Summary
Add the ability to route requests directly to a specific model and reasoning level (e.g. openai/gpt-5.5-low) instead of relying solely on dynamic task-based routing via the manifest/auto model. This would make Manifest a fully OpenAI-compatible API.
Motivation
I understand Manifest's intent is to dynamically route based on task using the manifest/auto model, and that's worked well for nearly every use case I've thrown at it.
However, I'm trying to use Manifest from my self-hosted environment in Warp as a custom endpoint, and Warp's harness errors out with an invalid JSON error. As far as I can tell, it doesn't know how to handle the way responses are formatted. Warp expects a standard OpenAI-compatible response from a specifically named model.
If I could route directly to a model and specify the reasoning level (e.g. gpt-5.5 at a given reasoning level), it stands to reason this would let clients like Warp work against Manifest as a drop-in OpenAI endpoint.
Proposed changes
- Support passing an explicit
model value that maps directly to a specific provider model and reasoning level (e.g. openai/gpt-5.5-low), bypassing dynamic routing.
- Add a
/v1/models endpoint so OpenAI-compatible clients can discover available models.
Together these would make Manifest a fully OpenAI-compatible API.
Additional context
Originally posted on Discord here: https://discord.com/channels/1089907785178812499/1516615058237423737/1516615058237423737
Summary
Add the ability to route requests directly to a specific model and reasoning level (e.g.
openai/gpt-5.5-low) instead of relying solely on dynamic task-based routing via themanifest/automodel. This would make Manifest a fully OpenAI-compatible API.Motivation
I understand Manifest's intent is to dynamically route based on task using the
manifest/automodel, and that's worked well for nearly every use case I've thrown at it.However, I'm trying to use Manifest from my self-hosted environment in Warp as a custom endpoint, and Warp's harness errors out with an invalid JSON error. As far as I can tell, it doesn't know how to handle the way responses are formatted. Warp expects a standard OpenAI-compatible response from a specifically named model.
If I could route directly to a model and specify the reasoning level (e.g.
gpt-5.5at a given reasoning level), it stands to reason this would let clients like Warp work against Manifest as a drop-in OpenAI endpoint.Proposed changes
modelvalue that maps directly to a specific provider model and reasoning level (e.g.openai/gpt-5.5-low), bypassing dynamic routing./v1/modelsendpoint so OpenAI-compatible clients can discover available models.Together these would make Manifest a fully OpenAI-compatible API.
Additional context
Originally posted on Discord here: https://discord.com/channels/1089907785178812499/1516615058237423737/1516615058237423737