Windows Harness

Use one CLI call per decision point, not per primitive:

windows-harness <<'PY'
app = "Notepad"
win.see(app)
win.type("hello", app=app)
print(win.see(app))
PY

The CLI preloads win, lazy browser, Path, and subprocess. Prefer bounded stdin programs; reserve windows-harness repl for manual exploration and always exit it.

How to invoke: exactly two forms

One-off questions — use a subcommand (identical in PowerShell, cmd, and bash):

windows-harness apps                 # app inventory, compact JSON (--all adds system windows)
windows-harness see "CC Switch"      # one screenshot -> JSON with its "path"
windows-harness state "Notepad"      # UIA tree as JSON (--screenshot adds image)

Anything multi-step — write a .py file, then run it. No quoting or encoding traps in any shell; this is always the most reliable channel:

windows-harness run task.py          # win, browser, Path, subprocess preloaded

Write generated task scripts to the harness scripts dir (printed by windows-harness doctor as scripts_dir, default %USERPROFILE%\.windows-harness\scripts); run resolves bare filenames there. Never write task scripts into the caller's working directory or invent ad-hoc folders like .windows-harness-temp/ — they pollute the user's repo and show up as untracked files. Never run them via python script.py / python -c either: bypassing windows-harness run also bypasses the preloaded win session, so every call pays a fresh interpreter start and loses the element table behind element_index.

Never use <<'PY' heredocs outside bash; never enumerate windows through hand-rolled C#/PowerShell — win.list_apps() already did it. Screenshots land in %TEMP%\windows-harness-*.png and persist: the returned "path" stays readable after the CLI exits, so just open it. Set WINDOWS_HARNESS_OVERLAY=off to skip the pointer renderer for minimum latency.

Delivery: foreground first

Every input primitive takes delivery="foreground" (default) or delivery="background".

Why foreground is the default: focus-driven UI (search boxes with suggestion popups, context menus, IME candidate windows) closes the instant its window loses the foreground, and CEF off-screen / XAML input stacks only accept message input while truly foreground and focused. Holding the foreground across a burst turns "sometimes works" into a procedure.

• win.release() hands the foreground back to the user's previous window. Call it when a task burst is done.
• For one quick invisible action use delivery="foreground", hold=False (cloaked ~150 ms takeover, restored after) or delivery="background".
• Foreground clicks use SendInput mouse events and do NOT summon the Windows touch keyboard; synthetic pen/touch injection (background clicks) can pop it up and wreck the focus chain — avoid pen taps into text fields.
• background results carry "mode": "pen" | "message" plus "verified": false when the effect cannot be confirmed. When a framework's input stack silently drops background events the call raises BackgroundUnavailable with a structured code; redo the SAME action with default foreground delivery — do not probe more background transports.
• windows-harness doctor reports input_health.sendinput: ok, swallowed (hook software such as audio enhancers or overlay recorders is eating injected events — suspects under possible_injectors), or unknown. When SendInput is swallowed, foreground actions automatically fall back to pen/message transports against the now-foreground target.

Use the small surface

Think in seven verbs: see, key, type, click, paste, ax, script — plus two inventory calls: win.list_apps() (processes with their window titles) and win.windows(app) (one app's windows).

frame = win.see("Notepad")
win.key("ctrl+s", app="Notepad")
win.type("Alessia Cara", app="Notepad")                    # type at the focused field
win.type("Alessia Cara", app="Notepad", x=640, y=420)      # click the field, then type
win.click(640, 420, app="Notepad")
win.click(640, 420, app="Notepad", clicks=2)               # double click
win.click(640, 420, app="Notepad", button="right")         # context menu
win.drag(200, 300, 500, 600, app="Notepad")                # press, glide, release
win.scroll(-720, app="Notepad")                            # wheel at window center (or x/y)
win.hover(640, 420, app="Notepad")                         # really hover -> tooltip appears
win.paste("Alessia Cara", app="Notepad")                   # clipboard paste (see ladder)
win.release()                                              # hand the foreground back

