Case study 04 · Gabi OS
Gabi OS
A personal AI operating system for reliable work across assistants.
AI assistants are individually powerful, but real work breaks down between sessions: context fragments, status goes stale, the human becomes the transport layer, and “done” becomes a claim instead of an evidenced state. Gabi OS is an internal product I built and continue to evolve to make that workflow explicit, bounded and verifiable.
AI systems · Product engineering · Internal product
The problem
Where AI-assisted work breaks down
Different assistants do not naturally share dependable state. Each new session starts by rediscovering what the last one knew, so long prompts and repeated explanations use up context and attention before any useful work begins.
When information has to move between tools, the person in the middle becomes the orchestration layer: copying instructions in, pasting results out, and remembering what actually happened.
Agents also need edges. Without explicit scope, permissions and stop conditions, a small task can turn into open-ended discovery or unintended change. And a model saying a task is complete is not evidence that it is. Missing or conflicting evidence has to stay visible rather than quietly becoming “done”.
The system
How a piece of work moves through it
- Human and chat control planeDecides what matters and approves consequential actions.
- Context compilerAssembles the smallest sufficient context from canonical state.
- Bounded Task EnvelopeFixes scope, exclusions, acceptance criteria, required evidence, limits and permissions.
- Approved workerCarries out only the bounded task, on an approved host.
- Execution receiptReturns structured evidence of what was actually done.
- ReconciliationChecks the receipt against canonical state and the acceptance criteria.
- Canonical stateRecords the accepted outcome, or keeps the gap visible. The next context package starts here.
- Canonical state feeds Control TowerThe human-readable view of current work and its integrity.
- Shared authenticated context (MCP) is shared with AssistantsDifferent assistants retrieve the same durable context instead of relying on copy and paste.
Architecture
The parts that make it dependable
One canonical state
Gabi OS keeps one canonical operational state instead of a separate tracker for each assistant. Workstreams, tasks and execution runs live in Supabase with source references and explicit lifecycle states, so any assistant can recover active work from the same record rather than from conversation history.
- DONE
- ACTIVE
- PARTIAL
- WAITING
- BLOCKED
- UNKNOWN
- RECONCILIATION_REQUIRED
Missing evidence is not silently converted into DONE.
Bounded Task Envelopes
Before anything executes, work is compiled into a bounded Task Envelope. It carries the relevant canonical state, the exact scope and exclusions, acceptance criteria, the evidence required, budgets and limits where relevant, stop conditions and permissions. Its job is to stop a small task from turning into uncontrolled discovery or mutation.
Context compilation
Instead of handing every model the entire history, Gabi OS assembles the smallest context package that is sufficient for the task. Large logs and raw output stay behind evidence pointers rather than being injected into every prompt.
Execution receipts and reconciliation
Workers return structured receipts, not prose. Receipts are reconciled deterministically against canonical state, and completion depends on accepted evidence and read-back. A successful run does not by itself complete a whole workstream, and an ambiguous result returns RECONCILIATION_REQUIRED instead of a guess.
Human and permission boundaries
Some actions deliberately stop and ask. The execution loop halts before consequential permissions rather than automating through them.
Automation should remove unnecessary human relay, not remove necessary human judgement.
Shared context through MCP
An authenticated Model Context Protocol endpoint gives different assistants access to the same durable context. Requests are authenticated, authorised and audited, so information no longer has to be carried between tools by hand.
Control Tower
The Control Tower is the human-readable operational surface for Now, Plan, Sessions and execution, Tasks, History, and Integrity and reconciliation, projected from canonical state. It is part of the v1 operating architecture and is still evolving.
What I built
- AI-system architecture and the orchestration model
- Product and frontend engineering in TypeScript and Next.js
- Supabase-backed canonical state and lifecycle model
- Authenticated MCP shared-context integration
- Deterministic workflow, governance and reconciliation logic
- Testing and acceptance gates
Verified evidence
Accepted, with a clear boundary
The v1 operating architecture reached accepted canonical completion in October 2026, after passing its acceptance gates.
AI Operating System v1 acceptance
- tests passed
- 1,115
- tests skipped
- 3
- TypeScript
- Pass
- scoped lint
- Pass
- production build
- Pass
Gabi OS is an internal system under active development. The core AI Operating System v1 architecture and implementation reached accepted verification, while source-integrity, scaling and continuing operational work remain active. This evidence describes verified architecture and implementation; it is not a claim that every component is deployed or live.
What it demonstrates
AI is most useful when the system around it knows what is true, what is allowed, and what still needs proof.
This is not a collection of prompts or an AI demo. It is AI designed as part of a software system: deterministic code handles deterministic work, AI is used where semantic reasoning earns its place, state is explicit, permissions have boundaries, outputs require verification, provenance matters, and people stay in the loop where judgement is required.
Practical AI systems
Building an AI workflow that needs to hold up outside the demo?
Let’s talk about internal workflows, agents, knowledge systems or human-review architecture, designed around real constraints. No magic, and no promise of full autonomy.
Discuss an AI workflow