v0.20.16

Fixed

• Delegated work comes back. Every subagent a session handed work to failed
  before it started, with contextWindow from task-parent-session requires a catalog physical limit: a child inherits its parent's window, so nothing
  needed to state one for it, and asking for one failed the delegation. This is
  the "subagents come back with nothing" report.
• A goal's elapsed time counts from the goal. It restarted whenever you
  switched away from a session and back, because it counted from the redraw
  rather than from the goal.
• A bundled app writes a log. Launched from Finder, the agent's own account
  of a failure went to a descriptor nothing read, so a bundle that failed had
  nothing to open.
• A session's first five lines are no longer dropped. The agent announces its
  commands a millisecond behind the session opening, and the app filed that
  answer only after applying the session's settings — about 240ms against a
  200ms window.
• Session titles exist. The title command passed --output-format to the
  wrong subcommand, so it printed its usage and exited: no session had ever been
  titled.
• A file link in a message opens the file. A relative path is resolved
  against the session's workspace as well as the project root, with a
  unique-suffix fallback, instead of one base that is right only sometimes.
• An exec run that was cut off keeps what it found, instead of returning a
  status with no output at all.

Added

• A delegated run has a ceiling that delivers rather than kills. It is held
  to 200 model requests, told when it crosses, and stopped at 1.5× — and a
  stopped run reports what it found. A run that will not wrap up is ended, and
  what it committed is still its answer.
• A child cannot end without producing something. A run that finishes
  silently is asked once, in the same conversation whose context it still holds,
  and the result says exit: reported or exit: silent — so "produced nothing"
  is no longer indistinguishable from "still going". A silent exit names both
  ways to get the rest.
• Export a session from its context menu: the transcript as Markdown, or its
  event log as JSON.
• The sidebar says more with less — counts in the numeral face, a line for
  the sessions waiting on you, and the agent's own face on a session that is
  working, where an orb used to be.

What's Changed

• fix(models): what a model sees, the blocks that said otherwise, and the effort it takes by @Hamertingo in #40
• fix(agent): a task child inherits its parent's window, and ends through a contract by @Hamertingo in #41
• feat(agent): a delegated run is held to a request budget that stops and delivers by @Hamertingo in #42
• fix(agent): an exec run that stopped short keeps what it found by @Hamertingo in #43
• fix(desktop): three things the app lost silently, and a Fail that can print by @Hamertingo in #44
• fix(desktop): a goal's elapsed time counts from the goal, not from the redraw by @Hamertingo in #47
• docs(desktop): what is worth building next, and what reading the code changed by @Hamertingo in #48
• feat(desktop): a path the agent names, and one the reader picks by @Hamertingo in #45
• feat(desktop): the sidebar's numbers, and the face that says a session is going by @Hamertingo in #46

Full Changelog: v0.20.15...v0.20.16

v0.20.15

The same content as 0.20.14, with the Windows installer that release shipped
without. 0.20.14 published its macOS assets and then failed in the step that
uploads the NSIS installer: the glob that names the version was written
*0.20.14-setup.exe, while tauri writes Hyze Code_0.20.14_x64-setup.exe — the
architecture sits between the version and -setup, so the pattern matched nothing
and the guard fired on a bundle that was complete. The pattern names that
separator now, which still leaves a stale 0.20.8 file out.

That failure also held the updater manifest. The job that writes
stable.json declares both builds as its needs, so a red Windows job skipped it —
and 0.20.14 sat published with nobody offered it. It now runs whatever a platform
build did: a missing Windows fragment was already an ordinary state to that step,
and the alternative is one platform's build failure deciding when another
platform's users hear about a release.

Full Changelog: v0.20.14...v0.20.15

v0.20.14

Added

• A Usage page — ⌘⇧U, or the row in the sidebar. What the agent has spent,
  read back out of each session's log: the period's total and how it splits
  between input, output, reasoning and cache, a day-by-day chart, the models with
  the gateway each one came through, and the newest records. A period row and a
  gateway row narrow all of it, and the gateway is read from the model's own wire
  ref — command-code and opencode-go serving the same model name are two rows,
  not one. Nothing on it is a counter this app keeps: it is the agent's own
  accounting, so a session deleted from ~/.hz takes its rows with it.

• A goal the agent owns. The runtime's own objective object — its budget, its
  turns, its status — drawn in the composer's row, with a receipt in the
  transcript when the work lands. The objective goes out through the composer's
  own path, so the reader's words and the agent's work arrive as an ordinary turn
  and the runtime does not redo it.

