — 114 pass, 0 fail (723 assertions)
• bun run gen:skill-docs --dry-run — all 12 SKILL.md files FRESH
• Preamble teammate awareness identical across all skills (single source of truth)
• EVALS=1 bun test test/agent-teams-llm-eval.test.ts — LLM quality evals (~$0.08/run)