Auto-Terminal separates the command you can see from the trust and policy that allow it to run. Run executes the current work, Run All drains the eligible local queue plus distinct typed input, Always Run stores bounded command trust, and Run Everything broadens eligible automation without bypassing safety or foreground gates.
Distinguish Run from Run All
Run is the focused action for the current Auto-Terminal command. Run All is a queue action: it runs the eligible work for the current repository scope and can include a distinct command you typed into the input when that typed command is not merely the visible active suggestion repeated a second time.
Run All does not turn every text fragment into a shell command. GDP keeps command parsing, repository/cwd identity, duplicate suppression, and execution scheduling explicit.
The visible active suggestion is not re-added as duplicate typed input.
Distinct typed input can still participate in Run All.
Repository and cwd remain part of execution identity.
A failed command remains failure evidence even if later queue work continues.
Use Always Run as bounded semantic trust
Always Run is explicit local trust for matching future commands. GDP prefers a semantic operation identity when the command analyzer can derive one, and otherwise falls back to a bounded executable/command signature rather than treating arbitrary argument changes as automatically trusted.
Trust is intentionally narrow. Allowing a read-only Git status operation does not authorize destructive Git reset, and trusting one test command does not grant blanket shell access to unrelated commands.
Review the displayed command and trust scope before storing Always Run. Convenience should remove repeated prompts for the same bounded operation, not erase command intent.
Understand what Run Everything changes
Run Everything lets GDP continue eligible queued terminal and Suggested Check work under the selected local profile. It changes automation behavior, not the underlying safety model.
Command validation, repository confinement, cwd rules, risk classification, timeouts, cancellation, and evidence recording still apply. Network-required, write-declared, unknown, higher-risk, or otherwise specially gated work can remain outside headless Auto-Run policy.
A Suggested Check submitted by an AI is still only a request until Desktop runs it.
Run Everything does not make the hosted gateway a shell.
A command can be visible and queued while policy still requires local action.
Exact run ID, exit code, stdout/stderr, and current repository state remain the completion evidence.
Keep foreground work user-controlled
Checks that require native UI, focus, or another foreground-only capability stay user-controlled even under broader automation. GDP can leave them ready for explicit Run or Run All instead of repeatedly stealing focus while the user is away or working in another surface.
If a command does not run automatically, inspect its policy/risk/foreground reason before assuming Auto-Terminal is stuck. Manually running an intentionally gated check can be the correct state transition.
Automation should reduce unnecessary interaction without converting focus-sensitive or high-impact work into invisible background execution.
