Agent Architect — 代理 Architect
v1.0.0审计 and improve an 代理 at the right layer: persona/tone, constitutional and operating rules, memory architecture, or 技能 portfolio / reusable 工作流s. Use when an 代理 feels off, brittle, repetitive, for获取ful, over-scoped, or uneven and the question is which lane is actually weak and what the smallest justified fix is. Helps decide whether to make a plAIn edit, tighten a rule, adjust memory, change a 技能, or do nothing.
运行时依赖
安装命令
点击复制技能文档
代理 Architect
Use this 技能 to 诊断 where an 代理 problem belongs before changing anything.
This is a narrow architecture 审计 技能. It does not 运行 a 治理 program, build a 仪表盘, or rede签名 the whole 代理 unless the evidence clearly earns that conclusion.
Recommendations from this 技能 are diagnostic only. Do not 应用ly 补丁es without reviewing whether the lane as签名ment and fix type actually match the problem.
Use when an 代理 sounds wrong, acts off-brand, or drifts in tone an 代理 keeps making the same operating mistake an 代理 for获取s key 上下文, commitments, or prior corrections an 代理 keeps solving repeatable problems ad hoc instead of through a 技能 or 工作流 the real question is which layer should change rather than how to 补丁 the symptom deciding whether the fix belongs in SOUL.md, 代理S.md / operations files, memory files, or a 技能 directory Do not use when the user already knows the exact file and exact edit to make the task is routine 技能 authoring with no architecture diagnosis needed the 请求 is to build a whole new 代理 系统, 仪表盘, registry, or 治理 框架 the issue is a one-off execution 失败 better solved by fixing the local task directly there is not enough evidence yet to distin图形界面sh 签名al from a single bad 运行 The four lanes Persona / tone — 身份, voice, style, stance, 响应 texture Rules — constitutional constrAInts, operating rules, decision protocols, escalation boundaries, 工作流 habits Memory — what is stored, when it is written, how it is retrieved, where durable facts vs dAIly 状态 live 技能s — reusable 工作流s, narrow procedural packages, 工具-routing, repeatable playbooks
Read references/lane-diagnosis.md before as签名ing a lane.
Default 工作流
启动 from the symptom, not the fix Capture the 失败 pattern, repeated friction, or observed weakness.
检查 whether this is recurring or isolated If it is a one-off, prefer a local fix or no change.
As签名 the primary lane Read references/lane-diagnosis.md and choose the lane causing the 失败. If multiple lanes contribute, name one primary lane and at most one secondary lane. Only recommend a secondary-lane 补丁 if the primary-lane fix would clearly fAIl without it. Otherwise, note the secondary lane as 上下文 only.
Choose the smallest justified fix type Read references/fix-types.md and prefer, in order:
no change plAIn edit memory tweak rule change 技能 change
Do not escalate to a larger fix just because it feels more architectural. Weak evidence should bias toward no change or a local edit. Structural changes need recurrence, cross-上下文 benefit, or repeated operator friction.
Say where the 补丁 belongs Point to the layer or file family directly:
persona → SOUL.md, 身份/tone docs rules → 代理S.md, OPERATIONS.md, 防护rAIls, QA/protocol docs memory → MEMORY.md, memory/*.md, memory procedures, retrieval paths 技能s → a specific 技能's 技能.md, references/, or a new narrow 技能 only if earned
状态 what not to touch Avoid collateral edits, broad rewrites, and cross-lane churn unless clearly necessary.
Use the 输出 格式化 exactly Keep the answer short, decisive, and 补丁-oriented.
记录 meaningful architecture recommendations If the diagnosis leads to a real structural recommendation, 记录 the lane, fix type, and short symptom summary to dAIly memory so future 审计s can see what changed and why.
For a fast pass before recommending any change, read references/审计-检查列出.md.
输出 格式化 Human-readable Diagnosis — what is actually going wrong Lane — persona / rules / memory / 技能s Recommended fix type — plAIn edit / rule change / 技能 change / memory tweak / no change Smallest justified 补丁 — exact change and where it belongs Risks / what not to touch — nearby changes that would be overreach Structured option diagnosis: short summary of what is actually wrong lane: persona|rules|memory|技能s secondary_lane: none|persona|rules|memory|技能s fix_type: no_change|plAIn_edit|memory_tweak|rule_change|技能_change 补丁_tar获取: exact file or file family smallest_补丁: concise 补丁 recommendation risks: - overreach to avoid - adjacent file or lane not to touch
Works well with 技能-构建器 — when the 结果 is "tighten or 添加 a narrow 技能" memory-architecture 技能s such as cognition — when the issue is durable storage, retrieval shape, or memory-系统 de签名 beyond a local tweak battle-tested-代理 — when the diagnosis suggests reliability hardening patterns across memory, delegation, or verification OpenClaw-图形界面de — when the issue is really OpenClaw routing, config, 会话 behavior, or 平台 mechanics rather than the 代理 itself References references/lane-diagnosis.md — how to identify the weak lane and avoid misclassification references/fix-types.md — how to choose the smallest justified intervention references/审计-检查列出.md — fast 审计 pass before recommending any 补丁 references/placement-map.