| name | workstream-about |
| description | Use when working with repository workstreams stored under `.workstreams/` and you need to understand their structure, files, lifecycle, and execution loop. Trigger phrases: "workstream about". |
Workstream About
Overview
Use this skill to understand the workstream model used in a repository.
A workstream is a unit of work stored under:
<repo>/.workstreams/<workstream-name>/
Each workstream has a few files:
.workstreams/<workstream-name>/design.md
.workstreams/<workstream-name>/plan.md
.workstreams/<workstream-name>/tasks.json
.workstreams/<workstream-name>/activity.json
.workstreams/<workstream-name>/review.md
Treat the workstream folder as the durable planning surface for that unit of work.
Workstreams are the primary planning surface for this unit of work.
.workstreams folder is usually gitignored in projects.
What the Files Mean
design.md captures discovery work:
- what is being built
- user intent
- constraints
- relevant architecture notes
- research findings that will help execute tasks
plan.md captures the approved execution design:
- ordered waves
- behavioral specs
- acceptance criteria
- scenarios
- verification
- per-wave review gates
tasks.json captures the durable execution ledger:
They strictly follow this type definition:
type Task = {
id: string;
name: string;
category: "setup" | "feature" | "testing" | "bugfix";
description: string;
acceptance_criteria: string[];
verification: string[];
steps: string[];
done: boolean;
};
type Wave = {
id: string;
name: string;
review_gate: string[];
checklist: Task[];
};
type Tasks = {
must_read_files: string[];
waves: Wave[];
};
There is a task example in tasks.example.json
review.md captures the execution review:
- contains details from a separate agent that reviewed recently completed work for this workstream
activity.json captures the execution's history:
- used by the agent as a memory of accomplished things during a workstream execution
- updated whenever agent does something meaningful or has an important thought or important finding
- initialize it to
[] if it does not exist yet
type ActivityFile = ActivityEntry[];
type ActivityEntry = {
agent: string;
at: string;
task: string;
message: string;
next_step: string;
};
There is a task example in activity.example.json
Workstream Lifecycle
Workstreams usually move through this sequence:
- understand the workstream concept
- brainstorm the workstream and write findings into
design.md
- build the ordered wave plan in
plan.md
- build or refresh
tasks.json from plan.md and review.md if it exists
- execute all waves serially from
tasks.json, while completing tasks inside each wave in parallel
- after each wave, satisfy that wave's review gate before starting the next wave
- run a fresh
workstream-review session after execution completes
- if review finds follow-up work, refresh
tasks.json and repeat the loop
Execution and review are both manually triggered by the user, typically in separate headless agent sessions.
Do not assume workstream-execute automatically invokes workstream-review.
Do not assume workstream-review automatically follows execution unless the user starts that session.
design.md informs plan.md.
plan.md is the source of truth for planning intent.
tasks.json is the source of truth for execution state.
When to Use
Use this skill when:
- the user refers to a workstream
- you need to brainstorm a workstream, execute a workstream or review a workstream.
- you need to execute a workstream from its track
- the structure or meaning of workstreams is unclear
Do not use this skill when:
- the work is unrelated to the repository's
.workstreams/ system
- you already have the full workstream model in context and do not need to reload it
Rules
- Keep all workstream-specific context inside the matching workstream folder.
- Agents may create
.workstreams/<name>/ when starting a new workstream.
- If the intended workstream name or path is ambiguous, resolve it with the user before writing.
- Do not treat ad hoc notes outside
.workstreams/<name>/ as the source of truth over design.md, tasks.json, activity.json or review.md.
- If another workstream skill applies, use this skill first or alongside it when the workstream model needs clarification.
- Treat
workstream-execute and workstream-review as user-invoked headless-session steps in the loop, not as automatic nested handoffs.