# Changelog What changed in each Hangar.AI release, newest first. The format loosely follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Versions follow [Semantic Versioning](https://semver.org/); before 1.0.0, a minor version could change behaviour as well as add to it. The 0.x history is grouped by minor version. Every release, with its installers, is on the [releases page](https://github.com/hangar-ai/hangar/releases). ## [Unreleased] ## [2.5.1] - 2026-11-19 ### Fixed - Codex: the reviewer set for a workspace (Settings › Agents, or the workspace's own choice) now reviews its deliveries, with an account of its own agent; it was always Claude Code. - Codex: accounts follow the agent everywhere. A Codex pane, reviewer or new workspace picks among Codex accounts, and the new-workspace dialog's account picker follows the orchestrator you choose. - Codex panes: the context gauge, the restart after an MCP change and Hangar Mobile's chat of a node's pane work for Codex too; a cloned repository gets your default orchestrator agent; the Discord presence no longer names Claude Code alone. ## [2.5.0] - 2026-11-19 ### Added - Codex (beta). HangarAI runs OpenAI's Codex CLI beside Claude Code, or on its own: Hangar works with either one installed, or both. - Setup: both agents on the first-run sheet, installed by their official installers and signed in from a pane; either one is enough. - Workspaces: choose the orchestrator's agent and the workers' agent when you create one; mixed teams work (a Claude Code orchestrator with Codex workers, or the reverse). Settings › Agents sets the defaults, Codex's launch command and models. - Panes: a Codex pane shows whether it works, waits or needs you, like a Claude Code pane. Hangar attaches to the pane's Codex server to follow its turns and answer its approvals (Settings › Attach Hangar to Codex, on by default); without it, the pane still runs, with less to show. - Board, memory, kits and Hangar's tools reach Codex too, through `-c` options: your `~/.codex/config.toml` is never edited. The playbook goes to `AGENTS.md` as well as `CLAUDE.md`. - Accounts: Codex accounts beside Claude ones, each in a home of its own, chosen per pane, workspace or folder. - Images: what Codex generates shows in the chat view, on Hangar Mobile and in Discord. - Delegate: an agent hands a sub-task to a Codex or Claude Code child that runs with no terminal, in its folder or a worktree of its own, and gets back its answer, files, images and branch (`delegate`, `delegation_status`, `delegation_cancel`, `delegation_list`). - Also: sessions resume and sleep for either agent, Hangar Mobile answers Codex's questions, Discord follows Codex panes, nodes run Codex, statistics per agent with OpenAI's prices, the reviewer can be Codex. - Codex is in beta: report what goes wrong with the « Codex bug » issue form. ### Changed - README, guide and site speak of both agents. ## [2.4.6] - 2026-11-19 ### Fixed - hangarai.dev also shows in the title bar with the navigator hidden and in workspace windows. ## [2.4.5] - 2026-11-19 ### Added - Pin a pane: a pinned pane keeps its edge and its size while workers and other panes open and close, and presets never close it. Pin and unpin it from the pane's menu or the command palette; a pin marks it in its header. - Window opacity, for every theme (Linux): Settings › Appearance › Window opacity lets the desktop show through the window, the panes and the terminals, while each theme keeps its own colours. Blur is the desktop's: KDE does it, GNOME with an extension such as Blur my Shell. ### Changed - HangarAI: the new name, with hangarai.dev always shown in the title bar (a click opens the site). The program, its folders and its settings keep their names, so nothing moves on an upgrade. - Contact: contact@hangarai.dev, for questions about your data, conduct reports and security reports by email. ### Fixed - Discord: a message written in a channel while its agent was busy could lose its place in the queue, and its ⚡ then did nothing. ## [2.4.4] - 2026-11-19 ### Changed - Hangar's address is now hangarai.dev: the website at https://hangarai.dev and the optional Hangar account's server at https://account.hangarai.dev. The old addresses (hangar.simnjs.fr, hangar-account.simnjs.fr) keep working and lead to the new ones. ### Fixed - Typing in a text field while a pane was selected went to the pane: the "Send the same instruction to all panes" bar, the terminal's search, renaming in the navigator and adding a card on the board lost the keyboard to the terminal, often after the first letter. A field that takes the keyboard now has it until you click a pane again, everywhere in the app. ## [2.4.3] - 2026-11-18 ### Added - Coloured badges for Claude accounts (Settings › Accounts, off by default): a colour and an icon per account, shown on the toolbar's account button and its menu, and wherever an account other than the default is named. Contributed by SimonB44470. - A Claude account per navigator folder: workspaces without an account