fix(provider): send Anthropic's dashed native slug through the AI Gateway

[superhighfives]

Issue for this PR

Closes #44280

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Every Anthropic model under cloudflare-ai-gateway fails with Error: model: claude-haiku-4.5, so none of them are usable.

The plugin already routes anthropic/* models through Anthropic's native Messages passthrough (createAnthropic), which forwards the sliced model id to Anthropic's API unchanged. models.dev lists Cloudflare's canonical dotted ids (anthropic/claude-haiku-4.5), but Anthropic's Messages API expects dashed native slugs (claude-haiku-4-5) — so it 404s.

The fix translates dots to dashes when slicing the anthropic/ prefix. OpenAI is unaffected because its native slugs already use dots (gpt-4.1), which is why only Anthropic was broken.

Worth noting: Cloudflare's own catalog-aware REST endpoints (/ai/v1/messages, /ai/run) accept the dotted id and do this exact dot→dash translation internally before calling Anthropic. The native passthrough route this plugin uses forwards the slug raw, so the plugin has to do the same translation. The dotted id in models.dev is correct — it's Cloudflare's canonical id (see https://developers.cloudflare.com/ai/models/anthropic/claude-haiku-4.5/), so this is fixed on the opencode side, not in the catalog. (Context: surfaced while testing anomalyco/models.dev#4922.)

How did you verify your code works?

• Built the single binary and ran against a live Cloudflare AI Gateway: anthropic/claude-haiku-4.5 went from Error: model: claude-haiku-4.5 to a normal completion; openai/* and workers-ai/@cf/* still work.
• Added a regression test in cf-ai-gateway-e2e.test.ts asserting a dotted Anthropic id reaches the Anthropic Messages wire as a dashed slug; confirmed it fails without the fix and passes with it.
• bun test test/provider/ in packages/opencode passes (one unrelated pre-existing flake in header-timeout.test.ts, which also fails on dev). tsgo --noEmit clean.

Screenshots / recordings

N/A (not a UI change).

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

Correct diagnosis and a tidy fix: models.dev lists Anthropic ids dotted (claude-haiku-4.5) while the Messages passthrough forwards them verbatim, so every anthropic/* model under the gateway was doomed to a 404. The new e2e assertion pins both the endpoint (v1/messages) and the translated upstream slug (claude-haiku-4-5) — exactly the regression guard this needs.

Two suggestions:

1. Extract the shared translation. The .replaceAll(".", "-") logic now exists verbatim in two places: the production loader (packages/opencode/src/provider/provider.ts, ~line 850) and the test harness's gatewayModel (cf-ai-gateway-e2e.test.ts, ~line 168). Mirrored helpers like this drift silently — if the rule ever changes (say, only version-position dots), the tests will keep validating the old behavior. Exporting something like anthropicNativeSlug(id) from the provider module and using it in both spots keeps the test honest.

2. Document why replaceAll is safe. It works today because no native Anthropic slug contains a dot, so blanket replacement can't corrupt a legitimate id. The code comment explains the dotted→dashed direction well; half a sentence noting the inverse invariant ("no native slug contains '.', so global replacement is lossless here") would preempt someone generalizing this pattern to prefixes where dots are meaningful — e.g., the workers-ai/@cf/moonshotai/kimi-k2.6 family, where dotted ids are apparently accepted as-is on the unified route.

Minor: worth a quick sanity pass over the other first-party branches in custom() to confirm none of them have the same catalog-vs-native mismatch — this one only surfaced because Anthropic is strict about unknown model names.

[superhighfives]

Passed this back to the robot and got the below, which added a comment and rationalized the other two. Happy to discuss!

---

Thanks for the review (re: #44281 (comment)).

2. Documented the safety invariant — done in 70ae461. Added a note that no native Anthropic slug contains a dot, so the blanket replacement is lossless, and contrasted it with the OpenAI branch above (whose native ids like gpt-4.1 keep their dots and must not be touched).

3. Sanity pass over the other custom() branches — checked; no other branch has the mismatch. openai/* is the only other native-passthrough route and it deliberately does not translate, because OpenAI's real slugs legitimately contain dots (gpt-4.1). Everything else rides the unified route, which accepts dotted ids as-is (e.g. kimi-k2.6). The dotted/dashed mismatch is Anthropic-only, which is why the fix is scoped to that branch.

1. Extracting a shared anthropicNativeSlug() — not changing this. The test's gatewayModel is a deliberate independent mirror of the full routing (all three branches), not just the translation line — it reconstructs routing so the test can inspect the wire. The assertion checks the resulting slug against a hardcoded literal (claude-haiku-4-5), so it catches drift whether or not it imports the prod function. Extracting just one of three mirrored branches into a new prod export is inconsistent and widens the module surface for marginal gain, so I'd rather keep the change minimal.