Repository navigation
Agent Durbar v0.1.18
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.
v0.1.18 (release, 2026-11-13)
- A faster, stricter pass in the long-chat check (spec 005 FR-210, issue 1610):
check-chat-visibilityreached the 600 s ui-suite limit on a loaded Linux run. Its one pass over the whole chat now steps 1.5 screens and asserts that every one of the 4,000 messages was laid out in some frame (a step of 6 screens fails withCVPASS=6); the two size entries also get a 900 s limit as a stated margin (it hides slowness, it does not remove it). - Fixed: a DPR 2 UI check on Linux no longer runs on a screen too small for its window (spec 008 FR-213, issue #1591):
scripts/ui-suite.jsstarted each driver underxvfb-runwith Xvfb's default 1280x1024 screen, so atUI_SUITE_DPR=2most of a 1440x900 window was off screen and painted-pixel checks could not see the lower rooms. The screen is now 1920x1100 times the density at colour depth 24 (UI_SUITE_XVFB_SCREEN=WxHxDoverrides it). macOS and Windows are unchanged. Checked bytest:uisuite; the Xvfb run itself was not done (no Linux here). - Fixed:
check-ui-textno longer reads a dialog's buttons in their hover colours (spec 008 FR-214): the real pointer rests where the window sits on the screen, so the larger Xvfb screen left it over End-all (4.14:1); the check now moves the pointer off the page first and fails withis hovered at the readif a control still is. - Fixed: two more colour-reading UI checks no longer depend on where the pointer rests (spec 008 FR-214):
check-097-chat-contrastandshots-102-arcademove the pointer off the page and assert no control is hovered before they read colours. - Fixed: the UI checks' pointer reset no longer reports a slow first run as a hovered control (spec 008 FR-215, issue #1648):
pointerOffre-sends the move to 0,0 every 250 ms for up to 3 s and, if a control is still hovered then, fails withthe pointer did not leaveand the list. - The UI shards are balanced again (spec 008 FR-189):
scripts/ui-suite-timings.json, the seconds per UI check that only decide which shard runs which check, is refreshed from the median of five green main runs (mean of Linux and Windows). 23 checks that were missing (the slowest arecheck-needs-panel,check-ui-textandcheck-chat-visibility) now have a time. With the new times the predicted slowest and fastest of 4 shards are 1.01 apart (they were 1.65 with the old file), and 1.01 apart of 3 shards (was 1.44). Checked bytest:uisuite; no app code changes. - Stacked PRs no longer clash in
npm run specs:numbers(spec 021 FR-094): a PR stacked on another inherits its spec numbers (only what it adds beyond the shared history counts; a fork PR never joins a stack); a separate PR taking the same number is still a clash. Tests:npm run test:specnumbers; to check afterwards, re-run theSpec numbers and picture budgetjob on #1232 once this is on main (#1240 still needs its own fixes). - The primary button reads at 4.5:1 in a user theme with the retro look, at rest, hovered and pressed (spec 052, issue #1614): a user theme that picked the retro look and left the button colours alone got Clean Office's pale pair, 3.8:1 at rest and 2.8:1 hovered (the hover mixed in black, which only lowered it). The hover now mixes toward the button's own text where that keeps 4.5:1, and away from it (black under light text, white under dark text) where moving toward it would lower the contrast, as on a blue, brown or grey fill; the theme repair holds the text to 4.5:1 on the rest and the hovered fill, for every fill and text colour tried. The built-in themes and the Clean look are unchanged in contrast terms (Retro Pixel 15.3 rest, 11.4 hovered).
- The keyboard focus ring on the Tools, Settings, Office and List buttons and on the ⋯ menu items reads on the retro look (spec 052, issue #1615): the ring was yellow on the cream strip and the cream menu (1.0 to 1.2:1; the pressed Office button was yellow on yellow). The menu items now use the page's ring colour and the four buttons an inset ring in the text colour of their own fill (15.3:1 in Retro Pixel; 3.8 to 12.1:1 in a user theme with the retro look). The buttons on the dark bar keep the yellow ring. The arcade bar is unchanged and stays a known gap.
- The focus-ring check no longer fails on a busy runner (spec 052, issue #1651):
check-052-focus-ringnow waits until its first Tab press has moved focus before it focuses the first control, instead of racing it. Test driver only, nothing changes in the app. - A failed server's row in the MCP list keeps a readable text column: the name and the help text no longer get squeezed by the buttons, which wrap below the text when they do not fit beside it.
