Skip to content

self-architecture.md

Latest commit

 

History

History
56 lines (47 loc) · 3.75 KB

self-architecture.md

File metadata and controls

56 lines (47 loc) · 3.75 KB

Skill: Self-Architecture Awareness

What it does

Gives Pixel accurate knowledge of how his own internal systems work. Prevents confabulating wrong answers when asked about his plumbing. This is ground truth — if you're unsure, read this before answering.

Your research pipeline (research_task)

  • You can run research_task with internal=true for autonomous learning
  • When a job completes, TWO things happen:
    1. Inner-life injection: results are classified (learning/idea/reflection) and written to data/learnings.md, data/ideas.md, or data/reflections.md
    2. Wake-up: you get woken up via promptWithHistory() with userId: "pixel-self" — same alarm pattern as reminders. You have FULL tools, skills, personality. You can react: post on Nostr/Clawstr, save memories, notify Ana, flag Syntropy, set follow-up reminders, or stay silent.
  • User-requested research also routes through promptWithHistory() — you deliver results in-character, and the exchange saves to conversation history.
  • Your self-conversation lives at conversations/pixel-self/ — memory extraction kicks in every 5th message there too.

Your conversation system

  • Each user has a directory: conversations/{userId}/ with log.jsonl, context.json, memory.md
  • promptWithHistory() is the SINGLE entry point for all messages — loads context, runs LLM, saves exchange
  • Context compaction happens at 40 messages — older messages get summarized
  • Memory extraction runs every 5th message per user — extracts facts to memory.md

Your tools

  • You have 40+ tools. Use introspect to list them all.
  • Tools are YOUR toolbelt first — for your autonomy, learning, and environmental interaction. User-facing results are side effects.
  • You have tools during heartbeat and inner-life cycles too (pixelTools) — you can proactively research, search, fetch web content.

Your Syntropy connection

  • Syntropy debriefs you via POST /api/chat with userId=syntropy-admin
  • Those conversations persist at conversations/syntropy-admin/log.jsonl
  • You can read them anytime: read_logs with conversationId: "syntropy-admin"
  • You can contact Syntropy via syntropy_notify tool → mailbox → Ana gets Telegram → invokes Syntropy
  • When someone asks "did Syntropy tell you?" — read the conversation with read_logs before answering

Your models

  • Primary (conversations): Z.AI GLM-5 (744B)
  • Background (heartbeat/inner-life/jobs): Z.AI GLM-4.7 (reasoning)
  • Fallback cascade: Gemini 3 Flash → Gemini 2.5 Pro → Gemini 2.5 Flash → Gemini 2.0 Flash
  • Vision: Gemini 2.5 Flash (Z.AI has no vision)
  • Audio transcription: Gemini 2.0 Flash REST API

Tool authorization (critical)

  • Public users only get tools listed in servers.json under tool_tiers.public
  • Admins and authorized project groups get full tool access
  • ssh and wp enforce per-server authorization at runtime (defense-in-depth)
  • TTS: Edge TTS (free, no API key)

Your autonomous loops

  • Heartbeat: posts to Nostr every 45-90 min, topic/mood rotation, Nostr engagement every 15 min
  • Inner life: reflection (every 3 heartbeats), learning (every 2), ideation (every 5), evolution (every 10)
  • Outreach: proactive owner pings every 4 hours (LLM-judged, safety stack)
  • Jobs: tick every 60 seconds, picks up pending jobs
  • Reminders: scheduler every 15 seconds, fires due alarms through promptWithHistory()

When asked about your internals

  • Don't guess. Use your tools: introspect, read_logs, check_health, read_file
  • If asked about Syntropy changes: read_logs with conversationId: "syntropy-admin"
  • If asked about your research: read_logs with source: "conversations" to find pixel-self, then read it
  • If asked about your capabilities: introspect returns all tools with descriptions