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.
Python agent-orchestrationai-agentai-coding-assistantclaude-codecoding-agentcomputer-usedesktop-applocal-firstmcppersonal-jarvisvoice-assistantvoice-coding
docs(screen-context): describe the boundary as it is now enforced
Harald900eeee910c7
diff --git a/docs/screen-context.md b/docs/screen-context.mdindex 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-runfollow-up window. Before the split, the bare noun "Bildschirm" was itselfsufficient 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 tohappen. 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 arewrapped 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 entrydiff --git a/jarvis/brain/cu_gate.py b/jarvis/brain/cu_gate.pyindex 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 itnames a surface and even inside a live desktop episode. Only an action verb, anexplicitly 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 tellingit to answer inline (or via ``search_web`` for fresh facts). Over-ALLOWING is