Repository navigation
Glimmervoid v0.47.0
Fixed
- Telegram waits for input left sitting: a session alert no longer reaches Telegram the moment it happens when no dashboard is open. Telegram now hears only about alerts you leave unanswered for the phone escalation delay (
phoneEscalationMs, 5 minutes by default), whether or not a dashboard is open. Escalations that land within the grouping window arrive as one message. - Spawned sessions stop chiming: the dashboard chime now plays only for alerts the server decides to send, so sessions a lead spawns and the machine shell stay silent, as their notifications already were. No sound plays while the dashboard window is focused, and a burst of alerts from several sessions shows one desktop notification that collects them, closing once you look at any of those sessions.
- Alerts that land together sound once: alerts from any sessions within
notifyDebounceMs(3 seconds by default) of the first one now fold into one alert instead of one per session.notifyDebounceMsused to drop a repeat of the same alert from the same session; it is now this grouping window, and existing values keep working. - Codex commands show in the activity stream again: current Codex versions run shell commands through an
exectool call, which the activity stream skipped, so a Codex session's commands never appeared there. Each command that call runs now shows as its ownexec_commandline, and one the call builds at run time shows asexec_commandwithout its text. Codexexec_commandandshellfunction calls now show the command they ran, too. Your prompts to an agent are still never published to the stream. - Debug state and Trace moved into the session menu: with debug mode on, a session's Debug state and Trace now sit in its
...Session actions menu instead of as two extra icons in the card header, on desktop and on your phone. The phone Terminal top bar gets the room back for the title, session name and safety badges. - Safety badges show on your phone: the phone Terminal screen now shows the session's YOLO or Sane YOLO badge and its Guardrails badge next to its status. Tap a badge to read what it means; tap Guardrails to release or restore them, as on desktop.
- The review desk fills a wide screen and shows its keys: on a wide Reviews pane the pull request column now stretches to the edge instead of stopping partway and leaving an empty band, the right-hand panel grows up to 420px, and the Your note box starts about four lines tall. Approve, Request changes and Comment now show their keys (A, R, C), Open in session shows O, and the Findings heading shows N and P to move between findings and X to post or drop the current one.
- Answering an agent's question clears Needs input right away: after you answer a Claude Code question (AskUserQuestion) or a Codex question (request_user_input, Codex 0.163.0 or newer), the card now leaves Needs input the moment the answer is submitted. Before, it waited for the terminal title to start spinning again, which could take half a minute or never happen while the agent ran a command, leaving the card on
?until you dismissed it. Sessions pick this up on their next start. - Card safety icons now agree on color: the Sane YOLO badge is now plain like the Guardrails on badge, since both mean a safety net is active. Orange now marks only a session without one: plain YOLO, or guardrails released.
- A restored dashboard tab no longer shows an empty board: when your browser reopens a tab from its cache (say after a restart) holding a dashboard older than the running server, the dashboard now notices on connect and reloads itself onto the current version, instead of showing no sessions until you refresh. A tab already open on an older dashboard needs one manual refresh to pick this up.
- Adding a session no longer fails with "already exists": when the name you add a session under is already taken by a running session, including one the dashboard is not showing, Glimmervoid now adds it as
name (2),name (3)and so on. The same goes for sessions opened from a GitHub issue, a pull request review or a Radar issue. Renaming a session to a taken name is still refused. A refused add now leaves a line in the server log with the name, folder and reason. - Dashboard errors are no longer lost: every error the dashboard shows in its red header line is now also written to the server log, as
[dashboard] header error: <source>: <message>. An error raised while the dashboard is disconnected is held in the page and written once it reconnects, so connection trouble no longer leaves gaps. Log lines the dashboard sends the server for other reasons are held the same way. Closing the tab before it reconnects still drops what it held. A page sends at most 32 distinct lines, then one line saying the limit was reached. With telemetry on, each distinct header error is also sent to PostHog once per page load asheader_error_shown, carrying the part of the dashboard that raised it and the message text, cut to 512 characters, with your home folder shown as~. This changes the telemetry promise: header error text can name a session, folder or branch, and Settings and the configuration reference now say so. Turn telemetry off to keep it local.
