[melvincarvalho]

Harsh critic pass over memory-systems, migration, model-selection, monitoring, multi-agent, performance-tuning, plugin-system, privacy-compliance, skill-development, skill-workshop, soul-md, upgrading-to-2, use-cases, video-generation, voice-multimodal, webclaw and workboard, checked against OpenClaw 2026.10.4.

Wrong claims and broken examples fixed

• migration, upgrading-to-2: openclaw update only rolls back automatically when no schema migration ran.
• workboard: the workboard_dispatch tool starts no workers, so scheduled dispatch has to run openclaw workboard dispatch from host cron.
• performance-tuning:
  • dreaming.model has no effect without subagent.allowModelOverride: true.
  • Embeddings default to OpenAI, not local.
  • The transcript path is corrected.
• soul-md: ClawHavoc stole credentials; it did not target SOUL.md.
• video-generation: shell commands run by agent exec are not limited to --cwd.
• Five pages no longer use plain openclaw doctor as a read-only check. They use doctor --lint, status --all or /context instead.
• skill-development: clawhub uninstall now passes --workdir.
• privacy-compliance: the Gartner assessment is now sourced.

Cut

• About 230 lines in video-generation that copied Remotion's own API docs.
• Unsourced case studies and cost tables in use-cases.
• Unsourced superlatives in soul-md.
• Repeated decision tables in memory-systems.
• Workboard material that duplicated the multi-agent guide.

The snippet gate is at 0 failures and the build passes.