Skip to main content

Release Notes

New features, improvements, and fixes in each CoderFlow release, newest first. To update your server, open Server Administration → Health and use Check for Updates.

0.14.1 — August 25, 2026

Added

  • An environment quick switch. Ctrl+E (Cmd+E on macOS) — or a bare e when you are not typing in a field — opens a palette on the dashboard and the environments page where typing an environment's two-letter avatar code, or its name, switches to it. A code only one environment answers to switches immediately; when codes collide the palette stays open so you can pick.
  • A Container Inactivity setting in Server Administration, controlling how many hours a task container may sit idle before it is stopped. The default stays at two hours; half an hour to an hour is recommended on CPU- or memory-constrained hosts.
  • Server hosts can now bound the whole task fleet's CPU at once through a shared control group instead of capping each container. Configuring it removes the per-container ceiling, so a lone task can still burst on an idle host while the fleet as a whole stays inside its budget no matter how many containers are running.

Improved

  • CoderFlow stays responsive when many tasks run at once. Live activity is served from state the server already keeps rather than re-read on every request, container state no longer costs a Docker round trip on every poll, task activity and API-key usage are written in batches instead of on every update, Docker is polled far less while a task's application server starts up, and Server Administration → Health now records request and operation timings so a slow response can be traced to its cause.
  • Container CPU limits no longer shrink as the agent limit rises. Deployment, test, prompt, and documentation jobs burst freely at a lower scheduling weight — finishing sooner is what frees the host — and agent containers get half the host's cores, never fewer than two, so a task's own build steps still run in parallel.
  • A failed approval push now shows the remote's own rejection in the Approve Changes dialog, with the full git output behind a Show full git output disclosure, instead of only "Push failed". A refusal that means the connection lacks write access to the repository is named as such and kept separate from an expired sign-in, since reconnecting cannot fix a missing permission.
  • The AI Judging dialog opens faster and shows its agent list loading, instead of briefly settling on whichever judge happened to be listed first.
  • The environment identity chip at the top-left now gets out of the way of every full-screen surface — a maximized composer, task section, file diff, file viewer, or the architecture diagram — as it already did for drawers and dialogs. The environment rail's tooltip is back as well, and a container badge's tooltip now reads as a titled card whose state line is tinted to match the badge, rather than three unrelated facts joined into one sentence.

Fixed

  • A task application answering with an unusual HTTP status could terminate the CoderFlow server process, dropping every connection it was serving. Such a failure is now contained to the one request, a backend that dies mid-response is reported as an error instead of delivering a truncated page that looks complete, and the built-in HTTP/2 listener is now opt-in. Installations that terminate HTTPS at an upstream Nginx or load balancer were never affected.
  • The server became slow to respond whenever several tasks were running. The account keepalive was refreshing tokens about once a minute regardless of when they actually expired, pushing credentials into every running container each time — work that, at dozens of containers, consumed a large share of every minute.
  • Provider account and token updates are now safe when more than one CoderFlow process is running, so simultaneous refreshes can no longer overwrite each other's stored accounts.
  • On phones and tablets, a downward swipe inside a scrolling panel — the file browser's tree or file viewer, tool output, the terminal — was taken as a page pull, so content could be pushed up but never brought back down, and a hard pull could silently reload the page from behind an open panel.
  • Opening the AI Judging dialog popped a tooltip over a button nobody had pointed at or tabbed to.
  • A forked task could surface a stale mid-run diagnostic copied from the task it was forked from as its Latest Update.
  • The Ask CoderFlow settings panel laid its fields out in two mismatched columns, splitting the Usage Telemetry heading from its description and pulling the Model row up beside the Enable toggle.
  • API-key usage recorded shortly before a server shutdown is now saved rather than lost.
  • Environment names in the environment theme bar had the tails of letters such as g, y, and p clipped.

0.14.0 — August 20, 2026

Added

  • The task page has a Sections panel on its left edge, beside the Tasks panel, listing every section on the page — including ones you have hidden. Each row jumps to its section, shows a status dot and a count of what changed since you last read it, and can hide or restore that section without scrolling to its header. Rows can be dragged (or moved with Alt+Arrow keys) to reorder the cards on the page itself, so the order you set in the panel is the order the task page uses. The panel's closed tab carries the unread total, and O toggles it.
  • Personal Automations in Profile Settings. You can now schedule your own recurring tasks — daily, weekly, monthly, a fixed interval, or a custom cron expression — in any environment you have access to, running as you with your own permissions, secrets, AI provider account, and Git connections. Each automation can launch from custom instructions or a task template, with optional agent, model, and effort overrides. Environment automations gained a Timezone field as well, so schedules follow the zone you choose and its daylight-saving changes.
  • An interactive Architecture diagram on the environment Overview tab, showing repositories, ports, secrets, and connected services as a map you can zoom and pan, with each node linking to the settings tab that configures it. Nodes can be dragged into a custom arrangement that is remembered per environment, Reset layout puts them back, and Export downloads the diagram as an SVG or a 2× PNG. On phones and tablets the controls collapse to icons and the diagram responds to touch gestures.
  • Reply to all on the task page's follow-up box and the dashboard's Quick Reply: when a task has variants, one correction can go to every eligible sibling at once instead of being retyped tab by tab. The toggle appears only when there are siblings that can accept a follow-up, names any that are being skipped, and reports per-variant results if one send fails.
  • Dashboard tiles now carry their environment as a colored corner mark, and clicking it opens an environment menu: move an objective to a different environment, filter the board down to one environment, choose the environment for new tasks, or open its settings.
  • A Slack Add as objective message shortcut. Right-clicking a message captures the whole thread, drafts a title and summary for you to edit, and creates an objective with the full transcript — plus any thread images — attached. Files the bot cannot carry over, such as videos and PDFs, are noted in the transcript with a permalink back to the originals.
  • Grok 4.6 model support.
  • Syntax highlighting for IBM i languages — RPG (free and fixed format, including embedded SQL), DDS, CL, and COBOL — in the task file browser, diffs, Git history, and the file editor, instead of those files rendering as plain text.
  • Application ports can now be exposed on your private network. Each port in an environment's configuration chooses between CoderFlow only (the default, unchanged) and Private network with a specific host interface IP, and invalid addresses are flagged inline as you type. The setting applies to newly created task containers.
  • Published documentation sites now render Mermaid diagrams as diagrams, matching how CoderFlow already renders them in the app, instead of showing the raw code fence.
  • Pin now directly in the pin menu of a snoozed task or objective, instead of only inside the Change snooze… dialog.
  • Server Administration → Health gained rolling host CPU, memory, and I/O pressure figures with peaks, and container initialization timings, so a slow start can be traced to host pressure.

