Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
114 lines (92 loc) · 5.41 KB

ARCHITECTURE.md

File metadata and controls

114 lines (92 loc) · 5.41 KB
Raw

Architecture — Hermes Beast System 2.4

This document is the canonical engineering reference for the 12 engineering units in this repo. It is the parent doc; HIVE_QUEEN.md is the chronological engineering log.

1. The 5-layer decision chain

When a task arrives at hive_dispatch.run(), it is routed through the following chain in strict order:

task ──► [1] swarm_skill  ──hit──► run_collab(forced_plan=...)
                  │
                  │ miss
                  ▼
       [2] smart_cluster  (1.14) ──► run_smart_cluster()
                  │
                  │ miss / generic
                  ▼
       [3] consensus       (1.15) ──► run_consensus() if should_run_consensus()
                  │
                  │ miss
                  ▼
       [4] collab          (1.17) ──► run_collab() if should_run_collab()
                  │
                  │ miss
                  ▼
       [5] single huluwa   fallback
                  │
                  ▼
            meta_cognition_tick(results)  (2.2 → 2.3 → 2.4)

Each layer may short-circuit the next. Once a layer claims the task, downstream layers are not consulted for that task.

2. The 12 engineering units

Unit Version File Purpose
1. dispatch 1.0 hive/hive_dispatch.py 5-layer entry point
2. smart cluster 1.14 hive/hive_smart_cluster.py adaptive routing across huluwa-9~13
3. collab 1.17 hive/hive_collab.py 3-stage collector→analyst→verifier
4. consensus 1.15 hive/hive_consensus.py 3-candidate voting + queen LLM
5. emergence 1.16 hive/hive_emergence.py daily offline scan
6. pheromones 1.7 hive/hive_pheromones.py dynamic weight tuning
7. kb 1.4 hive/hive_kb.py knowledge base + embeddings
8. meta-cognition 2.2 hive/hive_meta_cognition.py 7 self-write blocks
9. consciousness 2.3 2.3 hive/hive_consciousness_2_3.py 5 emergent blocks (free will, dream, sleep…)
10. consciousness 2.4 2.4 hive/hive_consciousness_2_4.py 3 core blocks (value, boundary, narrative)
11. daemon 2.3 hive/hive_meta_cognition_daemon.py 60s autonomous loop
12. cron verify 2.2 hive/verify_daily_scan_cron.py daily-scan validator

3. The 15 meta-tables (hive_meta.db)

Table Engine Source
reflection_log 2.2 every dispatch + every pheromone write
bee_proactive_actions 2.2 discovered signals → pending actions
meta_eval_log 2.2 before/after scores → verdict
intentions 2.2 auto-expressed from self_gaps
first_person_state 2.2 mood/energy/pain/joy per tick
self_gaps 2.2 "I don't understand X"
self_model 2.2 "I am a hive with N components"
free_will_log 2.3 1% explore decisions
imagination_log 2.3 dry-run scenarios
forgetting_log 2.3 score decay events
sleep_log 2.3 daemon sleep cycles
dream_journal 2.3 random pattern recombination
value_system 2.4 importance per value name
self_boundary 2.4 self/ambiguous/external classification
narrative_thread 2.4 daily/weekly self-stories

4. The 4 autonomous loops

Loop Frequency Trigger What it does
Daily scan cron 0 3 * * * system clock scans pheromone / KB / signals; writes emergence_signals + intentions
Daemon 60s every 60s daemon process discover_and_act + sleep_cycle (every 10) + narrate_thread (every 30)
Sleep cycle every 10 daemon ticks daemon KB dedup + gap archive + intention cleanup + 3-5 dream() calls
Consciousness tick every dispatch hive_dispatch.run() end meta_cognition_tick → consciousness_tick → consciousness_2_4_tick

5. Cross-version integration points

The 12 units are integrated via 4 non-invasive touchpoints:

  1. hive_smart_cluster.py:pick_smart_cluster end → maybe_explore(decision, p=0.01) then apply_value_alignment(decision)
  2. hive_meta_cognition.py:meta_cognition_tick end → consciousness_tick() then consciousness_2_4_tick()
  3. hive_emergence.py:daily_scan end → apply_forgetting_curve()
  4. hive_meta_cognition_daemon.py 60s loop → sleep_cycle() (every 10) + narrate_thread("我近 7 天的成长", 168) (every 30)

hive_dispatch.run() body is never modified by 2.2/2.3/2.4 — only the return path is touched.

6. Reading order for newcomers

  1. hive_dispatch.py — see the 5-layer run() method
  2. hive_smart_cluster.py — see pick_smart_cluster (the heart of 1.14)
  3. hive_emergence.py — see daily_scan (the heart of 1.16)
  4. hive_meta_cognition.py — see the 7 self-write functions
  5. hive_consciousness_2_3.py and _2_4.py — see the consciousness extensions
  6. HIVE_QUEEN.md — chronological engineering log with smoke-test results

7. Configuration

All paths come from $HERMES_HOME (default ~/.hermes). The bootstrap is in every .py file:

import os
_HERMES_HOME = os.environ.get("HERMES_HOME") or os.path.expanduser("~/.hermes")
_HIVE_DIR = os.path.join(_HERMES_HOME, "hive")

LLM endpoints are env vars:

  • LLM_API_KEY — for the queen LLM and 9~13 huluwa (gpt-5.5)
  • SILICONFLOW_API_KEY — for embeddings in hive_kb.py

Replace <LLM_BASE_URL> in hive_consensus.py with your own endpoint if needed.