Files
reverse-engineering-anything/docs/desktop-application-capture.md
T
Edgar ArakelyanandClaude Opus 5.5 970969cb53 feat(desktop): capture and export apps installed from a downloaded .deb (#8954)
* feat(desktop): capture a Linux desktop application in use

Add `capture-desktop-application`, a CLI-only, Linux and X11 capture that
launches an installed desktop application, watches it while it is used and
writes capture Evidence plus SHA-256 addressed content:

- inventory: desktop entry, launch command, executable, package and its
  apt sources (or snap, flatpak, AppImage), Electron markers and candidate
  state directories;
- session: the launched process tree, client windows from xprop with the
  first application window time and a window match, and files held open;
- state: a stat walk of the home directory before launch, at checkpoints
  read from stdin, at stop and after exit, with created, modified, deleted
  and retyped paths and the content of changed files and state directories.

It refuses non-Linux hosts, missing DISPLAY or xprop, a non-empty output
folder and an application whose window is already open, and signals only
processes it started.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): own launched processes, confirm walk coverage and keep cancelled captures

- Launch through the owned provider path with a run token; release with
  SIGTERM, then the token-verified group cleanup, after every session.
  Processes without the token are never signalled; each records whether it
  carries it, and the session records whether cleanup was verified.
- Report creations and deletions only where a walk listed the nearest
  existing directory; count the rest as unconfirmed.
- Retain a file only when it is still the walked version, through owned
  no-follow reads; charge the budget only for new content; create files
  exclusively in a private output folder owned by the caller.
- Walk content roots outside the home directory, compare paths in their
  physical spelling, and run executables from the caller's directory.
- Refuse when the window list cannot be read, re-evaluate window ownership
  each sample, handle one queued request per sample with a bounded stdin
  queue, and write the capture with end reason `cancelled` on interruption.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep owned cleanup failures and fresh digests

- Allow a capture cancelled before any requested checkpoint to hold only the
  baseline.
- Retry an incomplete stop once; if cleanup is still unverified, write the
  capture and fail with the run token and process group as cleanup
  resources.
- Classify path changes after a walk (ENOTDIR and other selected-file
  errors) as changed_since_walk instead of failing the capture.
- Treat a non-directory ancestor as proof that nothing existed beneath it.
- Report a before digest only for the file version it was read from.
- Derive stderr_truncated from stderr's own byte counts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* wip(desktop): state view over a capture folder

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep failed file-handle closes owned for retry

Content reads pass retainFailedClose, so a handle whose close fails stays
owned by the content store; the capture retries those closes after it is
written and fails with the output folder as the cleanup resource if one
still cannot be closed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(desktop): read a captured application's state

Add inspect-desktop-application-state, a CLI command that reads a capture
folder into the application's state: every retained or changed file with
its role and reasons, a typed JSON view of each version (SQLite through the
shared worker, dconf through a private dconf dump, JSON, XML, key files,
text) and the kinds of change between versions (lines, values, keys, rows).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): report process and handle cleanup failures together

When both the launched processes and a retained file's handle remain
unresolved after the capture is written, one cleanup error now lists both
reasons and both resource sets.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(desktop): export a captured application as an agentenv computer app

Add export-desktop-application, which writes an agentenv bundle with one
computer_app environment from a capture folder and its state view: the
apt package, a boot launch matched on the captured window class with a
window ready check, owned paths from the state roles, a seed of the
baseline files, data hooks that call agentenv's add/get/reset helpers
with a typed JSON view on get, a skill, a README and a deploy smoke task.
The desktop CLI commands now load their services lazily.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep state views faithful to the capture

- Read key files, XML attributes, dconf groups and SQLite tables into
  dictionaries without inherited properties, so captured keys such as
  __proto__ stay data.
- Version databases with their write-ahead log as of each checkpoint, and
  stage only digest-verified bytes for the SQLite and dconf readers.
- Detect the format per version, report format changes, keep JSON null and
  empty containers as values, keep XML text and preserved whitespace, and
  mark truncated views (including XML depth) and comparisons of them.
- Number versions and changes by checkpoint index, decide baseline
  presence from the captured type, and keep tagged SQLite integers, text
  bytes and blobs distinct.
- Carry the capture's limitations and omitted content forward, match state
  directories by their home path, index changes once per capture, keep
  failed handle closes owned, and finish the CLI through the shared owner.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep coverage, logs and handles honest in the state view