Improved

  • CoderFlow loads faster, especially on slow or high-latency connections. Content-hashed assets are now cached properly instead of being revalidated on every visit, static files no longer touch your session, pages ship route-specific bundles rather than one large module graph, supported deployments serve browser traffic over HTTP/2, optional tools such as the image editor load only when used, and the dashboard and task pages render as their data arrives rather than waiting for everything. Replaying a long task's activity history no longer holds up the rest of that page's data either.
  • Tooltips across the dashboard and task pages now read as structured cards — a heading, a muted status line, and a body — rather than several facts joined into one run of text. Keyboard shortcuts render as keycaps, the context usage meter shows a percentage over its token breakdown, and hovering a variant's agent icon shows an excerpt of that variant's own latest response, so a row of identical icons can be told apart.
  • The task launcher can be collapsed at any window size, not just on phones, and its state is remembered.
  • Server Administration → Health no longer polls every 30 seconds. Metrics are fetched when you ask for them, the card reports how old the figures are and marks them stale after five minutes, and a manual refresh keeps the current numbers on screen instead of blanking the panel. Automatic Docker cleanup is still tracked while the tab is open.
  • The middle review verdict is now labeled Changes Requested rather than "Needs Improvement", and judges are told it means a specific change is required — so acceptable work is no longer marked down for having optional suggestions.
  • Dashboard task tiles name the command an agent is running instead of showing a bare "Running", and a finished command no longer reports itself as "Completed: Error".
  • Failure banners on deployment, test, and documentation-publish tasks now name the process that failed and point at Logs, which is where those tasks record their output, instead of at Agent Activity.
  • The Bob agent has been updated for Bob Shell 2.0, which IBM shipped in August — without this, every Bob task fails once the base image picks up the new version. Bob's activity feed also no longer splits a single answer into a run of Thinking rows.
  • IBM i build-rule generation follows DDS field-reference chains: a physical or logical file's rule now carries the files its DDS borrows definitions from through REF, REFFLD, PFILE, and JFILE, at any depth, so a file rebuilds when a definition it depends on changes and its dependencies are built first. Objects no recipe can actually build — a program whose modules were all excluded, or a source member whose type cannot create the target — are now reported by name instead of having a rule written for them, and their buildable modules still get rules of their own.
  • The rebuild schedule and Build History on the environment Build tab no longer show two unlabeled clocks that disagreed. The schedule now names its own timezone ("Every day at 2:00 AM UTC") and restates the next run in your local time, and build timestamps carry a zone label with the exact instant on hover.
  • Auto-pin is honored for tasks started from Git History and for objectives created from Slack, matching tasks created everywhere else.
  • The Target Branch and Mainline Parent dropdowns in the Git History merge and cherry-pick dialog use the same pickers as the rest of the app, with working keyboard support and no clipping inside the dialog.
  • Repository names and paths are now validated where they are entered, and the values CoderFlow writes into generated build commands are quoted. The Add Repository dialog also explains what the Name and Path actually do — the Name becomes the workspace folder when Path is blank.
  • A transient agent error — a retried request or a rate limit the agent recovers from — no longer leaves a red badge on the task page's Sections rail for the rest of the run. Genuine failures still show through the task's status.

