Skip to content

Glimmervoid v0.62.11

Latest

Choose a tag to compare

@johncwaters johncwaters released this 21 Nov 00:57

Added

  • Chat shows each subagent as a row under the turn that started it. Where a subagent call used to be one more tool line, Chat now shows a row naming the subagent's type and task, with a pulsing dot and Running while it works and its run time once it finishes; a background subagent stays Running until it really stops. Click a row to unfold it: for Claude Code, the subagent's own steps and its final report once it finishes; for Codex, OpenCode, Kimi Code and Antigravity CLI, which report only that a subagent started and stopped, the row still shows when it ran. The subagent count under the message box now counts background subagents until they stop. The Trace tab lists the start and stop of every subagent.
  • HTML pages on a board scroll in place. Scrolling with the wheel or trackpad over a page box now scrolls the page, the way a scrollable area on any web page does, where it used to pan the board and leave a tall page showing only its top. A scroll that starts over a page with nowhere left to go pans the board, and a scroll that started moving the page stays with it until you pause, so reaching the end of a page mid-swipe does not jerk the board. Ctrl or Cmd with the wheel still zooms, a click still opens the page full size, and a drag still moves it.
  • A Claude Code session stopped by a rate or usage limit now waits and says so. It used to go Idle with only a plain line in Chat. Now the session waits on you, the session row reads Rate limited in red, or Limit until 4am when a plan limit hit and Claude Code reported its reset time, and the waiting notification goes out once. Other agents report no limit stop Glimmervoid can read, so their sessions are unchanged.
  • Chat shows a limit card when a plan limit stops a turn. In place of the plain limit line, a red card names the limit, the reset time and the time left, counting down. Switch account (1) opens the account switch, and Continue at reset (2) sets one wake that types continue into the session when the limit resets; the agent cannot cancel it, and it does not count toward the agent's own wake limit.
  • Packs can carry hooks that gate an agent's tool calls. A pack's pack.json can declare hooks: each names the tool kinds it watches (edit, shell) and a command, which may run a script from the pack through its interpreter, such as ["node", "{packDir}/hooks/check.mjs"]. You see each hook's description, tools and command when you approve the pack, and glimmervoid packs check reports a hook whose script the pack lacks. Once approved, the hook runs before every matching tool call in Claude Code, Codex, Kimi Code, OpenCode, pi and Antigravity CLI sessions: exit 2 stops the call and the agent reads the hook's stderr as the reason, while a crash, a timeout or any other exit lets the call run. Sessions started without an approved pack hook keep the hook setup they had. Glimmervoid releases up to 0.62.10 refuse a pack whose pack.json sets hooks.
  • A ship pack in Glimmervoid examples: spec, review and land in every session. Enable it from the Packs tab to give every agent the specifying, challenging-plans, gating-reviews, reviewing-code and landing skills, adapted from the dotfiles workflows under their MIT license. Its spec gate refuses the first edit in a checkout that no valid design spec under ~/.local/state/ship/specs names (a retry goes through), and its commit gate refuses a raw git commit or git push in favor of the landing script, which commits exactly the tree a review stamped; both gates catch accidents, not an agent set on getting around them, and keep their state in the repo's .git folder so they work with sessionSandbox on. Only Claude Code runs the review lanes in parallel; other agents run them one after another.

Changed

  • Changing the model from Chat opens a picker above the message box instead of a dialog. Clicking the model name under the message box opens the + session model picker in place, reading "Continue on Opus 5.5 at high" over a find field and the model list, so the conversation stays in view. Picking a model moves to its effort step, and picking the effort switches the session; a model without efforts, or one you type, switches as soon as you pick it. Escape or a click outside closes the picker without switching. While a turn runs the model name is plain text; switch once the turn ends. The session card's Change model dialog is unchanged.

  • Boards with thousands of wires lay out much faster. Drawing each wire used to compare it against every wire already drawn, so a board with a few thousand wires, such as a large code map from an earlier release, could keep the layout busy for hours while every other board waited on its tiles. Each wire is now checked only against the wires it could cross; the layout comes out the same, and a 5,000-element board lays out in well under half the time. A large board still laying out no longer holds up the others either: every other board gets its tiles while it works.

  • A Chat message sent while a session starts just looks sent. Sending from Chat to a dormant or starting session used to park the message under a Starting line, marked "Sends when the agent is ready" with a Cancel button. The message now shows at once as your prompt in the conversation and the activity line shows the agent at work, while Glimmervoid holds the message until the agent can take it; the agent's own copy of the prompt replaces it in place. A message sent during a running turn still shows "Queued until the turn ends" with Cancel.

