SmartX PRISM

Captures Delivery Loop activity and AI usage from the developer's IDE and posts it to the PRISM OS backend. Works in VS Code, Cursor, and Kiro (all VS Code-derived).

Install (dev)

cd prism_os_extension
npm install
npm run compile

Then in VS Code: F5 to launch an Extension Development Host with this extension loaded. In Cursor / Kiro the same dev workflow applies — they inherit the VS Code extension API.

To package for distribution:

npm install -g @vscode/vsce  # one-time
npm run package              # produces smartxprism-0.1.6.vsix

Install the .vsix via code --install-extension smartxprism-0.1.6.vsix.

Activation

The extension activates on onStartupFinished — i.e. in every window once installed. While signed in it sends a lightweight, sessionless presence heartbeat so the admin Extension-health dashboard goes green on install + sign-in, regardless of which folder is open.

Telemetry capture (file edits, git, AI events, the per-project DevSession) is a separate, narrower gate: it starts only for workspaces that contain a .smartx/session.json file at the root. Opening a personal repo therefore registers presence but collects no telemetry. The config file looks like:

{
  "project_id": "6498dd8f-18dd-42b0-9f9b-498e75b5da96",
  "current_lap_item": null
}

project_id is the UUID of the PRISM project this workspace maps to. current_lap_item is optional and can be set later via the PRISM: Link current work to a Lap item command.

First-run

1. Open a workspace with a .smartx/session.json.
2. Run the PRISM: Sign in to SmartX command (Ctrl/Cmd-Shift-P → search "PRISM").
3. Reload the window. A $(pulse) PRISM · <project> indicator appears in the status bar. The session is live.

What it captures

AI integration

Native AI tools (Cursor's composer, Claude Code, Copilot, Kiro's chat) do not expose hooks for third-party extensions to observe their calls. The integration model is opt-in: an AI extension or wrapping script calls prism.recordAIUse with the call metadata, and the PRISM extension handles scrubbing → policy decision → server post.

const outcome: { decision: 'ALLOW'|'ADVISE'|'WARN'|'BLOCK'; reason: string } =
    await vscode.commands.executeCommand('prism.recordAIUse', {
        provider: 'anthropic',
        model: 'claude-sonnet-4-6',
        prompt_text: '...',                  // scrubbed locally before send
        prompt_tokens: 1234,
        completion_tokens: 567,
        completion_excerpt: '...',
        accept_status: 'ACCEPTED',           // or REJECTED / PARTIAL / PENDING
        skill_kind: 'CODE_REVIEW',
        skill_versions: ['[email protected]'],
    });

if (outcome.decision === 'BLOCK') {
    // Policy said no — abort the AI call.
}

For shells / scripts outside the IDE, use the PRISM CLI (separate package, ships next).

Privacy

• Workspace path is hashed before send (sha256). Server never sees the raw path.
• Prompt text is scrubbed locally for AWS keys, JWTs, private-key blocks, emails, and phone numbers before send. The server re-runs the same sweep as defence-in-depth.
• JWT credentials live in OS keychain via vscode.SecretStorage — never on disk in plain text.
• Telemetry events are buffered locally only when offline; when network is healthy nothing persists on the dev machine beyond the JWT.

Settings

Offline behavior

Failed flushes spill to .vscode/.smartx-buffer.jsonl. The buffer is drained chronologically on the next successful flush. Past prism.offlineBufferMaxHours of unflushed events the extension shows a sticky warning and starts dropping new events until reconnect.

Commands

• PRISM: Sign in to SmartX
• PRISM: Sign out
• PRISM: Record AI use (programmatic) — invoked by other extensions, not the user
• PRISM: Link current work to a Lap item
• PRISM: Show session status

Releasing

One artifact, three destinations — npm run release packages a single .vsix from the current package.json version and ships that same byte-identical file to every channel so they never drift:

1. PRISM OS self-hosted — auth-gated; the in-extension updater pulls it, so it reaches VS Code + Cursor + Kiro uniformly.
2. VS Code Marketplace (vsce publish) — VS Code only.
3. Open VSX (ovsx publish) — where Cursor and Kiro look.

Every channel is attempted independently; one failure (e.g. "version already exists") doesn't abort the others, and the script exits non-zero if any channel failed.

# 1. bump the version first
npm version patch --no-git-tag-version    # or edit package.json

# 2. release everywhere
npm run release

Auth comes from env vars (a local ./.env, which is gitignored, is auto-sourced for convenience):

Skip channels with flags — e.g. internal-only:

npm run release -- --no-marketplace --no-openvsx                 # PRISM only
npm run release -- --min-version 0.1.6 --notes "Presence health" # force-upgrade + notes