fix(nvidia): surface provider error detail, raise timeout, retire EOL models

[aliihsaad]

Ports one VPS-only commit upstream and fixes the three defects that made NVIDIA NIM the worst-performing provider in production.

Production evidence

Read-only audit of the Hostinger VPS (/opt/llm-hub, 4618 logged requests, 47.7% error rate). NVIDIA accounts for 216 failures — 215 are explained by the three fixes here:

Count	Error	Fix
182	This operation was aborted	timeout 15s → 60s
20	NVIDIA NIM API error 410: Gone	migration V21
13	NVIDIA NIM API error 404: Not Found	error-detail fix will finally explain it
1	503	transient

1. Error bodies were being discarded

NVIDIA answers RFC 7807 problem+json ({type,title,status,detail}) with no error envelope, but all six call sites read only err.error?.message and fell through to res.statusText. A fully explained retirement surfaced as the useless NVIDIA NIM API error 410: Gone — the body actually said:

The model 'minimaxai/minimax-m2.7' has reached its end of life on 2026-07-27T00:00:00Z and is no longer available.

readProviderErrorText() now checks error.message → title/detail → message → title → statusText.

Probing every provider showed three more were silently affected: Cerebras ({message}), Mistral ({detail}), HuggingFace ({error:"string"}). Google (error.message) and Cloudflare (errors[0].message) were already correct and are untouched.

2. No timeout override

NVIDIA hosts almost exclusively 70B–550B models on a non-latency-guaranteed free-credit pool, yet ran on the 15s default while nine slower-tier providers already had 60s. Worse, classifyProviderError maps an abort to a 120s key cooldown, so one slow request benched the entire provider.

3. Four rows end-of-lifed upstream (migration V21)

Live-probed 2026-08-23. NVIDIA resolves the model before authenticating, so an unauthenticated POST cleanly separates 401/403 (exists) from 404 (never existed) and 410 (retired) — verified against a bogus-model control.

Retired: llama-4-maverick, deepseek-v4-pro, mistral-large-3, minimax-m2.7. Disabled not deleted, per V18/V19 precedent. Eight probe-confirmed replacements added, including deepseek-v4-flash-0731, minimax-m3, nemotron-3-ultra-550b-a55b and kimi-k3.

Also: ports a VPS-only commit

fix(routing): respect tool schema context size existed only on the production VPS, never pushed (original authorship preserved). It also introduced its own migrateModelsV20, which is exactly why the NVIDIA migration here is numbered V21 — porting it first surfaced the collision.

Verification

• Tests written first and confirmed failing against old behaviour — the 410 case reproduced the exact production string NVIDIA NIM API error 410: Gone
• 257/257 server tests pass, tsc clean
• dist rebuilt in a separate commit (repo ships it via package.json files)

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: ca2edce0-7fd4-4858-8bed-786c321c185c

---

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.}