Fixed

  • Launching a task with several agents failed all but the last one during repository sync. Every task creation was rotating your GitHub credentials, which revoked the credentials the previously launched container was still using to fetch. A group now causes at most one rotation, before any container starts syncing, and provider responses that mean "rate limited" or "provider is down" are no longer treated as a revoked token.
  • Launching three or more review judges at once could fail with expired Git authentication even though the connection was healthy, because the concurrent launches raced each other's single-use refresh token.
  • Gemini tasks could fail mid-run with "agy is not authenticated" after the agent had already done real work: credential files were being rewritten underneath running agents, so a read could land on a half-written file. Credential updates are now atomic for every agent, and a failure that is not actually a credential problem is no longer reported as one.
  • Grok and Kimi tasks showed an early scrap of narration as the agent's latest response instead of its actual final answer, in both the Latest Update banner and Agent Activity.
  • The Approve button could stay disabled with "No changes to approve" after a task page first loaded, until an unrelated update happened to refresh it — which is why reloading the page usually cleared it.
  • The Changed Files card could finish loading with a stale summary when its data arrived alongside the rest of the task page.
  • Undoing a pin or unpin from the dashboard now restores exactly the state you had: a snoozed item goes back to Snoozed rather than reappearing as pinned, and undoing a pin no longer wakes snoozed items underneath it.
  • Scheduled automations and rebuilds could fire an hour early or late around a daylight-saving change.
  • Automatic container CPU and memory limits are now sized from the host as Docker sees it. When CoderFlow runs inside a container, the previous behavior could ask for more CPUs than the daemon allows and fail container creation outright.
  • Building the base image needed an interactive login, so fresh installations and hosted updates failed with "No API key is configured" and rolled back. The build now runs directly when it is already running as the server's own account.
  • Unattended web-managed updates are no longer at risk of stalling on npm's install-script prompts for the server's native dependencies.
  • Opening AI Review from the mobile More Actions sheet silently did nothing.
  • The AI Providers section of Profile Settings lost its styling in released builds, and the expired-sign-in message now names the account that failed and says what to do when a running task is pinned to the server's fallback account.
  • The Base URL and API Version fields in the Add and Edit API Key dialogs no longer shrink to their content width or sit in a mismatched box.
  • A task with several pull request links no longer breaks the layout of the task header.
  • File links written as markdown by agents that format for an IDE now open in the task's File Browser like the plain paths other agents write, instead of trying to navigate the browser to a dead relative link.
  • Tooltips no longer feel aggressive: they wait their turn instead of appearing instantly after a click, dismiss when you press, never pop up over a menu they just opened, and no longer appear on a control the page focused by itself while loading. Tapping on a touch screen no longer opens a tooltip at all, and the dashboard no longer autofocuses its search box on phones and tablets, which popped the on-screen keyboard on arrival.
  • Continuing an imported task immediately after it finished restoring could be refused as "still being restored".
  • Repository names containing spaces no longer break environment image builds.

0.13.11 — August 10, 2026

Added

  • Tasks can now be exported and imported, so a run can be moved between CoderFlow servers or handed to someone else. More → Export Task on the task page downloads a package, and the dashboard imports one back. Only the files the task actually added, changed, or deleted travel with it, which keeps packages small by leaving out dependency folders and other untracked content. An imported task comes back in the state it was exported in — changed files, status, pinned marking, and conversation — and continuing one replays its transcript into a fresh session. Both ends are also available from the CLI as coder export and coder import.
  • An Auto (Task Classifier) option in the Claude and Codex model pickers. When it's selected, a small model reads the task instructions, rates the work as low, medium, or high complexity, and launches the task on the matching model tier — so routine work doesn't pay for the largest model and hard work isn't given the smallest. Follow-ups are classified on their own instructions. If the classification can't be made, the task launches on the middle tier rather than failing.
  • Environments can cap how many agents a single developer runs at once. Agent Limit per Developer on the environment's Overview tab applies on top of the server-wide limit; leaving it blank keeps the previous behavior.
  • Server logs are now kept on disk and can be searched by time range. The admin Logs panel gained a range picker — Recent, 1 hour, 6 hours, 24 hours, 3 days, 7 days, or a custom from/to window — with the active window, its data source, and the retention period shown beside it. History survives restarts and is kept for 14 days by default.
  • The usage dashboard gained a 1 day period, a Containers card breaking running/stopped/paused containers out with its own drilldown, and a Customize control for choosing which cards you see. Your selection is remembered.
  • A Task Tile Details preference in Profile Settings controls what dashboard task tiles show in their detail box and hover popup: the task's original prompt, its most recent prompt, or the agent's latest response.
  • Task details now show a Last Activity time.
  • The task page recognizes the agent's two standard output files: the summary opens expanded in place, and the proposed commit message opens in its own dialog with copy and raw-file actions.
  • The dashboard's instructions editor gained the full-page maximize toggle the task page's reply box already had, and a task can still be named, attached to, and launched without leaving full-page mode.
  • The branch indicator on the launcher gained a one-click control for putting every repository back on its default branch, instead of reopening the branch dialog to undo a selection.
  • The AI Judging dialog shows the winner, loser, and approval marks from the task tabs on every list of variants, so a judgment can be read against the variant you already picked. An approved variant is marked as such even when it was never starred as the winner.

Improved

  • Server Administration → Health now reports Active Users — the people with CoderFlow actually open in a browser right now — in place of the old Active Sessions tile, which counted week-old logins. The drilldown lists each user with when their visit started and when they were last seen.
  • The Docker Containers drilldown names the task and the user behind each container, links the task, lets every column be sorted, and pins its column headers flush to the top of the panel while you scroll.
  • Tasks opened from the dashboard now deep-link to exactly what was clicked — a specific variant within a group, or a specific judgment inside the AI Review dialog — and the browser's Back button steps out of the judging dialog rather than off the page.
  • Kimi has its own brand color, and Grok and Kimi are now tinted like every other provider across task cards, tabs, timelines, model pickers, comparison views, and the rest of the app, instead of falling back to grey.
  • Notifications stay up long enough to be read: dwell time now scales with the length of the message, hovering one pauses its timer, and every notification carries a dismiss button. A progress message is replaced by its own result instead of the two stacking, and notifications no longer swallow clicks aimed at whatever sits underneath them.
  • Exporting a task always ends in a result you can read, laid out as a table, rather than a message that disappears on a timer.
  • The AI Review timeline separates the actions you reach for — viewing a recommendation or its feedback — from rename, delete, and the raw run, which moved into an overflow menu.
  • When a task's Git credentials have expired, CoderFlow now points you at forking the task, which carries the uncommitted work forward automatically, instead of telling you to start over and copy files by hand.