• A retry is visible. The harness now says when it is re-asking a model, and
  the row and the working indicator built for it finally have something to draw.

• A new identity. Four cuts of one artwork — tiled, plain, flat, mono —
  replace the black cat on the dock and installer icons, the composer's
  watermark, and the site's wordmark, favicon and OG card.

Changed

• One agent process serves every session in a project. A second session in a
  repo opens a conversation on the child that is already running instead of
  paying the boot again — measured across five sessions, −73% time and −77%
  memory — and the process count stops growing with the number of chats.
  Stopping a session closes its conversation; the child goes when the last one
  does. The routing is the whole of what makes this safe, and the plan note
  that specified it required the app itself as the instrument, so that is how it
  was checked: two sessions in one project, a prompt in each, each answer landing
  in its own transcript — then the shared child killed with -9, after which
  both sessions resumed on a fresh one.

• HZ_SESSION_ID travels with the turn rather than with the process. A child
  serving several sessions has one environment and could not name five sessions
  with it, which is what hz issue link with no id defaults to. The agent exports
  the id of the session a tool child belongs to, and the app resolves either
  spelling — its own uuid or the agent's id — through one funnel, so the CLI and
  the skill did not change.

Fixed

• A weekly usage limit is classified as one. The provider's own words
  ("you've reached your weekly usage limit for your plan") matched none of the
  classifier's patterns, so the request was retried against a limit with hours to
  go.
• The installer attached to a release is that version's. The bundle directory
  sits inside the cached target/, so an older version's installer was still
  there and a glob uploaded it beside the right one — 0.20.8's 13MB setup was
  attached to 0.20.9, 0.20.11 and 0.20.12 alike, and a reader picking the wrong
  row installed a build from three releases ago.
• The slow-read line floats on the float veil. It wore --card, which under
  vibrancy is a 5.5% white veil with no blur, so session titles read straight
  through its own words. It also moves to the top of the conversation column,
  where it stops covering the sidebar it was never about.
• A #7 the repository knows is a chip. An agent that has just opened a pull
  request says so in prose, and the number is now pressable — opening the PR pane
  in the app rather than a browser. A number nothing holds stays the words it was
  written as.
• A session's slash commands survive a restart. The agent publishes them when
  a session opens, so a reopened session had no child to have published anything
  and the menu did not open at all. The last list is kept per session now.
• The wait a /compact takes is drawn. The wire says nothing about a
  compaction, so the indicator is derived from the prompt this app sent: the
  newest prompt names the command and the turn is still open.
• A prompt with nothing in it is never sent, and the composer's row keeps its
  three nodes in place.
• A project attached at / has a name, and a session's project is resolved
  from the main worktree rather than from the tree the caller stood in — which is
  what made a session created inside a linked worktree read a directory that does
  not exist.
• macOS 12 is the floor the browser already required. Three releases failed on
  the universal build with Undefined symbols: ___isPlatformVersionAtLeast,
  because the CEF objects are built for 12 while the app linked for 11 — and the
  clang runtime the ggml objects ask for is linked explicitly now.
• Windows stages the agent the way macOS does.

What's Changed

• fix(paths): name the root instead of nothing by @Hamertingo in #31
• fix(agent): apply picks before a prompt is queued by @Hamertingo in #32
• feat(agent): show the wait a compaction takes by @Hamertingo in #33
• feat(chat): open the pull request a message names by @Hamertingo in #34
• feat(composer): keep the command list a session was sent by @Hamertingo in #35
• feat(agent): give a session a goal the agent owns by @Hamertingo in #36
• feat(brand): bring the new identity to main by @Hamertingo in #37
• fix(ui): draw the slow-read line on the float veil by @Hamertingo in #38
• fix(release): publish this version's installer, not a stale one by @Hamertingo in #39

Full Changelog: v0.20.13...v0.20.14

v0.20.12

Fixed

• The Windows installer stages what the macOS one stages. It was extracting
  32,000 files and 524MB — node_modules\json-schema… under NSIS's progress bar
  — because the two launchers were two copies of one staging rule and only the
  shell script had been fixed. The rule lives in one script both call now, and
  the tree it produces is 1,286 files and 89MB. 0.20.11 installs correctly and
  slowly; this one installs correctly.

