Skip to main content
CHANGELOG

What shipped

Released versions of Harmony only, read from the product's Keep a Changelog file. Unreleased work is not listed here: if it is not in a version heading, it has not shipped.

Latest released v3.25.8 · · 65 releases

Download the current build from /download. Channel and exact version always come from the live release manifest.

WHAT'S NEW IN 3.25

The short version

  • One setup guide on first launch

    First launch walks you through it once: the agents already on your computer, an optional Harmony account, your workspace and privacy, instead of separate pop-ups.

    v3.25.5

  • Your own models and keys, checked before they are saved

    Add an API key or a custom endpoint in Settings, see the models it can actually reach, pick a default, and set the order to fall back through when a provider hits a limit. Keys are encrypted by your operating system.

    v3.25.5

  • The Harmony Agent on your existing subscription

    harmony auth connects Harmony's own agent to a model in a few questions, including through your Claude Code or Codex subscription, and the agent switches provider before a quota runs out when the provider reports it.

    v3.25.5

  • Agent Hours in the status bar

    Signed in to Harmony Pro, the app shows how many Agent Hours you have left, or the one thing blocking cloud work. Signing in stays optional.

    v3.25.5

  • Terminals that survive more

    A terminal Harmony had to restart now comes back inside the background terminal host, so it keeps running through later restarts, and no longer restarts again every time.

    v3.25.5

  • Linux on the stable channel

    Linux builds, as .deb and AppImage, ship on the stable channel alongside Windows. On Linux Harmony tells you when a new version is out and links to it.

    v3.25.7

  • Windows installs finish on their own

    The installer no longer sits on its INSTALLING window until you quit Harmony; it closes on its own when setup is done.

    v3.25.8

  • Agent tools that stay in their own workspace

    Harmony's MCP tools act only on the caller's own workspace, and several that used to report success for something that did not happen now say so.

    v3.25.5

Releases 1–5 of 65

Page 1 of 13

v3.25.8

Docs

  • Docs: persistent agents and memory, documented honestly. A new concept page explains the four things that are not one thing — persistent identity, provider session, checkpoint, memory — and says exactly which parts ship: agent memory with trust, scope, and sensitive-content screening is landed; durable identity + checkpoint save is implemented but off in current builds and refuses with NALA_CAPABILITY_UNSUPPORTED; embedded agent browser control, Cloud execution of persistent agents, and paid enforcement are not in the product. The glossary gains the matching terms (persistent agent, checkpoint, incarnation, rehydrate, host-derived trust, and the verbatim memory states), and feature status carries the honest rows. No prices, plans, or catalogue entries changed.

Fixed

  • Installing or updating on Windows no longer hangs on the INSTALLING window. The install finished in under a minute, but the installer then waited until you quit Harmony, because the app it started after installing kept the installer's output open. Windows now starts Harmony the normal way when the install completes, the installer closes on its own, and the PowerShell one-liner from ideharmony.com returns as soon as Setup finishes.
  • Quitting the Harmony Agent while it is still starting its model session no longer sends your message anyway: nothing reaches the model and nothing is saved for a retry after you have quit.

v3.25.7

Changed

  • Linux (.deb and AppImage) now ships on the stable channel alongside Windows. Linux stays a download-and-install update: Harmony tells you a new version exists and links to it, but does not replace itself on Linux yet.

v3.25.6

Added

  • Linux builds (.deb and AppImage) of the stable release.

Fixed

  • Each installer now ships the third-party licence notices for its own platform (for example the LGPL libvips attribution on Linux).
  • Image and speech features on Linux and macOS: packages now include the libvips library that sharp loads (it was pruned from every Linux and macOS package).

