The Ghost Vocabulary

In plain English: Ghosthread's core engine. It performs actions inside your apps without moving your cursor, without stealing your focus, and without bringing any window to the front. You can keep typing while it works.

Under the hood: Background input injection via Win32 PostMessageW into resolved child input sinks (RenderWidgetHostHWND for Chromium apps, WinUIDesktopWin32WindowClass for modern apps, RichEdit controls for editors). No SendInput. No hardware hooks. No foreground activation.

In plain English: A future mode where Ghosthread stops waiting for you to ask. It watches what you're doing, understands the context, and speaks up only when it matters — suggesting fixes, catching mistakes, and pointing out things you might have missed. It never executes anything without your say-so.

Under the hood: Continuous perception via WinEvent hooks and UIA subscriptions, with strict attention gating — flow-state detection, interruption budget, do-not-disturb signals, and confidence thresholds. Suggestion-only. Zero autonomous execution.

→ Full True Phantom breakdown

In plain English: The specific channel Ghosthread uses to talk to an app. Every app has a tiny "inbox" where it receives input — Ghosthread finds the right inbox for each app and drops commands there directly.

Under the hood: Decoupled child input sink resolution. The engine recursively inspects a window's hierarchy to identify the true active input surface — RenderWidgetHostHWND for Chromium, WinUIDesktopWin32WindowClass for WinUI, RichEditD2DPT for modern editors. Input posts to that sink, never the top-level window.

In plain English: A safety check that runs before Ghosthread types anything into any window. It verifies the target window is still where it's supposed to be — if not, it aborts rather than risk sending keystrokes somewhere wrong.

Under the hood: Foreground window ownership assertion executed immediately before any hardware pulse. Any mismatch between expected and active foreground HWND aborts the pulse entirely and logs the event. Prevents keystroke leakage into other applications when focus shifts mid-execution.

In plain English: Highlight anything on your screen — a word, a code snippet, an error — press a hotkey, and get an answer right there next to it. No chat window opening, no app switching, no context to explain.

Under the hood: Ephemeral region-anchored overlay triggered by global hotkey and drag-marquee selection. Content extraction via UIA node at cursor, falling back to Windows.Media.Ocr, then vision. Non-activating topmost overlay with capture exclusion. Dismiss with Esc. Nothing persists unless pinned.

In plain English: Record a complex task once, and Ghosthread turns it into a replayable recipe. Run it again anytime — on any machine — and it adapts to whatever changed.

Under the hood: Semantic session capture filtered from the perception stream. Actions stored as typed nodes (verb + target descriptor + payload + verification condition), not brittle coordinates. Replay routes through the existing DAG executor and applies the same policy gates as live execution. Secrets redacted before export.

In plain English: How Ghosthread breaks a request into ordered steps. Each step points to the steps that depend on it, forming a flowchart that flows one direction and never loops.

Under the hood: Kahn's topological sort validates cycle-free execution before any node runs. Each node declares typed input and output artifacts, verified against upstream producers at compile time. Independent branches execute in parallel across isolated IO, CPU, System, and GUI lanes.

In plain English: How one step passes its result to the next. Each step says "I produce a file path" or "I produce a block of text," and the next step can only accept a compatible result.

Under the hood: Dual-channel SQLite-backed blackboard. Machine-facing typed artifacts enforced against declared slot contracts at commit time; human-facing display text stored separately. Type mismatches raise at graph validation, not at runtime.

In plain English: A fixed set of safety rules that runs on every action, every message, every command. The AI can't override them, rewrite them, or talk its way around them.

Under the hood: Regex-based and pattern-based filters applied at the executor boundary, not at the planning layer. Financial, credential, legal, and crisis tripwires on inbound WhatsApp. DLP on outbound payloads. Path jails and destructive command patterns on shell operations. Immutable to the reasoning model.

In plain English: A watchdog that monitors a window or screen region in the background. When something changes — a build finishes, a download completes, an error appears — it speaks up.

Under the hood: 64-bit difference hashing (dHash) with localized region-of-interest cropping and Hamming distance threshold detection. Configurable polling interval. Optional spoken TTS alert on condition met. Supports screen-region and per-window targeting.

In plain English: One switch that cuts off all internet access. Ghosthread runs fully on your machine with no cloud calls of any kind.

Under the hood: Runtime socket interdiction layer. All multi-step reasoning routes strictly through local offline rule evaluators when the air-gap flag is set. Cloud LLM calls are structurally blocked, not just disabled.