Fixed

  • Attaching a file stored in OneDrive or SharePoint could make an upload fail outright, with a message blaming your network. Attachments are now read fully before the upload starts, so cloud-backed and recently moved files work, and a file that genuinely can't be read is reported by name before anything is sent. This affected every upload in the app — task creation, follow-ups, feedback, attachments, and skill and environment imports.
  • Screenshots and other images embedded in agent responses no longer render at their full pixel width and push the page sideways, and images that showed as broken in the Latest Update banner, judge recommendations, variant feedback, and the comparison table now load.
  • The superseded-review badge was nearly invisible in dark mode — a white icon on a near-white circle — and its icon looked like a stalled spinner. It now uses a readable slate chip and a history icon.
  • Judge tasks and groups you had unpinned could reappear pinned after a server restart. Unpinning is now saved before it's applied, so a failed save reports itself instead of leaving the two out of step, and stuck judge records left behind by an earlier failure clean themselves up.
  • Auto-review banners from a previous round no longer linger after a new judgment arrives.
  • Repeatedly opening or refreshing an admin drilldown could make Stop or Remove fire several times from one click.

0.13.10 — August 6, 2026

Added

  • Running tasks can now be followed straight from the dashboard: a compact conversation panel shows the agent's activity live and lets you send rich replies inline without opening the task page. Replies work with ACP-backed agents too, file references in the conversation open that task's File Browser at the right file, and on phones the panel no longer pops the on-screen keyboard over the conversation as soon as it opens.
  • Task views now carry their environment's color. The task page and dashboard are tinted with the environment's accent, and browser tab titles are prefixed with the environment name, so it's easier to tell work in different environments apart.
  • The Git history view now shows a CoderFlow task badge on commits that came from an approved task, linking back to the task and naming who approved it.
  • AI provider settings now warn when tasks are running on a pay-per-use API key while an OAuth account with subscription access is available, so unexpected usage-based charges are easier to catch.

Improved

  • Choosing which AI provider account tasks use is clearer: accounts show a Use for tasks button and a Used for tasks badge in place of the previous generic Active toggle.
  • The AI judging feedback list now says when the judge simply had no improvement suggestions, instead of showing the same message as when no feedback exists for a judgment.
  • The Profile Settings dialog is wider, giving its tabs and the Secrets and AI Providers forms room to breathe, and its empty states now line up with the rest of the content.

Fixed

  • Enabling HTTPS on an existing server silently broke background tasks: schedules, webhooks, and chat triggers launched containers that could no longer call back into the server. A new site_url setting lets you state the server's address explicitly and takes precedence over the address learned from past requests, and the server warns at startup when SSL is enabled but background tasks would still be handed a plain-HTTP callback URL. The installation guide's SSL section now covers this, along with certificate chains and renewal.
  • coder-server restart dropped a server configured for HTTPS back to plain HTTP — restarting, including for certificate renewal, now keeps the SSL configuration. The TRUST_PROXY environment variable now takes precedence over the config file when launching through the CLI as documented, and false-like values such as false or 0 disable the setting instead of crashing startup.
  • After updating a server, browsers holding stale cached files could show a blank dashboard, task, or environments page until the cache was cleared. Release builds now stamp asset URLs with a hash of their content, so every update reliably loads matching files.
  • Saving a task as an objective now preserves the task's attachments.
  • On phones, the All tab's pagination bar was clipped off the bottom of the page and couldn't be reached; its controls are now styled for narrow screens and always within scrolling distance.
  • The environment page's content tab bar could clip its last tabs — such as Memory and Automations — with no way to scroll to them. The scroll arrows now appear whenever the tabs overflow, and using them no longer snaps the strip back to the start.
  • Typing in the changed-files editor no longer triggers the task page's single-key shortcuts, which could open dialogs, expand quick reply, or navigate away while swallowing the typed characters.
  • Forms in narrow dialogs, such as the Create GitHub App wizard, no longer overflow horizontally and clip their inputs.

0.13.9 — August 5, 2026

Added

  • The task File Browser has a Refresh button that re-reads the container without closing the panel, so files an agent wrote while you were looking show up right away. Your folder, active filter, search, and open file are all replayed with their scroll positions, and an open editor is left alone so a refresh can never discard unsaved work.
  • Review feedback now carries a priority per item. The Send Feedback dialog lists the reviewer's points as a checklist with High/Medium/Low group toggles, so you can send only the ones you want, and the auto-review loop gained a minimum feedback priority (High and Medium by default) beside its round limit — items held back are reported in the follow-up rather than dropped silently.
  • Maximize and Restore on 28 more dialogs — build and server logs, the generated Dockerfile, Commit and Uncommitted Changes, Edit Memory, Assign Skills, Keyboard Shortcuts, the Jira composers, the attachment image viewer, template preview, permission checklists and others — chosen for dialogs holding long output, a rich-text editor, or a list to scan. Each remembers its state per page, and their panes now use the extra room instead of keeping their compact height caps.
  • A Resolve pull conflicts dialog: when pulling a setup repository from Environments or Skills stops on conflicts, you can now choose which side wins the conflicting hunks in each file and retry the pull from the dialog. Non-conflicting edits from both sides are kept.