- Read coverage flags only from formats whose readers write them, so JSON
  fields named truncated stay data.
- Fail a SQLite version whose write-ahead log existed but was not retained
  or does not match its digest, instead of reading the database alone.
- Close retained content on every exit, including cancellation, and report
  a close that still fails together with the primary failure.
- Finish the CLI with the canonical operation name.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): export only what the computer can install and seed

- Refuse packages not from the Ubuntu 24.04 archive, since the computer
  installs by name from that archive.
- Start the launch command in the captured working directory.
- Respell the home path only in files the state view read as text.
- Report files that were not retained instead of dropping them silently.
- Load captured dconf keys for apps that own no files.
- Keep home-level files as owned paths; rename reserved Python class names.
- Mark truncated SQLite and text views in the class's get view, and read
  file types the way the state view does.
- Release the output tree through a resource scope, and read seeds through
  an owner closed on every exit.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep staged folders owned, symlinked state roots and XML values

- Stage database bytes in a private runtime root released through a
  resource scope, so a failed removal keeps its owner and is reported with
  the read's outcome.
- Report cleanup facts and owners from desktop failures instead of
  wrapping them as a plain adapter error.
- Record the state directories as walked, in their physical spelling, and
  match roles against both spellings, so files in a symlinked state
  directory are not mistaken for selected documents.
- Keep whitespace-only XML leaf values; drop only formatting between
  elements.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): read XML in the exported app the way the state view does

Keep whitespace-only leaf values and xml:space="preserve" text in the
generated get view; drop only formatting between elements.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): release retained handles with their owners, read older captures

- Release retained-content handles whose close failed through a resource
  scope on every exit; a release that fails again reports the exact handle
  owners. A read that finished keeps its bytes when only its close failed.
- Accept captures written before state directories were recorded, and say
  in the state view's limitations that symlinked state directories may be
  misclassed for them.
- Compute each XML element's whitespace context once instead of scanning
  siblings for every text run.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): boot without the working folder, bound XML, accept mirrors

- Try the captured working directory only when an argument is relative,
  and start the app even when that folder is missing.
- Cut the generated XML view at REA's node and depth limits, and keep one
  unreadable file from failing data/get.
- Accept Ubuntu 24.04 mirrors, not only the canonical hosts; PPAs and
  other repositories are still refused.
- Read seeds through the retained-content owner's new signature.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): state view review round

- Stage a retained -journal beside its database so the SQLite worker
  accepts an inactive journal and refuses a hot one; a -wal or -journal
  that existed but was not retained (including one omitted at the
  baseline) makes the version unreadable.
- Keep tables that read when another fails, and let cancellation and
  unfinished cleanup propagate instead of becoming view errors.
- Read text as a key file only when every line parses as one.
- Rank a file named in a window title or selected by the caller as a
  document even when the application holds it open; match titles on
  file-name boundaries; leave the display name out of dconf matching.
- Bound JSON and key file views (4 MiB, 500 levels, exact numbers) and
  report too-large, missing and altered retained content separately.
- Add no version or change while a database stays absent.
- Read staged databases in place instead of copying them again.
- Give Evidence the same limitations as the result.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): export review round

- Load captured dconf keys without D-Bus: install dconf-cli, compile the
  existing and captured key files into the user database, and warn on
  failure instead of stopping the server.
- Start relative-argument launches in the captured working directory,
  the home included, since the framework launches from the app folder;
  respell the home itself and option values under it.
- Read views in a thread, skip files over 64 MiB, align the generated
  readers with the state view's, and close SQLite connections.
- Never own the home or a shared base folder; map a symlinked state
  directory's files back to their home spelling.
- Report omitted files and unretained opened folders.
- Name the module so it cannot shadow another, and keep the class off
  the module's own names; derive names the bundle check accepts.
- Write seeds as they are read, and remove an unpublished bundle on
  failure.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): capture review round

- Ask the launch process group and token carriers to exit before
  killing them, on errors too; report incomplete cleanup with its facts.
- Count descendants of the launch as started by the capture even when
  their environment cannot be read; treat a reused pid as a new process.
- Follow symlinked desktop entries; name a flatpak's state after its
  application id; match flatpak windows by name only.
- Stop on end of input; refuse empty, duplicate and reserved checkpoint
  names; require whitespace after the checkpoint keyword.
- Stay on each walk root's device, and never walk or retain `/` or a
  directory containing the home; say so in the limitations.
- Refuse a window manager that does not publish _NET_CLIENT_LIST.
- Remove retained content when a capture fails before it is written.
- Skip reading a file larger than the remaining content budget.
- Record apt's archive origin for each package source.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): read an unset client list as no windows when supported

xfwm4 leaves _NET_CLIENT_LIST unset while it manages no windows, so the
capture refused an ordinary desktop with nothing open. Read
_NET_SUPPORTED with the list: an unset list means no windows when the
window manager supports it, and the capture is refused only when it
does not. `xprop`'s "no such atom" output now reads as unset too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): accept packages by apt's archive origin

