@bastani/atomic 0.9.19-alpha.3 → 0.9.19-alpha.5

package/docs/computer-use.md

@@ -0,0 +1,390 @@
+---
+title: "Computer use"
+description: "Use desktop apps, browsers, and terminals with Atomic on macOS, Linux, and Windows."
+---
+
+# Computer use
+
+Atomic can work in applications, not just edit code. Computer-use automation, or CUA, can create a Blender scene, build a presentation, edit a video, operate a desktop app, or move work between applications. Browser and terminal automation cover related tasks with more direct controls.
+
+This guide explains tool selection, setup, and safe operation. To test a software change and attach the results to a PR, see [Verification and evidence](/workflows/verification).
+
+Jump to [application scripting](#application-scripting-and-apis), [desktop CUA](#desktop-automation-with-pyautogui-and-uv), [browser automation](#browser-automation-with-playwright-cli), [terminal automation](#terminal-automation-with-herdr), or [creative workflows](#creative-work-and-cua-workflows). Platform setup: [macOS](#macos), [Linux](#linux), [Windows](#windows).
+
+## Choose the right tool
+
+Start with the result you need, not the application you could click through. If a library, CLI, or supported API can produce that result directly, a short script is often simpler and more token-efficient than repeated screenshots and UI actions. Use computer use when the task needs visual judgment, a UI-only operation, or verification of the interface itself. Saving tokens is useful, but not at the expense of the requested behavior or output quality.
+
+| Task | Preferred tool | When to use something else |
+| --- | --- | --- |
+| Create or edit files, such as presentations, documents, spreadsheets, or media | **A file library or CLI** | Use an app API or UI when the library cannot preserve required features, or when you need rendering or visual adjustments. |
+| Interactive terminal or TUI | **Herdr** | Use tmux on macOS/Linux or native Windows psmux when Herdr cannot be used. Ordinary shell commands need no multiplexer. |
+| Browser page or web application | **playwright-cli** | Use desktop CUA for browser chrome or OS dialogs that browser automation cannot reach. Keep existing Playwright test suites for repeatable tests. |
+| Desktop application or work across apps | **PyAutoGUI, run with uv** | Use native accessibility tools, application scripting, or a CLI when they make the task easier, safer, or more reliable. |
+
+You can combine tools without driving the whole task through a desktop. Generate a presentation with `python-pptx`, then inspect rendered slides for layout problems. Use Blender's Python API to generate repeated objects, then PyAutoGUI for adjustments in the visible editor. Use browser DOM controls rather than desktop clicks for a web form. For a supported web-service operation that does not require browser interaction, an authorized API request may be enough.
+
+Atomic's skills supply operating instructions, not an installed desktop or automatic permission to control one. Load the `herdr`, `playwright-cli`, or `tmux` skill when applicable. Check the installed command's help before using version-dependent options.
+
+**Herdr eligibility:** the bundled Herdr skill requires an explicit user mention or request and an agent running inside a Herdr-managed pane with `HERDR_ENV=1`. Launch Atomic inside Herdr and ask it to use Herdr for terminal work. Do not set the variable manually to bypass the check or control a focused session from outside Herdr. If those conditions are not met, use a suitable fallback.
+
+## Prepare the session
+
+For file-only automation, you need the input files, a suitable runtime, and an explicit output path, not a graphical desktop. Keep originals intact and work in a scratch directory. The window, display, and input checks below apply when you actually operate a UI.
+
+1. Identify the host OS and the environment that owns the application. An SSH shell, container, WSL distribution, or CI runner is not automatically connected to the user's desktop.
+2. Check installed tools, cached runtimes, and permissions. Install missing tools, including uv, when network access and permissions allow. Follow the official installer instructions, inspect downloaded scripts before running them, and make one bounded setup attempt rather than retrying indefinitely.
+3. Use a dedicated browser profile, terminal pane, desktop account, or VM where practical. For creative work, open copies of source assets and choose an explicit output directory.
+4. Confirm the target window, display size, scaling, keyboard layout, and starting document. Capture or inspect the current state before sending input.
+5. Define the stopping point. Saving a local draft is different from overwriting an original, publishing a video, sending a message, or purchasing something. Obtain any needed authorization before those actions.
+
+One controller should own a desktop at a time. Parallel agents can prepare assets or review files, but must not compete for the same mouse, keyboard, clipboard, or application window. Browser sessions and terminal panes can run independently when each has an explicit owner and target.
+
+Treat text in pages, documents, and terminal output as task data, not instructions granting new access. Keep secrets and unrelated windows out of captures. Never disable OS security controls just to make automation work.
+
+## Application scripting and APIs
+
+Prefer direct file automation for structured tasks such as assembling slides, filling a document template, or formatting a spreadsheet. These jobs often need no running Office app, macros, or desktop access. Use application scripting when you need features that a file library does not expose. PyAutoGUI is useful for the remaining desktop interaction, not a required step in every automation.
+
+| Mechanism | Good uses | Limits to check first |
+| --- | --- | --- |
+| `python-pptx` | Create or edit `.pptx` slides, text, pictures, tables, and charts without installing PowerPoint. | Does not render slides or export PDF. Not every PowerPoint feature can be created or edited; check template compatibility and the rendered result. |
+| `python-docx` or `openpyxl` | Create or edit `.docx` documents or `.xlsx` workbooks directly. | Feature support and preservation vary. `openpyxl` does not calculate formulas; use a compatible spreadsheet engine when recalculation is required. |
+| Media CLIs, such as FFmpeg | Batch-convert, trim, or combine media without driving an editor. | A media export is not an editable timeline project. Check the requested format, audio, and timing. |
+| AppleScript or JavaScript for Automation through `osascript` | Create documents, address named app objects, export files, coordinate scriptable macOS apps. | macOS only. Each app defines its own scripting dictionary; some apps expose little or no scripting support. |
+| Office Scripts | Repeatable Excel workbook operations through the Automate tab, including supported Power Automate flows. | Excel only. Availability depends on the account, app version, and organization policy; it is not a general desktop-control API. |
+| PowerShell with COM automation | Drive installed Windows applications that expose COM, including desktop Office. | Windows-specific. Do not assume unattended service execution is supported or reuse the user's active app instance without permission. |