Improved

  • AI reviewers no longer grade work down for having suggestions. A review's verdict is now independent of its improvement list — a pass can come with a long list of purely optional ideas — and reviewers are told to judge against what the task asked rather than against what they would have built, so unrequested refactors and hardening no longer push the auto-review loop into extra rounds that churn working code.
  • Dashboard stat cards now count the cards you actually see on the board rather than raw tasks: variant groups collapse to one and judge tasks fold into their group's card. With the filter bar active each badge reads "shown of total", and the Running card keeps the number of agents currently executing as a sublabel.
  • The All tab is paged by card, so a variant group and its judge always land on the same page, and the pagination bar reads in the same unit — "Showing 21-40 of 137 cards", with the task count in its tooltip.
  • More of the app makes use of the wider page layout: Settings and Environments forms pair rows into two columns on wide screens while keeping a single column on laptops and inside panels, task list and Objectives rows spread across the full row instead of leaving an empty strip down the right, and Agent Activity previews now cut at the real edge of the row rather than at a fixed character count.
  • Dialog close buttons now use the same icon everywhere, with a consistent click target and a proper accessible name for screen readers.
  • Agents are now told to leave their work uncommitted, because a repository the agent has already committed looks unchanged and drops out of review and approval — the most common cause of "No changes to approve". The rule ships in the standard Output Requirements instructions, which are now on by default, and conflict-recovery prompts repeat it. The Changes card also warns you when a repository has local commits but no uncommitted changes, and offers the command to undo them.

Fixed

  • Deleting tasks got slower the longer a browser stayed open — eventually minutes per deletion, clearing only after a browser restart — because task pages kept live update streams open for tasks that no longer existed and filled the browser's connection pool. Streams are now closed when a task is deleted, and deletion itself is safer: if removing a task's files fails, the task is kept intact instead of being announced as deleted.
  • Creating an environment without a local repository failed on released builds.
  • The Judge dialog now tracks variant status as it changes, so a variant that has just finished becomes eligible for judging without reloading the task page, and the dialog and the task tabs no longer disagree.
  • Switching environments with the Memory tab open no longer leaves the previous environment's memory toggle, index, and file list on screen — which could show memory as enabled for an environment that has it turned off. A tab near the end of the row, such as Memory, also no longer ends up clipped when the page opens with it already selected.
  • Maximizing Edit Memory now grows the content editor to fill the dialog instead of leaving a band of empty space above the footer, and the source view of markdown editors no longer clips its bottom edge — which also affected Agent Instructions, README, task templates, Skills, and the task File Browser.
  • On narrow screens, the "Keep reviewing until Pass" row in the AI Judging dialog no longer has its controls overlapping the label.
  • With the task list panel open, the task page no longer scrolls sideways or lets the panel overlap page content in the widened layout.
  • The task page no longer comes up blank when the Docker daemon is unreachable: metadata, activity, output, and diffs are read from disk, and the container badge is simply left off. An unreachable daemon still never decides a task's outcome.
  • A setup repository pull that stops on conflicts now rolls back, so your local files are left intact while you decide how to resolve it, and a commit is refused while unresolved conflict markers are still staged.

0.13.8 — August 4, 2026

Added

  • Settings navigation groups can now be collapsed. Groups start expanded, your choice is remembered between visits, and jumping to a setting inside a collapsed group opens it automatically.

Improved

  • The page content column now scales with the width of your screen instead of stopping at a fixed width, so large monitors get more usable space without content stretching edge to edge on mid-size screens. Settings and Environments keep the classic narrower column, since their forms are designed for it.
  • Task cards on the dashboard are wider, and a long task name now keeps the full first line — the status, agent, and container badges wrap below it instead of squeezing the name down to a few characters.
  • The Docker Storage breakdown now reflects how your host actually stores data: on hosts where image layers, container layers, and build cache share a single store, those categories are nested under a Layer store row instead of being listed as if they added up. Every category that contributes to the reclaimable total now shows its own share, so large consumers such as stopped container layers are named rather than hidden inside the headline figure.

Fixed

  • Automatic Docker cleanup no longer removes tagged images at any disk-pressure level. Docker considers an image unused whenever no container is running from it, which is the normal idle state for CoderFlow environment images, so the previous behavior could delete the images tasks launch from — and, on a shared host, images belonging to other software. Automatic cleanup is now limited to dangling images, unused build cache, and unused networks; reclaiming tagged images, containers, or volumes remains a deliberate action through Docker System Prune.
  • Docker storage totals were overstated on hosts using the containerd image store, where container layers and build cache are counted inside the image layer figure — one host with 187 GiB in use was reported as 292 GiB.
  • Hosts that split Docker's images and volumes across two filesystems no longer lose their Docker Storage meters and automatic disk-pressure cleanup, which were disabled because neither filesystem held all of Docker's data.

0.13.7 — August 3, 2026

Added

  • Automatic Docker storage reclamation: when the disk holding Docker's storage runs low, CoderFlow now frees space on its own — dangling images and build cache, plus unused tagged images once space is critical — at most once an hour. Containers, volumes, and the base image are never touched. A new Docker Storage section in server settings controls the behavior, and cleanup progress appears on the Server Administration dashboard.

Improved

  • The Docker Storage card now takes its capacity reading from the Docker daemon itself, so the numbers are right on hosts where image layers live outside the Docker root directory, on Docker Desktop, and when CoderFlow runs inside a container. Layers and volumes get their own rows when they sit on different filesystems, and the usage breakdown matches docker system df. The Disk Usage card reports the filesystem holding the CoderFlow data directory.
  • Pulling from Git on the Environments page now reports how many new commits came in, or that the repository was already up to date, matching what the task page already showed.

