Environment

• Windows
• OpenClaw 2026.5.7
• OpenClaw dashboard endpoint: [local loopback OpenClaw dashboard endpoint]
• LM Studio provider (OpenAI-compatible)
• LM Studio endpoint: [local LM Studio OpenAI-compatible endpoint]
• Model google/gemma-3-4b
• API adapter openai-responses

Goal

Run OpenClaw in proposal-only mode with no file/system/tool/channel/automation authority.

What Works

• Dashboard works after LM Studio context increase.
• Proposal-only prompt passes.
• tools.web.search.enabled was disabled.
• LM Studio provider works using openai-responses.
• Thinking default is low.

Prior Blocker Fixed

• n_keep around 29k exceeded n_ctx 4096.
• Reloading LM Studio with larger context fixed dashboard response.

Current Blocker

• Exact config key paths for disabling remaining plugin/tool/channel/owner-elevated surfaces are unclear.
• Codex safely stopped instead of guessing config keys.

Surfaces Needing Authoritative Key Mapping

• browser
• device-pair
• file-transfer
• memory-core
• phone-control
• talk-voice
• channels
• cron/automation
• MCP
• hooks
• apps
• owner-only commands
• exec approvals
• dangerous action approvals

Safety Requirements

• no guessing config keys
• backup-first edits only
• no read-only mode until surfaces are locked down
• no tools/hooks/MCP/apps/channels/automation
• no file/system authority
• Codex validates any execution

Maintainer Questions

1. What exact config keys disable each listed surface?
2. Is there a native proposal-only/no-tools mode?
3. Is there a native read-only mode?
4. How do we disable owner-elevated commands and approvals?
5. How do we disable or isolate browser/file/phone/voice/plugin sidecars?
6. How do we prevent doctor from auto-enabling plugins or shrinking config?
7. How do we validate no tools/channels/automation are active?
8. Which surfaces are mandatory for dashboard-only chat?
9. Which surfaces can be safely disabled without breaking dashboard chat?

Redacted Evidence Summary

• No tokens, API keys, bearer values, cookies, or session IDs included.
• No full raw local config included.
• OpenClaw dashboard and proposal-only checks pass under current provider state.
• Remaining blocker is authoritative key mapping for capability lockdown fields.

What We Need Before Proceeding

• authoritative config key map
• second lockdown execute plan
• second lockdown execution
• dashboard retest
• proposal-only retest
• read-only manual test plan after lockdown