+| Application APIs, such as Blender's Python API | Generate geometry, set scene properties, apply repeated edits, and render or export. | Use the API and runtime for the installed app version. Some operations depend on an active document, selection, or editor context. |
+
+Before writing a script, identify the input format, required features, output path, and library or app version. Read the relevant API reference rather than guessing methods. Start with a read-only query or a disposable copy. Save to a new path and reopen the result to check its contents; use a compatible viewer or renderer when appearance matters. Scripts still need the same authorization as UI actions to overwrite, upload, or publish files.
+
+### macOS recipe: create a draft with osascript
+
+Open Script Editor and choose File > Open Dictionary to inspect an application's supported commands, objects, and properties. Apple's [scripting terminology guide](https://developer.apple.com/library/archive/documentation/LanguagesUtilities/Conceptual/MacAutomationScriptingGuide/AboutScriptingTerminology.html) explains how to read the dictionary. App scripting addresses document objects directly; `System Events` UI scripting instead drives accessible interface controls and needs Accessibility permission.
+
+Save this as `create-note.applescript`:
+
+```applescript
+on run argv
+    if (count of argv) is not 1 then error "Pass the draft text as one argument."
+    set draftText to item 1 of argv
+    tell application "TextEdit"
+        set draft to make new document with properties {text:draftText}
+        activate
+        return (text of draft) as text
+    end tell
+end run
+```
+
+Run it from a macOS shell:
+
+```sh
+osascript create-note.applescript "Draft outline for the presentation"
+```
+
+This creates a new, unsaved TextEdit document and returns its text to the shell. It does not overwrite a file. Check the returned text and inspect the document, then save to an agreed destination if required. macOS may ask permission for the launching app to control TextEdit; let the user grant it.
+
+Pass content as arguments rather than interpolating it into executable script text. For longer content, have the script read an explicit input file. JavaScript for Automation is another macOS option, invoked with `osascript -l JavaScript script.js`; it uses Apple's automation objects, not a browser DOM or Node.js APIs. Use whichever language fits the app's documentation and existing scripts.
+
+### PowerPoint recipe: create a draft with python-pptx
+
+Use [python-pptx](https://python-pptx.readthedocs.io/en/latest/user/quickstart.html) to assemble a `.pptx` directly instead of creating slides through desktop clicks or VBA. It runs on macOS, Linux, and Windows without PowerPoint or a graphical session.
+
+Save this as `create_deck.py` in a scratch directory:
+
+```python
+from pathlib import Path
+
+from pptx import Presentation
+
+deck = Presentation()
+title_slide = deck.slides.add_slide(deck.slide_layouts[0])
+title_slide.shapes.title.text = "Quarterly review"
+title_slide.placeholders[1].text = "Draft for discussion"
+
+summary = deck.slides.add_slide(deck.slide_layouts[1])
+summary.shapes.title.text = "Next steps"
+body = summary.placeholders[1].text_frame
+body.text = "Review the results"
+body.add_paragraph().text = "Agree on next quarter's priorities"
+
+output = Path("quarterly-review-draft.pptx")
+with output.open("xb") as stream:
+    deck.save(stream)
+print(f"Created {output.resolve()}")
+```
+
+Run it from that directory with [uv](https://docs.astral.sh/uv/):
+
+```sh
+uv run --no-project --with python-pptx python create_deck.py
+```
+
+`--no-project` keeps this one-off task separate from an unrelated Python project. uv may download Python and dependencies on the first run. The script creates two slides and refuses to overwrite an existing output file. Choose a new filename for another draft.
+
+This example uses the layouts and placeholder IDs in the library's default template. For a branded deck, load a copy of your `.pptx` template with `Presentation("template.pptx")` and inspect its layouts and placeholders before adapting the script. Do not assume their indices match the default template. See [working with presentations](https://python-pptx.readthedocs.io/en/latest/user/presentations.html) and [using placeholders](https://python-pptx.readthedocs.io/en/latest/user/placeholders-using.html).
+
+Reopen the saved deck to check slide count and text. Then view it in PowerPoint, LibreOffice Impress, or another compatible renderer to check clipping, fonts, and layout. `python-pptx` does not render slides or export PDF; use a compatible application for those steps. A successful save is not a visual check. If no renderer is available, hand off the draft and state that its appearance remains unchecked.
+
+For similar file-based tasks, use [python-docx](https://python-docx.readthedocs.io/en/latest/) for Word documents or [openpyxl](https://openpyxl.readthedocs.io/en/stable/) for Excel workbooks. Check feature support before editing a complex existing file. Use an app's own API when a library cannot make the required change, rather than forcing a lossy conversion. If an approved task requires macros, inspect the code and follow the organization's macro policy; never weaken security settings to run it.
+
+### Office Scripts, app runtimes, and file tools
+
+For Excel on the web or a supported desktop installation with the Automate tab, consider Office Scripts. Record a small action or create a script there, then use the `ExcelScript` workbook API for repeatable edits. These TypeScript scripts are not VBA and do not run as ordinary Node.js scripts. Check [Office Scripts versus VBA](https://learn.microsoft.com/en-us/office/dev/scripts/resources/vba-differences) for platform, licensing, and API differences. Creating a Power Automate flow can introduce scheduled runs and cloud access; do so only when that automation is part of the request.
+
+Use an application's own scripting runtime when it supplies the API. For example, Blender scripts normally run through Blender's Python Console, Text Editor, or command line. A plain uv Python environment does not automatically have the running application's `bpy` module or scene. With Blender on PATH, an existing `input.blend`, and a reviewed `scene-script.py`, a batch invocation is:
+
+```sh
+blender --background input.blend --python-exit-code 1 --python scene-script.py
+```
+
+Argument order matters. This loads the scene before running the script, and `--python-exit-code 1` makes a script exception produce a nonzero process exit. The script must explicitly save or export any intended output to a new path; exiting successfully does not imply a saved scene. See the [Blender Python quickstart](https://docs.blender.org/api/current/info_quickstart.html) and [command-line reference](https://docs.blender.org/manual/en/latest/advanced/command_line/arguments.html). Use uv for external orchestration or file-processing scripts, and Blender's runtime for Blender operations.
+
+For file-only work, a library can avoid opening the application at all. Know what it preserves: [openpyxl does not calculate Excel formulas](https://openpyxl.readthedocs.io/en/stable/simple_formulae.html), and a deck created with [python-pptx](https://python-pptx.readthedocs.io/en/latest/) still needs a layout check for clipping, fonts, and missing media. For video, FFmpeg can handle batch transforms while an editor's own scripting API can retain timeline structure. Check installed API/version or edition limits before assuming an editor exposes scripting.