item = win.ax.at(640, 420, app="Notepad")
win.ax.perform(item["element_index"], "invoke")
win.ax.click(item["element_index"])      # click the element's center — no coordinates
win.ax.hover(item["element_index"])      # hover it so its tooltip fires

win.script("Get-Process | Select-Object -First 3 Name")

win.click() is raw coordinate input; it never guesses a UIA action. For semantic actions use win.ax.at() + win.ax.perform(), and fill text fields with win.ax.set_value() (verified by read-back) when an element exposes a ValuePattern — no keystrokes at all. win.ax.click() / win.ax.hover() act on the element's own BoundingRectangle center: coordinates never leave the harness, so no coordinate-space mismatch is possible — prefer them over copying frame values into win.click() yourself.

win.hover(x, y) delivers a real hover (physical cursor, or pen hover when SendInput is filtered) so tooltips and hover-only UI appear; win.move(x, y) only animates the virtual pointer and delivers no input. Hover is coordinate-routed — the target needs to be unoccluded, not foreground — and the cursor stays put so the tooltip survives long enough for a win.see().

Modal dialogs and overlays

win.see(app) is window-scoped: it captures only that one window's client area, so an overlay sitting on top (a Save As / Open / Print dialog, a confirm prompt) is not in the shot even though it is on screen. To see a popup that just appeared, do not guess its title:

win.key("ctrl+s", app="Notepad")        # an untitled doc opens Save As
fg = win.foreground_window()            # the dialog's title/hwnd
frame = win.see(fg["hwnd"])             # capture the dialog itself
# or grab the whole desktop, which sees every window at once:
desktop = win.capture_screen()

win.capture_screen() captures the entire virtual desktop (all monitors) and its screenshot-space coordinates pair with coordinate_space="screen". A bare win.see() with no app now targets the foreground window, so it also catches a modal that is currently in front.

win.see(app, bring_to_front=True) fronts the target and keeps it fronted (restoring it if minimized) until win.release(); the default is a quiet background grab that never raises or activates the window, so see stays non-intrusive unless you opt in.

Text input ladder

1. win.ax.set_value() when the element exposes a ValuePattern — verified by read-back and works fully in the background.
2. win.type(text, x=..., y=...) — focuses the field and types in one call.
3. win.paste(text) — the clipboard route. Setting the clipboard is a plain data handoff no hook software can filter; only the paste trigger needs input (SendInput Ctrl+V, or a posted WM_PASTE when SendInput is filtered — never a posted Ctrl+V, whose modifier never reaches GetKeyState). Focus the target field first (a click, or win.type with x/y).
4. If the paste trigger also fails: right-click the field and click the context menu's Paste item (pen injection can drive the menu).
5. When doctor reports input_health.sendinput == "swallowed", tell the user which processes (possible_injectors) are eating injected input — removing the hook software is the real fix.

win.type() reports what actually landed, not just that keystrokes were sent. XAML/WinUI text boxes (Win11 Notepad, Settings) and CEF editors drop much of a KEYEVENTF_UNICODE burst while claiming verified: True; the harness reads the field back and, on a mismatch, retries via win.ax.set_value() (verified) or win.paste(). If the result carries "retried_via", the keyboard route dropped characters and the retry landed them. The result's verified_via says how verified was established: "value_pattern"/"clipboard" mean a read-back confirmed the text, "transport" means the target exposes no readable text state and only delivery was confirmed. A verified: false with a read_back field means the window accepted the keystrokes but the text did not appear (e.g. an app-launch/session-restore race) — click the field and retry, or switch to win.paste().

Replacing a field's existing text: there is no reliable select-all keystroke on hook-filtered machines (Ctrl+A is a combo and refuses; CEF ignores a posted Backspace). The dependable sequence is a triple click — win.click(x, y, clicks=3) selects the field's text — then win.type or win.paste, which replaces the selection. On machines with healthy SendInput, win.key("ctrl+a") first is fine too.

Bare keys vs combos: when SendInput is swallowed, win.key("enter") and other modifier-less keys still work (posted to the foreground target); only combos like ctrl+a refuse, honestly.