v3.25.5

Added

  • Settings › Models & providers: add an API key or a custom endpoint (checked with the provider first), pick the default model from what your key can reach, set the failover order, and see where each key is kept — the same settings as harmony auth. Keys are never shown again after you save them.
  • With Harmony Desktop running, provider keys are encrypted by your operating system; plain-text key backups are deleted once the moved keys are proven to work; gateway headers such as api-key are kept in the key store (harmony inference add --header).
  • The Harmony Agent switches providers before a quota runs out when the provider reports it, and after an automatic switch resends a message that failed before any answer or tool ran — once.
  • Account & Cloud: a switch to carry admitted agent memory to your own cloud workspace (off by default; no effect while signed out).
  • harmony auth walks you through connecting the Harmony Agent to a model in five questions or fewer: use your Claude Code or Codex subscription (open that CLI, or run its own sign-in; Harmony never sees its token), sign the agent in with your subscription, paste an API key (checked, then pick from the models it can reach), or a custom endpoint (preset or protocol, URL, key). /login with nothing after it runs the same guide inside the agent, with the key shown as dots while you type it. harmony login now says it is Harmony Cloud sign-in and points at harmony auth.
  • harmony auth status (and /auth in the agent) shows, for this machine only, each provider's route and billing, where its key is kept, when it was last checked and the result, how old its model list is, and the default model — plus whether Claude Code and Codex are installed and signed in. Anything not checked says "unknown". --json is available and it always exits 0. /auth routes shows the route table.
  • Provider failover: /model fallback add <provider>/<model> auto|ask|never (or harmony auth fallback …) sets the order the agent switches through when a provider hits a rate or usage limit. The switch happens between turns and writes "Switched from X to Y: ; X resets at " into the conversation. It switches on its own only to entries you set to auto (which is when you are shown how it is billed); otherwise it asks. It returns to your first choice once its limit resets.
  • Azure OpenAI endpoints are stored with a clear "not verified — Azure deployments cannot be listed; test with a prompt" note instead of being refused, and the agent is started with the api-version you gave.
  • The Harmony Agent checks an API key before storing it. /login key <provider> <key> and nala inference add list the provider's models with the key, and store it only if that works. A failure says what went wrong: bad key (401), no access (403), wrong base URL (404, with a hint about a missing or doubled /v1), a TLS problem, connection refused ("is Ollama / LM Studio running?"), host not found, or a timeout. --no-verify stores the key without the check and says it was NOT verified. Providers Harmony cannot check (for example Bedrock) are stored with a note saying so. Keys never appear in output or in the model cache.
  • /model shows where each provider's list came from: "live from your key", "cached (age)", or "catalogue (not verified for this key)". A catalogue model your key's live list leaves out is no longer shown as available. /model refresh lists models again with your stored keys. If a refresh fails, the old list is still shown but marked STALE. /model <provider>/<id> also accepts an id that is not in the list, and warns you when it does.
  • nala inference has endpoint presets for OpenAI, Anthropic-compatible proxies, OpenRouter, Azure OpenAI, LiteLLM, vLLM, Ollama and LM Studio, plus the existing ones, and --api for other protocols (OpenAI completions/responses, Anthropic messages, Azure OpenAI). It warns about base-URL mistakes. It uses the verified model list when --model is left out. The key comes from --key-stdin or a hidden prompt. --key <value> still works, with a warning that the key shows up in process listings. Local presets need no key.
  • Signing the Harmony Agent in with a consumer subscription (/login anthropic or /login claude, /login openai-codex, /login github-copilot) is your choice. The first time, it shows this once: "This signs the Harmony Agent in with your subscription. Your provider's terms apply to how you use it." /auth, /login list and harmony doctor all read the same auth-route registry, which lists these routes as "user choice". Before, it called the Claude route unsupported even though /login anthropic ran it.
Show the other 150 changes in v3.25.5
  • nala doctor has a session-survival section: how many terminals are held by the background terminal host and will survive a daemon restart, and a warning naming each terminal (with its folder) that runs inside the daemon and will restart with it. If the daemon's session list cannot be read, it says the answer is unknown instead of "none at risk".
  • First launch is now one setup guide instead of separate pop-ups: Welcome, your agents (what is actually installed on this computer), an optional Harmony account, your workspace, privacy, and a ready screen with the key shortcuts. Every step can be skipped, it picks up where you left off if the app closes, and you can reopen it any time from Settings > Getting started or the app menu's Setup guide. If you have used Harmony before, you will not see it again, apart from the privacy question if you never answered it.
  • Installed builds offer an optional Harmony account: sign in from Settings > Account & Cloud and your plan, Agent Hours and Cloud workspaces appear. Nothing is sent to the account server until you sign in, and Harmony Core never needs an account. Enterprise and air-gapped installs opt out with NALA_ACCOUNT_GATE_MODE=disabled. Development builds and tests keep the account layer off unless asked.
  • Signed-in Harmony Cloud accounts see their remaining Agent Hours ("48 of 60 Agent Hours left") in the status bar's Usage and system status panel, or the one thing blocking Cloud work (payment failed, trial used) when it is refused. It reads the account the app already holds and opens the Cloud panel on click. Nothing is shown when you are not signed in, and it adds no network requests.
  • nala logout (and nala cloud logout) now revokes this device on Harmony Cloud before deleting the local sign-in, then says exactly what happened: revoked, already signed out on the server, or unreachable (in which case the device may still be signed in, with the link to revoke it). The local sign-in is removed in every case. --json reports serverRevocation. Needs the site's new POST /api/v1/device-auth/revoke.
  • npm run make and npm run package refuse to start when the disk cannot hold the build. A full build that runs out of space fails at the END, after about forty-five minutes, and reports it as tail: write error: No space left on device rather than as a full volume. The preflight says what is free, what is needed, and the one command that reclaims it. It checks and does not clean: deleting build output stays the operator's call.
  • Equalize Pane Sizes is in the command palette. The action existed in the pane store and nothing in the tree dispatched it, so it was reachable only from the type system.
  • The daemon reports per-method RPC service time and what is running right now on daemon.ping, alongside the event-loop lag it already reported. A stalled loop now names the RPC holding it, or says inFlight=none when the cause is outside RPC entirely. Diagnostics only; no scheduling change.
  • A Law-3 poller inventory guard ratchets polling in the surfaces that can reach the daemon. The raw setInterval count across Desktop, main and TUI is 38; only 7 sit in files that reference a daemon RPC at all, and one of those is a component the orphan census records as deliberately unmounted. The scanner refuses to report a clean result unless it can still find a timer it is known to be able to find — its first version reported zero pollers because the detector was broken, not because the code was clean. This is a build guard, not a runtime change.
  • The theme engine exposes canvas(text), which paints a full terminal row with the active canvas/text colors and restores them after nested 39/49 selective resets. Terminal-native, monochrome, and depth-zero canvases are left untouched. This is not a new IPC channel or a packaged rebuild.
  • Bundled JetBrains Mono TTF assets and a design-research doc with rendered TUI previews back the terminal typography work. This is not a packaged font install.
  • Cloud transfers have a compact choose/review/send/follow flow, preserve pending work during account refresh errors, and open the created workspace directly. Queue refusal keeps the created workspace reachable. The workspace overview adapts to narrow windows; pane Bring Home targets the linked terminal. The pink dot uses canonical server states and requires a running VM. This is component QA, not packaged desktop or live Cloud.
  • Freestyle Cloud is a top-bar tool for paying Cloud accounts and confirmed site admins. The pane’s trailing menu can send that terminal’s work, open Cloud workspaces, or Bring Home. One dot turns pink only when a linked workspace is confirmed running. Local agent launch does not wait on Cloud.

