Hermes venv at /opt/hermes/.venv/ has no pip binary. Install packages with: uv pip install --python /opt/hermes/.venv/bin/python3 <package>. Verify with: /opt/hermes/.venv/bin/python3 -c "import <package>".
§
GrowReach: Rahul’s SaaS, AI LinkedIn engagement automation. Pricing 49/$69. Two personas: Personal Brand Builders + LinkedIn Lead Generators. linqin.ai = competitor.
§
Kanban-first: every substantive task MUST be on the board before work starts. Rahul flagged twice. Even quick research = create task first. kanban_complete when done.
§
Rahul strongly prefers TABLE format over bullet lists for any multi-item presentation. Not optional.
§
Tools: PostHog (project 136693 EU) for ALL analytics. 5 marketing profiles active: reddit-marketer, linkedin-marketer, cro, competition-analyst, brand-strategist. SEO and CRO separate.
§
Leverage existing work FIRST before new scraping — check board + GitHub. Rahul’s terminal (Telegram-SSH) breaks multi-line pasted commands (inserts ─, converts -- to —, breaks backticks): write .sh scripts to mounted volume, tell him bash <host-path>. NEVER use GitHub raw URLs (private repo=404). Host volume /docker/hermes-agent-fzqh/data → /opt/data.
§
AdsPower swarm: 13 Reddit profiles (11 active, 1 shadowbanned, 1 revived). Webshare+Decodo proxies. 4 layers: Local API (50325), ads CLI (profile mgmt only), MCP Server (Playwright, browser automation), RPA (GUI-only, NO API). Windows VPS (EC2) runs AdsPower 24/7 + Hermes desktop (adspower-operator) + Telegram gateway. This VPS=planner, Windows=operator. GitHub: rahul-10xers/hermes/adspower-dashboard/. Rahul on Windows uses CMD (setx for env vars). ads status=false negative in GUI mode, use curl localhost:50325/status.
§
Composio Reddit: 2/5 tokens expired (401), 1 still valid (429). RSS works but no scores. Reddit JSON API 403 on VPS IP. Incident report on GitHub.
§
CRITICAL: NO Composio Reddit API until July 25. Rules: 1 req/5 min, never batch accounts, never test multiple accounts in succession, always check rate limit first. Max 10 calls/account/day. Never use proxy_execute for bulk. Push all learnings to GitHub for cross-instance sync.