What's Changed

• Stage the Windows agent the same way the macOS one is staged by @Hamertingo in #26
• Release 0.20.12 by @Hamertingo in #27
• Give 0.20.12 the section the release reads by @Hamertingo in #28

Full Changelog: v0.20.11...v0.20.12

v0.20.11

Changed

• Windows draws its own window. The system title bar is gone
  (decorations: false, in a tauri.windows.conf.json overlay so nothing else
  moves), the corners are rounded by CSS from the app's own radius, and the three
  window controls are the app's — which is not decoration: with no system chrome
  they are the only way to minimise, maximise or close. Every drag region
  reserves their width. The default menu bar Tauri draws in-window goes with it.
  This shipped as 0.20.10, whose release never published.

• The agent's boot is cached, and a quarter to a half of it is gone.
  initialize was 1.6s cold and 2.0–2.6s on repeated children, all of it V8
  parsing the bundle. NODE_COMPILE_CACHE writes nothing for this tree; the API
  that works runs as a --import preload, and it is only written on a clean
  exit — which is why an earlier measurement of this found zero bytes. Measured
  through the launcher the bundle ships: 1.59s → 0.82s, and 1.34s → 0.96s on
  a repeat.

• A quarter off per-turn work. One lock per session instead of one global map
  lock; each agent line parsed once, with usage updates held for the frame; the
  transcript loads its newest turns first and pages backwards; the sidebar, split
  view and panels are memoized against per-delta renders; dialogs, pages and pane
  bodies load as chunks (2,436 kB → 2,268 kB); the release profile gets thin LTO,
  one codegen unit and strip, with devtools dev-only.

Fixed

• The macOS build links again, and the floor is honest. The app declared
  macOS 11 while the CEF build it ships is built for 12, so the linker asked for
  ___isPlatformVersionAtLeast and found nothing to provide it — three releases
  in a row failed on aarch64-apple-universal with Undefined symbols. The
  deployment target is now 12.0, which is what the browser has been requiring
  anyway: an app that cannot start on 11 should not say it can.

• A model's window is the model's, not the gateway's guess. A reseller
  serving deepseek-v4.1-flash under a provider id of its own had no catalog
  entry, so a session compacted at 200k while the model runs at a million — and
  the settings row drew 200k default because the window is only ever stated on
  a usage update, which needs a session already running the model. The agent now
  states it on the model itself (ACP _meta), the app reads it onto Model, and
  a stated figure also overrides what the gateway claims when a model ref is
  stamped.

• A model's name is decoded. A gateway that names its catalog
  publisher/model reaches the wire as deepseek%2Fdeepseek-v4.1, and the name
  drew with the escape in it — worse, the brand table then keyed off no vendor at
  all. The split still happens on the wire's own separator, which has its own
  test: decoding first would hand the model a name that is the provider's tail.

• No more console windows flashing on Windows. A console program started by
  an app that has no console gets a new console allocated, and that console is a
  window for as long as the child runs — so git, gh and the agent each
  flashed one, and attaching a project was a screenful.

What's Changed

• Stop the Windows gate from running for a React-only change by @Hamertingo in #15
• Keep the console trait import inside the function that uses it by @Hamertingo in #16
• Perfil de release com thin LTO + 1 codegen unit + strip, e devtools só by @Hamertingo in #9
• Cada linha do agente parseada uma vez (o demux do prompt reusa o `Valu by @Hamertingo in #10
• Decode the escapes a model's name arrives with by @Hamertingo in #17
• Know the window of a gateway model the catalog has never heard of by @Hamertingo in #18
• Apply a stated window where the provider's own would reach a ref by @Hamertingo in #19
• Carry a model's window from the agent to the screen that draws it by @Hamertingo in #20
• O lock do mapa guarda só o formato do mapa; cada sessão tem lock própr by @Hamertingo in #11
• Seleção carrega as 400 turns mais novas e pagina para trás no scroll; by @Hamertingo in #12
• Dialogs, páginas e painéis em chunks sob demanda. by @Hamertingo in #14
• Cache the compile of the bundle the agent parses on every boot by @Hamertingo in #21
• Release 0.20.11 by @Hamertingo in #22
• Write down what one child serving many sessions would take by @Hamertingo in #23
• Link for the macOS version the browser already requires by @Hamertingo in #24
• Link clang's runtime, which the ggml objects ask for by @Hamertingo in #25

Full Changelog: v0.20.10...v0.20.11

v0.20.9

Changed

• The bundle carries the agent, not its build tree. Staging ran
  pnpm install --prod inside the copy, which resolved every dependency of
  every workspace package: 32,077 files and 524MB, where the shipped CLI runs
  dist/ — 771 files that already have every JavaScript dependency inlined. What
  a bundle cannot carry is a native module, so the stage now copies exactly the
  packages esbuild marked external, read off the build's own metafile. The tree
  is 1,289 files and 205MB, and that number is not about disk: NSIS extracts
  file by file, so installing was minutes of Extract: agent\app\node_modules\@smithy\core\… and is now seconds. Verified by
  running the staged agent, not by trusting the copy — initialize and
  session/new both answered from the trimmed tree.

