System Report — Phase 21.5 (Historical)

Historical snapshot of Arc OS after Evolutionary Intelligence integration. Note: This document reflects the system state at Phase 21.5. The current system is at Phase 40.18 with 62+ API endpoints, CRM dashboard, OAuth, Sage Worker, and more. See the User Guide and Product Overview for current documentation.

System Summary

Metric	Value
Core TypeScript (shared + bots + scripts)	3,298 lines
Shared modules	7
Registered skills (registry)	18
Library skills	7
Total skills	25
Eval rule types	6
Pilot eval rulesets	3 (9 rules total)
Agents	6
Child bots (production)	2
Git commits	36
Phases delivered	15 → 21.5
Custom backend code	0 lines

Architecture

                    ┌──────────────────────────────────┐
                    │       Master Bot (Port 19210)      │
                    │       @citadel_master_bot          │
                    │                                    │
                    │  /status /health /deploy /watchdog  │
                    │  /new_project /remove_project       │
                    │  Approve/Reject improvements        │
                    └──────────┬───────────┬─────────────┘
                               │           │
                  ┌────────────┘           └────────────┐
                  ▼                                     ▼
        ┌───────────────────┐             ┌───────────────────┐
        │  Child: citadel-v2 │             │  Child: pt         │
        │  Port 19211        │             │  Port 19212        │
        │                    │             │                    │
        │  Intelligence:     │             │  Intelligence:     │
        │  ├ Context Router  │             │  ├ Context Router  │
        │  ├ Learnings       │             │  ├ Learnings       │
        │  ├ Binary Evals    │             │  ├ Binary Evals    │
        │  └ Quality Track   │             │  └ Quality Track   │
        └───────────────────┘             └───────────────────┘
                  │                                     │
                  ▼                                     ▼
        ┌───────────────────┐             ┌───────────────────┐
        │  MCP Bridge        │             │                    │
        │  Port 19200        │             │  (no bridge yet)   │
        │  SSE + HTTP API    │             │                    │
        └────────┬──────────┘             └───────────────────┘
                 │
                 ▼
        ┌───────────────────┐
        │  Phaser.js Frontend│
        │  Port 5173 (dev)   │
        │  Visual AI Office  │
        └───────────────────┘

Module Map

shared/ (7 modules, 1,124 lines)

Module	Lines	Phase	Purpose
`logger.ts`	189	20.5	Structured JSONL logging: system/dialog/error channels, daily rotation
`vault.ts`	216	20.5	AES-256-GCM encrypted secret storage: initVault, getSecret, setSecret
`quality.ts`	192	21.1+21.5	Execution/feedback tracking, aggregation, underperformer detection
`ui_templates.ts`	158	21.0+21.5	Telegram keyboards: master menu, project cards, child buttons, improvement proposals
`evals.ts`	142	21.5	Binary eval engine: 6 rule types, checkOutput, formatWarnings
`context-router.ts`	117	21.5	Skill scoring: trigger/keyword matching, SKILLS_HINT generation
`learnings.ts`	110	21.5	Persistent corrections: load/add/format learnings from markdown

Bot runtimes (1,937 lines)

File	Lines	Purpose
`child-bot/bot.ts`	944	Telegram-Claude CLI proxy with intelligence layer integration
`master-bot/bot.ts`	993	Orchestrator: commands, project cards, watchdog, improvement approval

Scripts (237 lines)

File	Lines	Purpose
`scripts/nightly-improve.ts`	237	Karpathy Loop: metrics analysis → improvement proposals → CEO Telegram

Skills Inventory

Registry Skills (18)

Skill	Category	Phase	Evals
handoff-protocol	complex, skill_injection	15.2	-
code-review	complex	15.3	3 rules
system-audit	complex	15.3	3 rules
git-manager	complex, skill_injection	16	3 rules
project-onboarding	complex, skill_injection	16	-
web-scaffolding	complex	15	-
code-review-protocol	complex	15	-
automated-testing	complex	15	-
deployment-flow	complex	15	-
knowledge-index	complex	15	-
document-analysis	complex	15	-
citadel-wrapup	cli_command	19.4	-
citadel-recall	cli_command	19.4	-
citadel-status	cli_command	19.5	-
citadel-task	cli_command	19.5	-
citadel-deploy	cli_command	19.5	-
figma-audit	complex, skill_injection	21	-
figma-apply-system	complex, skill_injection	21	-
figma-implement	complex, skill_injection	21	-

