Skip to content

Agent Durbar v0.1.9

Choose a tag to compare

@github-actions github-actions released this 04 Nov 09:40
· 3510 commits to main since this release
v0.1.9
da3dc75

v0.1.9 in 45 seconds

Unsigned build of Agent Durbar. Not code-signed or notarized: see the README for the macOS Gatekeeper and Windows SmartScreen steps. Needs at least one agent app signed in: GitHub Copilot, Claude Code or OpenAI Codex (Settings > Agent apps can install it). Independent project, not affiliated with GitHub, Microsoft, Anthropic or OpenAI.

v0.1.9 (release, 2026-11-04)

Highlights

  • The agent rail: every agent beside the chat, by folder, and who needs you first.

All changes

  • Still highlights (no motion) (spec 001 FR-192, issue #780, WCAG 2.2.2): Settings > Highlights has a switch that keeps the coloured glow and the New badge on desks that need you or have a new reply, but stops the movement that goes with them: the person's jumping and pulsing, the glow's breathing, the badge's pulse and the waiting badge's pings, the waiting person's waving hand and the tap on the desk after a minute. In the Real office the whole room stands still. It does not stop every animation (the system's reduced motion setting does that; some small ones, such as a theme's own figures, still follow it). Off by default and saved on this computer. A failed save says so and the switch shows what is really saved. durbar_settings (MCP) now reports stillHighlights next to seriousMode. Checked in check-unread (simulator, both sizes), test:stillhighlights (main's handler on a stand-in electron with a real store) and test:ipcchannels; not run on a packaged app, Windows or macOS.
  • Long chats open faster, and keep following the end (spec 005 FR-210..FR-215): when you leave a chat, its older messages are marked with their measured height, so reopening a chat of 4,000 simulated messages lays out only the ones near the view (the timing of each run is printed by the driver, never asserted; a long history opened for the first time is not faster, nor is a chat closed from its Browser tab). The marks are dropped when the width or a layout setting (density, theme, font) changed (FR-213), and for a message that changed while the chat was closed (FR-215). A chat scrolled up, even by a little, is no longer pulled back to the end by a streaming turn, and a late scroll event no longer stops the chat from following a turn (FR-214). Scrolling, copying across messages, find and keyboard focus are unchanged. Simulator only; not run on Windows or macOS.
  • Test driver only (issue #1102): the long-chat visibility check no longer fails on a loaded machine or on a 2x screen. No change for users.
  • Test driver only (issue #1127): the long-chat visibility check runs one window size per CI slot, so it no longer hits the suite time limit. No change for users.
  • The agent rail (spec 005 FR-190, FR-194, FR-216; spec 052 FR-215, FR-216): with a chat open, a rail on the left replaces the strip at the top. It lists every agent by folder, with the folder's count; whoever needs you comes first, in a taller row that says how long they have waited and what they ask, and their folder moves up. Your door at its top shows the Needs you count and opens the queue. A folder collapses and expands with a click, Enter, Space or a double click, still showing its count and how many in it need you; opening a chat in it expands it, and the collapsed folders are remembered per window. On a Mac the rail's top row and the chat header move the window. Checked in the simulator by check-rail at 1440x900 and 1280x800 (all 15 agents, no page overflow, two row heights, contrast) and test-railorder; not tried on a real Mac or Windows.
  • A short clip of a release's changes, as a trial (spec 008 FR-171..FR-178, owner request): npm run release:video -- --version <tag> --out <dir> records 30-60 s of the simulator (title, one captioned scene per chosen change, end card) with the captions in a band under the app, and writes an MP4 and a GIF (ffmpeg needed), refusing to finish when a scene's text shows a local path, the user or host name or a token shape. A fragment may now end with optional Highlight: and Scene: lines (changelog.d/README.md), which the fold drops. The release notes now start with a Highlights block built from those lines, and the release steps in changelog.d/README.md make the clip part of the flow, starting with v0.1.9 (the first clip is shown to the owner before it is published; older releases get none). Team tooling, no app change, 0 AI credits. Checked by test:releasevideo and a v0.1.7 trial on Linux; not run on macOS or Windows.
  • Open with menu in the chat header (spec 009 FR-019..FR-027, #685). The header's Open in VS Code button is now Open with: it lists the installed VS Code, VS Code Insiders and VSCodium, and Show in folder for the system file manager. It opens the agent's folder only and says the conversation stays in Durbar; Continue in another app… (More menu, when a second agent app is installed) is still the way to hand it over. A folder that is gone or a file is refused with a message; a failed launch says so. Something the file manager would run is selected, not opened. A project folder that is a symbolic link is opened at its real path, with a note saying so; a link to a file or to nowhere is refused. The folder status dialog's Open in editor and Continue in VS Code go through the same folder checks and use the first installed edition that can open the folder (on Windows, the next one when the first is only a .cmd and the folder's name has a character cmd.exe could misread); the profile editor's button is renamed Open in editor but opens the profile file as before. Simulator, unit tests and a driver at 1440x900 and 1280x800; no editor or file manager was launched by a test, and the one real check (a folder opened in Finder) ran before the later changes.
