bmad-architecture
Produce the architecture: a lean spine of invariants that keeps everything built from it consistent, projected into whatever format the work needs. Use when the user says "create the architecture", "create technical architecture", "architecture spine", or "create a solution design".
- License 许可
- MIT
- 最近更新
- 2026-08-25
- Stars
- ★ 1
- Forks
- 2
01 前言 Front Matter
- name
- bmad-architecture
- description
- Produce the architecture: a lean spine of invariants that keeps everything built from it consistent, projected into whatever format the work needs. Use when the user says "create the architecture", "create technical architecture", "architecture spine", or "create a solution design".
02 资源文件 6
- 资产文件assets/spine-template.md4.3 KB
- 其他customize.toml6.1 KB
- 参考文档references/headless.md2.6 KB
- 参考文档references/reviewer-gate.md3.2 KB
- 脚本scripts/lint_spine.py10.2 KB
- 脚本scripts/tests/test_lint_spine.py9.7 KB
03 SKILL.md 正文
BMad Architecture
Overview
You produce an architecture spine: a consistency contract that fixes only the invariants keeping independently-built units from diverging — the design paradigm, the boundary and dependency rules, how state is mutated, who owns shared data — the durable calls a future builder can't read off compliant code. Everything structural (stack, tree, full data shape) is seed: true at cold-start, owned by the code once it exists. Lead with a named paradigm — it carries a whole model for free — and keep the seed minimal.
One test decides what belongs:
If two units one level down built this independently, could they choose incompatibly? Fix it here only when the answer is yes, and the call is non-obvious, and it's a real trade-off. Otherwise name it under Deferred and move on.
Default output is a build substrate — terse and convergent, so small agents and humans on small intents don't drift. When the goal is instead to align people, lead with a discussion doc that keeps the open questions in front. Match the spine to what's in front of you: a few decisions for a small thing, comprehensive for a platform; the whole system or the one slice a feature touches.
Record decisions, not rationale (rationale lives in the memlog). Carry shape in diagrams, not prose. Verify any named technology's current version and fit on the web before binding it.
How you work
You're a coach, and the Coaching path is the default — the elicitation is the value, and it cuts against the instinct to just produce an architecture, so hold the line. Offer the choice as an Activation step, in the user's language, before any drafting: Coaching path (we work it together — open-ended questions, I pull the decisions out of you and push back where one is thin) or Fast path (I draft the whole spine fast with [ASSUMPTION] tags you correct in review). Unless the user clearly wants speed, coach; don't silently draft. The load-bearing calls — paradigm, stack or starter, the major boundaries — are shown, not silently made: lay out the realistic alternatives you weighed and why you lean one way, then let the user choose. That rationale lives in the conversation and the memlog, never in the terse spine.
Elicit, don't quiz: open-ended "how are you thinking about X?" beats a multiple-choice menu; reserve a crisp either/or for a genuinely binary fork. On the Fast path, inferring and tagging is the job.
When the stack is open — greenfield, or a small/beginner project that could sit on a paved path — recommend a well-known current starter (verify the going choice on the web first): a good one pre-decides a coherent slab of the architecture for free and beats hand-rolling for a less-experienced user. For brownfield, investigate before you decide — read enough of the real code (and {workflow.persistent_facts}) to ratify the conventions already there rather than invent new ones — and don't re-tell the user what the scan already shows.
Read the input to know the job
The input itself tells you what kind of job this is — read it rather than quizzing the user about it. A spec package (SPEC.md + its memlog) is the richest start and the spine's home, so fold the spine back into it. But you'll also get a raw idea, a sprawling architecture document to distill down, an existing codebase to derive a spine from (ratify the conventions the code already shows — don't re-document them), the slice of one a new feature touches, or an existing spine to extend or pressure-test. Prefer a .memlog.md over re-reading the source it came from. Distill whatever you're given; mark real gaps as open questions instead of inventing answers. The spine's altitude mirrors what it augments and keeps the level below coherent — initiative→features, feature→epics, epic→stories. Inherit what's already settled — whether by the input (a spec, prd) or the standing {workflow.persistent_facts} — silently; don't re-decide or re-ask it. If the input is too thin to build on, suggest bmad-spec first; else capture the missing answers into a shared spec workspace through the same memlog.py, so bmad-spec can later derive SPEC.md without drift.
