[ffahleraz]

Summary

Adds OpenScale as a first-class provider (MODEL_PROVIDER=openscale). OpenScale serves DeepSeek V4.1 Flash and GLM 5.3 Flash through an OpenAI-compatible Chat Completions API. The change follows #133 and the AtmoRouter provider file for file: same 15 files, same invoker shape, same tests.

Keys are created at app.openscale.so/app/keys. The model list, with context windows, limits and prices, is at GET https://api.openscale.so/v1/models (public, no key).

Changes

• src/config/index.js: OPENSCALE_API_KEY, OPENSCALE_MODEL (default deepseek/deepseek-v4.1-flash), OPENSCALE_ENDPOINT (default https://api.openscale.so/v1/chat/completions); added to SUPPORTED_MODEL_PROVIDERS, the primary key check, the fallback checks and reloadConfig().
• src/clients/databricks.js: invokeOpenScale() with the same Anthropic to Chat Completions conversion, tool handling, no replay of billable POSTs, and 429 escalation as invokeRuncrate(); registered in PROVIDER_INVOKERS and exported.
• src/routing/index.js, src/routing/model-tiers.js, src/orchestrator/index.js, src/orchestrator/sse-transformer.js, src/clients/provider-capabilities.js, src/api/providers-handler.js: same entries as the other OpenAI-compatible providers.
• bin/lynkr-init.js: OpenScale in the init wizard and PROVIDER_ORDER.
• .env.example, README.md, documentation/providers.md (section 8e): configuration and tier examples.
• test/openscale-integration.test.js (mirrors test/runcrate-integration.test.js), test/dispatch-registry.test.js, and the new test added to test:unit in package.json.

Testing

• node --test test/openscale-integration.test.js test/runcrate-integration.test.js test/atmorouter-integration.test.js test/cheaperinference-integration.test.js test/dispatch-registry.test.js test/init.test.js (with the test:unit env vars): 49 tests, 49 pass, 0 fail.
• npx eslint src index.js: no errors.
• npm run test:unit: 1561 pass, 94 fail, 10 cancelled on this branch; 1555 pass, 94 fail, 10 cancelled on unmodified main (dd73bbf) on the same machine. The failing test names are identical, so these failures are not from this change, and the 6 new tests pass.
• Live check: Lynkr from this branch with MODEL_PROVIDER=openscale against https://api.openscale.so/v1/chat/completions, using Claude Code-style /v1/messages requests. All returned HTTP 200:
  • deepseek/deepseek-v4.1-flash: plain message (end_turn, exact expected text), streaming (end_turn), get_weather tool call with {"city": "Jakarta"} (tool_use), tool result round-trip (end_turn).
  • z-ai/glm-5.3-flash: tool result round-trip (end_turn), streaming tool call with a thinking block and {"city": "Jakarta"} (tool_use).

This PR was prepared with AI assistance: an AI coding agent wrote the change and ran the checks above.

Disclosure: I work on OpenScale.

🤖 Generated with Claude Code

[github-actions]

✅ OpenCodeReview: Review complete: 0 finding(s) across 10 selected item(s).

[veerareddyvishal144]

Thank you so much for your contribution @ffahleraz