0.13.6 — July 31, 2026

Added

  • Container resource limits: a new Containers section in server settings lets administrators cap the CPUs, memory, and process count available to each task, interactive session, deployment, test, documentation publish, and prompt container. Leave the CPU and memory fields blank to use automatic limits of half the host's resources. Limits apply to new containers; existing containers keep the limits they started with.
  • Per-user Agent Preferences: each user can now save style preferences for agents — tone, phrasing, emoji use — in their profile, and agents receive them with every task. Environment rules always take precedence, and judges never see user preferences. Administrators can turn the new User Preferences entry on or off per environment in the Standard Instructions settings.
  • A "View effective instructions" option on the task page shows the exact combined instructions the agent received for that task.

Fixed

  • Long version strings no longer overflow the Server Health card on narrow screens, which previously forced the admin page to scroll horizontally on mobile.

0.13.4 – 0.13.5 — July 30, 2026

Added

  • Update channels for early builds: server administrators can list additional update channels in Update Management settings, and a channel selector then appears next to Check for Updates on the Server Administration Health tab. Channel builds install through the normal web-managed update flow and never affect servers that stay on the standard channel.
  • The terminal status bar has a new Shortcuts button that opens a keyboard-shortcut reference panel, in both the standalone and docked terminal.

Improved

  • CoderFlow-hosted repositories are now baked into environment images at build time instead of being cloned into every task container at startup, so containers start faster and no longer each hold a full copy of the repository — a significant disk saving for large repositories. The environment editor's Repos tab and Dockerfile preview reflect this. Takes effect after images are rebuilt.

Fixed

  • Several terminal issues: graphical corruption while scrolling tmux history (including rows missing their leading characters); mouse selections clearing on release instead of staying highlighted and being copied to the clipboard; and Ctrl+Shift+C now copies the terminal selection instead of opening the browser's developer tools.
  • Container ports for code-server and application servers are now bound to the local machine only, closing a gap where Docker could expose them directly to the network even on firewalled hosts. Existing containers keep their old bindings until recreated.
  • Duplicating a browser tab no longer causes problems: the file browser's filter box no longer picks up browser-autofilled values (which could show an email address and a stale "Showing 0 of 0 items" status), and abruptly closed tabs no longer leak server connections — a leak that could eventually leave new CoderFlow tabs stuck loading until site data was cleared.
  • Interrupting a Codex task now reports it as interrupted instead of failed.
  • Forking a failed task no longer misreads the original task's failure as its own and immediately reports that container initialization failed while the fork's container is in fact starting normally.
  • Containers that fall back to cloning a CoderFlow-hosted repository at startup (images built before the repository was added) no longer hang and time out after a successful clone.
  • The pin menu and the group tile's container menu now use the standard menu hover styling, with proper dark-mode colors, instead of a leftover purple tint.
  • Hints on settings forms no longer render squished into the label column.

0.13.3 — July 29, 2026

Added

  • An automatic review loop: when launching a review, enable "Keep reviewing until Pass" with a round limit and CoderFlow runs the cycle for you — feedback from non-passing judges is merged into a single follow-up, the same judges re-verify the fix, and the loop repeats until every verdict passes, the rounds run out, or a round produces no code changes. Notifications stay quiet while the loop runs, so you're notified once when it finishes rather than after every round.
  • The file browser panel can now dock beside the task window instead of floating over it. A toggle in the panel header switches between overlay and docked modes; while docked, the task content wraps to stay fully readable in the narrower column and reflows live as you resize the panel. The choice is remembered.
  • The quick reply and follow-up composers can now maximize to the full page, so composing or reviewing a long reply no longer means scrolling a small box. Escape or the same button restores the normal size.
  • Plain-text file references in agent output — such as paths with line numbers in responses and output summaries — are now clickable: activating one opens the file in the files panel and jumps to the referenced line.
  • Parent tasks now show links to their forked child tasks, with each fork's status.

Improved

  • The container run-state badge on dashboard task tiles is now a control, not just an indicator: on single-task tiles one click starts or stops the container (with a confirmation if an agent is still working), and on group tiles the badge opens a menu listing each variant's container with individual start/stop actions plus start-all and stop-all.
  • Agents now receive the task's custom detail field values as context — including on follow-ups, forks, and resubmits — so they can take that information into account.
  • The judgment timeline marks superseded judgments with a single quiet "Superseded" marker linking to the latest verification, replacing the previous cluster of overlapping badges and indent rails.
  • Obsolete models have been removed from the provider pickers.

Fixed

  • Approving changes on a long-running task no longer fails to push because the task's Git credentials have expired; fresh credentials are issued for the approval automatically.
  • Forking a task whose selected branch never reached the remote (for example, after an approval whose push failed) no longer launches a container that cannot initialize: branch selections are validated at fork time and fall back to the configured branch when needed, with a warning when work could not be carried over. When a branch really is missing from the remote, the task error now says so plainly instead of showing a generic sync failure.
  • With group completion notifications enabled, the "group finished" notification is no longer sent before automatic judges have launched; it now waits until every variant has fully settled.
  • Automation runs no longer stay stuck at "running" after their task completes if a transient error interrupted recording the result — finalization is now retried automatically.
  • Dashboard task tiles now show the newest review verdicts first.