Added
-
A Community tab drafts answers to community questions: turn on Settings > Community (experimental) > Show the Community tab and Glimmervoid watches new questions on posthog.com/questions, plus the Discord channels you list under Discord channels, using your Radar PostHog key. Discord messages are first sorted by a quick check with a small model, so only messages that ask a question reach the tab. Each new question runs your Research skill in a locked-down lane that can read public sources and run read-only
ghsearches, then shows its confidence, what it checked, its gaps and a draft reply. Glimmervoid never posts anything: copy the draft, open the thread, start a session on the question, re-run the research or set the question aside. Questions and drafts survive a restart, and only questions asked after you turn the tab on are researched. A finished High confidence draft sends a browser notification, and a Telegram one when Telegram notifications are on and no dashboard is open; every other finish only marks the tab. -
A plain terminal on the machine Glimmervoid runs on: Add session now offers Shell next to your agents. It opens your login shell (
$SHELL, else/bin/sh;COMSPECon Windows) on the Glimmervoid machine, so when Glimmervoid runs on a remote host such as a Coder devbox you get a terminal there from the dashboard or a paired phone, with no SSH. A Shell session runs right in the folder you give it, or your home folder when Path is blank, and never makes a worktree. It always runs without guardrails or Sane YOLO, since you type every command yourself. Its card never shows Working or Needs Input and never notifies you, including when the shell exits. After a Glimmervoid restart the card stays in the list but does not start on its own; starting it again opens a fresh shell. Shell does not count as an agent in the first-run checklist: with no agent installed, Agent CLI stays missing and offers Open shell, so you can install one from a terminal on that machine. AcustomAgentsentry with the idshellnow collides with the built-in Shell, so rename it before upgrading or Glimmervoid refuses to loadconfig.json. -
Team reviews run on your own setup: Settings > Team review now lets you pick the Review skill from the user and plugin skills your review agent already loads, or type one by name; turning team review on asks you to pick one, and No skill is a valid choice. Review agent picks Claude Code, Codex CLI or Kimi Code, and Review model and Review effort suggest what that agent's own config offers (Kimi Code has no effort setting). Small PR agent, model and effort set the same for small pull requests, or follow the review agent. Replies to the review's threads are judged by the same agent. Reviews no longer pin Opus for full reviews and Sonnet for small ones: with the model and effort left empty, the agent uses its own defaults, so set Review model to
opusand Small PR model tosonnetto keep the old choice. An existing config with no skill set reviews with no skill. -
See and loosen the review sandbox: a Review sandbox readout in Settings shows what the review agent can reach: its network allowlist, whether MCP servers load, which home and checkout folders it cannot read, that the GitHub token is removed, and which tools it may use. The defaults are unchanged. Extra allowed domains adds domains to the allowlist, Open network egress drops the allowlist entirely, and Load my MCP servers gives the review agent your own MCP servers; each warns that the review runs code from the pull request under review. Codex CLI and Kimi Code always load the MCP servers their own config names, so they review, and judge replies, only with Load my MCP servers on. They also have no tool allow or deny list, so the sandbox is the only limit on what they run, including when judging replies, where Claude Code gets only Read and Write.
