fix(server): Cursor turns send the picker's default options

[adampeterhiggins]

Problem

T3 only sends Cursor the options the user changed, and Cursor treats an omitted option as the model's standard tier. The picker shows Cursor's default variant: 1M for Opus, Sonnet, Fable and GPT-5.x, and 500K for Grok 4.7. But these all actually ran at about 300K:

• an untouched composer, which sends only fast=false
• delegate_task with only an effort, and delegate_task to another model
• thread launches and scheduled tasks

Separately, delegate_task with options on the parent's own model replaced the parent's options, so an effort-only child of a parent that chose 300K lost that choice. This is upstream pingdotgg/t3code#16149, which is triaged with no fix yet.

Fix

• withCursorDefaultParameters (apps/server/src/provider/cursorSdkModel.ts) fills each option the selection leaves out with the value the picker shows for that model, using the same getProviderOptionCurrentValue the composer uses. Chosen options always win. Fast stays off unless chosen. Cursor defaults Fast on for Grok, Composer and some GPT models, but T3 deliberately keeps it off (withImplicitFastModeDefault).
• CursorDriver applies it where it already wraps the SDK runner per instance, on agent open and on every send, using that instance's provider snapshot. That covers composer turns, resumes, delegate_task, launches and scheduled tasks without touching the adapter. No extra network call: the snapshot is the catalog the picker already uses.
• delegate_task on the parent's provider and model now merges the requested options over the parent's by id. On a different model it still sends only what was requested, which the driver then fills with defaults. This applies to every provider. The option description agents see says so.

Text generation (commit messages, titles) still sends only the configured options, so those small prompts don't move into Max Mode.

Cost: default Cursor turns on models whose default is the long tier now run at 1M (500K for Grok 4.7) in Max Mode, billed per token on usage-based plans. That's what the picker already claims, and what Cursor's own apps do.

Verification

• New tests: withCursorDefaultParameters goes through the real catalog → descriptor mapping, covering fill-all, keeping chosen options (including Fast on and 300K), and models with no options. The delegate_task merge test fails on the old code.
• vp test run on CursorProvider, CursorDriver, CursorAdapterV2, CursorAgentSdk, cursorSdk, OrchestratorMcpService (+ spawn, toolkit integration) and spawnPolicy: all pass.
• tsc --noEmit for apps/server and packages/contracts: 0 errors. vp lint and vp fmt clean on the changed files.
• Live, real Cursor account, Opus 5.5 with what an untouched composer sends (fast=false), effective window read from Cursor's checkpoint:

	Params sent	Window
Before	fast=false	300,000
After	fast=false, effort=medium, context=1m	1,000,000

Not exercised: a full T3 turn through the UI; mobile and desktop share this server path.

docs/fork-differences.md has a section to remove once upstream fixes pingdotgg#16149.

Model: Claude Opus 5.5 (1M context). Harness: Claude Code.

🤖 Generated with Claude Code