  • Short replies between tool calls ride the Chat activity line instead of stacking up. While a turn runs, a one-line reply of 60 characters or fewer, such as "Running.", no longer adds its own paragraph to the turn each time; it replaces the playful verb on the activity line under the turn, which already names the current tool call. Longer replies still show in the turn, the turn's final reply still shows when it ends, and the folded turn details keep every reply.

  • A board change in Chat is one quiet row with a shortcut. The card that read " changed X on " with a Focus this button is now a single row: the boards icon, "Changed X on Board" without repeating the agent's name, a hover tint, and a Show on board button that shows the board shortcut (Cmd+B on macOS, Alt+B elsewhere). Opening the board with that shortcut now picks the board the agent last changed and frames that change once.

  • The Chat status row under the message box shows the agent's logo and colors its meters. The model name now starts with the session agent's mark, in the agent's brand color where it has one. The plan usage and context bars fill green under 50%, amber from 50% and red from 80%, and at 80% or more the reading itself turns red, so a session close to a limit stands out at a glance.

  • Chat shows a terminal wait as a card like the others. When a session waits on something only its terminal shows, the card is now amber, names the agent ("Claude Code is waiting in its terminal"), says Glimmervoid cannot read what it shows and to answer it there, and holds the live terminal with one Open Terminal (1) button. The activity line no longer reads "Working on it" under the card while the agent waits on you.

  • Coherence 1.15.0 in development. The @posthog/coherence dev dependency moves from 1.14.0 to 1.15.0, which adds peer repositories, a testSkip setup key and a Components view that reads each component as its spec. No config or spec change was needed.

  • Focus this leaves a margin around the changed cards. The board camera stops 2% short of a tight fit, so the cards no longer sit against the edges of the board.

  • The Chat message input is focused when you enter a session. Picking a session shown in Chat from the rail, a shortcut or an attention tile, or switching a session to Chat, puts the caret in the message input, so you can type at once. On a phone nothing is focused, so the keyboard stays down. The turn keys [, ], B and Space still act only while the input is not focused.

  • Esc dismisses a question in Chat. With a question open, Esc or the card's new Dismiss control leaves every question of that ask unanswered and stops the agent, which then waits for your next message. The card keeps a Dismissed line. Esc no longer leaves Chat while a question is open.

  • Questions in Chat are easier to read. Each option is now one full-width row: its label and key on the first line, its explanation underneath at normal text size across the whole card, so every explanation starts at the same left edge. Clicking anywhere on a row answers with it. The question's topic sits beside its place, such as 1 of 4, above the question.

  • Agents can no longer write over board edits they have not read. When you or another session change a box, wire or note, an agent that tries to change or remove that element before reading the board again is refused, and told to read it first. An agent has to be connected to the board and have read it since connecting before it changes anything already there, and showing a file again or importing a board file over the board is checked the same way. Restacking several boxes counts as a change to each of them, and a large edit no longer locks agents out of their own boxes or hides earlier changes from the list an agent reads. Your own edits on the dashboard are never refused, and the code map and plan step updates still land. Undo and Redo on a board no longer revert an edit someone else made since: the step is dropped and the status line names the box that changed.

  • The code map board shows levels instead of a wall of wires. Each top folder is a collapsed frame you can open, stacked by level so a folder sits above what it depends on. Only imports that point up a level, the crossed wires, are drawn red with their file count, or the one file import holding them. A session overlay opens the folders it touches and draws its modules' imports. A folder matrix card beside it lists which folder imports which, with the upward ones marked. The map's own redraws no longer queue as board changes for a connected session.

  • Every repo is import gated by default. The first time Glimmervoid maps a repo's mainline it writes the code map setup with the gate on, allowing the module imports mainline already has, so no setup step stands between a repo and the gate; glimmervoid codemap setup <repo> --gate off turns it off for good. The gate is a built-in check: it runs with the other checks at every turn end, shows in the agent's session check, works with sessionSandbox on, and refuses a merge while it fails, with no verification entry in config.json. It judges the working tree, uncommitted files included, so an agent learns about a bad import in the turn it adds it, while a merge judges the commit that lands. The codemap check an earlier codemap setup --gate on wrote into config.json is removed on start, and codemap is no longer a check id config.json can use. A repo whose mainline cannot be mapped, such as a partial clone, shows the check as skipped and stays ungated.