Library Skills (7)

Skill	Domain
docker-ops	Container orchestration
figma-design	Design-to-code pipeline
github-actions	CI/CD workflows
odoo-expert	Odoo ERP development
odoo-owl-expert	Odoo Owl framework
plasthetica-style-guide	Project-specific style guide
postgres-pro	PostgreSQL optimization

Intelligence Layer (Phase 21.5)

Message Processing Pipeline

Message In
    │
    ├─ Context Router    scores 18 registry skills → top-5 SKILLS_HINT
    ├─ Learnings Loader  reads learnings.md → LEARNINGS block
    ├─ BTW Queue         prepends extra context from BTW mode
    │
    ▼
buildGsdPrompt() assembles: LEARNINGS → BTW → CONTEXT → SKILLS_HINT → TASK
    │
    ▼
claude -p <prompt> --max-turns 20
    │
    ▼
    ├─ Binary Evals      checks response against .evals.json rules
    ├─ Quality Tracker    logs execution event (skills, success, duration)
    │
    ▼
Telegram response + inline buttons + eval footnotes

Feedback Loop

Fix It button → addLearning(source: "fixit") → learnings.md → next prompt
Thumbs-down   → addLearning(source: "negative") + logFeedback → metrics
Thumbs-up     → logFeedback(positive: true) → metrics

Nightly Cycle

03:00 UTC → nightly-improve.ts
  → Read quality-metrics.json per child
  → findUnderperformingSkills(threshold: 80%)
  → Generate proposals
  → Send to CEO via Master Bot Telegram
  → CEO: [Approve] → backup skill.md → mark approved
  → CEO: [Reject] → mark rejected

Infrastructure

VPS (Contabo)

Service	Port	Health Endpoint
Master Bot	19210	`/api/master/health`
Child: citadel-v2	19211	`/api/child/health`
Child: pt	19212	`/api/child/health`
MCP Bridge	19200	`/api/health`
Frontend	5173 (dev) / Docker	N/A

Cron Jobs

Schedule	Job
Every 15m	Git sync (memory bridge)
Every 30m	Morty SRE health check
Every 2h	CFO metrics aggregation
Daily	CMO status report
Daily	DB backup
Daily 03:00	Nightly improvement loop

Logging

Format: Structured JSONL
Channels: system, dialog, error
Location: /var/log/citadel/<bot-name>/
Rotation: daily (logrotate)
Access: jq . /var/log/citadel/citadel-v2/dialog-$(date +%F).log

Security

Bot tokens: AES-256-GCM vault (config/vault.json, gitignored)
Encryption key: config/vault-key (auto-generated, chmod 600)
CEO-only access: all bots check CEO_CHAT_ID on every message
No credentials in git (.gitignore + .claudeignore)

Phase History

Phase	What Was Built
15	Skills system, agents, blueprints
16	Git manager, project onboarding
18	Parent umbrella issue
19	CLI commands, library, wrapup/recall
20	Federated bots, Master + Child architecture
20.5	Structured logging, watchdog, vault, onboarding
21.0	UI templates, subprocess control, quality tracking
21.1	Quality metrics aggregation, feedback buttons
21.5	Evolutionary Intelligence: evals, context router, learnings, nightly loop
22-26	CRM Backend API, Workspace, Workers, Specs
28-30	Knowledge auto-ingest, onboarding wizard
32-34	Wiki, Skills CRUD, Issues, ARC CLI
36	Cloud PM, Neural Skill Generator, NotebookLM Bridge
37-38	OAuth (Google/GitHub), email/password auth, ARC CLI binary
39	Mobile responsive CRM
40.10-40.18	Chat history, Sage Worker, benchmarks, marketplace, Lucide icons, Telegram vault, worker bots

For current system details, see Product Overview.