stage-launch

原文涵盖的主题

• Phase 2 — Stage → Launch
• Workflow
• Hard rules (API-key safety)
• Forcing-question library (recommend + cite)
• Tools

Turn the build sheet into runnable artifacts, then let the founder launch with their own key. No script here touches the network or the key — the user runs launch.sh. 1. Generate payloads. Agent toolset → alwaysallow; every MCP toolset → alwaysask (baked into the agent payload's permissionpolicies). 2. Write the launch script. launch.sh creates environment → agent → session → kickoff in order, chaining IDs, and resumes on re-run (each step skips if its .id file exists). It reads $ANTHROPICAPIKEY at runtime. 3. Validate before launch. FAIL blocks — especially a keyleak finding. Fix and re-run. 4. Minimal key step (never in chat). Check the shell first: Point the founder to platform.claude.com → API keys. Never print the key to chat, never write it to a file. 5. Launch + watch the first poll. Mark checkpoints with Console deep links. Then goalstate.py set --phase grade-iterate and advance. The key never enters chat, a file, a payload, or a log. launch.sh reads it from the environment; payloadvalidator.py scans for sk-ant-… leaks and FAILs. Sequential launch. environment → agent → session → kickoff. Watch the first poll foreground before declaring success. Resumable. Re-running launch

这里只展示节选，请以完整原文和当前要求为准。