Accept a package source when apt reports its origin as Ubuntu for a
noble suite, so mirrors count and PPAs or other repositories that merely
have "ubuntu" in their URL do not. Captures made before origins were
recorded fall back to Canonical's own archive hosts. Seed reads report
why retained content could not be used.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(desktop): capture and export apps from vendor apt repositories

A closed-source app installed from its vendor's apt repository, such as
Typora, can now be captured and exported as a computer app.

- The capture records each non-Ubuntu repository listing the installed
  version (uri, suite, components from apt's sources files, one-line or
  deb822) and the key that signs it: the signed-by key, or the trusted
  key that gpgv shows verifies the repository's signed index. The key
  is retained with the capture.
- The capture records the packages owning the libraries the executable
  loads (DT_NEEDED with RUNPATH and the linker cache, through REA's ELF
  readers), since a vendor package may not depend on them.
- The export declares a verified vendor repository as an AptSource with
  its key, installs a vendor package's library packages, and adds
  --no-sandbox to an Electron app's launch, since a container denies
  Chromium's sandbox.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): report a locked database at once in data/get

A database the running app holds locked made each data/get wait sqlite's
five-second busy timeout before reporting it. Record the Typora run in an
agentenv computer in the verification.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* perf(desktop): detect key files without splitting the whole text

Read lines one at a time and stop at the first that a key file view
would drop, instead of splitting a file of up to 64 MiB into arrays
before its format is known.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep symlinked homes and state folders under the home in exports

- Respell a path under its state folder's logical spelling only when that
  spelling is under the captured home, so a symlinked HOME keeps its
  files; pair walked and logical state folders only when the capture
  lists the same number, and match the innermost folder first.
- When two files map to one home path, keep the one present before the
  capture and name the other in the limitations.
- Read key files in the generated data/get with the state view's rule,
  checked against it by running the helper under python3.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): end captures on zombies and keep cleanup owners

- A zombie or dead process no longer keeps the session open.
- Remove only the partial capture file this run created.
- Keep the exact owners of handles and processes whose cleanup failed on
  the returned error, so a retry can release them; test the published
  capture's cleanup-failure branch with mixed outcomes and retries.
- Release the application once even when the exit checkpoint fails.
- A cancellation before the capture is written removes what it retained.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): vendor package review round

- Match apt sources through `apt-get indextargets`, which covers flat `./`
  and `/` suites and gives each release file and sources entry.
- Accept a key whose signature verifies even when the release carries
  another signature without a key here; keep only the verifying key from
  a keyring holding several (OpenPGP packet splitter).
- Install the package from its repository's release (`name/<release>`),
  declare one source per URI and suite, support only the capture host's
  architecture, and refuse Ubuntu Pro repositories.
- List only library packages the computer can install; walk libraries only
  for vendor installs, with RPATH inherited as ld.so does, normalized
  paths, and DT_NEEDED bounded while reading.
- Put --no-sandbox after the program an env wrapper launches; keep the
  gpgv folder's owner when its removal fails; respect --max-file-bytes for
  keys; wait 0.25 s on a busy SQLite database in data/get.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(desktop): keep the vendor spec's gpgv command on one line

A code span broken across lines left `<binary keyring>` for VitePress to
read as an unclosed HTML tag, so the docs build failed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(desktop): add a validated task to exported desktop applications

Each capture can now yield one agentenv task in its exported bundle. An
agent writes the task (a prompt plus declarative checks over the app's
data/get views); REA validates it and stays model-free.

