Three ways to check an answer
Evidence first (evidence-first, default): ask the model to separate facts supported by supplied sources from assumptions and missing information. Use it for unfamiliar code, an investigation or a factual question. Example: before changing an API call, identify the current contract and the failing request. It cannot create evidence that was not supplied.
Plan and check (plan-check): ask for a short plan, complete the work, then compare the result with the original request and available evidence. Use it for a change with several dependent steps. Example: add a form, validate fields and check the saved result. It does not add another model or turn automatic editing on.
Review the answer (self-review; setting ID: peer-review): ask for a draft, challenge its assumptions and edge cases, then correct it. This is self-review by the same model unless a separate reviewer is explicitly assigned. Use it to examine a proposed fix; it is not an independent jury and does not guarantee correctness or lower token use.
Chat mode is not a server team
Single, Auto and Council describe the extension chat workflow. Auto delegates routine development steps within the selected permissions and limits. Council asks for a bounded review; it is not the switch that installs persistent terminal workers. Auto Apply and command execution permissions remain separate settings. Read the run plan before relying on any mode to execute changes.
Service and task presets
Standard is the ordinary text/chat service identifier, not a quality certification or a premium tier. Coding is a separate server workflow with its own project snapshot and quote. A client must offer only workflows it can actually submit. If there is only one available service, there is no useful choice to make.
A pack is a reusable instruction and checklist, not extra model memory. Current IDs are engineer (code changes), osint (open-source research), trading (analysis with paper/live boundaries), gamedev (games), marketer (marketing) and optimizer (optimization). Leave Automatic when unsure. A pack can supply a model default, but an explicitly selected model takes precedence.
Your computer, your SSH host and private state
The extension runs on its selected host. In a Remote SSH workspace that is the remote Linux server; otherwise it is the local extension host. The orchestration state directory lives on that host, not automatically in Tvijo cloud storage. Permission 0700 lets only its operating-system owner access the directory. It is not encryption or a backup.
Python and controller paths are advanced overrides. Leave them empty for discovery from the installed host manifest. A path is not an installer: if the runtime or a subscription is absent, setup must report that prerequisite. Changing apiBase chooses another service endpoint; do it only when connecting a known self-hosted service.
Limits protect a task, not its success
The local Agent defaults to a USD2 task cap and 60 steps. A step is one iteration of work, not one completed feature. Near the spending cap, the extension can offer an explicit increase of USD2, USD5 or USD10. Declining does not reset usage. A larger context collection can help with larger projects, but the service still applies its request and model limits.