npxt.graph progress updates

This is a quick day-end update on npxt.graph. A few things are still in development and fine-tuning, but the recording is meant to show the current situation and where the graph is heading. The app is running locally as “wireframe v19” (browser tab title, localhost:8819), demonstrated on a sample company, Cerox Engineering · Pvt Ltd.

The walkthrough covers five things in order: the built-in onboarding tour (introduction, how to read the graph, and the rules), the graph window itself, builder mode and its side panels, the new-company wizard, and an honest note on the one part that isn’t wired up yet — the AI backend.

Full 6:40 walkthrough. Pick a quality below — the player switches without losing your place. If scripts are off, the middle quality plays by default and each quality is also a direct link.

Opening the app starts a short tour with three parts: 1 · How it’s built, 2 · Reading it, and 3 · The rules. It’s described as a simple introduction to what the graph is and how it works — “how to read the graph, what the items are” — followed by a slightly longer rule-set section covering the main rules already discussed. The tour can be skipped at any time to go straight to the graph.

Part 1 · How it’s built — the pipeline

The first part lays out how a company graph is produced, stage by stage. It opens on the owner’s brief and ends with the graph constructing itself:

This is the graph window that has been the ongoing subject of discussion — the Cerox Engineering map. Its State 0 is considered to look right here, because the data for this company was entered by hand. Stepping into a node “just expands to everything” as you move through each state — which is working as intended, though adjustments are still needed and some problems remain (which is partly why the new-graph flow, below, was built next).

How to read a node

These four node kinds and the “size = importance” rule are shown in the on-screen legend at the bottom of the graph.

What State 0 contains

The whole-company view is made of a few bands. For Cerox this is: the company base bar, six operating branches, two flagship products, the people who run it, four shared services, and one urgent “ready-now” strip across the top.

Four shared services the whole company works through: Telegram — “the company’s nerve centre” Zoho Books — “the canonical financial record” GitHub — source control G Drive — shared storage

People: the Directors cluster and Hanz. Urgent “ready-now” strip (dashed action cards along the top): “Nanosmoothies site · due Sun 12 Jul,” “Recruit marketing officer,” and “Approve monthly payment run.”

A builder-mode switch flips the graph from viewing to editing — “you can edit anything, change stuff, everything is there.” In builder mode the workspace shows three supporting views plus an AI panel:

• Navigator (left) — a nested tree of the whole company, “like in Figma”: branches, sub-nodes, flagships, and concepts.
• Notifications — a “Waiting at the gates” list of items needing attention (e.g. from Ruwan, Hanz, Udara, Claude). The speaker notes this is provisional — “we can discuss what to put here.”
• Presence — a “Who’s around” list showing who is active and where (Ruwan on the Nanosmoothies map, Udara on Mad Studios automation, Hanz logging invoices in Zoho, Niranjan on general upkeep).
• State Creator Assistant (right) — an AI chat panel that “helps with tasks on the graph itself”: add a node, set weight & load, name a cluster, expand a concept, or suggest a tool. It’s labelled “Sample only — nothing changes silently.”

Back in User mode, a “new graph” function handles working across more than one company — “say you work for two companies and need to enter the other one.” It launches a wizard that begins from the owner’s brief, following the same principle shown in the tour: AI proposes → human signs → checker confirms.

• Company name — the one required input (placeholder “e.g. Lumen Bakery Ltd”).
• Owner statement — a basic brief from the owner: what the company is, who the clients are, who does what, where the money moves, which apps hold the work. This seeds the AI’s understanding.

The intended flow after the statement: the AI fills in the company’s details (name, brand, and so on), then asks follow-up questions — telling you what’s missing and prompting for anything left out. It then prepares the sheet to sign; sign-off and a built-in checker/validation step gate the result, and only then does the graph construct. Additional stages exist, as laid out in the “How it’s built” tour.

The wizard is powered by AI that, in this version, is not yet connected to the backend. Because of that, the wizard can’t advance past the statement step — entering test data (e.g. company name “nanosmoothies LTD”) and pressing Next does nothing, since the AI plugin isn’t wired up. The step that should fire is npxt.organization.create.

The backend work is happening on the speaker’s other screen: the wizard’s AI calls route through a local “claude bridge” (development branch feat/v19-claude-bridge). As shown in the terminal, the bridge’s claude -p call was timing out at ~85 s — the likely cause being that each call boots the full Claude Code environment (all MCP servers, plugins, hooks), which is slow. The plan discussed earlier is that this can later be connected to a cloud API; for now it simply uses the user’s cloud account via the terminal.

The tour’s third part is the rule set. The headline rule for the whole-company view:

Summary in the speaker’s own words: the graph is “almost done.” The main remaining task is getting the AI part working so the flow can be properly tested.

• Get the AI wired up — connect the wizard to the backend so it advances past the statement step.
• Test on two or three more companies — e.g. Nanosmoothies and one or two others — to shake out problems.
• Keep refining the rules — not fixes so much as more rules, plus additions and adjustments, since the model differs a lot per company. A general rule set is already in place, but there’s room to improve it.

This document is that report.

Verbatim automatic transcription (Whisper large-v3), with timestamps. Minor speech artefacts are left as spoken; the sections above resolve them against what’s on screen.