Fixed

• The installer carries the app's icon. bundle.windows was empty, so NSIS
  drew its own defaults while installing — the title bar, the Add/Remove Programs
  entry and the uninstaller showed Tauri's mark rather than this app's.
  bundle.icon had always covered the installed app; nothing covered the
  installer.
• The Windows updater entry is written. The jq that builds it used
  windows-x86_64 as a bare object key, which jq cannot parse — the step failed
  with a compiler error and the release published macOS alone. It never ran
  before: the artifact guard above it failed first, every time.
• No more console windows flashing on Windows. A console program started by
  an app that has no console of its own — which every GUI app hasn't — gets a
  new console allocated, and that console is a window for as long as the child
  runs. git, gh and the agent are all console programs, so attaching a project
  flashed a screenful of windows titled hz-agent. Every spawn in the app now asks
  for CREATE_NO_WINDOW, through one trait in proc.rs; a new Command::new that
  forgets it is a black window on somebody's screen. Nothing to do elsewhere — a
  unix child has no console to be given.
• Staging refuses a runtime it cannot ship. The node copied beside the
  launcher was copied without being run, and better-sqlite3 is compiled for one
  ABI — so a Homebrew node (a thin binary plus dylibs at absolute paths) or a
  module built under a different major produced a bundle whose agent never
  answers. Both are checked now, with the cure in the message, because the
  failure they prevent is an app that opens and a session that does nothing.

Fixed

What's Changed

• Install in seconds: stage the agent's bundle, not its build tree by @Hamertingo in #5
• Stop every spawned child from opening a console on Windows by @Hamertingo in #6
• Make the console trait import survive a cfg, and the node copy readable by @Hamertingo in #7

Full Changelog: v0.20.8...v0.20.9

v0.20.8

Fixed

• The Windows release publishes. 0.20.7 came out for macOS alone: the job
  that uploads the installer expected a *.nsis.zip beside the setup.exe, which
  is the layout Tauri v1 produced — this bundle is v2, where the setup.exe
  is the updater artifact and carries its own signature. The guard failed a
  bundle that was complete, and the manifest job never ran. Nothing about the
  app changed here; the tag exists so the release could be cut again without
  moving one that had already been pushed.

What's Changed

• Say what this app is, and how to get it by @Hamertingo in #2
• Correct what the front page said that the tree no longer does by @Hamertingo in #3
• Look for the Windows updater artifact Tauri v2 actually writes by @Hamertingo in #4

Full Changelog: v0.20.7...v0.20.8

v0.20.7

Added

• hz runs on Windows. An installer, and the agent staged for that platform
  beside the one macOS already ships — so the app is still one download:
  configure a provider, pick a model, work. hz in a terminal reaches the app
  there too, over a named pipe rather than a socket, so an agent inside a
  session can still create and message sessions.

Changed

• The release manifest names both platforms. An installed copy on either one
  updates itself from the same channel, and the two entries are merged rather
  than one overwriting the other.

Not there yet

• The in-app browser is macOS only, which it always was. A Windows build has
  no Browser tab — the CEF bindings are not compiled there.
• "Open in app" and the terminal opener are macOS only. Windows is told so,
  with the path to copy, rather than handed a button that does nothing.

What's Changed

• Port hz to Windows by @Hamertingo in #1

New Contributors

• @Hamertingo made their first contribution in #1

Full Changelog: v0.20.6...v0.20.7

v0.20.6

Added

• The PR tab says how to turn pull requests on. Without the GitHub
  CLI the tab simply vanished, so nothing said that one install puts
  pull requests on screen. It now stays, draws the command to run with
  a copy button, and a terminal to paste it into.
• Dray can ask you a question in the app. A short survey card may
  appear for installs with usage reporting on; turning reporting off in
  Settings turns it off too.

Full Changelog: https://github.com/Hamertingo/hz/commits/v0.20.6