🔮Augur v0.7.0

Added

• VisibleProviders GitHub Copilot shows its credits used this month. Augur reads the count through the GitHub CLI’s sign-in (gh auth login) and compares it with a monthly spending cap you enter in settings, since GitHub does not report a cap. The card shows the credits used and when the month resets. Turn it on in settings; it is off until then.
• VisibleUI A Model rules page sets what agents may use each model for. Each provider and model lists its allowed activities, how often to pick it for each one, the most sensitive data it may see, whether it runs only when named, and an optional pause. Augur saves the rules to policy.json beside the usage file whenever one changes. A new model stays blocked until its rules are confirmed, and the dashboard shows how many are waiting. Every change is listed in History and can be undone. If a policy.json is already in place when the page first runs, Augur adopts its rules, pauses and confirmations instead of importing the starting set.
• VisibleProviders Model rules lists each provider’s current models. Once a day Claude, Codex, Grok and MiniMax add their newest models for review, and OpenRouter and fal keep theirs as a searchable list to add from. Only the newest version of each model is listed. A newer version starts with the older one’s rules, and confirming it hides the older one.
• InternalPlatform A local dispatch service and CLI run jobs under the model rules. augurd accepts a job, checks it against policy.json and the usage meters, and runs the harness in its own supervised process. A job outlives its caller, can be cancelled with its whole process tree, and is picked up again after the service restarts. The augur command starts, watches, cancels and reads jobs, and augur apply applies the patch a sandboxed job returns. Adapters cover Codex, Hermes, Copilot, Grok, MiniMax Code and OpenCode (the last two inside Docker Sandboxes, on a copy of the working folder), plus OpenAI-style and Anthropic-style APIs. The app does not use the service yet.
• InternalPlatform The dispatch service enforces the model rules with live usage. A job is refused when its provider’s session or weekly meter is over its limit, when a pay-as-you-go balance is under its minimum, or when the model is burning its window far faster than the window allows. augur pick ranks the models a task may use from the rules and current usage, and the service can require that a job’s model was picked for its caller within the last hour. A caller can skip the pick check or the usage check for one job, and each skip is recorded with the job.
• InternalPlatform The dispatch service checks a claim that a person named a model. A caller that says so is confirmed against a message the service was told a person sent (augur note-prompt keeps only which models it named, never the text), or against a terminal a person is typing at. The setting verifyNamed records an unconfirmed claim by default, and can refuse the job instead or turn the check off.
• InternalPlatform The dispatch service can classify a task and score model fit with Laya. Laya is an open System One model that runs on this computer, with a second machine as failover. A server that reports more than 80% GPU or video memory use steps aside, and the next server on the list answers. Jev with your own key, or no classifier at all, can be chosen instead, and a second backend can be run beside the first only to compare answers.
• InternalPlatform A Laya package serves the model with LoRA adapters, trains new adapters from labelled decisions, and copies adapters between machines. Rows carry a provenance so an adapter can be built without any Jev-labelled rows, and training can use the published loss or plain cross-entropy.
• InternalPlatform The dispatch service can keep a record of each pick and how the job that followed it ended, and a nightly run turns that record into new Laya adapters. Task text is written only when learning is switched on and the task’s data tier is public or internal. A new adapter goes live only when it beats the current one on rows held out from training, and it is then copied to the second machine.