The builder
What the agent can do
The agent's complete capability list. Each one is a real operation with real validation behind it, and each one shows up in the chat as a card.
There is no hidden capability and no general-purpose escape hatch. The agent can do exactly the things below, and every one of them is checked by code before it reaches your project.
Writes: things that change your project
These always appear as a visible card, because they have a consequence.
| Capability | Effect | What is checked |
|---|---|---|
| Build or edit the workflow | Adds, edits, removes, or replaces steps in Preview. | Step types must be real, loops must be paired, action titles must exist in your instruments’ catalogs, and a patch must stay inside the current edit scope. |
| Write a document | Replaces the document in Plans. | Body content only. It cannot inject its own stylesheet or scripts into the document. |
| Wire the schematic | Sets the DUT pins and the wires in Schematic. | Wires may only attach to configured instruments, and a pin must exist on the DUT. |
| Shape the report | Sets the Report’s groups, columns, units, limits, and charts. | Every column must be an exact header of the last run’s results file. A miss returns the real column list. Any spec carrying a measured value or a verdict is refused outright. |
| Propose a bench change | Shows a proposal card in chat. | Nothing is applied until you confirm it. Up to six changes per proposal, all or nothing. |
| Start a run | Opens Executor and starts the run immediately. | The workflow must have steps. One run per turn. It cannot start a verification run while holding a result that already has limit failures. |
| Flag an instrument problem | Posts an issue card about your bench. | Not used for instruments without a SCPI catalog, which are fully supported. |
Reads: things that gather evidence
These fold into one summary row in the chat, because a lookup has no consequence and evidence belongs collapsed.
| Capability | What it returns | Notes |
|---|---|---|
| Look up an instrument’s actions | The real action catalog for one configured instrument. | How it finds out what your specific model can actually do. |
| Search the instrument catalog | Matching instruments by name or type. | Used before proposing a bench change. |
| Check the run result | The outcome of the last run. | Reports “no outcome available” if the workflow changed since that run. |
| Read the run log | A budgeted slice of the log. | It is told, and tells you, that the slice is not the whole log. |
| Read a skill | A stored procedure for a specific instrument or technique. | Skills are curated know-how, not general web content. |
| Post a task list | Up to eight tasks, ticked off live. | Only on work long enough to need it. |
The one it is not allowed to use freely
There is a separate capability whose only job is to state that a test passed. It refuses unless the turn is holding a passing, current outcome, and its refusal hands the agent the honest sentence to use instead. This is why you will see “the run finished, two readings were outside the band” rather than a cheerful summary.