0.13.1 – 0.13.2 — July 26, 2026

Added

  • Environments can now use CoderFlow-hosted Git repositories: choose "Host in CoderFlow" when creating an environment or adding a repository, then launch tasks and approve changes without configuring an external Git provider.
  • Git History shows branches, commits, and diffs for CoderFlow-hosted repositories. This view is read-only because these repositories do not have a working tree on the server, so merge, cherry-pick, and revert actions are unavailable.
  • Removing a CoderFlow-hosted repository from an environment retains its history, and adding the same repository name again reattaches it. Deleting the environment permanently removes all of its hosted repositories, including detached ones. Back up the server data directory to protect this history; backing up coder-setup alone does not include it.
  • Server administrators can now host the Setup Repository in CoderFlow, either alongside its existing remote or as its new origin, without external provider credentials.
  • Review verdicts now appear directly on dashboard task tiles, with distinct passed, needs-improvement, failed, and superseded states; groups where every current review passed are highlighted at a glance.
  • Claude Opus 5 is available in the model catalog.

Improved

  • Git Providers settings now put the provider list first and show the Setup Repository in a compact status card. A guided dialog brings its hosting, authentication, and remote options together with clearer validation and consistent controls.
  • The task-tile overflow control for additional agents and judges can now be opened by touch or keyboard, dismissed with Escape, and navigated accessibly; its mobile tap target is larger without making the badge look oversized.
  • Microsoft Teams app packages now default to personal installation, so account linking opens in the expected one-to-one chat. Administrators can also set a stable Teams app ID so updated packages replace the existing app instead of creating duplicates.

Fixed

  • Multi-repository approvals now recover safely when only some pushes succeed: completed repositories are not repeated, and retained commits can be retried without creating duplicates. If the target branch moved before a push completed, CoderFlow pushes the already-created approval commit to a new branch instead of redoing the work.
  • Environment setup-script output now streams into the task log as it runs, and a failing setup script can no longer be reported as successful.
  • Agent pickers now appear with the correct configured providers in one step instead of briefly showing unavailable choices, and they remain usable if provider availability cannot be refreshed.
  • Deployment tasks now report a failed deployment script as failed instead of appearing to complete successfully.

0.13.0 — July 23, 2026

Added

  • The CLI can now reach containers retained by completed tasks: coder containers lists them alongside interactive containers (with --all and --json options), coder attach <task-id> connects to a task's retained container directly, and coder status <task-id> reports the task's state and its container's state separately, with --json output for scripting. Access follows normal task visibility and container-shell permissions.
  • Snooze multiple pinned items at once: selecting tasks or objectives on the dashboard now offers a bulk Snooze action, with an Undo that restores each item's previous snooze state.
  • The feedback widget now accepts file attachments via a new attach button, alongside the existing screenshot capture (the usual attachment count and size limits apply).

Improved

  • Task board tiles can be moved between lanes without a mouse: focus a tile (which now shows a clear focus ring) and press Shift+Left/Right Arrow, with the move announced for screen readers. On phones and tablets, tiles get a dedicated hold-to-drag handle, so dragging a card no longer fights with scrolling the board.
  • On phones, the task board's Sort action now sits in the bottom action dock next to Filter instead of occupying its own toolbar row.
  • In the resubmit dialog, the Branches section now appears below Attachments.
  • Third-party dependencies across the server, web UI, CLI, and docs tooling have been broadly updated, resolving known security advisories — including rebuilding the image editor and the Microsoft Teams integration on modern, supported libraries to clear the remaining high-severity audit findings in production dependencies.

Fixed

  • Kimi Code CLI tasks now display Edit and Write activity entries correctly — expanding them shows the file diff or written content instead of raw argument text — and rewinding or forking these tasks no longer bloats the handoff transcript with streamed argument fragments.
  • Right-clicking in the terminal no longer triggers the application's own mouse handling (such as tmux's menu) underneath the browser context menu, which could cause a paste to activate both.
  • Launching a judge from a snoozed pinned item, or resubmitting one, no longer immediately resurfaces the item in the Pinned list — the new task inherits the snooze.
  • Markdown editors' code view now follows light/dark theme switches immediately instead of keeping the previous theme.
  • Tasks that finish without any repository changes no longer produce a leftover commit message, so there is nothing stale to approve when a task turns out to require no code changes.
  • CLI SSO sign-in is more robust: the device flow now reads the API key from the server's approval response regardless of field naming, never saves an empty credential, and shows a clear error message if an approval response arrives without a key.

0.12.145 — July 21, 2026

Added

  • Environment repositories hosted on GitHub now show an "Open on GitHub" link in the repository editor.

Improved

  • The task page's pin control now matches the dashboard tiles: a single pin icon that opens a Snooze… / Unpin menu once the task is pinned, replacing the separate snooze clock icon.
  • Resubmitting a task now offers the same searchable branch picker used elsewhere, collapsed to "Same as original" by default.

Fixed

  • The mobile dashboard's section tabs no longer crowd together or truncate when the Snoozed tab is present; each tab now stacks its count above its label and the row scrolls when space runs out.
  • Returning to the dashboard from an environment page now keeps the environment you had selected.

0.12.144 — July 21, 2026

Added

  • Pinned tasks and objectives can now be snoozed until a quick preset or custom date and time, moving them to a dedicated Snoozed view before they return to Pinned automatically. An optional profile setting can show a browser notification when they return while the dashboard is open.

Fixed

  • Deploy quick actions now include environment variables from the selected deployment profile, so deployments that depend on those variables no longer fail.