- `describe-desktop-application-task` summarises a capture's net change
  per file (role, format, views before and after) and drafts checks.
- Checks: file presence, lines present/absent/counted, JSON-pointer values,
  SQLite rows and counts. REA keeps a task only if a check fails on the
  seed and every check passes on the captured end state; a file that was
  not retained at the baseline counts as absent, as the computer starts.
- `export-desktop-application --task` writes tasks/NAME.json (deploy,
  agent, skills, recording, prompt, verifier), a compiled verify.py that
  reads data/get, and an eval. A boundary test runs the verifier under
  python3 and checks it agrees with REA's evaluator on every check kind.
- Docs, spec and skill guidance, with the solve-rate recipe.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(desktop): keep the seed test in its own block

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(desktop): record the task verifier driven in an agentenv computer

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(desktop): record the Typora task run in an agentenv computer

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep task checks bounded, complete and attributable

- line_count counts lines holding a substring (contains) instead of
  matching a regular expression, which could stall the verifier on
  agent-written text and read differently in Python.
- value_equals of a whole object or array is undecided on a truncated
  view.
- The draft keeps a database whose rows changed under an unchanged hash,
  and a file whose truncated views match while it changed.
- A task file that cannot be read keeps the reader's detail, and a failed
  cleanup keeps the read failure as its primary.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): leave a file with one version out of the task draft

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(desktop): update the task verification's test counts

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(desktop): record the task solve rate with Claude Code and Sonnet 5.5

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(desktop): capture and export apps installed from a downloaded .deb

capture-desktop-application --package-file reads the file's control fields
with dpkg-deb and records its SHA-256, size and dependencies when it holds the
installed package; the file itself is never kept. The library walk now also
runs for local installs.

export-desktop-application --package-url declares agentenv's
Download(url, sha256) for that file, its dependencies and library packages as
apt packages, and a run script that installs it with dpkg.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): seed app-named dotfiles and respell UTF-16 seeds

The capture now treats dotfiles in the home named after the application
(.varicad-cfg.ini, .notesrc) as state roots, so a file the app only reads
during the capture is still seeded. The exported seeding respells the
captured home in UTF-16 text files with a byte-order mark, keeping the mark.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): hold readiness until the app's own window shows

An exported app is ready only once a window of its class that the taskbar
lists is shown, so a splash screen of the class, which skips the taskbar, no
longer counts as the app being up.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* docs(desktop): record VariCAD captured, exported and run in an agentenv computer

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep exported views strict JSON for infinite and inexact numbers

The exported data/get spells infinite and negative-zero SQLite reals as REA's
SQLite reader does ("Infinity", "-Infinity", "-0"), and refuses a JSON file
holding a number that reading would change (1e999, integers past 2^53, -0,
rounded fractions) with the state view's own error, instead of a view with
an infinity in it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep package-file facts tied to one file and URLs unechoed

- A --package-url refusal never repeats the URL, and credentials are checked
  before the scheme.
- The package file's identity must hold between dpkg-deb's read and the hash.
- A failed close keeps its cleanup owner even when the capture is cancelled.
- Readiness also ignores windows typed as splash screens.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): record database logs and keep state changes distinct

- A selected path that is not a directory brings its -wal and -journal as
  content roots, so the capture records whether they exist and retains them.
- The state view treats a file as absent only under recorded content roots
  and refuses a database whose log presence is unknown.
- JSON changes list containers too, so an array that becomes an object with
  the same members is a change.
- Key file entries escape the key as a JSON pointer segment, so [a/b] c and
  [a] b/c stay apart.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep watching through unread samples and late workers

- An unreadable window list is unknown, not zero windows, and is named in
  the limitations.
- Before ending the session, live run-token carriers no parent link reached
  are tracked; an unreadable process table leaves the exit unknown.
- An env prefix's assignments find the program and name its XDG folders.
- flatpak run options that take a value are skipped with it, -- is honoured
  and a /arch/branch suffix is dropped from the id.
- State directories include the entry's full id.
- Path options refuse empty values; roots count toward the entry limit.
- A capture file whose close failed stays owned and is retried on discard.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): bound apt sources, keep older captures and cleanup owners

- Sources files parse into one stanza per line or deb822 paragraph with its
  URIs and suites kept as lists, so no URI x suite product is built.
- Repository architectures and release are optional, so captures written
  before they were recorded parse; the export treats them as unrestricted
  and unknown.
- Library sources are matched against every index of a declared
  repository's URI, suite and components.
- Two failures side by side keep both causes and both cleanup owners, and a
  host file whose close fails reports a lease on its handle.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): match escaped key file entries and suggest containers through members

- The draft's key file lookup uses the state view's escaped `group/key`, so a
  key holding `/` or `~` still gets its suggestion.
- JSON and XML changes now list containers too; the draft suggests an added
  or removed container through its listed members, and a container only
  when none of its members is listed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): read relative PATH entries from the launch folder

A program, or an empty or relative PATH entry, is resolved from the folder the
program starts in, as the launch does, so `env PATH=. app` with the entry's
Path still finds the app.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* perf(desktop): compare JSON views on one shared path