  • An agent can turn a repo's import gate back on. Asked to, an agent calls codemap setup from its session: it turns the gate on for the session's repo, writing the setup from mainline when there is none and keeping an existing allow list. Turning a gate off and allowing an import stay yours, through glimmervoid codemap setup <repo> --gate off and glimmervoid codemap allow.

  • Crossed wires can only shrink. For two modules that import each other, the lighter direction is a crossed wire: codemap map lists each with the file imports holding it, cheapest first, and the setup records its count as a ceiling. A branch that adds file imports to a crossed wire is refused, naming the new import first, and when mainline drops one, the ceiling drops with it. Allowed imports mainline lost are dropped the same way as mainline moves, with no codemap shrink needed. A setup written by an earlier version keeps its gate as it was and gains its ceilings the next time mainline is mapped.

  • Settings shows agent accounts as rows. Under General, each agent gets its mark, name and a compact Add account button, and each account is one row with a check or warning mark, its email, a DEFAULT tag and its actions, in the same style as Ready to commit as you. Each lane is one row with its account picker on the right, and the follow-default choice now reads Default account.

  • A packs reference for pack authors. docs/packs-reference.md lists every pack.json field, panel bridge message, panel action, prFlows command output and glimmervoid packs command, generated from the same schemas Glimmervoid checks packs with, and docs/packs.md explains approval and what a pack can and cannot reach. Fields that had no description in glimmervoid packs guide --schema and the published pack.v1.json now have one.

  • OpenCode, pi and Antigravity CLI sessions get your approved packs' MCP servers. Executor and every other pack MCP server now reach these three agents too, vault bearer secrets included, the same as Claude Code, Codex, Kimi Code, Gemini CLI, Qwen Code and Copilot CLI. OpenCode reads them from the config Glimmervoid hands it at launch, laid over your own config, and pi from Glimmervoid's own extension. Antigravity CLI lists them in Glimmervoid's own plugin: a server with a vault secret goes through Glimmervoid's bridge, which reads the secret from the session and offers no tools in an agy you start yourself or in a project where the secret is denied, and a server with no secret is listed directly, so agy runs its own MCP sign-in for it, in every agy session, yours included. The secret value never lands on a command line or in a file. Nothing is written to your own OpenCode, pi or Antigravity config. A pack server url can no longer hold { or }, as it already could not hold $, so no agent fills another secret into it.

  • The board properties panel drops the divider lines around More. More and the Add note, Frame and Delete row now sit in the panel without a rule above or below them.

  • Change the model from the Chat status row. Click the model name under the Chat message box to open Change model for that session; a session that cannot switch models shows the name as plain text. The row also lines up: the model sits at the left, the other readings sit together at the right with even gaps, and the row is centered between its divider and the bottom edge of the message box.

  • The check result at turn end is one pill on the turn's summary row. The Checked at turn end card is gone. The last turn's row ends in a pill that reads the result, such as 4 checks passed or 2 of 4 checks failed: red when a check failed or timed out, grey when none failed but one was skipped, green only when every check succeeded. Click the pill to open the list below the turn: failed checks with their command and output, then skipped checks with the reason they did not run, then one line naming the rest. The list stays open while you read it and closes again at the next turn's checks.

Fixed

  • A failed save no longer stops the server. When Glimmervoid could not save its pack sources while following the built-in example packs, the server stopped. It now logs a warning and keeps running.

  • Chat no longer opens empty while subagents run. When the newest part of a session's history held only subagent work, such as review or workflow lanes, Chat showed just the activity line with nothing above it, and earlier history never loaded. Chat now loads earlier history until the view is filled.

  • A long or pasted message shows as you wrote it. Chat showed a long message, or text pasted into the agent's terminal, wrapped in raw <pasted_content id=...> tags with blank lines above it, and the session title and sidebar row read <pasted_content id=...> instead of the task. Chat now shows just the text, with anything typed around the paste kept, and the title comes from the pasted text.

  • A Claude Code session stuck on something only its terminal shows now waits on you as Check Terminal. When Claude Code shows a dialog in the middle of a session (such as its usage credits prompt), it fires no hook, so a session holding open for a background agent read Running for 30 minutes, then Complete, and nothing told you it was stuck. Now, when a turn has ended but a background agent sends nothing until the 30 minute background agent limit runs out, or a Chat message is not taken, the session waits, its row reads Check Terminal, the waiting notification goes out once, and Chat shows the live terminal so you can answer the dialog.