of their own run on their folder's, and follow it when moved; an explicit choice, "Default" included, still wins. A "+" on a folder's heading creates a workspace in it. Contributed by SimonB44470. - With two Claude accounts or more, the navigator's plan block shows each account's own limits and next reset. Contributed by SimonB44470. ### Changed - Choosing another Claude account no longer asks to restart running agents: each one moves by itself, on the same conversation, as soon as Hangar knows it has nothing to lose (its turn over, no draft in its prompt, no command of its own still running); meanwhile its header shows where it is going and why it waits. "Ask before moving running agents" in Settings › Accounts brings back the question. Contributed by SimonB44470. ## [2.4.2] - 2026-11-18 ### Changed - A quieter orchestrator on big projects. A worker waiting on a job of its own (a build, a benchmark, a CI run it watches) is no longer reported idle every ten minutes. A worker that is really idle is reported once, then again only after it did something or half an hour later. News that can wait (plain comments, idle workers) comes as one digest every five minutes at most, while deliveries, questions, a worker waiting for you or a crash still arrive at once. ### Fixed - Enabling a kit works again. With its activation sheet open over a workspace's Kits dialog or over Settings, the click on "Accept and enable" went to the dialog underneath: the sheet closed and nothing was enabled, and in Settings the click could flip another switch. Only the dialog in front now takes the mouse and the keyboard. ## [2.4.1] - 2026-11-17 ### Fixed - The Hangar account reaches its server at `hangar-account.simnjs.fr`: 2.4.0 pointed at an address that does not exist, so signing in failed. ## [2.4.0] - 2026-11-17 ### Added - **An optional Hangar account.** Settings › Hangar account (« Compte Hangar ») lets you sign in with GitHub, in your system browser, to link your machines. It is off by default: Hangar works as before without one, never calls the account server until you sign in, and still sends no telemetry. Signed in, the page shows who you are, your machines (name, system, last seen, « This PC »; rename one, remove another with a confirmation), Sign out, and Delete my account (type the word to confirm; it deletes everything on the server, then signs this PC out). A « Sync statistics » switch, off until you turn it on, sends totals per machine, day, model and project name (tokens, active time and cost; never a path, a prompt or any code) at once, every 15 minutes, when you open Statistics, and on « Sync now », without ever waiting on the network in the window. A day's totals already sent are never replaced by lower ones, so Claude Code cleaning old transcripts does not shrink your history. Errors read as plain sentences (server unreachable, signed out from another machine). The token stays in your system's keychain. The details are on the [privacy page](https://hangar.simnjs.fr/privacy.html) and in docs/native/cloud.md. - **Statistics of all your machines.** Once synced, the Statistics page has a « This PC / All machines » switch. This PC is the page as it was; All machines shows the merged totals of every machine of the account by machine, day, project and model, for the period chosen, with when the account was last synced. - **The account server, ready to run with Docker.** `ghcr.io/hangar-ai/hangar-cloud`, built, checked and published by CI, and `crates/hangar-cloud/compose.yaml` to run it behind a reverse proxy or through a Cloudflare Tunnel (`docker compose up -d`), with its health check, a read-only root and no capability. For whoever hosts it; docs/native/cloud.md, « Deploy », walks through it. - **Hangar on NixOS.** A Nix flake at the root of the repository: `nix run github:hangar-ai/hangar` builds and starts Hangar with its webview helper and Hangar Mobile, its package goes in a NixOS or Home Manager config, and `nix develop` opens a shell with everything a build needs. On Nix, the MCP configs and the login item name the `hangar-ai` on your `PATH`, not the store path, so they survive upgrades and garbage collections. Nodes are not in the package yet. Contributed by Kisakay. - On NixOS, the **Set up Claude Code** sheet offers Claude Code from nixpkgs instead of Anthropic's installer, whose `claude` does not start there without nix-ld: the line for the configuration (`environment.systemPackages = [ pkgs.claude-code ];`) and a `nix profile install` command, to copy, with what claude-code being unfree asks for, then **Check again**. The installer stays behind **Use Anthropic's installer anyway**. ### Changed - **Discord: a thread per worker, shared files, and ⚡ to send now.** A worker's thread opens by itself when it starts (its « Ouvrir le fil » button stays only if Discord refuses it), a file an agent shares (`phone_share`) is posted in its thread, or the workspace's channel for the orchestrator (one too big for Discord becomes a line saying it is on the phone), and a message waiting for the agent's turn to end gets a ⚡ that you click to stop the turn and send it at once, as « Send now » does in the chat view. - **The viewers open any file on the computer.