[Felixkw12]

Summary

Adds Opper as a gateway provider. Opper is an EU-hosted AI gateway that serves models from many providers behind one OpenAI-compatible API.

The entry sits at the end of the resale gateway block in raven/providers/registry.py and copies that block's shape: reached through OpenAI's driver at https://api.opper.ai/v3/compat, told apart by that address, OPENAI_API_KEY as the driver's variable, no default model and no shortlist. The gateway's own /v3/compat/models fills the picker once a key is entered. A bare model id reaches the pool of providers that serve that model and a provider/model id pins one provider (https://docs.opper.ai/control-plane/rules/routing.md), so strip_model_prefix stays off.

Also included: a ProvidersConfig section, a row in the onboarding wizard's key group, and a brand mark for the web UI. Neither bundled icon set carries an Opper mark, so the SVG comes from us, with a LICENSES/README.md entry beside AMD's. The registry totals written elsewhere (ui-web/CONTEXT.md, two comments on the settings provider page, a docstring in raven/providers/registry.py) now count fifty-six specs, twenty-two of them gateways.

Keys: https://platform.opper.ai
Docs: https://docs.opper.ai

Disclosure: Opper is my startup.

Type

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

Verification

Run by a maintainer on this branch, rebased onto current main:

• make check-commits, make check-large-files, make check-source-language: pass
• PR_TITLE="feat(providers): add Opper as an OpenAI-compatible gateway (#869)" make check-pr-title: pass
• make lint-python, make lint-imports, make lint-deps, make lint-types: pass
• uv run --frozen --python 3.12 --all-extras pytest -q -n 2 (macOS, a loaded machine): 34294 passed, 182 skipped, 2 xfailed, 4 failed, 1 error. Three of the failures also fail on main without this change on the same machine (a case-insensitive APFS volume, a local LibreOffice install, a process-stop timing test). The fourth, in tests/test_node_stderr_replay.py, timed out under load, and the error is its teardown finding the agent it left; run alone it passes three times out of three here and passes on main. None is in provider code.
• npm run gen:check, npm run lint, npm run type-check in ui-web: pass
• npm test --prefix ui-web: 253 files, 4215 tests passed
• npm run --prefix ui-web build, then ui-web/build.py: the page builds and both boot snapshots match their goldens
• The mark rendered in headless Chromium at 20px and 40px, in the light theme and inverted for the dark theme: legible in both

Not run: a live call through Raven against the gateway; no Opper key was at hand.

• Relevant tests pass locally
• Relevant lint / type checks pass locally
• User-facing docs or screenshots are updated when needed (no user-facing doc lists providers one by one)

Risk

• Security impact considered
• Backward compatibility considered
• Rollback path is clear for risky changes

Additive: one new provider section, nothing existing changes. Rollback is reverting the commit.

Related Issues

N/A

[LivXue]

We are closing this for now and will revisit Opper integration once LiteLLM has official native Opper support. Opper can currently be used through Raven's custom OpenAI-compatible provider configuration.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Technical review is clean. I checked the AGENTS.md and context vocabulary rules, the full diff, provider resolution/auth/protocol/catalog callers, the branch history and backward compatibility, test changes, and the provider architecture. Opper's current official documentation confirms https://api.opper.ai/v3/compat and provider-prefixed model IDs; the locked LiteLLM 1.85.0 has no native Opper route, so the borrowed OpenAI driver is internally consistent.

Verification:

• uv run --all-extras pytest tests/test_litellm_provider_model_names.py tests/test_provider_catalog.py tests/test_rpc_model.py tests/test_provider_wire_model.py tests/test_provider_resolution_invariants.py tests/test_cli_onboard_commands.py -q: 1686 passed, 104 skipped. An initial run without --all-extras produced 1443 setup errors because raven_everos was absent; the repository-correct rerun passed.
• npm test --prefix ui-web -- ProviderMark.test.tsx: 15 passed.
• npm run type-check --prefix ui-web: passed.
• Source-language and large-file scripts against github/main..HEAD: passed. (make is unavailable in this environment, so I ran the Makefile's underlying commands directly.)
• git diff --check github/main...HEAD: passed.

The maintainer's stated decision to close and defer this integration until LiteLLM has native support remains controlling as a scope decision; this review does not contest it.

[0xKT]

Not a blocker (review of 5dd2740b, main 43159933): the new row routes as its neighbours do. Across the stock catalogues' model ids and a set of configs (OpenAI key only, an OpenAI base, a custom base with opper.ai in its path, an Opper section with and without a key), nothing that went elsewhere before now reaches api.opper.ai, and with the network blocked no OpenAI key was sent there. One small thing: the counts that describe the registry are now one short. ui-web/CONTEXT.md:332 still says the gateways are "Twenty-one of the fifty-five" (the registry now has 56 specs, 22 of them gateways), and "fifty-five" is still written in ui-web/src/features/settings/providers/ProviderSide.tsx:8, Roles.tsx:317 and raven/providers/registry.py:1533.

Severity of this one finding, not a verdict on the pull request. Any blocking findings from this pass are review threads on the changed files; GitHub renders those collapsed, as a file name with no text.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

The rebased revision preserves the previously reviewed implementation, removes the redundant schema comment, and correctly updates the four live totals to 56 providers and 22 gateways. I checked the new delta, the resulting full PR diff, current context terminology, commit history, callers, compatibility, tests, and provider architecture; I found no new defect or weakened test. The older 55-provider references still present under dated design/acceptance specs are historical snapshots rather than live totals.

Verification:

• uv run --extra dev pytest tests/test_provider_catalog.py tests/test_litellm_provider_model_names.py tests/test_provider_wire_model.py -q: 712 passed, 7 skipped.
• npm test --prefix ui-web -- ProviderMark.test.tsx: 15 passed.
• npm run type-check --prefix ui-web: passed.
• Source-language check and git diff --check github/main...HEAD: passed.
• The repository's --all-extras setup could not build datrie because this environment lacks x86_64-linux-gnu-gcc; it stopped before test collection.

[LivXue]

Reopening. We closed this to wait for native Opper support in LiteLLM, but that is not the bar the registry already applies: the thirteen resale gateways and the six CN vendors in raven/providers/registry.py all reach their endpoints through LiteLLM's OpenAI driver, because LiteLLM routes none of them. Opper qualifies on the same basis.

Since maintainer edits are allowed on this branch, we finished it here instead of sending it back:

• Rebased your commit onto current main, unchanged and still authored by you. CI never actually ran on the old head (every run ended with zero jobs), so this push gives the first real CI run.
• Dropped the trailing # Opper gateway comment in raven/config/schema.py: it restates the field name, and AGENTS.md section 1.1 keeps comments to what the code cannot say.
• Updated the registry counts @0xKT pointed out: ui-web/CONTEXT.md now says twenty-two of the fifty-six, and ProviderSide.tsx, Roles.tsx and the _tables docstring in registry.py say fifty-six.
• Updated the description, which becomes the squash commit message: the Verification section now lists what we ran, the model-id sentence links to your routing docs, and the summary no longer quotes model and provider counts.

One thing we could not do is a live call through Raven against the gateway, since we have no Opper key. If you can, please run one (fetch the model list, send one chat completion) and post the result here.

Heads-up: #1009 (Atlas Cloud) adds a gateway to the same registry block and changes the same counts, so whichever of the two lands second rebases and updates them: 57 providers, fifteen resale gateways, and the fifty-six and twenty-two written in the four places above.