mawf-orchestration
Plan and coordinate cost-bounded multi-agent workflows for complex projects using mawf.
- License 许可
- MIT
- 最近更新
- 2026-11-09
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- mawf-orchestration
- description
- Plan and coordinate cost-bounded multi-agent workflows for complex projects using mawf.
03 SKILL.md 正文
MAW Orchestration (skill)
Use when planning or running a multi-agent workflow for a complex codebase.
MAW (multi-agents-workflow) dynamically picks an architecture per project and writes per-agent, independently-editable configs under .mawf/.
When to use
- A new complex project: after
mawf init, runmawf plan. - An existing project where a single agent is insufficient (many files, multiple languages, high risk, large context).
- You want cost-bounded multi-agent runs with Codex review gates.
The four paradigms MAW picks between (and combines)
- Loop engineering — open-ended single tasks; Think→Act→Observe until done.
- Subagents / orchestrator-workers — dynamic, parallelizable subtasks; each subagent gets its own context window and compresses findings.
- Multi-agent — high-value breadth-first work that tolerates ~15× token cost.
- Graph engineering — predictable, inspectable control with HITL, persistence, and review gates.
MAW also understands dynamic (prefer the host's native dynamic-workflow/multi-agent mechanism) and ultracode (complex coding with implement→codex-review→fix loops + graph checkpoints).
Quickstart
mawf init -u <user> # init workspace
mawf plan --project . # probe + plan + write .mawf/
mawf doctor # env + capability check
mawf cost # real cost rate from cc-switch
mawf run # execution guidance (batches + guards)
Core invariants
- Start simple; add complexity only when it demonstrably helps.
- The cost guard is authoritative: $5/min per agent, $10/min total (defaults), measured from real cc-switch spend — not token estimates.
- Codex review is risk-gated, not on every step.
- Every file under
.mawf/is editable;mawf planregenerates from fresh signals.
See the architecture report at docs/ARCHITECTURE.md for the theoretical grounding.
04 同仓库更多 Skills
mawf-planner
Choose an agent architecture from project signals, host capabilities, risk, parallelism, and cost constraints.
mawf-loop
Run open-ended coding or research work with an iterative act, observe, reflect, and decide loop.
mawf-knowledge
Retrieve prior decisions and solutions from the MAWF knowledge store and inject bounded context. Use when planning or taking over a task, before authoring decision/solution notes, or when a symptom smells like a known pitfall.
trellis-brainstorm
Guides collaborative requirements discovery before implementation. Creates task directory, seeds PRD, runs the grill-with-docs interview (grilling + domain-modeling) to resolve the decision tree in rounds, and converges on MVP scope. Use when requirements are unclear, there are multiple valid approaches, or the user describes a new feature or complex task.
mawf-graph
Plan and inspect graph workflows with review gates, branching, checkpoints, and persistent execution state.
mawf-decision
Author and maintain long-lived decision notes in the MAWF knowledge store (write-notes semantics). Use only for non-trivial changes that deserve a durable decision record; never for mechanical edits or task-local design.