** Images, PDFs, Markdown, code, video, web pages and 3D models shown by an agent (`preview_show`, `model_show`) or by you no longer have to be inside the workspace's folders, and a workspace whose folder is not a real path shows files too. What the viewer pages may load stays limited to what was shown (and, for a 3D model, the textures and buffers beside it): a page cannot read the rest of the disk. ## [2.3.0] - 2026-11-13 ### Added - **A Hangar with no window, on a server.** A node runs workspaces of its own now, with no desktop connected: their state is kept on the node, their panes start there with Claude Code, and their agents' requests (workers, deliveries, the board) are answered there, so they keep working with your computer off. Make one with `hangar-node workspace add ` on the server (or from Discord); `hangar-node workspaces` lists them, and a connected desktop lists them in its navigator among the others, opens them with their terminals live, and changes them there (a rename, a pane opened, closed or restarted, a preset, a deletion: each is sent to the node, which stays their owner; while the node is offline they can be looked at, not changed). Merge on one of their deliveries and Stop workers are carried out by the node. A workspace made from the desktop with « Runs on » set to a node is the node's own from the start, and one made before, which still stops with your computer, has « Move to {machine} » in its menu in the navigator: its agents are restarted on the machine on their sessions. If the machine's answer is lost on the way, the workspace waits as it is and the machine's own list settles it at the next read; and a workspace a machine came back without after a restart is kept in the list, marked, with « Run on {machine} again » in its menu. `hangar-node service install` makes the node a systemd service that starts on its own. - **The same repository on several machines (nodes, experimental).** Hangar knows a repository by its `origin`, however the address is spelled. When one is open on this computer and on a node, each of its rows says so in the navigator (a faint mark on a local row, « also on… » in the tooltip of a machine's mark), and its menu starts with « Also on », which goes to the other checkouts. « Open on another machine… » in that menu opens the new-workspace dialog ready to clone it where it is not yet. In that dialog « Runs on » is there for a clone too, and lists every machine (one that is offline or switched off is greyed, with its state; from four machines on, they are a menu): the node clones with its own keys, into a folder you pick there (`projects` in its home by default), shows its progress, and runs the workspace itself. If it cannot clone, git's own words are shown. - **Hangar on Discord (experimental).** Control your Hangar from a Discord server of your own, as from the phone: a channel per workspace (those on your nodes included), its orchestrator's turns as cards and its answers under its name, each agent with its own face; what you write in the channel goes to the orchestrator, waits in its queue while it works (⏳, then ✅ when it is taken), and an edit or a delete changes what waits; permissions, questions and plans as cards answered with buttons; workers announced with « Ouvrir le fil », which opens a thread with their work since they started, then live, where you can talk to them; deliveries with Land and Send back; a pinned dashboard; `/stop`, `/model`, `/effort`, `/plan`, `/status` and `/workspace`. Only the account you pair with a code is heard. Turn it on in Settings › Experimental, then give the bot's token and your server's ID in Settings › Discord; it runs while Hangar runs, window closed or not. The same bot serves the workspaces your nodes run by themselves: the desktop hands it to them, and they keep answering in their own channels with your computer off. A node with no desktop can have a bot of its own (`hangar-node discord setup --guild `). - **Answers reach Claude Code as answers.** A permission allowed or denied, or a question answered, in the chat view or on the phone now goes to Claude Code as the decision itself, through the hangar-bridge mod, as an Agent SDK host's would: no keys typed into the dialog, no screen read back. The dialog stays up in the terminal meanwhile and either place can answer; a pane on a node, or a Claude Code without the mod, is still answered with keys. - **Stop in the chat view.** While the orchestrator works, a button next to Send stops its turn, as Escape does in the terminal. - **Model, effort and plan mode from the chat view's input.** Next to the input, the orchestrator's model, its effort level and Plan: a click shows the choices, and the choice reaches its Claude Code as `/model`, `/effort` or `/plan`, at once when it is idle, at the end of its turn otherwise. The hangar-bridge mod reports what the session runs with as it changes, so the input shows it. As in the terminal, `/model` and `/effort` also become the defaults of new sessions. - **Files Claude names are links, in the chat view.