+
+Combine these approaches only where they help. Generate content with a script, inspect it in a viewer, and use CUA if it needs visual adjustments or UI-only export controls. You do not need a desktop interaction just to prove that a file script ran. If the task is specifically to verify a menu, dialog, or user flow, exercise that interface too; an API call is not proof that the GUI path works.
+
+## Desktop automation with PyAutoGUI and uv
+
+[PyAutoGUI](https://pyautogui.readthedocs.io/en/latest/) controls the real mouse and keyboard and captures screenshots. It does not understand the application by itself. Atomic must inspect the screen or another reliable state source between actions.
+
+Prefer [uv](https://docs.astral.sh/uv/) to manage Python and the script's dependencies. If `uv --version` fails because uv is missing, install it using the instructions for your OS below. An existing Python environment is a fallback when uv installation is blocked, not a reason to change an unrelated repository's dependencies.
+
+### Run an isolated script
+
+Save the following as `desktop_probe.py` in a scratch directory. It takes a screenshot and reports geometry without clicking or typing:
+
+```python
+from pathlib import Path
+
+import pyautogui as gui
+
+gui.FAILSAFE = True
+gui.PAUSE = 0.25
+
+output = Path("artifacts")
+output.mkdir(exist_ok=True)
+print(f"Screen: {gui.size()}; pointer: {gui.position()}")
+gui.screenshot().save(output / "desktop-before.png")
+```
+
+Run it from that directory:
+
+```sh
+uv run --no-project --with pyautogui --with pillow python desktop_probe.py
+```
+
+`--no-project` avoids discovering or syncing an unrelated Python project. uv can download a Python runtime and dependencies if needed, so this first run may require network access. OS screenshot and accessibility dependencies still need separate setup. For a reusable script, declare dependencies in [inline script metadata](https://docs.astral.sh/uv/guides/scripts/#declaring-script-dependencies), pin versions, and use uv's script locking support. Keep scratch environments and captures out of the application's repository unless they belong in the deliverable.
+
+Open the captured image and confirm it shows the intended desktop. A successful import is not proof that screenshots or input work. Test a harmless action in a disposable document before running a longer sequence.
+
+### Observe, act, and check
+
+- Use short action sequences. Inspect the result after opening a menu, changing focus, or switching applications.
+- Prefer named accessibility controls or application APIs where available. If using coordinates, derive them from the current screen rather than an old screenshot.
+- Keep the target on the primary monitor. PyAutoGUI's multi-monitor support is limited. Retina and DPI scaling can make screenshot pixels differ from input coordinates; compare screenshot dimensions with `gui.size()` before clicking.
+- `gui.write()` sends keystrokes to the focused window and is not a general Unicode text-insertion API. For non-ASCII content, prefer app scripting or a controlled clipboard paste. Clipboard contents may be sensitive, so preserve and restore them when appropriate.
+- For image matching, crop to the relevant region and use fixtures from the same theme and scaling. Handle a missing image as a failed observation, not a reason to click a default location. PyAutoGUI's `confidence` option requires OpenCV in the Python environment.
+- Wait for an observable result with a deadline. A fixed sleep alone does not prove a render, export, or save has finished.
+
+See PyAutoGUI's [keyboard controls](https://pyautogui.readthedocs.io/en/latest/keyboard.html), [mouse controls](https://pyautogui.readthedocs.io/en/latest/mouse.html), and [screenshot functions](https://pyautogui.readthedocs.io/en/latest/screenshot.html) for API details.
+
+### Stop and recover safely
+
+Keep `FAILSAFE` enabled and leave a pause between calls. Moving the pointer to a corner of the primary monitor causes a subsequent PyAutoGUI call to raise `FailSafeException`. Keep a separate way to interrupt the automation process available too.
+
+Prefer complete actions such as `press`, `hotkey`, and `click` over holding input across several steps. If a script must hold a key or mouse button, track what it holds and release it in cleanup. An interrupt or failsafe can itself prevent PyAutoGUI cleanup calls. Stop the script, check for held input, and release it manually or through a safe native mechanism before resuming. Do not disable the failsafe in order to keep clicking.
+
+After a timeout or interruption, inspect the current document and any output files. A save or export may have completed even if its acknowledgement was lost. Do not repeat destructive actions blindly.
+
+## Browser automation with playwright-cli
+
+For tasks that require browser interaction, prefer [playwright-cli](https://github.com/microsoft/playwright-cli) for websites and web apps on all three desktop platforms. Its snapshots expose page structure and element references, so automation can use actual controls rather than screen coordinates. For data retrieval or batch operations, consider a supported API first when it meets the request and you have permission to use it.
+
+### Setup and first session
+
+Load the `playwright-cli` skill and check `playwright-cli --help`. If the command is unavailable, check whether the project's installed Playwright exposes `npx --no-install playwright cli --help`. Otherwise install the CLI when permitted:
+
+```sh
+npm install -g @playwright/cli@latest
+playwright-cli --help
+```
+
+Use the installed CLI's browser setup guidance if a browser is missing. Do not add browser automation dependencies to an unrelated project just to run a one-off task.
+
+Create a uniquely named session, replacing `desktop-demo` if that name is already in use:
+
+```sh
+playwright-cli -s=desktop-demo open https://example.com --headed
+playwright-cli -s=desktop-demo snapshot
+playwright-cli -s=desktop-demo screenshot --filename=browser-before.png
+playwright-cli -s=desktop-demo close
+```
+
+For a real task, act between the snapshot and final capture. Read element references from the current snapshot, then use `click`, `fill`, `select`, or `press`. Do not reuse an example reference such as `e5` without discovering what it points to. Refresh the snapshot after navigation or substantial UI changes.
+
+### Best practices
+
+- Keep the same session name on every command. Close only sessions you created, not every browser on the machine.
+- Prefer a fresh profile. Attach to an existing personal browser only when authorized; stored sessions can expose private tabs and credentials.
+- Use headed mode for visual work. Headless mode can verify DOM behavior, but does not establish that desktop integration or native dialogs work.
+- Inspect visible results and relevant console/network output. Use semantic locators and assertions in a maintained Playwright test for repeatable regression coverage.
+- Use `upload` for supported file inputs rather than driving an OS file picker. Switch to CUA or native tooling only for UI outside the page, and then recheck focus before returning to browser control.
+- Treat cookies, saved authentication state, traces, and network logs as sensitive. Do not commit or attach a browser profile as evidence.
+- Browser mobile emulation tests a web viewport, not a native Android or iOS application.
+
