GitHub's device flow documents slow_down as a rate limit: a poll that
arrives inside the throttle window is answered with slow_down instead of
the token, and the response's interval field reports the new required
minimum ("adds 5 seconds to the last interval"). A client that only
tracks its own +5s increment can stay behind the server's ratcheting
requirement when its timers fire early - common with WSL/VM clock drift
(microsoft/WSL#10006) - so every subsequent poll keeps hitting the rate
limit and login appears to hang forever even after the browser reports
the device as authorized.
Adopt the server-provided interval when a slow_down poll result carries
one, falling back to the RFC 8628 section 3.5 +5s increment otherwise.
GitHub Copilot passes the interval field through.
This restores part of the #1994 mitigations that were lost in the #4788
device-code refactor.
refs #6187
Picked up upstream models.dev/OpenCode catalog drift while
regenerating for the Fireworks GLM 5.2 Fast fix: Cloudflare AI Gateway
and OpenCode gain Claude Sonnet 5; OpenCode gains Kimi K2.7 Code and
MiniMax-M3; MiniMax model name formatting normalized to MiniMax-Mx.x.
Broaden the Fireworks GLM 5.2 special-case in generate-models.ts from
an exact id match to a glm-5p2 substring match, so the router variant
accounts/fireworks/routers/glm-5p2-fast also uses the OpenAI-compatible
endpoint and thinkingLevelMap instead of falling back to the default
Anthropic Messages config.
closes#6195
The Codex provider previously loaded node:os asynchronously during module evaluation. A fresh process that imports the provider and immediately starts an SSE request can build headers before that promise callback runs, so the first request reports User-Agent: pi (browser) in Node/Bun.
Reproduced against parent fd6659dd with a stubbed-fetch harness: 50/50 immediate first requests reported pi (browser) in both Node and Bun. The same harness on this fix reports the OS-specific user agent 50/50 in both runtimes.
Use process.getBuiltinModule("node:os") behind the existing Node/Bun runtime guard so OS metadata is available synchronously while still avoiding top-level runtime Node builtin imports that break browser/Vite builds.
Add a shared normalizeProviderError helper in packages/ai/src/utils/error-body.ts
and route the 8 body-blind / status-only providers through it (amazon-bedrock,
azure-openai-responses, google, google-vertex, images/openrouter,
openai-codex-responses, openai-completions, openai-responses). Non-schema 4xx/5xx
responses from proxies / gateways now show the real reason carried in the response
body alongside the HTTP status, instead of "403 status code (no body)" or
"Unknown: UnknownError".
The helper probes status (statusCode, status, $metadata.httpStatusCode,
$response.statusCode) and body (body, parsed error object, $response.body) across
the Mistral, OpenAI, Google, and Bedrock SDK shapes, truncates the body at a 4000
char cap, and preserves error.message when the SDK already folded the body in
(Anthropic / Google happy path). mistral.ts and anthropic.ts are left untouched.
Provider prefixes and the OpenRouter metadata.raw append are preserved.
closes#5763
Add optional reasoning?: number to Usage as a subset of output. Populate
for Anthropic (output_tokens_details.thinking_tokens), OpenAI
Responses/Codex/Azure (output_tokens_details.reasoning_tokens), OpenAI
Completions (completion_tokens_details.reasoning_tokens), and Google
Generative AI / Vertex (thoughtsTokenCount). Bedrock Converse and Mistral
do not return a reasoning breakdown, so they stay unset.
closes#6057
Armin Ronacher