Coordinates

coordinate_space accepts:

• 'screenshot' (default) — pixels of the latest win.see() image; the returned scale_x/scale_y map them to the window.
• 'normalized' — a 0..1000 grid over the client area of the window from the latest win.see(), independent of screenshot resolution and DPI; use directly when the model outputs normalized coordinates:

  frame = win.see("Notepad")                          # anchors the window
  win.click(500, 48, app="Notepad", coordinate_space="normalized")
  

  
  The window's origin is resolved live at click time, so the window may move between see() and click(); only a resize invalidates the grid — take a fresh see() after any resize.
• 'client' / 'screen' — raw window-client or physical screen pixels.

Every positional primitive (click, double/right click, drag, scroll with x/y, type with x/y, hover) accepts coordinate_space; the normalized grid works for all of them.

Default to vision + normalized coordinates; only fall back to element centers from the ax tree when you need semantic identity or a layout shift makes a coordinate fragile. ax output reports frames in the pixel space of your latest win.see() screenshot of that window (the frame key, alongside "frame_space": "screenshot" in state), so they drop straight into the default coordinate space. Without a matching screenshot the raw physical screen pixels are reported under frame_screen instead — those only pair with coordinate_space="screen", never the default. Cleanest of all is keeping numbers out of the loop: win.ax.click(element_index) / win.ax.hover(element_index).

Target apps by exe name when titles can collide — resolution prefers an exact process-name match ("Discord.exe", or bare "Discord"), then an exact title, and only then substrings, so a window whose title merely mentions the name (a VS Code tab titled "… Discord …") never hijacks the query.

Action proof: look only when it misfires

Every coordinate primitive (click, drag, scroll, hover, type with x/y) stamps the intended point back onto the screenshot it was anchored to and returns the path under result["proof"]["path"]. The proof key is optional: it is present only when annotation is on (the default) and the proof was actually generated, so read it through result.get("proof") before using it.

out = win.click(640, 420, app="Notepad")
proof = out.get("proof")
print(proof["path"] if proof else "no proof")   # annotated PNG you can open

It stays absent for non-coordinate primitives (key, paste, move, ax.click, see) and for any call given annotate=False, or when WINDOWS_HARNESS_PROOF=off.

• click / double / right / hover / click-to-type draw a red reticle (circle-cross) at the point, labelled with the button and count.
• scroll draws an arrow from the anchor point (the window centre whenever x/y are omitted) in the direction of the wheel delta.
• drag draws the glide as a blue line with a reticle at the start and a square at the end.

Do NOT open the proof on every action — the input already ran. Open it only to decide WHY a coordinate action mis-landed: you aimed at the wrong spot, the effect did not happen, or you need to nudge a coordinate before retrying. A normal win.see() on the finished step is the usual confirmation.

The concrete trigger: an action reports success (or verified: true) but win.verify_change(app) says nothing changed, or a win.see() shows the page changed in a way you did not expect. In that case open result["proof"]["path"] FIRST — before retrying, nudging, or switching transports. The reticle on the anchored screenshot tells you at a glance whether the input even landed where you intended; guessing from coordinates alone just burns rounds.

Forgot to print the proof path, or need an earlier one? Every proof is journaled to proofs.jsonl under the harness config dir (default %USERPROFILE%\.windows-harness) as it is written, so history survives across CLI invocations:

win.last_proof()                      # newest proof whose PNG still exists
win.proofs(limit=5)                   # recent proofs, newest first
win.proofs(app="Notepad", kind="click")   # filter by process/title and kind

Each entry carries the same fields as result["proof"] plus ts and the target app (process/title); entries whose PNG was cleaned up are skipped. Pass annotate=False to any primitive (or set WINDOWS_HARNESS_PROOF=off) to turn the proof off for a coordinate you already know is good.

Minimize round trips

• Bundle deterministic, reversible steps into one program, then verify once.
• Stop at a genuine decision boundary: ambiguous identity, new coordinates, an irreversible action, or unexpected state.
• Do not screenshot merely to confirm a known shortcut opened a field before typing; let the final screenshot verify the whole sequence.
• Poll exact UIA state inside the same Python program when possible.
• Use the cheapest strong end-state check: one screenshot for visible state, one exact UIA query for semantic state.