  • Agent events are checked in the app (spec 010 FR-035, issue #1039): every event from Copilot, Claude Code or Codex now passes the same shape check the tests use before the window or any other part of Durbar sees it. A malformed one is dropped, with a log line that names only the event type and the fields at fault, never their values. Checked with test:runtime-core and one short real session per app (149 events, none dropped).
  • Settings: "Saved" only when it was saved (spec 011, issues #1094, #1047): asking again for a value already in force while the settings file cannot be written no longer says Saved, and a failed save no longer leaves a .tmp file next to the settings. Checked by test:credits (injected write and rename failures, mutation-checked).
  • Test driver only (issue #1136): the Codex stop-escalation test waits for the signals instead of a fixed sleep, so it no longer fails on a loaded machine. No change for users.
  • No MCP config left in the temp folder (spec 012, issue #1093): the private folder holding mcp-config.json (MCP env and header values) for a Claude session is now removed also when the app quits while Claude still runs; every other ending (spawn failure, error, kill, close) was already covered and now has a test. npm test no longer leaves durbar-* folders behind and fails a suite that does.
  • A helper started in the background no longer reads Done while it works (spec 015 FR-018, issue #1141): when an agent launches a helper in the background (Claude Code's run_in_background, Copilot's background task), the agent's own call returns at once; the helper's chat block, card and transcript now stay Running until the helper itself finishes or fails, and a helper still running when its session stops reads Failed ("Stopped with its session"). Foreground helpers are unchanged. Checked with mapper tests and a simulator driver; not seen on a real session.
  • Auto-follow work, opt in (spec 052 FR-182..FR-190, issue 686): a new switch in Settings (off by default, kept on this computer; main answers a real ok) makes a sent message take you back to the office with that agent's desk marked and focused, and open its chat (focus on the chat title, never on Allow or Deny) when it finishes, asks, errors or reaches its limit. It never answers or approves anything, waits while you type, have a dialog or menu open, or the window is in the background, stops following when you open another agent yourself, and respects reduced motion (no animation). A strip under the top bar shows Previous and Next over the attention queue (the Needs you inbox order, then unread replies) with the position and count (keys [ and ]), a Pause button, and, when nothing needs you, Next returns to the office with the last desk focused. Checked in the simulator (check:ui check-auto-follow, both sizes, 15 agents) and test:autofollow; not run against a real agent session.
  • Needs you text no longer runs words together (spec 052 FR-214): the inbox, the answer bar, the announcements and the Real office slips read "search_companies) Allow for this session ...", not "search_companies)Allow ...", and a card with several questions shows all of them. Checked by the simulator driver check-inbox-text.
  • MCP sign-ins kept fresh (spec 058 FR-204..FR-217, first part): before a chat's first message after an hour without a reply (or a wake from sleep), Agent Durbar asks Copilot or Claude Code to reconnect the chat's remote MCP servers, so the app renews an expired sign-in with its own refresh token. The message waits at most 5 s under one quiet line; Stop sends nothing; a server that still needs a sign-in gets a line with Sign in…. On by default (Settings > MCP servers); in Copilot's sandboxed modes (Read-only, Ask first) only the sign-in line, since a reconnect there did not renew the sign-in (Linux probe); Durbar never sees a token; no local server, no Codex chat, on macOS only servers that already need a sign-in. Simulator, fakes and unit tests; the apps' behaviour probed on Linux.
  • Organizations can require the OS sandbox on macOS and Linux (spec 076 FR-209, issue #734): Durbar reads the machine policy.json (/Library/Application Support/AgentDurbar/managed/ on macOS, /etc/agent-durbar/managed/ on Linux) for permissions.requireOsSandbox. Only a root-owned file in root-owned, not group- or world-writable folders counts, checked along the path as written and through any link on it. A policy folder whose file is missing or broken also requires the sandbox, and the log says why. Windows is not read yet. Checked with test:sandbox-floor; no real session.
  • Agent-app updates: a vendor "updates are disabled" answer is no longer shown as done (spec 081, #662): on Linux, claude update with updates disabled exits 0, and Durbar now reports it as Managed by your organization. The two Claude fixtures are recorded shapes; checked on Linux with an isolated copy, macOS and Windows still open.