+For verification captures and recordings, see [browser evidence](/workflows/verification#browser-changes).
+
+## Terminal automation with Herdr
+
+Prefer Herdr for interactive terminal work on macOS, Linux, and Windows, subject to the [eligibility requirements](#choose-the-right-tool). Use ordinary shell execution for builds, scripts, or commands that do not need interactive input. A long-running command alone is not a reason to add a multiplexer.
+
+### Setup and a dedicated pane
+
+Check `herdr --version` and `herdr --help`. Install missing Herdr using its [official installation guide](https://github.com/herdrdev/herdr/tree/v0.9.0#install) when permitted. macOS supports `brew install herdr`; macOS/Linux and Windows also have official shell and PowerShell installers. Inspect downloaded scripts before executing them.
+
+Once inside an eligible managed pane, check `herdr status` for client/server compatibility. Do not stop or replace an active server just to obtain a new feature. Atomic's automatic status reporting is documented separately in [Herdr integration](/herdr).
+
+Use the Herdr skill to discover the current pane and create a dedicated sibling without changing the user's focus. For example, in a POSIX shell inside Herdr:
+
+```sh
+herdr pane split --current --direction right --cwd "$PWD" --no-focus
+```
+
+Choose the split direction to suit the available space. Read the new pane ID from the creation response. In the commands below, replace `<pane-id>` with that returned ID and `<command>` with the intended command:
+
+```text
+herdr pane run <pane-id> "<command>"
+herdr pane wait-output <pane-id> --match "<expected output>" --timeout 10000
+herdr pane read <pane-id> --source visible
+```
+
+`pane run` sends text and Enter. `pane send-text` alone does not submit. `wait-output` can match text already on the screen, so use a fresh pane or a run-specific marker and inspect the result. The appearance of a marker is not a substitute for checking the command's exit status or the application's final state.
+
+For ordinary logs, `--source recent-unwrapped --lines 120` avoids soft-wrapped lines. For layout, inspect `visible` at the intended terminal dimensions. Alternate-screen content that has scrolled away may not be recoverable from host scrollback. Capture important states as they occur.
+
+### tmux and psmux fallbacks
+
+When Herdr is unavailable, cannot be installed, or has no eligible managed session, use tmux on macOS/Linux or [psmux](https://github.com/psmux/psmux) for native Windows terminals. Record the reason when it affects the requested coverage. Do not take over an unrelated pane to satisfy the preference.
+
+Load the tmux skill and check the installed help. tmux and psmux share familiar commands, but supported flags and behavior can differ. Herdr has a different CLI entirely.
+
+For tmux, create a dedicated session with a unique name, then discover its actual pane ID:
+
+```sh
+tmux new-session -d -s atomic-demo
+tmux list-panes -t atomic-demo -F '#{pane_id}'
+```
+
+Substitute the returned ID for `<pane-id>`:
+
+```text
+tmux send-keys -t <pane-id> -l -- "<command>"
+tmux send-keys -t <pane-id> Enter
+tmux capture-pane -p -t <pane-id>
+```
+
+Use literal text and a separate Enter to avoid interpreting arbitrary text as key names. On psmux, inspect `psmux list-panes` and use `psmux capture-pane -p -t <pane-id>` as documented in its [scripting guide](https://github.com/psmux/psmux/blob/master/docs/scripting.md). Discover IDs rather than assuming `%0` is your pane. Clean up only the session or pane created for the task. Never use a global server-kill command as routine cleanup.
+
+For modified-key setup in Atomic, see [tmux setup](/tmux). For behavioral checks and recordings, see [terminal evidence](/workflows/verification#terminal-changes).
+
+## macOS
+
+### Desktop and native tools
+
+- Install missing uv with `brew install uv` when Homebrew is available, or use the reviewed macOS installer from [uv installation](https://docs.astral.sh/uv/getting-started/installation/). Confirm `uv --version` in the launching shell.
+- Allow the application launching automation, such as Terminal or your IDE, in System Settings > Privacy & Security > Accessibility. Screenshot capture also needs Screen Recording permission, which may be labelled Screen & System Audio Recording on newer macOS versions. Relaunch the affected app if macOS requests it.
+- AppleScript automation may also prompt for Automation permission to control another app. Let the user grant permissions; do not script around consent dialogs.
+- PyAutoGUI depends on native Python bindings on macOS. If import or capture fails, check the installed release's [installation requirements](https://pyautogui.readthedocs.io/en/latest/install.html) before adding dependencies to the uv environment.
+- Check Retina scaling and keep the target on the primary display. A black or incomplete capture usually needs permission or display troubleshooting, not more clicks.
+
+Use `osascript` for AppleScript or JavaScript for Automation when an app's scripting dictionary exposes the operation you need. See [application scripting and recipes](#application-scripting-and-apis) for a runnable example and Office automation choices. `System Events` UI scripting and native accessibility APIs can address menus and controls more reliably than coordinates; consult Apple's [UI scripting guide](https://developer.apple.com/library/archive/documentation/LanguagesUtilities/Conceptual/MacAutomationScriptingGuide/AutomatetheUserInterface.html).
+
+`screencapture` is useful for native screenshots; Screenshot or QuickTime Player can record the screen or a selected area. Check permissions and the selected recording region before capture.
+
+### Browser and terminal
+
+playwright-cli uses its own browser session. WebKit coverage is not proof of every Safari-specific desktop behavior. Use an actual target browser when that distinction matters.
+
+Herdr is the first choice for interactive terminals when eligible. Homebrew provides Herdr and tmux. Preserve the shell, terminal dimensions, and keyboard behavior relevant to the task rather than silently changing them to make a scenario pass.
+
+## Linux
+
+### Desktop and native tools
+
+- Install missing uv through the [official Linux installer](https://docs.astral.sh/uv/getting-started/installation/) or an available distribution package. Review the installer before executing it, then confirm `uv --version`.
+- PyAutoGUI's Linux input backend uses X11. Run in an accessible graphical X11 session with the correct `DISPLAY` and authorization. Installing Python packages does not create a desktop session.
+- Check the distribution's screenshot and Python support packages. PyAutoGUI documents `scrot` and Python Tk/development packages for Linux; the required capture backend varies with the installed Pillow/PyScreeze versions. Use the distribution package manager with permission, not guessed cross-distribution commands.
+- A Wayland session is not equivalent to X11. XWayland does not grant access to every native Wayland app. Prefer compositor-supported capture/input tools, desktop portals, or native accessibility APIs when they support the operation. Do not weaken session security or claim PyAutoGUI has full Wayland support.
+- For unattended X11 work, a dedicated virtual display such as Xvfb can be useful. It does not prove behavior on a real Wayland desktop, GPU configuration, or physical display. Creative applications may require working graphics acceleration.