JSON and XML changes are found by walking both trees with one shared pointer
path, building a pointer only for a listed value, so memory follows depth
rather than every container's full path. A pointer over 4096 characters is
counted but not listed. A 256-level document with 4000-character keys went
from about 251 MiB of heap to under 1 MiB.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep JSON Lines as text in the exported reader

The exported reader parses the whole text before an inexact number makes it
JSON with an error, so JSON Lines with an integer above 2^53 stay text as in
REA. Over-long integer spellings are unsafe without converting them. The
guide no longer claims large integers keep their value.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep the state-path help free of dash-led words

The CLI alias check reads `-wal` and `-journal` in help text as short flags,
so the help names the write-ahead log and rollback journal instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): never record or export apt source credentials

- Source URIs lose any user name and password as they are parsed, the way apt
  splits them, and so do index-target URIs and descriptions and the sources
  apt-cache policy lists, so matching works whether or not apt prints them.
- A repository whose entry held credentials is marked
  `uri_credentials_removed`, and the export refuses it, or an older capture
  whose URI still holds them, with a reason that names no URL.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): look programs up without normalizing their paths

Program paths and PATH entries are joined rather than resolved, so `..` after
a symbolic link means what it means to the launch. The lookup moves to its own
module.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): note inexact JSON numbers with a flag, not a list

The exported reader only needs to know whether any number was inexact, so it
keeps a boolean instead of every rejected spelling.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): refuse older captures that still hold apt credentials

A capture written before credentials were removed counted the credentialed
repository's libraries as installable, so the export refuses it whole and asks
for a new capture rather than declaring packages no source offers. Recorded
policy sources are redacted in export messages too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test(desktop): keep the credential fixture on an example host

A login on a real host reads as a secret to the commit scanner, so the older
capture's fixture uses apt.example.com like the other credential tests.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): tidy looked-up paths and search relative entries only for env

Empty and `.` segments are dropped from a looked-up program path, which keeps
`/snap/bin/` entries recognisable as snaps, while `..` is still left for the
kernel. Relative and empty PATH entries are searched only in the PATH an `env`
prefix sets, which travels with the exported launch; the host's own PATH does
not.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): keep excluded and unwalked database logs unknown

The capture records the paths every walk left out, and the state view treats
a log as absent only under content roots the baseline walked completely and
outside those exclusions, so an excluded or unvisited log leaves the database
unread rather than read without it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): judge JSON numbers after duplicate keys settle

The exported reader marks an inexact number instead of flagging it during the
parse, then looks for marks in the parsed value, so a number a later duplicate
key overwrote no longer refuses the file, as in REA.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): redact a repository's policy source when it names it

An older capture whose policy source matched no sources entry kept that text
whole, and the export named the repository by it; the export now names it
without any user name or password.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): never signal the launch group on unknown membership

When the current read cannot observe a carrier's process group, the carrier
is signalled on its own instead of letting an older process-table row grant
a group-wide SIGTERM.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): look programs up in the launch's order and refuse host-relative finds

The lookup searches PATH as the launch does, an empty entry meaning the
launch folder, so inspection and launch name the same program. A program found
only through a relative or empty entry of the host's own PATH is refused,
since the exported computer would not have it; an env prefix's PATH travels
with the command.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): count unchecked paths and keep logs behind them unknown

Each walk counts the paths it could not stat and the directories on another
device it did not enter, and the checkpoint records it as `unchecked_paths`.
The state view treats a log as absent only when the baseline left nothing
unchecked, so a log whose stat failed stays unknown.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* perf(desktop): scan parsed JSON with one iterator per level

The check for marked numbers walks the parsed value with an iterator per
nesting level instead of copying each container's children; a million-entry
array went from about 15.7 MiB to 8.1 MiB of peak traced memory.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(desktop): start a relative program from its captured folder

An exported launch whose program path is relative, such as `./app` from an
entry's Path, now changes to the captured working directory first, as it
already did for relative arguments.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 15:14:04 +08:00

19 KiB

Desktop application capture

rea capture-desktop-application launches an installed Linux desktop application, watches it while a person or a driver uses it, and records what it is, which windows it opened and how the files under the home directory changed. The result is the starting point for turning the application into an agent environment: the inventory says how to install and launch it, the windows say how to recognise it, and the checkpoints hold the state it read and wrote.

The application runs with the current user's permissions, in the current desktop session, against the user's real files. This is not a sandbox. The capture folder keeps copies of real files from the home directory, so treat it as private data and keep it local.

