Skip to content

Glimmervoid v0.62.5

Choose a tag to compare

@johncwaters johncwaters released this 17 Nov 22:05
fbd9288

Added

  • Answer permission prompts from Chat. With the experimental Chat view switch on, a session launched without skip permissions holds each permission prompt from Claude Code, Codex, Kimi Code, OpenCode, pi and Antigravity for an Allow or Deny answer from the dashboard; read-only shell commands are never held. Unanswered, the agent shows its own prompt as before. On Kimi Code, Deny from Chat blocks the tool, while Allow still needs Kimi's own prompt answered, because Kimi takes only a deny from a hook. Turning the Chat switch off stops new holds at once, without restarting sessions. With Chat on, Claude Code questions from the main session are held for an answer too. Skip permissions plus Sane YOLO stays the default launch.

  • An experimental Chat view switch. Settings, Sessions, Chat view adds the switch for the Chat page being built as an alternative to a session's terminal. It is off by default and changes nothing yet.

  • Trace shows the conversation for every first-class agent. Codex, Kimi Code, OpenCode, pi and Antigravity sessions now fill the session's Trace with the prompt, thinking, replies and each tool call with its result, as Claude Code sessions already did. Glimmervoid reads each agent's own session log, or for OpenCode receives the turn from its plugin when the turn ends. A record that does not fit the shared trace format is refused and a notice names the agent and the field.

  • Walk a board's wires in the media viewer. With a picture, video or page open full size, the arrow keys show the box wired to it in that direction, in the same viewer and at the same zoom, so a flow of screens can be stepped through and compared without closing each one. Boxes without a picture, video or page are skipped, and inside a page an arrow first scrolls the page or moves through a field before it walks on.

  • Add session reads as one sentence that holds still: Start [agent] as [account] on [model] in [project]. The whole sentence shows from the first open, with the words not picked yet dimmed, so picking never moves the dialog; the list starts in the same place on every screen, with a title such as Agents or Opus 5.5 effort where a screen has no search, and All models or All projects as the first row of the list. Every word opens the same searchable list of rows. The model word shows the model and its effort, such as Opus 5.5 medium; its list holds the agent's own models with an OpenRouter group under them, and any typed model id is accepted. Picking a model that has efforts lists them as rows, Default first with the effort your agent settings name. The account word opens a list when the agent has more than one account. Several repos keeps a search box, lists picked repos first, and ticks rows instead of checkboxes. The GitHub issue field is gone: type an issue number such as #128 in the project search and pick Issue #128 in the project you want, the project already in the sentence listed first. The dialog has no dropdowns, tabs or chips left.

  • Every way of starting a session remembers your last picks. Glimmervoid now keeps, in start-picks.json in its home folder, which agent you last started in each project and the model and effort you last picked for each agent. The Add session dialog, the + quick add button and every browser or phone share it, so a quick add starts the same agent, model and effort you last chose there. A quick session copies the model and effort of the session it starts from. A remembered model that the agent no longer lists falls back to the agent's own settings. Glimmervoid never writes your agent's own config files. Picks remembered in the browser before this release are not carried over.

  • Edit any changed file, and preview the running app beside a session. The Changes panel has an Edit button on each changed text file that opens it in the session's doc editor. A session that starts a dev server can run glimmervoid preview <port> (or the preview tool) to show the app in a new Preview page on its card; Glimmervoid forwards plain HTTP to that local port under /app-preview/<session>/, in a sandboxed frame that never sees your dashboard sign-in.

  • Lockfile conflicts get their own merge steps, and a waiting merge shows its place in line. When a session's merge parks on a conflicting lockfile (npm, pnpm, yarn, bun, Cargo, Go, Poetry, uv and more), the merge prompt lists it apart from the other files with the command that regenerates it, instead of asking the agent to merge it by hand. A merge waiting behind other merges in the same repository shows how many merges are still ahead of it in the Changes panel, counting down as they finish.

  • Reactions to session events. A reactions block in the config file reacts to a session failing, ending, needing input or landing commits, by raising a note on its card or, for needing input and landed commits, typing a set text into it the next time it is idle, even when the event came while it was busy or waiting on you. Each reaction can be limited to one project folder and waits cooldownMinutes (default 10) before reacting to the same session again.