** A file of the workspace that Claude names in a message (`src/search/index.ts:42` in code, or a link to a relative path) opens in the workspace's preview, beside the conversation, with a click. Only names that are files there become links; code blocks are left alone. - **A queue you can see and change, in the chat view.** What you send while the orchestrator works now waits in its queue instead of being typed into the turn, and shows at the top of the input: each message can be edited (back into the input) or removed while it waits, and Send now stops the turn so the first one goes at once. Claude takes them when its turn ends. Through the hangar-bridge mod; without it, messages are typed as before. - **What a delivery brought, in the chat view.** A delivery that landed or waits to land shows how many files it changed and their +/− on its line; a click lists the files, and a file opens the Changes window on exactly what that delivery brought (its commits, not the working tree). The orchestrator is told the same range (`git diff ..`), so it can look at the changes too. - **Plans in the chat view.** When the orchestrator presents a plan (plan mode), the chat view shows it whole, formatted, in a card with Approve the plan and Keep planning; a field says what to change in it, and Claude plans again with those words, the turn going on. Through the hangar-bridge mod. ### Changed - **Several machines, at a glance (nodes, experimental).** A machine goes by one icon, its kind's (server, cloud, laptop, desktop, mini PC, tower), chosen when it is added and changed from its row. In the navigator a workspace on a node wears its machine's icon, the filter finds it by the machine's name or narrows the list to one machine, and a line at the foot counts the machines and, in the warning colour, those that are offline (those switched off are in its tooltip); the palette and the new-workspace dialog name machines the same way. A node now connects as Hangar starts and reconnects for as long as it is down (2 s doubling to a minute, at once when you come back to the window or press Retry now; an attempt that hangs is given up after a deadline). A first connection that takes a while, an install for instance, reads as connecting, not offline. A link that drops says nothing for 5 seconds; then its rows step back with an unplugged mark, and its panes say why with Retry now and, after 20 seconds, Switch off, which hides the machine's workspaces, in Needs-you too, until it is switched on again in Settings › Nodes. That page lists the machines as rows, this computer first, each with a switch and a menu (Retry now, Rename, Remove from this computer). A failed connection reads as one short sentence (« Offline · retrying in 8 s: unreachable over SSH (Connection timed out) »), with what ssh said under it. - **The chat view reads like a chat.** Claude's words sit open on the page, your messages in rounded bubbles on the right; its calls are quiet lines with an icon for what they do (a command, an edit, a read, a search, a task), folded in runs as before. While it works, a line at the bottom says so and for how long, and the header gives the model and how full its context is. The input is one rounded composer; while a question or a permission waits, its card takes the composer's place, with only the question being asked (no recap of the answers given before: Back goes to them). - **The orchestrator's exchanges with its workers, short and in your language.** What Hangar tells the orchestrator about the workspace shows as one line per thing that happened, with an icon and a colour for what it means (Ada delivered “Search index”: merged, task done; Bo started on “Filters”; Cy has been waiting for you for 4 min) instead of the block written for the model. A worker's message to the orchestrator and the orchestrator's message to a worker each show as one line with their first words (Bo: “…”, To Bo: “…”), the whole message a click away, and a message is never folded into a run of calls. The state of the work handed to a conversation that starts is one line too. ### Fixed - **A long question could not be answered from the chat view or the phone.** Claude Code now draws a bar down the left of a question's lines; when a question wrapped over several lines, Hangar no longer recognised it on the screen and refused to type the answer ("The question is not on the PC's screen…"), so the first, short questions went through and the longer ones that followed did not. The answer now goes through the mod with no screen involved, the screen check ignores the bar where keys are still typed, and a terminal the chat view hides starts at 160×50 instead of 80×24, so a long question is drawn whole. ## [2.2.1] - 2026-11-08 ### Added - **A delivery's progress is on its worker's pane.