Ascenda IDE Extension (VS Code & Cursor)

Privacy-first extension for pairing a local developer tool installation with the Ascenda mobile app and sending AI workload telemetry events. One extension, one package identity (ascenda-one.ascenda) published to both the VS Code Marketplace and Open VSX — Cursor installs the same VSIX from Open VSX. The host is detected at runtime (packages/ide-extension-core/src/host.ts); there is no separate Cursor build.

Part of ai-engineer-tools. For what these measurements do and do not establish, see What this measures.

What it measures

This package is the baseline IDE telemetry source for both hosts. Each row is a named, observable event — a proxy for the input beside it, not a measurement of it:

"Partial proxy" is meant literally: file-save cadence is not focus, and editor switches are not task switches. They are the observable traces those things leave in an editor. Whether the traces track the underlying construct is the open question — see What this measures.

Subjective strain (NASA-TLX-style check-ins), meeting load, and personalised baselines are handled by the Ascenda app and backend — not this extension.

Running in Cursor

Identical telemetry to VS Code, reported with source: cursor_mcp / toolType: cursor_mcp — detected automatically, no separate install or config. Cursor is treated as a first-class telemetry producer for AI engineers; for richer AI-agent signals today, also run Claude Code hooks. Deeper agent-level capture (prompts, tool calls, correction loops) is tracked as a separate future surface — see CURSOR_ADAPTER_PLAN.md — not a second editor extension.

No permission posture (autonomyMode) — on purpose

The hook adapters send an autonomyMode on their events: the posture an agent was working under — default, accept_edits, dont_ask, bypass_permissions and so on, each mirroring the runtime's own word — read from Claude Code's and Codex's permission_mode. This extension deliberately sends no such key, and its absence is not an oversight. Checked 28 Aug 2026:

• Its events are not agent actions. A file save, an active-editor change and a terminal command are things a person did, or things that appeared in a terminal with no record of who started them. There is no per-action approval to be in a posture about, so no value would be true of the row it rode on.
• The API surface exposes no posture. The extension targets the VS Code API baseline in engines.vscode and uses onDidChangeTextDocument, onDidSaveTextDocument, onDidChangeActiveTextEditor and the terminal shell integration events. None of them reports an approval mode, and there is no event for "an agent was allowed to do this without asking".
• A chat auto-approve setting would be the wrong thing. A host's agent settings are readable through workspace.getConfiguration, and reading one would be a mistake: it describes the configuration of a different agent whose actions this extension never observes. Attaching it to a human's file save would manufacture a link that does not exist, and would then pollute the cross-collector cohort comparison the field exists to enable.

Absence is the correct wire state and is already meaningful: autonomyMode distinguishes absent ("this runtime has no such concept") from "unknown" ("a posture arrived that we could not read"). If agent-level capture lands — CURSOR_ADAPTER_PLAN.md — the events it adds would be agent actions, and the question becomes live again for those events only.

What this version provides

• Common Ascenda telemetry event types
• Privacy-safe editor activity events
• Terminal command classification where VS Code shell integration exposes execution events
• Session start/end events
• After-hours AI session signalling
• Event queue + periodic flush, with the undelivered backlog kept on disk across reloads, crashes and a failed final flush (re-sent only when ascenda.telemetry.drainPersistedQueue is on)
• Loose-coupled pairing model:
  • extension stores only toolInstallationId
  • extension stores scoped eventWriteToken
  • backend privately maps tool → anonymous app user → push device

Loose coupling invariant

The extension does not know:

• user email
• phone number
• real name
• Ascenda app user ID
• mobile push token
• organisation identity

The extension only knows:

• toolInstallationId
• scoped eventWriteToken

The backend resolves:

toolInstallationId
  -> anonymous Ascenda user
  -> active app device
  -> APNs / FCM push token

Install

Requires VS Code 1.90+ or any recent Cursor.

Open the Extensions pane — ⇧⌘X (macOS) or Ctrl+Shift+X — search Ascenda, and click Install. The publisher is ascenda-one.

Cursor installs the same extension from Open VSX — same publisher, same build, no separate package to choose between. From a terminal instead:

code   --install-extension ascenda-one.ascenda   # VS Code
cursor --install-extension ascenda-one.ascenda   # Cursor

On macOS neither CLI is on PATH until you run ⇧⌘P → Shell Command: Install 'code' command in PATH (Cursor has the equivalent). The Extensions pane needs no setup at all, so it is the shorter route.

