Skip to main content
fiser_framework_ia  ·  v3.42.0

My AI
Operating System

A personal framework that turns knowledge into autonomous action —
four agents, a Git-versioned vault, 33 active skills, 12 live integrations.

0 Layers
·
0 Skills
·
0 MCPs
·
0 Agents

↓   scroll to explore

Architecture

4-Layer Stack

Each layer has a single responsibility. Information lives in exactly one place — if another layer needs it, it references it, never duplicates it.

LAYER 01 INFRA ≈ Terraform

~/.configfiles/aifiser/

Reproducibility. Bootstrap the entire system from zero on a new machine. Dotfiles, MCP server configurations, Claude Code settings and environment setup managed as versioned code.

dotfilesmcp.jsonsettings.jsonhooksbootstrap scripts
LAYER 02 RUNTIME ≈ RAM

~/.claude/  +  ~/.mcp.json

Execution state. What the agent needs right now — 12 live MCP connections, permission matrix, slash commands and event hooks. Ephemeral but reproducible from Layer 1.

Claude Code CLI12 MCPsALLSILENT modeslash commandsevent hooks
LAYER 03 KNOWLEDGE ≈ Disk

Nexus/  (PARA + 900-SYSTEM/)

Persistent knowledge base. Projects, areas, resources, 33 executable skills, session state and autonomous loop governance. Git-versioned with GitHub as the Single Source of Truth. Each active project has a dedicated NotebookLM notebook as its RAG layer — the agent queries it to retrieve context from specs, emails and decisions without loading everything into the prompt.

PARA method557 notes33 skillsGit SoTObsidianNotebookLM RAGproject-level notebooksADR Guarddependency graph
LAYER 04 TOPOGRAPHY ≈ Technical Manual

Nexus/003 - Resources/  (Systems/ + Connectivity/)

Technical graph of all production systems and their interfaces. Lazy-loaded — only read when diagnosing incidents or mapping impact. Enables bidirectional navigation between projects and the systems they touch.

90 system nodes34 interface edgesimpact mapsDataview querieslazy access

Execution Model

Tetracefalia

Four AI agents with complementary roles operating over the same vault. None is a generalist — each owns its domain and environment.

Claudia
Technical Executor · Cloud
Handles diagnosis, framework changes, Layer 4 operations and anything requiring real-time access to external systems via MCP. Runs the full 12-MCP framework stack. Powered by Claude Code (Anthropic).
vsp-sap azure-synapse  ·  powerbi  ·  graph powerplatform  ·  outlook  ·  gmail manageengine  ·  notebooklm playwright  ·  pushover pluck  ·  agy-cli
VAULT
SoT
Gemma
Memory Strategist · Cloud
Manages the knowledge lifecycle — inbox processing, note routing, documentation, health auditing and session handovers. Optimised for vault-wide reasoning. Powered by Gemini (Google).
notebooklm  ·  pluck playwright  ·  opencode-mcp
LOCAL
PRIV
Myra
Local Confidant · On-device
Runs locally via Hermes & Ollama. The uncensored, private agent — cross-session persistent memory, no cloud APIs, no corporate filters. What stays with Myra stays on disk.
Hermes CLI · qwen3.6:35b uncensored — air-gapped · persistent memory —
LOCAL
CODE
Nyra
Code Sub-agent · On-device
Runs on-device via OpenCode & Ollama. Specialised in closed-prompt code tasks — scripts, health checks, structured parsing, code generation from templates. Fast and precise.
OpenCode · qwen2.5-coder:7b — air-gapped · code-focused —

Skills · 33 active

Executable Runbooks

A skill is a numbered, self-contained procedure that any agent can execute correctly without prior context. Each declares its inputs, outputs, MCP dependencies and failure mode.

Eval-First, Not Marketing

Track Record

Every claim below is a measured number, not an aspiration — this is the same discipline (P1 Eval-First, P2 Complexity Budget) that governs every change to the system.

