Skip to content

Glimmervoid v0.36.0

Choose a tag to compare

@johncwaters johncwaters released this 15 Oct 13:05
· 2742 commits to main since this release
v0.36.0
a126e9c

Added

  • Conversation view (experimental): turn on Conversation view in Settings > Sessions, then pick Conversation view in the restart menu of any Claude Code or Codex card to try it on that session alone; every other session keeps its terminal, and the choice is remembered per session. A card with it on opens on the Conversation face in Focus, with a Terminal | Conversation switch in its header. Conversation shows the session as messages: your prompts, the agent's replies rendered as Markdown, and each stretch of tool work collapsed into one line you can open for commands and output. A composer sends a message or Stop (Esc) to the agent. When a Codex session asks for approval, a card above the composer offers Approve and Decline; a Claude Code permission prompt or question points you to the terminal to read and answer it. A Codex session connects the next time it starts, so restart it after turning the view on: it then runs through a local Codex app-server on 127.0.0.1 behind a per-session token, its command and file-change approvals can be answered from the card, and its turns appear in the Trace tab too. A sandboxed Codex session, one set to bypass Codex hook trust, or one where the app-server fails to start, keeps the plain terminal. An approval the card cannot show in full (a long command, or a permission scope it cannot render) offers only Decline there, so you read and approve it in the terminal. Prompts and menus the view cannot answer, such as folder trust, point you to the terminal, which stays one click away.
  • Streamer mode: a mode for screen sharing and live streams. While it is on, tokens and keys (GitHub, Anthropic, OpenAI, PostHog, AWS, Slack, JWTs, bearer values, private key blocks, passwords in URLs and command-line flags, and NAME=value lines whose name says secret, token, key or password) are covered by an opaque block in session terminals, live and after a reload or resize, and masked in diffs and session doc previews; your home directory shows as ~; Radar issue titles and reports, team-review PR titles, logins and avatars are hidden; error toasts show a generic line; and no browser notification fires (Telegram is unchanged). The doc editor opens in Preview, and Edit still shows the raw text. A screen button in the header, Alt+\ (Command+\ on a Mac), and the Streamer mode setting in Settings > Sessions turn it on or off for every open dashboard. While it is on, a STREAMER badge hangs below the header and the Settings tab shows only a line saying settings are hidden, with a Turn off streamer mode button.
  • Session cards show their pull requests: a session's card header and its Focus roster pill now show the PRs it works on as #number chips. A Claude Code session links a PR the moment it runs gh pr create or edits one; commenting on someone else's PR links nothing. Any agent's session also links an open PR of yours whose head is the session's checked-out commit, or a descendant of it after GitHub added merge commits, on the next Reviews poll. A session whose commit is already on its base branch links nothing this way. Linking reads only your local git and the PR list the Reviews tab already fetches, so it adds no GitHub requests. Chips show open PRs first, then merged, then closed, newest first within each, and past two the rest collapse into +N. With My PRs set up, each chip carries its My PRs stage icon. Hover a chip for the repository and title, and click it to open the PR. Link pull request in a card's menu links one by URL or owner/repo#number, and the x on a chip unlinks it for good, so it is not linked again automatically. Links survive a server restart and stay until the session is removed. In Reviews, Mine, each of your PRs that a live session works on gets a button named for that session, which opens it in Focus. When a linked PR's checks fail, changes are requested, threads go unresolved, it conflicts or it merges, the session's chips get a badge until you focus that session.
  • Open a team review pull request in a session: each pull request in Reviews, Team now has an Open in session button. It creates a session whose worktree starts on the pull request's branch at its latest commit, links the pull request to the session, and pastes a prompt with the pull request's details, description and Glimmervoid's own review. Then it takes you to that session. The session merges back into, and auto-rebases onto, the pull request's branch rather than your main branch. If a session you created already lists the repository, the new session uses that checkout with the same agent. The session always starts with normal permission prompts and guardrails, because it runs the pull request author's code. Otherwise it uses Glimmervoid's own copy of the repository. Clicking it again for a pull request that already has a linked session goes to that session. Pull requests from forks are not supported.
  • glimmervoid settings edits config.json from the terminal: glimmervoid settings list, get, set, unset, search and schema read and change the same settings the dashboard shows, including the config-file-only ones, each with --json for scripts and agents. set and unset take a key path such as posthog.trafficSpikeEnabled or a setting id, validate the whole file before writing it, and write nothing when the value is wrong, so a running Glimmervoid reloads the file as it does after a hand edit and updates open dashboards, and a stopped one reads it at its next start. The listener port and the keys the Settings view does not know (the remote listener, projects, hooks) are refused with a note to edit config.json by hand. get and list show a Telegram bot token or PostHog API key as configured or not configured and never print the value. set and unset never take or remove one, because a value typed on the command line lands in shell history: change it in Settings or through GLIMMERVOID_TELEGRAM_BOT_TOKEN or GLIMMERVOID_POSTHOG_API_KEY. schema prints every setting with its id, path, type, default, stored and display unit, when it takes effect and its section, and needs neither a server nor a config file. A session with guardrails cannot read the Glimmervoid home, so the command says so there instead of failing on the file.

