Upsell ai reasoning effort

[YoussefKeyrouz]

Task/Issue URL: https://app.asana.com/1/137249556945/project/1212810093780571/task/1214635053937863?focus=true

Description

Part 2 of 2 — Subscription upsell trigger for gated reasoning efforts.

Part 1 added schema support for the new reasoningEffortAccess field on the /duckchat/v1/models response. This PR wires the user-facing upsell: tapping a gated reasoning mode on an accessible model now routes the user into the native SubscriptionPurchase (FREE → PLUS/PRO) or SubscriptionUpgrade (PLUS → PRO) flow, mirroring the model-level upsell already shipped on #8625.

What changed:

• Data layer plumbing: AvailableReasoningMode now carries an optional access: ReasoningEffortAccess? plus a derived isAccessible flag. ReasoningResolver.availableModes(supported, effortAccess) picks the first candidate effort per mode that is supported and accessible; if none are accessible, it falls back to the first supported one so the row still renders for upsell routing. Missing per-effort access metadata is treated as accessible.
• Reasoning picker upsell wiring: ReasoningModePickerViewModel.onModeTapped(mode, surface) runs the same decision table as the model picker: accessible → select; gated → route to LaunchPurchase/LaunchUpgrade based on access.requiredTier + current userTier.
• Shared upsell types: PickerSurface and UpsellCommand live in a new shared file. The model picker was re-pointed to use these instead of its previously nested types.

Notes:

• Product request: Hardcoded gate for gpt-5.2 EXTENDED_REASONING until BE is ready.

Steps to test this PR

• Free account, gated model tap (already shipped, regression check): Tap a PLUS-only model from the model picker → SubscriptionPurchase opens with the model-picker funnel origin.
• Plus account, PRO-gated reasoning tap: Open GPT-5.2, tap EXTENDED_REASONING → SubscriptionUpgrade opens.
• Tier downgrade: While on PLUS/PRO with a gated reasoning mode selected, downgrade by removing the subscription. The picker auto-selects the first accessible model/mode and submitting a prompt sends that mode's effort, not the previously-cached gated effort.

UI changes

screen-20260520-202439-1779319456993.mp4

---

Note

Medium Risk
Introduces new per-effort accessibility logic for reasoning modes and routes users into purchase/upgrade flows, which can affect subscription funnel behavior and persisted selection state. Also adds a temporary hardcoded gate for gpt-5.2 that could mis-gate if model IDs or backend behavior change.

Overview
Adds per-reasoning-effort access awareness and upsell routing: AvailableReasoningMode now carries optional access metadata and ReasoningResolver.availableModes selects the first supported + accessible effort per mode (falling back to a gated effort so the row can still render for upsell).

Updates RealDuckAiModelManager to validate persisted reasoning selections against accessibility (auto-selecting/persisting the first accessible mode, refusing gated selections, and allowing selectedReasoningMode to become null when all efforts are gated). The reasoning-mode picker now triggers the native SubscriptionPurchase/SubscriptionUpgrade flows when a user taps a gated mode, using shared PickerSurface/UpsellCommand types (also adopted by the model picker).

Includes a temporary, hardcoded PRO-only gate for gpt-5.2 EXTENDED_REASONING until backend per-effort gating ships, and expands tests to cover gating/auto-promotion/upsell behavior.

^{Reviewed by Cursor Bugbot for commit c92886a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[YoussefKeyrouz]

• Add reasoningMode and getChatById to DuckAiChat Store #8659
• Upsell ai reasoning effort #8648 👈 (View in Graphite)
• Add reasoning effort access to the AI chat model #8625
• develop

This stack of pull requests is managed by Graphite. Learn more about stacking.

[malmstein]

nice refactor pulling routeUpsell and PickerSurface into a shared file — the reasoning picker now mirrors the model picker pattern cleanly, and the test coverage for the new per-effort gating is solid (the ReasoningResolverTest cases are particularly thorough). couple of small things: the hardcoded gpt-5.2 gate overwrites rather than merges and has no test of its own, and the silent else -> null in routeUpsell loses a useful debug log.