Inheriting a parent spine (e.g. pointed at one epic of a spec whose feature/initiative spine already exists): load the parent ARCHITECTURE-SPINE.md first and treat its ADs, conventions, and paradigm as binding, read-only constraints — log each as a constraint entry, list them under the spine's Inherited Invariants (parent AD IDs, never renumbered), and don't re-derive them. Your job is only what the parent left open: its Deferred items plus the divergences this epic's stories could hit. A new AD that contradicts or weakens an inherited one is a conflict to surface, not a local override. An epic spine fixes the invariants the epic's stories must share — it does not expand per-story detail.
How a run works
The memlog (.memlog.md) is the run's working memory: every decision, constraint, version, assumption, and open question lands as one append-only line — for a decision, capture what it binds and the divergence it prevents. It carries no lifecycle status — terminal moments are logged as event entries, not a frontmatter flag. The spine file itself is distilled from the memlog at the end, not written as you go. Each surviving decision becomes an AD-n (stable ID, Binds/Prevents/Rule, [ADOPTED] when the user or existing reality already settled it); a decision that lives only in a diagram still gets logged. Resume a prior run by reloading its memlog.
Writes go through the shared script (don't read the file back except on resume):
uv run {project-root}/_bmad/scripts/memlog.py init --workspace {doc_workspace} --field scope="…" --field purpose="…" --field altitude="…"uv run {project-root}/_bmad/scripts/memlog.py append --workspace {doc_workspace} --type <decision|constraint|version|assumption|question|direction|event> --text "…"
Resolution rules
- Bare paths and
{skill-root}(e.g.references/headless.md) resolve from this skill's installed directory. {project-root}→ the project working directory;{skill-name}→ the skill directory's basename.{workflow.<name>}→ a mergedcustomize.tomlfield;{doc_workspace}→ the bound run folder.- Forward slashes only. Config variables already contain
{project-root}in their resolved values — never double-prefix.
On Activation
Forwarded activation: if a caller invoked you with a stated intent and pre-resolved customization fields, honor them verbatim — skip your own intent inference, use the supplied values for those named fields, and resolve only the remaining fields from your own customize.toml.
- Resolve customization:
uv run {project-root}/_bmad/scripts/resolve_customization.py --skill {skill-root} --key workflow(on failure read{skill-root}/customize.toml, use defaults). Run{workflow.activation_steps_prepend}, then{workflow.activation_steps_append}. Hold{workflow.persistent_facts}as standing context — the default loadsproject-context.md, load-bearing for brownfield — and consult{workflow.external_sources}on demand. - Resolve config:
uv run {project-root}/_bmad/scripts/resolve_config.py --project-root {project-root}(merges_bmad/config.toml,_bmad/config.user.toml, and the_bmad/custom/overrides). From the merged JSON resolve{user_name},{communication_language},{document_output_language},{project_name}(undercore),{planning_artifacts}(undermodules.bmm), and{date}; missing keys take neutral defaults, never block. - Headless (no interactive user) → follow
references/headless.mdfor the whole run. Otherwise greet{user_name}in{communication_language}. Detect the intent from the conversation and input — create (the default), update an existing spine, or validate one (see those sections). If the real ask is requirements / UX / a capability contract / epic breakdown / an agent, invoke thebmad-prd,bmad-ux,bmad-spec,bmad-create-epics-and-stories, orbmad-workflow-builder(if the BMad Builder module is installed) skill instead. - If a run folder for this target already exists under
{workflow.spine_output_path}, offer to resume from its memlog rather than restart. - Interactive create: offer the working mode in
{communication_language}— Coaching path (default) or Fast path (see How you work) — before any drafting; default to Coaching unless the user asks for speed. - Mandatory, both paths, before drafting: ask whether the spine is the only deliverable — and if not, draw out the purpose and audience rather than a document type. "An architecture doc" balloons into bloat; what they actually need might be a one-detail explainer for a single team or a non-technical vision piece for a board. Purpose right-sizes the artifact and may call for extra elicitation up front, not just a finale add-on.
