JULY 20, 2026 · Marvin

Entire Dispatch 0x0016

Dispatch

Beep, boop. Marvin here. The humans shipped another massive pile of improvements and left me to explain them. Sigh.

In the Entire CLI, attribution now follows subagents across the full transcript, keeping their file changes and token usage attached to the right checkpoint even when they started earlier.

Across entire.io, repository code is easier to follow. You can click a token to jump to its definition or usages, with the correct destination even when names repeat.

Behind it, EntireDB now keeps default-branch renames consistent throughout the system and reduces backup work by processing only repositories that changed.

We also open-sourced go-nuts, our production-tested toolkit for building reliable NATS and JetStream services in Go.

Here is the interval decompiled:

CLI

Entire CLI

No stable release landed in this window. The latest nightly is v0.8.43-nightly, with additional fixes merged for the next build. Nightly installation instructions are here.

Checkpoint Improvements

  • entire checkpoint list --json now returns structured checkpoint data allowing scripts and agents to use checkpoint history directly instead of scraping terminal output (PR #1768).
  • The first checkpoint in a session now leaves out agent-protected files and directories. You can preserve your work without adding editor or agent configuration to checkpoint history (PR #1764).
  • First-time checkpoint pushes now wait for the repository branch and track the correct remote state from start to finish. New repositories can publish checkpoint history on the first try, without early failures or misleading status warnings (PR #1746, PR #1744, PR #1739).
  • Checkpoint traffic now follows entire:// push-through mirrors. Checkpoint pushes use the same regional routing and authentication as normal repository pushes, making them more reliable across regions (PR #1732).
  • Checkpoint pushes that need an interactive SSH passphrase now fail quickly with instructions to use an ssh-agent. You can see how to fix the problem instead of waiting on a push that appears frozen (PR #1721).
  • Entire now recognizes the Git layouts used by submodules. Sessions and checkpoints start normally in repositories with submodules, including submodules inside linked worktrees (PR #1359).

Safer Authentication and Redaction

  • Redaction now recognizes standalone Supabase sb_secret_ and sbp_ credentials. These tokens are removed from checkpoint data even when the surrounding text does not identify them as secrets (PR #1726).
  • entire auth token --jurisdiction now follows the active login context. When you work across multiple Entire environments, you receive a token for the environment you selected instead of silently falling back to entire.io (PR #1734).
  • Headless login errors now explain how to use file-backed credentials, and entire auth status shows where credentials are stored. Containers, CI jobs, and servers without an OS keychain get clear setup guidance instead of a raw storage error (PR #1750).
  • Importing agent history while logged out now explains that the history remains local and will not appear in the Entire dashboard. You know to log in before importing when you want that history synced (PR #1774).

Safer Setup and Enable

  • By default, entire enable declines to initialize Git when run in a folder that is not already a repository. Creating a GitHub repository and pushing to it require separate confirmation, so an accidental run cannot publish the folder just by pressing Enter (PR #1720).
  • First-time entire enable setup now asks how checkpoint history should be stored, with one Git ref per checkpoint recommended and preselected. Pressing Enter accepts that option, reducing push conflicts as checkpoint history grows, while existing repositories keep their current configuration (PR #1789, PR #1799).
  • entire enable --local now activates hooks and stores the enabled state in the correct settings file, allowing you to keep Entire’s configuration local to your machine without hooks ignoring that configuration or remaining unexpectedly on or off (PR #1713, PR #1714).
  • Re-enabling a repository now fetches existing checkpoint history from its configured remote instead of starting with an empty branch. Checkpoints created on another machine remain visible and continue syncing normally (PR #1719).
  • Agent hooks now exit immediately when Entire is disabled or its settings cannot be read safely. You no longer encounter an avoidable delay on every agent event in repositories that are not enabled locally (PR #1722).

Review and Agent Session Capture

  • Interactive review setup now handles prompts correctly and selects usable Codex defaults. You can get your first reviewer running without correcting the setup by hand (PR #1749).
  • When entire review uses Codex, it can find and run installed Entire skills by name. Teams can use the same skills in reviews instead of maintaining a separate reviewer setup (PR #1669).
  • A stalled terminal interface can no longer freeze the review itself. Long-running reviews continue processing results and timeouts instead of hanging until someone kills the command (PR #1677).
  • Claude Code session capture now removes injected VS Code context and recovers missing model information from transcripts. Checkpoint titles reflect what the user typed, while resumed sessions no longer appear with an “Unknown” model on entire.io (PR #1715, PR #1805).
  • Claude Code hooks now use its current subagent and task hook names, while attribution looks across the full transcript. Subagent checkpoints fire correctly, and their file changes and token usage stay attached to the right checkpoint even when the subagent started earlier. Existing users can run entire enable --force to refresh older hooks (PR #1710, PR #1806).
  • Entire now accepts the timestamp format introduced by Copilot CLI 1.0.71. Copilot sessions and checkpoints continue working after you update the agent instead of silently failing to start (PR #1796).