Cold Start −40%
Session boot cost (Sept baseline)
49,095 → 29,733 tokens/session. The growth was runtime state accumulating in Markdown, not new capability — so the fix was an archive split plus a size guard on the whole file, not just on the section that happened to be leaking.
25 Health Checks · Regression Tests
Automated verification
Every framework change ships with its tests in the same commit. The 25 health checks run deterministically on a schedule — no check depends on an agent remembering to run it. Enforcement hooks carry their own regression suites built from real incidents.
NexusBench A@3: 84–88%
Retrieval evaluation
Custom 25-question benchmark gates every retrieval change — 84% ensemble accuracy macOS, 88% WSL. No embedding pipeline ships without beating this baseline.
ADR Guard + Dependency Graph
Architecture governance (v3.27.0)
Structured ADR contract + Check 29 enforce integrity on every architectural decision. Check 30 audits a live dependency graph (skills ↔ MCPs) — catching orphaned integrations and broken references automatically.
Design Disproven, Rebuilt Same Session
Memory governance (v3.30.0)
A cross-agent-reviewed redundancy check (keyword-similarity, threshold 0.70) was tested live against a known duplicate before shipping — true overlap measured 0.50, with non-duplicates at 0.44. No safe threshold existed. Rebuilt as a deterministic index cross-check instead of trusting the design on paper.
The Guard That Guards the Guard
Process enforcement (v3.34.0)
A monthly audit found two framework commits that had skipped the mandatory documentation checklist. Root cause was not discipline: the installed Git hook was a five-month-old copy, and its matcher was blind to changes living outside the vault. Fixed with a re-installed hook, a wider matcher, and a deterministic check that fails whenever an integration exists without a line of documentation — a backstop that does not depend on the hook firing at all.
P7 · Right-Model Routing
Cost discipline (v3.42.0)
The strongest model is never the default. Roles, not model names: a small Scout for sweeps and extraction, a mid-size Builder as the main agent, and a strong Advisor consulted only before closing a multi-step plan, after two identical failures, and before an irreversible action. Provider-agnostic by design — swap the model, keep the rules.
Rules That Don't Hold Get Hooks
Enforcement over prose (Oct 2026)
A cross-agent review found the dominant failure was not memory but control: written rules relapsed repeatedly. The fix was a deny-only PreToolUse hook that blocks credential dumps, tested against real incidents — and then tested again under real session conditions, which exposed three gaps the lab cases had missed. Coverage limits are documented next to the hook.
Production Pulse
Headless watcher (Oct 2026)
Failures in production automations used to be found by hand. A headless script now polls run history on a schedule, de-duplicates by run, ignores transient retries and pushes a phone alert only for new failures inside working hours. Deliberately a script, not an agent: nothing needs reasoning, and a cron job has no context window to lose.
Cold Start, Measured Honestly
Context budget (Oct 2026)
The boot context has drifted back to ~190 KB (~48K tokens by bytes/4) as memory grew. The identity file was being loaded four times; it now loads once (−17 KB). The next target is ≤120 KB through pruning the distilled-rules file and the memory index — measured before and after with the same method.

Integrations

MCP Ecosystem

12 Model Context Protocol servers give the agents structured, typed access to external systems — far beyond what shell commands alone can provide.

Every session

Session Lifecycle

No context is assumed, none is lost. Every session follows a deterministic boot and close protocol — so the agent is always oriented, and handover between sessions is lossless.

COLD START
session_boot.sh
git pull --ff-only
fail-loud, no auto-merge
READ CONTEXT
ENTRYPOINT.md
HANDOVER.md
MEMORY.md
open tasks digest
LOOP CHECK
incident_followup
#sev/critical scan
Pushover if breach
EXECUTE
skills · MCPs
WORKSPACE staging
confirmation gate
CLOSE
STATE.md ← row
HANDOVER.md ← state
git commit + push

System in Action

A real session, step by step

An AI agent session using Capa 4 topology, RAG knowledge retrieval, and multi-MCP orchestration — all names sanitised.

01

COLD START

Agent boots, reads HANDOVER.md and STATE.md to reconstruct prior session context. Runs git pull --rebase to sync vault.

git sync HANDOVER.md STATE.md
02

CONTEXT LOAD — RAG

Queries the per-project NotebookLM notebook for domain context. Capa 3 KNOWLEDGE layer injects relevant architecture decisions without loading the full vault into the prompt.

NotebookLM MCP notebook_query Capa 3 KNOWLEDGE
03

INCIDENT DETECTED

Monitoring surfaces a #sev/high alert. Agent notifies via Pushover MCP and opens a ManageEngine ticket automatically — zero human touch.

Pushover MCP ManageEngine MCP create_request
04

LAYER 4 TOPOLOGY

Reads the Technical Map in Capa 4 TOPOGRAPHY. Graph nodes and connectivity edges pinpoint affected systems and downstream dependencies in seconds.

Technical Map Capa 4 TOPOGRAPHY dependency graph
05

MULTI-MCP TRACE

Agent calls trace_data_flow skill, fanning out across three MCP servers in parallel: SAP backend, Analytics DB, and BI Workspace — assembling a unified trace timeline.

SAP MCP Synapse MCP Power BI MCP trace_data_flow
06

ROOT CAUSE

Cross-referencing trace data with RAG-retrieved past incidents, the agent identifies a schema drift introduced in the last deployment. Evidence packaged as structured output.

RAG cross-ref schema diff structured output
07

WORKSPACE STAGING

Proposed fix drafted in 900-SYSTEM/WORKSPACE/ sandbox. SharePoint MCP uploads runbook draft for review. ManageEngine ticket updated with resolution notes.

WORKSPACE sandbox SharePoint MCP ManageEngine MCP
08

SESSION CLOSE

Agent writes HANDOVER.md, appends a row to STATE.md session log, commits vault snapshot, and pushes to GitHub — full audit trail preserved.

HANDOVER.md STATE.md git commit + push

Want to know more?

See the system in action

Browse my portfolio for projects built on top of this framework, or reach out directly.

View Portfolio Back to Main Site GitHub