Desktop application tasks

A desktop application export can carry one task: a prompt for a computer-use agent and checks that grade what the agent left in the application's files. An agent, not REA, writes the prompt and the checks from the change the capture recorded. REA validates the task against the capture and keeps it only when its checks fail on the seed and all pass on the captured end state. REA runs no model and no code the agent wrote: the checks are declarative, and the exported verifier evaluates them the same way REA does.

rea describe-desktop-application-task ./mousepad-capture --format json > draft.json
# edit the draft's task into task.json: a name, the user's request, the checks
rea export-desktop-application ./mousepad-capture ./mousepad-bundle --task task.json
agent-env run ./mousepad-bundle --task add-milk --agent claude-code

Describe the change

rea describe-desktop-application-task <capture-folder> reads the capture and its state view and returns Evidence with operation describe_desktop_application_task and predicate type rea.desktop-application-task-draft:

• application: the environment name the export would use, the title and the window class.
• owned_paths: the paths the export owns, the only ones whose files the app's data/get serves and checks can read.
• checkpoints: the two checkpoints compared, baseline for the seed and final for the end. They are the capture's first and last unless --seed-checkpoint and --end-checkpoint name others, as for the export, which takes the same options; a name the capture does not hold, or an end before the seed, is refused.
• files: each file under them whose state differs between those checkpoints, with its home path, role, its versions at both (baseline and final, the last recorded at or before each) and the net change between them, from the state view's own diff (lines, values, keys or rows with what was added, removed and changed, or created, deleted, format, content, unreadable; content also when the views hold only part of a file and match while the file changed). A database is compared by its rows, since its hash leaves out the -wal file. Text changes list each line once and add line_counts, the lines found at both ends whose number of occurrences changed. Views appear only for the files the suggestions cover, with arrays past 200 items and strings past 2,000 characters cut (view_cut).
• task: a draft with the name change, a placeholder prompt the export refuses, and suggested checks derived from the net change of the files with role document, or of the app's own state files when no document changed. A text file gains lines_present for its new lines and lines_absent for lines it lost (when either is at most ten), and line_count for each line holding a letter whose count changed; JSON, XML and key files gain value_equals and value_absent by pointer; SQLite tables gain rows added, changed and removed, by primary key, and their new row count. A zip archive gains checks inside the JSON, key file and XML views of its entries, by selectors and never through an element by position: an element added whole that it exists, one removed whole that it is absent, then values changed, added and removed, an entry or a JSON value added whole through its members, and an empty array or object as it is. Values that look like a time the app stamps are left out.

The suggestions are a starting point. Keep the ones that say what the user did; drop the ones that only follow from how the app saves its file. A LibreCAD capture that added four holes to a drawing rewrote it from DXF R12 to DXF 2007, so most of its lines changed; the draft suggests counts of SECTION, ENDSEC and CIRCLE, and only CIRCLE going from 5 to 9 is the user's change.

The task file

{
  "name": "add-milk",
  "prompt": "Add buy milk to my to-do list in Mousepad and save it.",
  "checks": [
    {
      "id": "milk",
      "kind": "lines_present",
      "file": "Documents/todo.txt",
      "lines": ["buy milk"]
    }
  ]
}

name is a portable bundle name other than smoke (lowercase letters, digits and hyphens); prompt is what the user asks for; checks holds 1 to 32 checks. Each check has a kind, a file given as a path under the home such as Documents/todo.txt, and an optional id (check-1, check-2 and so on by position otherwise). A check reads the typed view data/get returns for that file, in the shapes of the state view.