  • Antigravity CLI sessions open straight to the prompt again and take Chat messages. Antigravity CLI 1.5.13 trusts only the exact folder it starts in, not the folder above it, so every session opened on its "Do you trust the contents of this project?" dialog and the first Chat message was typed into that dialog and lost. Each session now trusts its own folder.

  • Cmd+B only shows the agent's latest board change. Cmd+B (Alt+B off macOS) used to open or close the board beside a session, so pressing it from Chat's Show on board row closed a board that was already open. It now does exactly what Show on board does: opens the changed board beside the session if needed, focuses it and frames the change. With no new change to show it does nothing; drag a board onto a session to open it, and close it from its pane header.

  • Cmd+0 on a focused board resets its zoom on macOS. It used to open the New session dialog as well, because the dashboard's Add a session shortcut also saw the key. The board now keeps the Cmd chords it handles.

  • A double tap on a zoomed board page works while the page is still settling. Right after a pinch, the page moved under your finger between the two taps, so they were taken as far apart and the page stayed zoomed. Taps are now matched by where your finger is on the screen.

  • A Chat message the agent never takes comes back instead of showing as sent. When Claude Code shows a dialog in the middle of a session (such as its usage credits prompt), it fires no hook, so Chat typed your message into the dialog and then showed it as sent while no reply ever came. Now a message whose turn has not started 15 seconds after it was sent is marked as not taken, with a pointer to the Terminal, and Cancel puts its text back in the message box. Later messages wait behind it instead of being typed into the same dialog.

  • A slow Telegram no longer sends the same ping to your phone every minute. When Telegram took longer than 10 seconds to answer, Glimmervoid counted the send as failed and sent it again each minute, even though Telegram had delivered it, so one notification could arrive a dozen times. A ping that reached Telegram now goes out once, answered or not. A ping that never left this machine, or that Telegram refused, is still retried every minute as before.

  • A Chat message sent to a session that is just starting now reaches the agent. Sending from Chat to a dormant session starts it, and the message used to be typed the moment the agent first drew its screen. Antigravity CLI ignores input that early, so the message showed as sent but never arrived and no answer came. A session now shows Starting until its agent can take input, and Glimmervoid types a message in at that moment with no fixed wait: Claude Code when its startup hook reports in, Antigravity CLI when it shows its cursor, Kimi Code when it draws its first full screen, and every other agent at its first output. If Antigravity or Kimi never shows that sign, the message goes in 30 seconds after the agent first draws. The same holds for the line that wakes a parent session when a child it spawned finishes.

  • Leftover folders Glimmervoid made are cleaned up at start. A removed worktree's folder could come back holding only an agent tool's bookkeeping files, written after the removal, and then sat in the worktree folder for good; a folder like that, at a path Glimmervoid removed before and with no checkout in it, is now deleted once nothing in it has been touched for a day. Folders Glimmervoid did not make are never touched.

  • Worktrees of sessions that are gone stop filling the disk. A worktree left by a removed session used to stay forever when it held uncommitted or unmerged work, with only a log line at each start. Once such a worktree has been untouched for 7 days, Glimmervoid now saves its tracked, untracked and .env files as a commit on refs/glimmervoid/parked/<session id>/<folder> in the repo, then deletes the folder. The server log prints the git worktree add command that brings it back, and a parked commit is dropped after 30 days. Other git-ignored files such as build output are not saved. A worktree whose git state cannot be read, or that holds another repository inside it, is still kept.

  • A worktree removal that gets interrupted no longer leaves a half-deleted checkout behind. Removing a clean worktree used to delete its files in place, so a server stop partway through left a folder with missing files that every later start read as uncommitted work and kept forever. A clean worktree is now moved aside in one step before it is deleted, and each start finishes deleting any moved-aside folder an earlier run left in a project's worktree folder.

  • A command that starts with node runs Node in the standalone install. A done line, repo check or pack command such as node scripts/check.js used to run Glimmervoid itself and print its help, because the standalone build sends any program named node back to its own binary. Glimmervoid now runs the node on your PATH, as the source install always did, so env node is no longer needed. With no node on PATH, the command fails with an error naming glimmervoid-node-not-on-PATH.

  • Chat shows a session's whole history again. Scrolling to the top of Chat while the agent was working could stop loading earlier turns for good, leaving only the latest one; the earlier page now loads as soon as the live update in flight finishes. A session whose first prompt carried a pasted image or file also lost that first prompt from its history, because recording started at the end of the conversation instead of the beginning; it now starts at the beginning. History already recorded without its first prompt stays as it is.