Add Requesty as an opt-in OpenAI-compatible model provider

[Thibaultjaigu]

Summary

Adds Requesty (router.requesty.ai) as an opt-in gateway provider alongside OpenRouter. Pi has no built-in Requesty id, so it is registered through the existing openai-compatible models.json path with REQUESTY_API_KEY and an optional REQUESTY_BASE_URL for the EU router. One key reaches 700+ models plus Requesty's managed routing policies (short ids such as claude-sonnet-4-5 or gpt-5.4-mini).

Changes

• New src/model/requesty.ts: constants, base URL resolution honoring REQUESTY_BASE_URL, catalog fetch with an 8s timeout, and field mapping to ModelsJsonModelConfig (context window, output limit, input/output/cache pricing converted to per million, reasoning and vision flags, chat models only). GET /v1/models/managed is the primary source; GET /v1/models is merged on request or used as the fallback when the managed list is unavailable.
• src/model/commands.ts: feynman model login requesty walks through base URL (EU hint), key paste or $REQUESTY_API_KEY, and the catalog choice (managed only, about 150 entries, or managed plus full catalog, about 900). CustomProviderSetup gains typed models and a verifyModelIds opt-out, since managed ids are not all listed by /models.
• src/model/models-json.ts: exports ModelsJsonModelConfig; ProviderConfigPatch.models is typed with it.
• src/model/api-key-providers.ts, src/model/catalog.ts, src/workbench/credential-catalog.ts: requesty entry after openrouter, label, credential section.
• .env.example, README.md, RELEASES.md (Unreleased, Model providers), website docs (setup.md new Requesty section, configuration.md, cli-commands.md).
• Tests: new tests/requesty.test.ts (8 tests: mapping, defaults, merge, env base URL, mocked managed first / merge / fallback, Pi registry round trip via models.json, label and credential catalog). tests/model-harness.test.ts gains REQUESTY_API_KEY in the env key list and a gateway test; tests/catalog-snapshot.test.ts adds requesty/deepseek-v4-pro to the bypass list.

Requesty is not added to any automatic model preference, judge or fallback list; it is only used when selected.

Validation

• npx tsc --noEmit and npm run typecheck (including workbench-web) pass. git diff --check clean, no lockfile churn.
• tests/requesty.test.ts 8/8 pass; catalog-snapshot, model-harness, content-policy, workbench-setup-decisions and workbench-secret-ledgers all pass. Full npm test: 1210 pass, with 1 failure (native-bundle-runtime offline repair, blocked by the sandbox) and 3 cancelled stream watchdog tests that behave identically on a clean main in this environment.
• Live, through the repo code path (fetchRequestyCatalog to upsertProviderConfig to createModelRegistry to pi-ai complete): managed catalog 154 models, merged 932, registry exposed 885 requesty models, and requesty/openai/gpt-4o-mini returned a reply with stop finish reason. The ids used in docs and tests (openai/gpt-4o-mini, claude-sonnet-4-5, gpt-5.4-mini, deepseek-v4-pro) were confirmed in the live listings.

Disclosure: I work at Requesty. Happy to adjust anything to match project conventions.

[vercel]

@Thibaultjaigu is attempting to deploy a commit to the Advait Paliwal's projects Team on Vercel.

A member of the Team first needs to authorize it.