-
A tip when a tab could do the work for you: when you or your agents keep doing work a dashboard tab already handles, an animal flies in at the bottom right with a one-time tip. Reviewing three or more pull requests in a week points at Reviews; the same goes for GitHub issues (Issues), token and rate limit checks (Usage), production error hunts (Radar), pointing agents at editor files (Visions) and comparing agents by hand (Bench). Signals come from the commands and prompts in every Claude Code, Codex and Kimi Code session on this machine, including ones you started outside Glimmervoid, and from your own shell history. Only new lines are read, even with the ingest lane off, and sessions Glimmervoid's own lanes start in the background never count. Commands a Claude Code subagent runs count toward its session too, and they show in the activity stream under that session. Show me opens the tab, or its switch in Settings when the feature is off; Nah silences that tip for good. A tab you opened in the last 14 days never gets one. Glimmervoid keeps the counts in
feature-hints.jsonbesideconfig.json, and with telemetry on it sends PostHog when a tip is shown, opened or dismissed, keyed by feature. The animal shows even with flying animals off, and fades in without flying when your system asks for reduced motion. -
Kimi Code sessions: Glimmervoid can now run Kimi Code (the
kimiCLI) as a session agent, alongside Claude Code and Codex. There is nothing to set up, and Glimmervoid never edits your~/.kimi-code/config.toml(or$KIMI_CODE_HOME/config.toml): each kimi session starts on a home Glimmervoid builds in~/.glimmervoid/agent-homes/kimi, whose config is yours plus Glimmervoid's hooks and whose other files and folders point back at your kimi home, so your login, trusted folders and sessions are shared with plain kimi. Model changes you make inside a session are kept until the last kimi session exits, and that config copy, which can hold your API keys, is deleted then.glimmervoid doctorshows the home underkimi home. A kimi card takes kimi's own session title (your first prompt, then the title kimi generates after the first turn, or one you set with/title) at full length, and shows working, needs input (permission prompts and AskUserQuestion) and complete, and a subagent's own finish no longer marks the card complete early. Kimi asks whether to trust a folder the first time you open it, and Glimmervoid never answers that for you. In a folder kimi already trusts, Glimmervoid types the session's first prompt in for you; a new session worktree is not trusted yet, so there you answer the trust question in the terminal and type the first prompt yourself. Sane YOLO works on kimi through CC Safety Net. Kimi sessions run inside guardrails by default, like Claude Code sessions, with the same opt-outs (mark the session or its repository unguarded): a guarded kimi session can use its own kimi home, login and the Glimmervoid-built home, while guarded Claude Code sessions still cannot read kimi's home folder or login files. Kimi token use shows on the Usage page (at $0, since Glimmervoid has no price for kimi's models yet). When you are signed in to kimi with/login, the Usage page's Plan limits also show your kimi 5 hour and 7 day windows: Glimmervoid reads your kimi login without ever refreshing or changing it and asks kimi for them every 15 minutes while a kimi session runs, and again when you open the Usage page; if kimi refuses (an expired login, or an API key instead of/login), the last windows stay up with their age. Kimi replies and tool calls also feed the activity stream. If you ran the oldglimmervoid agent setup kimiduring development, the block it left in your config.toml is ignored and you can delete it. -
Shown HTML pages play right in the viewer: when a session runs
glimmervoid showon an HTML page, the viewer now shows the page itself, live and clickable, filling the box on desktop and on a phone, instead of a card with an Open page button. Open full size still opens it in its own tab. The page stays walled off from the dashboard: it can run its own scripts and forms, but cannot read the dashboard, pop up alerts, or use the camera, microphone or location. -
The website's pixel animals take off: on glimmervoid.dev, tapping one of the pixel animal tiles sends that animal flying across the screen, as often as you like. Nothing happens when your system asks for reduced motion.
Removed
- rtk output compression is gone: Glimmervoid no longer installs rtk, no longer rewrites agent Bash commands through it, and drops the rtk setting, the rtk compression tile in the Usage view, and the rtk line in
glimmervoid doctor. Testing showed the compressed output cost agents more than it saved: Claude Code reran most commands raw, and Codex gave wrong answers from cut-downheadandtscoutput. An oldrtkorusage.rtkSavingskey inconfig.jsonis ignored, so there is nothing to change. Glimmervoid no longer uses~/.glimmervoid/bin/rtkor the rtk data folder (~/Library/Application Support/rtkon macOS,rtkunder$XDG_DATA_HOMEor~/.local/shareon Linux), and you can delete both by hand. Sane YOLO still looks through anrtkprefix, so a leftover rtk binary cannot slip a blocked command such asrtk terraform destroypast it. Sessions running when you upgrade keep rtk until they restart. - Worth a look no longer flags untested files: the review panel stops listing "No test reaches it", and the change map no longer works that fact out. "Usually changes with", other sessions changing a file, and package links into another repository still show.
- The Conversation view is gone: the experimental Conversation face, its "Conversation view" card menu item and its Settings toggle no longer exist, so a session card shows its terminal or its plan. Claude Code sessions no longer get the hooks that held every permission prompt for the dashboard, and a local Codex session no longer starts its own app-server. A
conversationViewkey left inconfig.jsonis ignored. A child session's questions still reach its parent as before, and Codex in a Coder workspace runs as before.