- Settings keeps its place when you switch an MCP server off or on (spec 069 FR-069, issue 1624): the switches of the Copilot CLI list and the Claude Code list rebuilt the whole Settings page, which jumped back to the top and dropped the focus. A repaint of Settings now keeps the scroll position and the focused control, also while agent app accounts, quota and policy rows arrive after it; only your own scrolling, a key or a click moves the page. New check
check-069-toggle-scrollreads twelve switches at both window sizes. Not verified on macOS or Windows. - Rename pencil no longer drawn over a plate's name or status while the office rearranges (spec 097 FR-365, issue #1513): in the Real office, when desks ease to a new layout (an agent moving to the lounge), each moving plate's rename pencil now hides for the ease and fades back in at its spot, instead of being drawn displaced over the name and the status icon for a few hundred milliseconds (seen in the release clip and wherever every pencil shows: touch screens, Linux without a pointer). A pencil with the keyboard focus stays shown, and one that gets the focus while hidden shows at once and does not fade in a second time when the ease ends. Reduced motion is unchanged. Checked in the simulator from real screen captures, Photo office and arcade, 1440x900 and 1280x800, on macOS.
- Retro arcade: a plate's status word and helper dots (spec 102, issue #1558): at 1280x800 a plate with 3 helpers read "+3 dots Compacti..."; now the +N and the status word stay whole and the helpers' dots drop out whole, a waiting or failed helper's dot last (the plate's dots follow the order of the heads). On the narrowest plate the longest word can still be shortened.
- The arcade's "Showing only … · Show all" notice can be read (spec 102, issue #1559): in the Retro arcade the line that says a status pill is filtering the office was dark purple on a dark page (1.09:1), and the same line in the List and the List's "No agents match these filters." were dark on dark (1.00:1); they now read in the arcade's light inks (9:1 and better) and Show all stays underlined. Clean Office, Retro Pixel and the other worlds are unchanged.
- No office scroll under a strip in the Retro arcade (spec 102, issue #1566): at 1280x800 with 15 agents and the MCP problems strip shown, the arcade scrolled and its bottom row was cut off; the hall now takes the floor's own padding while it would otherwise scroll, and gives it back when the strip goes. Checked in the simulator by
shots-058-mcp-launch(Photo and arcade, both sizes); macOS only. - Arcade keeps small folders whole at 1280x800 (spec 102, issue #1590): 10 agents in 3 or 4 folders no longer have a folder cut across two rows at 1280x800; the hall takes a lower lobby (106 px) to keep them together, and does not scroll. The arcade layout test now runs at the real room heights. With the MCP strip shown the split remains.
- Letter skeleton for team-template names (spec 111 T032, FR-158):
src/shared/skeleton.tsmaps letters that look Latin (spec 003's closed Cyrillic, Greek and Latin lookalike lists) to their Latin letter (NFKD, table, lower case, table again, lower case, NFC), with spec 109's digit folds on request, andnameKeybuilds on it (compatibility forms fold through the NFKD) and is idempotent; best effort, not a full confusables set. Pure code, nothing visible yet. Checked bytest:skeleton(178 checks: table equals the lists, a golden check of every row, an idempotence grid of 36,000 names, a sweep of every code point up to U+2FFFF, thirteen mutants seen to fail). - Team templates have their store (spec 111 T006, issue #1306): main keeps the templates in one file (
team-templates.json, formatagent-durbar-team-templates, at most 50), written by temp file, flush and rename (mode 0600 on POSIX; a folder the store creates gets 0700, an existing one keeps its mode), and a broken, newer, foreign-format or not-UTF-8 file is moved aside once asteam-templates.bad-<time>.jsonwith one notice (the newest 5 are kept). A read that fails once (at the load and before a save) is tried again before anything is moved or refused, and a file that still cannot be read at a save makes that save fail without touching it; a Windows rename held for a moment is retried. If another running instance saved meanwhile (the app takes no single-instance lock), a save (it re-reads the file and compares a SHA-256 of its bytes) is refused withchangedand the store reloads, so a template saved by the other instance is not lost (except two saves in the very same instant: there is no lock).isBlocked()is true while a file it could not read at the load, or a bad file it could not move, is still in the way, and stays so until a save finds the file readable (nothing re-checks between saves). Such a save reads the file again first: a file that is readable and valid by then is reloaded and the save answerschanged, never moved aside. A dangling symlink as the store file is moved aside like any bad file on Linux and macOS; on Windows (noO_NOFOLLOW) it reads as no file and the next save replaces it. Known gap: the folder is not flushed after the rename (same as schedules; see #1620). Leftover temp files of dead processes are removed. Nothing shows in the app yet. Tested with an in-memory file system (42 mutants killed) and with the real one in a temporary folder; not run on Windows or macOS (the 0600/0700 modes are POSIX only). - Team template validator and frame guard (spec 111 T004, FR-002..FR-005, FR-010..FR-012):
src/shared/teamtemplate.tsholdscheckTeamTemplate(value, level),checkMemberandframeCheck: limits (own brief 4,000 characters and 80 lines, imported 2,000 and 40), unknown member keys refused, own-key lookups only, hidden characters refused in role and brief and in every imported string (ZWJ and ZWNJ allowed between letters or marks of Arabic and Indic scripts), briefs masked withmaskOutgoing, names compared onnameKey, and a best-effort guard against text that forges the brief's frame (also on the joinedRole: <role>. <brief>, marks and look-alike brackets folded). Pure code, nothing visible yet. Checked bytest:teamtemplatewith 63 mutants.