+
+[AT-SPI](https://gnome.pages.gitlab.gnome.org/at-spi2-core/) can expose named controls in accessible applications. `xdotool` and `wmctrl` can help with focus and window placement on X11; they are not general Wayland replacements. On Wayland, choose tools for the actual compositor and inspect their permission requirements. Use app APIs where custom canvases do not expose useful accessibility controls.
+
+For recordings, use a supported desktop recorder or OBS with the appropriate display or portal source. Confirm the saved file contains the intended window, not a blank capture.
+
+### Browser and terminal
+
+playwright-cli may need browser binaries and system libraries on a minimal Linux install. A headed browser needs a display. Headless browsing remains useful on SSH or CI hosts but does not grant desktop access.
+
+Herdr is preferred when eligible; tmux is a practical fallback on local or remote POSIX shells. An SSH terminal can run terminal scenarios without access to the remote desktop. Record which host owns the pane and application.
+
+## Windows
+
+### Desktop and native tools
+
+- Install missing uv with `winget install --id=astral-sh.uv -e` or use the reviewed PowerShell installer from [uv installation](https://docs.astral.sh/uv/getting-started/installation/). Open a new shell if PATH changed, then run `uv --version`.
+- Run PyAutoGUI and uv in the Windows graphical session that owns the app. Running them inside WSL does not automatically control native Windows windows.
+- Keep the session unlocked and available during automation. A disconnected or minimized Remote Desktop session can change rendering or input behavior; verify the actual remote-session setup before relying on it.
+- Use a consistent display scale and primary monitor. Check coordinates again after moving a window between displays with different DPI settings.
+- Standard-user automation cannot reliably drive elevated apps or the UAC secure desktop. Stop for the user or choose an authorized non-elevated path rather than escalating just to force input through.
+
+[Windows UI Automation](https://learn.microsoft.com/en-us/dotnet/framework/ui-automation/ui-automation-overview) exposes controls by name and automation ID. Tools such as [pywinauto](https://pywinauto.readthedocs.io/en/latest/) can be easier than pixel matching for accessible Windows apps. For structured document operations, start with [file libraries and app scripting](#application-scripting-and-apis). Use PyAutoGUI for the remaining visual interactions.
+
+Snipping Tool or OBS can capture desktop evidence. Check the selected window and saved recording before sharing it.
+
+### Browser and terminal
+
+Use native Windows playwright-cli when the task depends on Windows browsers, downloads, or desktop dialogs. Quote paths and URLs for the shell actually in use; do not paste POSIX shell syntax into PowerShell.
+
+Prefer native Herdr when eligible. In PowerShell, check `$env:HERDR_ENV -eq '1'`, use `(Get-Location).Path` for the working directory, and read pane IDs from CLI responses. If Herdr cannot be used, install psmux through its documented Windows installation options and inspect its help. WSL tmux is useful for Linux programs, but is not native Windows ConPTY coverage.
+
+See [Windows setup](/windows) for Atomic's shell requirements.
+
+## Creative work and CUA workflows
+
+Choose the deliverable first. A library or application API may produce it without computer use at all. Add visual interaction when it helps create or inspect the result.
+
+| Task | Practical approach | Useful deliverables |
+| --- | --- | --- |
+| Blender 3D modeling | Use Blender Python for repeatable geometry or scene setup; use PyAutoGUI for visible editor operations and visual inspection. | Editable `.blend` file, exported model if requested, preview render. |
+| Presentations | Generate structured slides with `python-pptx`, inspect them in a compatible viewer, and use CUA for visual refinements or slideshow interaction when needed. | Editable deck plus PDF or slide previews exported through a compatible application. |
+| Video editing | Use the editor's scripting API or media CLI for repetitive operations; use CUA to adjust the timeline, inspect transitions, and review playback. | Editable project, required source references, final export. |
+| Work across applications | Use native scripting for named windows and file operations; use PyAutoGUI where the task needs visual interaction. | Saved documents and a concise record of completed steps. |
+
+Do not substitute a screenshot for the editable project or final export the user requested. Reopen saved files, check missing assets and fonts, and inspect the actual export. For video, check audio and timing as well as a still frame. Keep originals intact and use explicit save paths. Rendering, uploading, or exporting through a paid service may need separate authorization.
+
+When the user specifically wants a CUA workflow, include PyAutoGUI in the stage that operates the desktop. For artifact-only requests, keep script-based work outside the desktop session and omit UI stages that add no useful operation or check. A desktop sequence is:
+
+1. Prepare assets and confirm the intended application, output formats, and permissions.
+2. Open the dedicated desktop and inspect its starting state.
+3. Create or edit with PyAutoGUI and suitable native/app APIs, saving checkpoints.
+4. Reopen and inspect the deliverables, then make bounded corrections if needed.
+5. Hand off local files. Publish or upload only to an authorized target.
+
+Pass scripts, project files, and artifact paths between stages rather than long click transcripts. Give one stage exclusive desktop ownership and use finite deadlines for renders and exports. Stop on unexpected dialogs, lost focus, missing permissions, or failed observations. On resume, inspect the app and files before repeating an action.
+
+See [workflow authoring](/workflows/authoring) for stages and human-input gates. Use durable `ctx.tool` calls for workflow-owned external operations, with finite timeouts and cancellation; model stages can use the appropriate automation tools to operate the app. If the user asks to work inline, keep the same safety and deliverable checks without creating a workflow.
+
+## Troubleshooting
+
+| Symptom | What to check |
+| --- | --- |
+| uv or another command is missing | Install it when permitted, refresh PATH, and check its version in the same shell that will launch automation. |
+| Black screenshot or no desktop | Check screen permissions, display/session ownership, X11 versus Wayland, and remote-session state. |
+| Input reaches the wrong app | Stop. Confirm focus, window identity, scaling, and that no other controller shares the desktop. |
+| Browser element reference no longer works | Take a fresh snapshot and locate the current control. |
+| Herdr binary exists but control is unavailable | Check explicit request, managed-pane context, and client/server compatibility. Use a fallback rather than replacing the server. |
+| Save/export timed out | Inspect the file and app state before retrying. Preserve partial output for diagnosis. |
+| Install or graphical access is blocked | Continue work that can be done safely with available APIs or shell tools, and state what remains unverified or unfinished. |
+
+A tool being unavailable is a reason to choose another supported mechanism or report a limitation, not to invent a successful interaction.