Changed

  • Provider API keys are no longer written into auth.json or models.json. They go into Harmony's key store, and those files hold a $HARMONY_KEY_<PROVIDER> reference that the agent receives in its environment when it starts. Keys already in those files are moved on the next /login or with harmony auth migrate; the old files are backed up (the backups still contain the keys — delete them once the agent answers). The store is a file only your user account can read, not an OS keychain yet, and harmony auth status says so.
  • /model no longer waits on the network: an expired model list is shown straight away, marked STALE, and refreshed in the background.
  • The default dark theme is now Harmony Obsidian (renamed from NALA Obsidian): a warmer graphite/ivory palette with restrained gold, muted metadata that stays readable on real rasterizers, and bold shortcut tokens. The welcome card shows two rows of canonical lettering on tall Unicode terminals and a plain lockup otherwise, wraps the tagline instead of clipping it, and styles / and Ctrl+P action hints so narrow panes keep every key with its explanation. The transcript header uses PRODUCT_DISPLAY_NAME instead of a hardcoded NALA. Six-row banner art stays reserved for nala --version and the splash. This is not a packaged rebuild.
  • The MCP server key is now nala, not wmux. A client registered under the old key needs registering again.
  • The eleven-item panels menu in the title bar is gone. The same panels open from Settings and the command palette.
  • The Fleet panel no longer has a Steer control; it never steered. Use Open to jump to an agent's terminal.
  • The app icon is the // mark, replacing the >n monogram. Your operating system may show the old icon until its icon cache is cleared.
  • "Raise a fleet from a sentence" no longer starts agents by itself after you approve. It copies the template and the exact command instead, so nothing runs, or costs anything, until you run it.
  • Settings controls that never took effect are removed, along with a placeholder Approvals page and a duplicate first-run menu entry.
  • The Diagnostics menu item is now "First-run setup / repair"; it opens the same place.

Removed

  • The unused in-app plan catalogue and its three hidden screens (billing, plan comparison, usage). They described a "Pro" plan with a 14-day trial that Harmony does not sell and could never be shown. Your plan, Agent Hours and billing come from your Harmony account (Settings > Account & Cloud) and ideharmony.com, which is the only plan authority.
  • The unwired src/cloud terminal-UI module and npm run test:cloud. Nothing imported it. Its money and safety rules are now enforced and tested in the shipping nala cloud CLI and the Desktop Cloud panel instead (CLOUD-RETIRE #347).

Fixed

  • The standalone agent browser refuses a full Chrome or Chromium at launch, with a message that says to use chrome-headless-shell (via HARMONY_BROWSER_CHROMIUM_PATH). Before, a full browser opened pages but could never take a screenshot, so every click, fill or select was refused after an 8-second wait as a "stale observation". The popup guard stays as strict as it was.

  • nala update --check no longer says "up to date" when the installed build is newer than the release channel; it names both versions and says there is nothing to install.

  • A supervised agent terminal that crashed and was restarted by Harmony now comes back inside the background terminal host, so it keeps running through later daemon restarts. Before, the restart put it inside the daemon, and it died with the next daemon restart. Closing the pane while the restart is starting now cancels it cleanly. Terminals held by the terminal host now also update their folder, git branch and title in the sidebar, and send desktop notifications, like other terminals; before, those froze at the values from when the terminal started. Terminals the host brings back after a restart stay silent until the pane has its real size, so their first screen is not drawn at the wrong width. The terminal host also stops keeping every exited terminal in memory forever.

  • Settings > Workspaces now actually applies the warm grace window and the maximum held sessions to the running daemon. The setting never left the app before, because it was sent through a channel that refuses it. When the daemon adjusts a value to its allowed range, or cannot apply it, a notice says so.

  • Several MCP tools no longer report success for something that did not happen. A refusal from the app is now an error in browser_navigate, browser_navigate_back, browser_tabs close, a2a_task_send and the channel mission tools, instead of reporting the old page, "queued", or Error [undefined]. browser_navigate_back with no history now says it did not go back. agent_prompt no longer reports a confirmed submission as a failure. browser_tabs select now focuses the tab in Harmony instead of only inside the browser engine, and only for tabs in your own workspace. The workspace://identity resource now shows your workspace, pane and agent, or the reason it could not.

  • The Fleet panel shows the pending-question badge for the agent that asked, and Reconcile reports what actually changed (stale, superseded and archived rows, released provider slots) instead of always "0 retired". "Free slots now" counts only agents the daemon confirms it removed, and says how many it could not confirm.

  • The diff panel finds its task again, so the PR, close and apply controls and diff comments show up; it was reading the task list from the wrong part of the reply.

  • Support bundles no longer stamp the app and daemon version as 0.0.0; the daemon fills in the versions it actually knows.

  • Harmony's MCP browser tools (click, type, fill, evaluate, navigate, snapshot, screenshot and the rest) now work only on a browser in the caller's own workspace. Before this they picked the first browser page they could find in any workspace, so an agent in one workspace could click, type into or read another workspace's page. Naming a surface in another workspace is refused and says which workspace it is in. Tools that change a page need a caller Harmony has matched to one of its panes; read-only page tools accept the same identity the listing tools do, but never read another workspace's page. browser_tabs now lists and closes only your own workspace's browser surfaces. The browser tools also now tell Harmony which pane is calling, which the app requires before it lets an agent touch a browser. The nala CLI follows the same rule: focus-pane, focus-surface, focus-workspace, close-workspace, nala open and nala browser navigate | close need to run inside a Harmony pane and act only in its own workspace. nala browser close no longer falls back to whichever workspace is on screen, and close-workspace can no longer close a workspace you are not in.

  • A terminal that Harmony had to restart once (after a reboot, a crash, or losing the background terminal host) no longer restarts again on every later daemon restart. Harmony brought such terminals back inside the daemon itself instead of the background terminal host, so each daemon restart ended the shell and started a new one from its saved scrollback, and anything running in it or set in it was lost. They now come back inside the terminal host, like new terminals, and keep the same shell process through later daemon restarts. A terminal that is running inside the daemon right now still restarts one more time, because a running shell cannot be moved; after that it stays. If no terminal host is available, the daemon log now says that the terminal will not survive a daemon restart, and why. New terminals were already kept in the terminal host and are unchanged.

  • Windows now shows the installed app as Harmony. The desktop and Start Menu shortcuts are Harmony.lnk, in a Start Menu folder named Harmony. Settings > Apps lists Harmony with the app icon, and the taskbar, Task Manager, notifications and the dictation tray tooltip say Harmony too. Updating an existing install removes its old NALA.lnk shortcuts and the empty HiNala Start Menu folder, but only after the new shortcut exists, and only shortcuts that point into that install. Uninstalling removes both the old and new names. The install folder %LOCALAPPDATA%\NALA, the NALA.exe file name and the update identity have not changed, so updates keep working. A taskbar pin you made earlier keeps its old label until you pin the app again.

