TokenGovernor for VS Code

Budget control for AI jobs inside the editor. The extension talks to the TokenGovernor API (/v1/jobs) and shows spend, reservations, and risk live while work runs through the governor.

Overview

TokenGovernor is a spend firewall for AI agents. Every model call is routed through a deterministic count → price → reserve → cap → dispatch → settle pipeline: it measures the exact input size for free, holds the worst-case cost against a fixed budget, caps the output length, dispatches, then refunds the unused reservation. If a call can't fit the budget, it's refused cleanly instead of overspending.

This extension is the live cockpit for that engine. It also drives a governed CI-repair agent: point it at a broken repo fixture and it reads the failing tests, proposes patches through sandboxed tools, re-runs the suite, and stops the moment the build is green or the budget is exhausted — with an independent, sealed evaluator confirming the fix in a fresh checkout so success can't be gamed.

What it does

• Status bar: committed spend / budget for the active job, colored by the same risk thresholds the governor core uses (WATCH at 70% committed, AT_RISK at 90%).
• Sidebar view: a Governor icon in the Activity Bar opens a vertical job view — committed spend hero number, budget bar (settled + reserved), accounting breakdown, and every submitted request with its phase and status, so you can see which steps to look into. A one-click toolbar drives every action, and CI-repair results (verdict, spend, sealed-evaluation checks, and the full governed trace) render right in the panel. Sits alongside the file explorer like Copilot does; works in VS Code and Cursor.
• Commands (Cmd+Shift+P, category "Governor"):
  • Health Check
  • Create Job (prompts for a USD budget, e.g. 0.075)
  • Set Active Job
  • Open Job View
  • Submit Selection Through Governor (sends the editor selection as the model input for a work or final response)
  • Finalize Pending Response (accept/reject a settled final output)
  • Run Governed CI-Repair (runs the repair agent on a fixture under budget)

How it saves money

The governor bakes in a set of cost-control tactics. Implemented today:

• ✅ Count before you call — measure the exact size of each request for free, so we never overpay on a guess.
• ✅ Set money aside, then refund — hold the max a call could cost, then give back whatever it didn't use.
• ✅ Cap the answer length — limit how much the model can write so a reply can't blow the budget.
• ✅ Don't repeat dead-end work — if nothing changed, block the agent from running the same failing step again.
• ✅ Shrink giant logs — cut a huge test log down to just the lines that explain the failure.
• ✅ Reuse instead of resend — when a result is identical to last time, point back to it (content-addressed artifacts) instead of paying twice.
• ✅ Always keep enough to finish — reserve budget so the agent can deliver a final answer instead of running out mid-task.
• ✅ Quit while you're ahead — refuse calls that won't fit the budget and stop cleanly instead of wasting money.
• ✅ Use the cheapest model that works — default to the low-cost model and only upgrade when we truly need to.

On the roadmap:

• ◻ Only run the tests that matter — run the tests a change could break, not the whole suite.
• ◻ Look before you read — guess where the bug is and read just those lines, not the whole file.
• ◻ Ignore unreliable tests — don't waste tries "fixing" tests that randomly fail on their own.

Run it (demo)

1. Start the governor server from the repo root:

   npm start                      # scripted provider ($0)
   OPENAI_API_KEY=... npm start   # live OpenAI calls
   

   For live calls, set OPENAI_API_KEY first and create the job afterward.

2. Open the Governor view: click the Governor icon in the Activity Bar, or run Governor: Open Job View from the Command Palette.

3. Click Run CI-Repair to watch a governed repair run end to end, or Create Job → budget 0.075, select text in any file, then Submit Selection Through Governor. Watch the status bar and sidebar update as the governor reserves, settles, and (when over budget) refuses requests.

Settings

• tokenGovernor.apiBaseUrl (default http://127.0.0.1:8787)
• tokenGovernor.pollIntervalMs (default 2000)
• tokenGovernor.defaultFixtureId (default broken-sum)
• tokenGovernor.providerMode (fake | hybrid | live, default fake)