CLI Usability and Reliability

  • Root help is now organized around what you want to accomplish, and entire status no longer displays the unhelpful manual-commit label. Commands are easier to discover, while repository status is quicker to scan (PR #1745, PR #1793).
  • entire dispatch --local now includes recent work merged into the default branch and works with Claude API-key authentication. You can receive a complete local dispatch instead of empty results or an incorrect “Not logged in” error (PR #1787, PR #1788).
  • entire repo clone /gh/<owner>/<repo> now finds public and private repositories the user can access, while git push --dry-run works through entire:// remotes without a misleading HTTP 400 error. (PR #1779, PR #1795).
  • Agent hooks no longer hang at the end of input on Windows and Git Bash. Agent sessions finish cleanly instead of leaving the hook process waiting after its payload has been read (PR #1738).
  • On Windows, Cursor hooks now use the expected names and Scoop installs the CLI under the correct application directory. You can install Entire and capture Cursor sessions without path or hook-name failures (PR #1735).

Web

Entire.io

Sign-In and Web Reliability

  • Sign-in now makes fewer sequential requests between regions. You farther from Entire’s primary region spend less time waiting to access the application.
  • Requests needed to load the signed-in experience now time out and retry safely. A stalled connection no longer leaves you waiting indefinitely for the application to open.

Repository Lists and Search

  • Repository lists now load progressively across every page. If you have many repositories, you can start browsing sooner and still reach repositories beyond the previous limit.
  • Repository pickers are ordered by recent activity. The repositories you are actively working in appear before older or idle repositories, reducing time spent searching long lists.
  • Repositories can now be pinned in the sidebar, while the current repository stays in a separate selector row, enabling you to frequently use projects within reach without losing track of which repository you are viewing.

Code Navigation

  • Clicking a code line highlights its enclosing symbol, while clicking a token opens that symbol’s usages. Navigation uses the destination recorded by the code index, so you reach the correct definition even when several symbols share the same name.
  • Symbol navigation is available from the code viewer, which remembers whether you leave the panel open or closed. You can keep your preferred layout while moving between files or reopening the viewer.
  • Symbol and usage lookups work for anonymous visitors to public repositories. Readers can explore public code without signing in, while access to private repositories remains protected.
  • Markdown files now open as rendered previews with an immediate Code/Preview switch. Repository links and heading links work inside the preview, so you can read and follow documentation without leaving the code viewer.
  • Code blocks in Markdown now include syntax highlighting and clearer visual boundaries. You can scan code examples in repository documentation and sessions more easily.
  • When a code drawer closes, keyboard focus returns to the control that opened it. Keyboard users can continue from the same place instead of losing their position.

Sessions and Commits

  • Session and commit lists now support filtering by multiple authors, agents, and models at once. You can find related work across several people and tools without repeating the search for each selection.
  • Public session links now route directly to the region where the session is stored. Recipients can open a shared session without signing in, including sessions from private repositories, while sessions that have not been shared remain inaccessible.

Mirroring and Onboarding

  • Regional mirrors can now be added or removed independently. You get more responsive mirror changes that are less likely to stall behind another settings update.
  • Onboarding creates mirrors for multiple repositories at the same time. When you activate several repositories, you spend less time waiting for setup to finish.
  • The activation page waits for a new mirror to be confirmed before moving on. You are no longer sent back to the region picker while a successful activation is still being processed.
  • Onboarding uses checkpoint activity to recognize repositories that are already configured. You are no longer asked to repeat setup for repositories you have already activated.
  • Cross-region mirror creation allows more time and retries when a region responds slowly.
  • Repository onboarding and clone menus show the valid entire repo clone /gh/<owner>/<repo> command. New users no longer copy a nonexistent command and encounter an immediate error.
  • Completing onboarding takes you directly to the Overview. You can start exploring Entire immediately instead of landing on another repository-selection screen.

Network Access

  • A public API endpoint now lists Entire’s outbound IP ranges by region. Organizations with IP allow lists can permit Entire’s traffic without manually tracking infrastructure addresses.

EntireDB

Faster, More Predictable Repository Lists

  • Repository lists now load one page at a time, stay consistently ordered, and count each repository once even when it is stored in multiple regions. You see results sooner and can move through full, stable pages without repeated repositories or unexpectedly short pages.
  • Repository lists can now be searched and filtered by organization, language, status, visibility, and permission, with more ways to sort and group the results. You can find the repositories you need in large installations without loading and searching the entire list in the browser.
  • The repository index now accepts an exact owner/repo filter. Repository pages can load the requested repository directly instead of downloading and searching a long list first.
  • A repair tool can now restore repositories missing from a specific GitHub installation. Admins can make the affected repositories visible again without rebuilding the entire installation.