Host support

Capture runs on Linux with an X11 display: DISPLAY must be set, and xprop (from x11-utils) must be installed, or REA_XPROP_COMMAND must name it. The window manager must support the EWMH client list (_NET_CLIENT_LIST, listed in _NET_SUPPORTED), as xfwm4, Mutter, KWin and Openbox do; without it the capture is refused, because no window could be seen. A supported list that is unset, as xfwm4 leaves it while no window is open, means no windows. Applications that run under XWayland are visible; native Wayland windows are not. Other hosts are refused before anything is launched.

Capture an application

Name the application by its desktop entry id, the entry's absolute path, or an executable:

rea capture-desktop-application --desktop-entry org.xfce.mousepad --output ./mousepad-capture
rea capture-desktop-application --executable gnucash --argument /home/ada/books.gnucash --output ./gnucash-capture

Desktop entries are looked up in $XDG_DATA_HOME/applications and then each $XDG_DATA_DIRS directory; an entry installed as a symbolic link, as flatpak and Nix do, is followed to its file. The entry's Exec line is expanded with no files; --argument values are appended in its place, and the application runs in the entry's Path or the home directory. When the line starts with env NAME=value, the program is found on the PATH those assignments set, with an empty or relative entry read from the folder the application runs in, and their XDG folders name the state directories. Otherwise the program is looked up on the host's PATH in the order the launch searches it, and one found only through a relative or empty entry is refused, since the exported computer would not have that entry. An --executable runs from the current directory, as it would from a shell, and a relative path to it is resolved there.

Before writing anything the command refuses to start if a window of the application is already open, or if the window list cannot be read. A launch beside a running instance can hand off to it, and that instance's unsaved state would be missing from the baseline. It then records the starting state, launches the application in its own process group with a run token in its environment, and reads commands from stdin, one per line:

Line Effect
empty, checkpoint, or checkpoint <name> Record the current state as a checkpoint
any other text except stop Record a checkpoint named by that text
stop, or the end of stdin Record a final checkpoint, close the application and finish

An unnamed checkpoint is named checkpoint-<ms> after its time in the session. A name that repeats an earlier checkpoint, or is one of the capture's own names (baseline, stop, time-limit, exit), is refused with a progress line on stderr; no checkpoint is taken and the session continues.

The session also finishes when the launched process has exited and the application's processes and windows are all gone in two consecutive samples, or when --max-duration-seconds passes. Before deciding the processes are gone, the capture looks for any process still carrying the run token, so a worker whose launcher exited before it was sampled keeps the session open. A sample whose window list could not be read is not taken as an empty one; the limitations say how many there were. One queued line is handled between samples, and reading stdin pauses while 16 lines wait. Progress lines on stderr say when the window has opened and when each checkpoint is taken.

When the session ends the capture sends SIGTERM to the launched process, its process group and every process that carries the run token, waits up to five seconds for them to exit, and then removes every process that still carries the token, whether or not it stayed in the process group. If that cleanup cannot be verified after a second attempt, the capture is still written, and the command then fails, naming the run token and process group that may remain. Interrupting the command (Ctrl-C) ends the session the same way without a further checkpoint; the capture is still written with end reason cancelled, and the command exits with status 130. Interrupted before the application is launched, the command writes no capture and removes the content it had retained.

If the capture fails for any other reason before capture.json is written, the application is closed the same way and the content already retained is removed from the capture folder. The error names any process or file that could not be removed.

A person types these lines in the terminal. An agent can drive the application with a tool such as xdotool and write the lines to a FIFO that it passes as stdin. Because the end of stdin stops the session, the driver must keep the FIFO open for writing between checkpoints, for example with exec 3>capture.fifo in a shell, rather than open and close it for each line.

Option Default Meaning
--state-path none File or directory whose starting content is retained too (repeatable)
--exclude none Path left out of every checkpoint walk (repeatable); the output folder is always left out
--max-duration-seconds 3600 Stop the session after this long
--max-entries 250,000 Paths recorded per checkpoint walk, its roots included
--max-file-bytes 64 MiB Largest file whose content is retained
--max-content-bytes 1 GiB Total retained content
--package-file none The .deb file the application was installed from; its facts are recorded, not the file

--state-path, --exclude and --output refuse an empty value, which would otherwise name the current directory.

The capture folder

The output folder must be new, or empty and owned by the current user; the command makes it private (mode 0700). It holds:

  • capture.json: the Evidence record, also printed on stdout.
  • content/<sha256>: the bytes of each retained file, stored once per distinct content.

