[knowhycodata]

Summary

Adds LLMTR (https://llmtr.com) as a gateway provider. LLMTR serves other vendors' models, and a few of its own, behind one OpenAI-compatible API at https://llmtr.com/v1. Disclosure: I maintain LLMTR.

The change follows the Opper addition (#869):

• raven/providers/registry.py: a ProviderSpec at the end of the resale gateway block. It uses OpenAI's driver, is identified by llmtr.com in the address, uses OPENAI_API_KEY as the driver's variable, and has no default model and no shortlist. strip_model_prefix stays off. A model LLMTR publishes as llmtr/<name> is stored as llmtr/llmtr/<name>, so that segment is sent upstream. LLMTR declares a chat address only. A family guess that would send a model to /v1/responses steps aside, so a chat-only id such as qwen/qwen3.8-27b-free stays on chat completions.
• That chat-only address is native_api_bases=(("chat", ...),). Raven infers responses for qwen, deepseek, gpt, kimi and minimax ids, and a gateway with no declared addresses accepts that guess at its default base; many LLMTR models are chat-only (qwen/qwen3.8-27b-free returns a 400 on /v1/responses). It works the same way as for Moonshot and Groq. A user who sets protocol explicitly still gets what they set.
• raven/config/schema.py: ProvidersConfig.llmtr.
• raven/cli/onboard_commands.py: a row in the onboarding key group.
• ui-web/src/assets/providers/llmtr.svg and ProviderMark.tsx: the vendor mark, hybrid tone (crescent in currentColor, star kept red), with a LICENSES/README.md entry.
• Counts written elsewhere now say fifty-seven specs, twenty-three of them gateways (ui-web/CONTEXT.md, ProviderSide.tsx, Roles.tsx, the registry docstring and the gateway block comment).
• Tests: the pinned provider set and count, the LiteLLM name tables, NO_SHORTLIST_VENDORS, tests/data/wire_model_baseline.json (only the new llmtr block), and llmtr added to test_an_inferred_native_protocol_steps_aside_where_the_provider_has_no_address_for_it.

Keys: https://llmtr.com/dashboard/api-keys

Runcrate (#1143) and Atlas Cloud (#1009) also move the counts to fifty-seven. Whichever lands first, I'll rebase and bump the numbers.

Type

• Fix
• Feature
• Docs
• CI / tooling
• Refactor
• Other

Verification

• uv run --frozen --python 3.12 --all-extras pytest -q tests/test_provider_protocol.py tests/test_provider_catalog.py tests/test_provider_wire_model.py tests/test_litellm_provider_model_names.py: 839 passed, 7 skipped

• pytest tests -k 'protocol or provider or registry or factory': the same set of failures as on main, nothing new.

• ruff check raven tests: clean. ruff format --check on the touched Python files: already formatted.

• Live check with a real key through Raven's own factory: Config with providers.llmtr.apiKey and model llmtr/qwen/qwen3.8-27b-free gives provider llmtr and get_api_base https://llmtr.com/v1. make_provider logs that the responses guess steps aside to chat and returns a LiteLLMProvider. One chat(..., max_tokens=32) call went to POST https://llmtr.com/v1/chat/completions with model qwen/qwen3.8-27b-free and returned LLMTR_OK (finish_reason stop, 20 prompt / 5 completion tokens, about 1.3 s). Before the native_api_bases line was added, the same call went to /v1/responses and got a 400.

• Relevant tests pass locally

• Relevant lint / type checks pass locally (ruff; ty and the full make ci not run)

• User-facing docs or screenshots are updated when needed (no screenshot attached; the row shows up in Settings → Providers)

Risk

• Security impact considered (no new code path; the key travels the same way as for the other OpenAI-driver gateways)
• Backward compatibility considered (additive; no existing provider changes)
• Rollback path is clear for risky changes (revert the commit)

Related Issues

Fixes #1180

[LivXue]

Windows compatibility review — PR #1181

Reviewed head 0641738be644a4fa354e048f0577d0e262315293 against base f401dfccd3cca42dd33d58dbeb0857ec44d447bb (15 files, +69/−10). The full diff (raven/providers/registry.py, config/schema.py, cli/onboard_commands.py, the pinned test baselines, and the ui-web provider mark) was read.

Findings

No Windows-specific issues found.

The change is a provider registry entry for LLMTR plus its config field, onboarding row, pinned test counts/baselines, docs licence note, and an SVG asset. It adds no OS-specific API, path handling, subprocess, installer, or packaging surface; default_api_base/native_api_bases are ordinary HTTPS URLs, and the provider keeps its model prefix (strip_model_prefix=False) by design.

Verification (native Windows 11, isolated worktree)

• Detached worktree D:/tmp/rv-1181 at the reviewed head; uv sync --frozen --offline --all-extras --no-extra segmenter --no-extra browser --dev succeeded.
• uv run --no-sync pytest tests/test_provider_catalog.py tests/test_litellm_provider_model_names.py tests/test_provider_protocol.py tests/test_rpc_model.py -q → 751 passed, 8 skipped (the skips are the pre-existing "LiteLLM does not name an environment variable" cases); the updated 57-provider and wire-baseline assertions pass.
• ui-web: ./node_modules/.bin/vitest run src/components/ProviderMark.test.tsx → 1 file, 15 tests passed (node_modules borrowed through a junction from D:/livxu/Documents/Raven/ui-web; package.json/package-lock.json byte-identical).
• Gap: no live LLMTR endpoint was called (no credentials/network); provider behavior is verified through the registry/baseline tests only.

No issues to report; no changes requested.