  • Copilot gets a private temp folder (spec 087 FR-023, issue #755): Copilot's sandbox always keeps a temp folder writable. Durbar now starts the Copilot runtime with a private one (a new durbar-cprt- folder, mode 0700, removed when that runtime stops or the app quits) instead of the shared temp folder, with TMPDIR (and TEMP/TMP on Windows) pointing at it. The chip tooltip and the approval card say that approved commands can write there too. Checked for real on Linux and macOS (2026-11-04, Ask first): a write in the private folder was allowed, and a folder beside it in the shared temp folder (/tmp on Linux, the per-user temp folder on macOS), /tmp and the home folder were refused; no folder was left behind. Not run on Windows; Copilot sessions of the same run share the folder.
  • Fewer layout reads when an agent's status changes (spec 097 FR-184, T156, T159): fitting headers and usage lines and placing the rename pencils is done only for what a redraw changed, and in full whenever something that decides a fit changes (a status filter, Simple or Expert, the doors' signs). Layout reads for one status change in the simulator: 650 to 98 at 15 agents, 998 to 86 at 40; a redraw asks for the floor's style once. The compact desks' tapping hands now move (they showed one position because of a steps(1) that holds the first keyframe). New simulator check rows count renders, layout searches and reads, animations, stale fits and misplaced marks. The start (T157) and the halo's cost (T158) are measured and not changed.
  • Internal (Autopilot is not reachable in the app yet): a goal with a line starting with / is no longer refused on Claude Code, and the goal limits (4000 UTF-16 units, six reserved words) were checked against the real CLI (spec 099).
  • Autopilot core, types and event guard (spec 099 phase 1): the shared run counter (stop at limit + 1, end once, backstop, fail-closed guard), the goal text and end-reason helpers, the autopilot event and autoAnswered checks. Not reachable yet: no adapter reports it and there is no setting or button.
  • Previous sessions and chat history, Claude Code (spec 099, user-visible, live): a Stop hook feedback: message that Claude Code writes itself (a synthetic message, for example from your own Stop hook) is no longer shown as a prompt you typed and no longer becomes the session title. A message you typed with the same words still shows. Checked with npm run test:autopilot and mutation checks.
  • The next built-in world, the owner's third and fourth looks (spec 102, still behind its closed world flag): the floor is used at 3, 8 and 15 agents, in staggered rows that never mix two folders (a zigzag was tried and removed, see the spec).
    A folder of 4 or fewer stays one group under one sign; with 12 to 15 agents in 4 folders some still continue on the next row.
    A small room stands close and centred, drawn up to five times bigger with 4 agents or fewer, over a lower lobby when that lets it grow, its text growing with it. Cabinets come in four bodies (upright, cocktail table, racer, pinball) and four colours. Everything is dark at night, and nobody is cut by a cabinet.
    An asleep agent sits on a stool, while an idle agent stands awake. Waiting and limit plates and away notes are dark. The "needs you" pills keep to a lane under the lobby.
    Checked by test-realoffice-aisles, test-realoffice-assets and the shots-102-arcade driver at 1440x900 and 1280x800.
  • No "undefined" status (spec 102, every world, not only behind the flag): an agent at its limit or asleep with no label of its own showed "undefined"; it now says Limit reached or Asleep, the words spec 052 FR-076 and 097 FR-120 already use. A folder sign's path now gives way before the folder's name when the font is wider than expected (seen on Linux).
  • Asleep people in the next built-in world are drawn, not built (spec 102 FR-259, still behind its closed world flag): the seated sit frame of each of the 15 pixel people is now a Codex-drawn picture shrunk to the pack grid (scripts/build-arcade-sleep.py), not a standing frame with a stool drawn under it. Faces were kept readable (the standing frame's skin tone, closed-eye lines kept); looked at on contact sheets and in the simulator at both sizes, not yet by the owner.
  • World packs, phase 1 (spec 102): a strict zip reader and a world-pack validator for packs from outside the app (zip layout, names, sizes, picture headers, world.json and its colours), with a hostile fixture set in npm test. Nothing in the app calls them yet, so no user sees a change.
  • Website guards before the redesign (spec 104 phase 1, for contributors): the site build now fails if a world or feature that is not in a release is named on a page, in a caption or transcript, or in the README (docs/site/unreleased.json), leaves out the notes of draft specs from its specs index, takes its download only from a stable app release (never an asset-only release such as media-library, a pre-release or a test build), allows same-origin video (media-src 'self') and checks a video manifest (docs/site/media.json: budgets, hashes, captions, poster, licence and who made and approved each file). npm run blog:check (in npm test) guards the coming agent blog: tokens, private paths, addresses, chat markers, unknown hosts, owners' first names, evidence on main, a security hold, and a human read and an agent review where the spec asks. Nothing is published to a blog yet; no page shows a video yet. Checked by unit tests with stand-ins and mutants; the capture measurement ran on Linux under Xvfb only.
  • Compaction: the events of Claude Code, Codex and Copilot were checked on real sessions (manual compaction) and the shapes are recorded. After a Codex compaction the context ring waits for the next message instead of showing the old size, and a Copilot manual compaction ends its "Compacting" mark from the answer when no end event comes.