  • The terminal host no longer quits about a minute after a daemon restart. It was checking whether the daemon that first started it was still running, not the daemon now in charge of it, so after a restart it saw the old daemon gone and exited, even though the new one was connected. Harmony then had to start a fresh host, and new terminals could not open for a few seconds. The host now follows whichever daemon currently controls it and allows a two-minute grace period for a restart to finish. Terminals that are already running were never closed by this check, and they still are not. A host that nothing controls and that has no terminals still exits, so it cannot linger in the background. Interrupted writes also no longer leave pty-host-ownership.json.*.tmp files behind in ~/.wmux, and the next daemon start removes any that are more than ten minutes old.

  • Harmony's MCP tools that act (create, send, close, focus or write) now follow the same rule as the nala CLI: a caller that Harmony has not matched to one of its panes is refused, even when it names a workspace, pane or surface, and a matched caller can act only in its own workspace. Before this, pane_split and surface_new created panes in any workspace the caller named, and pane_close, pane_focus and surface_close acted on a pane in any workspace. A2A sends, agent_prompt, agent_open_terminal, browser_open/browser_close, pane_set_metadata and the Company send tools also accepted the WMUX_WORKSPACE_ID environment variable as the caller's identity. That variable is only a claim, so these tools no longer use it; an MCP server's own claimed workspace still counts. nala split --workspace and nala new-surface --workspace also stop bypassing the identity check, as the CLI documentation already said they did. Read-only tools and commands are unchanged. Refusals say what was refused, why, and what still works.

  • Stopping the daemon no longer logs session-host release on shutdown failed: Cannot read properties of null (reading 'close') after a release that worked. The terminal host was released and the handover for the next daemon was saved, but the daemon then tried to close its connection through a reference the release had already cleared. It now closes the connection it released, including when the release fails, so the terminal host frees the controller slot straight away instead of when the old daemon process exits. A release that does fail is still logged as a failure. When there was nothing to release, the log now says so: either no terminal host was connected, or the daemon had already lost its claim on the host. The sessionHostRelease shutdown timing line now includes the outcome.

  • The MCP pane_list and surface_list tools, and the workspace://state resource, no longer show the focused workspace when they cannot tell which workspace the caller is in. They now list every workspace, each labelled with its id and name, the same way nala list-panes and nala list-surfaces do. With a known caller they still list only that caller's workspace, and an unknown workspaceId is still a "no workspace found" error. pane_split and surface_new with no workspaceId and no caller identity now refuse and say what to do, instead of creating the pane in whichever workspace is on screen. nala split and nala new-surface already refused in this case.

  • nala run "<prompt>" and nala "<prompt>" now run a real Harmony Agent model turn, through the same pipeline as the Harmony Agent TUI. Before this, they created an empty draft task that nothing ever ran, then failed on their first poll, so no model was ever called. nala run streams the turn's events (--output text|json|jsonl) and returns a result naming the session and turn (output schema version 2). Its exit codes follow the documented contract: 3 when no model is configured (with the setup step), 4 when a permission policy or the billing route lock refused the turn, 5 when the turn failed, 6 when you cancelled, and 7 on timeout. Ctrl+C cancels the turn, and the command reports it cancelled only when Harmony confirms the cancel. Permission requests are denied, or with --non-interactive-approval fail they end the turn; nothing is ever approved on your behalf. --detach now refuses and says why: Harmony only confirms a turn once it has finished, so there is nothing to detach from. nala "<prompt>" has no deadline and prints the transcript as the turn runs. The flags these commands accepted and ignored (--provider, --profile, --model, --mode, --workspace, --read-only, --max-turns, --wait, --idempotency-key, --max-retries, and the allow-profile approval policy) are removed and now rejected. nala ask and nala plan now refuse instead of failing silently: Harmony cannot yet enforce a read-only or planning-only turn.

  • doctor --export now reports the running app's version, not the CLI's own, with the CLI version beside it, and says why when the app's version is unknown. It reports live and retained session counts separately and marks a missing count as not reported instead of leaving it out. The export now carries exportSchemaVersion: 2. --include-logs now includes the last 100 log lines; it used to read only the last 100 bytes.

  • nala doctor --workspace no longer shows "all held" for cleanup invariants the daemon never reported. Each named invariant is checked, and a missing or malformed preview shows as not reported (WARN). nala workspace clean and nala agents prune no longer print a missing count as 0, and --workspace current no longer sends the word "current" as a workspace id.

  • list-panes and list-surfaces for a workspace that does not exist now fail with "no workspace found" instead of printing "No panes found." A workspace that exists and is empty still lists as empty. The same answer reaches the MCP pane_list / surface_list tools. Listing every workspace from outside a pane now fails loudly if the workspace list cannot be read, instead of listing nothing.

  • set-status and set-progress now update the workspace of the pane you run them from, and refuse outside a pane, instead of writing to whichever workspace was focused. They report which workspace the app updated, and fail when the app cannot find it.

  • nala config validate now actually checks your settings. It read a field the daemon never sends, so it reported every settings file as valid. It now names each invalid setting and exits non-zero, and says the result is unknown if the daemon's answer cannot be read.

  • nala voice test now fails when nothing was spoken: when voice is muted or disabled, no speech engine is available, synthesis fails, or the phrase was suppressed. It used to print the refusal and exit 0.

  • nala tool test and nala tool call now exit non-zero with the daemon's reason when a dry run finds missing arguments or an invocation fails, instead of printing the failure and exiting 0.

  • nala artifacts list now shows each artifact's kind and status instead of ?, and --status filters by kind (plan, review, ...) or status (draft, done, ...). --status terminal now returns done and cancelled artifacts instead of always returning none.

  • nala sessions list now shows each terminal's state, detected agent and folder instead of blanks, applies --limit (and says when it cut the list), and says when it could not narrow the list to your workspace.

  • nala attach without a session id now explains that it needs a structured session id. It used to list terminal sessions, none of which it could attach to.

  • nala daemon status now shows the daemon's version, and shows live and retained sessions separately ("3 live (46 retained)"). It used to print every retained row, dead ones included, as the session count, and 0 while the daemon was still starting.

