The gitlab-ai-provider npm integration does not expose a reasoning
passthrough (see providers/gitlab/provider.toml), so advertising
reasoning_options caused opencode to inject reasoning_effort into
/v1/chat/completions requests. GPT-5.6 rejects reasoning_effort (unless
'none') when function tools are present on chat completions, producing a
400. Match the other duo-chat-gpt-5-* entries with reasoning_options = [].
Inherit metadata from anthropic/claude-sonnet-5 instead of duplicating
it, matching the amazon-bedrock/fable pattern. Override only the
gitlab-specific name, required reasoning_options, cost = 0
(proxy-billed), and the lower output limit (64k) GitLab enforces.
Inherit metadata from anthropic/claude-fable-5 instead of duplicating
it, matching the amazon-bedrock fable definition. Override only the
gitlab-specific name, the required reasoning_options, and cost = 0
(proxy-billed).
Adds duo-chat-gpt-5-5 to the GitLab provider. The GitLab AI Gateway
proxies this model to OpenAI's gpt-5.5-2026-04-23 backend with a
1.05M token context window (922k input + 128k output).
Source: gitlab-org/modelops/applied-ml/code-suggestions/ai-assist
models.yml (gpt_5_5 entry with proxy_provider: openai).
The gitlab-ai-provider npm package exposes this model id starting in
v6.7.0.
The output limit was set to 4,096 tokens which caused tool calls with
large content (like file generation) to be truncated mid-JSON.
Updated to match standard Anthropic model limits:
- duo-chat-opus-4-5: 4,096 → 64,000
- duo-chat-sonnet-4-5: 4,096 → 64,000
- duo-chat-haiku-4-5: 4,096 → 64,000
Vladimir Glafirov