** A chip says where it stands (queued, merging, checking with the time it has taken, in a batch of N, reviewing, landing) and then how it ended; the chat view's tasks panel shows the same on the task. - **Deliveries are checked together.** When several wait, Hangar merges them on the branch and runs the check once for all of them; if it fails, it splits them to find the one that breaks, and only that one is handed back. The branch stays exactly as safe: nothing lands on a tree the check did not pass. With five deliveries waiting and a one-minute check: 73 s instead of 315 s when all pass. The batch size is in the orchestration settings (4 by default). - **Claude Code is set up with Hangar.** When it is missing, a first screen installs it in one click with Anthropic's official installer (its output shown as it runs), then opens Claude Code to sign in; a workspace cannot be created without it, unless you choose to use Hangar with shells only (Settings › General). ### Changed - Releases ship Windows and Linux. macOS is no longer built for download: 2.2.0 is the last macOS release, and the README says how to build Hangar on a Mac. ## [2.2.0] - 2026-11-07 ### Added - **A new workspace starts with its orchestrator.** Creating a workspace now gives one Claude Code pane that is the orchestrator, ready to plan and hand work to workers; the layouts of several panes and shells are under More options. In a folder without git it still orchestrates, its workers sharing the folder (git init gives them worktrees). - **Chat view (experimental).** Turn it on in Settings › Experiments, then switch any workspace between its panes and a chat with its orchestrator. Claude Code keeps running in the panes behind it; a side panel lists the board's tasks with the progress their agents report, and kits, previews and the 3D view open beside the chat as they do in the grid. - Agents can report how far along a task is (`board_report_progress`, over MCP and HTTP): a percentage and a short note, shown with the task. Workers are asked to report at each real step. - **Nodes (experimental).** A headless Hangar, `hangar-node`, runs on a Linux machine you reach over SSH: add it in Settings › Nodes and Hangar installs it, starts it and keeps a tunnel to it. A workspace can then live on the node: its terminals, Claude Code and board run there, and the window shows them like a local workspace's, catching up on what happened while it was away. Its memory, transcripts, chat view and orchestration work too: the orchestrator and its workers run on the node, in worktrees there, while the connected window drives them. A local workspace can also send a worker to a node (`worker_spawn` with `node`), for a heavy build or a GPU box: the worker runs there and its branch comes back to be checked, reviewed and landed here. Design in `docs/native/nodes.md`. - **Pane scopes (experimental, Linux).** Each pane can run in a systemd scope of its own, so when memory runs out the system stops the pane that uses it (a build, say) rather than Hangar and every agent with it. ### Changed - Released builds use link-time optimization across the whole program, rarely used code is built for size and unused image codecs are left out: Hangar's executable is 44 % smaller (119 → 67 MB, hangar-node 17 → 10 MB) and parses terminal output a few percent faster. TIFF and ICO images no longer show in kit tabs. - Previews are lighter: a large image is shrunk when it is decoded and read in full only when zoomed (an 8000×6000 PNG: +554 → +143 MB), a Markdown page no longer colours its whole source unseen (38 % → 4 % CPU) and parses off the interface thread, memory freed by a closed preview goes back to the system, and the web view helper stops a minute after its last window. - At rest, the panes no longer wake the window several times a second: about once every 20 s (43 wake-ups a minute → 3 with a shell pane), and Hangar's CPU at rest falls by about 40 %. - Drawing costs half as much or less: a busy terminal 0.97 → 0.48 ms a frame, 16 printing panes 2.7 → 1.3 ms, scrolling back through history 4 → 1 ms (a scroll plans only the rows that came into sight), selection 3.9 → 0.9 ms, the statistics view 1.8 → 0.95 ms. The terminals' lock no longer makes the window wait behind panes that print. - Output from panes that are not on screen no longer makes Hangar rebuild and redraw its window 60 times a second: with eight agents printing, Hangar's CPU falls from 8.1 % to 2.9 % in the chat view and from 6.4 % to 5.8 % in the panes view. - On relaunch, the next wave of resumed sessions starts as soon as the previous one is under way (its agent printed and went quiet) instead of a fixed 2 s later: eight sessions in the chat view are all back in 3.2 s instead of 9.1 s. Agents that keep printing still hold the next wave as long as before. - The login shell's PATH is read from the first instant of the launch, beside the window's start: with a slow shell profile (0.35 s), the first prompt comes about 90 ms sooner. - In the chat view, the hidden terminals start their