Pair

1. Open the Command Palette — ⇧⌘P / Ctrl+Shift+P — and run Ascenda: Connect App.

   

2. The editor shows a QR code and a six-digit code, valid for a few minutes.

   

3. Scan the QR in the Ascenda app, or enter the code under Connections → Ingest telemetry. On a Dev backend with no phone, ascenda-pairing-sim stands in for the app:

   ascenda-pairing-sim confirm-device-code <6-digit-code>
   

4. Confirm with Ascenda: Show Status, then send one event end to end with Ascenda: Send Test Signal.

This pairing covers the editor. The CLI adapters — Claude Code hooks, Codex hooks, the MCP server — hold their own installation, paired against the same account with one command (npx -y @ascenda-one/claude-code-hooks pair); this extension's token stays in the editor's private secret storage and is not shared with them. See the repo README.

Settings

Preferences → Settings → search "Ascenda", or in settings.json:

{
  "ascenda.apiBaseUrl": "https://api.ascenda.one",
  "ascenda.telemetry.enabled": true
}

ascenda.apiBaseUrl defaults to https://api.ascenda.one and only needs changing to reach a development backend — http://localhost:5002 for a local build, or the Azure Dev host. Every collection toggle is listed under the same search; all default to on and all are per-user.

Events the backend could not take — it was unreachable, the token had lapsed, consent was paused — are kept in the extension's global storage as telemetry-queue.json, so a window reload, a crash or a shutdown while offline no longer discards them. Each payload carries the idempotencyKey minted when it was queued, so re-sending a backlog that overlaps what did get through is answered duplicate rather than counted twice. Re-sending is behind ascenda.telemetry.drainPersistedQueue (default off, until the deployed backend is confirmed to dedupe on that key); while off the backlog is kept and bounded but never sent. The file is bounded by ascenda.telemetry.queueMaxEntries and ascenda.telemetry.queueMaxAgeDays, and every discard is written into the file's own record and to the Ascenda Telemetry output channel, so a truncation is never silent.

Commands

Backend contract

See TOOL_PAIRING_API_REFERENCE.md. Tool-side routes use /v1 (not /api).

Privacy defaults

Metadata-only telemetry. Does not send source code, raw prompts, AI responses, file names, branch names, terminal output, or repository names.

May send:

• event type, timestamp, workspace hash, session ID
• language ID, file extension/type, changed line count bucket
• terminal command class (test/lint/build/typecheck/run/git/install)
• success/failure/cancelled outcome
• after-hours flag
• pairing label: editor name plus machine hostname (never a workspace or repository name — a pairing is editor-wide, so its label is too)

Disallowed metadata keys are stripped server-side.

Not a medical device: this measures workload patterns for self-awareness, not diagnosis or treatment, and makes no clinical claim. Collection rides the ide_telemetry consent scope, which is revocable from the app at any time — revoking takes effect immediately and further ingest returns 401.

Build from source

Not needed to use the extension — the Marketplace and Open VSX both ship it built. This is here so you can read and run what is actually collecting from your editor, which is a fair thing to want from a telemetry tool.

# from the repo root — the workspace install resolves @ascenda-one/* locally
npm install
npm run build:shared

cd ascenda-vscode-extension-telemetry
npm run compile

Debug it live. Open this folder as the VS Code workspace, press F5 (or Run and Debug → Run Ascenda Extension), and a second window opens — the Extension Development Host — running your build. Open any project in it and the commands above work normally. Point it at a development backend with ascenda.apiBaseUrl so you are not writing to production while testing.

Package a VSIX.

npm install -g @vscode/vsce
npm run compile
npx vsce package --no-dependencies
code --install-extension ascenda-<version>.vsix

<version> comes from this package's package.json. Releases stamp the git tag over it in CI without committing back, so a local build is normally numbered lower than the published extension — check the filename vsce prints.

In Cursor, install a VSIX through Extensions → … (Views and More Actions) → Install from VSIX…, which needs no PATH setup. Reload the editor afterwards. Uninstall from the Extensions view like any other extension.

Roadmap

The pairing QR never leaves your machine

The QR encodes the pairing secret, so rendering it is a privacy decision, not a display detail. It is generated in-process and inlined into the panel as SVG (packages/ide-extension-core/src/qr.ts) — the pairing panel makes no network request to draw it, and no image service ever receives the secret.

This was not always true: an earlier build handed the URL to a third-party QR image endpoint, which put pairing secrets in someone else's access logs. It is now a regression test rather than an intention — tests/qr.test.mjs includes "the pairing secret never reaches a remote URL", so a change that reintroduced the leak would fail the build.

Called out because "we render our own QR" is exactly the kind of claim worth being able to check yourself: qr.ts is about sixty lines.