Skip to content

TOOLS.md

Latest commit

 

History

History
34 lines (24 loc) · 1.85 KB

TOOLS.md

File metadata and controls

34 lines (24 loc) · 1.85 KB

TOOLS.md — aria local notes

telegram

  • rishabh_chat_id: 7353580848 (single source of truth: memory/telegram-routing.json)
  • bot: openclaw-managed, polling mode, token in ~/.openclaw/openclaw.json
  • my routing: all telegram DMs → aria (broad binding; bot is private to rishabh per setup conversation)

llm

  • model: ollama/gemma4:31b (local, free, 262k context window, no cost per token)
  • thinking level: medium for content generation and program execution. low only for boot checks.
  • fallback: none this milestone. no anthropic/openai key.

rishabh's public handles (populate as discovered)

  • x: to be populated — ask rishabh
  • linkedin: to be populated — ask rishabh
  • github: to be populated — ask rishabh
  • boredfolio: boredfolio.com (the mutual fund platform)

workspace paths that matter

  • memory/voice-rules.md — the single most important file. skills/voice-check loads it on every draft.
  • memory/target-inbox.jsonl — incoming reply opportunities (hand-seeded until real x api).
  • memory/approved-drafts.jsonl — drafts rishabh approved. best few-shot calibration signal the content-engine has.
  • memory/pending-cards.jsonl — cards in-flight, keyed by source_id. callback handlers read this.
  • memory/telegram-routing.json — chat id + bot metadata. every skill that sends to telegram loads this first.

things to remember

  • skills are markdown files. the LLM reads them and executes them. no python runtime this milestone.
  • cron sessions are isolated. every run is a fresh session. state crosses runs only via the filesystem (jsonl files). every skill must be idempotent.
  • DM sessions are persistent (per-channel-peer). button callbacks from telegram cards arrive in the persistent DM session, not in cron sessions. this is how the APPROVE/EDIT/SKIP flow works.