🔮Augur v1.13.0

Added

• VisibleProviders An API route waits out a gateway that preempts its request. When a local gateway such as Cogitator answers 503 preempted because a person’s request took the model, the call waits the Retry-After seconds and tries again up to three times, and a job that is still preempted goes to the route’s fallback instead of ending as an ordinary failure.
• VisibleProviders A new fal-api adapter makes videos on fal’s hosted queue. A route on it takes generate_video jobs, where the prompt describes the video, an attached image becomes the first frame, a second image the last frame, and an audio file the target audio. Before anything is sent, Augur multiplies the route’s price per second at the chosen resolution by the length and refuses a job whose estimate passes maxUsd, and the finished video is saved in the job folder. The key comes from the key store or a named environment variable, goes only to fal’s queue, and is never written to a job record; the route options are in docs/configuring.md.
• VisibleProviders A new reverse_engineering activity covers taking apart a shipped program, package, or web app, and each activity can bring its own tools and instructions. Reverse engineering counts as deep work in picks, like write_code and reason_critique, and a model is offered for it only when the rules give that model a weight for it. toolkits.json in ~/.augur ties an activity to a charter pack and to MCP servers: Codex and Copilot CLI jobs get the servers for that one job, and Grok jobs use a server registered once with grok mcp add. A job on a route that cannot carry the tools runs without them and records a warning; see docs/configuring.md.
• InternalProviders The Laya trainer can check a finished adapter for collapse. After training it checks whether the adapter’s logits and loss barely moved, and promote and set refuse an adapter it flags. Upstream’s label-tag positive control and an annealed rlcd loss are available as trainer options, and the old loss and the fit floor stay reachable through --legacy-temp and --calib-min.
• VisibleProviders Every route now reports what it can take. The route listings in augur routes, the routes call, and the MCP tools show which attachment kinds a route accepts, whether it reads images, embeds text, or transcribes audio, and its size limits. An embeddings route also shows its model and vector length, so a caller can rule out a route before it submits a job.
• VisibleProviders An embeddings route on the openai-api adapter can state its vector length with the new dimensions option. The value is reported beside the model in the route’s capabilities.
• VisiblePlatform A local app can run jobs over HTTP with POST /api/dispatch. Only connections from the same computer are accepted, and each call needs the web token. It runs as an agent, so it can submit and watch jobs and ask for a pick or the route list, but it cannot change rules.
• VisibleProviders An embed job takes a batch and answers with the vectors. A prompt that is a JSON array of strings becomes one input per string, with text files added after them, up to 256 inputs. The job’s answer is now one JSON object with the model, the vector length, and the vectors in input order, so augur run --wait prints them on standard output. The copy in the job folder stays.
• VisibleProviders A model limited to patch-only output can now run a text-only, read-only job on a plain API route. The rule asked for an isolated workspace copy on every job for such a model, so MiniMax could not take an image job over its API. A job that asks for text only with the read tier, on an adapter that holds the read tier itself, no longer needs one. A write tier or a patch request still does.

Fixed

• VisibleProviders The shipped balance rules for MiniMax now match the model under any label in its family. They were keyed to the label minimax/m3, so when the live model sync replaced MiniMax-M3 with MiniMax-M3.1-Flash, the balancer lost the model’s light tier, its long-context and summary preferences, and its second-opinion seat. The rules now use the pattern minimax/*m3*, which also covers the next version.
• InternalProviders The Laya server loads its checkpoints on the first request instead of at start. A restart no longer holds about 4.9 GB of video memory until a request, the busy, game, and idle unloads work as before, and the server ignores LAYA_PRELOAD and no longer sets it.
• InternalProviders The Laya server now uses the temperature fitted for each question type. It had stored the fit in the adapter’s manifest and never applied it, so confidence scores came from the base checkpoint’s temperature. The server reads temperature_by_type, restores the base values after each prediction, and reports them in its routing answer. The trainer also stops dividing logits by the base temperature, so the next adapters train on raw logits, and it refuses to fit a temperature from fewer than 50 items per type.