package/docs/docs.json

@@ -17,6 +17,7 @@
           "index",
           "quickstart",
           "usage",
+          "computer-use",
           "providers",
           "environment-variables",
           "llama-cpp",

package/docs/extensions.md

@@ -1074,7 +1074,7 @@ The workflows extension registers a publisher on activation and publishes this a
 | `workflow_stage_completed` | `WorkflowStageCompletedEvent`: the lifecycle envelope with a stage target whose status is `completed`. Shares the lifecycle event ID and cursor. Failed, skipped, cancelled, and killed outcomes do not produce this hook. |
 | `workflow_heartbeat` | `WorkflowHeartbeatEvent`: run/root/owner identity, `scheduledAt`, and `intervalMinutes`. Observation only, with no scheduler or cadence change. |
 
-Run control actions describe the caller's request: an already-aborted caller signal still emits `kill` after run registration, and a whole-run interrupt at a task-result checkpoint emits `interrupt` even though graceful suspension retains the existing paused outcome and `exitReason: "quit"`. A control event alone does not mean execution has drained.
+Run control actions describe the caller's request: an already-aborted caller signal still emits `kill` after run registration, and a whole-run pause at a task-result checkpoint emits `pause` while graceful suspension retains the paused outcome and `exitReason: "quit"`. A control event alone does not mean execution has drained.
 
 Use `ctx.observeWorkflowActivity` for status consumers. Registration captures a snapshot atomically with attaching the observer. Delivery is asynchronous, snapshot first, then FIFO updates. Each callback finishes before the next callback for that observer starts; a slow observer does not delay the publisher or other observers.
 
@@ -1199,6 +1199,8 @@ const response = await ctx.modelRegistry.complete(
 
 Use `getApiKeyAndHeaders()` only when an extension must inspect auth before dispatch; normal requests do not need to resolve or overlay auth themselves.
 
+OpenRouter Chat Completions and Anthropic Messages requests send `x-session-id` by default when `sessionId` is supplied and prompt caching is enabled. Set the model's `compat.sendSessionAffinityHeaders` to `false` to opt out, or set `cacheRetention: "none"` on the request to disable cache-related affinity. Explicit request headers override generated headers.
+
 `await ctx.modelRegistry.refresh(options)` returns `{ aborted, errors }`, not just completion. `errors` is a per-provider map, so extensions can report partial refresh failures; `aborted` reports cancellation. Host integrations that call `ModelRuntime.setRuntimeApiKey(providerId, apiKey, options)` must note that it records the runtime credential but does not refresh the catalog; call `refresh({ providers: [providerId], signal })` explicitly when a fresh catalog is needed.
 
