truefoundry/trueforge is an open-source agent harness that presents itself as a runtime layer that turns LLMs into agents that actually do work. The commit c291315 of November 6, 2026 synced the model catalog file packages/trueforge/catalog/model-catalog.yaml with the separate repository truefoundry/models. The newly added release note file (changeset) records the addition of 24 models and a patch change for the @truefoundry/trueforge package. The overall change is 161 added lines and 1 deleted line.
This catalog is a reference table containing per-provider model identifiers and specifications (context length, maximum output token count, reasoning effort step values). Reasoning effort is a step value that determines how much reasoning a model may use before it settles on an answer, and the range from none to max differs from model to model. Among the newly added entries are gpt-6.1-terra with a context of 1050000 and a maximum output of 128000, claude-opus-4-6 and claude-opus-4-7 with a context of 1000000, and gemini-4-argon. On the Fireworks side, kimi-k3p1-pro, muse-spark-open, minimax-m3p1-flash, and deepseek-v4p1-flash-1017 were added, and on the Together side, the gemma-4 series, MiniMax-M3.1-Flash, Kimi-K3.1-Pro, and the NVIDIA Nemotron 3 series grew. A none level was added to reasoning effort for several Gemini models, and the context of Fireworks' deepseek-v4p1-flash was changed from 1000000 to 1048576.
That said, this table is only a record of the values declared by the upstream repository, not the result of performance validation. The same commit message also preserves the history of removing Fireworks' gpt-oss-120b, inkling, nemotron-3-ultra-nvfp4, and nemotron-lightning-3p5-30b-a3b from the catalog in response to review, and the reason given is that Fireworks will decommission these models on serverless on November 7, 2026 (changelog 2026-10-19). google-gemini's antigravity-preview-latest was also dropped, with the explanation that the model is not a chat model but an Antigravity agent, and that the Gemini adapter (@ai-sdk/google createGoogle) has no agent/Interactions path. In other words, this commit only tells us that the harness knows the model IDs and the specification values; it says nothing about how well those models actually work. There are no evaluation figures and no adoption cases in this commit.