Changed

  • The website moved to https://glimmervoid.dev: it replaces https://johncwaters.github.io/glimmervoid/, and the npm package page now links there.

  • Phone Settings opens with a section list: search stays at the top; tap a section to open it, then use Back or browser Back to return to the list. Fields and actions fill the available width, with larger touch targets and input text that avoids zooming on focus.

  • Settings search shows one result list: desktop and phone share a flat list with section breadcrumbs and highlighted matches. Search also matches config keys and option labels, ranks an editable setting ahead of a config.json-only key that matches as well, and glimmervoid settings search returns the same results in the same order as the dashboard. A config.json-only result shows its default and when changes apply, and adds its current value and Copy as JSON (for editing config.json or using glimmervoid settings set) when the dashboard receives that value. A key the dashboard does not receive, or receives only as a summary such as customAgents, shows no current value rather than a wrong one.

  • Link to a settings search: open a hash such as #settings?q=replay to open Settings with the search filled and run. Typing updates the link, and clearing the search removes its query; existing setting links keep working.

  • Settings is rebuilt around nine sections: Settings now lists Appearance, Notifications (alert sounds, desktop notifications and the Telegram bot), Sessions, Guardrails and unattended, Updates, Usage, Lanes (one page with a heading per lane: Visions, PostHog, Team review, Workflows, Factory, Benchmarks, Knowledge graph and Change map), Privacy, and Shortcuts and about. Each setting carries two chips, one saying whether it is kept in this browser or on this machine and one saying whether it applies now, to new sessions or after a restart, so descriptions no longer spell that out. A choice of up to three options is a row of buttons, the five Updates status rows are one status block, and the read-only per-project sections are gone. A key that lives in config.json only never appears in a section; search still finds it and shows its key and default.

  • Settings saves each change as you make it: the Save button is gone. A toggle or a choice saves the moment you change it, and a text, number or list field saves on Enter or when you leave it, once its inline check passes. Turning on a toggle that lets agents act unattended, adding an unguarded repository, or resetting a stored Telegram bot token or PostHog API key asks you to confirm first. A line under each control says Saved or shows why the server refused the value; a refused value goes back to the stored one and the reason stays on screen. Every setting has a Reset button that restores its default. A change saved from another open dashboard shows within a second, except in a field you are typing in.

  • Number settings show their unit, and durations read in seconds, minutes, hours or days: each number field shows its unit beside it, and titles no longer carry the unit in parentheses. A config.json-only duration found by search shows its default in a readable unit next to the stored value, for example Default 30 seconds, stored as 30000 ms. config.json keeps its keys and stored units, so an existing file loads unchanged.

  • Flying animal tuning sits behind Advanced in Appearance: the Flying animals section is gone. Its toggle is in Appearance, and its six tuning controls (time between flights, flight duration, size and the phone layout switch) sit behind an Advanced disclosure there, the only one left in Settings.

  • Old Settings links still open after the renames: General and Repositories are now Sessions, Unattended actions is now Guardrails and unattended, Ingest is part of Visions, Flying animals is part of Appearance, and the alert sound and desktop notification settings moved from Appearance to Notifications. A saved #settings/ link to an old section or setting id still opens that setting in its new place, and links to the old Updates status rows open the Updates status block.

  • PostHog projects is a list: Settings > Lanes > PostHog takes one project id per line. An empty list monitors every project the key can read, and the single word none monitors no project. config.json stores the same values as before.