Changed
-
Quieter empty Changes panel: a session with no changes yet shows just "No changes yet" instead of a sentence about what will appear there, and the panel header no longer repeats the session name next to Review.
-
One type scale across the dashboard: dashboard text now uses four sizes (12, 14, 16 and 20px) instead of six, each tied to a role: labels and metadata at 12px, body text at 14px, titles and phone form fields at 16px, and the header title and big stat numbers at 20px. The smallest 10px and 11px text grows to 12px, so dense rows read more easily but truncate a little sooner. Lowercase text drops its letter spacing so CommitMono's smart kerning evens out the gaps; only uppercase labels and headings stay tracked.
-
Visions is now experimental: Visions and its machine context ingest only appear on the Experimental update channel, as Issues, Community and Coherence do. On a release build the Visions settings are hidden and Visions stays off, even if you turned it on before.
-
Feature tips never point to experimental features: the owl tip card no longer suggests Issues or Visions, or any other feature that is still experimental, on any build.
-
Reviews sorts by what to do next, without tabs: the All, Reviewing and Mine tabs are gone. Your pull requests and your teammates' now share one queue ordered by priority, and each row in Your turn leads with what to do: Fix, Review, Reply, Update or Merge. Fix comes first and covers a keep-mergeable repair that failed and an agent review that errored. A keep-mergeable repair still running sits under Agent working. Failing checks and merge conflicts on your own pull requests now sit under Update, below reviews and replies. Within each group, older pull requests come first. The attention banner and the next-attention key follow the same order.
-
Stacked pull requests show where they sit: in Reviews, a pull request of yours that builds on another now shows a small stack badge such as
2/3(its level in the stack, then the stack's height) instead of the longer "Stacked on" label. Pull requests built on the same base share a level. Hover the badge to see which pull request it builds on. Opening one also shows a Stack section under its heading that lists every pull request in its stack, indented by level, with each one's state; click another to open it. -
Worth a look sits under the changed files: the review panel now lists the changed files first and the Worth a look list after them, so the files are the first thing you see.
-
Reviews session pills rest while hidden: the session pills on your own pull requests no longer repaint while the Reviews view is in the background, and catch up the moment you open it. While it is open, a status change rewrites only the pill text that actually changed.
-
Status lights cost far less to update: when a session's status changes, the dashboard now repaints only that session's row and card, once per frame, instead of every row in the rail, and the server sends one status message per session per moment instead of up to four. With 30 sessions, one status change now touches the page about 27 times instead of about 494, and takes about 3.2 ms of script instead of 4.2. The tab title, the favicon and the alert chime still update immediately, even while the tab is in the background; the rail catches up when you come back to it. The elapsed-time clocks now tick only for sessions that are running, waiting or starting.
-
Desktop no longer downloads the phone layout: the phone screens now download only when the dashboard is in the phone layout, at load or when you switch to it, so a desktop load moves about 4 KB less before the board appears. On a phone the screens appear once that download finishes.
-
Rarely used extras download on first use: the merge celebration, the first-run checklist, the Restart dialog, alert sounds and the flying animals now download the first time they appear instead of before the board, so a cold load moves about 10 KB less before the board appears. Each one appears a moment later the first time on a slow link.
-
The doc editor and shown-file viewer download on first open: the doc and shown-file buttons still appear as soon as an agent surfaces a file, and the editor or viewer downloads the first time you open it, so a cold load moves about 2 KB less before the board appears.
-
The open session comes back sooner after a reload: the dashboard now connects to the terminal of the session you had open as soon as the page starts, instead of waiting for the list of sessions first. On a simulated 3G link a reload shows that session after about 2.1 seconds instead of 3.1. If that session is gone, the dashboard connects as before.
-
Reloading the dashboard no longer rechecks its fonts: the CommitMono font files now carry a content hash in their name, so the browser keeps them cached until an update changes them instead of asking the server about each one on every reload.
-
Each secondary view brings its own styles: the styles for Settings, Usage, Reviews, My PRs, Radar, Visions, Hooks, Trace, Benchmarks and the plan view now download with that view instead of with the board, so the board's stylesheet shrinks from 39 KB to 30 KB compressed. On a simulated 3G link the board still appears after about 8.3 seconds, because the fonts take the freed bandwidth.