Skip to content

Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.

docs(screen-context): describe the boundary as it is now enforced

Harald900eeee910c7
diff --git a/docs/screen-context.md b/docs/screen-context.md
index 3c09593773..bcc0e2ca15 100644
--- a/docs/screen-context.md
+++ b/docs/screen-context.md
@@ -254,8 +254,16 @@ two paths are enforced apart in three places, and all three are load-bearing:
mission even when it names a surface, and even inside the recent-run
follow-up window. Before the split, the bare noun "Bildschirm" was itself
sufficient authority — which is how a question moved the user's mouse.
-3. **`BrainManager.generate`** — a successful capture strips every tool from
- the turn, so the two paths can never both run for one utterance.
+3. **`BrainManager.generate`** — a successful capture narrows the turn's tool
+ surface: the spawn vehicles and the deterministic record writes go, the
+ read-only tools and `computer_use` stay. Until 2026-08-17 it stripped
+ *every* tool, which enforced the mandate by making a mixed utterance
+ impossible to answer at all — "look here, close that window" produced a
+ description of the screen and no action. Point 2 is what actually keeps a
+ look request off the desktop, and it does so on the user's words, not on
+ the surface, so it holds for a pure look even inside a live episode. What
+ the narrowed surface adds is that the ACTION half of a mixed utterance now
+ has a vehicle. A question is still never answered by driving the desktop.
Two consequences worth stating, because both were regressions waiting to
happen. A screenshot request ("mach mal einen Screenshot") is the *least*
@@ -356,9 +364,11 @@ Five layers, in the order they run:
Screen pixels, OCR, and accessibility text are untrusted evidence. They are
wrapped in fixed `SCREEN_EVIDENCE` delimiters with an instruction boundary, and
-the production brain exposes no tools on a captured look turn. Text rendered by
-a webpage or document therefore cannot turn a read-only inspection into an
-action or tool call.
+the production brain hides the unattended vehicles on a captured turn — the
+spawn tools and every deterministic record write. Text rendered by a webpage or
+document therefore cannot start a background agent or write a contact, calendar
+or wiki entry. It cannot unlock the desktop either: `llm_computer_use_allowed`
+reads the USER's utterance, never the model's reasoning and never the pixels.
Default patterns ship for card numbers, IBANs, API-key shapes, and
`Authorization:`-style headers. They are configurable, additive, and each entry
diff --git a/jarvis/brain/cu_gate.py b/jarvis/brain/cu_gate.py
index 92817fd982..fc963fb4c2 100644
--- a/jarvis/brain/cu_gate.py
+++ b/jarvis/brain/cu_gate.py
@@ -34,9 +34,14 @@ into ACTION verbs and SURFACE nouns, and a turn that
``jarvis.screen_context.intent`` reads as a look request is refused even when it
names a surface and even inside a live desktop episode. Only an action verb, an
explicitly named Computer-Use, or a surface noun in a turn that is NOT a look
-request may still drive the desktop. This gate is the enforcement half; the
-other half is that a successful Screen Context capture strips every tool from
-the turn (``BrainManager.generate``), so the two paths can never both run.
+request may still drive the desktop. This gate is where the boundary is
+actually enforced: it decides on the USER's words, so a look request is refused
+even inside a live episode. A successful Screen Context capture additionally
+narrows the turn's surface (``BrainManager.generate``) — the spawn vehicles and
+the deterministic record writes go, the read tools and ``computer_use`` stay.
+Until 2026-08-17 that capture stripped EVERY tool, which also left the action
+half of a mixed utterance ("look here, close that window") with no vehicle at
+all; the narrowing keeps the boundary and gives that half its vehicle back.
Everything else is blocked and fed back to the model as a tool error telling
it to answer inline (or via ``search_web`` for fresh facts). Over-ALLOWING is