[codex] stabilize openrouter defaults and app identity

[anand-testcompare]

Summary

This PR temporarily switches Sketchi's default OpenRouter model back to google/gemini-3-flash-preview and improves OpenRouter request identification so activity and provider behavior are easier to trace.

This is a stabilization branch, not the full agent-architecture fix. The longer migration plan for moving the thread flow to Convex agent components is documented locally in .memory/convex-agent-migration-plan.md.

User impact

Users have been seeing inconsistent diagram generation/edit behavior, especially around multi-step structured flows. The immediate effect is that some requests fall into timeout or AI_NoOutputGeneratedError paths without producing a persisted scene.

This temporary change improves the default model path for generation/restructure work and makes OpenRouter traffic clearly attributable to the Sketchi backend in activity logs.

Root cause

There are two separate issues in play:

1. The current inner structured-output harness is brittle. For intermediate generation/modification, the model is expected to generate JSON, call validation, and then restate the same JSON as the final output. That control flow is fragile and model-sensitive.
2. OpenRouter request identity was too thin. We were only sending HTTP-Referer plus X-Title: sketchi (env), which makes traffic attribution harder when several apps share an OpenRouter account.

What this PR changes

• Sets the default OpenRouter model to google/gemini-3-flash-preview
• Updates backend and e2e env examples to reflect that default
• Adds explicit OpenRouter app identity metadata:
  • appName
  • appComponent
  • appIdentifier
  • appTitle
  • normalized appUrl
• Uses X-Title: sketchi/backend (env) instead of the previous ambiguous title
• Adds test coverage for explicit app identity overrides and URL normalization

Validation

Local checks:

• bun x ultracite fix
• bun run check-types
• bun run build
• cd packages/backend && bun run test

Focused verification:

• node .../validate_mermaid_in_md.mjs .memory/convex-agent-migration-plan.md
• direct OpenRouter probes for plain text / object / tool-loop behavior
• real intermediate-path probes comparing google/gemini-3.1-flash-lite-preview vs google/gemini-3-flash-preview
• real thread-level probes using google/gemini-3-flash-preview

Residual risk

This PR does not fix the underlying harness/control-flow issue.

• Structural generation/restructure behavior is healthier on google/gemini-3-flash-preview
• Cosmetic tweak behavior is still inconsistent because the tweak/intermediate subflows still have the same brittle state transition problem

The intended follow-up is to replace the current nested flow with Convex agent components using a parent thread agent and child agents as tools, with deterministic terminal transitions for validated intermediate state.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
sketchi-web	Ready	Preview, Comment	Mar 6, 2026 6:49am

[github-actions]

Browserbase Replays

• Session f8718ad2-f96a-4b78-91da-b5fd343f36d1 (scenario: visual-sanity.ts): replay | debug
• Session 4385b3d9-98c0-4900-a28b-6e579eaab701 (scenario: auth-gates.ts): replay | debug
• Session c9da28e8-ac0f-4228-9746-c784ee97f498 (scenario: opencode-web-continuity.ts): replay | debug
• Session 3f394955-2f71-4b76-9b00-7737582e9d9a (scenario: diagram-studio-happy-path.ts): replay | debug
• Session da105f26-2ced-4b82-9609-efa2c7a95ec9 (scenario: diagram-studio-occ-conflict.ts): replay | debug

Workflow run: 22752563489