  • Scheduled pack launchers can catch up a missed run. A launcher with catchUp: true starts one run at once when a scheduled run fell while Glimmervoid was down or the host slept, instead of skipping it.

  • Sessions run your checks at turn end, and can refuse to merge until they pass. Set verification.repos["<repo path>"].checks in config.json to a list of { id, command } and each session in that repository runs them in its worktree whenever a turn ends and the tree changed. The Changes panel lists each check as passed, failed, timed out or skipped, with the output tail of any that did not pass. A command runs as a plain program with arguments, never through a shell. Turn on gate.enabled and glimmervoid push and a session finish refuse to merge until every required check passed on the current tree, naming the check that failed. A commit made after the checks passed, while the merge waits its turn, stops that merge instead of landing unchecked.

  • Start a session from a GitHub issue. Type an issue number in Add session and the agent starts in that project with the issue state and link, and the issue title and body fenced as untrusted text, typed into its prompt for you to read and send. The repository comes from the folder's GitHub origin, and the issue is read with your gh login.

  • Pause at budget. Turn on Settings > Usage > Pause at budget and, once the estimated daily or monthly spend reaches its budget, running sessions are snoozed (or stopped when they have no conversation to resume), and agent spawns, agent Arena races and scheduled wakes wait until the day or month rolls over. A paused session that an automation wakes is paused again; one you type into stays running, and starting a session yourself still works. The pause happens once per period, even across a restart.

  • Loop guard. Set Settings > Sessions > Turns without progress and a session whose worktree stays unchanged for that many turns in a row, with no input from you, another session or a wake, is flagged for your attention, or stopped when When a session is stuck is Stop.

  • Spawn a session on a chosen model. An agent can now pick the model and reasoning effort of a session it spawns: glimmervoid spawn --model <id> --effort <level> <prompt>, or the model and effort fields of the MCP session tool's spawn. Unset, the child runs on the agent's configured model as before. A value that is not safe on the agent command line is refused.

  • Slack alerts. Paste an incoming webhook URL in Settings > Slack and turn on Send session notifications, and session alerts post to that Slack channel when no dashboard is open or an alert sits unacknowledged, like Telegram. GLIMMERVOID_SLACK_WEBHOOK_URL can supply the URL instead, and is never written to config.json.

  • Answer sessions from Telegram. Turn on Settings > Telegram > Take replies and each alert carries a [gv:...] session tag. Reply to an alert to send that session text, or, when it holds a question, answer with an option's number or label. Only messages you send the bot in your own private chat with it are read, so the chat id must be your own Telegram user id; a group chat never answers or types into a session, whoever sends, messages sent while Glimmervoid was stopped are skipped, and tagged alerts arrive one per message. If Telegram refuses the bot's reply polling (another process polls the same bot, or the token is revoked), replies turn off and Settings > Telegram shows why.

  • Tour of a change. The Changes panel has a Tour section that walks a session's changed files in reading order: grouped by subsystem, the files more of the code depends on first, each with why it is in that place (how many files depend on it, which tests cover it, how often it is fixed, which partner file it usually changes with). It is built from the same change facts as Worth a look, with no model involved.

  • Approve plans on the dashboard. Turn on Settings > Sessions > Plan approval and a plan an agent proposes shows on its card with Approve, Reject and Answer in terminal, plus optional feedback for the agent. Claude Code waits on the plan until you decide, gets rejection feedback with the rejection, and gets approval feedback typed in once the plan is approved; Codex shows its proposed plan and your decision is typed into it. Other agents keep approving plans in the terminal.

  • Shared memory and rules for every session. Agents get a memory tool (and glimmervoid memory) to keep short notes under a key for later sessions in the same project, such as the test command or a decision, and to list, read or forget them. Notes live in memory.json beside your config, at most 200 per project, and agents are told to verify them rather than follow them as your instructions. Lane sessions cannot use memory. Set agentRules.rules in config.json to a list of rules every agent session receives with its Glimmervoid tools, whichever agent runs it.

