refactor(llm): complete() owns the request — and a test asserted the bug

[catomean]

This file already walked ai-kit's chain but kept its own fetch — so the chain came from the package and the request did not. That is where the remaining defects lived, and they are the fleet's most common ones (census 2026-09-06: 9 of 12 hand-rolled clients share them).

defect	what it did
content || ''	an empty 200 became the bot's reply, and the chain stopped there satisfied. A reasoning model that spends its budget thinking returns exactly that.
throw new Error('Groq API error: ' + status)	the body was read, logged, and discarded. Capacity, request-too-large and daily quota share that status code and want opposite responses: retry shortly, send less, or come back tomorrow.
no deadline	a vendor that accepted the connection and never answered held the chat open indefinitely — and the fallback beneath it was never reached.

It also removes the last decision this file was still making: the attempt callback branched on provider.id to pick a key. Keys now resolve through a synthetic env — the seam complete() reads — so a BYOK caller's key and the server's own take one path instead of two. Both per-vendor call functions and both model loops are gone.

Attribution is preserved, deliberately

The OpenRouter call sends HTTP-Referer and X-Title, which OpenRouter ranks apps by. complete() had no way to send them — so ai-kit grew extraHeaders (0.10.0) rather than this adoption quietly dropping Botsmann off that list. There is now a test naming it, because the loss produces no error, no log line and nothing to notice.

Three test defects, all surfaced by the conversion

One test asserted the bug. returns empty string when no content in response expected result.content === '' — it encoded the defect as the contract. Replaced by two tests: an empty 200 demotes to the next link, and an empty 200 at every link throws rather than answering with nothing.

The fetch mocks were { ok, json } literals with no text(). That quietly encoded an assumption about how the client reads a body, and broke the moment it read the text first (to keep the vendor's body in the error — the difference between "429" and "your daily budget is gone, resets in 4h"). Real Response objects now.

mockResolvedValue(new Response(...)) handed every link the same object, and a Response body can only be read once. Link two therefore always failed with Body has already been read — a fake failure standing in front of whatever the real behaviour was. Anything that walks a chain needs a factory, not a value.

The now-inert vi.mock('@/lib/constants') for endpoint URLs is removed: the client takes them from the same ai-kit provider record that supplies the model list, so mocking them asserted a copy nothing reads. Verified the resulting URLs are byte-identical to the constants they replace, so no endpoint moved.

Verified

check	result
verify	format + lint + typecheck + selfhost check + 277 tests + build, all green
net	−134 lines from llm-client.ts

🤖 Generated with Claude Code

https://claude.ai/code/session_01NHDLTKGirKU7He8q2YR2y4