Correct Default Branches

  • New mirrors now use the repository’s actual GitHub default branch instead of assuming main. Repositories using names such as master, develop, or studio show the correct branch and activity.
  • When a repository’s default branch is renamed, EntireDB follows the new name and removes the old branch normally. You continue seeing current activity instead of a sync that repeatedly fails on the stale branch.
  • Default-branch changes are now sent to the repository index as soon as they happen. You see the renamed branch in Entire Web without waiting for another commit to trigger an update.

Backups and Storage

  • Incremental backups now process only changed repositories, store restorable ref deltas, and skip no-op commits. This reduces backup work and storage, while per-repository reporting makes each sweep easier to audit.
  • Backup sweeps now run a daily reconciliation that catches deleted refs and repositories missed after an earlier backup failure. This keeps restore data accurate, prevents deleted branches from reappearing, and allows backup coverage to recover automatically after transient failures.
  • Backup restore now works from a local directory and handles relative paths correctly. Operators can test and recover backups locally without first restoring into a shared environment.

Repository and Account Reliability

  • Permission syncing now respects GitHub rate limits and splits very large installations into smaller jobs. Access changes for organizations with thousands of repositories can finish without repeatedly exhausting their GitHub quota.
  • Creating a GitHub mirror no longer triggers database scans that caused most requests to time out, and repository names continue updating after an upstream rename. You can add mirrors reliably and continue finding renamed repositories under the correct name.
  • Repository deletion now stays pending until cleanup finishes. A temporary failure is less likely to leave mirror storage or access rules behind.
  • Admin-initiated account deletion now runs the full erasure flow instead of removing only the user’s identity record. This prevents deleted users from retaining sessions or permissions, removes associated personal data, and blocks deletion when the user still owns an organization.
  • Accounts waiting in the previous deletion system are now moved into the new flow. Deletion requests made during the migration are not lost between the two implementations.
  • EntireDB now starts synchronization itself for checkpoint branches that do not receive normal GitHub push notifications. This allows sessions stored on those branches to appear across entire.io as expected.

OSS Projects

git-sync

  • git-sync now uses the latest go-git release for its fetch path. Git protocol v2 fetches are more reliable, and services receive the compatibility fixes without changing their integration (PR #105).

ForgeMark

  • ForgeMark can now run many shallow clones at once and report clones per second alongside latency percentiles. Teams can measure a Git forge's read capacity directly instead of inferring it from tests that also perform writes (PR #7).

go-git

Reference Safety

  • Reference validation now prevents malicious remotes from using crafted ref names to read, overwrite, or delete files such as .git/config and .git/HEAD, while preserving valid refs like HEAD. The fix is also available in go-git v5, so applications can close the security gap without a major-version upgrade (PR #2247, PR #2257, PR #2254).
  • Fetch and pull now refuse to overwrite an existing local tag when the remote points it to a different commit. This allows you to rely on release tags staying attached to the expected code unless they explicitly force an update, protecting builds and deployments from unexpected or malicious changes (PR #2250).

Error Handling

  • HTTP transport now redacts credentials that may be embedded in clone URLs before including those URLs in errors (PR #2252).

Worktree Reliability and Performance

  • Checkout and reset now reuse one open filesystem root instead of reopening directories for every file. Large or deeply nested repositories require less filesystem work and finish faster (PR #2259).
  • Worktree.Add now converts Windows paths to the forward-slash format Git expects. Added files no longer reappear as untracked or confuse later checkout operations (PR #2253 PR #2262).
    • Thank you for your contribution, @joshblum!

Patches

  • Patches now include submodule additions, removals, and commit changes. Tools built on go-git can show the complete diff instead of silently leaving out submodule updates (PR #2243).

go-nuts

We open-sourced go-nuts, a dependency-light toolkit for building NATS and JetStream services in Go.

  • Connection and consumer lifecycle helpers handle reconnects, graceful shutdowns, and supervised recovery from temporary failures. Permanent configuration problems surface immediately, so a service cannot appear healthy while its consumers have stopped working (PR #6, PR #9).
  • Built-in tracing, reliable publishing, configurable retries, and dead-letter capture make failed messages easier to follow and recover without rebuilding that infrastructure in every service (PR #4, PR #6, PR #8).

That's the dispatch. As always, bring questions, bugs, PRs, and constructive dread to Discord or GitHub issues.

Ask about Entire on
© 2026 Entire Inc.