Teams

opus killer feature

Quick reference

Field	Value
Command	`/brewcode:teams`
Arguments	`create <name> [prompt]`, `status <name>`, `update <name>`, `cleanup <name>`
Model	opus
Tools	Read, Write, Edit, Glob, Grep, Bash, Task, AskUserQuestion

Modes

Create

Step	Action
1	Analyze project with 3-5 Explore agents
2	Propose 3 team variants (Minimal/Balanced/Maximum)
3	User selects variant and default model
4	Create agents in batches of 3-4 via agent-creator
5	Set up `team.md` + `trace.jsonl` in `.claude/teams/`

Status

Step	Action
1	Read `trace.jsonl` via `trace-ops.sh read`
2	Calculate per-agent success rates
3	Classify health: green/yellow/red
4	Report issues and insights from trace data

Update

Step	Action
1	Analyze performance data
2	Identify underperformers
3	Tune instructions or replace agents
4	Update team.md roster

Cleanup

Step	Action
1	Archive old trace data to `trace-archive.jsonl`
2	Remove inactive agents
3	Reset trace cursor

CREATE Flow

Project AnalysisC1
3-5 Explore agents scan code structure, tech stack, CI/CD, domain boundaries in parallel.
Team ProposalC2
3 variants: Minimal (5) / Balanced (10-12) / Maximum (15-20). User confirms via AskUserQuestion.
Model SelectionC2.5
Choose default model: Opus (recommended) / Sonnet / Haiku / Mixed per-agent.
Agent CreationC3
agent-creator x N in batches of 3-4 parallel. Each agent gets unified template + text-optimize.
Framework SetupC4
team.md + trace.jsonl in .claude/teams/ + verify-team.sh validation.

Task Acceptance Protocol

🛡️

Domain Check

Agent evaluates: “Is this my domain?” If not, refuses and suggests a better-suited colleague from the team roster.

🔍

Duplicate Check

Agent checks: “Has this already been done?” If yes, refuses and links to the existing result in trace.jsonl.

👫

Best Candidate

Agent considers: “Is a colleague better suited?” If yes, refuses and redirects to the recommended agent.

Framework Files

📝

team.md

Agent roster: names, domains, missions, health status. Source of truth for the team composition.

📊

trace.jsonl

Unified JSONL log: tasks (took/refused/completed/failed), issues (by severity), insights (by category). Agents write via trace-ops.sh add (Bash append, ~96% token savings). Session-scoped via SID.

Agent Health

Health	Criteria	Action
Green	>70% success rate, active in recent tasks	No action needed
Yellow	30-70% success rate or many refusals	Review instructions, consider tuning
Red	<30% success rate or inactive	Replace or remove via update mode

Integration with Core Skills

Team agents are automatically preferred by /brewcode:spec, /brewcode:plan, /brewcode:start when matching domains are found. Priority: team agent → project agent → plugin agent. Agents can refuse tasks and suggest a better-suited colleague.

Skill	Relationship
`/brewcode:setup`	Run first to analyze project structure
`/brewcode:spec`	Create task specifications for agents to execute
`/brewcode:plan`	Build execution plans from specs
`/brewcode:start`	Execute plans using the created team