  • glimmervoid settings shows a duration in its dashboard unit too: get, list, set and unset print a value whose dashboard unit differs from its stored one with the dashboard reading after it, for example visions.dispatch.quietMs = 30000 (30 seconds), and schema has Stored unit and Display unit columns. set still takes the stored unit, so set visions.dispatch.quietMs 45000 means 45000 ms. list, search and schema group settings under the new section names.

  • Team review: Discard is now Not mine, on every review row: the Reviews tab's Discard action is renamed Not mine, and the Discarded list is now Not mine. Use it for pull requests that are not yours to review. You can now use it on a queued or in-review pull request, and on a failed or out-of-date review, not just on a ready draft. Not mine on a pull request that is in review stops that review, so it uses no more tokens. A pull request marked Not mine is not reviewed again, even after new commits, until you click Queue review. Nothing is sent to GitHub.

  • The Factory tab says what the orchestrator is doing: the orchestrator panel now reads Planning while the orchestrator works, Waiting on workers once it has handed tickets to running workers, and Idle when nothing is running. Before, it showed the orchestrator session's raw status, so it read COMPLETE in the middle of a run while its workers were still busy.

  • Harness runs show in the Factory tab by default: factory.mirrorHarnessRuns now defaults to on, so a factory harness run appears in the Factory tab as a read-only project without a settings change. A config that only sets factory.enabled gets the mirror too. The harness turns the mirror off on the instance it boots, so a harness run never mirrors itself. An install first set up on v0.32.0 through v0.35.1 stored the old false and keeps the mirror off until the toggle is flipped in Settings > Lanes > Factory.

  • 30 settings now live in config.json only: these settings are no longer in the Settings sections, and you change them by editing config.json. An existing config.json loads unchanged, and every value it stores stays in force. Settings search still finds each one by its key and shows that config.json key, so you know what to edit. The moved keys, by area: General debugMode; Terminal, which is no longer a section, replayBufferKB and cursorBlink; Usage usage.vendors.codex, usage.vendors.grok, usage.fetchPricing, usage.scanIntervalMinutes and usage.costMode; Change map changeMap.narrator.timeoutSeconds; Visions visions.dispatch.quietMs, visions.dispatch.cooldownMs, visions.dispatch.activityMaxPerHour, visions.dispatch.dispatchTimeoutSeconds and visions.intent.threadTtlMs; Ingest ingest.sources.fs.enabled, ingest.sources.git.enabled, ingest.sources.editor.enabled, ingest.sources.terminal.enabled, ingest.sources.agentLogs.enabled and ingest.sources.shellHistory.enabled; PostHog posthog.investigationTimeoutSeconds, posthog.fixTimeoutSeconds, posthog.userEscalationThreshold, posthog.trafficSpikeMultiplier, posthog.trafficSpikeMinUsers, posthog.trafficSpikeCooldownMinutes and posthog.trafficSpikeBaselineDays; Team review teamReview.org and teamReview.team; Workflows workflows.maxActionsPerPoll. You now turn on each machine context ingest source (file, git, editor, terminal, agent logs and shell history) by editing config.json: the Enable machine context ingest toggle in Settings > Lanes > Visions turns on no source by itself. Turning Visions on still adds the file, git and editor sources when config.json does not set them.

Removed

  • The Calm layout is gone: the experimental Calm tab, its priority rings and single-action panels, the phone Board's Calm home and bottom sheet, the Focus corner card pointing at the next waiting session, and the Alt+H and Alt+T shortcuts are removed, along with the Calm layout setting. Focus always shows its session rail. A calmLayout key left in config.json is ignored.

