workstream-create

Workstream Create

Create a new workstream with directory structure, shared project sandbox, and per-agent registrations. This is a captain skill — guides the principal through substance after scaffolding.

Arguments

• $ARGUMENTS: <name> — workstream name in kebab-case. Followed by:
  • --agent name[,class] — agent name and optional class (default: tech-lead). Repeatable — one flag per agent.
  • --description <text> — one-line description of the workstream
  • --worktree — also create a git worktree for this workstream
  • --scaffold-only — skip the Phase B guided discussion instruction

If no --agent flags provided, create a single agent with the workstream name and tech-lead class.

Examples

/workstream-create mdpal --agent mdpal-cli,tech-lead --agent mdpal-app,tech-lead
/workstream-create ops --agent ops-lead
/workstream-create folio --worktree
/workstream-create ci-tools --scaffold-only

Model

Value stream (repo) → workstreams (standalone deployable units) → 1+ agents as workers

One workstream = one PVR, one A&D, one Plan, one shared project sandbox, 1+ agents.

Phase A: Deterministic Scaffold

Step 1: Parse and validate

1. Parse workstream name from first positional arg.
2. Parse all --agent name[,class] flags. Split on comma: first part is agent name, second (optional) is class (default: tech-lead).
3. Parse --description, --worktree, --scaffold-only if present.
4. If $ARGUMENTS is empty, ask for a workstream name.
5. Validate workstream name: must match ^[a-z0-9][a-z0-9_-]*$, max 32 chars.
6. Validate each agent name: same regex.
7. Validate each agent class: check agency/agents/<class>/agent.md exists.
8. Check agency/workstreams/<name>/ doesn't already exist.

Step 2: Detect principal

Find the principal from agency/config/agency.yaml — match $USER key in principals: section, use its name field.

Step 3: Create workstream directory

agency/workstreams/<name>/
  qgr/                      — quality gate receipts
  rgr/                      — release gate receipts
  drafts/                   — WIP before ratification (per {P}-{A} subdir)
  research/                 — MARFI outputs, investigations
  transcripts/              — 1B1 records, summaries, verbatim
  history/                  — superseded current versions
  history/flotsam/          — uncategorized archive
  KNOWLEDGE.md              — patterns, conventions, key decisions

KNOWLEDGE.md scaffold:

# <name> — Workstream Knowledge

## Patterns and Conventions

<!-- Accumulate patterns discovered during development -->

## Key Decisions

<!-- Record architectural and design decisions with rationale -->

Step 4: Create shared project sandbox

usr/<principal>/<agent>/
  tools/                    — agent-written scripts (persisted, reusable)
  tmp/                      — scratch space (gitignored)
  history/                  — personal archive
  history/flotsam/          — uncategorized personal items

Create tmp/.gitignore with * and !.gitignore.

Note: Shared content (seeds, transcripts, research, receipts) lives in agency/workstreams/<name>/, NOT here. Per-agent handoffs are flat files: {agent}-handoff.md in the agent directory. Written by agent-create in Step 5.

Step 5: Create per-agent registrations via agent-create

For each --agent name,class flag, call the agent-create tool:

./agency/tools/agent-create <agent-name> <workstream> --type=<class> --description="<description>"

agent-create is the single write path for .claude/agents/ registrations, workspace scaffolding (tools/, tmp/), and bootstrap handoff templates. Do NOT write registrations inline — always delegate to agent-create.

Step 6: Create worktree (optional)

If --worktree was passed:

./agency/tools/worktree-create <name>

Step 7: Report + Phase B instruction

Output the scaffold summary:

Workstream created: <name>

  agency/workstreams/<name>/          — workstream directory
  usr/<principal>/<name>/             — shared project sandbox
  .claude/agents/<agent-1>.md         — agent registration (<class>)
  .claude/agents/<agent-2>.md         — agent registration (<class>)

If --scaffold-only was NOT passed, append the Phase B instruction:

Bootstrap handoffs contain TODO placeholders. Run /discuss with these items:
  1. What is this workstream?
  2. What does <agent-1> own?
  3. What does <agent-2> own?
  4. What seeds do we have?
  5. What's the first action for each agent?

Then update each handoff with the resolved content:
  - usr/<principal>/<name>/<agent-1>-handoff.md
  - usr/<principal>/<name>/<agent-2>-handoff.md

If --scaffold-only was passed, append:

Next steps:
  1. Add seed materials to agency/workstreams/<name>/seeds/
  2. Fill in bootstrap handoffs at usr/<principal>/<name>/<agent>-handoff.md
  3. Launch an agent: claude --agent <agent-name>

What this does NOT do

• Write handoff substance — the captain fills in handoffs via Phase B guided discussion
• Write PVR/A&D — part of the methodology flow, done by the launched agent
• Drop seed material — user responsibility