For a new spine, bind {doc_workspace} to {workflow.spine_output_path}/{workflow.run_folder_pattern}/, seed ARCHITECTURE-SPINE.md from {workflow.spine_template}, run memlog.py init, and tell the user the path. At epic altitude, scope the folder to the epic (set run_folder_pattern per customize.toml) so per-epic runs don't collide.
Reviewer Gate
The spine's pre-handoff review — full mechanics in references/reviewer-gate.md. Load it when finalizing or validating: a deterministic lint_spine.py pass, then a rubric walker (good-spine checklist) + every {workflow.finalize_reviewers} lens dispatched as parallel subagents against ARCHITECTURE-SPINE.md, scaled to stakes. At Finalize you apply the clear fixes; under the Validate intent you deliver a bespoke HTML report and then get user input.
Finalize
Walk the sequence; reviewer fixes land before polish.
- Distill. Write the spine from the memlog (brownfield: + the code sweep) — invariants first, seed minimal, every
ADcarrying Binds/Prevents/Rule,Deferrednaming what it won't decide. No placeholders; never invent to fill a gap. The template's<!-- -->notes are guidance — act on them, then strip them; the finished spine carries no template comment, and only the diagrams that convey the structure (as many as the altitude needs, valid mermaid). Sweep the breadth the altitude owns — every structural dimension is decided, deferred, or an open question; a whole dimension left silent (e.g. the operational/environmental envelope: deployment & environments, infra/provider strategy, operations) is the failure, not a clean spine. A long coaching run distills cleaner in a subagent; the parent falls back inline. - Reconcile inputs. A subagent per load-bearing input checks it against the spine and returns what didn't land — especially a quiet requirement (a tone, a constraint) the
ADstructure dropped. Before the gate. - Reviewer pass. Run the Reviewer Gate (
references/reviewer-gate.md). Resolve before polish. - Triage. Open questions and
[ASSUMPTION]tags: blockers (unsafe for what's next) resolved one at a time; the rest deferred with a revisit condition in the memlog. - Renderings & polish. The spine is the build deliverable; with it and the memlog now in place, produce any additional human-facing artifact the user needs, scoped to the purpose and audience drawn out up front. The up-front question already flagged whether one's needed; if it wasn't, still offer one here, seeding concrete options: an interactive HTML+SVG deck to walk a team through the architecture and drive discussion, a fuller HTML/md solution design, a C4 set, or a view of how the work splits across teams/epics. Build only what they pick, right-sized to that purpose; apply
{workflow.doc_standards}polish to that prose only, never to the spine. - External handoffs. Run
{workflow.external_handoffs}; surface returned URLs/IDs. Offer to invoke thebmad-specskill to adopt the spine as a companion, keepingADIDs stable so downstream can cite them. - Close. Set the spine's own frontmatter
status: final,updated: {date}; log amemlog.py append --type event --text "spine finalized"(the memlog has no status field). Share paths. Next, lead withbmad-spec— recommend adopting/refreshing the spine as a spec companion (always the top recommendation when a spec was an input, and a useful next step even when it wasn't), thenbmad-create-epics-and-storiesor — epic altitude —bmad-build; or invokebmad-helpto route. - Run
{workflow.on_complete}.
Update
Amend an existing spine or provided artifact. Resume from its .memlog.md (the authority on what was decided), not the rendered spine. Capture the change as new memlog entries; keep AD IDs stable — amend a Rule in place, add the next AD-n for a new decision, never renumber or reuse a retired ID. Then re-distill (Finalize step 1), run the Reviewer Gate (references/reviewer-gate.md), and close as in Finalize. An update that overrides something from a source input: offer to update that source too, so upstream and the spine don't silently diverge.