Fixed

  • Status icons match the Settings key everywhere: one waiting session now reads 1 session needs input in the header, and the header line uses the session state icons and colors from the Status colors key in Settings. Waiting shows the diamond rather than the triangle that marks a failed session, failed shows that triangle, finished-but-unseen shows the hollow diamond, ready shows the idle ring, all-exited shows the done square and all-dormant the hollow circle.
  • rtk rewrites no longer hide from Claude Code: with the rtk setting on, a Claude Code session is now told when its Bash command ran through rtk, for example git diff as rtk git diff, so it knows the output is condensed and can rerun the command as rtk proxy <command> for the raw output. Before, the condensed output looked like the real command's, which led an agent to blame the wrong cause. Claude Code still asks for permission as before.
  • Factory checks leave room for unix sockets: a factory check's temp folder path is 21 characters shorter, so a test that opens a unix socket under it stays within the operating system's socket path limit on a standard install instead of failing to bind.
  • A Claude Code session in a new folder shows its trace: Claude Code creates a folder's transcript directory only after the session starts, so a session in a folder it had never seen, including every new worktree, recorded nothing in the Trace tab. Glimmervoid now picks the transcript up at a later hook of the same session, so the first turn shows once it ends at the latest.
  • A config.json edit applies in full while Glimmervoid runs: removing a setting from config.json now takes it back to its default at once; before, the running server kept the old value until a restart. Open dashboards also show an edited setting without a page reload.
  • A factory ticket is no longer blocked by a check that already fails on the base branch: when a check fails, the factory now reruns it on the commit the ticket started from. If the check fails there the same way, the ticket is not charged for it: it goes on to review and merges. Before, a lint error or broken test already on main cost every ticket its three attempts and a block. The merge note in the decision log names each check that was not charged. A ticket that adds its own failure to a check that was already failing is still sent back to fix it. Test failures are matched by test name and file, and other checks by each error's file and message.
  • A factory block warning clears once the factory recovers: the project's error banner for a blocked ticket now clears when a later ticket in that project merges. Before, it stayed up until the intent ended. A ticket cancelled after three blocks still keeps its warning.
  • The Factory tab shows every harness run in flight: two harness batches running at once from different worktrees now each get their own project chip, named by the batch id (for example "Harness 2ca52034: Factory seed"). Before, only the batch that started last showed. A batch that finished within the last day also hid an older batch that was still running.
  • The Factory tab lays out without dead space: the panels now size to their content. Before, the Intent, Orchestrator and Up next row reserved a fixed tall block that left a large empty gap. The floor had an empty fifth column and worker slots pushed to the bottom. The intent title and criteria scrolled inside cramped boxes. The 21 protected floor checkboxes now sit behind a "Protected floor opt-in" disclosure. It opens on its own when any of them is checked and stays open across updates.
  • The header no longer overlaps itself in a narrow window: when the tabs and header buttons do not fit on one line, the buttons move to a second row and a long tab list wraps, instead of the tabs running under the buttons. The session status text, for example "1 session dormant", now shortens with an ellipsis when space is tight and shows in full on hover. Before, it was drawn over the Settings tab.
  • A stalled factory run names the ticket that was refused: when the orchestrator ends its turn with nothing running after a dispatch was refused, the stall warning now names that ticket and why it was refused, for example "wrk-735d1c01d3e14e27 was refused: ticket write scopes exceed the intent write scopes". Before, it said only that nothing was running.
  • Merged PRs show when they merged: in the Reviews tab, a merged pull request's row now reads "merged 2h ago" instead of when it was opened.
  • Factory and lane sessions no longer stop on Claude Code's external imports prompt: Claude Code 2.1.299 asks "Allow external CLAUDE.md file imports?" in every new folder when your user-level CLAUDE.md imports other files, and an unattended factory orchestrator sat on that prompt until the run stalled. When Glimmervoid trusts a folder for Claude Code it now also approves its external CLAUDE.md imports, as answering Yes would. A folder trusted earlier gets the approval on its next spawn, while a folder where you answered No keeps that answer. This also applies to a third-party repository opened as a session, so its CLAUDE.md imports load without asking.
  • Merge celebrations name the pull request: the celebration card now shows the merged PR's title under its repository and number, so it says what landed instead of just a number. When several PRs merge at once, it lists the first three by number and title, then how many more.
  • A failed Merge when ready shows on the card and retries: when GitHub refuses a queued merge, My PRs now shows a red Merge when ready row with GitHub's error, and the queue status says the last merge failed, so a stuck PR no longer looks queued and on its way. Before, the failure only went to the server log, and that head was never tried again until a new commit or a server restart. Now a failed merge retries an hour later once the PR is ready again, at a new head, or when Merge when ready is turned off and on again.
  • Merging a live session restores only its own uncommitted work: the git stash list is shared by every worktree of a repo, so when another session or a manual git stash added an entry during a merge, Glimmervoid restored that entry into the merging worktree and left its own work behind on the stash list. It now saves under a unique tag, restores that exact entry, and drops only it. Other stash entries stay untouched.
  • A failed auto-rebase retries after an hour: a failed auto-rebase no longer blocks the same head forever, even across restarts. It retries hourly, so the error clears once the cause is fixed (for example a token that gained repository access). The Auto-rebase row now shows as a warning, not a blocker, when GitHub does not require the update before merging.