feat(models): cap free workspaces to low-tier models

[saimanoj]

Summary

Free workspaces were resolving to top-tier models for chat and most LLM work. This change forces free plans to the "low" model tier when billing is enabled. PRO/MAX, self-hosted (billing disabled), and explicit model overrides are unaffected.

Two paths sent free users to top-tier models; both are fixed at their chokepoints in apps/webapp/app/services/llm-provider.server.ts:

Part A — routed path (getModelForUseCase) — memory, search, sub-agents, titles, summaries, etc. After the explicit-override check, if billing is enabled and the workspace isn't on a paid plan, the effective complexity is forced to "low". planType is read from the existing workspace query (added Subscription to the select — no extra round-trip).

Part B — main chat agent — previously used getDefaultChatModelId() = env.MODEL for every plan (the real "high model for everything"). New resolveDefaultChatModelId(workspaceId):

• free → DB "low" tier (honoring a modelConfig.chat override)
• paid / billing-off / no-workspace → env.MODEL (unchanged)

Wired into the 3 chat entry points (api.v1.conversation._index.tsx, no-stream-process.ts, api.v1.voice.turn.tsx), keeping body.modelId precedence so an explicit per-request model still wins.

Decisions encoded

• Cap level: free → "low" for everything
• Billing disabled (self-hosted): no downgrade
• Explicit overrides: honored over the cap
• Low model source: existing DB "low" complexity routing (no new env var)

Net effect

PRO/MAX and self-hosted are byte-for-byte unchanged; only free hosted workspaces drop to low-tier.

Test plan

• New llm-provider-plan-tier.test.ts — 12 tests, written test-first (confirmed RED → GREEN). Covers FREE cap, override honored, PRO/MAX untouched, billing-off, no-workspace, and the chat resolver.
• Typecheck: 0 new type errors (296 pre-existing, identical with/without this change).
• Lint: 0 errors in touched files.
• All runnable unit tests pass.

🤖 Generated with Claude Code