  • nala activity subscribe --after now prints the new events after your cursor; text output showed only the snapshot.

  • nala wait-ready now gives up at once when the pane's process has died, instead of waiting out its whole timeout. When the daemon cannot be asked, it reads the agent status the app shows for the pane, so it no longer waits the full timeout on an agent that is already waiting for input.

  • nala browser navigate now reports the address the page actually loaded, such as the page a redirect ended on, instead of repeating the address you typed. When the page address cannot be read, the command says so and exits non-zero.

  • nala doctor no longer reports a starting or degraded daemon as healthy. It shows the daemon's own status, counts live sessions separately from retained ones, prints an unmeasured event-loop lag as unknown instead of -1ms, and shows the running app's version next to the CLI's own. Before, "app version" was the CLI's version, so a stale app build did not show up.

  • A daemon that is still starting no longer reports zero sessions. The desktop app could read that zero as "no live work", which would let it replace a daemon that hung during session recovery.

  • nala list-workspaces now marks the active workspace, and nala new-workspace exits non-zero instead of printing "Created" when the app did not confirm the new workspace.

  • nala fleet release-if-unreachable no longer says released=true when the board refused the release. The daemon answers with the release outcome, and a refusal (for example, the task is claimed by another agent) now exits non-zero, names the reason, and says the claim is still held.

  • nala fleet supersede no longer reports "old claim released and cancelled" when the old claim was not released. The replacement task is still created, but the command now exits non-zero and says the old task is still claimed, so the work is not done twice.

  • nala fleet complete now shows why the evidence check refused a completion, instead of a bare "board.complete rejected".

  • nala workspace clean, reset-runtime and forget with --apply, and nala agents prune --apply, no longer print "Applied" and exit 0 when the cleanup failed part-way. They exit non-zero and name the error the daemon reported.

  • nala send --submit, nala clear-composer and nala fleet send can see the input line again when the screen comes from the daemon, which is the usual case. The daemon's screen read left out the cursor rows that these commands use to tell text still waiting in the input box from text already sent, so they could not confirm a submit from what was on screen. clear-composer also stops treating the whole screen as the input line when those rows are missing, which made it report "not cleared" after any long transcript line.

  • nala doctor no longer warns about plain shells. Its pane check treated every titled terminal as an agent pane, so a plain PowerShell tab was probed (the write check typed a zero-width space into its prompt) and warned "surface omitted agentStatus". The desktop does send the field; it is empty because no agent runs there. The default scan now checks agent panes only, and a shell you name with --pane reports "none (no agent on this surface)". An agent pane with no status still warns.

  • nala doctor's host-authority line no longer says "never armed" when the daemon did not report it. An older daemon, or one still starting up, sends no host-authority field; that now reads as unknown instead of OK.

  • nala doctor's capacity check now reports what it can see. Every lease used to show its process as UNKNOWN, including leases that were already released (one real install printed 18 such warnings while holding no slots), and the cap was always "unknown". It now reads each lease's process from the daemon's latest observation, lists only leases that still hold a slot, counts the settled ones in a single line, and compares the busiest workspace against the cap the daemon enforces. nala workspace reconcile shows the same "slots held: N of cap" line.

  • nala update works on a normal Windows install again. When Node.js was installed on the machine, the nala command ran under that Node, and update refused with "install owner=unknown" because it judged the install by the Node program rather than by where Harmony is installed. It now checks where its own files live, so a Harmony install updates itself and a Scoop, winget or Homebrew install still defers to that package manager. nala doctor --install reports the owner the same way.

  • Four command flags that did nothing are removed, so passing one now gives an error instead of silently doing nothing: nala daemon logs --follow (it never followed), nala dictate --copy (it never copied), nala dictate --print and --until-enter, and nala config validate --file (it checked the app's settings, not the file you named).

  • nala daemon stop now reports whether the daemon actually stopped. It used to answer stopped: true as soon as shutdown was requested, before the daemon had exited. It now watches the daemon's process for up to 5 seconds, says "daemon stopped" only once that process is gone, and otherwise says it is still running or that the stop could not be confirmed.

  • nala sessions list --include-archived is removed. It was accepted and sent, but the daemon never read it, and there is no archive of sessions to list, so it never changed the result.

  • nala sessions attach-metadata is removed. It stored no metadata, because Harmony has no place to keep it. It re-attached the session instead, which disconnected the terminal's live output in the desktop app, and then printed "metadata attached". Its --metadata flags were also always refused.

  • nala rollback no longer claims a rollback that never happened. It ran the install script with a flag that only prints the installer's version, then reported "Rollback complete" while nothing changed, aimed at a fixed 3.23.4. Harmony cannot downgrade itself, so the command now says so and exits non-zero. It lists any earlier version still installed on this computer, with the exact program to run it, and the steps to install an earlier release. --yes, --version and --dry-run are gone, since they promised an install that never ran.

  • Documented command flags written with dashes are accepted everywhere. nala update --dry-run, nala rollback --dry-run, nala login --no-browser, nala desktop --no-workspace, nala run --max-turns, nala tool call --input-json and nala site stats --window-from were refused as "Unrecognized key" while the same error printed them as examples. nala sessions show <id>, nala config get <key> and nala harness replay <path> put their argument in the wrong place and were refused too. A test now runs every example in the help through the real parser.

  • nala cloud exit codes can be scripted again: an unreachable Harmony Cloud, a timeout, a busy server (429, 5xx) or a temporary Cloud pause now exits with the retryable code (11) and says to check nala cloud ls before retrying a change. These all used to exit 10 with "run nala doctor", and a refused connection was reported as the local daemon being down.

  • nala cloud no longer signs you out when Harmony Cloud is down or busy while it refreshes your sign-in. It keeps the sign-in and says to try again; only a server that actually rejects the sign-in removes it.

  • Deleting a Cloud workspace whose name is blank now asks for its id, in both nala cloud rm and the Desktop Cloud panel. Before, an empty field (or pressing Enter) matched the blank name and deleted it.