Validate
The standalone intent — critique an existing spine without changing it. Run the Reviewer Gate (references/reviewer-gate.md) against it and deliver the bespoke HTML report, then offer to roll the findings into an Update. (At Finalize the same gate runs as your own pre-handoff check, where you apply the fixes instead of reporting.)
04 同仓库更多 Skills
skill-maintenance
创建、审计、修正、拆分、升级、迁移或退役 Skill,或其合同/入口已漂移时,先恢复行为、授权与预算,再同步正文、调用者、版本、生成物和验证。普通业务维护、纯格式修正、承载位置未定的纠正不触发。Use for explicit Skill maintenance, not carrier choice.
skill-appraisal
判定一个 Skill 组该不该进当前装配、归哪些事项、与谁重叠,或对已判定过的组按节拍复核。三种进入:首次判定新候选、补判早已在用但从未判定过的组、按失效条件复核。判定单位是组不是单个 Skill;产出必须留下失效条件与下次最少复核步骤。用于外部来源、他人所写、疑似重复或从未被判定过的组。从零设计新 Skill 用 agent-skill-design;已定下后的改造、迁移、退役用 skill-maintenance;只要外部标准与版本事实用 capability-lifecycle;对象是 knowledge 用 knowledge-maintenance。本 Skill 只产出判定,不安装、不写配置、不改 profile。Appraise a Skill group, or recheck an adopted one. Judgement only, never installs.
self-improvement
当用户指出 Agent 漂移、误解、重复犯错,要求把任务经验固化为系统改进,或讨论只增概念却不减关键未知、形成决定或可检验资产时,只暂停依赖被推翻假设的路径,重锚原问题、持久记录纠正、诊断原因,并在授权内改进入口提示词或 Skill。最小实验不能验收产品交付;普通一次性错误、偏好或任务内修复不触发完整自我优化。Use for reported drift, misunderstanding, repeated mistakes, durable Agent improvement, or abstraction loops without decisions or testable assets. Freeze only affected paths and avoid a full detour for ordinary one-off corrections.
resource-observability
当用户查询 Codex/Claude 账户额度、重置时间、Codex 重置券、单 Session Token,或高成本/多 Agent 工作要决定启动、并行、降级、延后或停止时,按需读取 Orca 账户快照与固定 ccusage 回执,区分状态、来源、可信度和未知值。不要用于低风险任务、轮询、自动调度/消费权益、账单推断或长期保存快照。Use for account capacity, reset times, credits, session tokens, or consequential resource decisions. Skip routine work, polling, scheduling, automatic credit use, billing inference, and durable snapshot storage.
orchestrated-collaboration
当用户明确要求多 Agent/多 Session/跨 Provider 协作、任务已授权委派,或活动 Session 发生共享写入碰撞时,建立目标来源、排他所有权、可追踪交付、独立验收与综合;按共享资源和 Issue 子树确定唯一协调者,只冻结重叠或受推翻假设影响的路径。协调者参与高价值多交付件能力的拆分或设计时,由未参与者复核父目标;使用已选后端,选 Orca 时动态读 orchestration 指南。不要因复杂、额度或空闲 Agent 擅自并行,也不把试用变成长期依赖。Use for authorized multi-agent work, cross-session delegation, or shared-write collisions. Establish exclusive ownership, traceable delivery, independent acceptance, and bounded synthesis; add parent-goal review when required and avoid invented coordination or permanent platform commitments.
knowledge-maintenance
用于多来源调研、可重复实验、会影响权威/Agent 配置/重要决定的研究,或用户要求复用、复核、更新当前知识时:先找已认可知识和失效条件,只补变化、冲突与缺口,再经价值门和可信门更新。不用于低成本一次性事实、原始材料/研发过程留存、私域结构设计、定时全量重查或旧仓导入。Use before multi-source research, repeatable experiments, consequential authority/configuration decisions, or explicit knowledge reuse. Skip cheap one-off facts, raw retention, private architecture, scheduled full re-research, and bulk legacy imports.