0.12.142 – 0.12.143 — July 19, 2026

Added

  • Docs Hosting: publish a repository's docs folder as a hosted, searchable documentation site. Configure it per repository in the environment's repo editor (folder, branch, and public or authenticated access), then publish on demand — including from a new Publish Docs quick action on the home page — or automatically through an automation. Each repository shows its publish status with links to the build task.
  • Kimi K3 is available in the model catalog, with managed OAuth device sign-in.
  • A "Require Branch Selection" environment setting that prompts you to choose a branch before launching work. When enabled, submitting a task, staging, running tests or a deploy, or opening a terminal or VS Code asks you to pick a branch and then continues the original action automatically once you've chosen. An inline warning appears if no repositories participate in branch selection.
  • Optional compact fork handoffs, plus a chip on forked tasks showing whether a fork used a compact handoff or a full history replay — the compact handoff opens in a viewer you can read and maximize.

Improved

  • Forking a task now shows live stage-by-stage progress in a modal, with a per-stage checklist, an elapsed timer, and a Retry option if a step fails. Dismiss the modal to keep the fork running in the background, tracked by a clickable status toast.

Fixed

  • Editing a markdown file in the Changed Files editor or file browser now preserves the rest of the file's formatting, so a small change no longer produces a large spurious diff.
  • The Environment Management page keeps its URL in sync with the selected environment and tab, so deep links and browser reloads no longer snap you back to a different environment.
  • Dragging files onto the Objective editor's Instructions field now works — documents, and mixed image-and-document drops, are attached instead of being discarded.
  • Codex Bash command entries in the activity feed no longer stay stuck and unexpandable after follow-up runs.

0.12.141 — July 16, 2026

Added

  • A new "IBM i Build & Compilation" standard instruction you can enable per environment, guiding agents to build with codermake from the sources in your workspace instead of compiling manually or pulling source from the IBM i host.

0.12.140 — July 15, 2026

Added

  • A guided setup wizard for IBM i reference-app environments after import, walking you step by step through configuration with saved progress, a host picker, inline validation, and a live installation log. It includes a built-in save-file (SAVF) installer that streams progress as it runs.

Fixed

  • Environment editing now reliably tracks unsaved changes, so you're warned before navigating away with edits you haven't saved.

0.12.139 — July 13, 2026

Fixed

  • Fixed bug in Generate IBM i Build Rules tool causing failure to pick up on some RPGLE /COPY and /INCLUDE sources.

0.12.135 – 0.12.138 — July 12, 2026

Added

  • Task activity is attributed to the person who submitted each prompt, and tasks with prompts from multiple people can be filtered by submitter.
  • A Home link in the task page's floating header, so you can navigate back to the board from anywhere on the page.
  • A Release Notes link on the Server Administration version card, next to the update check.
  • GPT-5.6 Codex is available in the model catalog.

Improved

  • Agent Activity type filters are now multi-select toggles — combine chips (e.g. Prompts + Responses) to view several entry types together in chronological order.
  • Ask CoderFlow responses no longer force the view to the bottom while streaming: scroll up to read at your own pace, then use the new "Jump to latest" button to catch up.
  • The Server Version card on the Server Administration Health tab appears immediately, without waiting for the slower health metrics to load, and stays visible even if metrics fail.

Fixed

  • Dragging task cards in the Objectives view no longer intermittently aborts and opens the task page instead, and completing a drag no longer opens the dragged task.
  • Editing an external connection on environments created via git or zip import no longer silently fails.
  • Launching, resubmitting, or forking multi-agent objectives no longer carries one agent's model and effort settings over to a different agent's task.

0.12.132 – 0.12.134 — July 8, 2026

Added

  • Swimlane Board view on the task board, driven by custom set-list fields, with lane setup in Environment Settings.
  • Custom task detail fields, definable per environment, shown in the Task Details card.
  • Task detail field search on the task board — type field:value tokens (e.g. status:failed) or use the filter builder dialog, which constructs the query for you with no syntax to memorize.
  • User-controllable sorting on the front-page task board.
  • Live container run-state (started/stopped) shown on the task board.
  • "Expand hidden context" and one-click "Full" view in the Changed Files diff viewer.
  • Changed-files filtering and IBM i changed-objects visibility on the task page.
  • Searchable pagination for long branch option lists.
  • Grok 4.5 is available in the model catalog.

Improved

  • Default board stages are simplified to match the CoderFlow task lifecycle.
  • The board is horizontally scrollable, and home custom fields lay out inline.
  • The dashboard search field-filter dialog is now mobile friendly.
  • Integration-backed search fields are only offered when the integration is configured.
  • Custom-field dropdowns use the standard select control throughout.
  • The Latest Update panel has a dedicated expand/collapse button in its top-right corner. Clicking the panel contents still expands it but no longer collapses it, so you can highlight and copy text without the panel closing.
  • Clearer task search placeholder text.

Fixed

  • Dragging tasks out of the Unassigned lane in the Board view no longer fails intermittently.
  • The container run-state badge now appears reliably, including for test, deploy, and restored tasks.
  • The All tab no longer renders empty under the Status sort.
  • The per-file "Full" diff expander toggle works correctly.
  • Mobile More Actions and Share sheets no longer close when tapping inside the menu.
  • The Unified/Split toggle stays reachable when a changed file is maximized.
  • Setup-repo pulls no longer get stuck after an interrupted rebase.
  • Improved theme contrast for accessibility (WCAG), including inline code in dark mode and diff controls.