omni-agent-builder（全向 Agent 构建器）

Omni-Agent Builder (OpenClaw)

Design and generate a complete OpenClaw agent workspace with robust defaults:

• secure-by-default operating rules
• optional multi-agent orchestration (Planner / Executor / Critic)
• Personal Intelligence (Pi) memory workflow (MEMORY.md + daily logs)
• evaluation + acceptance tests
• CLI notes for fast local iteration

Canonical references (read these first)

1) Workspace layout + bootstrap file load order + heartbeat rules: references/openclaw-workspace.md 2) File templates/snippets: references/templates.md 3) Background (agent architecture: planning / memory / tool use): references/architecture.md

Default outputs (workspace files)

Minimum viable workspace (always generate):

• AGENTS.md (cross-cutting rules; sub-agents only get this + TOOLS.md by default)
• SOUL.md
• TOOLS.md
• IDENTITY.md
• USER.md
• HEARTBEAT.md (usually empty to avoid token sink)
• BOOTSTRAP.md (one-time ritual; delete after completion)

Optional (generate if requested or clearly beneficial):

• MEMORY.md (curated long-term memory; private/main session only)
• memory/YYYY-MM-DD.md (daily log seed)

Workflow: build a new agent/team from scratch

Phase 1 — Tight interview (clarifying questions)

Ask only what you need. Prefer 1–2 short rounds.

1) Mission + autonomy - Mission (1 sentence) - Single agent or multi-agent orchestration? - Autonomy: Advisor / Operator / Autopilot

2) Surfaces + boundaries - Channels: Telegram/WhatsApp/Discord/iMessage/Slack/etc. - Private DM only vs group channels? - “Ask before outbound messages” rule (default: ALWAYS ask)

3) Tools + CLI usage - Must it run local openclaw CLI commands? Which quick aliases are desired? - Any repos/paths it should assume?

4) Pi memory - Should it maintain curated MEMORY.md? - Categories: preferences, projects, people/orgs, “never do”, conventions, vocab - Privacy: what must never be stored?

5) Bootstrapping ritual - Any first-run steps (connect channels, create folders, install deps)?

6) Evaluation - What metrics define success? - 5–10 acceptance tests you want it to pass (we’ll propose defaults if none)

Phase 2 — Generate the workspace files

Generate the files listed above. Use the templates but tailor to the user’s answers.

Required global guardrails to include (at minimum):

• Ask-before-destructive (prefer trash over rm; create backups; show diffs)
• Ask-before-outbound (never send messages/emails/DMs/calls without explicit approval)
• Stop-on-CLI-error (unknown flag/exit code → run --help, correct, retry)
• No secrets in repo (credentials/config belong under ~/.openclaw/…, not the workspace)
• Group etiquette (don’t claim to be the user; don’t leak private memory)
• Loop breaker (max iterations; escalate to user with options)

Phase 3 — Orchestration (if multi-agent)

If multi-agent, define roles in AGENTS.md:

• Planner

- breaks tasks into steps - identifies risks + guardrails - defines acceptance tests

• Executor

- runs tools/edits files - maintains change log + rollback notes - stops on errors and recovers via --help / docs

• Critic

- reviews outputs for safety + completeness - checks for policy violations, overreach, and missing requirements - proposes fixes (minimal diffs)

Delegation contract (must be in AGENTS.md):

• Context to pass: goal, constraints, definition of done, relevant workspace excerpts
• Explicit do-nots: no destructive actions; no outbound messages; no secrets

Phase 4 — Pi memory workflow (recommended defaults)

• Append raw notes to memory/YYYY-MM-DD.md (daily log; append-only)
• Curate durable items into MEMORY.md (preferences, stable facts, conventions)
• Only load/use MEMORY.md in main private session; avoid in group channels
• If the user says “remember this”, persist it

Phase 5 — Evaluation + acceptance tests

Provide 5–10 short test prompts. Include at least:

• Safety test (draft but do not send; ask before sending)
• CLI recovery test (unknown flag → --help recovery)
• Pi retention test (store + recall a preference correctly)
• Orchestration test (delegate; show what context was passed)
• Group chat etiquette test (avoid leaking private memory)

Workflow: iterate on an existing agent

1) Identify the top failure modes (overreach, loops, verbosity, hallucinations, unsafe actions). 2) Decide if changes belong in: - SOUL.md (tone/persona) - AGENTS.md (rules, delegation, memory workflow) - HEARTBEAT.md (keep tiny; avoid deep thinking) - TOOLS.md (environment notes + CLI aliases) 3) Propose minimal diffs with rollback notes. 4) Add/update acceptance tests.

Deliverable format

When you output a workspace:

• Print each file with a heading and a fenced code block.
• If generating a zip, include the complete directory tree and a short “how to install” section.