  • nala cloud ls, queue ls, new, rm and bring-home print server-sent names, statuses and prompts with escape sequences and control characters removed, as nala cloud status already did. A workspace name or queued prompt can no longer set your terminal title, write your clipboard or type a reply into your shell. --json output is unchanged.

  • The Cloud panel marks its Agent Hours figure "last known, not refreshed" when the account refresh fails, instead of showing an old number as current.

  • Device-code sign-in (Desktop and nala login) no longer spends one more poll on a code that expired while it waited, and an expired nala login code is reported as retryable, matching its exit code.

  • A signed-in desktop no longer gets signed out when two token refreshes cross (for example, waking from sleep with a flaky connection). When the server says the refresh already happened, the app now re-links from this device's enrolment and stays signed in; it signs out only if that re-link is refused.

  • nala login now saves your sign-in only inside a credentials folder that is proven readable by your user account alone: an owner-only access list on Windows, 0700 on macOS and Linux. If the folder cannot be made private, sign-in says so and saves nothing. An existing sign-in is picked up and moved on its next refresh. The help text no longer names a file mode that has no effect on Windows.

  • The CLI no longer fails with an expired sign-in after two refreshes cross. It re-links from this device's enrolment, as the desktop app does.

  • Saving a Harmony Agent export now writes only into a folder readable by your user account alone (an owner-only access list on Windows, 0700 elsewhere) and refuses otherwise. A development build saves into its own -dev folder, not the installed app's.

  • nala logout --json now answers in JSON when your sign-in comes from the NALA_TOKEN environment variable. Before, that one case returned plain text and broke scripts expecting the JSON envelope.

  • Harmony Cloud refusals from the new billing rules now say what to do. Hitting your own extra-hours cap, a used-up trial, a failed payment or an ended subscription each explain the next step in the app and in nala cloud, with a link, instead of "The cloud API refused: ". Running out of hours no longer suggests turning on extra hours, which can be bought only where the website offers them.

  • The Windows uninstall entry shows the Harmony icon. The installer fetched it from a private repository, which returned 404 on every install. Shortcut repair on install and update is now proven by a test that runs the real script, including against the zero-byte Start Menu shortcut an interrupted write can leave behind.

  • Quitting the app no longer signs you out. Shutdown used to revoke the session and erase it from secure storage, and a saved sign-in was never restored at launch; now the session survives a restart and sign-out is only ever an explicit action.

  • Provider launch buttons now explain that a workspace must be selected when launch is unavailable, including a tooltip hit area and accessible label.

  • Desktop update actions recheck install ownership before downloading or launching a staged installer, including after automatic checks are re-enabled.

  • CLI updates now refuse unknown install ownership and honor Homebrew receipts for linked app bundles before checking, downloading, or launching an installer.

  • Gatsby, Azure Glass and Magenta Glass get their own colours back, and every skin gets a pane header you can see. One :root block near the bottom of the stylesheet was overriding five theme blocks above it — :root and [data-theme="gatsby"] carry the same specificity, so whichever is written last wins. It took --border-soft (Gatsby's gold edge, Azure's blue, Magenta's pink) plus the two glass skins' tinted modal backdrops and coloured glow, which had been rendering as generic black. --border-soft also draws the terminal header's bottom edge, and the overriding value measured between 1.051:1 and 1.135:1 against the pane behind it in all thirteen skins — below the 1.194:1 this product already treats as invisible. The authored values run 1.202:1 to 1.805:1. The theme-independent defaults now sit above the skins, where a default belongs, and a guard reads the resolved value through a real cascade rather than searching the file for the declaration.

  • The Catppuccin skin paints upstream Catppuccin again. Its pane fill had drifted to #3E3F56 — whatever our derivation produces from the skin's surface colour — where Catppuccin publishes #45475A for that slot, and no arithmetic on the other tokens reaches it. A built-in skin ported from a published palette can now pin such a colour to the value its source states. The skin reads better for it too: the pane fill stands off its surroundings at 1.378:1 where the derived colour managed 1.227:1. The muted text colour is deliberately NOT restored — upstream's #7F849C measures 4.44:1 against Catppuccin's own background, under the 4.5 AA floor this product enforces, so it stays at the readable #969AAE.

  • The launch-recovery banner moves the terminal down instead of covering it. It was absolutely positioned, which contributes no height to its parent, so the box Terminal.tsx measures to offset the xterm mount collapsed to zero and the alert painted over the first lines of output. An earlier pass at this made the banner's background opaque, which fixed reading the banner and left the terminal hidden beneath it; position was the cause, not opacity.

  • The launch-recovery banner retracts once the agent is demonstrably running. An agent that came up after the readiness budget expired left a red alert on screen indefinitely, over a pane whose prompt was visibly waiting for input. The launch verdict is unchanged — a late detector publication still must not retroactively call a timed-out launch successful — but the banner stops asserting something that has stopped being true. A launch whose typed command was never accepted keeps its banner, because a busy-looking tab is not evidence the command landed.

  • The pane header says no signal instead of unknown when an agent cannot be seen. The state is still reported, deliberately: a blind agent renders identically to a healthy idle one on the activity dot, so the text is the only thing that can say we have lost sight of it. The word now matches the phrase the rest of the product already uses for the same condition.

  • /login codex reaches the Codex provider. People type the product name; the runtime registers openai-codex, so the login ran OAuth against a provider id that does not exist and reported a bare OAuth failed while the real provider sat in the picker one screen up. Names now resolve against what the runtime reports — exact id first, then a whole hyphen segment, a prefix, or a display-name word — and an ambiguous name is reported rather than guessed, because openai matches two providers that bill differently. An unrecognised name lists every registered provider and the API-key command instead of failing.

  • /login key <provider> <api-key> no longer stores a key under a provider that does not exist. /login key codex … wrote the key under the literal string codex, which nothing reads, then reported success and told the user to pick a model — so the key was lost and the failure was invisible. The provider is resolved against the whole catalogue first (every catalogued provider accepts a key, not only the four that register OAuth), and an unknown or ambiguous name is refused with the list rather than stored.

