The Week It Watched Itself

One engineer, one week, a graph that observes Amplifier

An observability system for the tool that built it

The Context Intelligence Server captures Amplifier's own sessions as structured events and persists them as a graph. It watches the thing it was built with.

First, how much of it got built — and by how many people.

5 / 8
node types / edge types in the property graph
7
calendar days of the build burst (03-13 to 03-19)

One engineer built the whole first week

Diego Colombo was the sole author across 2026-03-13 to 03-19. The numbers are real, and they landed fast.

But a commit count doesn't tell you what he actually shipped.

~250
commits in the week (102 feat)
~19.6k
Python LOC across 69 files
879
test functions by 2026-03-19
1
author: Diego Colombo
The Turn Against Volume

Volume isn't the story

Hundreds of commits don't tell you what was actually built or why it's different. Of the ~250, only 169 are first-parent and just one arrived as a merged pull request. Scale is the setup — not the point.

The point starts with a plan written before any code.

Day 1 was a plan, not a scramble

2026-03-13 held only two commits: the initial commit and a single design document for a FastAPI + Neo4j telemetry server — written before any code.

That plan defined the machinery: how sessions become a graph.

426
lines in the Day-1 design doc, covering Docker Compose topology, POST /events ingestion, per-session async workers, blob storage and a Cypher proxy
docs/plans/2026-03-13-context-intelligence-server-design.md

Sessions become a queryable graph

Live sessions POST in through a fire-and-forget ingest API and land in a Neo4j property graph — one path from event to structure.

Once ingestion worked, the build hardened around it.

Sessions ingest via POST /events — fire-and-forget, 5s timeout, failures logged as warnings
5
node types
8
edge types

From ingestion to an operable platform

Across five active days (03-14 to 03-19) the graph grew from plumbing into a system that could replay, route, and stay up under load.

Hardened enough to finally watch a running system: itself.

14
Foundations
FastAPI ingestion server, Neo4j store, handler pipeline, blob storage, dashboard
15
Intelligence service + A2UI
WebSocket exploration UI, session manager, landing navigation
16
Routing + composition
6-phase composition pipeline and routing matrix generation
17
Robustness
Replay resilience, cursor persistence, request idempotency guard
18
Operability
gunicorn+uvicorn graceful shutdown, uv tool install, YAML config

On 03-19 it turns the lens inward

The platform starts recording its own runtime into the graph — Amplifier watching itself run. Self-observation of runtime, made concrete.

Which reframes what the week was actually about.

Per-run token accounting
run_tokens accumulate at llm:response and flush at orchestrator:complete (verified at week-1 snapshot e5d88ce)
SystemEventHandler
tracks context:compaction, cancel:requested and cancel:completed events
Recipe-loop enrichment
recipe events enriched with recipe:start context
The Takeaway

The threshold wasn't speed — it was architectural

From a bundle hook to a standalone graph that lets Amplifier see itself. The ~250 commits and ~19.6k lines were the means; the keeper is a system that observes its own sessions.

Self-observation, not just scale.

Sources

Research Methodology

Feature status: Active on main

Repository: microsoft/amplifier-context-intelligence (branch main, HEAD 440bc5d, 549 total commits).

Data as of: week-1 window 2026-03-13 to 2026-03-19; snapshot metrics anchored to commit e5d88ce (2026-03-19 23:45).

Commands run:

Gaps & caveats: ~250 is the author-date grouping (235 under --since/--until commit-date semantics); 169 first-parent commits, 1 merged PR. run_tokens and SystemEventHandler were later refactored — anchored to snapshot e5d88ce, not main HEAD.

Primary contributor: Diego Colombo — sole author of the first week (533 repo-total commits).

More Amplifier Stories