 `ctx.scopedModels` is the read-only list of models scoped to the current session — the same set the `/scoped-models` command shows. It is resolved from the `--models` CLI flag and the `enabledModels` setting, matched against the available catalogue. It is empty when no scoping is configured, meaning every available model is usable. Each entry is `{ model, thinkingLevel? }`, where `thinkingLevel` is set only when a pattern pinned it (for example `anthropic/*:high`). Use it to populate a model picker that mirrors the built-in one instead of enumerating the whole catalogue.
@@ -2368,6 +2370,14 @@ pi.registerTool({
 });
 ```
 
+### Fireworks deferred tool loading
+
+Extensions making requests directly through `@bastani/pi-ai` can use native deferred tool loading with Fireworks `anthropic-messages` models. Supply the tool definitions in `context.tools` and record newly loaded tool names in the loader result's `addedToolNames` field. The provider serializes deferred definitions with `defer_loading` and inserts `tool_reference` content at the load point.
+
+Name the loader `ToolSearch` or `tool_search` to keep deferred schemas out of the initial prompt prefix. Other names work, but Fireworks includes the schemas in the prefix and loses that cache benefit. Fireworks GLM models and Kimi K3 still use Chat Completions; this feature does not change their routing.
+
+This is an AI SDK capability. Atomic's `pi.setActiveTools()` updates the active tool list but does not automatically populate `addedToolNames`. See the [AI SDK deferred tool-loading guide](https://github.com/bastani-inc/atomic/blob/main/packages/ai/README.md#fireworks-deferred-tools) for details.
+
 ### Overriding Built-in Tools
 
 Extensions can override built-in tools (`read`, `bash`, `powershell`, `edit`, `write`, `find`, `search`, `ask_user_question`, `todo`) by registering a tool with the same name. Interactive mode displays a warning when this happens.

package/docs/herdr.md

@@ -2,6 +2,8 @@
 
 Atomic reports its status to [Herdr](https://herdr.dev) automatically when you launch it in a Herdr pane. No extra extension is required. The pane identifies the agent as `atomic`.
 
+For interactive terminal automation, pane isolation, and tool fallbacks, see [Computer use](/computer-use#terminal-automation-with-herdr). This page covers Atomic's automatic status integration.
+
 ## Setup
 
 Use Herdr 0.8.2 or newer and launch Atomic inside it. Herdr supplies the pane connection settings automatically; you do not need to configure them yourself.

package/docs/index.md

@@ -59,6 +59,7 @@ For the full first-run flow, see [Quickstart](/quickstart).
 
 - [Quickstart](/quickstart) - install, authenticate, and run a first session.
 - [Using Atomic](/usage) - interactive mode, slash commands, context files, and CLI reference.
+- [Computer use](/computer-use) - work in desktop apps, browsers, and terminals across macOS, Linux, and Windows.
 - [Providers](/providers) - subscription and API-key setup for built-in providers.
 - [Environment variables](/environment-variables) - Atomic/Pi aliases, provider credentials, and bash session metadata.
 - [Security](/security) - project trust, sandbox boundaries, and vulnerability reporting.

package/docs/intercom.md

@@ -119,7 +119,9 @@ See auth.ts:142-156.
 
 The reply hint (enabled by default) points to `intercom({ action: "reply", ... })`, so recipients never need raw sender or `replyTo` IDs. Idle recipients get a new turn immediately; busy interactive recipients receive the message once they go idle. Attachment content is included in the agent-visible body, and messages are rendered inline and stored in Atomic session history.
 
-A busy non-interactive recipient can refuse a message without interrupting its task. A successful `send` receipt acknowledges transport delivery, not acceptance by the recipient's model. The refusal carries the original reply thread: a waiting `ask` returns an error; otherwise the sender sees **Intercom delivery failed** feedback with a `Sent:` timestamp. That feedback bypasses the ordinary idle queue and does not trigger a standalone agent turn. During an active turn, protected delivery makes it visible and reconciles it at a protocol-safe boundary. Its wording describes the refused send, not the recipient's later activity.
+Working subagents and live workflow stages treat `send` and `ask` as a priority interrupt queue. The recipient's current model call or cancellable tool is cancelled immediately, and the message is processed next within the same task, session, and stage generation; Intercom never launches another task or repeats the original prompt. A tool that ignores cancellation finishes first, and completed side effects are kept rather than undone or replayed. This works with foreground and background subagents. Messages received during startup join the original task, multiple messages retain arrival order, and an ask keeps its exact reply correlation after the cancelled turn. Explicit `interrupt`, owner cancellation, host stop, and terminal children or closed stages still win over later input. Use an exact connected child name or full session ID from `intercom list`; subagents are not workflow-stage paths.
+
+A busy non-interactive recipient that is neither an admitted subagent nor a workflow stage can still refuse a message without interrupting its task. A successful `send` receipt acknowledges transport delivery, not acceptance by the recipient's model. The refusal carries the original reply thread: a waiting `ask` returns an error; otherwise the sender sees **Intercom delivery failed** feedback with a `Sent:` timestamp. That feedback bypasses the ordinary idle queue and does not trigger a standalone agent turn. During an active turn, protected delivery makes it visible and reconciles it at a protocol-safe boundary. Its wording describes the refused send, not the recipient's later activity.
 
 Atomic treats ordinary `intercom` as a mandatory runtime tool in main chat and every workflow model stage. Tool allowlists, exclusions, `noTools`, optional-extension restrictions, and reloads cannot unload or deactivate it. Restrictions on every other tool are unchanged, and `contact_supervisor` remains subagent-only. Tool registration is lightweight; broker connection and heavy initialization remain lazy until an Intercom surface is used.
 
@@ -137,6 +139,12 @@ The session list and ALT+M picker show connected agent sessions, not every open
 
 Name sessions with `/name` so they can target each other (for example `/name planner` and `/name worker`). If a session is unnamed, Intercom exposes a runtime-only fallback alias like `subagent-chat-1a2b3c4d-1111-4222-8333-123456789abc` so other sessions can still target it. That alias is not persisted as the session title, so resume pickers keep showing the transcript snippet instead of a generic name.
 
+### Troubleshooting initialization
+
+`Intercom heavy initialization failed; a later call will retry: …` means initialization can be attempted again on a later Intercom call. Interactive sessions show this as a yellow warning in the chat pane, without a console stack trace; non-interactive sessions (print, JSON, and RPC) retain console diagnostics. Terminal relay and cleanup failures appear as error notifications in interactive sessions.
+
+If initialization keeps failing, check the reported cause and `~/.atomic/agent/intercom/broker.log` (or the Intercom directory under `ATOMIC_CODING_AGENT_DIR`). Do not automatically resend an operation reported with an unknown delivery outcome; check with the recipient first.
+
 ## The intercom Tool
 
 | Parameter | Type | Description |
@@ -392,7 +400,7 @@ workflow({
 
 When neither `enabled` nor `delivery` is set, direct `parallel` runs default to `control-and-result` when Intercom is available; otherwise delivery is off. Treat Intercom payloads from direct runs as user-visible workflow output.
 
-While a workflow stage generation is open, incoming Intercom messages are admitted through the stage session's native steering/follow-up queue. Parallel child asks, sends, and supervisor requests use destination-side reservation and the exact-child probe/commit observation-yield handshake, so terminal stage close cannot overtake an admitted delivery. A destination-side admission failure returns a correlated actionable error to a blocking asker instead of waiting for the 10-minute reply timeout. Claimed single-child parent handoffs remain source-side terminal handoffs.
+While a workflow stage generation is open, incoming Intercom messages are admitted as priority input: the stage's current model call or cancellable tool is cancelled and the message is processed next in the same stage generation. Parallel child asks, sends, and supervisor requests use destination-side reservation and the exact-child probe/commit observation-yield handshake before that cancellation, so a child's own message releases the parent's foreground observation rather than cancelling the child, and terminal stage close cannot overtake an admitted delivery. A destination-side admission failure returns a correlated actionable error to a blocking asker instead of waiting for the 10-minute reply timeout. Claimed single-child parent handoffs remain source-side terminal handoffs.
 
 ### Subagent Control Notices
 

package/docs/providers.md

@@ -118,7 +118,7 @@ Run `/login xai`, then select **Use a subscription**. `XAI_API_KEY` remains avai
 
 Atomic defaults xAI sessions to `grok-4.6`. Built-in workflow and subagent fallback chains use `xai/grok-4.6:xhigh`, `github-copilot/grok-4.6:xhigh`, and `openrouter/x-ai/grok-4.6:xhigh`; GitHub Copilot also exposes Grok 4.6 when the account's model policy enables it. Network-backed catalogs refresh and cache these newer entries independently of the bundled catalog snapshot.
 
-The `codebase-locator`, `codebase-pattern-finder`, and `codebase-research-locator` agents use GPT-5.6 Luna at `xhigh` and Grok fallbacks at `medium` instead. Goal and Ralph orchestration, Ralph research, and Open Claude Design use GPT-6 Astra at `medium`; Ralph prompt refinement remains at `high`.
+The `codebase-locator`, `codebase-pattern-finder`, and `codebase-research-locator` agents use GPT-5.6 Luna at `xhigh` and Grok fallbacks at `medium` instead. Goal and Ralph orchestration, Ralph research, and the debugger use GPT-6 Astra at `medium`; Ralph prompt refinement remains at `high`. Open Claude Design starts with Anthropic Fable 5.1 at `medium`, then GitHub Copilot Fable 5.1 and Codex, Copilot, and OpenAI Astra at `medium`.
 
 ### Radius
 

package/docs/quickstart.md

@@ -218,7 +218,7 @@ Named workflow runs execute in the background. After launch you get the full run
 /workflow status                  # list this session's active and terminal runs
 /workflow connect <run-id>        # see agents working; chat with or steer each stage (F2 also opens latest)
 /workflow attach <run-id> <stage> # chat with one stage
-/workflow interrupt <run-id>      # pause resumably
+/workflow pause <run-id>          # pause resumably
 /workflow resume <run-id> "go"    # send a steer message and resume
 /workflow quit <run-id>           # pause gracefully and keep the run resumable
 ```

package/docs/sdk.md

@@ -795,7 +795,7 @@ const { session } = await createAgentSession({ resourceLoader: loader });
 