  • The /api/v1/me capability advertisement is built from code rather than from the request, and the site's schema identifier is derived from the canonical host instead of being written down a second time.

  • The launch card no longer lists filesystem, shell, network, destructive-Git and browser-automation as a permission profile. Nothing enforced those five: this card was their only consumer anywhere, and the single production consumer of any profile field branches on the Harmony level alone. hostSandbox grades this host filesystem PARTIAL and process, network and desktop NONE, so a row reading network: ask contradicted the daemon's own verdict two panels away. The honest report still ships in the harness inspector's sandbox tab. Name and Harmony Level, which are real, are unchanged.

  • Telemetry stops retrying an unreachable endpoint on a flat five-minute timer. An endpoint down for a day was 288 attempts per machine, each able to wake a laptop's radio to achieve nothing. Retries now back off, and a 4xx that drops the batch counts as reaching the site rather than as a network failure, so a bad batch cannot keep a machine backed off.

  • Persona CLAUDE.md writes canonicalize through realpath (and the parent for new files) before writing, so a symlink named CLAUDE.md cannot target a sensitive path. This is not sender validation on FS IPC and not workspace-trust wiring.

  • Accidental-tab detach no longer treats a missing daemon result as success/recoverable. An explicit ok: true is required. This is not sender validation on PTY invoke handlers.

  • The settings-reader census suite is named *.runtime.test.ts so parallel Vitest excludes it and the serial runtime config runs it alone. Do not raise timeouts to hide contention. This is not a product behavior change.

  • Provider capacity and retention no longer invent active for a missing lifecycleState. Only an explicit active occupies a slot or keeps a durable owner present. Approval digests and CLI omitted-lifecycle send are unchanged. This is not inspector chrome or a packaged rebuild.

  • A missing agent lifecycleState is unevidenced, not live. Roster reads, cleanup, delegation, and runtime binding require an explicit active before treating a peer as addressable. A live process on startup is evidence and promotes that row to active. Approval capability digests are unchanged. This is not inspector chrome or a packaged rebuild.

  • Cloud artifact downloads classify http_401 via statusCode, refresh the access grant once, and retry that handoff immediately. Transient, quota, offline, and hash/path errors do not refresh. Production uses session.refresh() and never returns the token to the bridge. This is not inspector chrome, a new IPC channel, or live Cloud.

  • Cloud artifact delivery records a classified last error (transient, auth, quota, offline, permanent) on the task row and on the existing task-event payload. Task completion stays separate from failed download. retryArtifactDelivery retries immediately after give-up. This is not inspector chrome, re-auth, or live Cloud.

  • Cloud send can keep a reviewed snapshot across an app restart. Bytes are AES-256-GCM under userData; the wrap key is OS secure storage. Preview after restart reuses that snapshot for the same account, window, and repo until it expires or is sent, including the new-file names used for selection. Sign-out deletes the files. Without a wrap key, restart still requires a new capture. This is not git stash create.

  • Bring Home adopts a leftover tree whose patch already classifies as applied even when the journal still says it is not. Recovery is not restored over that patch, so matching exclusive-create files are not left beside a rolled-back tracked file. This is a leftover-tree fixture, not a transactional worktree or a true process kill.

  • Cloud artifact downloads wait exponentially after a failure (5s, cap 5 min, eight attempts) instead of retrying on the next poll. Hash and path errors stop retrying and keep the pending id. Tests pin jitter at zero; production jitter is up to one quarter of the delay. This is not inspector chrome or live Cloud.

  • Cloud tasks keep reading the runtime event log after completion until a late artifact.sealed arrives or six empty pages say none is coming. This is not exponential backoff, inspector chrome, or live Cloud.

  • Cloud task registry writes use durable atomic JSON with .bak rotation. A corrupt primary is quarantined and the last readable backup is loaded instead of starting empty. Pending artifact downloads still retry after a process restart. This is not exponential backoff, inspector chrome, or live Cloud.

  • Cloud send remembers the declared handoff id on the in-memory review after create succeeds. Retry skips a second declare and resumes remaining blob PUTs. Restart still requires a fresh review. This is not encrypted persist or a disk journal.

  • Cloud send review now lets you choose which captured new files to include. Tracked patches stay together, an empty new-files-only selection cannot send, and pending sends lock the selection. Preview names come from the retained capture; the renderer never receives file bytes.

  • Cloud send keeps the reviewed capture when upload or seal fails before workspace creation, so a retry does not recapture unseen local edits. A second send while one is in flight is refused. Once workspace creation has started, the review is consumed even if the client saw an error. This is not encrypted persist across restart or a durable handoff journal.

  • Cloud send can name a subset of reviewed untracked files. Omitted files are not uploaded. A name that was not in the review is refused and the review stays usable. Omitting the field still sends every reviewed untracked file. This is not a selection UI, encrypted persist, or site runtime result-capture parity.

  • Cloud handoff capture confirms HEAD, changed paths, patch bytes, and untracked files after the read. A writer during capture is refused (repository changed during capture; review again) instead of sending a torn snapshot. This is not a coherent git stash create snapshot or encrypted persist across restart.

  • Bring Home invalidates inspection when its repository changes, clears older results before a fresh export, and releases controls after transport failures. Cloud terminal attach failures close allocated sessions; late connections cannot replace the selected workspace, and Interrupt is available while connected. Prompt queue failures offer Retry instead of appearing empty.

  • Cloud pane actions reserve space for zoom/supervision controls, and billing-refusal links use the real shell bridge. Cloud account/status races and keyboard menu actions are exercised. Launch profile writes refresh the bar through the registered RPC names.

  • Cloud handoff capture refuses staged and unstaged patches that contain high-confidence credential shapes (AWS access keys, GitHub tokens, private-key blocks, and similar). Untracked files with those shapes are skipped, not uploaded. Error text does not echo the token. Labeled password= assignments and ordinary filenames are not this scan. This is not complete secret detection or site runtime result-capture parity.

  • Bring Home does not record appliedPatch: true until exclusive-create files are copied into the user tree. A throw after patch copy restores the recovery snapshot and unlinks exclusive files created in that attempt, so retry can apply both. This is not an atomic rename of every path.

