2026-03-10

Search/fetch setup

Enabled native web_fetch in OpenClaw config.
Confirmed local SearXNG at http://192.168.153.113:18803 is working.
Confirmed local Brave MCP endpoint at http://192.168.153.113:18802/mcp is reachable via mcporter as tool brave_web_search.
Added local skill skills/brave-mcp-search for direct Brave MCP-backed search.
Added local skill skills/local-meta-search to route search as:
- SearXNG first
- Brave MCP fallback on failure or weak/empty results
Search routing preference documented in TOOLS.md.

I missed that Telegram was already set up and incorrectly asked setup-style follow-up questions instead of checking current state first.
New rule: before suggesting setup/re-setup, first inspect current config, memory, and recent evidence; if already configured, treat it as ops/validation/debugging.

Source: workspace changes + gateway config patch on 2026-03-10.

Added Brave Search API key to OpenClaw config under tools.web.search.apiKey and set provider to brave.
Native web_search now works.
Important constraint: Brave free plan hit rate_limit: 1; searches should be serialized instead of sent in parallel bursts.
Rate-limit note was added to TOOLS.md and committed.

Telegram channel was already configured and enabled; validated instead of re-setting it up.
Confirmed Telegram outbound send works to allowed user id 8367012007.
Confirmed Telegram inbound DM works by locating exact inbound token zap-telegram-inbound-verify-2109 in session log /home/openclaw/.openclaw/agents/main/sessions/7d9bf35f-81ad-45ce-b2ca-369822459acc.jsonl.
Important routing behavior: with session.dmScope = "per-channel-peer", Telegram DMs land in a separate main session from the TUI/webchat session.
Added Telegram/TUI session-routing note to TOOLS.md and committed.

Added a new AGENTS.md startup step: inspect trusted inbound metadata first and internally classify the session with a short label like direct-webchat, direct-telegram, group-telegram, etc.
Purpose: improve privacy, memory loading, and reply behavior by making session type explicit at session start.
Committed as docs(agents): add session classification step.

Chose daily-brief as the next capability to build out, starting with a calendar-first approach.
Checked workspace, memory, and local docs: no existing calendar integration is configured yet.
Candidate calendar skills found via ClawHub include gcalcli-calendar, gog-calendar, caldav-calendar, icloud-caldav, and macos-calendar.
Next decision needed from Will: which calendar backend they actually use (Google Calendar vs CalDAV/iCloud/other).