  • Run a whiteboard plan. glimmervoid whiteboard run <board> (or the whiteboard tool's run action) spawns a session for every planned step whose earlier steps have landed and marks it building. When that session's work merges, the step is marked landed and the steps waiting on it start. A step added or edited after the run waits for the next run. A step whose session ended without merging, or that a restart forgot, goes back to planned and starts again on the next run. Spawn limits still apply, a step that could not start after a merge is raised to the operator, and a plan whose steps wait on each other in a loop is refused.

  • Sibling sessions message each other and claim files. A session another session spawned can message a sibling the same parent spawned with glimmervoid send --to <sibling> <text> (or the session tool's to), with the same mailbox caps and wake limits as parent mail. Any session can claim a file it is editing with glimmervoid claim add <path> or the claim tool; a claim never blocks, names who already holds the file, ends with the session, and shows in the other session's change map as a collision when it changes that file.

  • Second-opinion review of a session. The Changes panel has a Reviewer section with a Review button that starts a read-only reviewer on the session's diff and lists what it finds, each finding opening its file. Pick the reviewer's agent (Claude Code or Codex) and model in Settings > Sessions > Session review, so the review comes from a different model than the one that wrote the code. The reviewer cannot edit the session's files; a Claude Code reviewer also reads the checkout, while a Codex reviewer sees only the diff. A diff too large to send whole is marked as a partial review and never approved, and a result is reused until the worktree changes.

Changed

  • Every notice under the dashboard header is now recorded, not only errors and warnings. An available update, old reports left in a repository and the telemetry notice now each write a line to the server log when they appear, and send one notice_shown usage event with their text. That event replaces header_error_shown, which errors and warnings now send as notice_shown too.

Fixed

  • Session Trace shows what a tool printed when its output was too large. Claude, Kimi, pi and Antigravity save a large result to a file and hand back a note pointing at it; Trace now shows the kept preview, marked as cut, instead of that note.

  • A pack panel that types into a session now presses Enter. Text a panel sent with its sessions act was pasted into the agent's input box and left there unsent.

  • Wakes stay with the Glimmervoid that set them. Scheduled wakes were kept in one wakes.json in the Glimmervoid home folder, so two instances started with different --config files shared it, and one marked the other's wakes as missed or wrote over them. Wakes now live in wakes.json beside your config. An instance with its own config folder starts with no wakes, so set its wakes again.

  • Codex sessions on GPT-6 models are priced offline. The bundled price list now carries GPT-6 Sol, Luna, Astra, Astra Law, Cyber and Rosalind, GPT-6.1 Sol and Terra, and GPT-6.2 Sol, Astra and Astra Law at OpenAI list price, so the Usage tab no longer reports them as having no price, and counts their cost, when the live price list cannot be fetched.

  • A Claude Code session started right after Glimmervoid starts gets folder trust. The sign-in check Glimmervoid runs at start can leave Claude Code's config lock behind for up to 10 seconds, and a session started in that window was refused with "another process holds its lock". Glimmervoid now waits as Claude Code itself does: it takes the lock the moment its holder lets go, takes over a lock nobody has touched for 10 seconds, and refuses only when another Claude Code keeps holding it.

  • Add session shows the model and effort your agent will actually start on. The model picker read only the agent's built-in defaults, so Claude Code could show effort low while the session started on the effort set in your settings. Add session now reads the model and per-model effort from Claude Code's settings.json and the model and reasoning effort from Codex's config.toml, without running the agent and without writing either file.

  • The dashboard no longer jumps when a notice appears under the header. Update, telemetry, old report and error notices now float over the top of the dashboard instead of pushing every session down, and dismissing one leaves everything where it was.

  • A Glimmervoid Cloud GitHub token that GitHub refuses now says why in the doorman log. The log line used to end at the status code, such as GitHub answered 422; it now carries GitHub's own reason, such as a permission the installation has not accepted yet or a repository it no longer reaches.

  • A Glimmervoid Cloud server no longer floods its doorman when a GitHub token request fails. Every git push, gh call and settings read used to ask again at once, several times a second; the server now waits a growing, slightly randomized interval of up to five minutes between attempts, and longer when the doorman sends a Retry-After.