  • Bring Home writes exclusive-create files in .harmony/bring-home/isolated first, then copies them into the user tree with wx. A crash after those isolated writes does not create the files in the user tree. This is not a single atomic publish.

  • Bring Home applies the result patch in .harmony/bring-home/isolated at the recorded base, then copies touched paths into the user tree only if that tree still accepts git apply --check. A crash after the isolated worktree exists does not mutate the user tree. This is not a full transactional publish.

  • Bring Home snapshots patch-path bytes after the apply journal, before the isolated apply. A thrown apply restores that snapshot (concurrent bytes taken after the journal are kept). Success removes the recovery file.

  • Bring Home classifies an appliedPatch: false journal against the tree. If git apply --check still succeeds, retry replays the patch. If --reverse --check succeeds, retry skips apply and finishes exclusive-creates. If neither, retry is incomplete_apply and does not rollback. This is not a transactional worktree.

  • Bring Home takes a per-tree lock at .harmony/bring-home/apply.lock after git apply --check succeeds. A different handoff is apply_in_progress. The same handoff adopts. Success unlinks the lock and the journal. A thrown git apply keeps both. --check failure never takes the lock. This is not a transactional worktree.

  • nala activity subscribe --workspace and MCP nala_activity_subscribeWorkspace return a snapshot frame, then a catch-up page or a resync-required frame. An invalid cursor is not an empty success. Reading does not ack mail. This is not a push stream or inspector chrome.

  • Exclusive-create Bring Home files are written wx then fsynced. A tracked edit after the apply journal, before git apply, fails the apply, keeps the journal, and leaves the user's bytes. Retry is incomplete_apply. This is not a transactional worktree.

  • nala activity list --workspace accepts --query (MCP query) over summary, kind, source id, and actor id — never mail bodies. 10,000 fixture events page at limit 200. This is not a live subscription or inspector chrome.

  • Bring Home fsyncs the apply journal (tmp → fd sync → rename) before mutating the tree. A thrown git apply leaves that applying journal. This is not a transactional worktree.

  • nala activity list --workspace and MCP nala_activity_listWorkspace page durable messages and artifacts. There is no focused-workspace default. An invalid cursor is a resync refusal, not a successful empty list. Reading does not ack mail. This is not a live subscription or inspector chrome.

  • Top-bar agent launch reuses a single-flight PATH lookup and bounded concurrent stats so five clicks do not walk PATH five times. Provider settings changes refresh the bar without restarting the app. This is not a claim about external CLI startup time.

  • Bring Home adopts a clean dest worktree at the recorded base when a previous git worktree add finished and no apply journal was written yet. A plain folder is still worktree_exists. A dirty dest or a dest at the wrong HEAD is incomplete_apply and is not overwritten. This is a fresh apply into that tree, not a journal resume.

  • Bring Home can resume remaining exclusive-create files in an existing new-worktree destination when that dest still has a matching apply journal. This is not a transactional worktree.

  • After the system browser opens for checkout, the desktop polls that attempt on a bounded timer and again when the window refocuses. Completed is still only a paid GET (or explicit paid proof); expired/replaced stay pending. A completed attempt refreshes entitlements and subscription. Focus with no remembered attempt does nothing.

  • The workspace activity RPC pages durable messages and artifacts, including rows whose task is gone. An invalid cursor is a gap (resyncRequired), not a successful empty list. Reading the page does not ack mail. This is not a live subscription.

  • Asking the cloud to Bring Home now returns a request id. The desktop requires requestedAt on that receipt and does not invent a clock. A result stamped for a different request stays pending.

  • Bring Home resumes remaining exclusive-create files from a versioned apply journal after the patch is recorded as applied. Matching on-disk bytes are skipped, not overwritten. A crash during git apply, a leftover journal for a different handoff, and identity-less legacy journals still refuse. This is not a transactional worktree.

  • Stripe webhook no longer ACKs a routed billing event when the database URL is missing. The browser return from checkout is pending unless payment is proven. A paid Checkout Session marks that attempt completed without granting here. Desktop remembers the last attempt id for that poll. A Bring Home request stays pending until a result declared after it exists. Disputes are recorded and do not cancel access. Cloud restore checks the declared blob hash even without a response header. Desktop handoff skips only blobs the server already lists as uploaded, and a dropped cloud terminal reconnects with the last session id until the user closes it.

  • Cloud task event replay starts before sequence 0, so the first runtime event is no longer skipped. Empty event pages no longer mint a fake cursor.

  • Billing usage IPC fetches server rollups or fails. It no longer reports a successful empty list for a fetch that never happened.

  • Invoice history IPC fetches the server list or fails. A missing fetch is no longer shown as “no invoices yet.” Subscription and entitlement refresh fail closed instead of returning a silent empty cache.

  • Cloud PTY writes re-check the signed-in account, not only the window id. Switching the cloud workspace tears down the previous session.

  • The compiler gate includes the MCP TypeScript project, so an MCP-only type error can fail a commit.

  • Cloud PTY reassembles fragmented WebSocket messages and buffers output until the renderer attaches. Sign-out closes those sessions.

  • Bring Home returns ok: false when git apply cannot check or apply a patch, instead of throwing out of the shared executor.

  • A cloud task is not treated as locally committed until the daemon accepts a2a.task.update. A refused or missing daemon leaves a retryable pending sync instead of a completed task that never retries.

  • Sign-out clears billing display state, and a late invoice/seat/entitlement response from the previous account cannot refill the store.

  • A failed cloud artifact download stays queued after the remote task completes, so a later poll can still bring the files home.

  • Bring Home writes an apply journal after a successful patch check. A crash mid-apply is refused as incomplete_apply instead of silently writing a second half-tree. A patch that fails --check does not leave that journal behind. The journal path is confined like any other destination.

v3.23.4

Added

Fixed

  • Packaged updater self-version. The bundled runtime manifest now consumes the same build-time version injection as nala version. Installed copies no longer fall back to 3.20.0 and repeatedly offer their own release; missing build metadata is reported explicitly as 0.0.0 instead of silently impersonating an old version.