Kind	Fields	Passes when
file_present		The file exists
file_absent		The file does not exist
lines_present	lines	Every listed line is a whole line of the text view
lines_absent	lines	None of the listed lines is
line_count	exact or contains; equals, at_least or at_most	The number of lines equal to exact, or holding contains, compares so
value_equals	pointer, value; optional tolerance	The pointer names a value equal to value, or within tolerance of it
value_absent	pointer	The pointer names no value
row_present	table, where	A row of the SQLite table has every column in where equal
row_absent	table, where	No row does
row_count	table; optional where; equals, at_least or at_most	The number of rows, or of rows matching where, compares so

The rules both evaluators follow:

• An absent file reads as empty: no lines, values or rows. So lines_absent, value_absent and row_absent pass on it and a count is 0.
• A file that has no view (binary, over the read limit, or unreadable) fails every check but file_present and file_absent. Line checks need a text view and row checks a SQLite view; another format fails.
• Equality is JSON equality: numbers by value, never a boolean for a number, objects by their keys whatever their order, arrays in order.
• Pointers follow RFC 6901 (~1 for /, ~0 for ~) and must start with /; on an array a segment may be a selector, below. Key files read as {"groups": {"GROUP": {"KEY": "VALUE"}}}, so a setting is /groups/GROUP/KEY; XML as {"root": {"name", "attributes", "children"}}; SQLite as {"tables": {"TABLE": {"columns", "rows"}}}; a zip archive as {"entries": [{"name", "format", "view", …}]}.
• With tolerance, a non-negative number, value must be a number, and the value found passes when it is a number, or a string that is a decimal numeral such as "5.0000000000000000", at most tolerance from value. tolerance: 0 compares numerically, so "1.2500" equals 1.25. Anything else fails as not a number.
• contains matches a line that holds it anywhere, case and all. Checks take no regular expressions: a pattern run on what the agent wrote could stall the verifier, and JavaScript and Python read some patterns differently.
• A check decided by what a view leaves out (lines_absent, value_absent, row_absent, the counts, and value_equals of a whole object or array) fails when that view is truncated; a check satisfied by what the view holds is decided on it.

The export refuses, naming the check, a file outside the owned paths, a file spelled with a leading slash, . or .., a pointer into a file that is text at the seed, as the bundle ships it, or at the end checkpoint, the empty pointer (the whole file), a value that names the captured home, or the capture host's name or user@host anywhere (which the export removes from the seed), a tolerance with a value that is not a number, and a value that looks like an ISO 8601 date and time. Other volatile content, such as a counter the app bumps on every save, is for the author to leave out.

Pointers and selectors

An app that rewrites a document moves things: FreeCAD 1.0 records more properties than 0.21, so a hole's radius sits at another position among its siblings, and a plate's entry moves once a hole is added before it. A pointer segment applied to an array may therefore name a member by what it holds rather than by its position:

• [KEY=VALUE] names the member whose string field KEY is VALUE, such as a zip entry by [name=Document.xml] or an XML element by its tag, [name=Object].
• [@KEY=VALUE] names the member whose attributes hold KEY="VALUE", as the XML view keeps an element's attributes.
• Conditions can follow one another, [name=Object][@name=Hole], and all must hold. They compare strings only, the selector is escaped like any pointer segment (~1 for /), and a value cannot hold ][.

A selector must name exactly one member. When several match, or none does before the last segment, the check fails and says which selector matched what. When the last segment's selector matches none, the pointer names no value: value_absent passes, which is how a check says an element was removed, and value_equals fails. On an object, a bracketed segment is an ordinary key.

A FreeCAD task that checks the user cut a 10 mm hole through the plate, in FreeCAD 0.21 and 1.0 alike:

{
  "name": "cut-hole",
  "prompt": "Cut a 10 mm hole through the middle of the plate and save it.",
  "checks": [
    {
      "id": "cut",
      "kind": "value_equals",
      "file": "Documents/plate.FCStd",
      "pointer": "/entries/[name=Document.xml]/view/root/children/[name=Objects]/children/[name=Object][@type=Part::Cut]/attributes/type",
      "value": "Part::Cut"
    },
    {
      "id": "radius",
      "kind": "value_equals",
      "file": "Documents/plate.FCStd",
      "pointer": "/entries/[name=Document.xml]/view/root/children/[name=ObjectData]/children/[name=Object][@name=Hole]/children/[name=Properties]/children/[name=Property][@name=Radius]/children/[name=Float]/attributes/value",
      "value": 5,
      "tolerance": 0.01
    }
  ]
}