 Specify which tools to expose by name:
 
-- Built-in tool names enabled by default: `read`, `bash`, `edit`, `write`, `find`, `search`, `ask_user_question`, `todo`
+- Built-in tool names enabled by default: `read`, `bash`, `kill`, `edit`, `write`, `find`, `search`, `ask_user_question`, `todo`
 - `find` discovers filesystem paths by glob; `search` searches file contents with regex patterns across files, directories, globs, and internal URLs.
 - `tools` is an allowlist: when provided, only the listed built-in, extension, and custom tool names are exposed, plus mandatory ordinary `intercom`.
 - `excludedTools` is a blocklist: matching built-in, extension, and custom tool names are omitted from the final registry and active tool set, except mandatory ordinary `intercom`. If both are provided, `tools` is applied first and `excludedTools` subtracts from it.
@@ -832,6 +832,14 @@ const { session } = await createAgentSession({
 Atomic's built-in `bash` tool matches upstream pi: when `bash` is enabled, commands execute through the configured shell with the Atomic process permissions. Use `tools`, `excludedTools`, or `noTools` to decide whether a session exposes the `bash` tool at all. Atomic no longer provides a command-level allow/deny option for `bash`; use an operating-system/container sandbox or a custom tool/extension when you need command allowlisting or stronger isolation.
 
 
+#### Waiting for existing shell tasks
+
+Both Bash and PowerShell factories accept `{ action: "wait", id: taskId, budgetMs: 1000 }` with a trusted `taskOwner` binding. No command is executed. `BashToolInput` and `PowerShellToolInput` distinguish command launches from existing-task waits; narrow by `action` before reading `command`.
+
+`budgetMs` is optional, finite, and non-negative. Omission uses the owner's command wait policy and zero polls. Results keep the `WaitOutcome` in `details.observation`, available exit information in `details.exitCode`, and retained output in text content. Failure and cancellation metadata remain in the settled observation. Yielded waits advance through bounded retained-output pages for the same owned task, even when the tool is recreated. Partial UTF-8 characters continue on the next page. Settled waits return all retained output again, subject to labelled gaps and truncation. Aborting the call releases only its observation. A binding from `AgentTaskHost.ownerBinding` also releases waits for incoming owner messages.
+
+Do not mix wait arguments with launch fields. Unknown or foreign IDs and unbound waits are rejected before execution hooks. Custom `operations.exec` does not provide existing-task ownership. See [Background tasks](/background-tasks) for examples and lifetime rules.
+
 #### PowerShell tool behavior
 
 `createPowerShellTool()` and `createPowerShellToolDefinition()` provide the same tool used by interactive sessions. When their default local operations execute on native Windows, they prefer `pwsh.exe`, fall back to `powershell.exe`, and throw a clear error when neither executable is available. `createLocalPowerShellOperations()` and `getPowerShellConfig()` are also exported for custom integrations. The PowerShell factories expose the current `ATOMIC_*` and legacy `PI_*` session snapshot by default; set `exposeSessionEnvironment: false` to opt out.
@@ -910,6 +918,8 @@ If you pass `tools`, include each custom or extension tool name you want enabled
 
 Factory-created `createBashTool()` instances receive the same execution-time `ATOMIC_SESSION_*`/`PI_SESSION_*` model and session snapshot as the built-in bash tool. Set `exposeSessionEnvironment: false` only when the subprocess must not receive it. `MessageRenderOptions.outputPad` is likewise passed to normal and isolated custom message renderers.
 
+Normal sessions also expose `kill({ id: taskId })` for their owned bash and PowerShell background tasks. Include `kill` when using a `tools` allowlist if the agent should be able to stop those tasks. The exported `createKillTool` and `createKillToolDefinition` factories accept `KillToolOptions.taskOwner`, a trusted execution-time callback returning the same owner binding used by shell launch. Without a binding they reject execution. The result details preserve the supervisor's `CancelReceipt`, including its decision, execution outcome, and cleanup state. See [Background tasks](/background-tasks#stop-a-shell-task-from-a-tool-call).
+
 #### Structured output final results
 
 `structured_output` is not registered in normal agent sessions by default. Add it only when a caller needs a machine-readable final-answer contract by registering the exported factory as a custom tool:

package/docs/settings.md

@@ -188,6 +188,7 @@ On a genuine first run, Atomic previews available themes and asks whether to opt
 | `compaction.compression_ratio` | number | `0.5` | Fraction of compactable transcript **lines to keep** (`0 < value < 1`) |
 | `compaction.preserve_recent` | number | `2` | Exact number of newest context-visible messages kept outside the compactable region; `0` keeps none |
 | `compaction.query` | string | last user message | Optional relevance focus for selecting older lines to retain |
+| `compaction.modelOverrides` | object | `{}` | Exact `"provider/modelId"` keys with optional `reserveTokens` and `preserve_recent` overrides |
 
 ```json
 {
@@ -203,6 +204,25 @@ On a genuine first run, Atomic previews available themes and asks whether to opt
 
 The model emits numbered line ranges only; Atomic reconstructs retained text mechanically. `preserve_recent` is enforced client-side and is not a provider parameter. Atomic does not widen this exact message count to a user-turn boundary or force a final logical turn to remain outside compaction.
 
+Per-model budgets use exact, case-sensitive provider/model IDs (including any slashes in the model ID), not patterns or reasoning suffixes:
+
+```json
+{
+  "compaction": {
+    "reserveTokens": 16384,
+    "preserve_recent": 2,
+    "modelOverrides": {
+      "anthropic/claude-sonnet-4-5": { "reserveTokens": 32768, "preserve_recent": 4 },
+      "openai/gpt-5": { "preserve_recent": 0 }
+    }
+  }
+}
+```
+
+Each field resolves independently: the active model's override, then the ordinary compaction setting, then the built-in default. Global and trusted-project settings merge overrides per model and per field. Model switches take effect on the next compaction check or manual call; borrowing a fallback planner does not change the selected budgets. Explicit manual `preserve_recent` parameters still take precedence.
+
+Both fields must be non-negative safe integers, including ordinary settings. Invalid ordinary values are reported even if a model override exists; malformed matching entries and invalid override values are reported when that model is used. `enabled`, `compression_ratio`, and `query` remain ordinary settings, not per-model overrides. Unlike upstream pi's token-based recent-history budget, Atomic uses the exact-message `preserve_recent` setting, not `keepRecentTokens`, and retains its verbatim line compactor.
+
 ### Branch Summary
 
 | Setting | Type | Default | Description |