Driving Ru without a provider¶
Ru can run its whole pipeline with a scripted model instead of a provider. A run then executes real tools through the real Broker, in the real app, with no API key and no cost.
This exists because a suite can assert almost everything except what the user actually sees. An interface needs a live model to render a turn, so a defect that only shows up on screen — a panel that never draws, a row that disappears when a session is reopened — stays green in every test. The scripted model closes that gap by making the real app drivable.
Turning it on¶
Set RU_SCRIPTED_MODEL to a JSON file and start the app normally:
$env:RU_SCRIPTED_MODEL = "$env:TEMP\ru-scripted-model.json"
Nothing else is required, and unsetting the variable restores the normal provider path.
The runner fails closed. A path that does not exist, or a file that does not parse, stops the host at startup with an error naming the variable. It never falls back to a real provider, because that would quietly spend the user's credit on what was meant to be a fixture.
The script¶
A JSON array of steps. The model returns them in order, one per model step.
[
{ "text": "Em se lap ke hoach.", "tool": { "id": "todo_write", "arguments": { "todos": [
{ "content": "Doc file index.html", "status": "completed" },
{ "content": "Gop 3 file thanh mot", "status": "in_progress" },
{ "content": "Chay thu nghiem", "status": "pending" } ] } } },
{ "text": "Da lap ke hoach xong." }
]
| Field | Meaning |
|---|---|
text |
Assistant text for this step. Optional. |
tool |
A tool to call. Optional; its presence is what makes the step a tool step. |
tool.id |
Tool id, for example todo_write, list, read. Required inside tool. |
tool.version |
Tool version. Defaults to 1.0.0. |
tool.arguments |
Argument object. Defaults to {}. |
A step with a tool finishes as tool calls, so the runtime executes it and hands the result
to the next step. A step without one finishes the turn.
Running out of script ends the turn with a short message rather than failing the run, so a script one step short leaves a finished turn to inspect instead of a crash.
What it is good for¶
- Watching a feature render: activity rows, the task list, an approval prompt, a denied call.
- Reproducing a report by hand, then keeping the script as the regression fixture.
- Checking that state survives something the suite cannot simulate — leaving a session, restarting the app, reopening a conversation.
What it is not¶
It is a development seam, not a product feature. It replaces the model, so it cannot tell you anything about provider behaviour, streaming, token accounting or reasoning — those need a real endpoint. Do not ship a build that depends on it, and do not treat a scripted run as evidence that a provider path works.
Consecutive tool steps¶
The offline workflow runner registers one bounded executor for each of the eleven declared
scenarios: Vietnamese text, a ten-turn constraint sequence, restart continuation,
independent read batches, UTF-8 read paging, sparse grep paging, terminal
middle-output paging, dependent multi-file edits, edit recovery, failed-test repair, and
explicit denial. They invoke the actual CLI and managed Host with a temporary workspace,
use the real Broker, and check bounded event counts, receipt identities, and file oracles.
Approval-gated cases drive the existing explicit agent approve/deny commands and read
only public SSE history; the restart case restarts the managed Host against the
same data root. Paging scripts resolve $next_offset and $projection_spool_id
from the exact preceding tool call, and context oracles can require a marker to be present
or absent in the assembled model context. These predetermined steps verify wiring and
effects; they do not measure provider behavior or prove that a real model would find a
sentinel or follow a recovery note. A failed terminal command keeps a bounded spool_id
in its failure content so its retained output stays pageable without rerunning it.
Assertions about the prompt contract itself live in ScriptedModelReceiptTests, which
covers both capabilities renderings directly.