The Evidence operation is capture_desktop_application and its predicate type is rea.desktop-application-capture. Its result has four parts.

Application. The desktop entry (id, path, name, Exec, icon, StartupWMClass, categories, MIME types), the launch command and working directory, the executable as found on PATH and after resolving links, and the package that installed it. A Debian package records its name, version, architecture and the apt sources that list the installed version, with the origin (o=) apt reports for each source in source_origins, such as Ubuntu for the Ubuntu archive and its mirrors or LP-PPA-<name> for a PPA; an empty source list means the version came from a local .deb or the package lists are missing, which package_lists_present tells apart.

Each source whose origin is not Ubuntu is also recorded in apt_repositories. apt-get indextargets names the sources file entry apt read it from and its release, so flat repositories match too, whether their suite is ./ (Typora's deb https://downloads.typora.io/linux ./) or / (pkgs.k8s.io's). The record holds that entry (sources_file, uri, suite, components, none for a flat repository, and architectures, the entry's arch= or Architectures: restriction, empty when it has none), release, the release's Suite (apt's a=) or else its Codename, and the key that signs it. A user name or password in an entry's URI (https://user:password@host/…) is never recorded: REA removes it from every URI it reads, in the sources file and in apt-get indextargets and apt-cache policy output alike, so entries match whether or not apt prints it, and sets uri_credentials_removed on a repository whose URI held one (a capture written before this was recorded omits the field). The key is found by verification, not by name: the candidates are the entry's signed-by keyrings, or else every .gpg and .asc key in /etc/apt/trusted.gpg.d and /etc/apt/trusted.gpg, and each is checked with gpgv against the release's InRelease (or Release with Release.gpg) beside its index in apt's lists folder. A key counts when gpgv reports a valid signature by it and no bad, expired, revoked or unverifiable signature by it, whatever gpgv's exit status: a release also signed by a key the host lacks makes gpgv fail. Each candidate is copied, an armored key decoded to a binary keyring, into a private temporary folder that is removed afterwards; a folder REA cannot remove fails the capture with the folder as its cleanup resource. The key that verifies is recorded in signing_key with its path, SHA-256, format (binary or armored) and primary fingerprint, and its bytes are retained under content/. A key file holding one key is retained as it is; from one holding several, such as a legacy /etc/apt/trusted.gpg, only the transferable key with the verifying fingerprint is retained, in binary, so the bundle trusts no other key. When no candidate verifies, gpgv is missing, apt holds no signed release, no sources entry matches, the entry's Signed-By holds the key itself, or the key is larger than --max-file-bytes or the content budget, signing_key is null and signing_key_reason says why. Ubuntu Pro's ESM repositories (origins such as UbuntuESMApps, host esm.ubuntu.com) are not checked: apt reaches them only with the host's Ubuntu Pro credentials, which a computer does not have.

Unless the installed version comes from the Ubuntu 24.04 archive, so for a vendor repository's package and for a local install that no apt source lists, library_packages lists the packages that own the shared libraries the executable loads at start; the walk runs beside the repository check, and not at all for an archive package, whose export does not use it. REA reads DT_NEEDED from the resolved executable and from each library it finds, and looks each soname up as ld.so does: in the RPATH of the needing image and of each image up its chain of loaders to the executable when the needing image has no RUNPATH, then in its RUNPATH (with $ORIGIN expanded and paths normalized), and then in the ldconfig -p cache entries for the executable's machine. It asks dpkg-query -S for each library's package. Libraries in the executable's own folder belong to the application: they are walked but not listed, and neither is the application's own package. One batched apt-cache policy then keeps only the packages the Ubuntu 24.04 archive or a recorded vendor repository with a verified key offers, in any component its sources entry lists, since the export declares them all; the rest, such as packages from a PPA or a local .deb, are listed in uninstallable_library_packages. Sonames found nowhere are listed in missing_libraries, and libraries no package owns in unpackaged_libraries. library_packages_reason says why the lists are incomplete: the executable is not an ELF image (a launcher script, for example), ldconfig, dpkg-query or apt-cache is missing, a library could not be read or names more than 512 libraries, or the walk stopped after 512 libraries.

A sources file, key or library whose handle REA cannot close fails the capture with the file as its cleanup resource. When the repository check and the library walk both fail, the capture reports both failures and every folder or file either could not release.