shells at once instead of waiting for a size they never get: a workspace in the chat view is ready in 0.38 s instead of 0.66 s. - On Linux with a Vulkan GPU, Hangar no longer also sets up an OpenGL context it never draws with: about 27 MB less memory at rest and a faster start (0.40 → 0.36 s with 4 shells). A `WGPU_BACKEND` you set yourself is kept, and the one Hangar sets does not reach your panes' shells. - Terminal output is parsed faster (+6 % on a listing, +8 % on an agent's interface): the scan for title and link sequences skips plain text in one search. ### Fixed - Chat view: a question with several parts is asked one part at a time, as Claude Code asks it (1/3, Back, Next, then your answers and Send), its card scrolls instead of overflowing, and a missing answer or a refusal is said beside the buttons instead of Send doing nothing. Runs of tool calls fold into one quiet line ("3 commands, 2 tasks created · 1 failed") that opens on a click. - Windows: a pane whose folder is longer than 258 characters starts its shell (in the folder's short 8.3 name) instead of failing with "The directory name is invalid". - Windows: Hangar starts on a Windows without the Visual C++ Redistributable. The installer now puts the runtime's three DLLs beside the executables; before, they exited at once (STATUS_DLL_NOT_FOUND) on a fresh install. - A Hangar started in a folder inside a worktree no longer removes other workers' worktrees: the sweep of orphan worktrees at start only removes the worktree of a task that is done, and never one a live process works in; landing a delivery and stopping a worker keep a folder another pane or process uses. - Kits, Windows: removing a kit's clone folder retries for a moment when a git process that was just stopped still holds it, instead of failing. - Tests remove the temporary folders they create, even when they fail; a full test run used to leave thousands behind in `/tmp`. ## [2.1.0] - 2026-11-06 ### Changed - On relaunch, the Claude panes on screen resume their conversations first; the others follow two at a time, or at once when they come into view, so a workspace full of big sessions no longer starts every Claude Code at the same moment. - A change to the app's state no longer copies and compares the whole state on every action, the state snapshot is rebuilt only when it differs instead of on each keystroke, and saves are serialised off the UI thread. - Less work while agents run: the session hunt lists transcript file names and dates only instead of reading the head of every transcript every few seconds, skips panes the mod already tracks, and settles to a 5 s cadence for resumed panes; the session picker reads at most 256 KB of each transcript and caches its labels. - The context gauge and the phone's chat read transcripts without a lock shared by every pane, stream a transcript instead of loading it whole, and find it under `CLAUDE_CONFIG_DIR` and account profiles too. - Kits ask again for your consent when a script, hook or MCP server file they run has changed since you allowed them, not only when their command lines change. An update is checked in full (manifest, permissions) before it replaces the installed kit, and a kit installed from a tag notices when that tag later points elsewhere. - Each pane gets an API token of its own, kept across a restart or a hibernation and revoked when the pane is closed; a pane's token fixes who the caller is, so an agent can no longer act as another pane or as you through the local API. `~/.hangar/endpoint.json` now advertises a token for outside callers instead of Hangar's own. - The terminal's scrollback is capped so a pane cannot hold tens of MB of history; box-drawing characters are cached instead of being redrawn from scratch every frame, and an open search no longer rescans the whole history on every chunk of output. - The command palette filters once per keystroke instead of on every redraw; a native kit tab compiles off the UI thread; pictures, highlighted code blocks and the stats page are computed in the background. ### Security - `~/.hangar/endpoint.json`, which holds the API token, is readable by its owner only (0600, folder 0700); trusting a folder keeps `~/.claude.json`'s permissions. - The updater keeps its download in a private folder and checks the signature again right before installing, so the file cannot be swapped in between. - Phone actions (send, allow, answer, land) carry a key and run once even when the phone retries; allowing a permission from the phone applies only to the prompt the phone showed. - Voice models are pinned and checked against their SHA-256 before use. - A kit tab cannot make Hangar hold unbounded logs, images or messages, or block the window with a long chain of updates. ### Fixed - Hangar no longer replaces a `CLAUDE.md`, `.gitignore`, `board.json`, `.mcp.json` or `~/.claude/settings.json` it cannot read or parse (not UTF-8, unreadable, invalid JSON): it leaves the file alone and says so; a corrupt board is kept aside as `board.json.corrupt.