Before the edit the archive has no Part::Cut and no Hole, so both checks fail on the seed; after it, Radius is "5.0000000000000000" and both pass. A check decided by what a view leaves out depends, in an archive, on the entry its pointer goes into: it fails when that entry's view is missing or truncated, or when the entry is not listed and the listing is truncated. A selector must see every member to know it names one, so a check whose pointer searches an array the view may hold only part of fails as undecided: any array of a truncated XML view, of a truncated entry's view, or the listing when it is cut. A position is decided on what the view holds. Like a file, an entry that has no view (binary, not read, or a database or archive inside the archive) fails every check into its view; its null there is no value, while an entry holding the JSON document null reads as null. Checks on its name, size, sha256 and format still read.

XML that uses namespaces is named as the document spells it, in both evaluators: an element or attribute by its prefixed name, such as text:p or inkscape:label, an element in a default namespace by its bare name, such as svg, and namespace declarations as the attributes xmlns and xmlns:PREFIX. Text is a string among an element's children. These checks hold a poster's headline in an Inkscape SVG and a date in a LibreOffice Writer document:

[
  {
    "id": "headline",
    "kind": "value_equals",
    "file": "Pictures/poster.svg",
    "pointer": "/root/children/[name=g][@inkscape:label=Layer 1]/children/[name=text][@id=headline]/children/[name=tspan]/children/0",
    "value": "Summer Bake Sale"
  },
  {
    "id": "date",
    "kind": "value_equals",
    "file": "Documents/flyer.odt",
    "pointer": "/entries/[name=content.xml]/view/root/children/[name=office:body]/children/[name=office:text]/children/[name=text:p][@text:style-name=Subtitle]/children/0",
    "value": "Monday 24 November"
  }
]

Inkscape writes a line of text as a tspan inside its text element, and Writer keeps a paragraph's style in text:style-name. A prefix is the one the document uses, so a check names the prefixes the app writes.

Validation and export

rea export-desktop-application <capture> <directory> --task <task.json> evaluates the checks on the capture's state view: on each file as it was at the seed checkpoint, which is the seed, as the bundle ships it with the capture host's name and user removed, and at the end checkpoint, the captured end state. Those are the capture's first and last checkpoints unless --seed-checkpoint and --end-checkpoint name others. It refuses a task whose every check passes on the seed, and a check that fails on the end state, with that check's reason. A capture yields one task, so the option is taken once. The result's task names the task, its number of checks and those that fail on the seed. A refused task writes no bundle.

The bundle then holds, besides the app and tasks/smoke.json:

Path	Contents
tasks/<name>.json	Deploy the app, deploy the agent and its skills, record the screen, prompt the agent, run the verifier, stop recording
artifacts/<name>-check/verify.py	Reads the app's data/get through the computer and evaluates the checks, one result per check
evals/<env>.toml	The task, so a run of the bundle runs it; the smoke task runs with --task smoke

The verifier finds the app among the computer's child environments, calls its data/get under the run's grant of the computer's control token, and reads the state data part. It raises when the app cannot be reached, so the run fails rather than scoring the agent. The README lists the task, its prompt, its checks and the recipe below.

Solve rate

REA does not run agents. To measure how often an agent solves the task, run it k times, three by default, and count the runs whose summary line says passed:

for run in 1 2 3; do
  agent-env run ./mousepad-bundle --task add-milk --agent claude-code --model MODEL
done

The agent defaults to Claude Code and the model is the caller's choice (Claude Sonnet 5.5 is a reasonable reference). A run reported as scored below 1 failed the checks and counts as a miss; one reported as failed did not finish, as when the computer or the app could not be reached, and is run again rather than counted. Each run's instance keeps the verifier's results per check under verifications.