Choose the lowest useful mode

Vision-first: if you can read a screenshot, lead with win.see(app) + normalized coordinates (a 0..1000 grid) instead of the UIA tree. Many apps (Electron/custom-drawn UI, game launchers, some frameless and webview windows) expose no usable UIA tree, and even when they do, screenshot + coordinates need no per-app support. Reach for win.ax only when you need semantic identity or state.

1. Use win.script() for a known exact command (Get-Process, registry reads).
2. Otherwise use win.see(app) and vision: locate the target in the screenshot, then act with coordinate_space="normalized" (0..1000) so the numbers are resolution/DPI independent.
3. Prefer a known keyboard route; use a verified coordinate for a visible, low-risk target.
4. Use targeted win.ax only when semantic identity or state matters. For CEF/Electron custom-drawn UI (chat apps, music players, game launchers) the tree fills lazily: state/ax.query/ax.at automatically warm a near-empty CEF tree with a brief (usually invisible, cloaked) foreground round, so warm it once if you truly need a semantic handle. Icon-only controls (server/avatar images) often expose only generic names — hover them (win.ax.hover) and read the tooltip from a win.see(). If a dump still comes back near-empty with a note, go straight to screenshot + normalized coordinate clicks instead of spending more rounds on UIA.
5. Use delivery="background" only for apps already proven to accept it on this machine (classic Win32 controls, UIA patterns) — it is the quiet opt-in, not the default to probe with.

After a failed verified burst, switch mode or stop. Never repair uncertainty with repeated keys, clicks, deletion loops, or bulk input.

Verify, then teach the matrix

• After a "verified": false action, confirm the effect with win.verify_change(app) — a background pixel diff that costs no focus and no vision round trip. Use a full win.see() only when you need to know WHAT changed, not THAT it changed. Animated windows (video, games) report change on their own; verify those semantically through win.ax instead.
• When a background action provably had no effect (verify_change negative AND the task did not advance), record it once with win.note_drop(kind) — e.g. win.note_drop("text_input", app="..."). Future background calls against that window class refuse honestly instead of repeating a dead transport.

Keep the invariants

• Foreground hold=True keeps the target fronted across the whole burst; always end a task with win.release() (or leave the window fronted deliberately and say so).
• Background delivery NEVER activates, raises, or moves the cursor. Occluded targets refuse with background_occluded rather than being raised.
• ax queries against CEF/Electron windows (Chrome_WidgetWin*) whose tree is near-empty auto-warm it with a brief cloaked foreground takeover; the previous foreground is restored when the warm-up round ends.
• Elevated (Administrator) targets are detected up front (UIPI preflight) and refused with background_uipi_blocked; run the harness elevated to drive them.
• The animated pointer is click-through and never moves the physical cursor; non-held foreground rungs move it briefly and put it back.
• Minimized windows are restored before foreground input and stay restored under hold=True; non-held foreground rungs re-minimize afterwards.
• Prefer semantic UIA patterns over simulated window chrome: frameworks that draw their own titlebar (WinUI3, Electron) may ignore posted WM_CLOSE and caption-button Invoke. The framework-neutral close is the WindowPattern — win.ax.raw(i).GetWindowPattern().Close() — then answer the app's own confirmation through ax.perform(..., "invoke").
• XAML apps (Win11 Notepad, Settings, UWP frames) refuse background typing — use win.ax.set_value() there; WPF refuses background drag and scroll — use default foreground delivery.

Browser tasks

Use the preloaded lazy browser object whenever the task lives inside a web page. It exposes Browser Harness helpers and connects to the real browser on first use:

print(browser.page_info())
browser.new_tab("https://example.com")

Do not substitute OS input for CDP inside web content; keep Windows primitives for browser chrome, native dialogs, downloads, and everything outside the page.

Run windows-harness doctor to inspect the runtime without changing anything.