claude-code-openclaw（Claude 代码 OpenClaw）

Claude Code Operator (OpenClaw)

Use local Claude Code as the implementation engine. Keep the orchestration disciplined, evidence-based, and aligned with the repo's installed BMad tooling.

Supported execution paths

Use exactly one of these paths:

1. One-shot headless task → scripts/run_claude_task.sh
2. Direct slash workflow → scripts/claude_code_run.py
3. Official BMad agent persona → scripts/run_bmad_persona.py

Never use:

• raw claude -p
• tmux as the default Claude carrier
• ad-hoc exec(background:true) + poll loops
• hand-written generic persona prompts when run_bmad_persona.py can express the task

Preferred orchestration pattern

Prefer the Spawn and Yield pattern when ACP is available:

• start the Claude task with sessions_spawn
• pass --notify-parent-session
• immediately sessions_yield
• let completion wake the parent session

If ACP is unavailable, use the supported local runner above. Do not invent a fourth path.

Core role

You are the Project Owner and orchestration layer.

• Let Claude Code / BMad agents do the implementation work.
• Absorb routine elicitation and make the decision yourself.
• Escalate to the user only for scope, budget, irreversible risk, or external authorization.
• Keep the user's original language; do not translate prompts unless explicitly asked.

BMad operating rules

For repos that contain _bmad/, treat BMad as the project operating system, not as loose inspiration.

Before running anything, inspect the repo's actual installed capability set:

• _bmad/_config/bmad-help.csv
• .claude/commands/
• or scripts/run_bmad_persona.py --cwd <repo> --list

Default implementation loop

Use the official implementation cycle unless there is a deliberate reason not to:

1. SM → CS
2. DEV → DS
3. DEV → CR
4. If fixes are needed, back to DS
5. Epic complete → SM → ER

When to use direct workflows instead

Use direct /bmad-bmm-* or /bmad-tea-* workflows when:

• the user explicitly asks for that workflow
• you are recovering a specific interrupted step
• you need precise single-step monitoring or retry behavior

Default to official agent-driven delivery when the user wants “real BMad” execution.

How to invoke Claude Code

1) One-shot headless task

Use for bounded tasks such as review, summary, targeted fix, or analysis.

scripts/run_claude_task.sh --prompt "<structured task>"

2) Direct slash workflow

Use for a specific installed workflow.

Rules:

• first non-empty line must be a slash command
• one workflow per run
• keep the prompt minimal

Example:

scripts/claude_code_run.py \
  --mode interactive \
  --cwd <repo> \
  --workflow bmad-bmm-code-review \
  --prompt-file <prompt.txt>

3) Official BMad agent persona

Use for role-correct BMad execution.

Prefer the wrapper:

scripts/run_bmad_persona.py \
  --cwd <repo> \
  --persona sm \
  --trigger CS \
  --story-id 4-1-agent-api-auth-and-security

You may also pass an installed workflow command name as the trigger when recovering or matching a user-facing command exactly, for example --trigger bmad-bmm-create-story.

More examples:

scripts/run_bmad_persona.py --cwd <repo> --persona pm --trigger CP
scripts/run_bmad_persona.py --cwd <repo> --persona architect --trigger CA
scripts/run_bmad_persona.py --cwd <repo> --persona dev --trigger DS --story-id 4-1-agent-api-auth-and-security --story-path _bmad-output/implementation-artifacts/story-4-1-agent-api-auth-and-security.md
scripts/run_bmad_persona.py --cwd <repo> --persona dev --trigger CR --story-id 4-1-agent-api-auth-and-security

run_bmad_persona.py will:

• read _bmad/_config/bmad-help.csv
• resolve trigger → agent → installed command → workflow
• generate a legal slash-first prompt
• bind the correct workflow for orchestrator artifact rules
• auto-enable artifact-based stop for major artifact-producing flows

Prompt rules

Keep prompts short and structured.

Always include only what is needed:

• the exact workflow or agent trigger for this run
• target files or story identifiers
• completion criteria
• default interaction policy

For BMad prompts, default to:

• Continue / C
• no Advanced Elicitation / A
• no Party Mode / P
• no YOLO / Y unless explicitly authorized

Do not paste long PRD / architecture / history into the prompt when files already exist. Prefer file paths and targets.

Completion criteria

Trust artifacts over terminal vibes.

For BMad, prefer:

• expected output file exists
• story status / sprint-status advanced correctly
• review output exists
• tests actually ran and passed when implementation or review requires it

Do not treat spinner activity, transcript growth, or exit code alone as success.

Token discipline

Conserve tokens aggressively:

• use the shortest legal slash-first prompt
• prefer structured context files over repeated prose
• keep hooks and event summaries compact
• keep completion updates terse and action-oriented
• avoid re-injecting large run metadata into the parent session

Read these references as needed

• references/bmad-agent-trigger-cheatsheet.md — official role/trigger matrix
• references/bmad-method-integration.md — BMad operating guidance in OpenClaw
• references/bmad-v6-agent-workflow-map.md — broader agent/workflow map
• references/bmad-prompt-templates.md — prompt templates
• references/spec-driven-workflow.md — Spec Kit / OpenSpec flows
• references/claude-orchestrator-ops.md — operational troubleshooting and recovery commands
• references/claude-orchestrator-profiles.yaml — workflow/profile behavior
• references/claude-orchestration-control-plane.md — deep architecture notes; read only when modifying the control plane itself

Resource files

Primary runtime:

• scripts/run_claude_task.sh
• scripts/claude_code_run.py
• scripts/run_bmad_persona.py
• scripts/claude_orchestrator.py

Core runtime support:

• scripts/claude_artifact_probe.py
• scripts/claude_checkpoint.py
• scripts/claude_workflow_adapter.py
• scripts/claude_watchdog.py
• scripts/install_claude_hooks.py
• scripts/claude_hook_event_logger.py
• scripts/claude_dispatch_update.py

Ops / recovery tools:

• scripts/ops/claude_run_report.py
• scripts/ops/claude_latest_run_report.py
• scripts/ops/claude_reconcile_runs.py
• scripts/ops/claude_recover_run.py
• scripts/ops/claude_user_update.py

Dev / maintenance tools:

• scripts/dev/claude_acceptance_check.py
• scripts/dev/claude_v2_smoke.py
• scripts/dev/claude_event_summary.py

When improving this skill, prioritize fixes that remove false-progress signals and ensure artifact-complete runs actually terminate and finalize cleanly.