An application installed from a .deb file you downloaded, such as VariCAD from varicad.com, has no apt source a computer could install it from. Name that file with --package-file so the export can have the computer download it instead. REA reads the file's control fields with dpkg-deb --field FILE Package Version Architecture Depends Pre-Depends, and when its Package, Version and Architecture are the installed package's, it reads the file once more for its SHA-256 and size and records package_file: the file's name without its folder, sha256, size, package, version, architecture, and the depends and pre_depends entries as the file declares them. Otherwise package_file_reason says why, for example that dpkg-deb is not installed, that the file holds another version than the installed one, or that the file changed between the two reads (its device, inode, size and times are compared), since its fields and its hash could then describe different files. The capture never copies the file: it is the vendor's, and only these facts are kept.

Snaps, flatpaks and AppImages are recognised from their paths, and flatpaks also from a flatpak run launch, whose application id is the first operand after flatpak's options and their values. Electron applications are marked by the fuse sentinel in their binary or a resources/app.asar beside it. State directories are the XDG config, data, state and cache folders and the dot-folder named after the package, the executable, the desktop entry's full id and its last part, and the entry's name, with whether each existed at the start. Dotfiles in the home named after the application (.<name>rc, or .<name> followed by -, . or _, for a name of three or more characters) that exist at the start join them, since an application may only read such a file during the capture. A flatpak run launch is named after its application id rather than the flatpak launcher, and adds ~/.var/app/<id>.

Session. When it started, how long it ran and why it ended; the launched process's exit status; every process seen in the application's tree; every client window with its class pair, titles, owner pid and type, marked when it belongs to the application; when the first application window appeared; the most common class pair of the application's windows, for matching its window later; files the application's processes held open under the walked roots; the first 8 KiB of the launched process's stderr; and whether cleanup was verified. Each process records whether the capture started it: the launched process, a process carrying the capture's run token, or a descendant of one, even when its environment cannot be read, as for a zombie or a sandboxed renderer. Processes are told apart by their start time, so a pid reused by another process is not counted as the application's.

A window belongs to the application when its owner pid is one of the application's live processes or its class matches the desktop entry's StartupWMClass, the executable's name, the entry's full id or its last part, or a flatpak's application id. A flatpak's windows report pids from its sandbox, so they are matched by class only. Ownership is checked again at every sample, because a window can appear before the process that owns it is found. A matching window's pid joins the application's processes, so a launch handed to a process the capture did not start, such as a D-Bus activated service, is still followed. That process does not carry the run token, so the capture never signals it and leaves it running.

State. The walked roots (the home directory plus any content root outside it), the content roots (state directories, --state-path values and absolute --argument values, except / or a directory that contains the home directory, which are named in the limitations instead; a selected path that is not a directory brings the -wal and -journal paths beside it, so a database selected on its own is captured with its logs), the state directories as walked (their physical spelling, which differs from the inventory's when one is a symlink), the paths every walk left out (the output folder and --exclude values), and the checkpoints in order: baseline before launch, each requested checkpoint, stop or time-limit, and exit after the application has gone. Each checkpoint records how many paths its walk saw, whether it was truncated, how many paths it could not read or directories on another device it did not enter (unchecked_paths), which paths were created, modified, deleted or changed type since the previous checkpoint, and which files it retained. A path counts as created or deleted only where the walk listed its nearest existing directory completely; differences under a directory that could not be listed, or that a truncated walk did not reach, are counted as unconfirmed instead. A walk does not enter a directory on another filesystem than its root, such as an NFS, sshfs or other FUSE mount under the home directory; the limitations list each such mount point. The baseline retains every file under the content roots; later checkpoints retain every changed file. A change records the file's type, size, mode, modification time and, when known, the SHA-256 of its content before and after, plus whether it lies in a state directory and whether the application held it open. A file is retained only if it is still the version the walk recorded and does not change while it is read. A file larger than the remaining content budget is not read, unless stored content of the same size could make it a duplicate. Files that could not be retained are listed with the reason: too large, over the content budget, changed since the walk, changed during the read, or unreadable. Directories the application held files open in, but whose starting content was not retained, are listed so a second capture can pass them with --state-path.

Limitations. What the capture could not establish, such as a truncated walk or a missing window.

What the capture does not establish

Changes are found by comparing walks, using each path's size, modification time, inode and mode, so a change undone between two checkpoints is not seen and a rewrite that keeps all four is missed. A change is not attributed to the process that made it: other desktop services write under the home directory too, such as the dconf database, the window manager's settings and GTK's recently used list. The markers for state directories and open files are evidence for attribution, not proof.

Clicks and keystrokes are not recorded. Window titles are sampled every half second, so short-lived titles can be missed.