Demote Cursor/OpenCode drivers to experimental (preserve future provider expansion)

[badcuban]

Cursor and OpenCode are not actively supported provider paths, but we want to keep the door open to re-adding them (or other providers) later. Rather than deleting the drivers, demote them to an explicitly experimental tier with zero runtime and UX surface.

Direction

• Codex and Claude Code remain the maintained provider paths.
• Cursor/OpenCode drivers stay in-tree and compiling (per AGENTS.md) as living reference implementations of the ProviderAdapter contract — useful both for future first-party providers and for OSS contributors.
• They should carry no default runtime or UI surface: not selectable, not registered, not marketed.
• Reactivating one later should be a deliberate, small change (re-register + re-test), not a resurrection of stale deleted code.

Checklist

• Stop registering Cursor/OpenCode as built-in runtime providers; old persisted settings/threads referencing them degrade to an "unavailable provider" state instead of crashing.
• Remove Cursor/OpenCode from provider creation/settings UI and any marketing/README positioning.
• Mark both drivers as experimental in code comments and CONTRIBUTING/README so their status is clear to contributors.
• Keep their test suites compiling; consider trimming to smoke-level coverage so they don't tax CI.

Notes

• OpenCode is not only a driver: OpenCodeTextGeneration powers branch-name/title generation. That dependency is out of scope here and stays.
• The earlier revision of this issue also covered removing remote/hosted surfaces; that direction was superseded by the phone-link and hosted-web work that has since shipped.

[badcuban]

Started the first local slice in the working tree:

• Removed Cursor/OpenCode from the visible provider picker/settings metadata.
• Stopped registering Cursor/OpenCode as built-in server drivers.
• Kept non-default legacy Cursor/OpenCode settings routable as unavailable provider shadows.
• Narrowed README/marketing copy to Codex + Claude.

Validation run locally:

• bun run test src/provider/Layers/ProviderInstanceRegistryHydration.test.ts
• bun run test src/components/settings/ProviderSettingsForm.test.ts
• bun run test:browser -- src/components/settings/SettingsPanels.browser.tsx --reporter=verbose
• bun fmt
• bun lint (passed with existing warnings)
• bun typecheck

[badcuban]

Phase 1 status update: local identity separation is complete and released in nightly v0.0.18-nightly.20260601.47. Also completed the checklist item to hide/disable the remote Connections UI. Next implementation slice: remove Cursor/OpenCode from provider creation/settings UI while preserving unavailable/legacy handling for any existing persisted provider/thread state.

[badcuban]

Landed the next provider-settings slice on main in fa64165e. The add-provider flow and maintained provider settings metadata now expose only Codex and Claude. Deprecated Cursor/OpenCode defaults stay hidden unless an existing explicit/default legacy setting or live unavailable snapshot exists, and custom deprecated instances still render as preserved compatibility rows with read-only driver config instead of crashing or being recreated as active providers.\n\nValidation: bun fmt, bun lint, bun typecheck passed. Targeted web tests passed for ProviderInstanceCard, SettingsPanels.logic, and ProviderSettingsForm. A full bun run test attempt hit an unrelated MessagesTimeline timeout under the dirty local worktree, and that file passed when rerun directly, so I did not cut a nightly from this slice yet.

[badcuban]

Rewrote this issue: instead of deleting the Cursor/OpenCode drivers, we're demoting them to an experimental tier (kept compiling, no default registration or UI) so future provider expansion stays cheap. The old remote-feature removal items were dropped — they conflicted with the phone-link/hosted-web direction that has since shipped.